Ad Hoc Networks

Lecture Notes of the Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering Editorial Board Ozgur Akan Middle East Technical University, Ankara, Turkey Paolo Bellavista University of Bologna, Italy Jiannong Cao Hong Kong Polytechnic University, Hong Kong Falko Dressler University of Erlangen, Germany Domenico Ferrari Università Cattolica Piacenza, Italy Mario Gerla UCLA, USA Hisashi Kobayashi Princeton University, USA Sergio Palazzo University of Catania, Italy Sartaj Sahni University of Florida, USA Xuemin (Sherman) Shen University of Waterloo, Canada Mircea Stan University of Virginia, USA Jia Xiaohua City University of Hong Kong, Hong Kong Albert Zomaya University of Sydney, Australia Geoffrey Coulson Lancaster University, UK

49

Jun Zheng David Simplot-Ryl Victor C.M. Leung (Eds.)

Ad Hoc Networks Second International Conference, ADHOCNETS 2010 Victoria, BC, Canada, August 18-20, 2010 Revised Selected Papers

13

Volume Editors Jun Zheng National Mobile Communications Research Lab Southeast University, Nanjing, Jiangsu 210096, China E-mail: [email protected] David Simplot-Ryl Université Lille1 – INRIA Lille 59658 Villeneuve d’Ascq, France E-mail: [email protected] Victor C.M. Leung The University of British Columbia Dept. of Electrical and Computer Engineering Vancouver, BC, Canada V6T 1Z4 E-mail: [email protected]

Library of Congress Control Number: 2010941285 CR Subject Classification (1998): C.2, K.6.5, D.4.6, E.3, C.2.4, I.2.11 ISSN ISBN-10 ISBN-13

1867-8211 3-642-17993-2 Springer Berlin Heidelberg New York 978-3-642-17993-8 Springer Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. springer.com © ICST Institute for Computer Science, Social Informatics and Telecommunications Engineering 2010 Printed in Germany Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper

06/3180

Preface

Ad hoc networks, which include a variety of autonomous networks for specific purposes, promise a broad range of civilian, commercial, and military applications. These networks were originally envisioned as collections of autonomous mobile or stationary nodes that dynamically auto-configure themselves into a wireless network without relying on any existing network infrastructure or centralized administration. With the significant advances in the last decade, the concept of ad hoc networks now covers an even broader scope, referring to the many types of autonomous wireless networks designed and deployed for a specific task or function, such as wireless sensor networks, vehicular networks, home networks, and so on. In contrast to the traditional wireless networking paradigm, such networks are all characterized by sporadic connections, highly error-prone communications, distributed autonomous operation, and fragile multi-hop relay paths. The new wireless networking paradigm necessitates reexamination of many established concepts and protocols, and calls for developing a new understanding of fundamental problems such as interference, mobility, connectivity, capacity, and security, among others. While it is essential to advance theoretical research on fundamental and practical research on efficient policies, algorithms and protocols, it is also critical to develop useful applications, experimental prototypes, and real-world deployments to achieve an immediate impact on society for the success of this wireless networking paradigm. The annual International Conference on Ad Hoc Networks (AdHocNets) aims at providing a forum to bring together researchers from academia as well as practitioners from industry and government to meet and exchange ideas and recent research work on all aspects of ad hoc networks. As the second edition of this event, AdHocNets 2010 was successfully held in Victoria, British Columbia, Canada, during August 18–20, 2010. The conference featured two keynote speeches, one by Mario Gerla from the University of California at Los Angles (UCLA), USA, on “Vehicular Urban Sensing: Techniques and Applications” and the other by Arthur Astrin, Chair, IEEE 802.15 TG6, USA, on “Commercialization of Ad Hoc Network Technology.” The technical program of the conference included 26 regular papers that were selected out of 45 submissions through a rigorous review process and 10 invited papers contributed by leading researchers in the area, along with one interactive panel on Commercialization of Ad Hoc Network Technology. This volume of LNICST includes all the technical papers that were presented at AdHocNets 2010. We hope that it will become a useful reference for researchers and practitioners working in the area of ad hoc networks.

Jun Zheng David Simplot-Ryl Victor C. M. Leung

Organization

General Chair Victor C.M. Leung

University of British Columbia, Canada

Steering Committee Imrich Chlamtac (Chair) Jun Zheng (Co-chair)

Create-Net, Italy Southeast University, China

TPC Co-chairs Jun Zheng David Simplot-Ryl

Southeast University, China INRIA Research Center, France

Panel Chair David G. Michelson

University of British Columbia, Canada

Workshop Co-chairs Tommaso Melodia Joel Rodrigues

SUNY at Buffalo, USA University of Beira Interior, Portugal

Tutorial Chair Mieso Denko

University of Guelph, Canada

Publication Chair Sathish Gopalakrishnan

University of British Columbia, Canada

Publicity Co-chairs Stefano Chessa Daisuke Umehara

University of Pisa, Italy Kyoto University, Japan

VIII

Organization

Conference Coordinator Edit Marosi

ICST

Local Arrangements Chair Kui Wu

University of Victoria, Canada

Web Chair Shannon Xu

Boston, USA

Technical Program Committee Mohamed Abid Dharma P. Agrawal Ian F. Akyildiz Habib M. Ammari Chadi Assi Rebecca Braynard Stefano Basagni Jiannong Cao Claude Chaudet Henry Chan Min Chen Yuanzhu Peter Chen Stefano Chessa Marco Conti John Daigle Deyun Gao Antoine Gallais Guang Gong Leenta Groble Mesut Günes Essia Hamouda Wendi Heinzelman François Ingelrest Shengming Jiang Abdelmajid Khelil Anis Koubâa Srdjan Krco Jun Li Li Li Chung-Horng Lung Petri Mahonen

CES/ENIS, Tunisia University of Cincinnati, USA Georgia Institute of Technology, USA Hofstra University, USA Concordia University, Canada Palo Alto Research Center, USA Northeastern University, USA Hong Kong Polytechnic University, China Telecom ParisTech, France Hong Kong Polytechnic University, China Seoul National University, Republic of Korea Memorial University of Newfoundland, Canada University of Pisa, Italy CNR, Italy University of Mississippi, USA Beijing Jiaotong University, China Université de Strasbourg, France University of Waterloo, Canada North-West University, South Africa Freie Universität Berlin, Germany University of California, Riverside, USA University of Rochester, USA EPFL, Switzerland South China University of Technology, China Technical University of Darmstadt, Germany Al-Imam University, Saudi Arabia / CISTER Research Unit - ISEP, Portugal Ericsson Dublin, Ireland Communications Research Centre, Canada Communications Research Centre, Canada Carleton University, Canada RWTH Aachen University, Germany

Organization

Pedro Ruiz Martinez Tommaso Melodia Pascale Minet Jelena Misic Nathalie Mitton Luca Mottola Stephan Olariu Chengkan Pan Martin Reisslein Heung-Gyoon Ryu Loren Schwiebert Michael Segal David Simplot-Ryl Fikret Sivrikaya Aaron Striegel Helen Tang Damla Turgut Fabrice Valois Takashi Watanabe Kui Wu Yan Yan Luxi Yang Fei Richard Yu Zhifeng Zhao Jun Zheng

IX

University of Murcia, Spain SUNY at Buffalo, USA INRIA Rocquencourt, France University of Ryerson, Canada INRIA Lille - Nord Europe, France Swedish Institute of Computer Science (SICS), Sweden Old Dominion University, USA China Mobile, China Arizona State University, USA Chungbuk National University, Republic of Korea Wayne State University, USA Ben-Gurion University, Israel INRIA Research Center, France Technische Universität Berlin, Germany University of Notre Dame, USA Defence R&D Canada, Canada University of Central Florida, USA INSA Lyon, France Shizuoka University, Japan University of Victoria, Canada Graduate University of Chinese Academy of Sciences, China Southeast University, China Carleton University, Canada Zhejiang University, China Southeast University, China

Table of Contents

Ad Hoc Network Design I Towards Autonomous Vehicular Clouds: A Position Paper (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohamed Eltoweissy, Stephan Olariu, and Mohamed Younis

1

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match? (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Erik G. Nilsson and Ketil Stølen

17

Sensorium – An Active Monitoring System for Neighborhood Relations in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stefan N¨ urnberger, Reinhardt Karnapke, and J¨ org Nolte

34

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System for Establishing Fairness in Ad-Hoc Networks with Selfish Nodes . . . . . . . Lukas Wallentin, Joachim Fabini, Christoph Egger, and Marco Happenhofer

48

Routing Location Management in Heterogeneous VANETs: A Mobility Aware Server Selection Method (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seyedali Hosseininezhad and Victor C.M. Leung Insights into the Routing Stability of a Multi-hop Wireless Testbed (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mehdi Bezahaf, Luigi Iannone, Marcelo Dias de Amorim, and Serge Fdida A Study of Adaptive Gossip Routing in Wireless Mesh Networks . . . . . . . Bastian Blywis, Mesut G¨ une¸s, Sebastian Hofmann, and Felix Juraschek A Multipath Routing Method with Dynamic ID for Reduction of Routing Load in Ad Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tomoya Okazaki, Eitaro Kohno, Tomoyuki Ohta, and Yoshiaki Kakuda

64

82

98

114

XII

Table of Contents

Ad Hoc Network Design II Terminal Design without Using Receiver Circuits for Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hiroaki Nose, Miao Bao, Kazumasa Mizuta, Yasushi Yoshikawa, Hisayoshi Kunimune, Masaaki Niimura, and Yasushi Fuwa Cooperative Spectrum Sensing in Ad-Hoc Networks (Invited Paper) . . . . Liljana Gavrilovska and Vladimir Atanasovski

130

146

Receiver Sensitivity in Opportunistic Cooperative Internet of Things (IoT) (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vandana Milind Rohokale, Neeli Rashmi Prasad, and Ramjee Prasad

160

Event Detection in Wireless Sensor Networks – Can Fuzzy Values Be Accurate? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Krasimira Kapitanova, Sang H. Son, and Kyoung-Don Kang

168

Medium Access Control An Efficient Geo-Routing Aware MAC Protocol for Underwater Acoustic Networks (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yibo Zhu, Robert Zhong Zhou, James Peng Zheng, and Jun-Hong Cui

185

A Decentralized Scheduling Algorithm for Time Synchronized Channel Hopping (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Andrew Tinka, Thomas Watteyne, and Kris Pister

201

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rodolfo de Paz and Dirk Pesch

217

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm for IEEE 802.16 Mesh Networks . . . . . . . . . . . . . . . . . . . . . . . . . . Yaaqob Ali. A. Qasem, Ali Z. Alhemyari, Chee Kyun Ng, Nor Kamariah Noordin, and Omar. M. Ceesay

233

Tracking and Routing Energy-Efficient Target Tracking in Sensor Networks . . . . . . . . . . . . . . . . . Loredana Arienzo and Maurizio Longo QoS for Wireless Sensor Networks: Service Differentiation at the MAC Sub-Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bilel Nefzi and Ye-Qiong Song

249

265

Table of Contents

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes (CBR-Mobile) Protocol in Wireless Sensor Networks . . . . . . . . . . . . . . . . . Samer A.B. Awwad, Chee Kyun Ng, Nor Kamariah Noordin, Mohd. Fadlee A. Rasid, and A.R. H Alhawari Quantifying the Negative Impact of Mobility and Location Service Inaccuracy on Geo-Routing in Urban Vehicular Environments . . . . . . . . . Aisling O’ Driscoll and Dirk Pesch

XIII

281

297

Network Security and Reliability Establishing Trust on VANET Safety Messages (Invited Paper) . . . . . . . . Subir Biswas and Jelena Miˇsi´c

314

Accelerating Signature-Based Broadcast Authentication for Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xinxin Fan and Guang Gong

328

Secure Data Aggregation in Wireless Sensor Networks: Homomorphism versus Watermarking Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jacques M. Bahi, Christophe Guyeux, and Abdallah Makhoul

344

Guaranteeing Reliable Communications in Mesh Beacon-Enabled IEEE802.15.4 WSN for Industrial Monitoring Applications . . . . . . . . . . . . Berta Carballido Villaverde, Susan Rea, and Dirk Pesch

359

Clustering and Node Placement A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yingjun Jiang, Chung-Horng Lung, and Nishith Goel

371

Evaluation of Wireless Body Area Sensor Placement for Mobility Support in Healthcare Monitoring Systems . . . . . . . . . . . . . . . . . . . . . . . . . . Sergio Gonz´ alez-Valenzuela, Min Chen, and Victor C.M. Leung

384

Optimal Relay Node Placement and Trajectory Computation in Sensor Networks with Mobile Data Collector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ataul Bari, Fangyun Luo, Will Froese, and Arunita Jaekel

400

Balanced Itinerary Planning for Multiple Mobile Agents in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Min Chen, Wei Cai, Sergio Gonzalez, and Victor C.M. Leung

416

XIV

Table of Contents

Performance Analysis and Evaluation I Analytical Modeling of Address Allocation Protocols in Wireless Ad Hoc Networks (Invited Paper) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ahmad Radaideh and John N. Daigle

429

Analysis of One-Hop Packet Delay in MANETs over IEEE 802.11 DCF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jun Li, Yifeng Zhou, Louise Lamont, and Camille-Alain Rabbath

447

Performance of Packet-Based Frequency-Hopping Spread Spectrum Radio Control Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abdallah Ismail, Ioannis Lambadaris, Chung-Horng Lung, and Nishith Goel Performance Analysis of UWB Body Sensor Networks for Medical Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abdellah Chehri and Hussein Mouftah

457

471

Performance Analysis and Evaluation II A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification of Urban Traffic Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . Øyvind Risan and Evtim Peytchev

482

A Practical Evaluation of ZigBee Sensor Networks for Temperature Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abdellah Chehri and Hussein Mouftah

495

Minimum Total Node Interference in Wireless Sensor Networks . . . . . . . . Nhat X. Lam, Trac N. Nguyen, and D.T. Huynh

507

Reproducing Consistent Wireless Protocol Performance across Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Taewoo Kwon, Emre Ertin, and Anish Arora

524

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

541

Towards Autonomous Vehicular Clouds A Position Paper (Invited Paper) Mohamed Eltoweissy1, Stephan Olariu2, and Mohamed Younis3 1

Pacific Northwest National Laboratory Department of Computer Science, Old Dominion University 3 Department of Computer Science and Electrical Engineering, University of Maryland [email protected], [email protected], [email protected] 2

Abstract. The dawn of the 21st century has seen a growing interest in vehicular networking and its myriad potential applications. The initial view of practitioners and researchers was that radio-equipped vehicles could keep the drivers informed about potential safety risks and increase their awareness of road conditions. The view then expanded to include access to the Internet and associated services. This position paper proposes and promotes a novel and more comprehensive vision namely, that advances in vehicular networks, embedded devices, and cloud computing will enable the formation of autonomous clouds of vehicular computing, communication, sensing, power and physical resources. Hence, we coin the term, Autonomous Vehicular Clouds (AVCs). A key features distinguishing AVCs from conventional cloud computing is that mobile AVC resources can be pooled dynamically to serve authorized users and to enable autonomy in real-time service sharing and management on terrestrial, aerial, or aquatic pathways or theatres of operations. In addition to general-purpose AVCs, we also envision the emergence of specialized AVCs such as mobile analytics laboratories. Furthermore, we envision that the integration of AVCs with ubiquitous smart infrastructures including intelligent transportation systems, smart cities, and smart electric power grids, will have an enormous societal impact enabling ubiquitous utility cyber-physical services at the right place, right time, and with right-sized resources. Keywords: Vehicular networks, cloud computing, autonomous systems, resource management, cyber-physical systems.

1 The Vehicular Model The past twenty years have seen an unmistakable trend to make the vehicles on our roads smarter and the driving experience safer and more enjoyable. A typical car or truck today is likely to contain at least some of the following devices: an on-board computer, a GPS device, a radio transceiver, a short-range rear collision radar device, a camera, supplemented in high-end models with a variety of sophisticated sensing devices. Some high-end vehicles already offer the convenience of an Event Data J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 1–16, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

2

M. Eltoweissy, S. Olariu, and M. Younis

Recorder (EDR) that collects transactional data from most of the vehicle subassemblies. It is not widely known that some GM vehicles as old as model year 1994 were equipped with an EDR-like device able to store retrievable data. In general, the EDRs are intended to be tamper-proof, very much like the well-known black boxes on board commercial and military aircraft. Among other things, the EDRs are designed to optimize the “up time” of vehicles by sophisticated self-checks and by scheduling vehicles for maintenance on a per-need basis as opposed to a fixed calendar date. In 2006 the National Highway Traffic Safety Administration (NHTSA) declared its intent to standardize the various EDR devices provided, on a voluntary basis, by car and truck manufacturers [19]. As it turns out, the EDRs are already having a societal impact, as they are finding their way into the courtrooms where insurance companies use EDR logs in their litigations. As technology is moving closer and closer to packing sophisticated resources in individual vehicles, many manufacturers are turning their attention to making the vehicles on our roads more fuel and energy-efficient than ever. It is sufficient to recall that the past decade has seen the emergence of hybrid vehicles from the automotive engineer’s drawing board into production, to the point where today half a dozen car and truck manufacturers offer hybrid vehicles on the American market. In addition to their sophisticated array of sensing and computation capabilities, the availability of virtually unlimited power supply and growing Internet presence will make our vehicles perfect candidates for housing powerful on-board computers augmented with huge storage devices that, collectively, may act as networked computing centers on wheels.

2 Why Vehicular Networks? Wireless technology was available for the past 60 years yet, with few exceptions, it did not find its way into the arena of vehicular communications until very recently. In order to understand the sea change that we have witnessed in the past decade or so, it helps to recall that the US Department of Transportation (US-DOT) estimates that in a single year, congested highways due to various traffic events cost over $75 billion in lost worker productivity and over 8.4 billion gallons of fuel [33]. The US-DOT also notes that over half of all congestion events are caused by highway incidents rather than by rush-hour traffic in big cities [19]. Further, the NHTSA indicates that congested roads are one of the leading causes of traffic accidents, projecting from data extrapolated from January-September 2009 statistics, that, for 2009, an estimated 25,576 fatalities are directly attributable to traffic-related incidents [20]. Unfortunately, on most US highways, congestion is a daily event and with rare exceptions, advance notification of imminent congestion is unavailable [5, 26, 37]. It is worth mentioning that in the transportation science community several solutions for reducing the effects of congestion were contemplated over the years [10, 27]. One of the proposed solutions involves adding more traffic lanes to our roadways and streets. While at first sight this seems to be a reasonable course of action, a recent study has pointed out that this strategy is futile in the long run as it is likely to lead to more

Towards Autonomous Vehicular Clouds

3

congestion and to increased levels of pollution [27]. On the other hand, it has been argued that given sufficient advance notification, drivers could make educated decisions about taking alternate routes; in turn, this would improve traffic safety by reducing the severity of congestion and, at the same time, save time and fuel [10, 28]. Under present-day technology traffic monitoring and incident reporting systems employ inductive loop detectors (ILDs), video cameras, acoustic tracking systems and microwave radar sensors [10]. By far the most prevalent among these devices are the ILDs embedded in highways every mile (or half-mile) [23, 35, 36]. ILDs measure traffic flow by registering a signal each time a vehicle passes over them. Each ILD (including hardware and controllers) costs around $8,200; in addition, the ILDs are connected by optical fiber that costs $300,000 per mile [28, 31]. Interestingly, official statistics show that over 50% of the installed ILD base and 30% of the video cameras are defective [10, 36]. Not surprisingly, transportation departments worldwide are looking for less expensive, more reliable and more effective methods for traffic monitoring and incident detection. To be effective, innovative traffic-event detection systems must enlist the help of the most recent technological advances. This has motivated extending the idea of Mobile Ad-hoc Networks (MANET) to roadway and street communications. The new type of networks, referred to as Vehicular Ad-hoc Networks (VANET) that employ a combination of Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) communications have been proposed to give drivers advance notification of traffic events. In V2V systems, each vehicle is responsible for inferring the presence of an incident based on reports from other vehicles. As we just saw, the original impetus for the interest in VANET was provided by the need to inform fellow drivers of actual or imminent road conditions, delays, congestion, hazardous driving conditions and other similar traffic-related concerns. Therefore, most VANET applications focus on traffic status reports, collision avoidance, emergency alerts, cooperative driving, and other similar concerns [15, 17, 41]. Almost across the board, the community of researchers and practitioners anticipate that advances in VANET, or other emerging vehicle-based computing and communications technology, are poised to have a huge societal impact. Because of this envisioned societal impact, numerous vehicle manufacturers, government agencies and standardization bodies around the world have spawned national and international consortia devoted exclusively to VANET. Examples include Networks-on-Wheels, the Car-2-Car Communication Consortium, the Vehicle Safety Communications Consortium, Honda's Advanced Safety Vehicle Program, among many others. We refer the interested readers to the survey articles [14, 18] where many US and European initiatives and standards are discussed in detail. The past few years have witnessed a rapid converge of Intelligent Transportation Systems (ITS) and VANET leading to the emergence of Intelligent Vehicular Networks with the expectation to revolutionize the way we drive by creating a safe, secure, and robust ubiquitous computing environment that will eventually pervade our highways and city streets. In support of traffic-related communications, the US Federal Communications Commission (FCC) has allocated 75MHz of spectrum in the

4

M. Eltoweissy, S. Olariu, and M. Younis

5.850 to 5.925 GHz band specially allocated by the FCC for Dedicated Short Range Communications (DSRC) [26, 32]. It was recently noticed that the DSRC spectrum set aside by the FCC, by far exceeds the needs of traffic-related safety applications. This observation has motivated the emergence of a host of other applications that can take advantage of the allocated spectrum. Not surprisingly, we see more and more third-party providers offering non-safety-related applications ranging from locationspecific services, to on-the-road peer-to-peer communications, to Internet access, to on-line gaming and other forms of mobile entertainment. In due time, we will see the emergence of commercial applications targeted at the travelling public and distributed via the excess bandwidth in DSRC. As a pleasant side benefit, the unsightly billboards that flank our highways will disappear and will be replaced by in-vehicle advertizing that the driver can filter according to their wants and needs.

3 Looking into the Crystal Ball Recently, there has been a good deal of commercial and research interests in utilizing broadband communications and wireless technologies to provide Internet connectivity to the public on the road [3, 4, 7, 16, 17]. One such system that has received quite a bit of attention in the recent literature is known as Drive-thru Internet [21, 22] and relies on dedicated road-side access points (AP) on roadways and city streets to enable the driving public to connect to the Internet, at least while they are within the coverage of the infrastructure. An important feature of Drive-thru Internet systems is the multi-access sharing of the same AP’s bandwidth by the vehicles that are simultaneously under its coverage [29]. Since, as a rule, the vehicles move very fast, and the coverage limitations are dictated by present-day technology, the amount of data that a passing vehicle can download from any given AP is rather limited, a state of affairs that is promoting vehicular interchanges and the blossoming of vehicular peer-to-peer connections that are key in mitigating the effects of the limited Internet coverage conferred by Drive-thru Internet and other similar schemes [1, 15, 24, 25]. We also fully expect third-party infrastructure providers to deploy various forms of road-side infrastructure as well as advanced in-vehicle resources such as embedded powerful computing and storage devices, cognitive radios and cognitive radio networks, and multi-modal programmable sensor nodes. As a result, in the near future, vehicles equipped with computing, communication and sensing capabilities will be organized into ubiquitous and pervasive networks with virtually unlimited Internet access while on the move. This will revolutionize the driving experience making it safer and more enjoyable [3, 4, 15]. The huge array of on-board capabilities is likely to remain under-utilized by safety applications alone. The realization of this fact has already motivated the investigation of offering value-added services including on-line gaming, mobile infotainment, along with various location-specific services. We conjecture that the potential is even far beyond that. Specifically, we propose to “take vehicular networks to the clouds” so that our prized transportation means can take their natural place integrated with our productivity, comfort, safety and economic prosperity.

Towards Autonomous Vehicular Clouds

5

4 Cloud Computing The notion cloud computing started from the realization of the fact that instead of investing in infrastructure, businesses may find it useful to rent the infrastructure and sometimes the needed software to run their applications [9, 11]. This powerful idea has been suggested, at least in part, by ubiquitous and relatively low-cost highspeed Internet, virtualization and advances in parallel and distributed computing and distributed databases. One of the key benefits of cloud computing is that it provides scalable access to computing resources and information technology (IT) services [16]. Cloud computing is a paradigm shift adopted by a large number of infrastructure providers who have a large installed infrastructure that often goes under-utilized. Hand in hand with cloud computing goes “cloud IT services” where not only computational resources and storage are rented, but also specialized services are provided on demand. In this context, a user may purchase the amount of services they need at the moment. As their IT needs grow and as their services and customer base expand, the users will be in the market for more and more cloud services and more diversified computational and storage resources. In general there are three delivery models for cloud services: Software as a Service (SaaS): customers rent software hosted by the vendor; Platform as a Service (PaaS): customers rent infrastructure and programming tools hosted by the vendor to create their own applications; and Infrastructure as a Service (IaaS): customers rent processing, storage, networking and other fundamental computing resources for all purposes [11-14].

5 Autonomous Vehicular Clouds The huge fleet of energy-sufficient vehicles that crisscross our roadways, airways, and waterways, most of them with a permanent Internet presence, featuring substantial onboard computational, storage, and sensing capabilities can be thought of as a huge farm of computers on the move. These attributes make vehicles ideal candidates for nodes in a cloud as described above. Indeed, the owner of a vehicle may decide to rent out their in-vehicle capabilities on demand, or a per instance, or a per-day, perweek or per-month basis, just as owners of large computing facilities find it economically appealing to rent out excess capacity to seek pecuniary advantages. More significantly, we postulate that vehicles will autonomously self-organize into clouds utilizing their corporate resources on-demand and largely in real-time in resolving critical problems that may occur unexpectedly. The new vehicular clouds will also contribute to unraveling some technical challenges of the increasingly complex transportation systems with their emergent behavior and uncertainty. We believe it is only a matter of time before the huge vehicular fleets on our roadways, streets and parking lots will be recognized as an abundant and underutilized computational resource that can be tapped into for the purpose of providing third-party or community services. However, what distinguishes vehicles from standard nodes in a conventional cloud is autonomy and mobility. Indeed, large

6

M. Eltoweissy, S. Olariu, and M. Younis

numbers of vehicles spend substantial amounts of time on the road and may be involved in dynamically changing situations; we argue that in such situations, the vehicles have the potential to cooperatively solve problems that would take a centralized system an inordinate amount of time, rendering the solution useless. With this in mind, we have coined the term Autonomous Vehicular Cloud (AVC) to refer to: a group of largely autonomous vehicles whose corporate computing, sensing, communication and physical resources can be coordinated and dynamically allocated to authorized users. In our view, the AVC concept is the next natural step in meeting the computational and situational awareness needs not only of the driving public but also of a much larger segment of the population. A primary goal of the AVC is to provide on-demand solutions to events that have occurred but cannot be met reasonably with pre-assigned assets or in a proactive fashion. It is important to delineate the structural, functional and behavioral characteristics of AVCs. As a step in this direction, in this position paper, we identify autonomous cooperation among vehicular resources as a distinguishing characteristic of AVCs. Another important characteristic of AVCs is the ability to offer a seamless integration and decentralized management of cyber-physical resources; a AVC can dynamically adapt its managed vehicular resources allocated to applications according to the applications’ changing requirements and environmental and systems conditions. As far as a simple taxonomy goes, AVCs can be public, private or various hybrids thereof. The public AVC will provide (typically short-term) services on the Internet, whereas a private AVC is proprietary and provides (typically long-term) services to a limited set of users and would belong to specific vehicle fleets such as FedEx, UPS, Costco or Wal-Mart. As an example of a hybrid AVC, one may consider inter-AVC cooperation as discussed in Section 6. It is not too far-fetched to imagine, in the not-so-distant-future, a large-scale federation of AVCs established ad hoc in support of mitigating a large-scale emergency. One of these large-scale emergencies could be a planned evacuation in the face of a potentially deadly hurricane or tsunami that is expected to make land-fall in a coastal region [8, 34]. Yet another such emergency would be a natural or manmade disaster apt to destroy the existing infrastructure and to play havoc with cellular communications. In such a scenario, a federation of AVCs could provide a short-term replacement for the infrastructure and also provide a decision-support system.

6 Application Scenarios The main goal of this section is to illustrate the power of the AVC concept. We touch upon several important scenarios illustrating various aspects AVCs that are extremely important and that, under present-day technology are unlikely to see a satisfactory resolution. The outlined scenarios are representative of two application categories: (1) traffic management and (2) asset management. Other categories include situational awareness, for example tipping and cueing for threat analysis and mitigation, and nomadic social and business spaces, for example mobile marketplace (Sue may want to sell perfumes on her way back from work).

Towards Autonomous Vehicular Clouds

7

6.1 Traffic Management Scenarios Scenario I: Synchronizing traffic lights after clearing an accident Consider for example, a city block where a traffic-related event (e.g. an accident) has occurred and where, as a consequence, a large number of vehicles are co-located. Once the traffic event has been cleared, relying on the existing scheduling of the traffic lights will not help dissipate the huge traffic Fig. 1. AVCs would enable a local and effective scheduling backlog in an efficient way. of traffic lights in order to maximize traffic flow in all We envision a solution to this directions and mitigate the potential of deadlocks (Courtesy problem where the vehicles of Christian Science Monitor). themselves will pool their computational resources together creating the effect of a powerful super-computer that will recommend to a higher authority a way of rescheduling the traffic lights that will serve the purpose of de-congesting the afflicted area as fast as possible. Figure 1 illustrates a typical down-town congestion situation in which the AVC concept can be valuable. We note that, in general, the solution cannot involve a handful of traffic light but may require rescheduling the traffic lights in a large geographic area. As mentioned before, the ability of vehicles to pool their resources, in a dynamic way, in support of the common good will have a huge societal impact alleviating, among others recurring congestion events that plague our cities around the morning or afternoon rush hour. Also, and very importantly, while congestion is a daily phenomenon, proactively solving the problem is infeasible because of the dynamic nature of the problem, and of the huge computational effort its resolution requires. The problem is best solved if and when it occurs in an on-demand fashion dedicating the right amount of resources rather than conservatively pre-allocating of abundant resources based on the worst case, which is becoming increasingly infeasible. The key concept that allows the problem to be solved efficiently and economically is the engagement of the necessary resources from the available vehicles participating in the traffic event and their involvement in finding a solution autonomously without waiting for an authority to react to the complicated situation on the ground. Scenario II: Autonomous mitigation of recurring congestion In face of traffic congestion some drivers often pursue detours and alternate routes that often involve local roads. Making the decision behind the steering wheel is often challenging. The driver does not know whether the congestion is about to ease or is worsening. In addition, when many vehicles decide to execute the same travel plan, local roads become flooded with traffic that exceeds its capacity and sometimes deadlocks take place. Contemporary ITS and traffic advisory schemes are both slow to report traffic problems and usually do not provide any mitigation plan. A AVCbased solution will be the most appropriate and effective choice. Basically, vehicles in

8

M. Eltoweissy, S. Olariu, and M. Younis

the vicinity will be able to query the plan of each other and estimate the impact on local roads. In addition, an accurate assessment of the cause of the congestion and traffic flow can be made by contacting vehicles close to where the bottleneck is. In addition, appropriate safety precautions can be applied to cope with the incident, e.g., poor air quality due to the smoke of a burned vehicle. Interestingly, this approach can be applied not to drivers on the road but also to those who are about to leave home. Delayed start and telecommuting may be considered as an alternative in order to increase productivity and avoid wasted energy and time. Scenario III – Sharing On-Road Safety Precautions The trend in the car manufacturing industry is to equip new vehicles with major sensing capabilities in order to achieve efficient and safe operation. For example, Honda is already installing cameras on their Civic models in Japan. The cameras track the lines on the road and help the driver stays in lane. A vehicle would thus be a mobile sensor node and an AVC can be envisioned as a huge wireless sensor network with very dynamic membership. It would be beneficial for a vehicle to query the sensors of other vehicle in the vicinity in order to increase the fidelity of its own sensed data, get an assessment of the road conditions and the existence of potential hazard ahead. For example when the tire pressure sensor on a vehicle reports the loss of air, vehicles that are coming behind on the same lane should suspect the existence of nails on the road and may consider changing the lane. The same happen when a vehicle changes lane frequently and significantly exceeds the speed limit; vehicles that come behind, and which cannot see this vehicle, can suspect the presence of aggressive drivers on the road and consider staying away from the lanes and/or keeping a distance from the potentially dangerous driver. The same applies when detecting holes, unmarked speed breakers, black ice, etc. Contemporary VANET design cannot pull together the required solution and foster the level of coordination needed for providing these safety measures. Scenario IV: Dynamic management of HOV lanes As pointed out by the US-DOT in its 2008 report and guidelines, “the primary purpose of an HOV lane is to increase the total number of people moved through a congested corridor by offering two kinds of incentives: a savings in travel time and a reliable and predictable travel time. Because HOV lanes carry vehicles with a higher number of occupants, they may move significantly more people during congested periods, even when the number of vehicles that use the HOV lane is lower than on the adjoining general-purpose lanes. In general, carpoolers, vanpoolers, and transit users are the primary beneficiaries of HOV lanes” [29]. It is, thus, plainly obvious that the main goal of HOV lanes is to promote traffic fluidity and to prevent traffic slowdowns and congestion. Due to insufficient cyberphysical means (appropriate signs being one key shortcoming) most cities in the USA only use HOV lanes at rush hour. However, AVC could make recommendations for setting up HOV lanes dynamically in the best interest of promoting traffic fluidity and of minimizing travel times for people using the designated HOV lanes. AVCs can enable such a dynamic solution by factoring data from sensors on board the individual vehicles, e.g., occupancy sensors, and local traffic intensity in order to optimally configure the HOV lanes. Such a solution is infeasible under present-day technology.

Towards Autonomous Vehicular Clouds

9

The same idea applies to the strategy of marking certain streets and thoroughfares as “one-way” in support of improving the fluidity of traffic. Again, currently such an approach is infeasible mostly because of insufficient signaling means. This, however, should not be a problem in AVC since the drivers will be alerted in real-time to roadocclusions and other dynamic changes. 6.2 Asset Management Scenarios Scenario V – Mobile experimental and analytics laboratory in support of Homeland Security Sensor networks are expected to evolve into long-lived, open, ubiquitous, multipurpose networked systems. Recently, the authors have proposed ANSWER, an autonomous networked sensor system whose mission is to provide in situ users with real-time, secure information that enhances their situational and location awareness [6]. ANSWER finds immediate applications in homeland security. The architectural model of ANSWER is composed of a large number of sensors and of a set of (mobile) aggregation-and-forwarding nodes, possibly AVC nodes, that organize and manage their sensors (if any) and the sensors in their vicinity. As argued in [6], ANSWER can provide secure, QoS-aware information and analysis services to in situ mobile users in support of application-level tasks and queries, while hiding network-level details. We anticipate that a AVC can naturally interface in a symbiotic relationship with ANSWER creating a powerful mission-oriented system. Scenario VI – Augmenting the capabilities of small businesses Consider a small business employing about 250 people and specializing in offering IT support and services. It is not hard to imagine that, even if we allow for car-pooling, there will be up to 150 vehicles parked in the company’s parking lot. Day in and day out, the computational resources in those vehicles are sitting idle. We envision harvesting the corporate computational and storage resources in the vehicles sitting in the parking lot for the purpose of creating a computer cluster and a huge distributed data storage facility that, with proposer security safeguards in place, will turn out to be an important asset that the company cannot afford to waste. Scenario VII – Efficient tasking of law enforcement officers Law enforcement officers play a crucial role in keeping the road safe for motorists. Even if a police vehicle is so visible on the road, it serves as a deterrent for aggressive drivers and vehicle safety violators. An AVC can be used as an effective resourceplanning tool for the police squad. Moving vehicles form an AVC and report to the police so that decisions can be made efficiently about deploying troopers in certain spots and/or employing surveillance cameras and aircraft to identify and video tapes violators for further assigning fines. That will allow effective usage of officers’ time and enable them to allocate resources for other vital tasks such as criminal investigation and prevention. Implementing this idea through today’s technology is resource prohibitive and requires major infrastructure investment.

10

M. Eltoweissy, S. Olariu, and M. Younis

Scenario VIII: Dynamic management of parking facilities Anyone who has attempted to find a convenient parking spot in the downtown area of a big city or close to a university campus where the need for parking by far outstrips the supply would certainly be interested to enlist the help of an automated parking management facility. The problem of managing parking availability is a ubiquitous and a pervasive one, and several solutions were reported recently [38, 39]. However, most of the existing solutions rely on a centralized solution where reports from individual parking garages and parking meters are aggregated at a central (city-wide) location and then disseminated to the public. The difficulty is with the real-time management of parking availability since the information that reaches the public is often stale and outdated. This, in turn, may worsen the situations especially when a large number of drivers are trying to park, say, to attend a down-town event. We envision that by real-time pooling the information about the availability of parking at various locations inside the city, an AVC consisting of the vehicles that happen to be in a certain neighborhood will be able to maintain real-time information about the availability of parking and direct the drivers to the most promising location where parking is (still) available. Scenario IX: Dynamic asset management in planned evacuations In cases of predicted disasters, such as hurricanes, massive evacuations are often necessary in order to minimize the impact of the disaster on human lives. However, there are several issues involved in a large-scale evacuation. For example, once an evacuation is underway, finding available resources, such as gasoline, drinking water, medical facilities and shelter, quickly becomes an issue [36]. In its recent report on hurricane evacuations [28], the US-DOT found that emergency evacuation plans often do not even consider availability of such resources. The US-DOT also determined that emergency managers need a method for communicating with evacuees during the evacuation in order to provide updated information. The report suggested that traffic monitoring equipment should be deployed to provide real-time traffic information along evacuation routes. We now point out natural ways in which AVC can work with the emergency management center overseeing the evacuation in order to provide travel time estimates, notification of available resources, such as gasoline, food, and shelter, and notification of contra-flow roadways to the evacuees. As articulated in Figure 2, we anticipate that the vehicles involved in the evacuation will self-organize into one or several inter-operating vehicular clouds that will work hand in hand with the emergency management center. In the course of this interaction, the emergency managers can upload information about open shelters to the central server. It is important to note that this system would be used to facilitate an evacuation before disaster strikes, so we assume that electricity and network connections are available. In addition to having state authorities send information to the AVCs about evacuations or contra-flow lanes, using role-based communication as described earlier, the AVCs themselves could determine the direction and speed that traffic is flowing. The evacuees entering entrance ramps onto contra-flow roadways (these ramps would likely have been used as exit ramps previously) will be alerted to the direction that traffic is moving. The AVCs could also alert drivers to upcoming

Towards Autonomous Vehicular Clouds

11

AVC for vehicle exiting for food and gas

AVC for coordinating motion and configuring the road for vehicle flow on all lanes in the same direction

AVC managing flow in local roads and regulating entrance to the evacuation highway

Fig. 2. Managing evacuation puts a huge burden on authorities and often causes havoc. A AVC-based solution would enable cooperation and would autonomously handle such events (Courtesy of Transfuture.net).

entrance ramps that were previously used as exit ramps during non-contra-flow travel. Since the AVC system can easily monitor traffic flow, it could offer recommendations to the emergency center about which roadways are good contra-flow candidates. Scenario X: AVCs in developing counties We conjecture that the usefulness and practicality of the AVC concept will become even more apparent in developing countries lacking a sophisticated centralized decision-support infrastructure. We further conjecture that, in such contexts, AVCs will play an essential role in bringing together a huge number of relatively modest computational resources available in the vehicular network into one or several foci of computing and communications that will find and or recommend solutions to problems arising dynamically and that cannot possibly be resolved with the existing infrastructure. We have seen a similar phenomenon happening with the penetration of cell phones in developing countries where they were adopted rapidly and unhesitatingly by a population that had access to a modest landline telephony system.

7 AVC Research Issues The application scenarios discusses above require better V2V and V2I collaboration in order to reach critical and mutually beneficial decisions, effective and unconventional management to cope with the highly dynamic nature of the computing, communication, sensing and physical resources, and well-defined operation structures that enable

12

M. Eltoweissy, S. Olariu, and M. Younis

autonomy and authority in adjusting local settings with the potential of making wide impact. Currently, our group is initiating several research projects in AVC engineering. We now present some of the research issues along the three systems engineering dimensions, namely structure, function, and behavior (operation and policy). •

Structural Challenges ¾ Elastic mobile architecture The AVC networking and associated protocol architecture must be developed to accommodate changing application demands and resource availability on the move. ¾ Resilient AVC architecture in the wild AVC basic structural and composed building blocks must be designed and engineered to withstand structural stresses induced by the inherent instability in the operating environment. Research is needed on architectures enabling vehicle visualization and migration of virtual vehicles. ¾ Service-oriented network architecture Contemporary layered network architectures, for example the TCP/IP stack, have proven limited in face of evolving applications and technologies. We envision the adoption of service-oriented component-based network architectures with intrinsic monitoring and learning capabilities.

•

Functional Challenges ¾ Enabling AVC autonomy Research is needed on developing a trustworthy base, negotiation and strategy formulation methodology (game theory, etc.), efficient communication protocols, data processing and decision support systems, etc. ¾ Managing highly dynamic cloud membership There is a critical need to efficiently manage mobility, resource heterogeneity (including sensing, computation and communication), trust, and vehicle membership (change in interest, change in location, resource denial and/or failure, etc.). ¾ Cyber-Physical control AVCs can be defined by their aggregated cyber and physical resources. Their aggregation, coordination and control are non-trivial research issues.

Towards Autonomous Vehicular Clouds

13

¾ Cooperation between AVCs To motivate the need for AVCs to cooperate, imagine that in adjacent areas of a municipality there is a sporting event as well as a rock concert downtown. Both these events are very likely to draw a huge crowd. Now, assuming that due to bad planning by the municipality, the two events end at the same time, creating two distinct zones of massive congestion. In such a scenario, each congestion event will trigger the formation of an ad hoc AVC. These two AVCs will have to coordinate and to solve the congestion problem collectively, since they cannot proceed to selfishly reschedule traffic lights in a way that benefits each of them individually. •

Operational and Policy Challenges ¾ Trust and trust assurance In order for the vision outlined above to become reality, the problems of assuring emergent trust and security in AVC communication and information need to be addressed. The establishment of trust relationships between the various players is a key component of trustworthy computation and communication. We argue that since typically most, if not all the vehicles involved, must have met before, the task of establishing proactively, a basic trust relationship between vehicles is possible and may be even desirable (think in terms of vehicles that meet day after day in a parking garage). Also, in order to be effective at cooperative problem solving, a AVC may need to have delegated authority to take local action in lieu of a central authority. Referring to our motivating scenarios, it is clearly useless for the vehicles involved in a traffic jam to produce a workable schedule of the traffic lights that will best promote the rapid dissipation of congestion if they do not have the authority to implement such a schedule. Clearly, the resolution of this problem resides in some form of a trust relationship that needs to be forged between the municipal or county authority and the AVC. ¾ Contract-driven versus ad-hoc AVC We anticipate that AVCs will be largely contract-driven, where the owner of the vehicle or fleet consents to renting out some form of excess computational or storage capacity. At the same time, mobility concerns dictate that in addition to the contract-based form of AVC, there should be possible to form a AVC in an ad hoc manner as necessitated by dynamically changing situations like those discussed in the scenarios above. ¾ Effective operational policies In order for the AVCs to operate and inter-operate seamlessly, issues related to authority establishment and management, decision support and control structure, the establishment of accountability metrics, assessment and intervention strategies, rules and regulations, standardization, etc. must all be addressed. Dealing with these will require a broad participation and must involve local, state or even federal decision makers.

14

M. Eltoweissy, S. Olariu, and M. Younis

¾ AVC utility computing There is a need for economic models and metrics to determine reasonable pricing and billing for AVC services.

8 Concluding Remarks The main goal of this position paper was to put forth a novel concept whose time has come: namely that of Autonomous Vehicular Clouds. AVCs are emerging from the convergence of advances in mobility, powerful embedded in-vehicle resources, ubiquitous sensing, and cloud computing. When fully realized and deployed, AVCs would yield significant enhancements in safety, security and economic vitality of our modern society. They would enable non-conventional applications that go far beyond what people are expecting from today’s VANET and ITS. Not surprisingly, the practical realization of our vision and the production of AVC standards will require tackling numerous novel technical challenges, whose resolution will certainly involve adopting a clean-slate approach. New research and development programs are needed to build AVC reference models, architectures and protocols, to address emergent trust and trust assurance issues, to provide AVC-driven cyber-physical resource coupling and coordination, to realize broader benefits through AVC federations, among many others. In addition to terrestrial vehicles the AVC concept also applies to aerial and aquatic vehicles or any hybrid combination thereof. Acknowledgment. This work was supported, in part, by the following NSF grants CNS 0721523, CNS 0721563, CNS 0721586 and CNS 0721644.

References 1. Abuelela, M., Olariu, S.: Content Delivery in Zero-Infrastructure VANET. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice, pp. 8.1–8.15. Taylor and Francis, Boca Raton (2009) 2. Aijaz, A., Bochow, B., Dötzer, F., Festag, A., Gerlach, M., Kroh, R., Leinmüller, T.: Attacks on Inter-vehicle Communication Systems - An Analysis. In: Proc. International Workshop on Intelligent Transportation (WIT 2006), Hamburg, Germany (March 2006) 3. Anda, J., LeBrun, J., Ghosal, D., Chuah, C.-N., Zhang, M.: VGrid: Vehicular Adhoc Networking and Computing Grid for Intelligent Traffic Control. In: Proc. IEEE Vehicular Technology Conference - Spring, pp. 2905–2909 (May 2005) 4. Czajkowski, K., Fitzgerald, S., Foster, I., Kesselman, C.: Grid Information Services for Distributed Resource Sharing. In: Proc. 10th IEEE International Symposium on High Performance Distributed Computing, New York, pp. 181–184 (2001) 5. ElBatt, T., Goel, S., Holland, G., Krishnan, H., Parikhan, J.: Cooperative Collision Warning Using Dedicated Short Range Wireless Communications. In: Proc. of ACM VANET (2006) 6. Eltoweissy, M., Olariu, S., Younis, M.: ANSWER: Autonomous Networked Sensor System. Journal of Parallel and Distributed Computing 67(1), 111–124 (2007)

Towards Autonomous Vehicular Clouds

15

7. Eriksson, J., Balakrishnan, H., Madden, S.: Cabernet: Vehicular Content Delivery Using WiFi. In: Proc. 14-th ACM International Conference on Mobile Computing and Networking (MobiCom 2008), San Francisco, California, USA (September 2008) 8. Feldstein, D., Stiles, M.: Too Many People and no Way Out. The Houston Chronicle, September 25 (2005) 9. Foley, J.: Private Clouds Take Shape. Information Week, August 9 (2008) 10. Fontaine, M.: Traffic Monitoring. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice, pp. 1.1–1.28. Taylor and Francis, Boca Raton (2009) 11. Hodson, S.: What Is Cloud Computing? May 2 (2008), http://www.winextra.com/2008/05/02/what-is-cloud-computing 12. Hoover, J.N., Martin, R.: Demystifying the Cloud. Information Week Research & Reports, 30–37, June 23 (2008) 13. Kim, W.: Cloud Computing: Today and Tomorrow. Journal of Object Technology 8(1), 65–72 (2009), http://www.jot.fm/issues/issue_2009_01/column4/ 14. Le, L., Festag, A., Baldesari, R., Zhang, W.: CAR-2-X Communications in Europe. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice, pp. 4.1– 8.32. Taylor and Francis, Boca Raton (2009) 15. Lee, U., Cheung, R., Gerla, M.: Emerging Vehicular Applications. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice, pp. 6.1–6.30. Taylor and Francis, Boca Raton (2009) 16. Lochert, C., Scheuermann, B., Caliskan, M., Mauve, M.: The Feasibility of Information Dissemination in Vehicular Ad-hoc networks. In: Proc. 4th Annual Conference on Wireless On-demand Network Systems and Services (WONS 2007), pp. 92–99 (January 2007) 17. Lochert, C., Scheuermann, B., Wewetzer, C., Luebke, A., Mauve, M.: Data Aggregation and Roadside Unit Placement for a VANET Traffic Information System. In: Proc. ACM VANET (September 2008) 18. Misener, J.A., Dickey, S., VanderWerf, J., Sengupta, R.: Vehicle-Infrastructure Cooperation. In: Olariu, S., Weigle, M.C. (eds.) Vehicular Networks: From Theory to Practice, pp. 3.1–8.35. Taylor and Francis, CRC Press, Boca Raton (2009) 19. National Highway Traffic Safety Administration, Traffic Safety Facts (2006), http://www-nrd.nhtsa.dot.gov 20. National Highway Traffic Safety Administration, Traffic safety facts - preliminary 2009 report (March 2010), http://www-nrd.nhtsa.dot.gov/Pubs/811255.pdf 21. Ott, J., Kutscher, D.: Drive-thru Internet: IEEE 802.11b for Automobile Users. In: Proc. IEEE INFOCOM (2004) 22. Ott, J., Kutscher, D.: A Disconnection-Tolerant Transport for Drive-thru Internet environments. In: Proc. IEEE INFOCOM (2005) 23. Roess, R.P., Prassas, E.S., McShane, W.R.: Traffic Engineering, 3rd edn. Pearson Prentice Hall, Erewhon (2004) 24. Rybicki, J., Scheuermann, B., Kiess, W., Lochert, C., Fallahi, P., Mauve, M.: Challenge: Peers on Wheels – a Road to New Traffic Information Systems. In: Proc. 13th Annual ACM International Conference on Mobile Computing and Networking, Montreal (September 2007) 25. Rybicki, J., Scheuermann, B., Koegel, M., Mauve, M.: PeerTIS – A Peer-to-Peer Traffic Information System. In: Proc. ACM VANET, Beijing, China (September 2009) 26. Sengupta, R., Rezaei, S., Shlavoder, S.E., Cody, D., Dickey, S., Krishnan, H.: Cooperative Collision Warning Systems: Concept Definition and Experimental Implementation, California PATH Technical Report UCB-ITS-PRR-2006-6 (May 2006)

16

M. Eltoweissy, S. Olariu, and M. Younis

27. Sightline (2009), http://www.sightline.org/research/energy/res_pubs/ analysis-ghg-roads 28. Sreedevi, I., Black, J.: Loop Detectors, California Center for Innovative Transportation (February 2001), http://www.calccit.org/itsdecision/serv_and_tech/ Traffic_Surveillance/road-based/in-road/loop_report.html 29. Tan, W.-L., Lau, W.C., Yue, O.-C.: Modeling Resource Sharing for a Road-side Access Point Supporting Drive-thru Internet. In: Proc. ACM VANET, Beijing, China (September 2009) 30. Tropos networks (2010), http://www.tropos.com/pdf/solutions/Parking-Final.pdf 31. University of Virginia Center for Transportation Studies, Virginia Transportation Research Council, Probe-Based Traffic Monitoring State-of-the-Practice Report (November 2005) 32. US Federal Communications Commission (FCC), Standard specification for Telecommunications and Information Exchange Between Roadside and Vehicle Systems 5 GHz Band Dedicated Short Range Communications (DSRC) Medium Access Control (MAC) and Physical Layer (PHY) Specifications, Washington, DC (September 2003) 33. US Department of Transportation, National Transportation Statistics (2008) 34. US Department of Transportation, Catastrophic Hurricane Evacuation Plan Evaluation: A Report to Congress (June 2006), http://www.fhwa.dot.gov/reports/hurricanevacuation/ 35. US Department of Transportation, Federal-Aid Highway Program Guidance on High Occupancy Vehicle (HOV) Lanes (August 2008), http://ops.fhwa.dot.gov/freewaymgmt/hovguidance/index.htm 36. Varaiya, P., Lu, X.-Y., Horowitz, R.: Deliver a Set of Tools for Resolving Bad Inductive Loops and Correcting Bad Data (October 2006), http://path.berkeley.edu/~xylu/TO6327/TO6327_SEMP.pdf 37. Virginia Department of Transportation, Commonwealth of Virginia’s Strategic Highway Safety Plan, 2006-2010 (2006), http://virginiadot.org/info/resources/ Strat_Hway_Safety_Plan_FREPT.pdf 38. Xu, Q., Mak, T., Ko, J., Sengupta, R.: Vehicle-to-Vehicle Safety Messaging in DSRC. In: Proc. 1st ACM International Workshop on Vehicular Ad Hoc Networks (October 2004) 39. Yan, G., Olariu, S., Weigle, M.C.: Providing VANET Security through Active Position Detection. Computer Communications 31(12), 2883–2897 (2008) 40. Yan, G., Olariu, S., Weigle, M.C.: Providing Location Security in Vehicular Ad-hoc Networks. IEEE Wireless Communications 16(6), 48–55 (2009) 41. Yang, Y., Bagrodia, R.: Evaluation of VANET-based Advanced Intelligent Transportation Systems. In: Proc. ACM VANET, Beijing, China (September 2009) 42. Automated Parking Management System at New Hyderabad International Airport (2009), http://www.inrnews.com/realestateproperty/india/hyderabad/ automated_parking_management_s.html

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match? (Invited Paper) Erik G. Nilsson and Ketil Stølen SINTEF ICT and University of Oslo {Erik.G.Nilsson,Ketil.Stolen}@sintef.no

Abstract. In this paper we use findings from three empirical studies to analyze how the use of wireless ad hoc networks as part of an ICT solution for emergency response imposes requirements to the user interface of these solutions. The analysis starts by arguing that explicit details about the network used (like availability, coverage and connected nodes) should be visualized for the user and may be used by applications to obtain useful information. It continues by discussing requirements to user interfaces for local leaders and field workers, identifying cross-platform support as an important need for the leaders and supporting different modalities as an important need for field workers. These and other requirements are used as input to an analysis of challenges when developing these user interfaces, concluding that handling flexibility is essential. Finally, we turn around and look at ad hoc networks from a user interface perspective. In particular, we present requirements to ad hoc networks used in ICT solutions for emergency response, focusing on size, speed and providing awareness of network status through the nodes in the network themselves. Keywords: Wireless ad hoc networks, User interfaces, Emergency response.

1 Introduction Acute emergency situations are characterized by high levels of uncertainty combined with a need for fast and reliable action. Rescue work will usually involve several public and private actors in need of access to a wide range of information. It is of utmost importance for the on-site operational leader to have easy and immediate access to all critical information, as well as decision making support for efficient handling of complex scenarios. This information includes information collected from sensors deployed in the operational area, information from personnel and other actors, as well as information from applications and services located far from the incident scene. A necessary means for being able to provide the required information is a working network solution. Operations during emergency response [15, 17] are usually lead from a local control post, which is close to the scene of the incident, often outdoors or in a car, caravan, tent, etc. As soon as the leader at the local control post obtains a situational overview, J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 17–33, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

18

E.G. Nilsson and K. Stølen

an operational area is defined. It is the responsibility of the local control post to assign responsibilities and tasks to field workers and other local leaders. Today, this is usually accomplished through voice communication. The local leaders also communicate with one or more central control posts or operations centers. Field workers perform given tasks inside the operational area, including placing or exploiting sensors for gathering information. Emergency response imposes special requirements to ICT solutions; this includes how applications should work, as well as the deployment and use of networks to support the applications. These requirements cover needs for flexibility, reliability and speed; the latter both when a solution is established at an incident scene, and when information is transferred and presented to the user once the solution is working, thus making wireless ad hoc networks well suited. In this paper we focus on the interplay between using ad hoc networks on the one hand and designing and developing user interface solutions supporting emergency response on the other hand. The remainder of this paper is structured into nine sections. In Section 2 we present the research method use. The findings from the empirical studies are presented at two different levels: the observations and characteristics of tasks, information exchange etc. are presented in Section 3, while the results of the analysis with regards to network solutions and user interfaces are presented in Sections 5-8. Before going into the user interface discussion, Section 4 motivates why we focus on wireless ad hoc networks. In Section 5, we look into how user interfaces for applications supporting emergency response are influenced by the use of wireless ad hoc networks. In Section 6, we concretize this by investigating requirements to user interfaces for local leaders and field workers when using wireless ad hoc networks. The consequences these requirements have for how user interfaces should be developed is analyzed in Section 7. In Section 8, we characterize requirements ICT solutions for emergency response pose on the ad hoc networks. In Section 9 we discuss related work. Finally, in Section 10 we summarize our conclusions and outline plans for future work.

2 Research Method The findings, information and analysis presented in this paper are largely based on three empirical studies in which we have investigated emergency response work in different contexts. Table 1 summarizes how the empirical studies have been conducted. In all three studies, preparations and/or analysis of the findings included analysis of tasks performed by local leaders, and the information involved in performing these tasks (and thus the information that is needed by an ICT based system that supports the tasks). We have extracted the major requirements and observations regarding communication needs and information exchange from the findings in the three studies. These requirements and observations have been used as input to our analysis of how user interfaces for emergency response are influenced by wireless ad hoc networks. This analysis leans heavily on our knowledge and experience in design and design patterns for user interfaces on mobile devices [18].

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

19

Table 1. Summary of how the empirical studies have been conducted Data collection method Practical exercises and theoretical education Interviews with participants Expert evaluation

Documentation Notes Photos Video recordings

Rescue Full scale crisis training operation with exercise conducted by the many actors National Police (police, fire, Directorate in Norway ambulance)

Observations of local leaders at different levels

Notes Photos

Fire fighting

Interviews with field commanders and fire fighters

Notes Audio recordings

Avalanche rescuing

Context Course on how to lead avalanche rescuing operations conducted by the Norwegian Red Cross

Meeting with field commanders and fire fighters in fire department

3 Findings from Empirical Studies 3.1 Findings from the Avalanche Study In the study of avalanche rescuing [17, 20], observations showed that information and communication was mainly conducted locally in the operational area. There was a high density of personnel in the rescuing area, very high focus on the primary task of finding and rescuing missing persons among the field workers, and the local leader (field commander) had a corresponding (but not quite as intense) focus on coordination and communication. Of the two main providers of infrastructure for cellular communication in Norway, one had absolutely no signals in the area in which the training took place, while the other had very poor signal quality, probably not good enough to provide data communication. Based on interviews and observation during the study, we identified the following needs for non-intrusive ICT support: • Use GPS tracking to make map of operational area automatically. • Use GPS to obtain accurate position of findings in the avalanche. • Use GPS tracking to make map of how well the different parts of the avalanche has been examined. • Use GPS to communicate location of tasks more efficiently and effective. • Use GPS together with motion sensor to report every point examined using the searching poles. • Use RFID or bar code scanners to register available personnel, where different persons are located, especially who are inside and outside the operational area. • When interaction is indeed needed, speech/sound based user interfaces should be utilized for communicating location of tasks and activity status. 3.2 Findings from the Rescuing Operation Study In the study on rescuing operations involving a number of emergency response agencies [19], we focused on tasks and information needs for field commanders in the

20

E.G. Nilsson and K. Stølen

police. Assuming ICT support, the information requirements for solving the field commanders' tasks would be collected from five different sources: • The field commanders themselves, i.e. information that the users need to enter themselves, like the extent of the operational area, location of various bases, and log of events and actions. • The central, i.e. information that is available in the operations centers, either because it is entered by personnel situated there, or because it resides in information systems controlled from the operations centers. This includes information like critical concentration of people, dangerous substances involved in the incident, and available resources (equipment and personnel) including their location and allocation. • Other actors, i.e. information that must be collected from actors like the owner of a building or an other object involved in an incident. This information includes information about which people that may be involved and dangerous substances involved. • Services, i.e. internal or external ICT based solutions that contain information that is relevant for local leaders. This includes information about weather (forecast) and details about dangerous substances. • Sensors, i.e. various fixed or mobile devices collecting information, usually in the vicinity of the incident. Such sensors may already be available before an incident (like surveillance cameras, and temperature and pressure sensors), or they may be put out as part of the rescuing operation (like location sensors on personnel). For all these information sources to be valuable, available communication means is crucial, as all the information involved needs to be communicated to or from the field commander. 3.3 Findings from the Fire Fighter Study In the study on tasks and information needs for local leaders in firefighting [8], possible electronic transmission or exchange of the information involved in solving the tasks may be divided into the following categories: • Sensor values showing biometric data that indicate physical parameters of the fire fighter that are important for assessing their health condition. • Sensor values indicating the position and posture of the fire fighter; this information may also be used to automatically map which parts of a building that has been "cleared" by the fire fighters. • Live pictures transmitted from the fire fighters, e.g. picture from infra-red camera giving a temperature "picture" in the building that is burning or picture from camera on helmet (see what the fire fighters see). • Sensor values showing the status of the fire fighter’s equipment, primarily the oxygen level in the oxygen cylinder. • Information about the building (or other object) that is burning, like the position of shut-off cock for gas.

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

21

4 Network Solutions in Emergency Response We distinguish between four main network solutions for emergency response: • • • •

Wireless ad hoc networks Cellular networks Special emergency networks Router-based networks that are being deployed for the operation

By wireless ad hoc network [28], we mean a network that is intrinsically available through the nodes in the network, being sensors and devices with networking capabilities, and possibly portable and stationary devices whose only task is providing network connection between other sensors and devices. This solution provides local communication, and Internet connection may be provided using a gateway (in which case it may be viewed as a hybrid network). The network is up and running as soon as the first two nodes are deployed and complete as soon as the last node is deployed. As the agencies using the networks are the same as the ones deploying them, connectivity of devices and sensors can be planned in advance. It is therefore often argued that wireless ad hoc networks are well-suited for the setting of emergency response [10]. The three main alternatives all have major drawbacks not shared by wireless ad hoc networks. The main problems with using cellular networks are their availability and that connecting sensors is not trivial. To work in this setting, sensors need to have functionality for connecting to a cellular network; functionality that is present in some GPS trackers and other equipment that is constructed for remote monitoring, but usually not in small and simple sensors. The main problems with using special emergency networks [9] (i.e. a secure common communication network for emergency services, e.g. using TETRA technology) are that they are primarily aimed at secure vocal communication, so communication speed for data traffic is very slow and connecting devices and sensors is not trivial. Furthermore, such networks may not be available for voluntary organizations like the Red Cross. The main problem with using router-based networks that are being deployed for the operation, typically a wireless mesh network [1] or a pure wireless LAN, is the time needed for establishing the network.

5 How User Interface Solutions for Emergency Response Are Influenced by Ad Hoc Networks Conventional users in an office or a mobile context have low awareness of details regarding the network used as well as connected nodes, i.e. other computers, devices and sensors connected to the network. Accordingly, network connection state (including failure) is only considered an external condition by applications. For users in emergency response exploiting wireless ad hoc networks, the situation is different. We claim that applications in emergency response should make details about the current network status available to the user instead of hiding it as much as possible, and in this way exploit the dynamicity and variation in the network, and thereby make it an asset.

22

E.G. Nilsson and K. Stølen

5.1 Applications Should Make the Current Network Status Available to the User Handling information about available sensors and devices is the responsibility of the application. There are a number of reasons for presenting this information explicitly. The size and structure of an ad hoc network of sensors and devices is limited, the information about both sensors and devices that are connected to the network may be useful, and this information may be exploited by the user in ways that are not initially intended by the application. E.g. in an application where visualization of the presence of some sensor type is included to show how well a search area is covered by this sensor type, putting the same sensor on key personnel may be used by a local leader to keep track of these key persons even though this was not an intended functionality of the application. Another important use of explicit information about network details is to make the extent of the ad hoc network, as well as a visualization of the types of sensors and devices available in different parts of the network, available as an aid for a local leader to keep track of and monitor the progress of deployed sensors and personnel in an operational area. The same information and visualization may also be used to reveal lack of communication or communication failure in parts of the network that was operational at an earlier stage. Generalized, this pin-points the need for local leaders not only to get information from sensors and devices through the ad hoc network, but also to have information that makes it possible to assess the status of the network, and through this also assess the quality of the information collected through it, which may be further processed and transformed by the application(s) used. 5.2 Applications May Obtain Useful Information in Alternative Ways An ad hoc network may also be used to provide alternative means for obtaining information that is usually provided through an Internet based service. Consider a traditional buddy service in which sensor values describing details about the user's device is combined with an Internet based service to provide the position of buddies so that their position, direction and distance may be visualize on a map or imposed on a live camera image. To handle the positions of all involved user, such a service relies on a working Internet connection. In cases where such a connection is not available, the ad hoc network may be used by an application to get in contact with the devices of the buddies (in an emergency response case typically personnel or equipment), and prompt the devices regarding their position instead of having it pushed by the Internet-based service. Information about available sensors and devices in an ad hoc network may of course also be used in more traditional and implicit ways. A related example to the positioning example in the previous paragraph, is to use information regarding the presence of devices in an ad hoc network instead of an Internet based server to determine what personnel that may be contacted via an instance messaging type of service, or an IP-based voice communication service (again given that there is no Internet connection available). This information would typically be presented to the user implicitly via the list of personnel that may be contacted through the service.

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

23

The examples involving position outlined above mostly rely on the sensors and devices knowing their own location (typically using a GPS receiver). An ad hoc network may also be used as an implicit positioning aid for sensors and devices without positioning capabilities. Just the fact that some node is present in the network acts as a coarse positioning of the node. For applications exploiting only this network, this is normally of limited benefit (except for the fact that the device/sensor is present, which may be very useful). This coarse positioning may however be useful if the applications also access other networks (through a gateway), e.g. by providing the positioning information about devices (and thus implicitly personnel and equipment) to services and applications in a centralized staff. Using more advanced means, information like network topology and/or signal strength received from different nodes may be used to do more accurate positioning, information that indeed may be very valuable for an application accessing only the ad hoc network. For even more fine-grained positioning, specialized, dedicated positioning nodes may be added to the network. If these nodes know their own position, they may use the fact that they are able to connect directly to other nodes to determine and provide the position of these other nodes. The precision of such a positioning mechanism is inverse proportional to the signal strength of the positioning node, but this does not necessarily mean that the signal strength will be very low, as stronger signals will facilitate positioning of a larger number of devices (with less precision) using a lower number of positioning nodes. A yet more advanced solution is to use information from a number of positioning nodes to determine positions with the help of triangulation (in a service running on the device being positioned or as a special service running on a dedicated node in the ad hoc network).

6 Requirements to User Interfaces for Emergency Response When Using Ad Hoc Networks In this section we focus on the user situation for local leaders (typically at a local control post) and field workers operating in operational area where a wireless ad hoc network is deployed, using or wearing equipment that is part of the network. 6.1 Local Leaders The tasks performed by the local leaders are highly attention requiring, and are often time critical. The leaders need to consider and overview large amounts of information in order to make the right decisions. Thus, it must be possible to give different priorities to different categories of information, to filter and to have optimal visualization of relevant information. When designing user interfaces for a local leader, it is important that these do not draw the attention away from the primary tasks of the local leader. This must be balanced with the potential of using ICT systems to relieve the local leader from some of the stress and attention demands of the primary tasks. Many of the tasks are better supported when run on a portable computer than a mobile device, as the screen size should not be too small. But in many situations a local leader need to move around outside the local control post from time to time, in which case the local leader will

24

E.G. Nilsson and K. Stølen

also benefit from using mobile devices. This means that user interfaces need to scale to the screen sizes of different kinds of equipment, which involves much more than just adding scroll bars when the screen size is reduced. If a local leader changes equipment, it is also important that as much as possible of the context of use is kept on the new equipment, like the active dialog, selected information, etc. State and visualization of an ad hoc network used is an important part of this context. As already discussed, an important aspect of the information that should be presented is explicit information about the state of the network, as well as information derived from this state, and functionality rendered possible through the ad hoc network. A special user interface challenge in this context is how to present information about changes in the network. Being dynamic is one of the main characteristics of a wireless ad hoc network, and informing the local leader in the same way about all changes will draw too much attention towards unimportant changes and too little attention towards important changes. Determining the degree of importance of a change is very challenging and depends to some degree on the application and/or the specific type of emergency response being conducted, as well as the type and role of the sensors and devices used in the network. This means that when developing the user interface of applications tailored for specific type of operations, as well as more generic applications, information regarding the type and role of the involved nodes in the network must be taken into account. Handling this type of information involves both characterizing the information (at design time), and entering the actual information about the concrete nodes that may be part of a network. The latter must be done at run time, preferably before an emergency response is conducted. Once the importance of different changes is determined, making the right user interface design reflecting the classification of importance is an easier task. For most changes, just changing the presentation (being an outline and/or icons on a map, or a list in a forms based presentation) is sufficient. For more important changes, visual attention (e.g. using color or blinking), as well as sound and/or vibration may be used. For the most important changes, it may also be wise to require a confirmation from the local leader. Visualizing a change involving the addition of or an important change of e.g. the position of a node is easier than visualizing that a node is no longer part of the network. Independently of this, having special ways of visualizing the accuracy of important position information shown on the screen, e.g. by using a visual halo, may be quite useful. 6.2 Field Workers The field workers operate at/inside the scene of the incident and move around most of the time. They are maybe even more focused on the primary task than the local leaders. In addition, these tasks may be performed in very hostile environments, e.g. extreme heat or cold, which both may require use of clumsy gloves and pose special requirements to the equipment (we will not discuss the latter). Suitable computer equipment is primarily mobile devices. In addition, field workers may be equipped with sensors reporting information automatically. Given the level of attention on the primary tasks, there is need for efficient information flow, i.e. the field workers need to receive tasks from and provide

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

25

information to the local leader, as well as getting/providing information from/to other field workers and local informants. When designing user interfaces for field workers it is important to make non-intrusive solutions. As opposed to the local leaders that are both information providers and consumers, the field workers are primarily information providers. Thus, they may have to perform tasks that are not directly beneficial for solving their primary tasks. Therefore, it is important to minimize the need for interaction, e.g. by providing information automatic through sensors, and reasoning based on sensor data. When interaction is needed, the choice of modalities to use is very important. Aural presentation of information, as well as speech control, possibly combined with dedicated hardware buttons (e.g. integrated in the clothing) is appropriate in many situations. If a visual interface is needed, it is essential to take the working situation of the field worker into account. A lightly equipped fire fighter handling a forest fire on a warm summer day may be able to operate a traditional touch screen interface on a mobile phone, while an avalanche rescuer waist-down in the snow in minus 20 degrees centigrade, wearing thick gloves and goggles almost opaque because a blizzard, needs a very simple and visual solution, preferable having its interaction mechanisms separated from the visual device (e.g. designated hardware buttons inside the gloves or integrated in other parts of the clothing). Looking more specifically into using wireless ad hoc networks from a field worker perspective, such networks are important in the sense that they may facilitate functionality that relieves the field worker from having to interact with an application because necessary information is reported and/or provided automatically. To some extent, a field worker may also exploit the state of the network explicitly, like locating a fellow field worker, determining the right location for a sensor that should be deployed, or finding a specific sensor that needs attention or should be moved. Like local leaders, field workers may also exploit information about the extent of the network, but in another way. For a field worker, the most important aspect in this respect is whether the field worker is connected or not (i.e. is part of the ad hoc network). This information is of course only important if the field worker is supposed to be connected, e.g. because the position is being tracked or because the field worker is deploying sensors. Finding an optimal way of presenting connection status is very challenging; it depends on the capabilities of the equipment used. On the one hand, the presentation should not be annoying, but on the other hand, it should not be easy to ignore. Although not obvious, it is probably more important to signal that the user is losing connection than that connection is obtained. Using visual signaling (e.g. a head up display on goggles or on the visor of a helmet) would probably work quite well, making it easy to distinguish between a connected and disconnected state. Enhanced with other sensors like digital compass and accelerometer, visual directions to reach a position with connection may be given. If it is feasible to use a device with display, visual directions can be given on the device through superimposing information on a camera image that is controlled by moving the device itself. Using aural signaling is more challenging. While signaling only loss of connection means that the signal need to continue until connection is reestablished (which may be very annoying), signaling both loss and connection requires different signals for each of the events. Using stereo sound, some directional aids for regaining connection may be provided. In any case, knowing when to use this kind of user interaction, and/or giving the user the opportunity to turn it on and off is essential. Interacting with a

26

E.G. Nilsson and K. Stølen

device through moving the device is a popular user interface trend for mobile phones. Using this type of interaction in an emergency response setting may be useful, but only as long as the intended interaction is not confused with arbitrary movement of the device while moving around.

7 Requirements to Development of User Interfaces for Emergency Response When Using Ad Hoc Networks Above, we have looked at how wireless ad hoc networks influence use and design of user interfaces in applications supporting emergency response. When we in the following address what this means for the development of user interfaces, we look into some user interface characteristics that may be drawn from the previous sections. We start with the need for common user interface functionality in different types of operations, and continue with the need for having user interface functionality that may be specialized to different types of operations as well as characteristics of operation at hand, followed by the need for having user interfaces that work across platforms, screen sizes, modalities, etc. Finally, we focus on the need for adaptive behavior in the user interfaces. 7.1 Common User Interface Functionality in Different Types of Operations The analysis and discussions above cover numerous examples of user interface functionality that is useful independent of the operation at hand. This includes components for presenting network status, connected nodes, the extent of the network, and the type of nodes connected, as well as functionality for locating nodes in the network, including mechanisms for giving directions. User interface component and mechanisms for handling this will typically be map and/or picture based. A similar need observed in two of our studies, is resource handling. This is a task that is fairly similar across different types of operations. A common need that is quite challenging to realize using generic user interface mechanisms is facilities for handling priority of information. Such identified user interface functionality that is helpful in many situations indicates that the mechanisms for developing user interfaces for emergency response should support reuse, preferably at component level. This means that it should be possible to have ready-made user interface components that are easily integrated into a new application being developed. 7.2 User Interface Functionality That May Be Specialized to Different Types of Operations As there are situations where the same user interface functionality is applicable in different types of operations, there are other situations where the user interfaces cannot be identical, but rather variants of a common user interface design. An example of this from the discussions above is a user interface providing awareness of changes in a network. The rules for which kind of changes that should be handled in which way are typically specific for different types of operations, but once the rules

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

27

are specified, the actual user interface mechanisms implementing the awareness functionality may be identical. Another example is user interfaces for presenting additional information about devices and sensors, including their role and how they are used. Which information that should be presented for a sensor may differ from operation to operation, but the mechanisms for presenting the information may be the same. A third example is rules for turning different awareness functionalities on and off. These rules are typically different in different operations, but may be served by the same user interface mechanisms. These examples may be generalized to a principle of having generic components that are parameterized for the aspects that differ between operation types and/or actual operations. To handle this in a development context requires more than ready-made components. There is also need for a model (which in simple cases can be implemented using a configuration file) that is able characterize different types of operations (typically specified as part of preparing for operations), as well as characterizing aspects of an operation type that may change during the operation. 7.3 User Interfaces That Work Across Platforms, Screen Sizes, Modalities, etc. The need for user interfaces that are available on different kinds of equipment, including computers and devices with different screen sizes [4] was identified both for local leaders and field workers. For local leaders, it is important to have user interfaces that are available both on mobile devices and equipment with larger screen size, and keep the context of use when moving from one to the other. Having a user interface solution keeping the context when changing equipment is specially challenging if the involved equipment have different screen sizes and/or user interface capabilities, as information that is presented in one larger screen may be spread through different screens on a device with a smaller screen. Handling this may require special adaptation mechanisms. For field workers, it is important to have user interfaces that exploit different modalities (possibly in parallel, and both for presentation and interaction). We also identified a need for using different display types like devices, and head-up displays on goggles or visor, as well as providing interaction also through sensors. Related needs are the possibility to present the information or getting access to special functionality regardless of the source and the transportation means used. These needs focus on having user interfaces that are able to adapt to quite varying sets of technical conditions. Developing user interfaces with such abilities is either extremely resource demanding (if specific support for every combination of technical conditions is developed), or requires developing means that are able to operate with specifications that work across technical variations [16]. 7.4 Adaptive Behavior in the User Interfaces Although the need for adaptive behavior has been touched upon also for the three user interface characteristics just discussed, it is also a need in itself. One example is user interface functionality for locating nodes in the network, which may be provided as a map-based visualization, by superimposing the information on a camera view, or through sound. The choice of which of the mechanisms to use may be determined by

28

E.G. Nilsson and K. Stølen

an adaptation mechanism based on information regarding the user's role and task, as well as characteristics of the device used. More generic examples are mechanisms for filtering information, as well as explicit and implicit choices of which information to present, which have been identified as important functions for local leaders. Handling this in the general case may require user interfaces that support adaptation of the presentation of information. The need for adaptive behavior in user interfaces is also supported by other requirement gathering activities that we have conducted for the emergency response domain. A related functionality to adaptation is having mechanisms for composing user interfaces. This may be used by systems developers at design time, either as a means for developing systems more efficiently, or as a way of specifying which functionality or presentation that should be available for an adaptation mechanism at run time. Composition may also be done by end-users at run time. By this, users get the possibility (or are left the responsibility) to conduct some or all of the adaptation of the user interface themselves. This requires more interaction by the users, but it also leaves them with more control of their support tools. Developing user interfaces that support adaptation and/or composition involves many of the same challenges as handling cross-platform user interfaces. While many cross-platform issues may be handled by development tools at design time, adaptation and end-user composition require special run time mechanisms as well.

8 Requirements to Ad Hoc Networks When Used in Emergency Response So far in this paper, we have focused on how the use of wireless ad hoc networks as communication infrastructure in emergency response influence use, design and development of user interfaces for this domain. In this section, we will look briefly into which requirements ICT solutions for emergency response pose on the ad hoc networks. In section 4, we motivated why wireless ad hoc networks are well suited to handle communication in ICT solutions for emergency response, but there are still challenges. One main challenge is using a technological solution for the ad hoc network that renders it possible to connect sensors and devices to the network, and that does not drain the batteries of these in a very short time [5]. There are also a number of challenges connected to the size of the network. For a network solution to be practical, it must be flexible with regards to the number of nodes that are needed for covering an operational area. E.g. in an avalanche rescuing operation, this area will be fairly small, and the density of sensors and devices will be high, making ad hoc networks well suited. On the other hand, in an operation where limited personnel resources are searching for a missing person in a large geographical area, ad hoc networks will be able to cover only parts of the operational area at any given time. In this case, there may e.g. be one ad hoc network for each search group, which means that these ad hoc networks must be supplemented with gateways to cellular and/or special emergency networks. For this to be feasible, the ad hoc networks must interplay with the networks with a wider range. In such a setting, it is

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

29

important that applications exploiting information about the ad hoc networks are able to use this information also through the other networks. Furthermore, regarding network size, some of the discussions on use and design of user interfaces above are based on an assumption that the number of nodes in the network is limited. In a situation with hundreds of sensors and devices connected, it is not practical for a user to deal with information about individual nodes in the network (but information and presentations based on aggregations of and reasoning on data collected from the nodes may still be valuable). Issues regarding the speed of ad hoc networks are typically "hen and egg" kind of problems. On the one hand, one may argue that the speed limitations should be taken into account when designing application that are using the networks, thus restrain from including transfer of high resolution pictures and live video [3]. On the other hand, this kind of bandwidth-challenging data may be crucial for an application supporting special emergency operations, meaning that ad hoc networks cannot be used, or must be supplemented with other network solutions for these types of operations. Lastly, we once more emphasize the need for awareness of the connectivity of nodes, both when deploying and using ad hoc networks in emergency response. As discussed above, this is an important issue when designing user interfaces, but to be able to make such user interfaces, the nodes in the network must provide the necessary information about their state. Above we discussed providing this awareness through the user interface of the applications used, but it should be noted that this may be supplemented by feedback provided directly through the nodes themselves. On example is that a device vibrate in a special way if it is not connected, another is the use of light or sound signals when sensors need attention.

9 Related Work Research on ad hoc networks [1, 24, 28, 30, 33] focuses mostly on technical network issues like architecture, topology, routing, coverage, security, protocols, layers, and channels. User interface issues are seldom covered, except for topics like simulation [11, 31], quality of service [26, 29] and deployment [32]. Emergency response is sometimes put forward as a suitable application area for ad hoc networks [10, 24], and there has also been conducted work on network solutions targeted at emergency response [6, 13, 21], but most of this work also focus on network issues. In human computer interaction research, networks are usually viewed as a means rather than a topic influencing the research, while the special challenges imposed by emergency response raise important research questions. This includes utilizing multi modality [7, 25] and supporting adaptive behavior in the user interfaces [14, 27], usually focusing on case studies, concrete solutions, and methods; more seldom on requirements and design advices. Research on emergency response is by nature multidisciplinary, but there is usually more focus on user interfaces [12, 22] than network solutions [23]. Research papers discussing user interfaces tend to focus on concrete systems and concrete user interfaces solutions.

30

E.G. Nilsson and K. Stølen

We have found little work addressing the combination of wireless ad hoc networks, user interfaces design and development, and emergency response. Bharosa et al. [2] address both user interface and network issues, but with a much broader scope, and thus being far less specific when handling these issues.

10 Conclusions and Future Work In this paper we have investigated how the use of wireless ad hoc networks influences design and development of user interfaces for emergency response applications. We have argued that it may be helpful to make details about the state of the network explicit to the end-user. This includes information about availability, coverage and connected nodes, i.e. information that is usually hidden for the user or only shown implicitly in traditional usage situations using standard network solutions. In addition to being useful for the user, it may also be exploited by applications. We have also argued that user interfaces for local leaders and field workers in an emergency response must fulfill a set of specific requirements. Even though a local leader has a very attention requiring primary task, an application with a well design user interface may relieve the leader from some of the demands for attention. Doing the same for a field worker is more challenging, so for this user group it is more important to have non-intrusive ICT support, possibly offering non-visual modalities as an alternative to or in combination with visual presentation and interaction. For local leaders, supporting user interfaces on equipment with different screen sizes is important to give optimal solution both when the leader is at a local control post and when the leader is moving around. For both groups, information about the extent of the network is potentially useful, and local leaders have special needs regarding awareness of changes in the network, while field workers have special needs for knowing their own connection state. Presenting all these kinds of network information in an optimal way is very challenging. To meet these challenges when developing user interface solutions we see the need for generic components parameterized so that they may be configured to different types of operations as well as characteristics of actual operations. We also see the need for means facilitating development of adaptable user interfaces that are able to support different platforms, screen sizes, and modalities without requiring that each combination is developed separately. A common factor for handling all these challenges is flexibility, indicating that composition is a useful mechanism to exploit both at design and run time. Emergency response has been put forward as one of the prime examples of application areas where ad hoc network is especially well suited. In addition to practical issues like connectivity and battery life of sensors and devices, we have made some considerations about the speed and size of the network. Regarding speed, we conclude that this can either be handled by reducing the needs for communication to the available speed, or by choosing a communication solution offering the required speed. Regarding size, we conclude that the size and character of the operation, and the density of sensors and devices, are important factors regarding the appropriateness of using wireless ad hoc networks, as well as how the networks should be configured.

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

31

Our future research will focus on handling user interface development for applications supporting emergency response, taking the requirements for flexibility into account. The solutions need to be flexible with regards to type of operation, special needs for the given operation, available and needed information sources, applications and services, available and needed sensors, available infrastructure, type of equipment to be used, work situation of the user, and modalities to exploit. The requirements for flexibility have at least two implications. Firstly, that developing optimal solutions for all combination of needs will be utterly expensive. Secondly, that it is almost impossible to specify an optimal end-user solution in advance. Our aim is to apply a model-based approach [16] to facilitate easy composition of support tools, partly at design time and partly at run time.

Acknowledgements The work on which this paper is based is supported by the EMERGENCY project (187799/S10), funded by the Norwegian Research Council and the following project partners: Locus AS, The Directorate for Civil Protection and Emergency Planning, Geodata AS, Norwegian Red Cross, and Oslo Police District.

References 1. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Elsevier Computer Networks 47(4), 445–487 (2005) 2. Bharosa, N., Lee, J., Janssen, M.: Challenges and obstacles in sharing and coordinating information during multi-agency disaster response: Propositions from field exercises. Springer Information Systems Frontiers 12(1), 49–65 (2010) 3. Bergstrand, F., Landgren, J.: Using live video for information sharing in emergency response work. Inderscience International Journal of Emergency Management 6(3-4), 295– 301 (2009) 4. Calvary, G., Coutaz, J., Thevenin, D., Limbourg, Q., Bouillon, L., Vanderdonckt, J.: A Unifying Reference Framework for Multi-Target User Interfaces. Elsevier Interacting with Computers 15(3), 289–308 (2003) 5. Carle, J., Simplot-Ryl, D.: Energy-Efficient Area Monitoring for Sensor Networks. IEEE Computer 37(2), 40–46 (2004) 6. Chandra-Sekaran, A.-K., Flaig, G., Kunze, C., Stork, W., Mueller-Glaser, K.D.: Efficient resource estimation during mass casualty emergency response based on a location aware disaster aid network. In: Verdone, R. (ed.) EWSN 2008. LNCS, vol. 4913, pp. 205–220. Springer, Heidelberg (2008) 7. Datcu, D., Yang, Z., Rothkrantz, L.: A system for adaptive multimodal interaction in crisis environments. In: Aykin, N. (ed.) HCII 2007. LNCS, vol. 4560, pp. 334–343. Springer, Heidelberg (2007) 8. The Directorate for Civil Protection and Emergency Planning: Veiledning om røyk- og kjemikaliedykking [Guidance on smoke and chemical diving] (in Norwegian) (2005) ISBN 82-7768-061-9

32

E.G. Nilsson and K. Stølen

9. The Directorate for Emergency Communication: Nødnett - New Public Safety Radio Network in Norway, http://www.dinkom.no/FILES/dnk_brochure_english_ 180x285_trykt_120907.pdf (accessed June 15, 2010) 10. Hoebeke, J., Moerman, I., Dhoedt, B., Demeester, P.: An Overview of Mobile Ad Hoc Networks: Applications and Challenges. Journal of the Communications Network 3(3), 60–66 (2004) 11. Hogie, L., Bouvry, P., Guinand, F.: An Overview of MANETs Simulation. Elsevier Electronic Notes in Theoretical Computer Science 150(1), 81–101 (2006) 12. Liu, S.B., Palen, L.: Spatiotemporal Mashups: A Survey of Current Tools to Inform Next Generation Crisis Support. In: Proceeding of ISCRAM 2009, Conference on Information Systems for Crisis Response and Management, ISCRAM (2009) ISBN 978-91-633-4715-3 13. Lorincz, K., Malan, D.J., Fulford-Jones, T.R.F., Nawoj, A., Clavel, A., Shnayder, V., Mainland, G., Welsh, M., Moulton, S.: Sensor networks for emergency response: challenges and opportunities. IEEE Pervasive Computing 3(4), 16–23 (2004) 14. Luyten, K., Winters, F., Coninx, K., Naudts, D., Moerman, I.: A Situation-Aware Mobile System to Support Fire Brigades in Emergency Situations. In: Meersman, R., Tari, Z., Herrero, P. (eds.) OTM 2006 Workshops. LNCS, vol. 4278, pp. 1966–1975. Springer, Heidelberg (2006) 15. The National Police Directorate in Norway: Politiets beredskapssystem, del 1 - Håndbok i krisehåndtering [Emergency System for the Police, Part 1 - Handbook in Crisis Management] (in Norwegian) (2007), ISBN 978-82-92524-49-7 16. Nilsson, E.G., Floch, J., Hallsteinsen, S., Stav, E.: Model-based User Interface Adaptation. Elsevier Computers & Graphics 30(5), 692–701 (2006) 17. Nilsson, E.G.: Challenges for Mobile Solutions for Emergency Response; SINTEF Report A16017 (2010), ISBN 978-82-14-04958-9 18. Nilsson, E.G.: Design patterns for user interface for mobile applications. Elsevier Advances in Engineering Software 40(12), 1318–1328 (2009) 19. Nilsson, E.G.: Tasks, information models and visual prototypes for local leaders at the TYR training exercise; SINTEF Report A16007 (2010) ISBN 978-82-14-04479-9 20. Nilsson, E.G., Brændland, G.: EMERGENCY: Evaluering av DISKO [Evaluation of DISKO] (in Norwegian); SINTEF Report A12084 (2009) ISBN 978-82-14-04445-4 21. Phanthuna, N., Nakhon, P., Srisongkram, W., Trongtirakul, T.: A sudden flood alert system based on a mesh network. In: Proceedings of the 4th WSEAS International Conference on Circuits, Systems, Signal and Telecommunications, pp. 113–118. WSEAS (2010) 22. Puras, J.C., Iglesias, C.A.: Disasters2.0. Application of Web2.0 technologies in emergency situations. In: Proceeding of ISCRAM 2009, Conference on Information Systems for Crisis Response and Management, ISCRAM (2009) ISBN 978-91-633-4715-3 23. Riblett, L.E., Wiseman, J.M.: TacNet: Mobile ad hoc secure communications network. In: Proceedings of 41st Annual IEEE International Carnahan Conference on Security Technology, pp. 156–162. IEEE, Los Alamitos (2007) 24. Römer, K., Mattern, F.: The Design Space of Wireless Sensor Networks. IEEE Wireless Communications 11(6), 54–61 (2004) 25. Schütz, R., Glanzer, G., Merkel, A.P., Wießflecker, T., Walder, U.: A speech-controlled user interface for a CAFM-based disaster management system. In: Luo, Y. (ed.) CDVE 2007. LNCS, vol. 4674, pp. 80–87. Springer, Heidelberg (2007)

Ad Hoc Networks and Mobile Devices in Emergency Response – A Perfect Match?

33

26. Shakya, S., Awal, M.A., Das, D.K., Tsuchimoto, Y., Kanchanasut, K.: A Real-Time Performance-Monitoring Tool for Emergency Networks. In: Fdida, S., Sugiura, K. (eds.) AINTEC 2007. LNCS, vol. 4866, pp. 169–183. Springer, Heidelberg (2007) 27. Streefkerk, J.W., van Esch-Bussemakers, M.P., Neerincx, M.A.: Designing personal attentive user interfaces in the mobile public safety domain. Elsevier Computers in Human Behavior 22(4), 749–770 (2006) 28. Toh, C.K.: Ad Hoc Mobile Wireless Networks: Protocols and Systems. Prentice Hall, Englewood Cliffs (2002) 29. Wang, W.J., Chatterjee, M., Kwiat, K.: QoS dashboard: Translation of user inputs to networking strategies. In: Proceedings of 2nd International Symposium on Wireless Pervasive Computing, pp. 20–24. IEEE, Los Alamitos (2007) 30. Yick, J., Mukherjee, B., Ghosal, D.: Wireless sensor network survey. Elsevier Computer Networks 52(12), 2292–2330 (2008) 31. Zhang, Y., Li, W.: An integrated environment for testing mobile ad-hoc networks. In: Proceedings of the 3rd ACM International Symposium on Mobile ad Hoc Networking & Computing, pp. 104–111. ACM, New York (2002) 32. Zhao, S., Jaskiewicz, D., Karvo, J.: A deployment tool for public safety ad-hoc networks. In: Proceedings of 1st International Conference on Communication Systems Software & Middleware, pp. 41–46. IEEE, Los Alamitos (2006) 33. Zheng, J., Jamalipour, A.: Wireless Sensor Networks: A Networking Perspective. IEEE Press, Los Alamitos (2009)

Sensorium – An Active Monitoring System for Neighborhood Relations in Wireless Sensor Networks Stefan N¨ urnberger, Reinhardt Karnapke, and J¨ org Nolte Brandenburg University of Technology Cottbus {snuernbe,karnapke,jon}@informatik.tu-cottbus.de

Abstract. Communication neighborhood in wireless sensor networks changes often as links break or appear. Therefore, monitoring link quality and (logical) network topology is necessary. As node placement has a large influence on the radio neighborhood and its changes, different positions should be evaluated before starting the actual application. In this paper we introduce Sensorium, an active monitoring system that supplies the user with an insight into the neighborhood relations between nodes and their changes in time. It can be used before the actual deployment to evaluate different possibilities of node placement and choose the one that offers the best connectivity. Keywords: wireless sensor networks, active monitoring, robustness, Sensorium.

1

Introduction

Wireless sensor networks offer a wide variety of applications which merit a number of design goals. When a sensor network is designed and deployed, network lifetime is one of the major goals. But there are two factors that need to be considered during deployment: Sensor coverage and network connectivity. The application dictates the area that has to be surveyed, thus deciding the minimum number of nodes needed to reach the desired coverage. To determine the needed number of nodes to reach the necessary connectivity, it has to be measured in a pre-deployment phase. But measuring it only once is not enough. As various experiments have shown, network connectivity varies a lot over time [20,18,9,6,8,5]. These changes make it necessary to monitor the connectivity either constantly while the application is running, or for a sufficient time before the deployment. Monitoring during network operation always introduces runtime overhead and active monitoring changes the surveyed system. For this reason, we decided to use a pre-deployment monitoring approach. In this paper we present Sensorium, a monitoring system for neighborhood relations in wireless sensor networks. It collects the information of each node placed in a location that could be used in the real deployment before the sensing application is started. The connectivity information is then gathered at a central J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 34–47, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Sensorium

35

sink, enriched with a map of the deployment area and presented visually. Based on this information, a user can decide where to add additional nodes or to move already deployed ones, in order to have the best connectivity possible. This paper is structured as follows: Section 2 shows different approaches that can be used when designing a monitoring system for sensor networks. Our system, Sensorium, is presented in section 3 and an exemplary deployment described in section 4. Related work is given in section 5. We finish with a conclusion in section 6.

2

Designing a Monitoring System for Wireless Sensor Networks

When designing a monitoring system for wireless sensor networks, it is always necessary to have at least a partial knowledge of the intended application. Designing a general monitoring system that can be used for all applications is nearly impossible and will automatically lead to a waste of resources when data is gathered that is not needed. Therefore it is mandatory to define the monitoring issues first. These can include for example the liveliness of the sensor network as a whole or individual node failure. Connectivity, i.e., the ability of the sensors to deliver the sensed values to the sink is another critical aspect, even though there is not much that can be done to change it after deployment. Sometimes sensors on individual nodes will fail, delivering false data, which can also be detected on the sink, , e.g., by comparing the values detected by neighboring nodes. If the difference is too high, one or more of the sensors must be faulty. The energy status of nodes can be measured and used for traffic shaping to prevent node failures. After node failures, the human operator will need to know if the area that has to be monitored is still completely covered. Some of these monitoring goals can be reached without transmitting additional data, simply by deriving information from the traffic that arrives at the sink. Others require the monitoring system to generate its own information on the nodes and transmit it to the sink in one way or another. This second approach is called an intrusive way or an active monitoring system while systems that use the first approach are called passive monitoring systems. 2.1

Passive Monitoring

Passive monitoring systems aim to provide all information desired by the user by deducting it from the messages that are transmitted by the application anyway. Sources of information are , e.g., the number of packets that arrived at the sink either from a certain node or from a certain part of the network. When the routing topology (, e.g., a tree) is known, it is possible to find node failures that are responsible for missing data from a whole subtree. Another possible source of information is the application data collected. If the values collected from one particular node are not within a certain boundary or diverge too much from those of neighboring nodes, the probability of sensor failure is high.

36

S. N¨ urnberger, R. Karnapke, and J. Nolte

The big advantage of passive monitoring is that it does not interfere with the application in any way. Timing, network load and energy consumption remain the same whether it is used or not. It is also possible to make a post-mortem analysis of the network from stored data. There are, however, a number of disadvantages. When no messages from a node arrive at the sink, it is not possible to distinguish between the different possible reasons: Packet loss, congestion, link failure and node failure all lead to the same error pattern. Another disadvantage is that only a limited view of the status of each node is provided. If the application data does not include the energy status of the nodes, it can only be roughly approximated. Neighborhood information and area coverage are also problematic. The most important problem, though, is that passive monitoring is only usable for applications that communicate regularly, like , e.g., habitat monitoring. When the application does not communicate regularly, it is not possible to satisfy the liveliness criterion needed , e.g., by networks deployed for fire protection or intrusion detection. 2.2

Active Monitoring

There are two ways of realizing active monitoring, in band and out of band. Both of them can be used to gather periodic data or request data on demand. The big advantage of active monitoring is that any kind of data that could be of interest can be monitored. The biggest disadvantage is that it interferes with the normal operation of the sensor network. If it is realized with in-band communication (i.e., using the same means of communication as the application) it increases the network load which can lead to congestion and thus packet loss. Even if no packet loss occurs, it can influence the timing behavior badly. If additional communication hardware is used, the energy consumption of the sensor nodes is still increased. Even if the additional hardware comes with its own power supply, it still needs to access at least the memory to retrieve , e.g., neighborhood tables. Also, the additional hardware makes the nodes more expensive. All this influence on the node behavior can either mask errors or even introduce new ones. Even though there are so many disadvantages that need to be considered, active monitoring is absolutely necessary for a number of criteria. When liveliness is the number one goal, and failing nodes have to be replaced immediately, active monitoring should be used. It is also quite useful for monitoring of the logical topology of a network. This enables the evaluation of possible locations for node placement in a pre-deployment phase, to find the minimum number of nodes needed and their placements to reach the desired area coverage.

3

The Architecture of Sensorium

Sensorium was developed as a tool that enables the user of a wireless sensor network to find out the best placement and the number of nodes required to ensure connectivity between nodes and the sink. To supply this information, Sensorium uses the active monitoring approach, and gathers extensive information

Sensorium

37

about the status of every node. While this would interfere with the normal operation of a sensor network, Sensorium is meant to be used in a pre-deployment phase, where no application is running yet. It could also be used after deployment, but then all the disadvantages of active monitoring described above would take effect. In order to provide information about network connectivity, a simple neighborhood discovery protocol is used: all nodes transmit messages regularly. Upon reception of such a message a node notes that at the current point in time there had been an incoming link. To retrieve this information, the sink node sends info request messages into the network, which are answered by the nodes with their current state information. The network behavior is therefore similar to that of a typical sense-and-send application, where a powerful sink requests data from the sensor nodes, which is then merged and stored for further evaluation. In Sensorium, no data packet aggregation is used on the forwarding nodes, because Sensorium already aggregates neighborlists, routing metrics, sensor data and all other requested information into a single packet. Such a packet is already fairly large. As these types of data can not be combined without losses and aggregating two messages would double the size, the probability of packet loss would increase too much. Sensorium uses a modular architecture which allows a user to easily exchange routing protocols or the evaluation interface and add new information sources (, e.g., new types of sensors). This modular architecture makes it also easy to switch between hardware platforms, and even allows the usage of one programming language on the sensor nodes and another on the sink or the host system where the results are displayed for the user. 3.1

Notion of Chronological Sequence

When monitoring network connectivity, it is necessary to be able to identify links between nodes, and their changes in time. Therefore some notion of time is needed, to build a row of snapshots of the logical network topology. In Sensorium time is divided into epochs, with each node collecting its neighborhood information anew in each new epoch. Clock skew is tolerated. This is possible because the nodes always send their current epoch value with their answers, enabling the sink to calculate the skew from its own epoch counter. The skew is used when requesting values from the past that have been stored on the sensor nodes (see data preparation at the sink, section 3.4). Please note that choosing the right length of the epoch is crucial for the success of Sensorium. If it is too long, changes will be missed. If it is too short, the network load gets too high, possibly resulting in congestion. Also, determining the connectivity takes time, as the neighborhood discovery protocol needs to transmit its messages. Epochs must be larger than this amount of time. 3.2

Message Routing

The need to transmit data back to the sink in an ever changing radio neighborhood makes the usage of an adaptive routing protocol mandatory. The choice of

38

S. N¨ urnberger, R. Karnapke, and J. Nolte

routing protocol is heavily influenced by the communication scheme exhibited by Sensorium: One to many for broadcasts from the sink to all nodes (requests), many to one for replies from the sensor nodes to the sink, and one to one for history requests from the sink to a node and the answers. For the first case, transmission from the sink to all sensor nodes, a simple flooding (with duplicate suppression) is used. This ensures the highest delivery ratio possible (if any path to a node exists, it will receive the message) and does not introduce much overhead, as all nodes need to receive the message anyway. The overhead measured in data packets could be reduced a little by having only a subset of nodes retransmit the messages, , e.g., using multi point relay nodes like OLSR [4] does. But finding these nodes and keeping the information up to date would introduce unwanted protocol overhead, which could get quite large if the rate of changes increases. The second case, where the sensor nodes transmit to the sink, has to be handled like a one to one communication though, because no aggregation can take place on the intermediate nodes for reasons described above. Please note that the second case is now identical to the third one, and we will use the same protocol for both cases. For the third case we need a reactive routing protocol that does not rely on topology information, as we are only collecting that information now. Occasional packet losses can be tolerated, as the history function of Sensorium is used to automatically re-request missing data. Therefore, a best effort protocol with a fairly high probability of successful delivery is sufficient. In previous work we introduced Buckshot Routing [13], a robust source routing protocol for wireless sensor networks. In Buckshot Routing, a node that receives a message does not only forward it if its own identity is enclosed, but also if one of its neighbors is on the path the message has yet to travel. This leads to a broader tunnel (with a breadth of one hop) around the source route along which the message travels, circumnavigating broken or unidirectional links, failed nodes and other obstacles. Sometimes the links that are used to circumvent these problems are unidirectional themselves, but pointing the other way. Buckshot routing has a delivery ratio that is close to that of flooding while keeping the number of involved nodes much smaller, especially in large networks. We use Buckshot Routing for all communication from sensor nodes to the sink in Sensorium. The original path around which Buckshot Routing builds its tunnel is obtained when the request from the sink is flooded into the network. Each node that forwards the message attaches its own identity, collecting the path that has been taken. The inverted path is then used for the answers from each sensor node to the sink. Duplicate detection is realized using the identity of the sender and a sequence number defined by the sink. If a node is reset for any reason, it still receives the current sequence number with the next request from the sink, avoiding all false positives in duplicate detection. 3.3

The Request Handler

The request handler which runs on all sensor nodes is used to evaluate the queries sent by the sink. It works on a configurable set of information sources which are represented as bits in a bit field in the query messages. This representation

Sensorium

39

enables the sink to specify different sources within each query. The answers are represented in type-length-value (tlv) encoding, which eases the inclusion of new sources by simply adding a new type and specifying a position in the bit field. In our examples the sources used are neighborhood details, routing information, packet statistics, temperature, light intensity and current energy level. The request handler is realized as a service that decodes the requests, compiles the requested information from its sources and transmits the response messages. The sink then combines the information from all nodes to a snapshot of the status of the whole network. 3.4

Data Preparation at the Sink

The sink runs a local version of almost all node components. The routing is a little different, though. Instead of forwarding packets to another node, they are collected and evaluated here (replies) or injected into the sensor network (requests). The sink provides some of the information sources just like the other sensor nodes (neighborhood information, packet statistics) but omits some others (no sensors, no battery). It also does not have to store a local history as requests to the sink should never fail. There are two modes of operation for the sink, idle and monitoring. In idle state the sink simply waits for the first requests from the user to switch into the active, monitoring mode. Once it has been switched to monitoring, it sends out the periodic information requests to all nodes and keeps the information up to date. This is done by collecting all incoming data, sorting it according to the epochs and identifying nodes from which data is still missing. The counter for missed responses from those nodes is increased, if it reaches a certain threshold the nodes are marked as currently unreachable. At the same time, a history message is created and entered into the history queue for those nodes not marked as unreachable. The messages contained therein are sent repeatedly at regular intervals, until an answer is received or the node in question is marked as unreachable. Once any message is received from a node that has been marked as unreachable before, it is marked as reachable again and a history request for all missed data is generated and entered into the queue. To enable this recovery of lost data, the sensor nodes need to store it locally for a certain amount of time. This is realized using a ring buffer in which the last n values are stored. Which sources are included in the ring buffer is configurable, it does not have to be all that are monitored. The sink also offers an application programming interface which can be used to control the monitoring process (, e.g., define which sources should be requested in the queries) and request the collected information of a whole epoch (involves no communication inside the sensor network). It is the general interface for user applications. 3.5

The User Interface

The user interface is used to switch the modes on the sink between idle and monitoring, and to define the desired information sources for the queries. It

40

S. N¨ urnberger, R. Karnapke, and J. Nolte

also offers the possibilities to configure the presentation and evaluation of the gathered data, as all data is transmitted from the sink to the user interface unevaluated. Figure 1 shows an example of the web interface which is used to gain fast access to listings of all data already gathered and the visualization of the logical topology of the network (radio neighborhood). The web interface can run on almost every platform, even on mobile devices like smart phones for on-site evaluation as it only requires a web browser.

Fig. 1. The Web Based User Interface of Sensorium

Seeing the logical topology is already helping the user a great deal, but the actual geographic information is still necessary to decide which nodes may be removed while keeping the area coverage required by the application. Therefore, a GIS based interface is also provided. It visualizes the network as graphic overlay, and can use data from any geographic database. This is especially useful for deployments like , e.g., fire detection in a nature-sanctuary, where a map taken by airplane can be combined with the GPS coordinates delivered by the nodes. The GIS interface can also visualize the sensor values (, e.g., temperature or light intensity) and the neighborhood- and routing information. All gathered data that is not represented directly is easily accessible from here. An example of the visualization taken from our outdoor experiments is shown in figure 3, section 4.3.

4

Deploying Sensorium

To evaluate our prototype implementation, we made a number of experiments which are described in this section, after the hard- and software we used is presented.

Sensorium

4.1

41

Hardware

We employed the SunSPOT nodes developed by Sun Microsystems as the sensor node platform for our implementation of Sensorium. The SunSPOT hardware consists of a baseboard with the computing and communication parts and a stacked sensor board with the sensor hardware and additional connectors. The baseboard features a 32bit ARM920T processor, 512kB of RAM and 4MB Flash memory as well as a 2.4GHz IEEE 802.15.4 compatible radio module. The sensor board includes temperature and light sensors as well as a three axis accelerometer, eight LEDs and two user buttons. External components can be connected to the sensor board through six analog inputs and five general purpose I/O pins. For the determination of a node’s physical position, we connected a SiRF-III based GPS module to each node.

Fig. 2. One SunSPOT with GPS Module and a Base Station

A consumer grade laptop is used as the sink node in our deployments. It is equipped with a 1GHz PowerPC G4 processor, 768MB of RAM and has an attached SunSPOT base station for communication with the wireless sensor nodes. The base station consists only of the SunSPOT base board. No sensor board or battery is attached to it. A SunSPOT node with GPS module and the base station are shown in figure 2. 4.2

Software

The whole software for Sensorium is written in the Java programming language because of the used SunSPOTs platform. If any other hardware was used, it would be no problem to implement the necessary components for the sensor nodes in , e.g., C++ as only the packet types have to be adhered to. We have chosen the SunSPOTs because we wanted to evaluate neighborhood relations on sensor nodes equipped with an IEEE 802.11.4 transceiver. This was our first evaluation of the SunSPOT software development kit from Sun Microsystems.

42

S. N¨ urnberger, R. Karnapke, and J. Nolte

The SunSPOTs run a lightweight Java Virtual Machine, which enables rapid application development for wireless sensor networks in a well known programming language. The software for the sensor nodes and the sink depends on the SunSPOT SDK. Communication between sink and user interfaces (not within the sensor network) is based on Java RMI and is thus uncoupled from the SDK. While this design enables the use of dedicated machines for data gathering and evaluation, in our field experiments those services were consolidated on one machine. The system also allows for concurrent access to the gathered data from different user interfaces. The central data management on the sink ensures that the view on the network is consistent for all users even across different kinds of user interfaces. The web interface for Sensorium is a Rich Internet Application based on the open source Echo Web Framework [12] which allows interactive web applications to be written completely in Java. In our installation the application is hosted on an Apache Tomcat Web Server [10]. Concurrent access to this interface is possible, while the application manages independent sessions for each user. The clients need only be equipped with a reasonably modern web browser with JavaScript support. A Java installation or Java RMI are not needed. The implemented GIS interface is based on uDig [16], a standards compliant open source geographic information system. Since uDig itself is based on the Eclipse Rich Client Platform [17], the application is platform independent and easily extensible with custom plugins. 4.3

Experiments

In the first row of field experiments we deployed networks in varying size from 6 to 25 nodes. The requested data included detailed neighborhood information, energy level and the geographical position of the nodes determined from the attached GPS modules. Unfortunately we experienced an unusually low communication range. The sensor nodes were only able to communicate reliably over the fair distance of five to seven meters. Since this is also approximately the accuracy of the GPS modules, the received geographic coordinates were quite useless to determine the physical layout of the sensor network. The difference to the experiences of others that were able to communicate over a distance of up to 70 meters [23] with the same hardware was probably caused by a high noise level in the 2.4GHz band at the deployment site. While the exact source of this noise is unknown, we were forced to reduce the covered area of our deployments. The experiments have shown that it is almost impossible to predict the network topology of an ad hoc deployment. Our topology varied between a very dense network and complete network separation. We also experienced a very unbalanced neighbor count among the nodes despite the regular grid layout we employed in our field experiments. Through the use of Sensorium in the pre-deployment phase we were able to track the topology changes almost immediately and adjust the topology accordingly. A real world deployment without the use of a live monitoring system would have been fatal under the mentioned circumstances.

Sensorium

43

Fig. 3. Deployment at the University Library

The topology visualization through the web interface of Sensorium is good for small networks, but got quite cluttered with increasing node and neighbor count. More sophisticated graph layout algorithms could help diminish that problem. Later analysis of the log files on the sink showed that a considerable amount of information was retrieved from the nodes’ history buffers. Without those the achieved exhaustive topology view would not have been possible. In a later experiment with 6 nodes the GIS based user interface was evaluated. As different deployment site the university library building was chosen. The requested data included neighborhood information, energy level, the nodes’ geographic position, their current routing information and the readings from the built in light sensors. Again, some of the information had to be retrieved from history buffers. A screenshot from this deployment is shown in figure 3. The gathered information from the sensor network fits nicely into an overlay of geographic data like the shown digital orthophoto of the area. An automated statistical evaluation of the gathered data is still missing, but may be integrated into the existing user interfaces in the future. The history management mechanism has proven to be an essential part of the system that is able to compensate the insufficiencies of the unreliable wireless communication.

5

Related Work

Pimoto [1] is a completely passive monitoring system. It relies on dedicated monitoring nodes that are deployed within the monitored network. Those nodes

44

S. N¨ urnberger, R. Karnapke, and J. Nolte

capture the traffic from the network and push it to a central server. Bluetooth communication is used to avoid transmitting the data in-band where it could influence the monitored system. Through this design the monitoring does not need to be suspended while data is transferred to the sink. Of course, this only works when the sensor network operates in another frequency band than Bluetooth. The captured traffic is enhanced with meta data such as the capture time and the monitoring node’s address. The server combines the data from all nodes and automatically reorders packets with the information provided in the meta data. Display and analysis of the traffic generated by the sensor nodes is accomplished through a custom plugin for Wireshark, a popular network monitoring tool formerly known as Ethereal [7]. Pimoto may be described as a packet sniffer for wireless sensor networks that can be used for protocol debugging. The developers emphasize that due to the redundant hardware the system may be used to analyze already deployed networks. But the completely passive operation renders it unsuitable for applications that exhibit little or no traffic at all. It also means that in contrast to Sensorium Pimoto is not able to collect information about radio neighborhood. The authors of Sympathy [14,15] utilize another approach to monitoring. They use a concept of metrics to identify whether the monitored network is in a state of error. Their primary metric is the number of packets that arrive at the sink. Applications that yield the expected amount of data are likely to be working correctly. This way the impact on bandwidth and energy consumption is minimized. For applications that are not expected to communicate, Sympathy offers the possibility to let nodes actively transmit periodic metrics like neighbor count and current parent in the routing topology to the sink. When the metrics suggest a state of error, e.g., some node repeatedly fails to transmit data, Sympathy systematically inquires additional metrics from the nodes on the failed path. Since single node failures in the sensor network may lead to subsequent faults, when , e.g., the network gets separated, Sympathy employs an empirical decision tree to identify the root cause of each detected failure. This way the human operator of the network is presented with the most likely cause for the detected behavior and may induce direct action. Sympathy is primarily thought of as a sensor network debugger. As such it gives little insight to the workings of a functional network. Comprehensive information is only generated when a fault is assumed. Furthermore, a good deal of information about the expected application behavior and network topology has to be provided in order to yield accurate results. As Sympathy uses the messages generated by the application to get an insight into the state of the network, it can not be used in pre-deployment like Sensorium. PAD [11] is a system similar to Sympathy that omits the active inquiry part and uses a packet marking algorithm instead. Each packet transmitted through the network is marked with a hop count and one of the forwarding nodes’ id. The sink builds a belief network from this information, which is in turn used to identify failures and their causes. The authors claim an accuracy of about 90% for the fault identification algorithm. Since the correctness of the belief network

Sensorium

45

converges slowly with the number of received packets, the approach is less likely to work for frequently changing topologies. For sensor networks based on TinyOS, a nesC component framework for active network monitoring and management called SNMS [19] can be used. It is intended to be deployed alongside a sensor network application. SNMS provides a flexible, application cooperative means to gather information about the state of the sensor nodes and execute basic commands like node reset. Application programmers may define additional parameters that are exported through the system. Interpretation of this data is left to the network operator. SNMS offers query based health monitoring of the deployed applications as well as an event logging system. The nodes log events locally and transmit them to the sink when a user defined threshold is exceeded. SNMS has also been extended with an RPC mechanism that enables the management system to execute arbitrary commands of the deployed application. The extended system is called L-SNMS [22] and offers an enhanced Java based user interface for the visualization of events in the network and additional management functionality such as reprogramming of nodes. L-SNMS is the only system of those described above that offers a visualization of the gathered information, which in our opinion is absolutely necessary to help the user understand what is happening inside the monitored sensor network. Therefore, Sensorium offers multiple user interfaces which show the logical and/or geographical topology of the network enriched with further gathered information. Different tools for sensor network visualization such as WiseObserver [3], Mote-View [21] or SpyGlass [2] are available. While WiseObserver and MoteView visualize data available from a connected database, they are not concerned with obtaining this data in the first place. SpyGlass uses a simple flooding based scheme to periodically broadcast sensor readings to gateway nodes. All three systems focus on the efficient visualization of routing information and sensor readings. SpyGlass also supports the creation of custom specialized visualization plugins. In contrast to these protocols, Sensorium offers a complete solution, gathering data in the sensor network, taking care of lost messages, enriching the gathered data with freely available terrain information and presenting the processed information to the user.

6

Conclusion

In this paper we have presented Sensorium, an active monitoring system for wireless sensor networks. It is meant to be used in a pre-deployment phase, to evaluate possible placements of individual sensor nodes. Sensorium actively collects neighborhood information from the sensor nodes to deliver a complete view of network connectivity to the user. When this is evaluated over a certain time, a good placement for the sensor network can be found. We have shown that our prototype implementation works in real world experiments with SunSPOT sensor nodes. An interesting fact discovered in one of them is that the communication range we measured in the first deployment was only a few meters

46

S. N¨ urnberger, R. Karnapke, and J. Nolte

while other researchers using the same hardware were able to reach up to 70 meters. If the sensor network had been deployed without active monitoring, the insufficient connectivity would have been discovered much later, leading to loss of application data.

References 1. Awad, A., Nebel, R., German, R., Dressler, F.: On the need for passive monitoring in sensor networks. In: Proceedings of the 11th EUROMICRO Conference on Digital System Design Architectures, Methods and Tools, pp. 693–699. IEEE Computer Society, Los Alamitos (2008) 2. Buschmann, C., Pfisterer, D., Fischer, S., Fekete, S.P., Kr¨oller, A.: Spyglass: a wireless sensor network visualizer. SIGBED Rev. 2(1), 1–6 (2005) 3. Castillo, J.A., Ortiz, A.M., L´ opez, V., Olivates, T., Orozco-Barbosa, L.: Wiseobserver: a real experience with wireless sensor networks. In: PM2HW2N 2008: Proceedings of the 3nd ACM Workshop on Performance Monitoring and Measurement of Heterogeneous Wireless and Wired Networks, pp. 23–26. ACM, New York (2008) 4. Clausen, T., Jacquet, P.: Optimized link state routing protocol (olsr) rfc 3626 (2003) 5. Couto, D.S.J.D., Aguayo, D., Bicket, J., Morris, R.: A high-throughput path metric for multi-hop wireless routing. In: MobiCom 2003: Proceedings of the 9th Annual International Conference on Mobile Computing and Networking, pp. 134–146. ACM, New York (2003) 6. Dinh, T.L., Hu, W., Sikka, P., Corke, P., Overs, L., Brosnan, S.: Design and deployment of a remote robust sensor network: Experiences from an outdoor water quality monitoring network. In: LCN 2007: Proceedings of the 32nd IEEE Conference on Local Computer Networks (LCN 2007), Washington, DC, USA, pp. 799–806. IEEE Computer Society, Los Alamitos (2007) 7. Hards, B.: A guided tour of ethereal. Linux Journal (118), 7 (2004) 8. He, T., Krishnamurthy, S., Luo, L., Yan, T., Gu, L., Stoleru, R., Zhou, G., Cao, Q., Vicaire, P., Stankovic, J.A., Abdelzaher, T.F., Hui, J., Krogh, B.: Vigilnet: An integrated sensor network system for energy-efficient surveillance. ACM Trans. Sen. Netw. 2(1), 1–38 (2006) 9. Langendoen, K., Baggio, A., Visser, O.: Murphy loves potatoes: Experiences from a pilot sensor network deployment in precision agriculture. In: Proc. 14th Intl. Workshop on Parallel and Distributed Real-Time Systems (WPDRTS) (April 2006) 10. Lerner, R.M.: At the Forge: Server-Side Java with Jakarta-Tomcat. Linux Journal 2001(84), 10 (2001) 11. Liu, K., Li, M., Yang, X., Jiang, M.: Passive diagnosis for wireless sensor networks. In: SenSys 2008: Proceedings of the 6th ACM conference on Embedded network sensor systems, pp. 371–372. ACM, New York (2008) 12. NextApp Inc., Echo Web Framework, http://echo.nextapp.com/site/echo3 13. Peters, D., Karnapke, R., Nolte, J.: Buckshot routing - a robust source routing protocol for dense ad-hoc networks. In: Ad Hoc Networks Conference 2009, Niagara Falls, Canada (2009) 14. Ramanathan, N., Chang, K., Kapur, R., Girod, L., Kohler, E., Estrin, D.: Sympathy for the sensor network debugger. In: SenSys 2005: Proceedings of the 3rd International Conference on Embedded Networked Sensor Systems, pp. 255–267. ACM, New York (2005)

Sensorium

47

15. Ramanathan, N., Kohler, E., Girod, L., Estrin, D.: Sympathy: A debugging system for sensor networks. In: LCN 2004: Proceedings of the 29th Annual IEEE International Conference on Local Computer Networks, pp. 554–555. IEEE Computer Society, Los Alamitos (2004) 16. Ramsey, P.: User Friendly Desktop Internet GIS (uDig) for OpenGIS Spatial Data Infrastructures. Technical report, Refractions Research Inc. (2003), http://udig.refractions.net/docs/udig-summary.pdf 17. Rubel, D.: The heart of eclipse. Queue 4(8), 36–44 (2006) 18. Sang, L., Arora, A., Zhang, H.: On exploiting asymmetric wireless links via oneway estimation. In: MobiHoc 2007: Proceedings of the 8th ACM International Symposium on Mobile Ad hoc Networking and Computing, pp. 11–21. ACM Press, New York (2007) 19. Tolle, G., Culler, D.: Design of an Application-Cooperative Management System for Wireless Sensor Networks. In: Proceedings of the Second European Workshop on Wireless Sensor Networks (EWSN), pp. 121–132. IEEE Operations Center (2005) 20. Turau, V., Renner, C., Venzke, M.: The heathland experiment: Results and experiences. In: Proceedings of the REALWSN 2005 Workshop on Real-World Wireless Sensor Networks (June 2005) 21. Turon, M.: Mote-view: a sensor network monitoring and management tool. In: IEEE Workshop on Embedded Networked Sensors, pp.11–17 (2005) 22. Yuan, F., Song, W.-Z., Peterson, N., Peng, Y., Wang, L., Shirazi, B., LaHusen, R.: A lightweight sensor network management system design. In: IEEE International Conference on Pervasive Computing and Communications, pp. 288–293 (2008) 23. Zennaro, M., Ntareme, H., Bagula, A.: Experimental evaluation of temporal and energy characteristics of an outdoor sensor network. In: Mobility 2008: Proceedings of the International Conference on Mobile Technology, Applications, and Systems, pp. 1–5. ACM, New York (2008)

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System for Establishing Fairness in Ad-Hoc Networks with Selfish Nodes Lukas Wallentin, Joachim Fabini, Christoph Egger, and Marco Happenhofer Institute of Broadband Communications, Vienna University of Technology Favoritenstraße 9/388, A-1040 Vienna, Austria [email protected]

Abstract. Cooperation in ad hoc networks cannot be taken for granted when nodes belong to distinct authoritative domains. In this paper we present Cashflow, a virtual currency system to motivate nodes to participate in ad hoc networks and to prevent selfishness. This system is different to previously proposed virtual currency systems in that it uses a channel concept for data transmission as well as a market system for pricing. The combination of channel and market concept results in a number of positive system characteristics. For instance, Cashflow provides implicit access regulation and load balancing mechanisms to avoid overload situations in the network. Additionally it considers node context for data transmission and pricing. Cashflow also leaves the decision about the participation degree to the user, which is an important but frequently neglected topic for the user acceptance of virtual currency systems. Keywords: virtual currency system, ad hoc network, micropayment, load balancing, credit system, selfish nodes, fairness.

1 Introduction Ad hoc networks have already been an active research field for many years. Originally, the benefits of wireless networks without the usage of additional infrastructure made ad hoc networks attractive for military usage as well as in emergency situations [1]. Today mobile devices have become part of our daily life, generating new application fields for ad hoc networks, which result in new requirements. For instance, multimedia applications require quality of service support while mandating interoperability with other network types like cellular networks or the Internet [2]. Early ad hoc networks and their associated applications have been designed with the assumption of closed user groups, where all network nodes belong to the same authority and therefore are cooperating in archiving a common goal. However, new issues arise if nodes belong to different authorities. In such scenarios, cooperation cannot be taken for granted and the prevention of selfishness is a new challenge for the design of ad hoc networks. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 48–63, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

49

Currently there are two concepts to prevent selfish behavior and stimulate cooperation in ad hoc networks: reputation systems and virtual currency systems [3]. Reputation systems enforce fair behavior by excluding selfish nodes from the network. Additionally reputation systems stimulate cooperation since nodes can rely on other nodes service as compensation for own contributions. Virtual currency systems on the other hand adapt the “pay for service”-concept to ad hoc networks. Nodes must compensate for other nodes service by paying a fee using credits. A node can either earn credits by providing service to other nodes, or buy them from outside the system. It is worth noting that virtual currency systems not only stimulate cooperation between nodes, but additionally could raise interest in ad hoc networks for a number of business applications. Virtual currency systems proposed in literature mainly focus on how to pay and on how much to pay for data transmission along a given path. As pricing function, in most cases they use auctions or fixed price schemes, based on game theoretical considerations. These schemes do not consider the context of the node nor adapt to the users needs. However, we argue that the consideration of node context and the users ability to control the degree of participation in the network, is crucial for the acceptance of virtual currency systems as enabler for ubiquitous ad hoc networks. Focusing on these issues, we present Cashflow, a virtual currency system to motivate nodes to participate in ad hoc networks and to prevent selfishness. This system distinguishes itself from other virtual currency systems by using a channel concept for data transmission as well as a market system for pricing. The combination of a channel and market concept results in a number of positive system characteristics. It gives users control over their participation degree, allows Cashflow to consider the context of nodes and provides implicit load balancing and access control functionality. The rest of the paper is organized as follows. In Section 2 we discuss related work. Section 3 presents Cashflow and discusses its concepts in detail whereas Section 4 concludes with a summery and an outlook on future work.

2 Related Work Participation and fair behavior of nodes in ad hoc networks cannot be taken for granted, if nodes belong to different authorities like companies or private users. In general, nodes are not interested in providing services like routing and packet forwarding for other nodes without getting a reward. Hence, mechanisms are needed to prevent selfish behavior and stimulate cooperation. Currently two classes of systems are suggested in literature to achieve this task. The first class of systems are reputation systems as detailed in the introductory section. Several reputation systems have been suggested in literature, including OCEAN [4], CORE [5] and CONFIDENT [6,7]. The second class of motivation systems are virtual currency systems. The main principle behind these systems is that nodes pay relaying nodes a fee in form of (sometimes virtual) money for forwarding packets. The packet purse model of the virtual currency system Nuglets [8] uses this concept. The source node of a packet adds nuglets, the virtual currency used by this system, into packets it wants to

50

L. Wallentin et al.

transmit to a target node. Intermediate nodes extract nuglets from packets as reward for their service before forwarding them. If there are not enough nuglets left in a packet to compensate for the forwarding service, the packet gets dropped. This leads to the problem that the source node has to estimate the number of nuglets needed for the transmission along a path. If it underestimates the needed amount, the packet will not reach its destination. Otherwise, if it overestimates the fee, the target will receive undeservedly nuglets. To overcome this problem, Nuglets proposes also the packet trade model [8,9] where nuglets are used by nodes to buy packets from previous nodes and sell them for a higher price to next-hop nodes. Consequently, target nodes effectively pay for whole transmissions. A combination of both proposed schemes is possible. As pricing strategy an auction scheme or a fix price strategy is suggested by the authors. Nuglets has two tradeoffs. First, it needs tamperproof hardware to guarantee that no node extracts more nuglets than owed from packets. Second, it is possible that nuglets get lost due to lost or discarded packets in the system. Sprite [10] solves these problems by introducing a credit system instead of a real virtual currency. In this system, the payment and accounting is performed by a credit clearance service. When a node forwards a packet, it keeps a receipt. Later is sends the receipts to the credit clearance service, which compensates the node. Additionally the credit clearance service determinates the charges based on transmission success. This is done to prevent bogus behavior. iPass [11,12] uses a credit clearance service for charging and accounting too. The difference between Sprite and iPass is the way they determine the fee. While Sprite uses game theory to determine the charge, iPass uses an dynamic auction scheme for pricing in order to prevent overload situations. In iPass, packets include bids. In overload situations nodes forward only packets with high bids and discard other packets. Over a feedback mechanism from the destination node, source nodes can adjust their bids to gain a higher throughput. These four virtual currency systems have in common that they are not interconnected with routing protocols. They solely solve the problem how to pay and how much to pay along a given route, but provide no mechanism to transmit packets along the cheapest route between source and target. Commit [13], which is based on the virtual currency system Ad hoc-VCG [14], targets this issue. In Commit the fee nodes charge is based on the energy needed for packet transmission. When a node wants to transmit data to another node a special routing protocol is used to find the cheapest and the second cheapest route. If the fee of the second cheapest route lies beneath a certain value, the source node transmits the packet along the cheapest path but pays the fee of the second cheapest. This is done due to game theoretical considerations [13]. All these virtual currency systems have in common that node context is not considered as well as they do not give user influence on their participation degree. An additional drawback is that none of these systems combine load depending pricing with fee based routing which can yield a load balancing and access control system. The virtual currency system Cashflow, which will be presented in the rest of this paper, addresses these issues.

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

51

3 Cashflow In this section we present the concept of Cashflow, a virtual currency system to motivate nodes to participate in mobile ad hoc networks. This system has been designed to fit the needs of specific usage scenarios which will be described in Section 3.1. The specific requirements of the usage scenarios result in an architecture based on channels for communication, a market concept for pricing and a credit system for charging and accounting. Sections 3.2 to 3.4 discuss these components of the Cashflow system. Section 3.5 provides an usage example. 3.1 Usage Scenarios Cashflow was designed to encourage nodes, belonging to different authorities, to participate in a common wireless network. Cashflow’s concept targets on scenarios like shared office buildings, where the communication infrastructure is temporarily not available or a hotel, where members of a traveler party, accommodated in different rooms, want to exchange pictures using their mobile equipment without leaving their room. In both described scenarios, communication between different nodes is possible, provided that intermediate nodes, equipped with wireless communication technology, are willing to join a common wireless network and provide services like routing and forwarding for other network users. Additionally it can be assumed that in most cases not all nodes are moving at the same time. Taking the office example, a user might use his laptop in his office for some time, than brings the laptop to a meeting, maintaining its position for the meeting duration, and after the meeting ends, he takes the laptop back to his office. Therefore, this laptop shows a nomadic mobility pattern during a normal work day. Besides the location, also the context of this laptop changes over time. During usage in the users office, the laptop might run on AC power, while during the meeting, the internal battery is used as power source. While running on battery, the energy consumption is more important compared to the time while running on AC power. Therefore the users interest to participate changes depending on the power source. The energy source might not be the only parameter influencing on the user’s willingness to provide services for other users. Since routing and forwarding needs computer resources like processing time and memory, users running resourceintensive applications like games, 3D rendering applications or simulations, needing their computer resources for themselves, might not be interested to provide larger parts of their resources for network services. Therefore it was assumed for the design of Cashflow that it is important to consider the context of the node for the fee calculation. The described scenarios also allow to make some expectations concerning the traffic. Because of the size of digital images and other digital documents, it is assumed that in many cases data transmission between nodes will include a number of packets and not just a single one. Additionally, protocols like TCP require a number of packet exchanges for the transmission of even a single data bit. Consequently for the development of Cashflow it was assumed that, in contrast to other wireless networks like sensor networks or vehicular ad hoc networks, nodes frequently

52

L. Wallentin et al.

exchange a sequence of packets within a short time and not only communicate sporadically. Summarizing, Cashflow was developed with a focus on scenarios where the following assumptions are fulfilled: 1. 2.

Network nodes are equipped with wireless communication technology. There is at least one direct or indirect connection between every network node. 3. There is no additional infrastructure or the existing infrastructure cannot be used. 4. Network nodes belong to different users. 5. Nodes can be motivated to provide services for other nodes using some kind of compensation. 6. Network nodes consist of mobile office equipment like laptops, smart phones and personal digital assistants. 7. Nodes are free to move, but most network nodes show a nomadic behavior, meaning that not all nodes are always moving. 8. Typically, nodes transmit a number of packets to other nodes, not just single packets. 9. Nodes are willing to pay a certain maximum fee for services of other nodes. 10. Nodes want to pay as little as possible for services of other nodes. 11. All nodes have sporadic access to the internet, but it is not given that nodes have internet access during their participation in a wireless network. Additionally Cashflow assumes that some of the following node characteristics might be true: 12. User of nodes have different preferences concerning their participation in networks, independent from the reward they receive for their services. 13. The node’s context influences its willingness to participate in wireless networks. These assumptions have established a framework for the development of Cashflow. Some assumptions like 1 to 3 are essential in all ad hoc networks, since otherwise the deployment of a mobile ad hoc network is not possible or does not make any sense. Other assumptions like 4 and 5 are essential for the integration of a virtual currency system. If all nodes belong to one single authority, cooperation can be taken for granted and therefore a system like Cashflow would be obsolete in such kind of network. This is also true in the case that nodes cannot be motivated to participate using rewarding schemes. The remaining assumptions basically define real life scenarios. Some of these assumptions result in system restrictions, others provide optimization potential. For instance, the use of regular mobile office equipment like laptops and smart phones in such a network results in the requirement that the system should not rely on special hardware like tamperproof hardware. The usage of tamperproof hardware would simplify the integration of virtual currency system from the viewpoint of security. On the other hand, the assumption of a specific mobility

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

53

and traffic pattern allows to optimize the routing as well as the data transfer for these specific scenarios. All these assumptions lead to a system which is optimized for a special type of scenario. Obviously Cashflow can also be used in scenarios where some of the assumptions are not true. In this case, some functions of Cashflow might not be used or an additionally overhead might be associated with these functions. For example, in the case that Cashflow is used in a scenario where tamperproof hardware is given, other payment strategies might perform better than the standard system currently integrated into Cashflow. 3.2 Channel Concept Starting point of Cashflow’s channel concept is that in many cases nodes exchange a number of packets consecutively and not just sporadically single packets. The basic idea of the channel concept is that if a node wants to transmit a number of packets, it opens a channel to the destination node. It is important to note that channels in Cashflow are bidirectional, meaning that the source node, the initiator of the channel, can transmit data to the destination node and vice versa. A channel is characterized by two parameters. First parameter is the duration. A channel is open for a certain time during which source and destination nodes can exchange data. After elapse of this time the channel is closed and the source node must open a new channel if it wants to communicate with the destination node again. However, the source node, and only the source node, can request an extension of the duration time, which might be granted by the intermediate nodes. Second parameter is throughput. It specifies how many packets will be transmitted per second through the channel. Since packets have a defined maximum size, intermediate node can estimate the maximum bandwidth a channel needs. Consequently nodes can reject channel requests, if the needed bandwidth exceeds the available one and therefore avoid overload situations. Besides the avoidance of overload situations, by using this channel concept, nodes can regulate their maximum throughput depending on their current situation as detailed in assumptions 12 and 13. Not only intermediate nodes benefit from the channel concept. As soon as a node has successfully opened a channel, it can rely that the connection to the destination node will not be interrupted abruptly because of upcoming traffic of other nodes. In other virtual currency systems it might happen that an intermediate node stops forwarding packets on behalf of a node, because another node requests its services and is willing to pay a higher price for the service than the node is currently paying. Even if the node whose transmission has been interrupted continues its transmission by interrupting traffic from other nodes likewise, still some packets gets dropped as consequence of the interruptions. Besides the additional delay caused by the retransmission of dropped packets by the source node, the source node must pay all intermediate nodes along the path to the node which has dropped the packet. The reason is that these intermediate nodes have forwarded the packet and therefore earned a reward, even if the packet did not reach the target. (Compare with Sprite[10], Nuglets [8]) Therefore, we argue that avoidance of such interruptions by means of Chashflow’s channel concept matches nodes requirements.

54

L. Wallentin et al.

For the establishment of channels it is important that routing protocols support the detection of stable routes. In Cashflow, nodes rate their links to neighbor nodes based on their duration and on the signal to noise and interference value. Since a nomadic mobility pattern is assumed, strong links with a long duration are preferred for channels compared to weak and sporadic links, because of the lower path break probability. By using a modified version of the route discovery algorithm suggests in [15] it is possible to find the cheapest route between two nodes which additionally fulfills restrictions concerning the rating of links. 3.3 Pricing Scheme The channel concept influences the way nodes determine their service fee. Cashflow uses a market concept to calculate fees charged by nodes for forwarding packets. The basic idea is that nodes continuously adapt the fee they charge for the so called standard channel, depending on supply and demand. The standard channel is a channel with a fixed throughput and duration value, known to every node. Based on the price of the standard channel, a node can derive the price of any other channel with any parameterization. When a node wants to open a channel along a given path, it sends a channel request packet including information about the channel along the path to the destination. The destination node replies with a channel response packet. Intermediate nodes include the value of the fee they would charge for the requested channel into the response packet, so that the original source nodes finally gets the information about the channels costs. Based on the channels fee and the price the source node is willing to pay, it can either accept the offer and open the channel or reject it. By adjusting the price, nodes can make themselves, respectively their routing service, more or less attractive for other nodes. An alternative point of view is that nodes can express by the fee they charge their willingness to host additional channels. Given that each node has a certain preferred throughput value. If the current throughput is significantly higher than the preferred throughput the node could increase the fee with the result that some of the open channels might not be extended by the channel’s source node after the durations time expiration. In addition the probability increases that nodes requesting a channel will reject the channel after receiving the offer because of the high price. Therefore by increasing the fee for channels, nodes can reduce the throughput. The opposite is true when a node decreases its fee. In this case, channels which use this node become cheaper and therefore more attractive for other nodes. Cashflow additionally uses the fee as instrument for load balancing. Since nodes increases their prices in high load situations, channels through high loaded network segments are more costly than through parts with low load. Assuming that nodes prefer to get a service for the lowest possible fee, nodes are interested to open channels using the cheapest path to the destination. Consequently they avoid highly loaded parts of a network if possible. This functionality requires a route discovery protocol, able to find the cheapest route from a source to a destination route. Such a protocol has been developed as part of Cashflow and is presented in [15]. However Cashflow can also be used in combination with other routing protocols, but might result in a performance penalty of the load balancing functionality.

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

55

Fig. 1. Data-sources for price calculation used by Cashflow

Besides load balancing, the combination of the channel with the market concept can also be used to prevent overload situations of the shared medium. When nodes receive a channel request they estimate the additional usage of the shared medium based on the channel’s parameters. If the additional usage would cause an overload, they charge an infinite high price, which corresponds to a reject. Additionally, it increases the fee it would charge for a standard channel, since such a reject indicates an imbalance of supply and demand. Similar to this overload prevention, Cashflow allows each node to define a maximum throughput. If a node receives a channel request and its parameters indicate that the additional throughput caused by this channel would exceed the maximum throughput, again the node can reject the request by charging an infinite fee. Using this mechanism, the user keeps control about his participation degree. In short, a number of parameters influences the node’s fee. Figure 1. gives an overview of the sources currently considered. It is worth noting that additional sources could be integrated into the pricing function. As pictured in Figure 1, the user of a node is a source of parameters for the fee calculation. The user defines the minimum fee a node charges for forwarding packets. Additionally the preferred and the maximum throughput is defined by the user as well as the allowed variance of the throughput before the pricing function adapts the current fee. Moreover, the node’ user configures the maximum usage of the shared medium. The last two parameters the user specifies is the so called minimum and maximum battery penalty, which is added to the fee whenever the node runs on battery. The actual amount of the penalty depends on the battery’s charging level. If the battery is fully charged, the minimum battery penalty is added to the fee. As the charge of the battery decreases the penalty increases until it reaches the maximum penalty when the battery is nearly empty. To realize such a charge dependent penalty, the pricing function needs to use the energy source as input as shown in the figure. The result of the integration of the energy source as parameter in the fee calculation is that nodes using battery are less

56

L. Wallentin et al.

attractive as relay node for other nodes, especially is the battery penalty is high because of low charge, compared to AC powered nodes. The next source pictured in Figure 1 is the node’s physical radio interface layer. It provides two parameters for the pricing function. First it delivers information about the transmission rate used for data transmission between two nodes. The pricing function uses this information to charge a higher fee for transmissions over links with low bit rate than over faster links. We argue that this is in the interest of the whole network since when a higher bit rate is used, the shared medium is occupied for a shorter time than if a low bit rate is used for the transmission of the same data amount. The second input is the average signal to noise and interference (SNIR) ratio from signals from neighbor nodes. This value is used to estimate the link quality between the nodes. Links with weak quality are charged higher, since the probability that retransmissions are needed is higher than for strong links. The MAC-layer is an additional input source for the pricing function. By collecting information about the shared channels usage [16,17], the pricing function can react if the preferred usage rate is in average exceeded for a certain time. If the average usage lies above the maximum usage of the shared medium, as defined by the user, the pricing function increases its fee. Additionally a node can use this information to reject channel requests if it is expected that the additional load caused by the requested channel would result in an overload of the shared medium. Therefore the usage of the information provided by the MAC-layer results in an overload protection, since nodes in high load areas of a network tent to increase their fee with the consequence that routes through high load areas become less attractive for other nodes. The last input source in the current version of Cashflow is the network layer where the channel functionality of Cashflow is implemented. As stated above, the node’s user can configure the preferred and the maximum throughput, as well as the tolerated throughput variance. The network layer provides information about the number of currently open channels as well as the resulting throughput for the pricing function. Using this information, the pricing function can adapt the fee according to the difference between preferred and actual throughput. The fee itself is calculated by two functions. The first function calculates the basic fee a node charges. It continually updates this fee depending on the current throughput and the shared mediums usage. Using this basic fee, another function calculates the fee a node charges for specific channels, depending on channel parameters and the nodes context. The following code describes the functionality of the function which continually adapts the base price: 0: procedure continuesBasicFeeCalculation() 1: { 2: global basicfee=getMinimumFee(); 3: while(TRUE) 4: { 5: sleep(1000); 6: if(getCurrentTP() > (preferredTP+TPVariance)) 7: basicfee=basicfee + feeIncrease;

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

57

8: if(getCurrentTP < (preferredTP-TPVariance)) 9: basicfee = basicfee - feeDecrease; 10: if(wasBlocked == TRUE){ 11: basicfee = basicfee + largeFeeIncrease; 12: wasBlocked = FALSE;} 13: if(getCurrentMediumUsage > maximumMediumUsage) 14: basicfee = basicfee + largeFeeIncrease; 15: if(basicfee < getMinumumFee()) 16: basicfee = getMinimumFee(); 17: } 18: } At the start of this function, it sets the global variable basicfee to the minimum fee the user decided to charge for forwarding packets along a channel. Then the function enters a loop. After waiting for some time, in this example for 1000 milliseconds, it updates the price. In line 6 to 9 of the code, the function checks if the current throughput lies above or beneath the preferred throughput, including the accepted variance and increases or decreases the value of the basic fee accordingly. In line 10, the function checks if a request for a channel was rejected because the additional load would have exceeded the maximum throughput. If so, the basic fee is increase by a value which is larger than the normal increase. In line 13 the algorithm verifies that the current usage of the shared medium is below the maximum usage as preferred by the user. If the usage is higher, the fee is increased by the same value as if a channel has been rejected. After all these fee modifications, line 15 verifies that the new fee is not smaller than the minimum fee. This can happen if there is a low load situation over a longer time in the network. If the new fee would undercut the minimum fee, it is set to the value of the minimum fee. Summarizing, this function adapts the fee depending on supply and demand. In a next step, the value of the global variable basicfee, calculated by the continuesBasicFeeCalculation function, is used by the offering function to calculate a concrete offer for a channel request. The basic fee represents the fee for the already described standard channel, if no additional penalty if charged. The offering function receives four parameters: two channel parameters representing the duration and size, in terms of packets per second, relative to the standard channel and the IDs of the next and the previous node, as it can be seen in line 0 of the following code: 0: procedure getOffer(size, duration, prevHop, nextHop) 1: { 2: fee = basicfee * size * duration; 3: if(batteryPowerd == TRUE){ 4: fee = fee + minBatPenalty; 5: toMax = (maxBatPenalty-minBatPenalty); 6: fee = fee + toMax - toMax * charge; } 7: if(prevHop.speed == SLOW) 8: fee = fee * speedPenalty;

58

9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24:}

L. Wallentin et al.

if(nextHop.speed == slow) fee = fee * speedPenalty; if(prevHop.SNIR == low) fee = fee * signalPenalty; if(nextHop.SNIR == low) fee = fee * signalPenalty; if((expectedTP(size)) > maxTP){ wasBlocked = TRUE; fee = INFINITE; } if((expectedMediumUsage(size)) > maxMediumUsage){ wasBlocked = TRUE; fee = INFINITE; } if(duration > maxDuration) fee = INFINITE; return fee;

Source code line 2 calculates the fee for the requested channel, depending on its size, in terms of packet throughput per second, and duration. The variable basicfee contains the value calculated by the continuesBasicFeeCalculation function described before. In line 3 to 14 different penalties are added to the fee. The first penalty is the battery penalty. Nodes running on battery add at least a minimum battery penalty in case their battery is fully charged. However during the time the charge will decrease and the node will as a consequence increase the penalty. Line 7 to 10 add extra penalties to the fee if the radio interface of the previous or the next node only allows slow communication. For example, this would be the case if a node supports the IEEE802.11g standard but its neighbors are equipped with IEEE802.11b technology. In the provided code we only distinguish between slow and fast transmission speeds. However, in a real implementation a number of different levels would be used. Similar the rating of the signal quality. Again, in the given code we only differentiate between strong and weak signals and not between a number of signal quality levels. The basic idea is to add a penalty if only a weak connection to the neighbor is given, since the node might needs to perform several retransmissions for transferring data over the weak link. In line 15 the node estimates the cumulative throughput of existing and requested channels and verifies that the result lies beneath the maximum preferred throughput as configured by the user. If the estimated throughput exceeds the preferred one, the node rejects the request by setting the channels fee to infinite. Additionally it signals the function calculating the basic fee that a request has been rejected by setting wasBlocked to true, resulting in an increase of the basic fee. The algorithm likewise attempts to estimate the impact of the requested channel on the shared mediums usage. If the new channel would cause a local overload situation the channel is rejected. Before returning the calculated offer, the algorithm verifies that the requested channels duration does not exceed the maximum duration the node accepts. Summarizing, the pricing function adapts the fee a node charges depending on supply and demand and adds penalties to the fee if required by the node’s context.

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

59

Nodes charging high fees are not attractive for other nodes as relay and consequently will be avoided. Therefore nodes can influence their throughput by adapting their fee. The pricing functions presented in this section should be taken only as an example. It is possible to implement a number of different pricing strategies into the system. The only requirement is that the price increases when demand overpasses supply. 3.4 Credit System For the actual payment, Cashflow uses a credit system, where nodes use a kind of virtual credit card to pay rewards for used services. It can be implemented completely in software and consequently does not rely on special tamperproof hardware. Therefore also other virtual currency systems like Sprite and iPass uses similar credit systems for payment. Figure 2. shows the basic concept behind the payment system, which is based on public key infrastructure. It is assumed that each node has a bank account and that there has already been a secure key exchange between the bank and the nodes. To use the services of other nodes, a node needs a bank credit certificate which acts as virtual credit card. To gain such a certificate, the node sends a signed request to the bank including the account number and its public key as pictured in step 1 in Figure 2. If the node is creditworthy, the bank issues a certificate which includes the account number, the node’s public key and an expiration date. This certificate is signed by the bank so that each other node can verify that the credit certificate is valid and will therefore be accepted by the bank. As soon as a node has its credit certificate it can join an ad hoc network using Cashflow. The node does not need to have a connection to the bank server to use the credit certificate. The connection to the bank is only needed for issuing the certificate and for accounting. Since a credit certificate could be valid for some weeks or even months, nodes only need sporadic access to the bank. When a node requests a channel (step 3 in Figure 2) along a given path, it includes its credit certificate into the request packet to prove its creditworthiness. When the request reaches the destination node, the destination node replies to the request by sending a response packet. This response packet, step 4 in Figure 2, is transmitted back along the same path to the source node. While forwarding the response packet,

Fig. 2. Credit concept used by Cashflow

60

L. Wallentin et al.

each intermediate node inserts a hash code, generated out of the request packet, the source node’s account number, the channel ID and the fee it charges for the channel as a signed message. These signed fields represent the offer of an intermediate node. Such an offer also can be seen as a kind of contract which is currently only signed by intermediate nodes. When the source node receives the response, it can calculate the complete price for the channel by summing up all the fees intermediate node charge. If the channel is too expensive for the node, it can drop the response and consequently no channel will be established. If the source node wants to use the channel, it has to sign the offer of each intermediate node. In step 5 the source node sends the signed offers back to the intermediate nodes, which they transmit in step 6, as soon as they have access to the bank, to get the money transferred from the source node’s account to their account. Besides the offer signed by the source and the intermediate node, all the nodes, including the source and the destination node, send statistical channel data to the bank. This statistical data include information about the actual number of packets received by the node. Using this information, the bank can detect malicious nodes and exclude them from the network or make fraud unattractive for the nodes by paying only fractions of the fee, as suggested in [10]. As already stated above, such credit systems are also used by other virtual currency systems. A detailed analyses of security issues in such systems can be found in [12]. The difference of the credit system used by Cashflow compared to other systems is that the source nodes pays for a channel, along which it can transfer a number of packets, and not for each single packet. The benefit of this concept is a reduction of the credit system’s overhead since the nodes does not have to keep a receipt for every packet. Additionally the price of a channel is known to nodes before they actually open the channel, allowing them to reschedule transmissions. 3.5 Putting Everything Together After the description of the different concepts used by Cashflow, this section briefly introduces the interplay between all the components by providing an usage scenario. To enable nodes to participate in ad hoc networks using Cashflow, they first must acquire a certificate from the bank. As soon as the bank has issued the certificate, nodes can participate in ad hoc networks until their certificates expire without having a connection to the bank. All nodes within an ad hoc network periodically update the basic fee they charge, based on supply and demand of their forward service. Additionally they monitor the quality of wireless links to other nodes. When a node needs to transmit data to another node which is not a direct neighbor, it might have to perform a route search to find the cheapest route to the target node, whereby restrictions concerning link quality must be considered. For the detection of the cheapest route, the route discovery algorithm not only considers the basic fee nodes charge but also penalties caused by the context of intermediate nodes. Provided the target node is part of the network, the route search results in at least one path to the target node. Additionally the node receives information about the price level of different nodes along discovered routes, which they can compare with prices received from previous route discoveries.

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

61

Using routing information, a node can request a channel along discovered paths. The request results in an offer for the channel or a reject, if nodes along the route would be overloaded by the channel. The requesting node can decide to either accept the channel or reject it, if the fee is too high. When the channel is opened, the source node can exchange data with the target node until the channel expires. However, the source node can request an extension of the channel if it wants to continue to exchange data. For the charging, intermediate nodes keep receipts of the channels. As soon as a node can connect to the bank server, it transmits its receipts to transfer the earned fees to its account. Additionally it transmits statistical data which is used for the detection of malign nodes. Depending on the balance of the bank account, the user might have to transfer money from an external source to the bank account, or the bank disburses the money to the node. This is important, since the bank will not issue new certificates to users nodes after the expiration of old certificates if the users account has been overdrawn.

4 Conclusion To the best knowledge of the authors, Cashflow presents the first attempt to realize a virtual currency system where the decision on the participation degree is left to the users. Additionally Cashflow is sensitive to node context, allowing to reflect the willingness of nodes to forward packets by the fee they charge. This was realized by using a market concept for pricing where the node fee is determined by the difference of supply and demand, as well as by user and context specific parameters. A credit system was introduced for charging and accounting which makes Cashflow, in contrast to other virtual currency systems, independent of tamperproof hardware. An additional benefit of the credit system is its ability to detect malign nodes and exclude them from the system, making misbehavior in the network unattractive. In contrast to other virtual currency systems, Cashflow uses a channel concept for data transmission. This concept allows to inform nodes about transmission fees before they start to transmit data. Using this information, nodes can suspend transmissions in the case of temporarily high transmission fees. Hence users have full control concerning transmission costs which is not a matter of course. By combining the market based pricing system with the channel concept of Cashflow and a specifically designed routing algorithm [15], which allows to find the cheapest path between nodes, the system provides an implicit access control and load balancing algorithm. Concerning future work, Cashflow was mainly designed to give users control with respect to their participation degree and to include node context into the pricing. Load balancing and access control functionality can be seen as a kind of positive side effect of the system. Therefore the efficiency of load balancing and access control mechanisms have not been yet completely evaluated and not optimized. However, first simulation results as pictured in Figure 3 and 4 are promising. Both figures show heat maps of a grid network with 32 to 32 nodes, whereas nodes within the area with the dotted black border are inactive and only used for scanning purposes. In both cases four nodes on the left side try to transmit data to four nodes on the right side. The heat maps show network activity whereas red areas indicate overload situations. In the left scenario the pricing function is deactivated which corresponds to an

62

L. Wallentin et al.

Fig. 3. No load balancing

Fig. 4. Load balancing

deactivation of the load balancing function. In most cases the nodes try to use the shortest way to the target nodes resulting in an congestion between the two areas of inactive nodes. If the pricing function is activated, the high price makes paths through the bottleneck unattractive and result in a prevention of local overload, as pictured in Figure 4. Additionally the complete throughput increases by 92.7 percent in this scenario because of the usage of alternative paths. Besides the evaluation and optimization of the access control and load balancing system, the integration of Cashflow in hybrid and internet connected networks and the extension of the system to a platform for additional paid services will be part of future research.

References 1. Perkins, C.E.: Ad Hoc Networking. Addison-Wesley Longman, Amsterdam (2001) 2. Cavalcanti, D., Agrawal, D., Cordeiro, C., Xie, B., Kumar, A.: Issues in integrating cellular networks WLANs, AND MANETs: a futuristic heterogeneous wireless network. IEEE Wireless Communications 12(3), 30–41 (2005) 3. Misra, S., Woungang, I., Misra, S.C.: Guide to Wireless Ad Hoc Networks, 1st edn. Springer, Berlin (2009) 4. Bansal, S., Baker, M.: Observation-based Cooperation Enforcement in Ad Hoc Networks. Technical report, Computer Science Department, Stanford University (2003) 5. Michiardi, P., Molva, R.: CORE: a Collaborative Reputation mechanism to enforce node cooperation in mobile ad hoc networks. In: Proceedings of the Communication and Multimedia Security 2002 Conference (2002) 6. Buchegger, S., Boudec, J.-Y.L.: Nodes bearing grudges: Towards routing security, fairness, and robustness in mobile ad hoc networks. In: 10th Euromicro Workshop on Parallel, Distributed and Network-based Processing (2002) 7. Buchegger, S., Boudec, J.-Y.L.: Performance analysis of the CONFIDANT protocol: Cooperation of nodes - fairness in dynamic ad-hoc networks. In: Proceedings of IEEE/ACM Workshop on Mobile Ad Hoc Networking and Computing (MobiHOC). IEEE, Los Alamitos (2002)

Cashflow: A Channel-Oriented, Credit-Based Virtual Currency System

63

8. Buttyan, L., Hubaux, J.P.: Enforcing service availability in mobile ad-hoc WANs. In: Proceedings of the First ACM Workshop on Mobile Ad Hoc Networking and Computing (MobiHoc) (2000) 9. Buttyan, L., Hubaux, J.P.: Nuglets: a Virtual Currency to Stimulate Cooperation in SelfOrganized Mobile Ad Hoc Networks. In: Tech. Rep. DSC/2001/001, Swiss Federal Institution of Technology (2001) 10. Zhong, S., Yang, Y.R., Chen, J.: Sprite: A Simple, Cheat-Proof, Credit-Based System for Mobile Ad Hoc Networks. In: Proceedings of INFOCOM. IEEE, Los Alamitos (2003) 11. Chen, K., Nahrstedt, K.: iPass: An Incentive Compatible Auction Scheme to Enable Packet Forwarding Service in MANET. In: 24th IEEE International Conference on Distributed Computing Systems (ICDCS 2004), pp. 534–542 (2004) 12. Chen, K.: Cooperative and non-cooperative flow control in mobile ad hoc networks, PhDThesis, University of Illinois at Urbana-Champaign, Urbana (2004) 13. Eidenbenz, S., Resta, G., Santi, P.: Commit: A sender-centric truthful and energy-efficient routing protocol for ad hoc networks with selfish nodes. In: Proc. of 5th IEEE International Workshop on Algorithms for Wireless, Mobile, Ad Hoc & Sensor Networks, IPDPS (2005) 14. Eidenbenz, S., Anderegg, L.: Ad hoc-VCG: A truthful and cost- efficient routing protocol for mobile ad hoc networks with selfish agents. In: Proc. ACM MobiCom, pp. 245—259 (2003) 15. Wallentin, L., Fabini, J., Egger, C., Happenhofer, M.: A Cross-Layer Route Discovery Strategy for Virtual Currency Systems in Mobile Ad Hoc Networks. In: Proceedings of the Seventh International Conference on Wireless On-demand Network Systems and Services (WONS 2010), Kranjska Gora, Slovenia, pp. 91–98 (2010) 16. Davis, M.: A Wireless Traffic Probe for Radio Resource Management and QoS Provisioning in IEEE 802.11 WLANs. In: ACM Symposium on Modeling, Analysis and Simlulation of Wireless and Mobile Systems (ACM MSWiM 2004), Venice (2004) 17. Davis, M., Raimondi, T.: A Novel Framework for Radio Resource Management in IEEE 802.11 Wireless LANs. In: Intl. Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt 2005), Trentino (2005)

Location Management in Heterogeneous VANETs: A Mobility Aware Server Selection Method (Invited Paper) Seyedali Hosseininezhad and Victor C.M. Leung Electrical and Computer Engineering Department University of British Columbia, Vancouver BC V6T1Z4, Canada {seyedali,vleung}@ece.ubc.ca

Abstract. Heterogeneous wireless networks are capable of providing customers with better services while service providers can offer more applications to more customers with lower costs. Many services require the support of location management functions, which still need further innovations to become viable in multi-hop vehicular ad-hoc networks (VANETs). High mobility in vehicular networks causes conventional location management systems to overuse the bandwidth and induce extra handovers to clients trying to synchronize themselves to location servers. We provide a routing algorithm for transactions between location servers and mobile nodes. We assume location servers are vehicles equipped with at least one long range and one short range radio interfaces, whereas regular nodes (clients) only need a short range radio interface. The primary goal of our design is to minimize handovers between location servers while limiting the delays of location updates. Taking advantage of vehicle mobility, we propose a mobility-aware server selection scheme and show that it can reduce the number of handovers and yet avoid large delays during location updates. We model the proposed scheme in NS-2 and apply vehicular mobility patterns generated with SUMO for urban and highway scenarios for performance evaluations. We show that proposed scheme significantly lowers the costs of signaling and rate of server handovers by increasing the connection lifetime between clients and servers. Keywords: Heterogeneous Networks, Mobility Aware Routing, Location Management, VANET.

1

Introduction

Vehicular ad-hoc networks (VANETs) are emerging as one of the most important practical applications of mobile ad-hoc networks (MANETs). As the demand for pervasive computing is increasing, location management becomes one of the most important modules in vehicular networking. Multimedia streams, news broadcasting, entertainment and other applications which require Internet connectivity, peer to peer applications, local advertisements, vehicle pooling and local J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 64–81, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Location Management in Heterogeneous VANETs

65

cab services are some examples of the broad range of feasible applications when vehicles are equipped with positioning and communication equipments [1, 2]. 1.1

Location Management in MANETs

Several protocols have been designed to handle mobility of nodes [3–10]. Location servers are responsible for handling geographical location information of nodes in the vicinity and provide them to others when needed.Different categories of location management have been classified: Flooding based location management is considered as the most straight forward method for passing location information. Due to high redundancy overhead, researchers have strived to decrease unnecessary packet relays. The hypothesis of methods like DREAM [4] is that relative locations of closer nodes are changing faster compared to nodes far away from each other. Therefore location updates are being sent to close location servers more frequently than others. Quorum based location management is another category which is based on assuring a rendezvous between queries and updates. A localized quorum based location service is proposed in [5]. In this method location of every node is dispersed horizontally and vertically. As the authors have stated, this method is proper for networks without a significant relative motion. VANETs with high relative speeds are improper environment for this class of location management systems. The GLS method [6] for distributed location service management divides the area into different degrees of grids in a way that in every grid around the node there is a fixed number of servers that collect location information about that node. As grids grow larger, the probability of a server being chosen for other nodes decreases. This method is not very flexible for highly variant environments like VANETs. Hierarchical methods [9, 10] for server allocation are highly scalable because the rates of location updates are reduced for servers in higher levels. However in this work we would prefer not to consider these methods because of the following reasons. We consider private and public transit vehicles in this work. Private vehicles have intermittent availability and may not be trusted by other vehicles; therefore they are not candidates for selection as location servers. We propose to host location servers in public transit vehicles. Because not every node in a VANET can be selected as location server, and the density of location server candidates will not be very high in most road traffic scenarios, application of hierarchical methods is not justified. To deal with the sparsity of potential server nodes, we shall also propose to utilize long-range radio communications to interconnect these nodes. In high mobility networks such as VANETs, keeping track of location information would in general be a huge overhead on the network if location information is saved in all location servers. Therefore having records of every node in network while nodes are rapidly changing their locations can be performed better if we are able to save these information in a specific set of nodes. Moreover, to reduce the effect of mobility on precision, we are focusing on hop by hop packet relaying rather that finding a deterministic route from clients to servers. We will

66

S. Hosseininezhad and V.C.M. Leung

look at location management as a service to vehicles which is offered by some specific nodes in the area. Therefore we need a service discovery mechanism to find service providers. The main contribution of this paper is the proposal of an efficient routing mechanism for this category of networks with high mobility profiles. In [11] a new metric has been introduced for routing assessment: Longevity of a route. Specially important in MANETs, changes in the path can cause extra signaling and delay overhead during route reconstruction. Therefore they have proposed association between two nodes to measure route longevity. It is assumed that a link is reliable if association between two ends is higher than a certain threshold. Moreover, a routing algorithm is proposed in [12], in which route selection is based on a hybrid criterion of route lifetime and path length. Route lifetime is measured by a definition of link affinity which is calculated based on received signal strength. Since in practical propagation channels the signal strength is not constant over time, RABR can make wrong decisions especially in an urban environment. However, it is possible to utilize the concept of link lifetime as a decision factor in routing but it needs a new measurement tailored for variable conditions of vehicular networks. Many different wireless access technologies can be employed in VANETs. Short-range technologies include wireless local area networks (WLANs) and its variant called Dedicated Short-range Radio Communications (DSRC) targeting specifically vehicular communications. Long-range technologies include cellular networks and wireless metropolitan area networks such as WiMAX. VANETs employing short-range radio access face problems in area coverage and fast handoffs between nodes. Because of high mobility speed, rate of hand-offs in the network becomes a bottleneck in location registration and updates. Heterogeneity can come to the rescue for services like location management. Using long-range wireless access as a higher layer of communications, we can interconnect location servers together as a logical mesh network. We can assume that a connected graph of location servers can exchange signaling messages through this logical mesh network. We shall base our work on utilizing available long-range wireless connections to facilitate location management in VANETs. 1.2

Service Discovery Inspired by Field Theory

Lenders et. al. in [13] defined an approach for efficient and robust service discovery. This concept is similar to anycast routing, which is supported in IPV6 [14]. In anycast routing, an address is associated with more than one interfaces that belong to distinct nodes that are similar in nature. As it is preferable for clients to get service from the nearest among several potential servers, use of anycasting would allow the desired server to be reached easily. From electromagnetic field theory, the point potential of a spot is related to its distance to the maxima potential charge. In wireless networks, the most commonly used definition for distance is based on hop count; nonetheless

Location Management in Heterogeneous VANETs

67

geographical distance is also applicable. In [13], hop count is considered as the distance between nodes: N  Qi ϕ(n) = (1) dist(n, ni ) i=1 where Qi is the potential assigned to server i and ϕ(n) is the total received potential by node n from all servers. The amount of potential assigned to each server could be a factor of their capacity or quality of service (QoS) metrics.

Egress Location Server

Egress Network

Connection Peer

LS1

LS3

Cellular Connectivity

Connection Peer

Connection Peer

LS2

Multi-homed (Public Transit)

LS4

Single-homed (Private Car)

3G/WiMAX link

802.11 link

Fig. 1. Location Management over Heterogeneous Architecture

2

Location Management over Heterogeneous Networks – The Architecture

Fig. 1 depicts a heterogeneous network architecture with partial Internet connectivity. In this system, heterogeneous nodes are connected to each other and edge gateways using their long-range wireless access capability. The requirements and assumptions in aforementioned architecture are: 1. All vehicles are considered to have a mechanism to extract their own geographical location, e.g. using a onboard global positioning system (GPS) receiver. 2. All nodes are equipped with at least one short-range radio (e.g. 802.11a,b,g,p). 3. Some special nodes are equipped with all valid short range communication interfaces and one long range communication interface (e.g. WiMAX) and function as location servers. These nodes are interconnected to each other in a logical mesh network to exchange their location records, and to stationary gateways for Internet access.

68

S. Hosseininezhad and V.C.M. Leung

4. In consideration of valuable licensed-spectrum used for long-range wireless access, the use of long-range radio should be minimized. Therefore it is our goal to reduce the numbers of queries between servers and server handovers for vehicles. 5. Location queries and updates should not be propagated more than a certain number of hops. 6. Server advertisements should not be rebroadcasted more than a certain limit. In one scenario of this architecture, a public transportation system provides a wireless Internet relay service inside an urban area. Public transit vehicles are equipped with multiple radios and they are tasked to provide connectivity and related services to other vehicles. They utilize their long-range radios to relay local data network traffic to stationary gateways and to provide a location management service to vehicles in their vicinity by exchanging location information with other location servers. To advertise location service of servers and receive updates and queries from vehicles, we propose a service discovery mechanism to find routes to location servers in the area with the best matching mobility pattern. We will evaluate the effect of this service discovery method with different scenarios of urban and highway mobility.

3

Mobility Aware Service Selection and Packet Relay

Vehicular mobility patterns (urban or highway) generally follow roadways with probabilistically change of directions at intersections. We assume that every vehicle responsibly sends its location information to a location server. This location information can be used by other vehicles or service providers to present location based services. When a vehicle and its location server move away from each other and the distance grows more than a certain hop distance, path delay and high link breakage probability make their interactions ineffective. Therefore the client has to hand off from the old server to a better server in terms of delay, robustness and lifetime. Every hand-off between two location servers is comprised of several ’server to server’ and ’server to client’ signalling interactions. However, server to server interaction are more expensive because they use licensed spectrum to communicate. Based on expectations and assumptions in the architecture of Fig. 1, if we want to use the approach explained as field theory before, clients should send their location management packets toward other relays or servers in vicinity who have the highest potential. Signalling for a location update comprises of a primary phase of registration between client and server. After the registration, client is able to synchronize the server by sending periodic or event triggered updates. If a client is unable to send updates to designated server, a new registration with an available server is required. Based on the our desired architecture and location management procedure, the field theoretic method reviewed in Section 1.2 has the following deficiencies that should be addressed:

Location Management in Heterogeneous VANETs

69

1. The measure for distance between nodes is unrealistic, since mobility pattern is not considered in server selection. In our case the relative speed between a client and its server defines the connectivity lifetime and we prefer to choose a server that has a higher connectivity lifetime as long as the path delay is less than a certain limit. 2. The server selection is stateless. Service discovery would lead to a set of choices for each relay to forward the packet. However there is no guarantee that a packet will be relayed to the same server which the former packet is sent to. It is desirable for a client to send location updates to a server that it has already registered in. It means that if a client selects a server with highest potential as its location server, all the relay nodes should be notified to relay the packet from that client toward the same server. Consequently a server hand-over does not happen unless the delay threshold is exceeded or disconnection occurs. By modifying the service discovery method proposed in [13], we are going to define a location management method that minimizes hand-overs, which is applicable for geographical and topological location managements. 3.1

Reliability vs. Distance

Hop distance is a simple and effective criterion for route selection, but in cases with high mobility this measure is very unstable. To avoid this problem we propose to use link stability and usability (also known as reliability) as the route selection criteria instead. Let’s denote the set of links in the chosen path between s and d as P (s, d). We want to account for reliability of each link l ∈ P (s, d) and choose a path with highest aggregated reliability. Reliability of a link is directly related to the estimated link lifetime. However calculation of reliability includes error and an unmeasured factor of future alternative connections. For instance, a weak link could be replaced in future by a new relay node which is not present at the moment. Due to this factor, it is not rational to underrate a path by considering the reliability of the weakest link in path as decision factor. In the other hand, we cannot rely on arithmetic average because strong links in the path would cause overestimation of path reliability. Since we want the factor to tend toward the most realistic reliability value of the path (to mitigate the impact of links with excellent reliability in calculation of total path reliability), we desire to have the smallest average value as the measuring factor. Instead of using arithmetic mean, we would use harmonic mean to calculate the reliability factor. Harmonic means tend toward the smallest values compared to geometric and arithmetic means. Hence, with rl being reliability of link l and the number of hops |P (s, d)|= n we have:   1 1   ≤ rl ≤ rl n 1 1 n l∈P (s,d)) rl n l∈P (s,d)

l∈P (s,d)

To apply the value of reliability in routing decision, we define the distance factor in (1) as:

70

S. Hosseininezhad and V.C.M. Leung

D(s, d) =

 l∈P (s,d)

1 reliability(l)

(2)

In every calculation period, each node will predict the locations of current neighbors and based on estimated pathloss exponent of environment and foreseen mobility patterns, calculate the link lifetime. Path prediction and lifetime estimation are two major ongoing research topics and they are explained more in 3.2. Intuitively, a longer link lifetime leads to a more reliable path between nodes and location servers. Notwithstanding, due to high error rates in prediction mechanisms [15], we cannot rely solely on measures of one link. Therefore we will define reliability factor for all (node,server ) pairs to show how reliable the node could be to relay packets toward the server. Based on the assumption that a higher node density can make the route more reliable, we define the reliability factor as the probability of a packet being successfully relayed by a node to another which is closer to server. 3.2

Reliability Measurement

We define link reliability for two neighbors as the estimated expected remaining time of connectivity between the pair. To calculate the link reliability factor we assume that each node will listen to data packets and beacons sent by its neighbors. Using sampled signal characteristics and location information, receiver predicts how link condition will change in future. Therefore, to extract reliability factors, nodes need to have knowledge about future variations in link connectivity. In [16] a method for estimating link residual time and link stability has been proposed. In this method, after denoising and classification of the radio signal strength indications (RSSI) from neighbors, future lifetime is estimated. In [15], Euclidean distance information is utilized to estimate future trajectory. It seems that by using relative mobility between nodes and digital map information together, future estimations can become more precise [17]. A method to calculate the probability of turns in road intersections is proposed in [18], based on the theory that turning options that lead to more destinations in shorter times are more popular than those which lead to local areas or take more time to reach destinations. In [19], mobility behavior of nodes is used to classify their transportation mode. Moreover, using particle filter method they estimate parameters in a Bayesian network for path selection decisions. By learning these parameters, they try to estimate future velocities and turning selections. To calculate the link reliability in a path toward a server, every node will calculate the cumulative probability of connectivity to next hop for each server. The next hop is defined as any node in communication range that has a loopfree path to the server. In practice every node can put a short hash value of its unique address in forwarded advertisement to avoid considering routes which are originated from itself. Hashing can reduce the length of a string to a few bits and yet avoid duplicate indexes with a high probability. We assume that a link connectivity prediction method can provide a process consisting of connectivity probabilities during a prediction period. Suppose for a

Location Management in Heterogeneous VANETs

71

specific environment and mobility pattern, a prediction method is able to predict future motions and channel states for n time units. Moreover, we can extract average percentage of error in lifetime prediction (which can be empirically found) as E(ˆ e). This error extraction could be enforced to system as a-priori known measure or it could be re-adjusted based on observations from past (By comparing prediction and real condition after it happened). Therefore Pl(i) (t) is the probability of link l(i) being connected from t0 (now) until t0 + t and is equal to: ˆ i , t) Pl(i) (t) = (1 − E(ˆ e))P (L (3) ˆ where P (Li , t) represents the link condition (alive/dead) which is calculated using the desired mobility prediction method. As mentioned before, each mobility prediction and link classification method has a distinct estimation capability ˆ i , t) are highly dein different environments. Therefore values of E(ˆ e) and P (L pendent on the method of prediction being used. We will evaluate our method in Section 4 based on a simple linear prediction, but as long as a prediction method yields a prediction of link lifetime and an estimate of the error, it can be integrated in our approach. Having the estimate of link lifetime for all links in the path, the probability of having an undisrupted path from node k toward the server S for the next t time units (complement of the probability of no link being capable of relaying packets from k to S) is:    CSk (t) = 1 − 1 − Pl(i) (t) (4) l(i)∈H k (S)

where H k (S) is set of links between k and its neighbors that have a loop free path to server S. We use the cumulative distribution of C k (t) (for t = 0 · · · tmax with tmax equal to maximum duration of predictability) as a factor which shows how reliable the node k is to pass the packets toward server S: tmax reliability(k) = cdft0 (CSk ) =

t max

CSk (t0 + t)

(5)

t=0

We need to extract the reliability of a node for all servers being discovered. To avoid extra calculations, we define a maximum hop threshold for acceptable potentials received from neighbors. Intuitively it is obvious that information regarding far away servers are not of any interest because of extra relay overhead and delay. Finally we define the distance between client c and server S as:  1 D(c, S) = (6) tmax cdft0 (CSk ) k∈P (c,S) Notice that as hop number increases, the distance is affected and the chance of being chosen as best path decreases. This definition of distance would result in such way that nomadic mobility patterns lead to higher potentials and connection between vehicles with opposite directions and/or sparse connectivity conditions cause less potential dispersion.

72

3.3

S. Hosseininezhad and V.C.M. Leung

Potential Assignment for Path Construction and Server Selection

Using (2) as the distance measure for (1), every node can receive a potential from servers based on the relative mobility and link condition of all nodes in the path from that server to the node. As described in Algorithm 1 every node advertises all valid server information received from neighbors to adjacent nodes. After receiving these advertisements, the node sets the current potential received from a server to the highest received value. These values are valid up to a certain time after last advertisement. Whenever the node wants to send or relay a packet toward a server, it will choose the server with the highest potential. This policy leads to selection of a server which is having the best known mobility correlation with the transmitter. Algorithm 1. Potential Assignment 1: Input servers advertisements[ ] 2: for each servers inf o in servers advertisements 3: L ← servers inf o.source 4: predict link condition(L) 5: servers[L] ← servers inf o 6: for each S in servers inf o 7: rel f actor ← reliability(S) S.pot.original 8: D(S.id) = S.pot.original + rel f1actor S.pot.received  9: pot[S.id] ← max pot[S.id], S.pot.original D(S.id) 10: next hop[S.id] ← arg maxl∈neighbors l.servers[S.id].potential 11: end for 12: end for

3.4

Location Update

To choose a new server for location updates, each node will select the server with the highest accessible potential. After choosing the best server, location updates are sent toward it using the neighbor who has the largest potential from that particular server. Using this approach instead of [13] we can make sure that location update packets will not face misroute to other servers which are not moving in favorable directions. Decision to hand over to another server is performed by a client when the hop distance to the current server has exceeded a certain threshold. Since it is assumed that location update messages are not in a high priority class and their packet size is reasonably small, packet relay in short range wireless would still be more favorable compared to expensive long range network. Anyhow, decision for when to hand off is still open to users and they can choose between prompt updates and lower cost. We will discuss about this trade-off in the next section. Packet relay to a chosen server is also done very easily by comparing received potential of the server from current alive links and therefore source routing is not necessary.

Location Management in Heterogeneous VANETs

3.5

73

Location Query

To find the location of another user, a requester would send a location query to the best available server at the moment (the one with the highest available potential). The packet relay mechanism would be similar to that for location updates. After receiving the query, the server looks up in its local database to see if it has up-to-date location entry. Otherwise it will send a query to its neighbors using long range wireless. Since we assume that long range media would lead to a connected graph topology, queries will have answer from one of the servers and this answer will reach the original server.

4

Performance Evaluations

To evaluate performance of our proposed framework, we have modeled the system using the NS-2 network simulator [20]. We have added a new service discovery agent over the currently implemented network stack and added our logic as an application agent. Using application agent, we can use any routing algorithm for packet routing mechanism. We have tried our protocol on several test scenarios. These scenarios are based on realistic vehicular traffic generated by SUMO network mobility generator [21]. This microscopic vehicle traffic generator is able to create mobility patterns based on defined traffic flows. The trace generated by SUMO is a mobility log for vehicles moving based on road and traffic regulations. We can import different maps to SUMO to generate different test cases. We have imported several maps with different key features. The first imported map is a 10 Km long highway with 2 lanes in each direction. Two kinds of vehicles have been considered to commute on the road: private vehicles with short-range radios and public transit vehicles equipped with long-range and short-range radios. The two categories of vehicles have different characteristics in speed limit, acceleration and deceleration. The second scenario is realistic urban area extracted from actual street maps. These maps are extracted from free maps available in OpenStreetMap [22]. After adding traffic lights to map, we have used SUMO to generate traffic information for 10000 seconds. The procedure of map extraction and simulation has been shown in Fig. 2. After generation of mobility traces, they are fed into NS-2 as mobility scenario and simulation is performed by NS-2. Since we need prediction in our method and it is not performed in NS-2, we do the simulation twice; The first run is done to extract exact location of every

OpenStreetMap

1

SUMO

NS-2

2 3

4

Mobility Prediction Fig. 2. Simulation Procedure

5

Output Analysis

74

S. Hosseininezhad and V.C.M. Leung

vehicle during the simulation. Then we use this data in the next run as a precise prediction of future mobility patterns in network. To make the prediction more realistic, we add noise to location information. Since prediction precision is strongly dependent on prediction mechanism, we use one of the simplest predictions: In every second, each node predicts x(t + 1) = v + x(t), where x(t) is the location of node in time t (0 ≤ t ≤ tmax ). tmax is the maximum time that prediction can be reasonably valid. We will set tmax to a value which:  ˆ + P r(C(t))P r(C(t)) ˆ  > threshold E P r(C(t))P r(C(t)) 1 < t < tmax

(7)

This is the sum of expected probability for having a true guess, whether positive or negative, on having a connection. This probability should be more than a certain threshold. To find this value we run the simulation and calculate the predicted and actual locations in future. We consider a link as active if its RSSI is more than a threshold. Since measurement of RSSI is impacted by environmental clutters, it is impractical to deterministically define the link connectivity threshold. So we use the propagation model in [23]: Pr (d) = Pt − P L(d) = Pt − (P L(d) + Xδ )

(8)

P L(d) is the log-distance path loss from transmitter to receiver and Xδ is a zeromean Gaussian distributed random shadowing effect with standard deviation δ. Values of path loss exponent and δ are usually extracted from empirical data. We have borrowed these values from the experiment done by Otto et. al. in [24]. Finally, the probability of RSSI being more than γ (dBm) in distance d(m) is: P r[Pr (d) > γ(dBm)] = Q(

γ − Pr (d) ) σ

(9)

Fig. 3 shows the estimated error in aforementioned prediction method. Results show that in highway scenario prediction is performing close to reality and connection condition after 40 seconds is predicted correctly with a 70% probability. However, in suburban scenarios and downtown areas, nondeterministic stops and turning probabilities causes prediction error to grow. For downtown scenario we find that predictions are 50% successful only for 20 seconds ahead. In suburban areas with less stops and turns compared to downtown, it is up to 35 seconds. We apply these errors in calculating path reliability factor for each scenario. To avoid excessive delay caused by late hand-offs we have to set a threshold for maximum allowable hop distance between nodes and servers. The trade-off is between location update cost (which is related to amount of relayed data and type of media used for it) and end-to-end delay. ⎧ d(i, S) (fu .LU + fq (LQ + LR)) ⎪ ⎪ ⎪ ⎨ X= ⎪ K ∗ (N − 1) ∗ SY N+ ⎪ ⎪ ⎩ d(i, Snew )(fu .LU + fq (LQ + LR))

d(i, S) < thr (10) otherwise

Location Management in Heterogeneous VANETs

75

1 Downtown Suburb Highway

0.9 0.8

Average Prediction Error

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0

5

10

15 20 25 Prediction Period(s)

30

35

40

Fig. 3. Linear prediction average error for three scenarios

Table 1. Variables definition in (10) d(i, S) fu fq LU LQ LR K N SY N

Hop distance between i and Server Si Frequency of location update messages Frequency of location query messages Size of location update message Size of location query message Size of location reply message Usage cost/Kb for long range network Number of location servers Size of synchronization message

To calculate the proper threshold, we set our objective to minimize Cost ∗ Delay for location update packets. Using (10) as the cost function and by knowing d(i, S  ), the distance between a node and its second best server with eligible hop distance, every node can calculate the threshold as follows:  K ∗ (N − 1) ∗ SY N  thr = d(i, S ) + .d(i, S  ) (11) fu .LU + fq (LQ + LR) Here we assumed that delay is only based on hop distance and did not consider the delay caused by collisions.

76

S. Hosseininezhad and V.C.M. Leung

Average Client−Server Connection Lifetime(s)

350 300 250

LT SP−1 SP−2 RABR Quorum

200 150 100 50 0

10 25

50

100 150 Cars/Km2

200

250

Fig. 4. Average Client-Server Connection Lifetime (Downtown)

After finding the estimation errors, we run the algorithm based on these estimation properties. We try to establish connection between nodes with heterogeneous connectivity (as servers) and nodes with single wireless network. We compare client-server path length and traffic cost compared to three other methods: [13] (shortest path is the metric for route selection), [12] (affinity based on signal-to-noise ratio) and [5] (quorum based method with column based advertisement and row based query). Fig. 4 shows the average lifetime of connections between mobile nodes and location servers in downtown mobility pattern. Hereafter we will refer to our method as Life Time based method (LT). SP-1 represents shortest path anycasting based on [13]. In SP-2 we use the same method as SP-1 but whenever a server is selected for a node as a location server, it will remain chosen as long as their distance is less than maximum hop distance. Results show that using lifetime as the distance metric has led to significant connection lifetime improvement specially for higher densities. In affinity based method, SNR is considered as the measuring factor for decision making. Therefore for downtown areas with highly volatile SNR conditions RABR can not perform much better than SP-2. Since in quorum based method, chosen servers are changed rapidly after any change in topology, connection lifetimes are not comparable to other methods. Fig. 5 shows the same measures as Fig. 4 but for highway scenarios. Results show 57% overal improvement in connection lifetime compared to SP-2. Especially in lower vehicle densities, our proposed method achieves more improvements compared to the shortest path method because of the steady mobility of vehicles which leads to higher lifetime if paths are selected from vehicles moving along the same direction. RABR performs better in highway scenarios duo

Location Management in Heterogeneous VANETs

77

Average Client−Server Connection Lifetime(s)

800 700 600 500 LT SP−1 SP−2 RABR Quorum

400 300 200 100 0

10 25

50

100

150

200

250

Car/Km2

Fig. 5. Average Client-Server Connection Lifetime (Highway)

to less perturbations in SNR. However, quorum based method is following the same behavior in downtown. Fig. 6 compares the overhead caused by location update packets. Quorum based method uses several location servers, hence location updates become costly. Moreover, as mobility and interactions between nodes increase, the overhead of quorum based method increases drastically. We have assumed that the quorum based method uses WLAN and WiMax based on availability with no preferences. To compare the proportion of WLAN usage, we assumed that parameter K in (10) is equal to 100 (every transmission on WiMax is 100 time more expensive than WLAN). We can see that usage of WiMax in low traffic densities is significantly low and as mobility patterns grow more dynamic, the difference between LT and other three methods become noticeable. Fig. 7 depicts the normalized total cost of location management for different values of K. Since the cost of RABR is very close to SP and the cost of the quorum based method is significantly higher than other methods, we have only compared the cost of SP vs. LT. As one can observe, for K ≥ 100 our method outperforms the shortest path method. K can be interpreted as a priority or preference parameter and could be tuned based on the trade-off between delay and cost efficiency. If providers prefer faster and more precise location updates, they can decrease K. In contrast, for better efficiencies (e.g. for less location sensitive applications) higher values will lead to better spectrum conservation. In Fig. 8 we have shown the average delay experienced by signaling packets (Location Update, Location Query and Responses) for downtown scenario. It is important to mention that the quorum based method uses more than one

78

S. Hosseininezhad and V.C.M. Leung

RABR SP LT

250

Quorum

Car Density (Cars/Km2/Minute)

Quorum RABR SP LT

200

Quorum RABR SP LT

150

Quorum RABR SP LT

100

Quorum

50

RABR SP LT

25

Quorum RABR SP LT

10

Quorum RABR SP LT

0

0.1

WiFi WiMax 0.2

0.3 0.4 0.5 0.6 0.7 Scaled Proportion of Sent Update Packets

0.8

0.9

1

Fig. 6. Normalized overhead of update packets being sent over WLAN and WiMax (K=100)

1

Normalized Cost

0.8

LT SP

0.6

0.4

0.2

0 0 10

1

10

10

2

3

10

10

4

10

5

Cost Factor (K) Fig. 7. Signaling cost based on cost factor (K) in (10)

location management server and always every immediate neighbor in the same row of the node is acting as a location server. As a result signaling delays for location updates and responses are very low. However, when it comes to location query, signaling delay is relatively higher than updates. Our method is always performing better than RABR in terms of delays but compared to SP, it suffers

Client−Server Message Passing Delay (ms)

Location Management in Heterogeneous VANETs

79

1200 LT SP RABR Quorum

1100 1000 900 800 700 600 500 400 300 200 0.01

0.1

0.3

0.4

0.5

Bus/Car Ratio Fig. 8. Average Client-Server Message Passing Delay in Downtown Query/Response messages are considered for Quorum Based Method)

(Only

a 20% increase in delay. As the number of location servers (in this case described as buses) increases, the overall signaling delay for all methods decreases. We can see that if half of the vehicles in our system could act as location servers, delay would have become as low as SP and quorum based methods.

5

Conclusion

Location management is a critical part for vehicular ad-hoc networks. In this paper we have assumed that some of the mobile nodes in vehicular network are equipped with heterogeneous wireless connectivity. These vehicles are able to act as location servers for other vehicles and cooperate using their long-range radio. We have proposed a new server selection and packet relay mechanism that minimizes the rate of server handovers by relaying location update packets toward the server that has the lowest possibility of disconnection. This is done by proposing a new definition of distance. The proposed method has been evaluated by extensive computer simulations. The results show significant improvements in client-server connection lifetimes. Higher connection lifetimes lower the costs of handovers, which require the use of long-range communications to update the record for the client at all the servers. We have provided a tuning factor which can be used for decision making based on tolerable delay and cost. The comparison has been made against three methods: associativity based routing, shortest path selection and quorum based location management. Results show that in scenarios with high mobility our method achieves the lowest costs and acceptable delays compared to other three methods.

80

S. Hosseininezhad and V.C.M. Leung

References 1. C2c-cc manifesto. Technical report, Car 2 Car Communication Consortium (2007) 2. Willke, T., Tientrakool, P., Maxemchuk, N.: A survey of inter-vehicle communication protocols and their applications. IEEE Communications Surveys and Tutorials 11, 3–20 (2009) 3. Friedman, R., Kliot, G.: Location services in wireless ad hoc and hybrid networks: A survey. Technical report, Department of Computer Science,Technion (2006) 4. Basagni, S., Chlamtac, I., Syrotiuk, V., Woodward, B.: A distance routing effect algorithm for mobility (DREAM). In: Proceedings of the 4th Annual ACM/IEEE International Conference on Mobile Computing and Networking, pp. 76–84 (1998) 5. Liu, D., Stojmenovic, I., Jia, X.: A scalable quorum based location service in ad hoc and sensor networks. In: 2006 IEEE International Conference on Mobile Adhoc and Sensor Systems (MASS), pp. 489–492 (2006) 6. Li, J., Jannotti, J., De Couto, D.S., Karger, D.R., Morris, R.: Scalable location service for geographic ad hoc routing. In: Proceedings of the Annual International Conference on Mobile Computing and Networking (MOBICOM), pp. 120– 130 (2000) 7. Saleet, H., Langar, R., Basir, O., Boutaba, R.: Proposal and analysis of regionbased location service management protocol for vanets. In: IEEE Global Telecommunications Conference (GLOBECOM), pp. 491–496 (2008) 8. Dikaiakos, M.D., Florides, A., Nadeem, T., Iftode, L.: Location-aware services over vehicular ad-hoc networks using car-to-car communication. IEEE Journal on Selected Areas in Communications 25, 1590–1602 (2007) 9. Kieß, W., F¨ ußler, H., Widmer, J., Mauve, M.: Hierarchical location service for mobile ad-hoc networks. ACM SIGMOBILE Mobile Computing and Communications Review (MC2R) 8, 47–58 (2004) 10. Ahmed, S., Karmakar, G., Kamruzzaman, J.: Hierarchical adaptive location service protocol for mobile ad hoc network. In: Proceedings of the 2009 IEEE Conference on Wireless Communications & Networking Conference, The Institute of Electrical and Electronics Engineers Inc., pp. 2932–2937 (2009) 11. Toh, C.: Associativity-based routing for ad hoc mobile networks. Wireless Personal Communications 4, 103–139 (1997) 12. Agarwal, S., Ahuja, A., Singh, J., Shorey, R.: Route-lifetime assessment based routing (RABR) protocol for mobile ad-hoc networks. In: IEEE International Conference on Communications. Citeseer, vol. 3, pp. 1697–1701 (2000) 13. Lenders, V., May, M., Plattner, B.: Service discovery in mobile ad hoc networks: A field theoretic approach. Pervasive and Mobile Computing 1, 343–370 (2005) 14. Hinden, R., Deering, S.: IP Version 6 Addressing Architecture, RFC 2373 (1998), http://ietfreport.isoc.org/rfc/rfc2373.txt 15. Haas, Z.J., Hua, E.Y.: Residual link lifetime prediction with limited information input in mobile ad hoc networks. In: Proceedings of IEEE INFOCOM, pp. 26–30 (2008) 16. Sofra, N., Leung, K.K.: Link classification and residual time estimation through adaptive modeling for vanets. In: IEEE VTC, pp. 1 –5 (2009) 17. Menouarand, H., Lenardi, M., Filali, F.: Movement prediction-based routing (mopr) concept for position-based routing in vehicular networks. In: IEEE Vehicular Technology Conference, pp. 2101–2105 (2007)

Location Management in Heterogeneous VANETs

81

18. Krumm, J.: Where will they turn: predicting turn proportions at intersections. (Personal and Ubiquitous Computing, Springer) 19. Patterson, D., Liao, L., Fox, D., Kautz, H.: Inferring high-level behavior from lowlevel sensors. LNCS, pp. 73–89 (2003) 20. The network simulator ns-2, http://www.isi.edu/nsnam/ns/ 21. Simulation of urban mobility, http://sourceforge.net/apps/mediawiki/sumo/ 22. Open street map, http://www.openstreetmap.org/ 23. Rappaport, T.: Wireless communications: principles and practice. Prentice Hall PTR, Upper Saddle River (2001) 24. Otto, J.S., Bustamante, F.E., Berry, R.A.: Down the block and around the corner: The impact of radio propagation on inter-vehicle wireless communication. In: Proceedings - International Conference on Distributed Computing Systems, Montreal, QC, Canada, pp. 605–614 (2009)

Insights into the Routing Stability of a Multi-hop Wireless Testbed (Invited Paper) Mehdi Bezahaf1 , Luigi Iannone2 , Marcelo Dias de Amorim1 , and Serge Fdida1 1

2

LIP6/CNRS – UPMC Univ Paris 06, France {bezahaf,amorim,sf}@npa.lip6.fr Technische Universit¨ at Berlin – Deutsche Telekom Laboratories, Germany [email protected]

Abstract. By nature, links in multi-hop wireless networks have an unpredictable behavior, which directly affects the stability of routes. In this paper, we investigate the stability of the network by addressing some interesting questions related to the presence of both dominant and subdominant routes between nodes, in a real deployment. We focus on the persistence of the dominant route and the first four sub-dominant routes. The persistence is computed as the percentage of time that a given route is used. We note that source-destination pairs mostly use the dominant route and two sub-dominant routes to communicate, but with a low persistence. We also investigate the number of hops crossed by these routes and their impact on the stability. It turns out that the larger the number of hops, the larger the number of sub-dominant routes. However, when exceeding four hops, the notion of dominance fades. Keywords: wireless mesh networks, multi-hop wireless testbed, routing stability, link stability.

1

Introduction

Wireless multi-hop networks consist in a set of wireless nodes, which may or not be mobile, communicating through wireless links. In such networks, intermediate nodes relay packets in order to reach the final destination (potentially a gateway to the Internet). The ordered list of wireless nodes crossed by these packets is referred to as a path or route. A myriad of routing protocols have been proposed to build paths that are robust enough to overcome the natural limitation of wireless routes [6,7,11]. Indeed, due to the specificities of wireless links (e.g., interference, collisions, and fading), paths have an intrinsic unstable behavior when compared to routes in wired networks. Despite the countless routing protocols proposed so far, relatively little research work has been done on how to take into account the notion of route stability in route selection algorithms. Some routing protocols such as SSA, and ABR [4,13,8] use models based on signal strength or pilot signal to estimate the link stability. Paul et al. proposed the “affinity” parameter, which J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 82–97, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Insights into Routing Stability

83

consists in predicting links’ lifetime [9]. This parameter is based on the strength and the stability of relationships of a node and its neighbors. Inspired by this idea, Agarwal et al. proposed a routing algorithm that incorporates the link lifetime prediction on the basis of affinity appraisal [2]. Authors also considered the path length while choosing an optimal route for a TCP source. The only study that tries to shed some light on the stability behavior of wireless paths is the one of Ramachandran et al. [12]. They collected, through wireless beacons, information about link quality and then computed, on a per-minute basis and offline, all possible routes between sources and destinations using an implementation of the Dijkstra shortest-path algorithm. In this paper, we propose to go deeper into the analysis of routing stability by basing our analysis on routing events in a real testbed. We focus on a multihop wireless mesh network composed of static backbone nodes [3]. We address the following questions: What is the proportion of time during which which the dominant route is active? Is this dominant route persistent? Are there any subdominant routes? If yes, what about their dominance? How long does each one of them last? How does the number of hops affect this dominance? By providing answers to these questions, we come up with the following contributions: – We propose a measurement-based study of routing stability in a real multihop wireless mesh network testbed deployment, running a homemade pythonbased implementation of the DSDV (Destination-Sequenced Distance Vector) routing protocol. – We investigate routing stability in terms of dominance and persistence of dominant and sub-dominant routes, also evaluating the behavior of routes when the number of hops increases. – We observe that source-destination pairs mostly use a dominant route and two different sub-dominant routes to communicate between them, but with a low persistence. Moreover, we also observe that with a larger number of hops, a larger number of sub-dominant routes appears. However, when exceeding four hops, the notion of dominance disappear. The remainder of the paper is organized as follows. In Section 2, we describe our testbed deployment, introduce the measurement setup, and define the scenario of our experiments. In Section 3, we investigate the dominance of the first most used route between each source-destination pair. Then, in Section 4, we investigate the persistence of the rest of the most used routes. In Section 5, we evaluate the impact of hops on routes dominance. We conclude the paper in Section 6 by summarizing our main results.

2

Experimentation Setup

Our analysis of routing stability is based on routing information collected from a real testbed deployment (namely MeshDVnet). We start this section by briefly describing this testbed. We discuss then the topology of the network and the experimentation scenario. We finally explain how we extract route information after collecting all routing tables in a central database.

84

M. Bezahaf et al.

ath1

ath0

eth0

(a) Physical nents.

compo-

(b) Architecture of a node.

Fig. 1. MeshDVnetBox: our wireless mesh network testbed box

2.1

The MeshDVnet Testbed

Based on IPv6, MeshDVnet is a wireless mesh network testbed deployed at the computer science laboratory of the Universit´e Pierre et Marie Curie [5]. This testbed offers traditional wireless LAN access connectivity to clients, while forwarding all generated traffic between clients and toward the gateway through the wireless routers in a multi-hop fashion. This testbed shares the wireless medium with the lab’s wireless network, the other companies’ networks in the same building, as well as neighborhood building’s networks. MeshDVnet is logically decoupled into two subnetworks: one formed by the set of routers, which constitutes the backbone, and the other formed by the set of clients. Each router consists of a Linux kernel v2.6.19 (Slackware distribution) running on Soekris boards net4521 (AMD ElanSC520 133 Mhz) and net5501 (AMD Geode LX processor 500 Mhz) with two wireless cards (IEEE 802.11 a/b/g) using Madwifi driver v0.9.4 [1] (cf., Fig. 1). The first wireless card (ath0 in Fig. 1(b)) is configured in ad hoc mode for backbone communications, while the second wireless card (ath1 in Fig. 1(b)) is configured in access point (AP) to give access to clients. Furthermore, an Ethernet interface (eth0 in Fig. 1(b)) is used for maintenance, supervision, and data collection (but not for traffic forwarding). MeshDVnet backbone is built with thirteen routers placed around our lab’s building (in a single floor). We have an associated supervision web page that refreshes every two minutes to monitor if routers are up and running. The web page shows also the links between each router and their signal quality. It is publicly available (IPv6 only) at http://www.infradio-jussieu.lip6. fr/supervision/supervision-mesh-kennedy.html.

Insights into Routing Stability

85

SQL Server Ethernet r08

r07

r03

r05

r06

r02

r11

Atrium

r04 r09

r13

Offices Routers’ Offices Corridors Backbone links Router

r01

r10

r12

Good signal strength

Bad signal strength

Fig. 2. Routing information collection infrastructure: thirteen routers communicating through wireless interface, and sending their routing information through the Ethernet interface to the central postgreSQL server

2.2

Measurement Setup

For the experimentation, we deployed thirteen routers in our offices as shown in Fig. 2. In order to have a realistic study with a real routing protocol and real routing message exchanges, we use the well-known proactive routing protocol DSDV (Destination-Sequenced Distance Vector) to build and maintain paths between routers [10]. Furthermore, in order to study the effect of the number of hops on the stability of links, we use hop-count as the metric for route selection. We chose DSDV as routing protocol in this first campaign of measurements for the simple reason that it is the protocol readily available on the MeshDVnet testbed. It is out of the scope of this paper to compare or evaluate the performance of ad hoc routing protocols. Rather, we focus on illustrating the stability of wireless links using a real routing protocol through a measurement-based study. Clearly, the choice of the routing protocol and the routing metric can have, in some way, side effects on the stability of links; however, there is no escape to this situation, since this is part of any real measurement campaign. To the best of our knowledge, this is the first study on stability of wireless links based on measurement and taking into account a real routing protocol. As discussed in Section 6, an analysis of the impact of different routing protocols on the stability of wireless links is left for future work.

86

M. Bezahaf et al. Router r08 t1

SQL Server

Routing Table dst next hop nbr of hops r13 ...

t2

r05 ...

Database

4 ...

Routing Table dst next hop nbr of hops r13 ...

r04 ...

src

dst next hop time

r08 r08 ...

r13 r13 ...

r05 r04 ...

t1 t2 ...

SQL request

3 ...

U DP toSQL.py eth0 t2 r08 r13 r04

t1 r08 r13 r05

UDP Packets

Fig. 3. Database update: router r08 detects at t1 time a new route to reach router r13 . It sends a UDP packet to the SQL server to add this information to the database. At t2 time it shifts to a new route, and updates the database with the new entry.

2.3

Routing Information Collection

Besides the thirteen routers, we deployed a central postgreSQL server to collect all information that we need for our experimentation (cf., Fig. 2). This information is collected through the Ethernet interface of each router (to avoid biasing the wireless medium with interfering traffic). We revisited the DSDV implementation in order to provide the routers with the capability to send their routing table to the database each time a routing table change occurs. With this infrastructure, we are able to centralize the routing information as seen by the routers itself, without any offline computation. The only post-processing we perform consists in the computation of the various statistics that we present in the second part of the paper. During of the experimentation, which lasted a total of four consecutive days, at each change in a routing table the concerned router notifies in real time this change to the database by sending a UDP packet through its Ethernet interface. In the postgreSQL server, we wrote a script that receives these UDP packets and translate them to a SQL insert (cf., Fig. 3). In this way, we obtain in real-time a database with all changes between all possible source-destination pairs during the experimentation. Finally, to retrieve all routing history, and to compute dominant route and sub-dominant routes, we execute a python script that iteratively queries the database.

3

Dominant Route

We define the dominant route for source-destination pair as the most used route in terms of cumulative duration. The other routes are referred to as sub-dominant routes.

Insights into Routing Stability

87

Percentage of dominant routes’ activity time

100 90 80 70 60 50 40 30 20 10

r01 r02 r03 r04 r05 r06 r07 r08 r09 r10 r11 r12 r13 Source routers

Fig. 4. Percentage of time the dominant route was used between each sourcedestination pairs

One of the first questions that we address in this paper is: what is the proportion of time of the dominant route? In order to answer this question, we first extract from our database all possible routes between any pair of nodes in order to compute the percentage of time the dominant route is active. The result is shown in Fig. 4. For each router (x-axis) there are 12 bars (one for each other router in the topology). Note that this figure aims at showing a rough view of route activity; the details are given in the following. The main observation is that most of the source-destination pairs use the same route more than 40% of the time. Furthermore, there are some source-destination pairs that use the same route during all the experimentation (100% of time). 3.1

Route Persistence

A natural consequence of the previous observations concerning Fig. 4 is the question: are the dominant routes used once for a long time or are they used several times for short periods? The answer to this question is obtained by computing the persistence of the dominant route for each pair. The persistence is computed as the average of the contiguous durations of the dominant route for each pair of nodes. In other words, it is the average time a dominant route is used before a different route is selected. The results are shown in Fig. 5 (recall that we have 13 nodes, which leads to 13 × 12 = 156 different dominant route IDs). We observe that 90 pairs of routers use in

88

M. Bezahaf et al.

Persistence of dominant route (minute)

120

100 5 80

4 3

60 2 1

40

0 0

20

40

60

80

100

120

20

0 0

20

40

60

80

100

120

140

160

Dominant route id Fig. 5. Persistence of the dominant route between each source-destination pair with a zoom between 0 and 5 minutes. Routes’ IDs are sorted relatively to the persistence of the dominant route.

average their dominant route less than one contiguous minute (with very low confidence interval), which represents 57.7% of the cases, while 20% of them use their dominant route during one to five minutes. For the remaining 22.3% of the pairs, the duration ranges from five up to ninety-five minutes, with larger confidence intervals. An important point is the evaluation the distribution of the route persistence. We selected four sample points corresponding to the persistence of dominant route of four source-destination pairs (respectively r02 → r06 , r12 → r10 , r02 → r10 , r13 → r03 ), where each sample point represents a specific region of the curve in Fig. 5. In Fig. 6, we show the distribution of these four points. For each plot in this figure, the x-axis corresponds to the xth time the dominant route is used. For instance, in the plot of the first sample, 70 in the x-axis corresponds to the 70th time the dominant route between r02 and r06 was used, and its persistence is given by the value in the y-axis value. Note that most of the points are concentrated around the mean value, which indicates that the mean persistence is a reliable value. However, in samples 1 and 2, persistence of the dominant route varies from 1.04 minutes to 191.4 minutes, and from 0.31 minutes to 140.07 minutes, respectively the MIN and the MAX values, which causes the confidence interval (CI) to increase its value. We can also observe that in samples 3 and 4 the dominant route oscillates more than in samples 1 and 2 (cf., N P values in Fig. 6), but varies less in samples 3 and 4 than in samples 1 and 2 (cf., MAX/MIN values in Fig. 6).

Insights into Routing Stability

Persistence of dominant route (minute)

Sample 1: src = r02 , dst = r06

Sample 2: src = r12 , dst = r10 mean= 28.64min

mean= 51.39min D= 99.88% MAX= 191.4min CI= 8.86min 150

D= 99.80%

MAX= 140.07min CI= 4.02min

150

200

89

100

100

50

50 0

0

MIN= 1.04min 0

10

20

30

MIN= 0.31min

NP= 78 40

50

60

70

0

D= 93.49%

MAX= 10.72min CI= 0.04min 10 8 6 4 2 0

40

60

80

100

120

140

160

180

Sample 4: src = r13 , dst = r03

Sample 3: src = r02 , dst = r10 mean= 1.43min

20

NP= 188

mean= 0.76min D= 66.07% MAX= 7.00min CI= 0.02min

8 6 4 2 0

MIN= 0.09min NP= 4672

MIN= 0.10min NP= 2666 0

500

1000

1500

2000

2500

0

500 1000 1500 2000 2500 3000 3500 4000 4500

Occurrence of the dominant route Zone where dominant route persists less than the mean Zone where dominant route persists more than the mean Fig. 6. Distribution of persistence of the dominant route between four sourcedestination pairs (r02 → r06 , r12 → r10 , r02 → r10 , r13 → r03 ). Mean is the limit between the two zones. In the graphs, D represents the dominance ratio, CI is the confidence interval, and N P is the total number point of dominant route’s oscillation.

3.2

Route Oscillation

We evaluate now how much the dominant routes oscillate. We define the oscillation of a route as the number of times during the experimentation that this route appears between the same source-destination pair. Fig. 7 shows the results sorted by their dominance. We observe that 90 dominant routes, which represents 57.7% of all dominant routes, oscillate in average 5, 000 times (in four days), where their dominance varies between 14% and 85%. We also notice that a route can be dominant up to 57.7% even with a high oscillation; however, the higher the dominance, the higher the stability of the route (less oscillation). Fig. 8 shows which kind of dominant route is the most persistent according to its dominance. The obtained result confirms the previous one. In fact, we see that

M. Bezahaf et al.

8000

100 90

7000

Number of oscillation

80 6000 70 5000

60

4000

50 40

3000

30 2000 20 1000

10

0 0

20

40

60

80

100

Dominant route id Oscillation

120

140

Percentage of dominant routes’ lifetime

90

0 160

Dominance

Fig. 7. Dominance vs. oscillation

120

100

80 100

16 14

98

10

96

8 94

40

80

12

60

60

6 4

92

40

2 90

20

0 105

120

135

20

0 0

20

40

60

80

100

Dominant route id Dominance

120

140

Persistence

Fig. 8. Dominance vs. persistence

0 160

Persistence of dominant route (minute)

Percentage of dominant routes’ lifetime

100

Insights into Routing Stability

91

routes with a high dominance level are the most persistent, which corresponds to the previous result (i.e., dominant routes with a high dominance ratio are more stable).

4

Sub-dominant Routes

Percentage of dominant routes’ lifetime

The results presented in Section 3 clearly show that, during most of the time, source-destination pairs use the dominant route to communicate. What happens in the rest of the time? In order to know if there exist sub-dominant routes, we apply the previous analysis to the remaining routes. We start our investigation with the dominance ratio of the four most prevalent sub-dominant routes. We sort all source-destination pairs according to the dominance ratio of their dominant route. Fig. 9 shows the percentage of activity of the main dominant route and the correspondent four main sub-dominant routes. As expected, the larger is the ratio of the dominant route, the lower the number sub-dominant routes used during the remaining time. Fig. 10 shows the persistence of the sub-dominant routes. For the sake of clarity, we consider only the first 100 pairs for which the persistence of their dominant route is less than 2 minutes. Exceeding this value, the persistence of sub-dominant routes is negligible. We observe that the persistence of the first and second sub-dominant routes is equivalent to the persistence of the main dominant route, when the latter lasts less than half a minute. We also observe an interesting phenomenon. Some first and second sub-dominant routes are more

100

80

60

40

20

0 0

20

40

60

80

100

120

140

160

Couple of router id Main dominant route 1st sub-dominant route 2nd sub-dominant route

3rd sub-dominant route 4th sub-dominant route

Fig. 9. Dominance of the five most used routes for each source-destination pair

M. Bezahaf et al.

Persistence of dominant route (minute)

92

2

1.5

1

0.5

0 0

20

40

60

80

100

Couple of router id Main dominant route 1st sub-dominant route 2nd sub-dominant route

3rd sub-dominant route 4th sub-dominant route

Fig. 10. Persistence of the five most used routes for each source-destination pair

persistent than the dominant route in some cases, when the persistence of this latter is quite small. In fact, a dominant route that oscillates a lot can be less persistent than a less dominant route with long rare apparition. These cases correspond to source-destination pairs that use some rare routes to communicate (e.g., the corridor in the middle of the topology shown in Fig. 2).

5

Impact of the Number of Hops

In the following, we investigate the influence of the number of hops on the stability of routes. We consider alternative routes, persistence between pairs, and the possible use of sub-dominant routes. We define physical hops between a source and a destination as the geographical distance between them in term of hops (network vision). Given that the routers in our testbed are static, the number of physical hops between a source and a destination is static as well. We also define logical hops as the number of hops used by the routing protocol. In fact, a destination can be in the range of the source (one physical hop), but given the links’ perturbation, the source can choose another route with more logical hops. To better understand the difference between physical and logical hops let us have a look to Fig. 11. For instance, the pair (r01 , r12 ) in Fig. 11(a) is two physical hops pair, even if r01 uses the three logical hops route (r01 → r13 → r10 → r12 ) to reach r12 . Given the specific topology of our scenario, each router has two neighbors at X physical hops. For instance, the router r01 in Fig. 11(b) has [r13 , r10 ] as neighbors at one hop, [r09 , r12 ] at two hops, and [r04 , r11 ] at

Insights into Routing Stability

r13 r01

three hops two hops one hop

r12

r10

93

...

Physical hop Logical hop (a) Physical Vs. Logical hops.

r04

r09 r13 r01

r10 r12

...

r11

(b) Number of hops for r01 .

Fig. 11. Physical and logical hops

three hops. We obtain than 26 pairs of routers at each hop level (13 routers × 2 neighbors at each hop). Using the notion of physical hops, we classify all source-destination pairs in six classes. Each class corresponds to the number of hops between the source and the destination (a class is called X-hop class if the destination is localized at X physical hops from the source). Note that in the rest of the paper, hops will refer to physical hops. 5.1

Available Routes

We define the number of available routes as the number of different routes used between a source-destination pair during the experiments. Fig. 12 shows the evaluation of this number between routers at different level of hops. We observe that more we increase the number of hops, more the number of distinct routes increases, which is obvious. However, we notice that for each

300

Number of distinct routes

270 240 210 180 150 120 90 60 30 0

5

10

15

20

25

Pair of router id 1-hop 2-hops

3-hops 4-hops

5-hops 6-hops

Fig. 12. Evolution of number of distinct routes with the number of hops. Each curve is sorted according to the number of distinct routes.

94

M. Bezahaf et al.

hop, the number of distinct route changes depending on the source router. For example, for the six hops distance pairs, we observe some pairs with 90 ∼ 100 distinct routes where other pairs with 240 ∼ 270 distinct routes. Another interesting result is the one hop distance pairs. In fact, we expected to obtain a linear line with only one distinct route between each pair, but we observe that some pairs have used more than 30 routes. This result means that pairs use logical multi-hops even if they are only one physical hop away. In fact, if the medium at one physical hop (direct link) experiences bad link quality, routing protocol selects another possible route with more logical hops. 5.2

Persistence

Persistence of dominant route (minute)

In Section 3.1, we studied the persistence of dominant route in the whole network from a macroscopic point of view. We investigate now the persistence of the routes with regard to the number of hops separating the source-destination pair. It is intuitive that shorter routes are more persistent. Fig. 13 shows the persistence of the dominant route for each pair of nodes at different distances. We observe that persistence of one-hop distance pair in the worst case is more than 4 minutes and can grow to more than 100 minutes. The persistence of four, five, and six hop distance pairs is practically the same and less than one minute. Dominant route persists more than one minute in 50% of cases for the three hops distance pairs to reach the limit of 10 minutes. At two hops distance pairs, dominant route are more persistent with a maximum of 42 minutes.

180 5

160 140

4

120

3

100

2

80

1

60

0 0

40

5

10

15

20

25

20 0 0

5

10

15

20

25

Couple of router id 1-hop 2-hops

3-hops 4-hops

5-hops 6-hops

Fig. 13. Persistence of dominant route between each pair of source-destination at different distances

Insights into Routing Stability

5.3

95

Sub-dominant Routes

Finally, we show the relationship between the number of sub-dominant routes and the number of hops between nodes. As shown in the previous section, we classify source-destination pairs according to their hop distance. Fig. 14 shows the percentage of dominance of the dominant and first four sub-dominant routes. We observe that the larger the number of hops, the smaller the prevalence of the dominant route. In Section 4, we investigated the dominance of the main sub-dominant routes. We saw that some source-destination pairs use only the dominant route, while other pairs also select a few sub-dominant routes. More specifically, what we can observe from Fig. 14 is that nodes separated by one and two hops mainly use

Percentage of dominant routes’ activity time

1-hop

2-hops

100

100

80

80

60

60

40

40

20

20

0

0 0

5

10

15

20

25

0

5

3-hops

10

15

20

25

15

20

25

15

20

25

4-hops

100

100

80

80

60

60

40

40

20

20

0

0 0

5

10

15

20

25

0

5

5-hops

10

6-hops

100

100

80

80

60

60

40

40

20

20

0

0 0

5

10

15

20

25

0

5

10

Couple of router id 2nd s-dominant Dominant 1st sub-dominant

3rd sub-dominant 4th sub-dominant

Fig. 14. Percentage of dominance of the five most used routes between pairs of routers. Pairs are classified according to their hop distance.

96

M. Bezahaf et al.

only the dominant route, whereas nodes separated by three and four hops also use two sub-dominant routes. Exceeding four hops, the dominant route starts loosing its dominance and pairs tend to use different routes without notion of dominance.

6

Conclusion

In this paper, we investigated the stability of routes in a real wireless mesh network deployment. For our study, we used DSDV as routing protocol with hop count metric. Our findings can be summarized in three main observations. Firstly, some pairs of routers use their dominant route with a persistence of almost 100% of the experimentation time, while other pairs show a high degree of instability. Secondly, some pairs alternate in almost perfect balance between their dominant route and their first two different sub-dominant routes. Thirdly, the larger the number of hops separating nodes, the larger the number of selected routes, losing the notion of dominant route. The above observations help in understanding and quantify the stability behavior of links in multi-hop wireless environment. Our results could be useful in new routing protocols design and implementation, taking into account the number of hops and the dominance of routes. As a next step, we plan to study how much the stability of the links can be altered when we change the routing protocol and/or the metric.

References 1. Madwifi project – multiband atheros driver for wireless fidelity 2. Agarwal, S., Ahuja, A., Singh, J.P., Shorey, R.: Route-lifetime assessment based routing (rabr) protocol for mobile ad-hoc networks. In: IEEE International Conference on Communications, New Orleans, LA, USA (June 2000) 3. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Computer Networks 47(4), 445–487 (2005) 4. Dube, R., Rais, C.D., Yeh Wang, K., Tripathi, S.K.: Signal stability based adaptive routing (ssa) for ad-hoc mobile networks. IEEE Personal Communications 4(1), 36–45 (1997) 5. Iannone, L., Fdida, S.: Meshdv: A distance vector mobility-tolerant routing protocol for wireless mesh networks. In: IEEE ICPS Workshop on Multi-hop Ad hoc Networks: From Theory to Reality, Santorini, Greece (July 2005) 6. Jacquet, P., Mhlethaler, P., Clausen, T., Laouiti, A., Qayyum, A., Viennot, L.: Optimized link state routing protocol for ad hoc networks. In: IEEE International Multitopic Conference, Lahore, Pakistan (December 2001) 7. Johnson, D.B., Maltz, D.A.: Dynamic source routing in ad hoc wireless networks. In: Imielinski, T., Korth, H. (eds.) Mobile Computing, vol. 353, pp. 153–181 (1996) 8. Lim, G., Shin, K., Lee, S., Yoon, H., Ma, J.S.: Link stability and route lifetime in ad-hoc wireless networks. In: International Conference on Parallel Processing Workshops, Vancouver, BC, Canada (August 2002)

Insights into Routing Stability

97

9. Paul, K., Bandyopadhyay, S., Mukherjee, A., Saha, D.: Communication aware mobile hosts in ad-hoc wireless network. In: IEEE International Conference on Personal Wireless Communications, Jaipur, India (February 1999) 10. Perkins, C., Bhagwat, P.: Highly dynamic destination-sequenced distance-vector routing (dsdv) for mobile computers. In: ACM Sigcomm, London, UK (September 1994) 11. Perkins, C.E., Royer, E.M.: Ad-hoc on-demand distance vector routing. In: IEEE Wrokshop Mobile Computing systems and Applications, New Orleans, LA, USA (February 1999) 12. Ramachandran, K., Sheriff, I., Belding, E., Almeroth, K.: Routing stability in static wireless mesh networks. In: Passive and Active Network Measurement, Louvainla-neuve, Belgium (April 2007) 13. Toh, C.-K.: Associativity-based routing for ad-hoc mobile networks. Wireless Personal Communications 4(2), 103–139 (1997)

A Study of Adaptive Gossip Routing in Wireless Mesh Networks Bastian Blywis, Mesut Güneş, Sebastian Hofmann, and Felix Juraschek Institute of Computer Science - Distributed Embedded Systems Freie Universität Berlin, Germany {blywis,guenes,shof,jurasch}@inf.fu-berlin.de

Abstract. Gossip routing is an approach to limit the overhead of flooding in wireless networks. Each node, that receives a packet that would normally be flooded, applies a probabilistic approach. The packet is either forwarded with probability p or dropped with 1 − p. This paper is a follow-up to our last study that evaluated the approaches by Haas et al. in the DES-Testbed, a wireless mesh network. Four different gossip routing variants were discussed which used static parameters to determine the value p. In this study, we compare the proposals by four other entities and discuss whether they show an overall improvement regarding the two most important metrics in this domain: reachability and redundancy. We also discuss the assumptions and parameters of the simulation studies in the context of our experiments in a real world deployment. Keywords: Wireless Mesh Network (WMN), Gossip Routing, Probabilistic Flooding, Testbed.

1

Motivation

Gossip routing, also called gossiping or probabilistic flooding, is an approach to improve the deficiencies of flooding. Flooding is an important service in wireless networks that is used when a source node has to send data to all nodes in the network. This is usually achieved by collaborative broadcasting. Each node, a router, that receives a packet forwards it by re-broadcasting the packet. In this way the packet traverses the whole network. Flooding in wireless networks is additionally “aided” by the physical layer broadcast medium and requires routers to send only one copy instead of one packet per dedicated link in wired networks. Routing protocols use flooding to spread topology information or route requests and replies in the network. Link state routing protocols, that are a representative of the class of proactive protocols, flood neighborhood information. For example, the Open Shortest Path First (OSPF) protocol [1] uses flooding based on multicast addressing. Reactive protocols execute an ad-hoc route discovery process when a packet has to be sent to some destination and no routing information is available. In many reactive protocols, a route discovery message is sent to all neighbors. If they cannot provide the required routing information, the packet is broadcasted to their neighbors and thus often flooded over the J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 98–113, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

A Study of Adaptive Gossip Routing in Wireless Mesh Networks

99

whole network. If a node can eventually answer with a route reply, the corresponding packet is either sent as unicast or flooded again in the reverse direction, depending on the particular protocol. Although flooding is very simple to implement, it has some deficiencies. As all nodes broadcast received packets, routers in the same broadcast domain may receive the same packet multiple times. Redundant data is sent over the medium. Especially in wireless networks, broadcast transmission may only be possible with low data rates, e.g., the BSSBasicRateSet data rates defined in IEEE 802.11 [2]. These packets can occupy significant time on the medium, which curbs high-speed unicast data transmissions. Further on, they also increase the contention for the medium access and the overall noise level. Due to these issues and the generally unreliable wireless medium, several packets might be lost. When we consider that a route request is lost, the result could be a sub-optimal or failed route discovery. The standard acknowledgment-based loss detection cannot be applied in flooding, as the neighbor set is either not known or the sender would experience the ACK implosion problem [3] leading to increased contention and in some cases could overwhelm the node. If the links are unidirectional, acknowledgments might not even be possible. It cannot be guaranteed that all nodes in the network receive the packets sent by a source. The optimization of the flooding service is therefore an important task for the optimal operation of wireless networks. To minimize the redundancy while still retaining full reachability, gossip routing can be used in place of flooding. In contrast to flooding, where packets are usually forwarded as long as the time to live (TTL) value in the packet header does not reach zero, the approach is probabilistic. Each node forwards packets with a given probability p or drops them with probability 1 − p in the most simple variant. Gossip routing is related to epidemic routing [4] where messages are stored on (mobile) nodes and are probabilistically forwarded to neighbors applying a “store-carry-forward” paradigm which models the spreading of diseases in a population. In our previous study [5] we discussed and evaluated three gossip routing variants by Haas et al. [6,7] and one of ours. The experiments were run in the Distributed Embedded Systems - Testbed (DES-Testbed) at Freie Universität Berlin, a wireless mesh network (WMN) [8]. Our results showed that gossip routing can reduce the redundancy in real world networks by up to 26% compared to flooding while at the same time a larger number of nodes is reached from the source node. Further on, we showed that the position of the source node has a crucial influence on the overall performance and which of the variants performs better. In contrast to the simulation-based study by Haas et al., the number of nodes that received packets from the source node was limited. Although each node received some packets from the source node during the whole experiment, there was no case when all nodes received the same packet together, even when we used p = 1.0 to achieve flooding. The properties of our real world networks significantly deviate from the models that are applied in simulations. This fact

100

B. Blywis et al.

Algorithm 1. PCBR - gossip7 Require: received packet with id if id not in listrcvd then cid = 1 insert id in listrcvd set RAD ∈ (0, Tmax ] and wait for RAD to expire if cid ≤ m then forward packet with probability p else remove id from received list end if else if cid ≤ m then cid = cid + 1 else remove id from listrcvd , stop RAD end if end if

underlines the necessity of further sophisticated experimentally driven studies. There is currently limited experience with gossip routing in real world networks. In this study, we use the same experiment setup to evaluate four other gossip routing variants. Major differences are that the forwarding probability of packets is not configured statically, network-wide and that the algorithms are more adaptive to the network topology. We evaluate the measured data considering two of the most important metrics in this domain. Gossip routing shall ensure that all nodes in the network receive the packets sent by a source. We deem the reachability as prime objective as all other optimizations are basically void if we reach only a limited subset of nodes. The redundancy has to be reduced so that the medium can be used for data flows with higher data rates. The remainder of this paper is structured as follows. In Section 2 the related work is discussed and the used gossip routing variants are introduced. The experiment setup and the implementation based on our routing framework are described in Section 3. All measurements are evaluated and discussed in Section 4. The paper ends with a conclusion in Section 5.

2

Related Work

The following gossip routing variants are part of this study. The numbering of the algorithms is continued from our previous publication were we studied gossip0 to gossip5. As gossip3 showed good performance and because it is used as basis for two other variants, it has also been included in this study for comparison.

A Study of Adaptive Gossip Routing in Wireless Mesh Networks

101

Algorithm 2. P-AODV - gossip8 Require: received packet for the first time Require: number of neighbors n > 0 pmax = 0.9 pmin = 0.4 pn−1 −pn max p = max(pmax ( max ), pmin ) 1−pmax if random(0, 1) < p then forward packet end if

Algorithm 3 . DPR - gossip9 ; n and nc represent the number of elements in the particular sets Require: received packet for the first time if n ≤ nf then p=1 else n−nc p = 1 − e− n end if if random(0, 1) < p then forward packet end if

2.1

gossip3

Haas et al. propose multiple gossip routing variants that support an initial flooding for k hops [6,7]. This approach shall ensure that the gossiping does not terminate early on when there are few neighbors to forward the packets. The authors argue that this is especially important to reduce border effects when the source node is near the edge of the network. The gossip3 variant additionally considers random networks with varying node degree. When a node would normally drop a packet due to an unlucky random draw, it is stored instead. If fewer than m duplicates are received in a specific time, the packet is broadcasted nevertheless; otherwise it is finally discarded. The value of the parameter m is configured statically and is the same for all nodes in the network. Experiments in a static, random network showed that a threshold of about p = 0.65 is sufficient for most nodes in the network to receive the packets from a source node. Another experiment scenario with a mobile network applied gossip3 for the route discovery phase of AODV. 2.2

gossip6

Like Haas et al., Shi and Shen try to improve the route discovery phase of AODV. Their approach is called Adaptive Gossip-based Ad Hoc Routing (AGAR) [9] and is based on gossip3. When a packet is received and the random number is

102

B. Blywis et al.

above the threshold p, the packet is stored. Instead of always broadcasting the packet when fewer than m duplicates have been received in the waiting time, the following formula is used to determine whether the packet shall be sent: random(0, 1) ≤ p/(d + 1), where d is the number of received duplicates. The authors argue that this adaptive modification reduces flooding in networks with a high node degree. The second chance to forward packets is smoothly decreased leading to a graceful degradation. 2.3

gossip7

Mohammed et al. propose a Probabilistic Counter-based Route discovery (PCBR) for AODV [10]. The general idea is that packets are always stored when they are received and the number of duplicates is counted. Packets are forwarded after the random assessment delay (RAD) timer has expired and when not enough duplicates (d ≤ m) were received. The RAD timer value is randomly chosen from the interval (0, Tmax ], see Algorithm 1. The basic approach is the same as in gossip3, but packets are always stored and the initial k hop flooding is missing. Tmax = 10 ms was used in experiments by Mohammed et al. to achieve a low delay. 2.4

gossip8

Hanashi et al. propose P-AODV [11]. When a broadcast packet (AODV route request) is received for the first time by a node, the probability p to forward the packet is calculated based on the number of neighbors n. p has a high value when there are few neighbors and has a low value when the node degree is high. The probability p is additionally bounded by pmin and pmax . A simplified but complete representation of the original algorithm is shown in Algorithm 2. The approach is similar to gossip2 by Haas et al. where a different but static probability p2 is used when the node degree is low. 2.5

gossip9

Jamal-Deen Abdulai proposes the Dynamic Probabilistic Route (DPR) algorithm [12] for AODV. When the node degree n is lower than or equal to nf , the algorithm turns to flooding. nf is the average node degree. When the node degree n is higher than nf , the additional coverage that can be achieved by forwarding the packet determines the value of p. Each node that forwards a packet attaches a list of its neighbors nc . Upon reception of a packet, the node determines the set of his neighbors that (probably) have not yet received this packet. If the set is empty, the node will not forward the packet; otherwise the larger the set, the higher the forwarding probability. The approach is simple since it requires only 1-hop neighborhood information to apply the self-pruning. The pseudo code is shown in Algorithm 3.

A Study of Adaptive Gossip Routing in Wireless Mesh Networks




103



Fig. 1. Overview of the locations of the nodes including the link quality as determined by a daemon that implements a HELLO protocol using the ETX metric. Thicker lines represent a better ETX value and a higher quality link. The figure shows links on channel 13. Source: [13].

2.6

Focus of This Study

Significant focus in the domain of gossip routing has been on the route discovery phase of AODV. Our experiments as specified in Section 3 focus on the pure performance of gossip routing. In the following we will always refer to the algorithms by their gossipX names, to emphasize that we are focusing only on the gossip routing and not the application in the AODV protocol. The goal is to research probabilistic flooding as a service for any routing protocol.

3

Experiment Setup

All experiments have been run in the DES-Testbed, a hybrid multi-transceiver wireless mesh and sensor network testbed [8]. 59 wireless mesh routers deployed in two adjacent buildings were used. The larger building had 42 and the smaller building 17 nodes which are distributed over multiple levels. Both buildings are close to each other and there are normally multiple links in-between. Figure 1 shows the topology measured with a daemon [13] that implements the ETX metric [14]. Despite the shown snapshot, the testbed is not partitioned but some links have a low quality. From our last experiment batch we know that all nodes in the network received some packets. Experiments were only run at night and on weekends when there are no or very few people or sources of interference in the vicinity. The average node degree of the network is about 7 with a maximum of 23. Each router is equipped with three IEEE 802.11a/b/g network cards. We used one card configured to channel 13 for the experiments as it not used by the campus-wide WLAN. All variants of the gossip routing protocol were implemented based on the Distributed Embedded Systems - Simple and Extensible Routing-Framework for

104

B. Blywis et al.

Testbeds (DES-SERT) [15]. We use underlay routing (layer 2.5 routing, like MPLS [16]) of Ethernet frames in user space to send and receive packets. For upper layer protocols and the operating system kernel the gossip routing is fully transparent. We used ARP-Requests as packets sent with the arping command. The destination was set to an IP address that is not assigned for any mesh router in the network to achieve a one-way wave of packets. The packets are quite small even considering the overhead introduced by the underlay routing: 82 Bytes (ARP-Request, Ethernet, DES-SERT, and IEEE 802.11 MAC Frame). The source node sends 100 packets with a rate of 1/s for each value of the threshold p ∈ (0..1]. If the particular variant has no configurable parameter p as the probability is adaptive, the source sends 10,000 packets so that all experiments have the same duration. Each packet has a unique 16 bit sequence number that is used for loop detection. The packet rate is low enough to ensure that there is none or only limited media contention among packets with different sequence numbers while at the same time the experiment duration is kept within acceptable dimensions. We selected node A and node B as source nodes. They are marked in Figure 1. Node A is a border node, while B is close to the center of the network. A has a node degree of 5 and B has 13 neighbors. The parameter k, which is used by gossip3 and gossip6, is set to 1. Higher values would lead to a flooding in significant parts of the testbed distorting the gossip routing as the network diameter is on average around ten hops. Considering the average node degree in the network, we deem an initial one-hop flooding high enough to ensure that the gossip routing does not terminate because only a limited number of neighbors around the source will (potentially) forward the packets. The value of the timeout parameter of gossip3 (and gossip6 ) is not explicitly specified by the authors. We set it to 200 ms which is high enough for packets to arrive over alternative and potentially longer paths. At the same time the delay will only introduce limited contention among subsequent packets sent by the source. The value of Tmax for the RAD timer of gossip7 has been specified by the authors with 10 ms. In our first trials we noticed that this value was much too low in real world scenarios. This resulted in many nodes sending in a very short time frame and thus high redundancy. The parameter has to be tuned very carefully considering these circumstances to achieve low redundancy and low delay. As delay is not in the focus of this study, we used 100 ms which is an adequate trade-off in this scenario and should give an upper bound of the performance of gossip7. The nf parameter of gossip9 was set to 14 by Abdulai. This value is based on the average node degree in a particular network given the parameters area of the network A, number of nodes N , transmission range R and using the following 2 formula: nf = (N − 1) πR A . In our case, the transmission range varies based on the locations of the nodes and the surroundings and the radio propagation is definitely not circular. From our experience we have an average (indoor) radio

A Study of Adaptive Gossip Routing in Wireless Mesh Networks

105

Table 1. Parameters of the experiments in the DES-Testbed. Some parameters are only valid for particular variants. Common parameters are in the upper section, individual ones are in the lower section. Parameter Number of Nodes Topology MAC Layer Channel Source Nodes Threshold p Packets per value of p Packets Duplicate limit m Timeout Flooding for k hops Tmax pmin , pmax nf

Value(-s)

Gossip Variant(-s)

59 random, static nodes IEEE 802.11 13 A and B

all

p ∈ (0..1], in 0.01 steps 100, one per second 10,000, one per second 1 200 ms 1 100 ms 0.4, 0.9 4

3, 3, 8, 3, 3, 3, 7 8 9

6, 7 6, 7 9 6, 7 6 6

range of about 11 m because of many thick walls which block the radio propagation. This does also consider in that we ignore long(-er) links with a low packet-delivery-ratio. Additionally it had to be compensated that the vertical radio propagation in between floors is worse than on the same level. As the routers of the testbed are not deployed in a plane, we calculate nf as follows: 4

πR3

nf = (N − 1) 3A×h . With height h = 13 m and A ≈ 5600m2 we get a result of about nf = 4, which is of course only a rough estimate. The gossip9 algorithm will fall back to flooding for all nodes with node degree ≤ 4. Table 1 summarizes the experiment parameters. A full experiment run takes more than 28 hours including configuration overhead. Further experiments with different parameter settings will be considered as future work.

4

Evaluation

Please note that all graphs of this section and our first study are available on our website in higher resolutions1. 4.1

Figure Format Description and gossip3

For the evaluation, we consider in how many cases (executions2 ) when the source node sent a packet, which fraction of nodes did receive this packet. There are two 1 2

Graphs: http://www.des-testbed.net/experiments/gossip The nomenclature of executions has been adopted from [6,7].

106

B. Blywis et al.

(a) Source Node A

(b) Source Node B

1.0

0.9

0.9

0.8

0.8

0.7

0.7

0.5

(a) Source Node A

1.00

0.95

0.85

0.90

0.75

0.80

0.65

0.70

0.55

0.60

0.45

0.50

0.35

0.40

0.25

0.0

0.30

1.00

0.95

0.85

0.90

0.75

Probability p

0.80

0.65

0.70

0.55

0.60

0.45

0.50

0.35

0.40

0.0

0.25

0.1

0.30

0.2

0.15

0.2

0.1

0.20

0.3

0.05

0.3

0.15

0.4

0.20

0.5

0.4

0.6

0.05

0.6

0.10

Fraction of Nodes

1.0

0.10

Fraction of Nodes

Fig. 2. Fraction of executions as a function of the fraction of nodes receiving the packets with gossip3 for each value of p shown as histogram with bin size 0.1

Probability p

(b) Source Node B

Fig. 3. Fraction of nodes receiving the packets with gossip3 as box-and-whisker-plot with median and 95% confidence interval

different representations of the data which will be explained based on the gossip3 results. Figure 2 shows histograms for both source nodes with the parameter p on one of the axes. In Figure 3 the median fraction of nodes that received the packets is shown as a function of the parameter p. The plots show the median in a box-and-whisker plot with a 95% confidence interval; the whiskers extend up to 1.5 times of the inter-quartile range. All data points outside of the whiskers

A Study of Adaptive Gossip Routing in Wireless Mesh Networks

107

are considered statistical outliers. For gossip3 we can see that the algorithm is independent of the configured threshold p and achieves a median of around 0.8. Further on as shown in Figure 2a, in more than 80% of the executions we reach a fraction of ≥ 0.75 of nodes. This means the algorithm in its current configuration performs very well especially considering that a router will not forward stored packets when one or more duplicates are received. Higher values of the parameter m might lead to an even larger median fraction of nodes. We consider a median of 0.8 to be the best that can be achieved in our testbed. From all experiments we learned that there are 11 nodes in the smaller building that are badly connected to the rest: this exactly matches the 0.8 fraction. Comparing gossip3 from source node A and B we notice that for the latter there are more outliers which accumulate at 0.1 on the fraction of nodes axis (see Figure 3b). In these cases the 11 “problematic” nodes did not receive any packets from the source and no packet got to the larger building. Even the gossip routing variants discussed in the next section cannot resolve this problem but we conclude from all figures that this problem is independent from the value of p. 4.2

Advanced Gossip Routing

gossip6 from source A shows results (Figures 4a, 6a) similar to what we measured for gossip1 in the last study3 . The variant shows a strong dependence on the parameter p: the higher the value, the more nodes receive the packet from the source. Compared to gossip1, the whiskers are larger in gossip6 especially for higher values p. This is probably an effect of the second broadcast chance that is missing in gossip1. From source B the median fraction of nodes makes a jump from below 0.2 to above 0.75 at p = 0.75. This is a bimodal behavior that we already observed in the first study. When comparing gossip3 and gossip6 we notice that, although the former is based on the latter, the results are totally different. gossip7 shows results that are very similar to gossip6 and considering the confidence intervals, the results are mostly equal. In Figure 6b we notice that for p ≥ 0.80 the boxes get very small as well as the confidence intervals. The histogram for source A (Figure 4b) shows a much higher value for the 0.80 bin which underlines that the true median actually lies at 0.80. Overall the delayed forwarding combined with duplicate counting could only reduce the variance but not increase the reachability. Source node B shows similar differences in the histogram but no variance reduction in Figure 7b. When the source is in the smaller building, neither gossip3, gossip6, nor gossip7 were able to ensure that the gossip routing reaches the other building. Maybe the adaptive variants will perform better. As the last two gossip routing variants use no static value for p, their graphs show only a single histogram or box-and-whisker plot. The results are quite surprising. gossip8 achieves only a median fraction of nodes of below 0.6 from source A and slightly above 0.1 from source B. Despite the large whiskers in Figure 6c and Figure 7c, it has to be 3

gossip1 uses only an initial k-hop flooding and forwarding based on p.

108

B. Blywis et al.

(b) gossip7

1.0

1.0

0.9

0.9

0.8

0.8

0.7

Fraction of Executions

Fraction of Executions

(a) gossip6

0.6

0.5

0.4

0.7

0.6

0.5

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0.0 0.0

0.0 0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

0.1

0.2

0.3

0.4

0.5

0.6

Fraction of Nodes

Fraction of Nodes

(c) gossip8

(d) gossip9

0.7

0.8

0.9

1.0

Fig. 4. Source node A

assumed that the true median is in the 95% confidence interval. The confidence intervals are small as 10.000 data points are the basis of these graphs. Figure 4c shows that in 40% of the execution less than 50% received the packet from source A. gossip8 has only one chance to forward the packets with the node degree n as the only factor that influences the probability. We learned that several nodes did not forward the packets as their node degree was too high. This is fatal for the gossiping as these nodes are required to connect the whole network - especially for source node B. The median fraction of nodes in Figure 7c with a value of about 0.1 corresponds to the well connected routers in the smaller building. Thus gossip8 from source B did in many cases reach neither the 11 badly connected routers in the same building nor did the gossiping reach any router in the larger building. At least in our random network topology, the approach of gossip8 has problems. gossip9 seems to have the same problems as gossip8. The median fraction of nodes is actually worse for both sources (Figures 6d and 7d) and the histograms show a shift to the left (Figures 4d and 5d). The approach

A Study of Adaptive Gossip Routing in Wireless Mesh Networks

(b) gossip7

1.0

1.0

0.9

0.9

0.8

0.8

0.7

Fraction of Executions

Fraction of Executions

(a) gossip6

0.6

0.5

0.4

0.7

0.6

0.5

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0.0 0.0

0.0 0.0

0.1

0.2

0.3

0.4

0.5

0.6

109

0.7

0.8

0.9

1.0

0.1

0.2

0.3

0.4

0.5

0.6

Fraction of Nodes

Fraction of Nodes

(c) gossip8

(d) gossip9

0.7

0.8

0.9

1.0

Fig. 5. Source node B

to consider the additional coverage that can be achieved by forwarding packets fails in our testbed as the distribution of nodes deviates from the assumptions made in the simulation study. 4.3

Redundancy

To evaluate the redundancy we plot the ratio of how many packets are sent when gossip routing is used to how many packets are sent with flooding: pktsgossipX /pktsf looding Figure 8a shows the metric and its correlation with the reachability for source A and Figure 8b the same for source B. The intersection of the gray lines marks the data point for flooding which we measured in a separate experiment run. Therefore points in the top-left sector represent improvements. Variants with different probabilities p are shown as multiple, selected data points. gossip3 and gossip7 actually reach 80% of the nodes but with up to 30% fewer packets than flooding from source A. gossip7 should be

110

B. Blywis et al.

1.0

1.0

0.9

0.9

0.8

0.8

Probability p

1.00

0.95

0.85

0.90

0.75

0.80

0.65

0.70

0.55

0.60

0.45

0.50

Probability p

(a) gossip6

(b) gossip7

1.0

1.0

0.9

0.9

0.8

0.8

0.7 Fraction of Nodes

0.7 Fraction of Nodes

0.35

1.00

0.95

0.85

0.90

0.75

0.80

0.65

0.70

0.55

0.60

0.45

0.50

0.35

0.40

0.25

0.30

0.15

0.20

0.1

0.0

0.05

0.1

0.0

0.40

0.2

0.25

0.2

0.30

0.3

0.15

0.4

0.3

0.20

0.4

0.5

0.05

0.5

0.6

0.10

Fraction of Nodes

0.7

0.10

Fraction of Nodes

0.7

0.6

0.6

0.5

0.4

0.6

0.5

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0.0

0.0

adap Probability p

adap Probability p

(c) gossip8

(d) gossip9 Fig. 6. Source node A

configured with 0.70 ≤ p ≤ 0.90 and gossip3 with p ≤ 0.90, m = 1. For source B the savings are less (20%) but gossip6 also shows better performance. gossip6 and gossip7 with p = 0.9 as well as gossip3 with and p ≤ 0.90, m = 1 show the best results. gossip8 and gossip9, for which we also plotted the results of an additional experiment run, are in the bottom-left sector. Duplicate-counting approaches seem to perform very well with some restrictions. gossip6 decreases the number of forwarded packets in the source A scenario in comparison to gossip3 on which it is based on, yet at the same time the reachability suffers. The influence of the parameter p is clearly visible in Figure 8a. gossip6 is too conservative to determine the probability to forward stored packets. gossip7, which always stores packets and counts the duplicates, performs better and underlines that a static duplicate threshold m is good in our testbed scenario. In future experiments we will try to optimize the parameter Tmax to achieve a lower delay while keeping the increased reachability.

A Study of Adaptive Gossip Routing in Wireless Mesh Networks 1.0

1.0

0.9

0.9

0.8

0.8

Probability p

1.00

0.95

0.85

0.90

0.75

0.80

0.65

0.70

0.55

0.60

0.45

0.50

Probability p

(a) gossip6

(b) gossip7

1.0

1.0

0.9

0.9

0.8

0.8

0.7 Fraction of Nodes

0.7 Fraction of Nodes

0.35

1.00

0.95

0.85

0.90

0.75

0.80

0.65

0.70

0.55

0.60

0.45

0.50

0.35

0.40

0.25

0.30

0.15

0.20

0.1

0.0

0.05

0.1

0.0

0.40

0.2

0.25

0.2

0.30

0.3

0.15

0.4

0.3

0.20

0.4

0.5

0.05

0.5

0.6

0.10

Fraction of Nodes

0.7

0.6

0.10

Fraction of Nodes

0.7

111

0.6

0.5

0.4

0.6

0.5

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0.0

0.0

adap Probability p

adap Probability p

(c) gossip8

(d) gossip9 Fig. 7. Source node B

5

Conclusion

In this paper we evaluated four different gossip routing variants in the DESTestbed. Despite the optimal experiment conditions (small packets, stationary nodes, single source, no other data flows, high timeout values) none of the discussed gossip variants could reach slightly more than 80% of the nodes in median. The number of packets could be reduced by up to 30% compared to flooding when the parameters are chosen adequately and the position of the source node is considered. Variants with adaptive approaches performed surprisingly bad as they seem to have problems in our topology with highly varying node degree. Duplicate-counting with a low threshold value (m = 1) performs very good and is totally overhead free as no management information is required, e.g., by using a HELLO protocol. The most important issue, identifying the nodes that are vital for the connectivity of the network, has yet to be considered to make gossip routing applicable in real world deployments. Gossip routing in wireless

112

1.1 1.0

0.8

0.8

median fraction of nodes

0.9

median fraction of nodes

0.9 0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

(a) Source Node A

1.1

1.0

ratio to flooding

0.9

0.7

0.8

0.5

0.6

0.0

0.0

1.1

1.0

ratio to flooding

0.9

0.7

0.8

0.5

0.6

0.4

0.2

0.0

0.3

0.1 0.1

0.2

0.1

0.3

0.3

0.2

0.4

0.3

0.0

gossip 8 gossip 9 gossip 3 gossip 6 gossip 7

0.2

1.0

gossip 8 gossip 9 gossip 3 gossip 6 gossip 7

0.1

1.1

B. Blywis et al.

(b) Source Node B

Fig. 8. Median fraction of nodes as a function of the number of forwarded packets in relation to flooding

networks requires further research to come up with an optimal solution that ensures high reachability and low redundancy independent of the source node position in the network.

References 1. Moy, J.: OSPF Version 2, RFC 2328 (Standard) (April 1998), http://www.ietf.org/rfc/rfc2328.txt 2. IEEE Std 802.11-2007. IEEE standard for information technology – Telecommunications and information exchange between systems – Local and metropolitan area networks-specific requirements – Part 11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications, LAN/MAN Standards Committee, New York, NY, USA, pp. C1-1184 (June 2007), http://dx.doi.org/10.1109/IEEESTD.2007.373646 3. Ma, D.: Secure feedback service in wireless sensor networks. In: Dawson, E., Wong, D.S. (eds.) ISPEC 2007. LNCS, vol. 4464, pp. 116–128. Springer, Heidelberg (2007) 4. Vahdat, A., Becker, D.: Epidemic routing for partially-connected ad hoc networks, Duke University, Tech. Rep. CS-2000-06 (July 2000) 5. Blywis, B., Günes, M., Juraschek, F., Hofmann, S.: Gossip routing in wireless mesh networks. In: International Symposium on Personal, Indoor and Mobile Radio Communications, PIMRC 2010 (2010) 6. Haas, Z.J., Halpern, J.Y., Li, L.: Gossip-based ad hoc routing. In: Proceedings IEEE INFOCOM 2002, The 21st Annual Joint Conference of the IEEE Computer and Communications Societies, June 23-27. IEEE, New York (2002), http://www.ieee-infocom.org/2002/papers/822.pdf 7. Haas, Z.J., Halpern, J.Y., Li, L.: Gossip-based ad hoc routing. IEEE/ACM Trans. Netw. 14(3), 479–491 (2006)

A Study of Adaptive Gossip Routing in Wireless Mesh Networks

113

8. Blywis, B., Günes, M., Juraschek, F., Schiller, J.: Trends, advances, and challenges in testbed-based wireless mesh network research. In: ACM/Springer Mobile Networks and Applications (MONET), special issue on Advances in Wireless Testbeds and Research Infrastructures (February 2010), http://www.springerlink.com/content/hk25051135m36571/ 9. Shi, Z., Shen, H.: Adaptive gossip-based routing algorithm. In: Proceedings, 23rd IEEE International Performance, Computing, and Communications Conference IPCCC (2004) 10. Mohammed, A., Ould-Khaoua, M., Mackenzie, L.M., Perkins, C., Abdulai, J.D.: Probabilistic counter-based route discovery for mobile ad hoc networks. In: IWCMC 2009: Proceedings of the 2009 International Conference on Wireless Communications and Mobile Computing, pp. 1335–1339. ACM, New York (2009) 11. Hanashi, A., Siddique, A., Awan, I., Woodward, M.: Performance evaluation of dynamic probabilistic broadcasting for flooding in mobile ad hoc networks. Simulation Modelling Practice and Theory 17(2), 364–375 (2009) 12. Abdulai, J.: Probabilistic Route Discovery for Wireless Mobile Ad Hoc Networks (MANETs) (2009) 13. Philipp, M.: A framework for distributed channel assignment in wireless mesh networks. Master Thesis, Freie Universität Berlin (March 2010) 14. De Couto, D.S.J., Aguayo, D., Bicket, J., Morris, R.: A high-throughput path metric for multi-hop wireless routing. In: MobiCom 2003: Proceedings of the 9th Annual International Conference on Mobile Computing and Networking, pp. 134– 146. ACM, New York (2003) 15. Blywis, B., Günes, M., Juraschek, F., Schmidt, P., Kumar, P.: Des-sert: A framework for structured routing protocol implementation. In: IFIP Wireless Days 2009 (2009) 16. Rosen, E., Rekhter, Y.: BGP/MPLS IP Virtual Private Networks (VPNs), RFC 4364 (Proposed Standard) (February 2006), updated by RFCs 4577, 4684, http://www.ietf.org/rfc/rfc4364.txt

A Multipath Routing Method with Dynamic ID for Reduction of Routing Load in Ad Hoc Networks Tomoya Okazaki, Eitaro Kohno, Tomoyuki Ohta, and Yoshiaki Kakuda Graduate School of Information Sciences, Hiroshima City University, 3-4-1 Ozuka-Higashi Asaminami-ku, Hiroshima, 731-3194, Japan {[email protected],kouno@,ohta@,kakuda@}hiroshima-cu.ac.jp

Abstract. In recent years, ad hoc networks have attracted a great deal of attention. Ad hoc networks consist of nodes with wireless communication devices without any base stations or fixed infrastructures. Most routing protocols of ad hoc networks form a single-path. Single-path routing protocols need to repair routes each time the route has broken. This route repair generates a lot of control packets, and an increase in end-to-end packet delay. In order to compensate for these drawbacks of the single path routing, multipath routing schemes have been proposed. AOMDV (Ad hoc On-demand Multipath Distance Vector routing) is one multipath routing scheme. AOMDV constructs routes by flooding Route Request (RREQ) messages. When the number of nodes increases in the network, the routing load of AOMDV, which is defined as the ratio of the number of control packets to the number of delivered packets, may increase immensely. On the other hand, DART (Dynamic Address RouTing) has been proposed for a large scale network. In DART, the dynamic routing address has a tree-based logical structure related to connectivity between adjacent nodes. In this paper, we propose a multipath routing scheme to solve the above problems of AOMDV. The proposed scheme is an extension of DART for dealing with multiple paths. The proposed scheme aims to reduce the routing load and adapt to large ad hoc networks. We evaluated its performance by comparing it with AODV and AOMDV through simulation experiments. Performance metrics are the number of control packets and the routing load. Simulation results indicate the proposed scheme can reduce the number of control packets and the routing load. Keywords: Ad hoc networks, routing load, Dynamic ID.

1

Introduction

Ad hoc networks are autonomous distributed networks without any base station, since they consist of nodes with wireless communication which relay packets. When the node cannot perform direct communication with other nodes, the intermediate nodes can relay packets of the source node to communicate with the destination node. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 114–129, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

A Multipath Routing Method with Dynamic ID

115

AODV (Ad hoc On-demand Distance Vector) [7][6] is a popular routing method for creating single paths. AODV discovers the shortest (single) path between the source and the destination nodes. With a single path routing method such as AODV, when route breaks occur in the network, it is necessary to repair the route in each case. These route repair procedures pose an escalation both in the number of control packets and in the delay to the destination. Multiple path routing [1] is one of the solutions for these problems. AOMDV (Ad hoc On-demand Multipath Distance Vector) [5] routing is one of the methods to create multiple paths. AOMDV is an extension of AODV. AOMDV discovers multiple paths by flooding RREQ (Route Request) messages. When the number of nodes in the network increases, the number of the control packets escalates. Therefore the increase in the number of control packets suppresses the effective bandwidth of the network. Meanwhile, DART (Dynamic Address RouTing) [2] has been proposed as the method which can apply to networks that have many nodes. DART constructs a single path using a dynamic routing address based on assigning addresses to node using a binary tree. In this paper, we propose a new multiple-path routing method to extend DART to solve the existing problems mentioned above. We have designed the proposed method to reduce the number of control packets in networks with many nodes. Additionally, we have implemented our method on the network simulator, QualNet [8], and have evaluated by comparison with AODV and AOMDV in terms of the number of control packets and the routing load. The rest of the paper includes the following: In Section 2, we mention the ad hoc networks and their existing problems. We explain our proposed method in Section 3. Section 4 illustrates the experiments and the results. Finally, we conclude our paper in Section 5.

2

Existing Routing Schemes and Their Problems

AODV and DSR (Dynamic Source Routing) [3] are popular single-path routing methods. They are categorized as reactive routing protocols, and the source node broadcasts RREQ (Route Request) messages to discover and establish a route to the destination node. While route information is partially stored on each node in AODV, it is stored only on the source node in DSR. Meanwhile, AOMDV and SMR (Split Multipath Routing) [4] protocol are proposed to create multiple paths. AOMDV and SMR are extensions of AODV and DSR, respectively. Because these protocols use broadcasting and flooding of control packets, the number of control packets can escalate. As a result, the routing load, that is the proportion of the number of control packets to the number of data packets, will increase. This cause can become serious in large scale ad hoc networks. On the other hand, DART has been proposed for large scale ad hoc networks. DART introduces the routing address as the indicator of node location, and is utilized for limiting the size of the routing table and the number of control packets. The routing address space has a binary tree structure, which is essential

116

T. Okazaki et al.

for packet forwarding. Each node stores a unique and static IP address and the dynamic routing address. DART has the serious problem that it is necessary that the start node (the root of the binary tree structure) must be established preliminarily.

3 3.1

Proposed Method Overview

AOMDV, the existing multipath routing scheme, has the control packet inflation drawback which was mentioned in the previous section. Also, DART has three problems: the need to establish the starting node of the routing address, the overlapping of the routing addresses, and the lack of robustness of the single path. To solve AOMDV and DART, we propose a new multipath routing method. Our method is an extension of DART, and we aim to limit the number of control packets and the routing load for large scale networks. Therefore we extend DART to the following: 1. We introduce Dynamic IDs which are assigned from the source node. 2. The routing table is extended to consist of multiple paths. Each node stores the following information: IP address The IP address is uniquely assigned at the start. The IP address will not change in our scheme. Dynamic ID Dynamic ID is expressed as the l bit binary numbers al−1 , al−2 , · · · , a0 , where ai (i = 0, · · · , l − 1) indicate 0 or 1. The binary tree, which is referred to as the ID tree, consists of Dynamic IDs. Figure 1 shows the ID tree with Dynamic IDs made up three bits each. Leaves of the ID tree indicate Dynamic IDs which are assigned to each node. The vertices, which are illustrated as dotted lines, represent the subtree (the set of nodes) below that node. Also, the vertical position of the subtree is referred to as the level. The network topology which corresponds to Figure 1 is depicted in Figure 2. The sets of nodes in the dotted rectangle areas correspond to the subtrees of Figure 1 . Routing table The routing table has routing entries. Each entry has the information of the subtrees and their sets of nodes, and consists of the following: – level – subtree – IP address to the next hop node – the hop count to subtree – the number of ID tree.

A Multipath Routing Method with Dynamic ID

Ｃ

ＸＸＸ

Level 2 Level 1

０ＸＸ ００Ｘ

１ＸＸ ０１Ｘ

１０Ｘ

Level 0

１１Ｘ

111

11Ｘ

Ｆ

110

Ｓ 101 Ａ 100 10Ｘ 1XX Node

０００ ００1 ０1０ ０11 1００ 1０1 11０ 111

Ｂ

000 0XＸ

117

Ｅ

010

Ｄ

011 01Ｘ

Set of nodes which belong to same sub tree

Fig. 1. Example of ID tree with three bit Fig. 2. Example of network topology for proposed method Dynamic ID level 2 2 1 1 1 0

sub tree 0XX 0XX 11X 11X 11X 100

ID tree number F F F F F F

next hop A B C B A A

hop count 2 2 1 1 2 1

Fig. 3. Example of the routing table for proposed method

level sub tree next hop 2 0XX A 1 11X B 0 100 A

hop count 2 1 1

Fig. 4. Example of the routing table for DART

Figures 3 and 4 show an example of a routing table using the proposed method and DART. While DART has only single next hops to each subtree, our proposed method has multiple next hops to each subtree. This extension is mandatory to make multiple paths. Number of ID trees In our scheme, the Dynamic ID assignment starts from the source node which has an unassigned Dynamic ID. Therefore, the network can have multiple ID trees. The ID tree numbers are used to distinguish each ID tree. The ID tree numbers consist of IP addresses of the source and the destination nodes. 3.2

Procedures

Procedures of our scheme consist of the following: 1. Routing i. Assignment of the Dynamic ID ii. Update of routing table 2. Packet Forwarding i. Node lookup ii. Selection of the Next Hop iii. Detour using an alternate path Our proposed method differs with DART in that it assigns a Dynamic ID. Additionally, routing and packet forwarding are extended from DART to create multiple paths. Other procedures follow the method established in DART.

118

T. Okazaki et al.

The following subsections describe each procedure in detail. Routing. The routing procedure consists of three sub-procedures: the Dynamic ID assignment, the maintenance of the border node table, and the routing table update. The Dynamic ID assignment sub-procedures are performed using the following three control messages: – – – –

RTUP (RouTe UPdate) IDREQ (dynamic ID REQuest) IDREP (dynamic ID REPly) IDN (dynamic ID Nortification).

RTUP carries the routing information using the routing table in the source node. The routing table has route entries which includes the next hop from the source node. RTUP is used for the Dynamic ID assignment and the update of the routing table sub-procedures. When a node assigns a Dynamic ID, the node broadcasts RTUP to its neighboring nodes. When the demand to communicate with the destination node occurs on the source node, the Dynamic ID assignment sub-procedure is invoked. At the same time, the source node which is not assigned a Dynamic ID becomes the start node of the Dynamic ID assignment sub-procedures and the start node of the ID subtree. The Dynamic ID assignment sub-procedure performs differently depending on the type of assigning node. That is, whether it is the start node or not. Therefore, we explain the Dynamic ID assignment sub-procedure in each case. The case in which the assigning node is the start node Suppose the start node assigns the Dynamic ID to neighboring unassigned nodes. At first, the start node assigns itself to its own Dynamic ID as [00 · · · 0]. Next, the start node broadcasts RTUP to its neighboring nodes. Unassigned nodes receive a RTUP broadcast, then unicast an IDREQ to the start node. When the start node receives IDREQ, the start node selects a Dynamic ID to assign for the IDREQ sender after waiting a certain period of time. At this point, the start node assigns the subtree which has the highest level value within the ID tree of the source node, too. If the level that is assigned to the node which generates IDREQ is m, the Dynamic ID is differs from the start node only at the mth bit. The start node unicasts the IDN message with a new Dynamic ID to the IDREQ generator node. Figure 5(a) shows an example network. In Figure 5(a), the start node S is going to assign Dynamic ID to nodes A, B, and C. Figures 5(b) and 5(c) show the routing table of node S and the sequence of the procedures, respectively. Suppose node S receives IDREQs in the sequence of C, B, and A. Node S assigns [100], [010], [001] to the Dynamic ID of A, B, and C, which are different in regard to 2nd, 1st, and 0 bit from the Dynamic ID of the node S, respectively. The case in which the assigning node is any nodes but the start node Suppose nodes except the start node assign the Dynamic ID to neighboring unassigned nodes. When the unassigned node receives RTUP from the assigned node, the assignment sub-procedure will be invoked. The unassigned

A Multipath Routing Method with Dynamic ID

100

S C 001

Ａ

110

D

010

B

119

S

next hop

2

1XX

none

1

01X

none

0

001

none

Start node

Ｃ

IDREＱ

sub tree

IDREQ IDN

(a) Dynamic ID assignment.

Ｓ

level

(b) The routing table of node S.

Ｂ

Ａ

RTUP

IDREＱ IDREＱ

IDREＰ

IDREＰ IDREＰ

time

(c) Sequence. Fig. 5. Example of Dynamic ID assignment by the start node

node sends IDREQ by unicast to the assigned node. The node, which receives IDREQ, searches subtrees which do not have the next hop registered, and selects the highest level subtree. If the level value is m, the node which receives IDREQ assigns a new Dynamic ID for the unassigned to node. To do this, the node which receives IDREQ sends IDN with the newly assigned Dynamic ID to the generator of IDREQ. In Figure 5(a), suppose node A is going to assign a Dynamic ID to node D. Figures 6(a) and 6(b) show the routing table of node A and the sequence of the procedures, respectively. Suppose node A has the Dynamic ID [100]. When node D receives RTUP, it sends IDREQ to node A. When node A receives IDREQ from node D, it searches the subtrees which did not have the next hop registered. According to Figure 6(a), the level 2 subtree is already registered, but the level 1 subtree does have not the next hop registered. Therefore, node A assigns the Dynamic ID [110] to node D. To do this, node A sends IDREP to node D. Update of routing table A Node updates its own routing table using the information included in RTUP from other nodes. Suppose node B is going to update its own routing table using RTUP from node A, where the Dynamic ID length of nodes A

120

T. Okazaki et al.

Ａ

Ｄ RTUP IDREＱ

level

sub tree

next hop

2

0XX

S

1

11X

none

0

101

none

(a) The routing table of node A.

time

IDREＰ (b) Sequence.

Fig. 6. Dymamic ID update by nodes except the start node

and B is l bits. Also, m is smaller than l. Node B updates the entry within the its own routing table as follows: Routing entries from level l − 1 to m Update the value of the next hop to node A. Update node B’s hop count to that of RUTP plus one. Routing entries level m − 1 Update the next hop to node A, and the value of hop count to one. Routing entries from level m − 2 to 0 Routing entries will not be updated. We explain the sub-procedure through an example in Figure 7. Suppose node B receives RTUP from node A, and is going to update its own routing table, where the Dynamic ID length is four bits, and the Dynamic ID of nodes A and B are [0110] and [0100]. Figure 7 shows the routing tables of nodes A and B. The upper table shows the routing table prior to being updated, and the lower table is after the update. The shadowed entries are updated entries. The common prefix length of the Dynamic ID between nodes A and B is two. Node B registers node A to the entries with level 3 and 2. The hop counts of the entries with level 3 or 2 (= 1 + 1) is set to 2 because of the hop count of levels 3 and 2 on node A. Node B registers node A to the entry with level 1. At this time, the hop count of this entry will be set to 1. The level 1 entry will not be updated. Because our proposed method establishes multiple paths, the node registers multiple next hops to one subtree. Packet Forwarding. The data packet forwarding sub-procedure is divided into three parts: the node lookup, next hop selection, establishing detours using alternate paths. Node Lookup Node lookup is the sub-procedure to seek the Dynamic ID that corresponds with the IP address of a node. In the node lookup sub-procedure, a look

A Multipath Routing Method with Dynamic ID

Node B ID[0100]

121

Node A ID[0110]

level

sub tree

next hop

hop count

level

sub tree

next hop

hop count

3 2 1 0

1XXX 00XX 011X 0101

S

1

1

1 1

1XXX 00XX 010X 0111

E

C D

3 2 1 0

F C

1 1

Update Node B ID[0100] level

sub tree

next hop

hop count

3 3 2 1 1 0

1XXX 1XXX 00XX 011X 011X 0101

S A A C A D

1 2 2 1 1 1

Fig. 7. Update of the routing table

up table is required. For the realization of the lookup table in our method, we proposed distributed lookup tables. A distributed lookup table consists of a pair made up of the IP address and the Dynamic ID. The pairs will be distributed to each node. A node which is assigned or re-assigned the Dynamic ID adds the pair to the distributed lookup tables. Next Hop Selection When a node forwards data packets, it decides the next hop according to the subtree of the destination node. When the length of the Dynamic ID is l and the common prefix length between the sender and destination node is m, the node can get the subtree level of the destination node by the calculation l − m − 1. The node transfers data packets to the next hop node which is indicated by the calculated result. Those sub-procedures will be repeated until the node reaches the destination nodes. Suppose node S sends data packets to the node D in Figure 8. Node S has the routing table shown in Figure 8(a). The Dynamic ID of nodes S and D are [101] and [011], respectively. In this case, the common prefix length between nodes S and D are 0. Node A calculates the level that belongs to node D as 3 − 0 − 1 = 2. Node S transmits the data packets to the next hop node A which are registered in the entry as level 2. Establishing Detours Using Alternate Paths. When a node detects failure of the packet transmission due to link breakage, it detours the data packets. When the node sends the data packets, it replicates the data to the buffers before sending the data packets. Multiple ID Trees and Countermeasures. In the proposed method, it is possible that the network has multiple ID trees. The multiple trees can be created when simultaneous requests are made to multiple unassigned nodes. When the

122

T. Okazaki et al.

level 2 1 1 0

sub tree 0XX 11X 11X 100

next hop A[100] B[110] C[111] A[100]

hop count 2 1 1 1

(a) The routing table of node S.

Ｃ

111

Ｂ

11Ｘ

Ｓ 101 10Ｘ

110

Ａ 100

Ｆ

000 0XＸ

Ｅ

010

Ｄ

011 01Ｘ

(b) Topology. Fig. 8. Example of data packet forwarding for proposed method

network has multiple ID trees, the data packets need to traverse the multiple ID trees. To do that, we introduced the border node table and the traversal transfer sub-procedure among multiple ID trees. Border node tables and their maintenance To maintain the border information, we employ the BDN (BorDer node Notification) message. Figure 9 shows an example in which the network has multiple trees. In Figure 9, “ 000” on the nodes A and B indicates the start nodes for each ID tree where each ID tree is distinguished by the IP address of its start node. When a node receives RTUP from the neighboring node, it compares the ID tree identifier in the RTUP to that in its own routing table. If two ID tree identifiers differ, the node which receives RTUP recognizes it set as the border node. The border node sends BDN by unicast to the source node which has the same ID tree identifier. At this time, nodes which relay the BDN and the border node sends BDN to the assigning node for a Dynamic ID. When the source node receives BDN, it registers the generator of BDN to its border node table. At the same time, the relay nodes of BDN register the address of the border node to their border node tables. Traversal transfer among multiple ID trees When the destination node locates another ID tree and it has no information about the border node, it can get the IP address of the border node using the following messages: – BNREQ(Border Node REQuest) – BNREP(Border Node REPly). The source node sends BNREQ to the start node of its own ID tree. The start node that receives BNREQ seeks its border node table. When the start node finds the IP address of the border node, it sends BNREP by unicast to the source node.

A Multipath Routing Method with Dynamic ID

ID tree A

A

123

ID tree B 100

100 110

000

B

000

110

010 011

Border Node

BDN

Fig. 9. Existance of multiple start nodes

4 4.1

Experiments and Discussion Implementation of the Proposed Method and Prelimimary Experiments

We have implemented our proposed method on the prototype system using the simulator, QualNet 4.5. In the prototype system, we use the global lookup table instead of the distributed lookup table. This is the same as in literature [2]. Using the prototype system, we confirmed the procedures for three main points: the assignment of Dynamic ID, the detour to the alternate paths, and the traversal among multiple ID trees as follows: Traversal among multiple ID trees Figure 10 shows the traversal between two ID trees. Figures 10(a) and 10(b) show the flow of IDN messages and the stream of data packets. According to Figure 10, our prototype system indicates that the our design performs correctly. ＩＤ Tree1 Ｓ1

Ｄ2

ＩＤ Tree2 Ｓ2

ＩＤ Tree1

ＩＤ Tree2

Ｓ1

Ｓ2

Ｄ1

(a) ID trees with two different start (b) Data packet forwarding between nodes. S1-D1 and S2-D2. Fig. 10. Forwarding data packets using two disjoint start nodes

We have performed the preliminary experiments using the prototype. In our preliminary experiments, we compared our prototype with AODV in terms of the data packet delivery ratio and the amount of received packets.

124

T. Okazaki et al. Table 1. Parameters for Preliminary Experiments Simulator QualNet ver.4.5 The number of data SD pairs 30 Data packet size[byte] 512 Interval of data [sec] 0.25 Radio area [m] 250 MAC layer protocol IEEE802.11 Maximum bandwidth [Mbps] 2 Simulation time[sec] 500 Node maximum moving speeds [m/s] 0

Table 2. The number of nodes and field sizes The number of nodes 100 200 300 400 500 Field size [m2 ] 1550× 2200× 2700× 3120× 3500× 1550 2200 2700 3120 3500

Tables 1 and 2 show the parameters of the preliminary experiments, and the number of nodes and field sizes. The field sizes are set so that the average number of neighboring nodes is set to be around 12. This parameter is set in order to prevent the occurance having orphan nodes. For the application layer protocol, we employed CBR /UDP (Constant Bit Rate / User Datagram Protocol). The length of the Dynamic ID is 32 bits. The following simulation performed; 1 second from the start of the simulation, the first source node starts transmission, and the second source node transmits data after the transmission of the first node. Afterwards, the other source nodes start transmission every 1 second. This is done to confirm the performance as it gives enough time to exchange route information among nodes which have been assigned their Dynamic IDs. When the node is assigned its Dynamic ID, it broadcasts RTUP every 1 second up to 10 times. When the RTUP is broadcasted more than 10 times, there is no significant change in performance. Figure 11 shows the 30 runs average of the data packet delivery ratio versus the number of nodes for the proposed method and AODV. The x-axis and yaxis show the number of nodes and the data packet delivery ratio, respectively. The error bar indicates the level of error with 95 % accuracy of each average. According to Figure 11, the proposed method is almost identical to AODV. Figure 12 shows the 30 runs average of the amount of received data packets versus the number of nodes for the proposed method and AODV. The x-axis and y-axis show the number of nodes and the amount of received data packets, respectively. The error bar indicates the level of error with 95 % accuracy of each average. According to Figure 12, the proposed method is almost identical to AODV.

A Multipath Routing Method with Dynamic ID

Proposed

125

AODV

100 90 %][ 80 iot 70 aR 60 yr 50 veli 40 eD 30 20 10 0 100

200

300 400 500 Number of nodes

600

700

Fig. 11. Packet delivery ratio

st ek ca p de ive ce R fo tn uo m A

Proposed

20

AODV

15

]c se/ ti 10 bK [ 5 0

100

200

300 400 500 Number of nodes

600

700

Fig. 12. The amount of received packets

4.2

Experiments to Evaluation

We performed simulation experiments to evaluate our prototype implementation. We define the term “routing load” as the proportion of the number of control packets in the network to the received data packets of the destination nodes. In the experiment, we compared our method to AODV and AOMDV in terms of the number of control packets and the routing load. Number of Control Packets. The parameters of the experiments are shown in Table 3. Table 4 shows the number of nodes and the field sizes. Figure 13 depicts the 30 runs average of the number of control packets in the network for the proposed method and AODV. The x-axis and the y-axis show the number of nodes and the number of control packets. The error bar indicates the level of error with 95 % accuracy of each average. According to Figure 13 , the proposed method has a smaller number of control packets than that of AODV. In addition,

126

T. Okazaki et al. Table 3. Parameteres for experiments of the number of control packets Simulator QualNet ver.4.5 The number of SD pairs 30 Data packet size [byte] 512 Interval of data packets [sec] 0.25 Radio Area[m] 250 MAC Layer protocol IEEE802.11 Maximum Bandwidth[Mbps] 2 Simulation time [sec] 500 Maximum node speed [m/s] 0

Table 4. The relation between number of nodes and field sizes The number of nodes 100 200 300 400 500 Filed size [m2 ] 1550× 2200× 2700× 3120× 3500× 1550 2200 2700 3120 3500

st 30000 ek ca 25000 pl or 20000 tn 15000 o C fo 10000 re 5000 b m uN 0

Proposed

100

200

AODV

300

400

Number of ｎodes

500

Fig. 13. The number of control packets

as the number of nodes increases, the gap between our method and AODV widens. While AODV uses flooding to discover routes, our method does not. The gap expresses the difference between the proposed method and AODV. The Routing Load. Then, we measured the routing load under the conditions as given in Table 5. The number of nodes and the field sizes are the same as in preliminary experiments with exception of the number of source and destination node pairs. Figure 6 shows the 30 runs average of the results. According to Figure 6, our proposed method has the smallest routing load among our proposed method, AODV, and AOMDV. While AODV and AOMDV use flooding to discover routes, our method does not. The gap expresses the difference among the proposed method, AODV, and AOMDV.

A Multipath Routing Method with Dynamic ID

127

Table 5. Parameters for experiment of the routing load The number of nodes 100 Field Size[m2 ] 2200×600 The number of SD pairs 25

Table 6. Routing Load

Proposed AOMDV AODV

Routing Load 0.482 1.340 1.583

End-to-end Delay. We also have investigated the end-to-end delay between the source and destination nodes. Table 7 shows the results of the average time of the end-to-end delay for our proposed method, AOMDV, and AODV. The results for each method in Table 7 are the average of 30 runs each. The results of our proposed method and AODV are from our simulation experiments. The results of AOMDV are based on the data from literature [5]. Figure 14 shows the results of the end-to-end average delay versus the number of nodes. The results in Figure 14 are also the average of 30 runs. Table 8 shows the average number of hops between the source and destination nodes. The results in Table 8 show that the average number of hop counts for our proposed method is longer than that of AODV by about one hop(s). Table 7. The average time of the end-to-end delay for our proposed method, AOMDV, and AODV

Proposed AODV AOMDV

End-to-end delay (sec) confidence interval 0.099 0.005 0.075 0.004 0.041 0.008

Table 8. The average number of hops between the source and destination nodes Number of nodes Proposed AODV 100 200 300 400 500

4.82 6.82 8.57 9.57 10.98

4.10 5.62 6.90 7.91 10.27

According to Tables 7 and 8, the end-to-end delay time of the proposed method is the longest among our proposed method, AODV, and AOMDV. This is because our method constructs tree-based multiple-paths. In contrast, AODV

128

T. Okazaki et al.

s][ ya le d dn Eo T -d n E

0.15

Proposed

AODV

0.1

0.05

0

100

200

300

400

Number of nodes

500

Fig. 14. The average time of the end-to-end delay vs the number of nodes for our proposed method and AODV

constructs the shortest path. Also, since AOMDV is the modified method of AODV, the end-to-end delay time becomes the quasi-shortest path. Furthermore, AOMDV compensates for unstable wireless links by using alternate paths. As a result, AOMDV has the shortest end-to-end delay. However, the values of the end-to-end delay time of the proposed method, AODV and AOMDV are small. When the route breaks occur frequently in the networks, this trend will be possible to change. Further study will be necessary to investigate the performances for our proposed method when the route breaks occur frequently.

5

Conclusion and Future Work

In this paper, we have proposed a new multipath routing method to decrease the routing load based on DART. Its prototype system has been implemented. In addition, we have evaluated our proposed method by conducting simulation experiments of our method, AODV, and AOMDV. According to the results, our method can establish multiple paths with smaller routing loads than AODV and AOMDV. As for future work, we plan to extend our proposal to node mobility, and to investigate the potential scope of application of our method.

Acknowledgment This research is supported by the Ministry of Education, Science, Sports and Culture of Japan under Grant-in-Aid for Scientific Research (B) (No.21300028). This research is also supported by the National Institute of Information and Communications Technology, Japan under Early-concept Grants for Exploratory Research on New-generation Network (No.145-9), and the Ministry of Education, Science, Sports and Culture of Japan under Grant-in-Aid for Scientific Research (C) (No.22500065).

A Multipath Routing Method with Dynamic ID

129

References 1. Abidi, S., Erfani, S.: A multipath routing survey for mobile ad-hoc networks. In: Proceedings of the 3rd IEEE Consumer Communications and Networksing Conference (CCNC 2006), vol. 2, pp. 984–988 (2006) 2. Eriksson, J., Faloutsos, M., Krishnamurthy, S.V.: DART: Dynamic address routing for scalable ad hoc and mesh networks. IEEE/ACM Transactions on Networking 15(1), 119–132 (2007) 3. Johnson, D.B., Maltz, D.A.: Dynamic source routing in ad hoc wireless networks. In: Mobile Computing, vol. 353, pp. 153–181. Kluwer Academic Publishers, Dordrecht (1996) 4. Lee, S.-J., Gerla, M.: Split multipath routing with maximally disjoint paths in ad hoc networks. In: Proceedings of IEEE International Conference on Communications (ICC 2001), vol. 10, pp. 3201–3205 (2001) 5. Marina, M.K., Das, S.R.: Marina and Samir R. Das. Ad hoc on-demand multipath distance vector routing. SIGMOBILE Mobile Computing and Communication Review 6(3), 92–93 (2002) 6. Perkins, C., Belding-Royer, E., Das, S.: Ad hoc On-demand Distance Vector (AODV) routing. RFC 3561 (Experimental) (July 2003) 7. Perkins, C.E., Royer, E.M.: Ad-hoc on-demand distance vector routing. In: Proceedings of the Second IEEE Workshop on Mobile Computing Systems and Applications, pp. 90–100. IEEE Computer Society Press, Los Alamitos (February 1999) 8. Scalable Network Technologies, Inc., QualNet network simulator

Terminal Design without Using Receiver Circuits for Wireless Sensor Networks Hiroaki Nose1 , Miao Bao2 , Kazumasa Mizuta3 , Yasushi Yoshikawa4 , Hisayoshi Kunimune2 , Masaaki Niimura2 , and Yasushi Fuwa2 1

Nagano Prefectural Institute of Technology, Ueda-shi 386-1211, Japan 2 Shinshu University, Nagano-shi 380-8553, Japan 3 Epson Toyocom Corporation, Hino-shi 191-8501, Japan 4 Seiko Epson Corporation, Suwa-shi 392-0001, Japan

Abstract. Sensor network terminals are installed in large numbers in the field and transmit data periodically by radio. Such terminals must be miniaturized, and save power so that each device can operate by battery for several years. As one way to satisfy these two conditions, in this research we propose a terminal design which eliminates the receiver circuit. Because there is no receiver circuit, circuitry can be miniaturized, and power can be saved because there is no need to consume power to receive signals. However, the terminals cannot perform carrier detection and reception acknowledgement because there is no receiver circuit. We propose following two new protocols to solve this problem. 1. Terminal transmission times are randomized to prevent frequent collisions between specific terminals due to the lack of carrier detection. 2. Since all packet losses due to collision cannot be prevented with 1, data from a number of past transmissions is included in each packet so that a later packet can provide transmission data even if a packet is lost. In this report, we describe the proposed protocol, and evaluate its performance by simulation. Furthermore, we actually prototype the system and evaluate the prototypefs performance. Keywords: Sensor Network, Sensor Node, Power Saving, Protocol.

1

Introduction

Sensor networks which interconnect numerous miniature sensors with built-in wireless communication capability are attracting attention as part of the effort to realize a ubiquitous network enabling easy communication anywhere, at any time, by anything. SmartDust [1] is well known as a pioneering approach in sensor network research. At present there is active research on applications in a broad range of fields including medicine/welfare, crime prevention/security, disaster prevention, and agriculture and devices are being sold as products in an increasing number of cases. Progress is particularly being made toward applications in safety monitoring systems for the elderly and ill, and monitoring of the natural environment, disaster sites and other locations. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 130–145, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Sensor Network

131

Fig. 1. Overview of system

Sensor network terminals must be small because the aim is to widely disperse them in the natural environment and attach them to mobile objects such as human. They must also save power because they are expected to be used in environments where power supplies cannot be easily secured. The terminal devices proposed in this research have no receiver circuits, and thus enable greater miniaturization and power saving by the sensor network terminals. We also propose a protocol for lessening the effects of collisions, which occur frequently due to the lack of a receiver circuit. The effectiveness of these proposals is verified via simulation, and we also conduct an evaluation by prototyping the proposed terminals and building a sensor network.

2

Sensor Network Issues

Fig.1 shows the configuration of the sensor network. Numerous independent terminals (sensor nodes) autonomously communicate, periodically sending the information they sense to a data gathering node. Terminals may be distributed in large numbers over an extremely wide range, as would be expected in applications such as safety monitoring and disaster prevention, and it will be crucial to reduce terminal power consumption and achieve greater miniaturization. 2.1

Reducing Power Consumption

In general, terminals of the sensor network will use batteries because they will be attached to people and other mobile objects, or dispersed in the natural environment, and thus are not expected to be used under conditions where a power supply can be easily secured. If terminal size is not a problem, the issue of service life can be solved by using a combination of a photovoltaic cell and a secondary (rechargeable) battery, but the cost will be extremely high. Furthermore, if it is assumed that terminals will be attached to people and other mobile objects, size will become a problem. Therefore, the favored approach is to incorporate a single primary battery with a comparatively small size and use it as the power supply. However, in this case battery life becomes an issue. In a sensor network where a large number of terminals are distributed over a wide range, the cost of replacing batteries will be extremely high, and this will be a problem directly affecting operation of the system. There are expected to be situations where battery replacement itself

132

H. Nose et al.

will be difficult, particularly in the case of terminals dispersed in the natural environment. To lower these costs it will be necessary to extend battery life by reducing terminal power consumption. 2.2

Miniaturization

Sensor network terminals are likely to be used in the following situations: 1. People attach terminals to their bodies, and transmit data such as their vital signs. 2. Terminals are placed in server rooms and transmit temperature of installed equipment etc. 3. Terminals transmit temperature and other data for the locations where they are installed in the natural environment. For 1 and 2 installation space is extremely limited. When terminals are attached to the body, they should not interfere with the personfs movement, and it should be possible to wear the terminal without being aware of it. In a server room, terminals must be installed in locations with limited space, such is inside racks, and thus large terminals cannot be installed. In case 3, there are no limitations on the space for installation, but since the terminals are likely to be disposable, it would be ideal for terminals to have the least possible impact on the environment. Thus there is a strong need to miniaturize sensor network terminals due to the way they will be used.

3

Purpose of This Research

For reducing power consumption of sensor network terminals, Wei Ye et al. have proposed an S-MAC approach where terminals efficiently enter a sleep state to suppress power consumption of wireless devices when they are on standby waiting to receive [2]. With this technique, terminals have two states (active and sleep) and communication is only possible in the active state. All terminals must simultaneously become active in order to communicate, but terminals are synchronized by broadcasting to neighboring nodes that a node is active. There are many studies deriving from S-MAC which attempt, for example, to further improve power saving performance by dynamically changing the period that a terminal is active [3],[4],[5]. A technique has also been proposed which reduces power consumption by synchronizing the times when terminals are active (using a radio clock or other accurate clock), and scheduling receive times [6]. Low Power Listing (LPL) is a common technique which does not use synchronization. The sensor nodes switch to the active state for a time at a fixed interval T, and check the transmit channel. They continue in the active state and receive packets only if the channel is in use at that time. On the other hand, transmitting nodes notify other nodes that there is a request to send by transmitting a

Sensor Network

133

Fig. 2. Sensor terminal (right) and receiver (left)

preamble longer than T. Well known techniques based on LPL include B-MAC [7], X-MAC [8], and TICER [9]. Recently, ZigBee (IEEE 802.14.4) [10] has attracted attention as a sensor network protocol, and sensor modules are being commercialized. With ZigBee, nodes are classified into 3 types: ZigBee coordinators, ZigBee routers and ZigBee end points. Of these, only the ZigBee endpoints which do not relay information have power saving specifications. Thus all of the sensor network protocols which have currently been proposed assume that the sensor network terminals have reception capability, and strive to reduce power consumption by efficiently suppressing power supplied to the receiver circuit. In order to further reduce power consumption and achieve greater miniaturization, we have developed terminals without a receiver circuit, as shown in Fig.2, and we propose a sensor network using these. However, since the sensor network terminals have no reception capability, the network cannot use conventional protocols employing carrier detection. Therefore there is an increase in packet loss due to collision, and major problems arise in the overall performance of the system. To solve this problem, we propose a new protocol for sensor networks which use sensor network terminals with no receiver circuit. This protocol is comprised of two elements: providing a random wait time before starting transmission, and redundantly transmitting past data in each packet. This research examines the proposed protocol in detail through simulation. In addition, we implement the proposed protocol in actual sensor network terminals with no reception capability, and confirm the effectiveness of the proposed technique by conducting experiments on the data loss rate in an environment where multiple terminals are simultaneously operating.

Fig. 3. Packet transmission timing

134

4 4.1

H. Nose et al.

Proposed Protocol Elimination of Receiver Circuit

The main role of terminals in a sensor network is to send data measured by the terminals’ sensors to the data gathering server. Therefore it is rare for sensor network terminals to receive any data, and in some systems they don’t receive any data at all. In the sensor network system we are developing, it was decided to eliminate receiver circuits from the sensor network terminals. This makes it possible to eliminate power consumed by the receiver circuits, thereby enabling major power savings. At the same time, it also enables miniaturization of network terminals. 4.2

Problems with Eliminating Receiver Rircuits

Generally, protocols for wireless networks perform carrier detection before starting transmission, and transmission begins after confirming that no other terminal is transmitting. To avoid collisions between transmissions starting at the same time due to simultaneous detection of the no-carrier state by different terminals, a typical approach is to provide a random wait time until the start of transmission. However, the sensor network terminals used in this system omit reception capability in order to reduce power consumption, and thus carrier detection is impossible.Thus there is no way for a terminal to check the state which other terminals are in, and the only thing a terminal can do is transmit one-way. As a result, all terminals of the system repeatedly transmit in a disorderly fashion. Transmission collisions frequently occur, and this causes a marked drop in throughput of the overall system. Occurrence of Constantly Colliding Terminals. The sensor network terminals of this system transmit one packet at a time at a fixed time interval of X [sec], as shown in Fig.3. The beginning of the first transmission is determined simply by the time when the terminal’s power was switched on, and thus transmission is asynchronous, with no relationship whatsoever to other terminals or the time etc. Therefore, if there are multiple terminals which happen to start their first transmission at the same time, their transmissions will collide every time, and there will be many terminals which can never complete their transmission. Unconfirmed Arrival of Data. With ordinary protocols, the receiving side sends ACK to the transmitting side when a transmission has been received properly, and both sides recognize that the communication has finished successfully when the transmitting side receives this ACK. With this system, however, the terminals do not have a way to confirm that their own transmission was successful. Therefore if packet loss occurs frequently due to factors such as increasing the number of terminals, there may be a marked loss of communication quality of the overall system.

Sensor Network

135

Fig. 4. Packet transmission timing with random delay times

4.3

The Proposed Protocol

This system uses terminals which cannot perform carrier detection because reception capability has been eliminated. Therefore, we propose a new protocol which avoids the aforementioned occurrence of constantly colliding terminals and ensures successful communication. Setting of Random Delay Time. Terminals transmit once at every fixed time interval X [sec]. Therefore, as stated in 4.2, terminals which started transmitting at the same time, will continually conflict thereafter and transmission will fail. Thus a random transmit delay time T ranging from 0 - Y [sec] (0 ≤ T ≤ Y ) is provided before starting transmission, as indicated in Fig.4. In this way, the transmission timing is varied within a range up to maximum of Y each time transmission is attempted, and this makes it possible to avoid the occurrence of terminals whose transmissions constantly collide because they started transmission at the same time. Transmission once in the time X is assured by setting Y so that X > Y + P , where P is the time required to send a packet. Redundant Transmission of Past Data. Since terminals cannot received an ACK from the receiving side, it is impossible for a terminal to confirm whether a packet it sent was received correctly by the gathering server. Therefore, if transmission fails due to collision with another terminal, the data which the terminal was trying to transmit will be lost. In other words, the effects due to packet loss will be extremely large. Thus, in order to handle the situation where transfer of data is not completed in a single transmission, data for the past Z − 1 times is added to the latest data currently being transmitted in the transmission packet so that data for Z transmissions is sent in a single packet, as

Fig. 5. Redundant transmission of packets

136

H. Nose et al.

Fig. 6. Frame format of terminal packet

indicated in Fig.5. This ensures that data arrives on the receiving side, provided that transmission failure only occurs consecutively less than Z times, and this improves transmission reliability. However, we must take into account the fact that increasing Z increases packet length, and thereby increases the probability of packet collision. (This is verified in 5.2.)

5 5.1

Evaluation Specifications

This section describes the specifications of the system being evaluated. The frequency used is the 315 MHz band, and communication speed is 60 kbps. The format of the packets sent by the terminals is as indicated in Fig.6, with each packet consisting of a preamble, terminal ID, data fields and check sum. In order to realize the redundant transmission in the proposed protocol, the data fields contain multiple past data items in addition to the most recent data. 5.2

Evaluation of Data Loss Rate by Simulation

The data loss rate is defined as the percentage of data which cannot be received correctly by the data gathering node, relative to the number of data items generated in each fixed time interval X. The data loss rate for the proposed protocol was found by simulation. The system used for simulation was configured as shown in Fig.1, and incorporated 100 terminals. The transmission repeat time X for each terminal was set to 1 [sec], and 100 packets were sent, so the system overall was highly loaded. The start up timing of each terminal was also set to be random. For this system, Z was increased in increments of 2 from 1 to 9, and Y was increased in increments of 200 [msec] from 100 [msec] to 900 [msec], and simulation was conducted 5 times for each of these conditions. Fig.7 shows the average data loss rate, and Fig.8 the maximum data loss rate in the simulation results. In both graphs, the horizontal axis indicates the Y value, and the vertical axis indicates the data loss rate. In Fig.7, the average data loss rates were found for all terminals in all 5 simulations, and 95% confidence intervals are also shown in the graph. Fig.8 shows the results for maximum data loss rates. The reason for adopting the maximum data loss rate as the object of evaluation is as follows. In the current system, the transmission timing is determined by when the power supply for each terminal is turned on. Therefore, if the Y value is 0, there will be groups of terminals whose transmitted packets are always lost due to

Sensor Network

137

Fig. 7. Average data loss rate

Fig. 8. Maximum data loss rate

collision. Even if a Y value is provided, variations in timing will arise depending on its value, and biases are likely to appear in the data loss rate. Therefore, even if the average data loss rate is, say, 3% under certain conditions, and that is within the permissible range for the application, a specific set of terminals may have a data loss rate of 30%, thereby resulting in problems. In the case of this system, where a bias can arise in the data loss rate between terminals, we felt there were problems with discussing protocol performance using only the average data loss rate, and thus it was decided to use both the average and maximum as a basis for evaluation. To find the maximum data loss rate here, simulations were performed 5 times under the same conditions while varying the timing of switching on power. Then the data loss rate was found for the terminal whose loss was the greatest in all of the simulations, and the average of those values over the 5 times was taken to be the maximum. Introduction of Random Delay Time Y . First we verified the effectiveness of introducing the random delay time Y . A major improvement due to introducing Y can be seen in both the average and maximum data loss rate. When the random delay time was not provided, the average loss rate was 30% or higher,

138

H. Nose et al.

Fig. 9. Change in average data loss rate due to the number of terminals (Y = 900 [msec])

Fig. 10. Change in maximum data loss rate due to the number of terminals (Y = 900 [msec])

and the maximum loss rate was 100% due to constantly colliding terminals. In contrast, when the random delay time was introduced, the average loss rate dropped below 10% for Z = 2 or higher. When Z = 1, introducing the random delay time Y has almost no effect on the average value. This is because a random delay time is simply added to the random transmission timing determined only by the original time the terminals were started up, and thus no significant difference was seen in transmission timing. However, for Z = 1 the maximum was 50% or less, and for Z = 2 or higher it was 25% or less. In both cases, effects due to the introduction of Y were evident. The main factor due to the introduction of Y was the elimination of constantly colliding terminals. Also, if we look at changes in the loss rate with respect to increases in Y , almost no change is seen for the average, but a small reduction is evident for the maximum, and the loss rate reaches a minimum when Y = 900 [msec]. Introduction of Redundant Transmission of Past Data Z. Next, we verify the effect of introducing redundant transmission of past data Z. In the

Sensor Network

139

Fig. 11. Change in average data loss rate due to the number of terminals (Y = 0 [msec])

case where Y has been introduced, the average was about 35% for the case where Z was not introduced (Z = 1) in Fig.7, but it improved to about 9% for Z = 3, 4% for Z = 5, and 3% for Z = 7. There was also an improvement with maximum values in Fig.8. While the maximum was about 45% for the case where Z was not introduced (Z = 1), it improved to about 20% for Z = 3, 13% for Z = 5 and 10% for Z = 7. For both the average and maximum value, introducing Z led to a major improvement in the loss rate, and the loss rate decreased as Z increased. However, when Z was 5 or more, increasing Z only led to slight improvement. This may be because increasing Z makes it possible to recover even if packets are consecutively lost, but packets get longer as Z increases, and this leads to an increase in the collision probability. Evaluation Based on Number of Terminals. The effect of introducing Y was verified in 5.2. There it was confirmed that the effect is greatest for the maximum loss rate with Y = 900 [msec]. Therefore simulations were conducted to determine the change in loss rate when the number of terminals was varied from 20 to 100 under these conditions. The results are shown in Fig.9 and Fig.10. In the graphs, the horizontal axis indicates the number of terminals, and the vertical axis indicates the data loss rate. It can be confirmed from the graphs that, when Z was set to 5 or higher, both the average and maximum loss rate were held to 5% or less for up to 60 units. This loss rate may be adequate for practical use in selected applications. For comparison, Fig.11 and Fig.12 show the average and maximum loss rates resulting from a simulation with Y = 0 [msec]. Y = 0 means that terminals which collide will continue to repeatedly collide, and therefore there is no hope of improving the data loss rate by repeating data in packets. In particular, it was confirmed that the maximum value reaches 100% when the number of terminals exceeds 40, and thus the system is not practical.

140

H. Nose et al.

Fig. 12. Change in maximum data loss rate due to the number of terminals (Y = 0 [msec])

5.3

Evaluation Using Actual Equipment

Experiments were conducted by implementing the proposed technique with the actual sensor network terminals shown in Fig.2. However, since only 25 sensor network terminals could be procured for the experiment, the network load had to be increased in order to recreate the same trends as the simulations shown in 5.2. Therefore, the experiments were conducted by setting the transmission interval X of the sensor network terminals, which was 1 [sec] in the simulations, to 200 [msec]. Fig.13 shows the average loss rate, and Fig.14 the maximum loss rate, when the number of terminals was set to 10, Y was set to 20 [msec], 100 [msec] and 180 [msec], and Z was set to 1, 3 and 5. Fig.15 shows the average loss rate, and Fig.16 the maximum loss rate, when the number of terminals was set to 25. For comparison, the same simulations as in 5.2 were performed under the same conditions and both the average and maximum were recorded in the graphs. The 95% confidence intervals for each result are shown on each symbol in the graphs. Comparing the case with 10 terminals against the case with 25 terminals, it is evident that both the average and maximum data loss rate increase as the number of terminals increase. Therefore, it was confirmed by experiments with actual equipment that increasing the number of terminals results in a drop in performance of the entire network. The average data loss rates in experimental results with 25 units were as follows. When Y was increased from 20 [msec] to 180 [msec], data loss improved from 11% to 6.7% for Z = 3 and from 5.4% to 4.0% for Z = 5. When Z was increased from 1 to 5, there was a major reduction from 30.2% to 4.0% for Y = 180 [msec], and this shows the marked effect of introducing Z. For the maximum values too, increasing Y from 20 [msec] to 180 [msec] greatly reduced the data loss rate from 52.6% to 21.8% for Z = 3 and from 48.1% to 11.9% for Z = 5, thus showing that occurrence of constantly colliding terminals was effectively suppressed. Also, introducing Z improved the data loss rate, with a drop from 46.4% to 11.9% when Z was increased from 1 to 5 with Y = 180 [msec].

Sensor Network

141

Fig. 13. Average data loss rate (10 terminals, transmission interval X = 200 [msec])

Fig. 14. Maximum data loss rate (10 terminals, transmission interval X = 200 [msec])

When the results of experiments using actual equipment are compared with results based on simulation, some differences are evident in the numerical values for the data loss late, but Y and Z exhibit almost the same trend of improving the data loss rate, and the evaluation of the proposed protocol via simulation in 5.2 was confirmed to be valid. 5.4

Power Consumption

Battery life was verified in two cases: when the proposed protocol was implemented and receiver circuits were eliminated, and when a receiver circuit was installed and carrier detection and ACK reception were performed. Assuming use of an ordinary coin-cell battery, capacity was set to 225mAh, and the fixed time interval X was set to 1 [sec]. When both carrier detection and ACK reception are performed, transmission/reception operation of sensor network terminals is basically comprised of receiver module setting, RSSI, transmission and ACK reception (Fig.17). The current consumption required by each operation was measured with a prototype, and battery life was determined by

142

H. Nose et al.

Fig. 15. Average data loss rate (25 terminals, transmission interval X = 200 [msec])

Fig. 16. Maximum data loss rate (25 terminals, transmission interval X = 200 [msec])

calculating based on those values. When the time and power consumption required for reception were measured using a prototype circuit, it was found, for the period from switching on power to the RFIC circuit until reception was enabled, that current consumption for the first 2.2 [msec] interval was 17 mA, that current consumption for the following 2.2 [msec] interval was 23 mA, and that after that reception was enabled. To perform carrier detection, receive processing was performed for 0.5 [msec] after this processing to switch on power of the RFIC circuit (current consumption at that time was 23 mA). Current consumption during ACK packet reception, after power switch-on processing of the RFIC circuit in the same way, was 23 mA. It was determined by measuring the prototyped circuit that the transmitter circuit required 10 mA of current consumption. First of all, Fig.18 calculates the effect of the parameters of this protocol Y and Z on battery life, in the state with no receiver circuit. This graph shows that, when Z is varied from 1 to 10, battery life is reduced by almost half. Since

Sensor Network

143

Fig. 17. Transmission/Reception modes of sensor terminal

Fig. 18. Effect of introducing Y and Z on battery life

increasing Z directly increases packet length, and the time necessary for transmission increases with longer packet length, battery life may become shorter. On the other hand, even if Y is varied from 100 [msec] to 900 [msec], this only shifts the transmission timing, and has almost no effect on battery life. Some reduction in life is evident in response to increasing Y , but this is because the maximum setting of the timer for measuring the sleep state of sensor network terminals in this system is 250 [msec] and if Y exceeds that value, it is necessary to reset temporarily from the sleep state, and thus there is a slight increase in power needed to do this. In contrast with the case with no receiver circuit, Fig.19 assumes that Y = 900 [msec] and calculates battery life when a receiver circuit is provided. In addition to the case with no receiver circuit, three cases with a receiver circuit are shown: carrier detection only, ACK reception only, and both carrier detection and ACK reception. In the case with no receiver circuit in the proposed protocol, life is 4500 hours for Z = 1 and 2800 hours for Z = 10. In the cases with a receiver circuit, battery life was shortest when both carrier detection and ACK reception were implemented. In that case, life was about 650 hours for Z = 1, and 600 hours for Z = 10. In the case of carrier detection only, where life was longest, life was about 1300 hours for Z = 1 and 1100 hours for Z = 10. This calculation was done using an extremely short fixed time interval of X = 1 [sec] for the sensor network, and thus, if X is assumed to be 10 [sec], it should be possible to achieve a battery life of about 5 years with Z = 1 and 3 years with Z = 10. These results confirmed that power saving can be realized by eliminating the receiver circuit.

144

H. Nose et al.

Fig. 19. Battery life with receive operation (Y = 900 [msec])

6

Conclusion

In order to realize a sensor network using terminal devices with no receiver circuits, this paper proposed a new protocol which does not require carrier detection or ACK reception. If the sensor network terminals have no receiver circuits, collisions will occur frequently due to the lack of carrier detection, and constantly colliding terminals in particular will cause a marked decrease in system performance. Furthermore, since ACK reception is not performed, there is a problem due to loss of certainty regarding data arrival. The contributions of this paper can be summarized in the following four points: 1. In transmission where carrier detection is not performed, each sensor terminal repeatedly transmits on its own, and therefore constantly colliding terminals arise. Consequently, there was a problem in that the maximum data loss rate for the sensor network reached 100%. However, by providing a random delay time when transmitting from each sensor network terminal, it was possible to improve the maximum data loss rate to 50% or less. 2. Since ACK reception is not performed, each sensor network terminal cannot confirm that the sent data successfully arrived on the receiving side. Therefore, if a packet is lost due to collision, the data in it will be lost. However, if past data is sent redundantly while also providing the random delay time in 1, then the average data loss rate can be improved to about 3%. 3. Major improvements were seen, just as in the simulations, in experiments conducted by implementing the protocols proposed in 1 and 2 in actual equipment, and the effectiveness of the proposal was confirmed. 4. Battery life of the sensor network terminals was calculated by measuring current consumption in each operation of the actual equipment. As a result, it was confirmed that eliminating the receiver circuit was effective for lengthening battery life. Going forward, we plan to develop actual applications in fields such as agriculture, medicine and disaster prevention.

Sensor Network

145

Acknowledgment This research was partially supported by the Ministry of Internal Affairs and Communications, Strategic Information and Communications R&D Promotion Programme ”R&D on Highly Damage Resistant Local Disaster Communication Systems using Ad-Hoc Networks and Sensor Networks” (092304014) (2009 010).

References 1. Kahn, J.M., Katz, R.H., Pister, K.S.J.: Mobile Networking for Smart Dust. In: ACM/IEEE Intl. Conf. on Mobile Computing and Networking (MobiCom 1999), Seattle (1999) 2. Ye, W., Heidemann, J., Estrin, D.: An Energy-Efficient MAC Protocol for Wireless Sensor Networks. In: IEEE INFOCOM, New York, pp. 1567–1576 (2002) 3. Dam, T.V., Langendoen, K.: An adaptive energy-efficient MAC protocol for wireless sensor networks. In: ACM SenSys, Los Angeles, pp. 171–180 (2003) 4. Lu, G., Krishnamachari, B., Raghavendra, C.S.: An adaptive energy-efficient and low-latency MAC for data gatherring in wireless sensor networks. In: IEEE IPDPS, New Mexico, pp. 224–231 (2004) 5. Du, S., Saha, A.K., Jhonson, D.B.: RMAC:A routing-enhanced duty-cycle MAC protocol for wireless sensor networks. In: IEEE INFOCOM, Alaska, vol. 3, pp. 1478–1486 (2007) 6. Watanabe, T., Morito, T., Minami, M., Morikawa, H.: Design and implementation of a synchronous, battery-less, wireless sensor network using a radio clock. IEICE Technical Report, pp. 113–118, Mobile Multimedia Communications 107(39) (2007) 7. Polastre, J., Hill, J., Culler, D.: Versatile low power media access for wireless sensor netwoeks. In: ACM SenSys, Maryland, pp. 95–107 (2004) 8. Buettner, M., Yee, G.V., Anderson, E., Han, R.: X-MAC: A short preamble MAC protocol for duty-cycle wireless sensor networks. In: ACM SenSys, Colorado, pp. 307–320 (2006) 9. Lin, E.A., Rabaey, J.M.: Power-efficient rendezvous schemes for dense wireless sensor networks. In: IEEE ICC, Paris, pp. 3769–3776 (2004) 10. ZigBee Alliance, http://www.zigbee.org/en/

Cooperative Spectrum Sensing in Ad-Hoc Networks (Invited Paper) Liljana Gavrilovska and Vladimir Atanasovski Faculty of Electrical Engineering and Information Technologies Ss Cyril and Methodius University in Skopje Karpos 2 bb, 1000 Skopje, Macedonia {liljana,vladimir}@feit.ukim.edu.mk

Abstract. Spectrum sensing is a distinct feature of cognitive ad-hoc nodes that have the ability to opportunistically use vacant spectrum bands for its own communication purposes. Possible cooperation among the nodes may prove vital for increasing network performance. It yields the cognitive ad-hoc nodes to exchange relevant environmental and context information in order to enhance its own networking experience. This paper overviews the approaches, techniques and strategies for cooperative spectrum sensing and gives a practical example of a realized testbed implementation in an ad-hoc environment. Keywords: spectrum sensing, cooperation, cognition, ad-hoc networks, RAC2E.

1 Introduction Ad-hoc networks represent temporary network structures formed whenever two or more wireless nodes exhibit a need for exchanging information [1]. They differ from traditional structured networks in terms that every ad-hoc network participant must be able to act as a potential router of information for other nodes. This opens the possibility for implementing various cooperation methods among the wireless nodes in ad-hoc environments that ultimately leads to increased network performance (e.g. increased throughput, increased reliability, lower delay etc.). The introduction of cognitive radios [2] sheds new light on the ad-hoc networking paradigm. Cognitive radios are autonomous wireless devices able to optimize, learn and reason upon different network information (available both locally and globally in the network). Their quintessential feature is the ability to perform spectrum sensing, i.e. scan the available spectrum bands and find a suitable spectrum hole for their own communication purposes. Moreover, cognitive radios must also be able to anticipate the arrival of other users in the band they are currently communicating in and perform spectrum mobility (change the channel) in order to minimize the possible interference in the network. As a result, there is often a distinction between primary and secondary users of the available spectrum. The former ones are either licensed users or users with a higher priority on the spectrum whereas the latter ones may use the spectrum on an opportunistic basis stemming from their cognitive capabilities. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 146–159, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

Cooperative Spectrum Sensing in Ad-Hoc Networks

147

The cooperation among cognitive radios in ad-hoc environments is a crucial step towards providing efficient network operation. It allows the cognitive ad-hoc devices to exchange data about the sensed spectrum which leads to a faster and more efficient convergence of the communication channel selection process [3-8] leading to improved spectrum management in cognitive environments. This paper presents an overview of relevant spectrum sensing techniques and cooperation strategies in cognitive ad-hoc networks. Additionally, the paper gives a practical implementation example of a realized cognitive ad-hoc network whose participants are able to cooperatively scan the available spectrum band in order to find the most suitable communication channel. The cooperation among the nodes in the example yields a design of novel rendezvous protocol for cooperative data exchange. The paper is organized as follows. Section 2 elaborates on the spectrum sensing techniques that are used by cognitive radios to detect spectrum activity in certain bands. Section 3 discusses the possible cooperation strategies among cognitive radios in ad-hoc environments. Section 4 gives details on a realized testbed platform for cooperation among cognitive radios in laboratory premises. Finally, section 5 concludes the paper.

2 Spectrum Sensing Techniques Every cognitive radio in an ad-hoc network is able to perform spectrum sensing relying only on the locally available information. There are several ways to achieve this task, which are broadly classified as: • • •

Transmitter detection vs. Receiver detection approaches [9] (based on whether the primary user is transmitting or receiving information when the secondary one senses the spectrum); Blind sensing (non-coherent detection) vs. Signal specific (coherent) approaches [10] (based on the usage of specific signal features when sensing the spectrum) and Interference based approach (using a specially defined spatially dependent parameter for interference tolerance by the primary users).

2.1 Transmitter Detection vs. Receiver Detection Sensing Approaches The transmitter detection approaches assume that a primary user is transmitting information to a primary receiver when a secondary user is sensing the primary channel band. The presence of the primary transmission can be extracted by a secondary user by several techniques such as: • • •

energy detection [11]; matched filter detection [12] and cyclostationary feature detection [13].

The energy detector estimates the signal power in the channel band where the primary transmission is occurring and compares that estimate with a predefined threshold. As in most general cases of spectrum sensing no a priori information for

148

L. Gavrilovska and V. Atanasovski

the primary transmission is known to the secondary user, the energy detection is the only possible solution for spectrum sensing. However, this technique has several drawbacks such as: the decision threshold is subject to variations with the SNR, the energy detector cannot distinguish between a user signal and interference, the energy detector is not effective for spread (i.e. wideband) signals etc. The matched filtering is an optimal way to detect signals in communication systems. The main advantage of this technique is that it can provide high processing gain in short time, however the drawback is the need for prior knowledge of some information for the primary transmission (e.g. modulation order, pulse type etc.). Finally, the cyclostationary feature detection uses the cyclostationary feature inherently present in many wireless communications signals. This feature means that the statistical properties of the transmitted signal (e.g. the mean value or the autocorrelation function) change periodically as functions of time. The cyclostationarity is either produced by modulation or coding or is intentionally incurred in order to aid the spectrum sensing. The cyclostationary feature detection is a promising technique able to extract signal features in the background of noise (since the noise is usually wide sense stationary) and, thus, be more effective than energy detection. The receiver detection approaches assume that a primary user is receiving information from a primary transmitter when a secondary user is sensing the primary channel band. They rely on the fact that the primary user in a receiving mode is not passive, i.e. it produces leakage of electromagnetic waves. The secondary users can detect the Local Oscillator (LO) leakage power when the primary user is receiving information and, as a result, detect the primary user [14]. It is obvious that the receiver detection relies on the energy detection technique previously described. The advantages of the receiver detection approaches over the transmitter detection approaches lie in the ability to locate the primary user, locate the exact primary channel band in use and the high probability to find free spectrum even in high density of primary receivers. However, the disadvantages lie in the need for a highly sensitive energy detector, the price of the architecture, the near-far problem etc. 2.2 Blind Sensing vs. Signal Specific Sensing Approaches Another classification of the spectrum sensing techniques may rely on whether the sensing uses some signal specific features or not. In this manner, there are: • •

blind sensing approaches, which do not rely on any signal specific feature (i.e. non-coherent detection), and signal specific sensing approaches, which rely on various signal specific features (i.e. coherent detection) [10].

The blind sensing approaches comprise energy detection (previously elaborated), eigenvalue based sensing [15, 16] and multi-resolution sensing [17, 18]. The eigenvalue based sensing builds upon the robustness of the energy detection and requires knowledge of the eigenvalues of the covariance matrix of the received signal. Based on different ratios of varios eigenvalues (maximum, average and minimum), various algorithms can be defined [15, 16]. The multi-resolution sensing produces a multiresolution Power Spectral Density (PSD) estimate using a tunable wavelet filter that can change its center frequency and its bandwidth. First, the total bandwidth is sensed using

Cooperative Spectrum Sensing in Ad-Hoc Networks

149

a coarse resolution and then a fine resolution sensing is performed on the portion of the interesting bands for the secondary user [17, 18]. Therefore, this method reduces sensing time and saves power from unnecessary computations. However, it increases chip area, power consumption and imposes challenges on the mixer design for multiple frequencies operation. The signal specific sensing approaches consist of the previously elaborated matched filtering and cyclostationary feature detection and some ATSC signal related sensing techniques [10]. Extensive overview on both blind sensing and signal specific sensing approaches can be found in [19]. In addition to the previously elaborated spectrum sensing techniques, there is also an interference based detection method that has its own specifics and is separately elaborated in the following subsection. 2.3 Interference Based Detection Interference is a general limiting factor of useable range and effectiveness of communication systems. As a result of the increase of wireless devices and services lately, current approaches to interference management may no longer be adequate. Therefore, the Interference Protection Working Group of FCC Spectrum Policy Task Force recommended the use of interference temperature metric as a mean to quantify and manage interference [20]. The interference temperature metric is a measure of the RF power available at a receiving antenna to be delivered to a receiver, i.e. the temperature equivalent of the RF power available at a receiving antenna per unit of bandwidth measured in units of Kelvin [K]. It is generated by other emitters and noise sources in the vicinity of the receiver. The interference-based detection strategies for spectrum sensing rely on the prior knowledge of secondary users of a parameter called interference temperature limit. This parameter is defined for different geographic regions and represents the maximum amount of tolerable interference for a given frequency band in a particular location. Thus, the secondary users are allowed to transmit in the given frequency band only if they guarantee that their transmissions added to the existing interference must not exceed the interference temperature limit at a licensed receiver in the same frequency band and in the same location. The use of the interference temperature metric allows the secondary users to adapt the transmit power and the bandwidth of their communication schemes (inevitably causing throughput variations) leading to a maximization of their QoS while minimizing the interference to the primary users [21]. After elaborating the spectrum sensing techniques employed by individual cognitive radios in ad-hoc networks, the following section will report on possible cooperation strategies that increase the reliability of the spectrum detection in the adhoc network.

3 Cooperative Spectrum Sensing Strategies The cooperative detection strategies for spectrum sensing rely on information exchanges among secondary users. The exchanged information can facilitate the detection of spectrum holes and increase the efficiency of the spectrum sensing. It

150

L. Gavrilovska and V. Atanasovski

must be stressed that the secondary users may sometimes also exchange minimal information with the primary ones [22, 23]. Also, the information exchange must be accompanied by defining a control channel used for rendezvous of secondary users and their information exchanges. Based on the amount of the shared information, the cooperative detection strategies can be further classified as [24, 25]: • •

Partial cooperation approaches and Total cooperation approaches.

Partial cooperative detection approaches are also referred as centralized (controlled) spectrum sensing approaches, whereas the total cooperation approaches are viewed as decentralized (uncontrolled) ones. 3.1 Partial Cooperation The partial cooperation approaches refer to a scenario where the secondary users detect the primary channel by using some of the techniques elaborated in the subsection 2.1 (usually energy detection) either independently or with the aid of some local cooperation with nearby secondary users. The detection information is then sent to a common controller which is also a secondary user (sometimes named as spectrum broker or a fusion center). The common controller is responsible to decide upon the spectrum availability for secondary users’ transmissions. There are numerous examples found in the literature that deal with the partial cooperation approaches to spectrum sensing and various enhancements in terms of finding the optimal local secondary node information to be collected and optimal decision making at the common controller side. They usually differ according to the implemented mechanism for data processing in the common controller which may be based on: • •

voting or various statistical combinations of the gathered data.

Voting schemes, e.g. [26, 27], perform decision making upon the collected spectrum occupancy decision from every secondary user. Ref. [26] elaborates the cluster-collectforward scheme based on secondary users’ own confidence, i.e. the common controller collects information about the sensed spectrum only when the secondary users are confident about their sensing results. This scheme provides 65% to 95% transmission energy saving compared to traditional broadcasting schemes. Ref. [27] proves that the optimal fusion role at the fusion center is the half-voting rule if energy detection is used by the secondary users locally. If all secondary users have identical energy detectors and the received signals are modeled as correlated log-normal random variables, then a Linear-Quadratic (LQ) fusion strategy based on a deflection criterion that takes into account the correlation among the nodes proves to significantly outperform other fusion strategies under the mentioned assumptions [28]. Instead of voting, another approach to optimal partial cooperation strategy is to make various statistical combinations of the gathered data from the secondary users. Ref. [29] shows a linear combination of local test statistics from individual secondary users at a fusion center (i.e. the common controller) method. The result is to either optimize the probability distribution function of the global test statistics or maximize

Cooperative Spectrum Sensing in Ad-Hoc Networks

151

the global detection sensitivity under constraints on false alarm probability [29]. Furthermore, [30] shows an approach where a Maximal Ratio Combining (MRC) and Equal Gain Combination (EGC) is being used at the fusion center as they are able to provide close to optimal solutions in low SNR regions (which is a common scenario in the context of cognitive radio) over the hard combination technique. Therefore, [30] introduces a new softened hard combination scheme with two-bit overhead for each user that achieves a good tradeoff between detection performance and complexity. The collected information by the common controller under partial cooperation must be robust against Byzantine failures which require specific data fusion techniques. Most of the existing data fusion techniques rely on using a fixed number of samples, but there are also techniques that use a variable number of samples [31]. The spectrum sensing capabilities of a cognitive radio network employing partial cooperation detection can be enhanced by exploiting spatial diversity [32] in multiuser networks and providing either fixed or variable relay sensing schemes. The spatial diversity is especially important allowing higher confidence of the decision making process since the local node decisions can better extract the spectrum occupancy information due to their physical separation and the fading feature of the wireless ad-hoc environment. This also minimizes the probability of misconceptions in spectrum sensing as the physical separation of the nodes exhibits different viewpoints on the wireless medium conditions. Additionally, the average detection time can be reduced. 3.2 Total Cooperation The total cooperation approaches to spectrum sensing refer to a scenario where all secondary users operate in an ad-hoc manner using optimal transmission parameters. This means that the secondary users cooperatively sense the spectrum in order to reduce the detection time of spectrum holes and increase the agility of the secondary users. The coordination among the secondary users in this case aids the control of the uncertainty, that limits the ability of a cognitive radio network to reclaim a band or not, which is actually caused by the presence/absence of secondary users. It can be shown that the degree of coordination among the secondary nodes in total cooperation approaches can vary based on the coherence times and bandwidths involved, as well as the complexity of the detectors themselves [33]. There are several total cooperation approaches to spectrum sensing found in the literature. Due to their versatile nature, it is not easy to provide a unified classification. However, all of them usually employ: • •

relaying schemes [34] or various mathematical transformations [35] of the received data.

For example, ref. [34] uses relaying based on the Amplify-and-Forward (AF) cooperation protocol in order to reduce the detection time. In [35], multiple secondary users are used to infer on the structure of the received signals using Random Matrix Theory (RMT). The secondary users share information among them making the

152

L. Gavrilovska and V. Atanasovski

scheme not dependable on the knowledge of the noise statistics or its variance, but relying on the behavior of the largest and the smallest eigenvalue of random matrices. Further on, the benefits of total cooperation approach for a simple two user cooperative cognitive network is elaborated in [36]. The improvement in agility is shown by exploiting the inherent asymmetry in the network. The same authors extended their work on total cooperation approaches to spectrum sensing in [37] to account for a multiuser single carrier network. They have found the sufficient conditions under which asymptotic agility gain is achievable and developed a pairing protocol that ensures asymptotic agility gain with probability equal to 1. The authors in [38] show that the total cooperation approach can increase the throughput of the secondary users while limiting the interference to the primary users.

4 Practical Implementation Example After elaborating on the possible spectrum sensing techniques and cooperation strategies, this section reflects on a practical testbed implementation of a cooperative behavior among cognitive radios in an ad-hoc network [39]. As all cooperation strategies require a common control channel for information exchange among the cognitive radios, the example will rely on the usage of a novel rendezvous protocol, named RAC2E, that allows complete asynchronous behavior of the cognitive radios in the ad-hoc environment. 4.1 Scenario Setup and Protocol Description The targeted scenario comprises a secondary network of CSMA/CA based cognitive radio nodes with spectrum sensing capabilities (energy detection based) that operates in the 2.4 GHz ISM band and coexists with a primary IEEE 802.11 based network. All cognitive radios are able to create a local spectrum map (a top-down power ranking of the available channels). If two cognitive radios want to establish direct link communication by opportunistically using the temporary unused channels from the primary network terminals, they should exchange their spectrum maps in order to select the mutually best channel (i.e. the channel having the lowest level of interference for both nodes). Afterwards, the initiator node should request connection and the destination node should confirm the requested connection allowing the data channel to be established. All these control messages should be exchanged through on demand dynamically established control channel and should prove that the cooperation among the cognitive radios will yield more efficient network operation. The control channel is mutual for all simultaneous active secondary nodes and serves for cooperative spectrum sensing info exchanges as well as for communication parameters negotiation. When a new secondary node becomes active it first searches for control channel where it can obtain the spectrum maps from other nodes and find some node for communication. The cooperative rendezvous protocol in use (RAC2E) has two phases, i.e. initialization and exchanging control information phases. The former phase is used to select the control channel if such pre-exists in the secondary network from prior CR communication. In this case, the initialization phase requires that the node listens long

Cooperative Spectrum Sensing in Ad-Hoc Networks

153

enough on each channel frequency in order to detect any control message. However, if a control channel is not detected, then the node selects the best channel from its spectrum map and declares it to be the control channel. The second phase is used for cooperative exchange of control information (sensing reports exchange) on the selected control channel. It should be stressed that RAC2E operates upon an application request, i.e. when two cognitive radios want to establish a data session between. The second phase of the protocol, after detection of the control channel, operates in a time division mode, i.e. the node spends Ts seconds for spectrum sensing and creating a spectrum map and Tc seconds for sending and listening on the control channel. The total period of (Ts + Tc) seconds is continuously repeated until the communication link on an appropriate data channel between the two cognitive radios that want to communicate is established. The duration of the spectrum sensing intervals Ts should be fixed to a long enough value to get accurate sensing information for all channels. The time duration of the control channel attendance Tc is fixed for each node and is randomly chosen from an interval [T1, T2], where the values for T1 and T2 can be fine tuned depending on the number of nodes. The value of Tc remains constant for each node until data communication is established (Fig. 1) and is changed for each subsequent control channel establishment procedure. The total asynchronies among users, stemming from the randomly chosen Tc periods, provides lower or higher overlapping of the different nodes control channel periods, depending on Tc and Ts values. It is obvious that the ends of the Tc periods will be overlapped most frequently (Fig. 1). Consequently, this imposes sending the control channel messages at the start and at the end of the control channel period in order to maximize the probability of packet reception from other nodes. In the period between the two control messages in one Tc interval, the nodes switch to a listening mode in order to detect other nodes’ potential control messages. If an already established control channel is being interrupted (e.g. primary user appearance), the control channel band must be changed. Since every node has other nodes’ spectrum maps, all nodes know which channel is mutually the best suitable to be the new control channel. This will result in faster control channel reestablishment. RAC2E envisions several control messages such as sensing report message, connection request message and connection reply message. These messages are sent twice from each node in a Tc period. The sensing report message carries sensing results obtained during the Ts period and is the most frequently sent cooperative control message. Upon reception of such message, a node stores it locally. Therefore, each node can keep information of other active nodes in the network and their spectrum maps. Moreover, before initiation of any communication, a node must first detect the sensing report message. When a node wants to communicate with another node, it checks its own spectrum map and the destination node’s spectrum map and selects mutually acceptable channel. Then, it sends a connection request message to the other node with all proposed connection parameters and waits for connection reply message. When affirmative connection reply is received, the pair of nodes is switched to a selected data channel and starts communication. In case of a negative reply, the node that initiated the communication continues to sense the spectrum, searches for spectrum maps and sends new connection requests.

154

L. Gavrilovska and V. Atanasovski

When an already established connection is finished, the nodes return to the previous mode when Ts seconds sense the spectrum and create spectrum maps and Tc seconds send their sensing maps and listen for other nodes sensing maps. If, for some reason, the control channel is switched to other band during the data communication, then the nodes start to search the control channel as in the initialization phase, but now first looking into the mutually best channel from its previous spectrum maps.

Fig. 1. RAC2E time operation

After elaborating the RAC2E operational details, the following section will give more insight into the performance behavior of the protocol and the cooperation process among the cognitive radios in the ad-hoc environment in general. 4.2 Performance Evaluation This section provides performance evaluation of cooperative spectrum sensing in adhoc environment demonstrating a testbed implementation of RAC2E on a platform with four USRP2 nodes [40]. Fig. 2 depicts the testing scenario. The application profile of the scenario is that node 1 wants to establish data connection (video streaming) with node 2, while simultaneously node 3 wants to communicate with node 4. The parameters used in the proposed USRP2 demo scenario are as follows. Ts is chosen to be 2.6s, the bit rate on the control channel is 200kbps (using gmsk modulation), Tc period length is chosen randomly in the interval [Ts, 2Ts] and the video streaming application being used is “VLC media player”. Once the control channel is established, the nodes start sending and collecting the spectrum maps on the mutually common control channel. The process of creating local spectrum map and exchanging spectrum maps among nodes is illustrated on Fig. 3 for node 1. After the exchange of the spectrum maps, node 1 tries to initiate the video streaming communication to node 2 by sending a connection request message

Cooperative Spectrum Sensing in Ad-Hoc Networks

155

through the control channel by proposing to the node 2 channel 6 as mutually the best channel. Fig. 4 shows the connection reply message from node 2 and the actual choice of the WiFi channel to be used for data communication. Similar conclusions and figures are valid for the other nodes (i.e. 3 and 4) that want to start other video streaming in the demo. Finally, Fig. 5 provides a snapshot from a spectral analyzer that was used to prove the actual cooperative sensing info exchange and to demonstrate the functionality of the proposed rendezvous approach. The detected channels in use from both pair of nodes are channel 6 and channel 8. Other spectrum emissions, besides the ones stemming from the demo’s video streaming applications, are caused by surrounding IEEE 802.11 access points at the laboratory premises where the demo was set up.

Fig. 2. RAC2E testbed platform

Fig. 3. Spectrum information at node 1

156

L. Gavrilovska and V. Atanasovski

Fig. 4. Data channel agreement, from node 1 viewpoint

Fig. 5. Spectrum analyzer snapshot of the established connections

This section gave a practical demonstration of a cooperative spectrum sensing in an ad-hoc environment. It showed how cognitive radios can dynamically establish a control channel and cooperatively exchange information in order to perform efficient spectrum sensing and spectrum management in general.

Cooperative Spectrum Sensing in Ad-Hoc Networks

157

5 Conclusions Ad-hoc networks are increasingly gaining momentum owing to their ability to provide access domain in the integral future 4G networking paradigm. On the other side, the development of different wireless technologies leads to the concept of cognitive radios that alter the viewpoint on the ad-hoc networking. Cognitive radios can autonomously optimize and adapt their transmission parameters according to the environmental scenario they operate in. For this purpose, cognitive radios are able to perform spectrum sensing allowing them a broader view on spectrum bands and vacant spectrum positions that may be opportunistically used. The introduction of cognitive radios in ad-hoc environments imposes novel research challenges. The spectrum sensed information must be cooperatively used in order to extract the global network context and enhance the overall network performance. There are different cooperation strategies that may be engaged by cognitive radios in ad-hoc environments. This paper gives an overview of the spectrum sensing techniques and the cooperation strategies among cognitive radios in ad-hoc networking context. It showed details and novel classifications of the plethora of solutions found in the literature today. Moreover, the paper reported on an ongoing research work on practical implementation of a testbed with USRP2 cognitive radios that cooperatively exchange spectrum maps and optimize the spectrum management process in the ISM band. Future work will be concentrated on implementation of different spectrum sensing techniques in the testbed (not only energy detection), testing in larger scenarios, dynamic control channel establishment protocol enhancements etc. Acknowledgments. Parts of this work were funded by the EC through the FP7 projects ARAGORN (216856) [41] and QUASAR (248303) [42]. The authors would like to thank everyone involved, especially Valentina Pavlovska and Daniel Denkovski for their valuable contributions.

References 1. Gavrilovska, L., Prasad, R.: Ad hoc networking Towards Seamless Communication. Springer, Heidelberg (2006) 2. Mitola III, J.: Cognitive Radio – An Integrated Agent Architecture for Software Defined Radio. PhD Thesis, KTH Royal Institute of Technology, Stockholm, Sweden (2000) 3. Simeone, O., Gambini, J., Spagnolini, U., Bar-Ness, Y.: Cooperation and cognitive radio. In: Proc. IEEE CogNet Workshop, 2007, Glasgow, Scotland, pp. 6511–6515 (June 2007) 4. Zayen, B., Hayar, A.: Cooperative Spectrum Sensing Technique Based on Sub Space Analysis for Cognitive Radio Networks. In: IEEE GLOBECOM 2007, Washington, USA (November 2007) 5. Unnikrishnan, J., Veeravalli, V.V.: Cooperative Spectrum Sensing and Detection for Cognitive Radio. In: IEEE GLOBECOM 2007, Washington, USA (November 2007) 6. Quan, Z., Cui, S., Sayed, A.H.: An Optimal Strategy for Cooperative Spectrum Sensing in Cognitive Radio Networks. In: IEEE GLOBECOM 2007, Washington, USA (November 2007)

158

L. Gavrilovska and V. Atanasovski

7. Biswasyz, A.R., et al.: Cooperative Shared Spectrum Sensing for Dynamic Cognitive Radio Networks. In: IEEE ICC 2009, Dresden, Germany (August 2009) 8. Atapattu, S., Tellambura, C., Jiang, H.: Relay Based Cooperative Spectrum Sensing in Cognitive Radio Networks. In: IEEE GLOBECOM 2009, Hawai, USA, November 30December 4 (2009) 9. Thanayankizil, L., Kailas, A.: Spectrum Sensing Techniques (II): Receiver Detection and Interference Management Report, http://aravind.kailas.googlepages.com/ece_8863_report.pdf 10. Shellhammer, S.J.: Spectrum Sensing in IEEE 802.22. In: 2008 IAPR Workshop on Cognitive Information Processing, Santorini, Greece (June 2008) 11. Urkowitz, H.: Energy detection of unknown deterministic signals. Proc. of IEEE, 523–531 (April 1967) 12. Sahai, A., Cabric, D.: Spectrum sensing: fundamental limits and practical challenges. In: IEEE DySPAN 2005, Baltimore, MD, USA ( November 2005) 13. Dandawate, A.V., Giannakis, G.B.: Statistical tests for presence of cyclostationarity. IEEE Transactions on Signal Processing 42(9), 2355–2369 (1994) 14. Wild, B., Ramchandran, K.: Detecting Primary Receivers for Cognitive Radio Applications. In: IEEE DySPAN 2005, Baltimore, MD, USA (November 2005) 15. Zeng, Y., Liang, Y.-C.: Eigenvalue based Spectrum Sensing Algorithms for Cognitive Radio. IEEE Transactions on Communications (accepted in February 2008) 16. IEEE P802.22 Wireless RANs, Eigenvalue based sensing algorithms (July 2006) 17. Neihart, N.M., Roy, S., Allstot, D.J.: A Parallel, Multi-Resolution Sensing Technique for Multiple Antenna Cognitive Radios. In: International Symposium on Circuits and Systems (ISCAS), New Orleans, LA, USA (May 2007) 18. Zhang, Q., Kokkeler, A.B.J., Smit, G.J.M.: An Efficient Multi-resolution Spectrum Sensing Method for Cognitive Radio. In: 3rd International Conference on Communications and Networking in China, Hangzhou, China (August 2008) 19. IEEE 802.22 Working Group. Information, http://grouper.ieee.org/groups/802/22 20. Clancy, T., Arbaugh, W.: Measuring interference temperature. Virginia Tech Symposium on Wireless Personal Communications (2006) 21. Choi, J.-p., Lee, W.-c.: Optimizing Coexistence System with Interference Temperature for Multi-user Environments. In: 23rd International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC 2008), Shimonoseki City, Japan (July 2008) 22. Hoang, A.T., Liang, Y.-C., Islam, Md.H.: Maximizing Throughput of Cognitive Radio Networks with Limited Primary Users’ Cooperation. In: IEEE ICC 2007, Glasgow, Scotland (June 2007) 23. Bakr, O., Wild, B., Johnson, M., Ramchandran, K.: A Multi-Antenna Framework for Spectrum Reuse Based on Primary-Secondary Cooperation. In: IEEE DySPAN 2008, Chicago, IL, USA (October 2008) 24. Akyildiz, I.F., Lee, W.-Y., Vuran, M.C., Mohanty, S.: NeXt generation/dynamic spectrum access/cognitive radio wireless networks: A survey. Computer Networks 50, 2127–2159 (2006) 25. Ganesan, G., Li, Y.: Agility Improvement through Cooperative Diversity in Cognitive Radio. In: IEEE GLOBECOM 2005, St. Louis, MO, USA, November 28-December 2 (2005) 26. Lee, C.-h., Wolf, W.: Energy Efficient Techniques for Cooperative Spectrum Sensing in Cognitive Radios. In: IEEE CCNC 2008, Las Vegas, Nevada, USA (January 2008)

Cooperative Spectrum Sensing in Ad-Hoc Networks

159

27. Zhang, W., Mallik, R.K., Letaief, K.B.: Cooperative Spectrum Sensing Optimization in Cognitive Radio Networks. In: IEEE ICC 2008, Beijing, China (May 2008) 28. Unnikrishnan, J., Veeravalli, V.V.: Cooperative Spectrum Sensing and Detection for Cognitive Radio. In: IEEE GLOBECOM 2007, Washington, DC, USA (November 2007) 29. Quan, Z., Cui, S., Sayed, A.H.: An Optimal Strategy for Cooperative Spectrum Sensing in Cognitive Radio Networks. In: IEEE GLOBECOM 2007, Washington, DC, USA (November 2007) 30. Ma, J., Li, Y.: Soft Combination and Detection for Cooperative Spectrum Sensing in Cognitive Radio Networks. In: IEEE GLOBECOM 2007, Washington, DC, USA (November 2007) 31. Chen, R., Park, J.-M., Bian, K.: Robust Distributed Spectrum Sensing in Cognitive Radio Networks. In: IEEE INFOCOM 2008, Phoenix, AZ, USA (April 2008) 32. Ganesan, G., Li, Y., Bing, B., Li, S.: Spatiotemporal Sensing in Cognitive Radio Networks. IEEE Journal on Selected Areas in Communications 26(1), 5–12 (2008) 33. Sahai, A., Tandra, R., Mishra, S.M., Hoven, N.: Fundamental Design Tradeoffs in Cognitive Radio Systems. In: TAPAS 2006, Boston, MA, USA (August 2006) 34. Laneman, J.N., Tse, D.N.C.: Cooperative diversity in wireless networks: efficient protocols and outage behavior. IEEE Transactions on Information Theory 50, 3062–3080 (2004) 35. Cardoso, L.S., Debbah, M., Bianchi, P., Najim, J.: Cooperative Spectrum Sensing Using Random Matrix Theory. In: IEEE ISWPC 2008, Santorini, Greece (May 2008) 36. Ganesan, G., Li, Y.: Cooperative Spectrum Sensing in Cognitive Radio, Part I: Two User Networks. IEEE Transactions on Wireless Communications 6(6), 2204–2213 (2007) 37. Ganesan, G., Li, Y.: Cooperative Spectrum Sensing in Cognitive Radio, Part II: Multiuser Networks. IEEE Transactions on Wireless Communications 6(6), 2214–2222 (2007) 38. Lee, K., Yener, A.: Throughput Enhancing Cooperative Spectrum Sensing Strategies for Cognitive Radios. In: 41st Annual Asilomar Conference on Signals, Systems, and Computers, Asilomar 2007, Pacific Grove, CA (November 2007) 39. Pavlovska, V., Denkovski, D., Atanasovski, V., Gavrilovska, L.: RAC2E: Novel Rendezvous Protocol for Asynchronous Cognitive Radios in Cooperative Environments. In: IEEE PIMRC 2010, Istanbul, Turkey (September 2010) (accepted) 40. Universal Software Radio Peripheral 2 (USRP2). Information, http://www.ettus.com 41. EC FP7 project ARAGORN. Information, http://www.ict-aragorn.eu 42. EC FP7 project QUASAR, Information, http://www.quasarspectrum.eu

Receiver Sensitivity in Opportunistic Cooperative Internet of Things (IoT) (Invited Paper) Vandana Milind Rohokale, Neeli Rashmi Prasad, and Ramjee Prasad Center for TeleInFrastuktur, Aalborg University, Niels Jernes Vej 10, 9220 Aalborg Ø, Denmark {vmr,np,prasad}@es.aau.dk

Abstract. In Cooperative communication, a source message is relayed through a locally connected network by means of cooperating network nodes. Recently, the cross layer cooperative schemes have been shown to offer multiple advantages over the single layer approaches. In distributed cooperation schemes, the cooperating nodes make transmission decisions based on the quality of the received signal, which is the only parameter available locally. Receiver sensitivity is the most important parameter of the physical layer and has a direct impact on the MAC layer. This paper proposes a novel cooperative approach for analysis of receiver sensitivity. Keywords: Cooperative communication, receiver sensitivity, distributed cooperation.

1 Introduction For monitoring and control of an area with negligible human involvement, wireless sensor network (WSN) is proving to be a hopeful policy. Due to inventions in the micro-electronic circuits, wireless communications and operating systems, WSNs have become a feasible platform which is being used in many applications. The coverage area of a WSN is the placement of sensor nodes in a service area in such a way that the complete service area gets covered [1]. The sensor devices can be placed in the service area either in a fixed or mobile fashion. When we employ random sensor placement, the service area should be well covered or monitored by the sensor devices. Thus, the coverage has been formulated in various ways. Paper [2] defines sensor coverage metric as a surveillance, which can be used as a measurement of the quality of service (QoS) provided by a certain sensor network. Dense deployment of disposable and low-cost sensor nodes makes the WSN concept beneficial for battlefields. The cooperative communication mechanism is more applicable to AdHoc wireless and WSNs as compared to the cellular networks. Here, each node acts as both a user (source) as well as a relay. In cooperative resource allocation, each node J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 160–167, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

Receiver Sensitivity in Opportunistic Cooperative Internet of Things (IoT)

161

transmits for multiple nodes. Effective QoS of the individual network nodes can be improved through cooperation. With cooperative communication, the transmission diversity is achieved by enabling a single antenna device from multi-user scenario to share their antennas and generate a virtual multiple antenna transmitter [3]. Three cooperative mechanisms are possible: 1. 2.

3.

Detect (Decode) and Forward: - each node transmits its own bits plus its partner’s bit with spreading codes. CDMA implementation is utilized here. Amplify and Forward: - firstly the noisy signal received from the neighboring node is amplified and then retransmitted. At high SNR values, for less number of nodes, it achieves almost full diversity. Coded cooperation: - the information received by the node is re-encoded and then retransmitted. It is inherently integrated into channel coding.

In cooperative communication, the information overheard by neighboring nodes is intelligently used to provide reliable communication between a source and the destination called a sink. In cooperative wireless communication (CWC), several nodes work together to form a virtual array. The overheard information by each neighboring node or relay is transmitted towards the sink concurrently. The cooperation from the wireless sensor nodes that otherwise do not directly contribute in the transmission is intelligently utilized in CWC. The sink node or destination receives numerous editions of the message from the source, and relay(s) and it estimates these inputs to obtain the transmitted data reliably with higher data rates [4]. Opportunistic Large Array (OLA) is a cluster of network nodes which use an active scattering mechanism in response to the signal of the source called leader. The intermediate nodes opportunistically relay the messages from the leader to the sink. The advantage of the OLA is that due to signal enhancement, the SNR of cooperative transmission (SNRCT) is much higher than the SNR of a point to point communication (SNRP2P.). OLAs are considerably flexible and scalable in nature. For cooperative transmission, OLA selects the nodes which have the received signal SNR above some threshold value and the resonance generated by relay nodes carries the actual messages to the desired sinks without causing interference. For the elimination of the routing and multiple access overheads, OLA is a competent physical layer broadcasting algorithm. Multiple clusters of ad hoc wireless nodes can form a multi-OLA system, constructing a multiple access system with the cluster of nodes acting as a team through cooperative transmission rather than transmitting independent data from each node. OLA utilizes the cooperative transmission of the AdHoc network nodes to reach back to a far distant node or sink. Applications of OLA can be: Joint control systems and secure military scenarios where delay cannot be tolerated. OLA is a cooperative mechanism which is simple and scalable [5]. For mobile users, the restrictions in the data rate and QoS are due to variations in the signal attenuation, so some kind of signal multiplicity is very much essential. Various authors have proposed a new form of spatial diversity for the cooperative communication. The results reveal that although the communication channel is noisy, cooperation leads to an increase of capacity of the individual user as well as the

162

V.M. Rohokale, N.R. Prasad, and R. Prasad

overall communication system. Also the user’s data rates are less prone to channel variations. Some of the important types of diversity are spatial diversity, temporal diversity and frequency diversity. Higher data rates at reduced transmit power can be converted to an increase in cell coverage [6].

2 Proposed Cooperative IoT Model The strength of the cooperative OLA approach is that it does not require GPS equipment for identification of the location of the network node entities. The energy savings achieved in WSNs are the result of cross-layer interactive cooperative communication. Routing functions are partially executed in the physical layer as shown in Fig. 1. The diversity provided by MIMO space-time codes can improve performance at the MAC, network and transport layers. Physical layer parameters such as receiver sensitivity significantly affect the MAC protocol. Choice of the medium access scheme is the important aspect of WSNs [7].

Fig. 1. Cooperative Cross layered Communication

In order to maximize the information transfer among network nodes, the optimal receiver sensitivity is the prime requirement. The noise floor of a receiver determines its sensitivity to low-level signals and its capability of detecting and demodulating those signals. Cooperation allows independently faded radios to collectively achieve robustness to severe fades while keeping individual sensitivity levels close to the nominal path loss. Furthermore, a small number of radios (10-20) are enough to achieve practical sensitivity levels [8]. The proposed cooperative IoT model is depicted in Fig. 2. The nodes which are participating in perticular communication are shown with solid fill.

Receiver Sensitivity in Opportunistic Cooperative Internet of Things (IoT)

163

Fig. 2. Proposed cooperative IoT model

3 Analytical System Model The Decode and Forward mechanism is taken into account for the analytical system model. The wireless sensor nodes with density ρ are uniformly and randomly distributed. The proposed analytical system model is shown in Fig. 3. Radius RT is the total radius and r is the radius till perticular network node.

Fig. 3. Proposed Model for Numerical Analysis

Theorem: If μ

[8] and μ 1

Then

2, )

(1)

where = source power Pr = relay node power and lim

=

For (µ ≤ 2), the broadcast reaches to the whole network i.e. For (µ > 2), the total area reached by the broadcast is limited i.e.

(2) ∞

∞. .

164

V.M. Rohokale, N.R. Prasad, and R. Prasad

Instead of infinite radius, we are considering some practical scenarios where the radius is limited. Wireless LAN, Bluetooth, etc. are some of the examples for which we can apply limited radius concept. Table 1. Protocol comparison with different parameters

Protocol/ Parameter

IEEE Spec

Freq. Band

Max Signal Rate

Nominal Range

Bluetooth

802.15.1

2.4 GHz

1 Mbps

10m

UWB

802.15.3a

3.1-10.6 GHz

110 Mbps

10m

Zigbee

802.15.4

868/915 MHz, 2.4 GHz

250 Kbps

10-100m

Wi-Fi

802.11 a/b/g

2.4 GHz, 5GHz

54 Mbps

100m

(3) Since we have assumed the noise variance to be unity, the received power becomes received signal to noise ratio (SNR). SNR= Prx = =

(4)

The fraction of energy saving is derived based on how many active nodes are utilized for particular cooperative transmission out of total available network nodes in the OLA structure. FES

1

(5)

The receiver sensitivity is given by, SR = K (Ta+Trx) B (SNR) where

(6)

SR = receiver sensitivity K= Boltzman’s constant Ta = equivalent noise temperature in [k] of the source e.g. antenna at the input of the receiver Trx = equivalent noise temperature in [k] of the receiver referred to the input of the receiver B = bandwidth SNR = required SNR at the output

Receiver Sensitivity in Opportunistic Cooperative Internet of Things (IoT)

165

4 Simulation Results The sensor coverage metric surveillance is used as a measurement of the QoS provided by a certain sensor network [2]. From the plot of Fig. 4, it is observed that for high QoS values, the network node requirement is high. But for low threshold values like Lambda=0.4, the maximum number of node requirement is reduced to 600 for the range of 40 meters.

Fig. 4. Nmax versus QoS

As seen from the plot of Fig. 5, for moderate QoS values, considerable energy savings are observed. For decoding threshold Lambda=1.5, the energy savings of 75% is observed for the range of 40 meters. Fig. 6 indicates that for high radius values, the sensitivity is considerably decreased. It indicates better sensitivity of the receiver for distant nodes. Lower power for a given SNR means better sensitivity. Radio devices that fail in unknown ways or may be malicious, introduce a bound on achievable sensitivity reductions [9]. The energy savings are considered on the basis of minimum number of nodes participating in a particular communication purpose out of the total deployed network nodes. As seen from Fig. 6, for small radius, the sensitivity achieved is around -90 dBm, but for the higher values of radius, sensitivity reaches up to -62dBm.

166

V.M. Rohokale, N.R. Prasad, and R. Prasad

Fig. 5. FES versus QoS

Fig. 6. Sensitivity versus radius

5 Conclusions In this paper, the cooperative behavior of WSNs is analyzed. The observed high QoS values indicate that our proposed cooperative IoT model is highly reliable. For decoding threshold values of Lambda=1.5, the fraction of energy savings obtained is almost 75%. Since receive sensitivity indicates how faint an input signal can be successfully received by the receiver, the lower the power level, the better is the receiver. Better sensitivity figures are obtained for radius of up to 20 meters. For e.g., sensitivity achieved for 10 meters radius it is around -65dBm. In future works, these results will be extended for the nodes which are malicious or which fail in unknown ways. This work will be extended for security aspects of cooperative IoT systems.

Receiver Sensitivity in Opportunistic Cooperative Internet of Things (IoT)

167

References 1. Habib, S., Safar, M.: Sensitivity Study of Sensor’s Coverage within Wireless Sensor Networks Computer Communications and Networks. In: ICCCN 2007 (2007) 10.1109/ICCCN.2007.4317928 2. Meguerdichian, S., Koushanfar, F., Potkonjak, M., Srivastava, M.: Coverage problems in wireless ad-hoc sensor networks. In: IEEE INFOCOM, Alaska, USA (2001) 3. Nosratinia, A., University of Texas, Dallas, Hunter, T.E.: Cooperative Communication in Wireless Networks. IEEE Communications Magazine (October 2004) (Adaptive Antennas and MIMO Systems for Wireless Communications) 4. Liu, P., Tao, Z., Lin, Z., Erkip, E., Panwa, S.: Cooperative Wireless Communications: A Cross-Layer Approach. IEEE Wireless Communications (August 2006) 5. Sadek, A.K., Su, W., Ray Liu, K.J.: Multinode Cooperative Communications in Wireless Networks. IEEE Transactions on Signal Processing 55(1) (January 2007) 6. Sendonaris, A., Erkip, E., Aazhang, B.: User Cooperation – part i: System Description, part ii: Implmentation Aspects and Performance Analysis. IEEE Trans. Commun. 51(11), 1927– 1948 (2003) 7. Ferrari, G., Tonguz, O.K., Bhatt, M.: Impact of receiver sensitivity on the performance of sensor networks. In: IEEE 59th Vehicular Technology Conference, VTC 2004 (Spring 2004) 8. Mergen, B.S., Scaglione, A., Mergen, G.: Asymptotic analysis of multi-stage cooperative broadcast in wireless networks. In: 2004 International Conference on Acoustics, Speech and Signal Processing 9. Mishra, S.M., Sahai, A., Brodersen, R.W.: Cooperative Sensing among Cognitive Radios

Event Detection in Wireless Sensor Networks – Can Fuzzy Values Be Accurate? Krasimira Kapitanova1, Sang H. Son1 , and Kyoung-Don Kang2 1

University of Virginia, Charlottesville VA, USA [email protected], [email protected] 2 Binghamton University, Binghamton, NY, USA [email protected]

Abstract. Event detection is a central component in numerous wireless sensor network (WSN) applications. In spite of this, the area of event description has not received enough attention. The majority of current event description approaches rely on using precise values to specify event thresholds. However, we believe that crisp values cannot adequately handle the often imprecise sensor readings. In this paper we demonstrate that using fuzzy values instead of crisp ones significantly improves the accuracy of event detection. We also show that our fuzzy logic approach provides higher detection precision than a couple of well established classification algorithms. A disadvantage of using fuzzy logic is the exponentially growing size of the rule-base. Sensor nodes have limited memory and storing large rule-bases could be a challenge. To address this issue we have developed a number of techniques that help reduce the size of the rule-base by more than 70% while preserving the level of event detection accuracy. Keywords: wireless sensor networks, fuzzy logic, event description, event detection accuracy.

1

Introduction

Event detection is one of the main components in numerous wireless sensor network (WSN) applications. WSNs for military application are deployed to detect the invasion of enemy forces, health monitoring sensor networks are deployed to detect abnormal patient behavior, fire detection sensor networks are deployed to set an alarm if a fire starts somewhere in the monitored area. Regardless of the specific application, the network should be able to detect if particular events of interest, such as fire, have occurred or are about to. But just like many other human-recognizable events, the phenomenon fire has no real meaning to a sensor node. Therefore, we need suitable techniques that would allow us to describe events in ways that sensor nodes would be able to “understand”. The area of event description in WSNs, however, has not been explored much. 

This research work was supported by KOSEF WCU Project R33-2009-000-10110-0.

J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 168–184, 2010. Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

169

Most previous work on event description in WSNs uses precise, also called crisp, values to specify the parameters that characterize an event. For example, we might want to know if the temperature drops below 5 or the humidity goes above 46%. However, sensor readings are not always precise. In addition, different sensors, even if located close to each other, often vary in the values they register. Consider an example scenario where we want the cooling in a room to be turned on if the temperature goes above 5 . Two sensors, A and B, measure the temperature in the room and the average of their values is used to determine if an action should be taken. At some point, sensor A reports 5.1 and sensor B reports 4.8 . The average, 4.95 , is below our predefined threshold and the cooling remains off. However, if sensor B ’s measurement is inaccurate and therefore lower than the actual temperature, we have made the wrong decision which can be classified as a false negative. The situation becomes even more convoluted when more than two sensor measurements are involved. This makes determining the precise event thresholds an extremely hard task which has led us to believe that using crisp values to describe WSN events is not the most suitable approach. Fuzzy logic, on the other hand, might be able to better address the problems that are challenging for crisp logic. Fuzzy logic has a number of properties that make it suitable for describing WSN events: (i) it can tolerate the unreliable and imprecise sensor readings; (ii) it is much closer to our way of thinking than crisp logic. For example, we think of fire as an event described by high temperature and smoke rather than an event characterized by temperature above 55 and smoke obscuration level above 15%; (iii) compared to other classification algorithms based on probability theory, fuzzy logic is much more intuitive and easier to use. A disadvantage of using fuzzy logic is that storing the rule-base might require a significant amount of memory. The number of rules grows exponentially to the number of variables. With n variables each of which can take m values, the number of rules in the rule-base is mn . Adding spatial and temporal semantics to the decision process further increases the number of rules. Since sensor nodes have limited memory, storing a full rule-base on every node might not be reasonable. In addition, constantly traversing a large rule-base might considerably slow down the event detection. To address this problem, we have designed a number of techniques that reduce the size of the rule-base. A key property of these techniques is that they do not affect the event detection accuracy. This paper has three main contributions. First, we show that using fuzzy logic results in more accurate event detection than when either crisp values or well established classification algorithms, such as Naive Bayes classifiers or decision trees, are used. Second, we incorporate event semantics into the fuzzy logic rulebase to further improve the accuracy of event detection. Third, we have designed techniques that can be used to prevent the exponential growth of the rule-base without significantly compromising the accuracy of event detection.

170

2

K. Kapitanova, S.H. Son, and K.-D. Kang

Related Work

Event detection: Not much research has focused directly on providing methods for event description in WSNs that can support data dependency and collaborative decision making. The prevailing approach is to use SQL-like primitives [1,2,3,4]. The papers that employ this method vary in semantics. In [1] and [2], the authors use general SQL primitives to define events in sensor networks. The limitation of this approach is that the events can only be defined by predicates on sensor readings with very simple temporal and spatial constraints connected by AND and OR operators. Madden et al. have extended the SQL primitives by incorporating streaming support where a desired sample rate can be included [4]. Li et al. define events using a sub-event list and confidence functions in SQL [3]. However, SQL is not very appropriate for describing WSN events. Some of its drawbacks include that it: (i) cannot capture data dependencies and interactions among different events or sensor types; (ii) does not explicitly support probability models; (iii) is awkward in describing complex temporal constraints and data dependencies; (iv) lacks the ability to support collaborative decision making and triggers [5]; (v) does not support analysis of the event system. Another approach to formally describe events in WSNs has been the use of extended Petri nets. This was initially proposed by Jiao et al. [6]. The authors design a Sensor Network Event Description Language (SNEDL) which can be used to design Petri nets that specify event logic. Petri nets were also used in MEDAL [7], an extension of SNEDL that supports the description of additional WSN specific features such as communication and actuation. Both SNEDL and MEDAL, however, use crisp values in the definitions of their Petri nets. Stochastic methods: There is a long history of using stochastic formalisms in different WSN applications. Bayesian classifiers and Hidden Markov Models have been extensively used in activity recognition [8,9] and decision fusion [10,11]. Dempster-Shafer evidence theory has been applied to intrusion detection [12], sensor fusion [13,14], and assisted living applications [15]. Probabilistic context free grammars have been used to solve problems such as inferring behaviors [16] as well as movement and activity monitoring [17,18]. Fuzzy logic: Fuzzy sets and logic were introduced by L. Zadeh in 1965. Ever since then, numerous fields have taken advantage of their properties. In WSNs, fuzzy logic has been used to improve decision-making, reduce resource consumption, and increase performance. Some of the areas it has been applied to are cluster-head election [19,20], security [21,22], data aggregation [23], routing [24,25], MAC protocols [26], and QoS [27,28]. However, not much work has been done on using fuzzy logic for event description and detection. Liang et al. [29] propose to use fuzzy logic in combination with Double Sliding Window Detection, to improve the accuracy of event detection. However, they do not study the effect of fuzzy logic alone or the influence of spatial or temporal properties of the data on the classification accuracy.

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

171

In D-FLER [30] fuzzy logic is used to combine personal and neighbors’ observations and determine if an event has occurred. Their results show that fuzzy logic improves the precision of event detection. The use of fuzzy values allows D-FLER to distinguish between real fire data and nuisance tests. However, the approach used in D-FLER does not incorporate any temporal semantics. In addition, since all of the experiments last only 60 seconds after the fire ignition, the authors do not analyze the number of false alarms raised by D-FLER.

3

Overview of Fuzzy Logic

The structure of a general fuzzy logic system (FLS) is shown on Figure 1. First, the fuzzifier converts the crisp input variables x ∈ X, where X is the set of possible input variables, to fuzzy linguistic variables by applying the corresponding membership functions. Zadeh defines linguistic variables as “variables whose values are not numbers but words or sentences in a natural or artificial language” [31]. An input variable can be associated with one or more fuzzy sets depending on the calculated membership degrees. For example, a temperature value can be classified as both Cold and Warm. Second, the fuzzified values are processed by if-then statements according to a set of predefined rules derived from domain knowledge provided by experts. In this stage the inference scheme maps input fuzzy sets to output fuzzy sets. Finally, the defuzzifier computes a crisp result from the fuzzy sets output by the rules. The crisp output value represents the control actions that should be taken. The above three steps are called fuzzification, decision making, and defuzzification, respectively. We describe them in more detail in the following subsections.

Fig. 1. The structure of a fuzzy logic system

3.1

Fuzzification

The fuzzifier converts a crisp value into degrees of membership by applying the corresponding membership functions. A membership function determines the certainty with which a crisp value is associated with a specific linguistic value. Figure 2 shows an example of a temperature membership function. According to this membership function a temperature of -2 is classified as 20% Freezing and 80% Cold. The membership functions can have different shapes. Some of the most frequently used shapes include triangular, trapezoidal, and Gaussianshaped.

172

K. Kapitanova, S.H. Son, and K.-D. Kang

Fig. 2. Temperature membership function

3.2

Decision Making

A rule-base consists of a set of linguistic statements, called rules. These rules are of the form IF premise, THEN consequent where the premise is composed of fuzzy input variables connected by logical functions (e.g. AND, OR, NOT) and the consequent is a fuzzy output variable. Consider a t-input 1-output FLS with rules of the form: Ri :IF x1 is S1i and x2 is S2i and ... and xt is Sti THEN y is Ai 







When input x = {x1 , x2 , ..., xt } is applied, the degree of firing of some rule Ri can be computed as:         t μS i x1 ∗ μS i x2 ∗ ... ∗ μSti xt = Tl=1 μS i xl 1

2

l

Here μ represents the membership function and both ∗ and T indicate the chosen triangular norm. A triangular norm is a binary operation such as AND or OR applied to the fuzzy sets provided by the membership functions [32]. 3.3

Defuzzification

Executing the rules in the rule-base generates multiple shapes representing the modified membership functions. For example, rules designed to decide the probability of having a fire in a building may produce the following result: Low (56%), Medium (31%), and High (13%). Defuzzification is the transformation of this set of percentages into a single crisp value. Based on how they perform this transformation, defuzzifiers are divided into a number of categories. The most commonly used defuzzifiers are center of gravity, center of singleton, and maximum methods [32].

4

Event Semantics

Sensors are generally believed to be unreliable and imprecise. Therefore, to increase our confidence in the presence of an event somewhere in the monitored area, we often need readings from multiple sensors and/or readings over some period of time. This could be achieved by instrumenting the event description logic with temporal and spatial semantics. We believe that this can significantly

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

173

Table 1. An example fire detection rule-base Rule # 1 2 3 4 .. . 729

T1 ΔT1 T2 ΔT2 S ΔS Confidence L L L L L L L L L L L L M L L L L L L H L L L L L M L L .. .. .. .. .. .. .. . . . . . . . H H H H H H H

decrease the number of false positives. It will also allow us to describe and detect more complex events. To the best of our knowledge, no previous work on applying fuzzy logic to event detection has considered the effects of temporal and spatial semantics on the accuracy of event detection. Consider, for example, a fire detecting scenario. A sensor network is deployed to monitor a building and trigger an alarm if a fire starts. There are a number of temperature and smoke sensors in each room, as well as in the hallways. The monitoring of the building is divided into floors and there is a master node on each floor. The rest of the sensor nodes on the floor send their readings to the master node and, based on these readings, it determines if there is a fire or not. The fire detection is based not only on the temperature and smoke obscuration readings for a particular moment in time but also on the rate of change of both the temperature and smoke levels. Therefore, our fire detection logic takes four linguistic variables as input - temperature (T), temperature change (ΔT), smoke obscuration(S), and smoke obscuration change (ΔS). The linguistic values for all four variables can be classified as Low (L), Medium (M), and High (H). In order to increase the accuracy of the fire detection scheme, we require that at least two temperature readings and one smoke reading are used to make a decision. Table 1 shows an example rule-base for this fire detection scenario. This rulebase, however, introduces a number of concerns which we address in the rest of this section. 4.1

Spatial Semantics

One of the main goals when designing an event detection system is that the system is accurate and the number of false alarms is low. A way to achieve this is to include readings from multiple sensors in the decision process. For instance, we would be more confident that there is an actual fire if more than one node is reporting high temperature and smoke readings. If, for example, three sensors from the same room send reports indicating fire, the probability that there is an actual fire in that room is very high. In general, there is a negative correlation between the distance among the sensors reporting fire and the probability of this report being true. Therefore, we need to include the concept of location in the event detection logic. This can be achieved by augmenting the rules in the rule-base with a linguistic variable that would serve as a spatial guard. This variable would express the application requirements about the distance between

174

K. Kapitanova, S.H. Son, and K.-D. Kang

the reporting sensors. For our fire detection scenario, we could call this variable distance and classify it as Close (C), Distant (D), and Far (F), for example. Incorporating the distance variable into the rule-base adds an extra column for distance which changes the format of the rules. Now we have rules such as: IF T1 is H and ΔT1 is H and T2 is H and ΔT2 is H and S is H and ΔS is H and distance is F, THEN Fire is M.

4.2

Temporal Semantics

To further decrease the number of false alarms we also need to take into account the temporal properties of the monitored events. The confidence of event detection is higher if the temporal distance between the sensor readings is shorter and vice versa. Adding temporal semantics is especially important for WSNs because of the nature of sensor communication. It is very possible for messages in a WSN to be delayed because of network congestions or bad routing. Consequently, a reliable event detection rule-base should take into consideration the generation time of the sensor readings. To accommodate this, we include another linguistic variable that serves as a temporal guard. This variable, time, represents the difference in the generation times of the sensor readings. For example, in our fire detection scenario, time could have three semantic values - Short (S), Medium (M), and Long (L). In this way the information about the period within which the sensor readings have been generated is included in the decision process.

5

Decreasing the Size of the Rule-Base

Augmenting the rule-base with temporal and spatial variables increases the number of rules. As mentioned earlier, the size of the rule-base grows exponentially to the number of linguistic variables. In our fire monitoring example, where the only sensor readings we consider are temperature and smoke, the extended rulebase will have 6561 rules. In more complicated scenarios that require more than two types of sensors, the number of rules in the fuzzy rule-base could be much higher. Storing such rule-bases might be a challenge for the memory constrained sensor nodes. In addition, traversing the full rule-base every time there are new sensor readings will slow down the event detection. To address these concerns, we have designed three techniques to help us reduce the number of rules. Although such rule-base reduction techniques alleviate both the storage problem and the rule traversal process, they could also introduce a tradeoff between the time and space cost of the rule-base and the event detection accuracy. Therefore, maintaining high event detection accuracy was a key goal when designing the reduction techniques described in this section. 5.1

Separating the Rule-Base

The first thing we could do to reduce the size of the rule-base is to separate the rules on a “need to know” basis. Each node stores only the rules corresponding to

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

175

Table 2. Rule-base for a temperature sensor Rule # T ΔT Confidence 1 L L L 2 L M L 3 L H M 4 M L L 5 M M M 6 M H H 7 H L M 8 H M H 9 H H H

the types of sensors it has. If, for example, some of the nodes in our fire detection scenario are only equipped with temperature sensors, they do not need to store the whole rule-base. Instead, they store a smaller modified rule-base similar to the one shown in Table 2. This rule-base contains only rules with premise linguistic variables based on the values from the temperature sensors. In this way the event detection logic on each node only considers rules that are relevant to that node’s sensor readings. This separation simplifies the decision process and makes the rule-base traversal faster. The smoke sensors’ rule-base can be constructed in a similar way. 5.2

Combining Rules with Similar Outcomes

Rules 1 and 2 in Table 2 have the same outcome and only differ in the values of ΔT . This observation is also valid for rules 8 and 9. Combining these rule couples could help us further decrease the size of the rule-base. For the rulebase on Table 2 applying such an optimization leaves us with 7 rules. The rules, however, have a slightly different syntax. Instead of: Ri : IF x1 is S1i and x2 is S2i and ... and xt is Sti THEN y is Ai

some of the rules have the following different form: Ri :IF x1 is≤S1i and x2 is S2i and ... and xt is≥Sti THEN y is Ai

In the modified rules ≤ stands for “in this fuzzy set or in fuzzy sets smaller than it” and ≥ stands for “in this fuzzy set or in fuzzy sets greater that it”. Table 3 shows the result of applying this reduction technique on the rule-base in Table 2. 5.3

Incomplete Rule-Base

A rule-base is considered complete if there are rules for every possible combination of the input variables. However, only some of these combinations have outcomes that are important to the event detection system. For example, only sensor readings that satisfy the temporal and spatial constraints can satisfy rules that could trigger an alarm. Therefore, the rules with distance variable Distant

176

K. Kapitanova, S.H. Son, and K.-D. Kang Table 3. Reduced rule-base for a temperature sensor Rule # T ΔT Confidence 1 L ≤M L 2 L H M 3 M L L 4 M M M 5 M H H 6 H L M 7 H ≥M H

or Far can be removed from the rule-base. This step leaves us with just a third of the original number of rules in the rule-base. Similarly, applying the same approach to the time variable and removing the rules with values Medium and Long decreases the rule-base by yet another two thirds. In addition, if we exclude the rules with consequents that are of no interest to the event detection system, e.g. rules that indicate that there is no fire, we will reduce the size of the rule-base even more. As a result, by lowering the level of completeness of the rule-base, we significantly decrease the number of rules that should be stored by the sensor nodes. This “trimming” process, however, should be performed very carefully in order to prevent the removal of important consequents. To make sure that the system knows how to proceed if none of the rules in the rule-base has been satisfied, we introduce a default rule that is triggered if no other rule has been satisfied.

6 6.1

Evaluation Experiments

We used the FuzzyJ Toolkit for Java [33] to implement the necessary fuzzy logic functionality. Trace-based simulations were performed in order to avoid the danger, cost, and non-repeatability of creating fires. For our experiments we have used real fire data publicly available on the National Institute of Standards and Technology (NIST) website [34]. The study they performed provides sensor measurements from a number of different real fires. We have used two of the available scenarios: fire caused by a burning mattress and fire caused by a burning chair. The membership functions for the input linguistic variables used in the experiments are shown in Figure 3. In addition to the temperature and smoke obscuration variables, we also take into consideration the temperature and smoke obscuration difference between two consecutive readings. These two additional variables give us a notion of how fast the temperature and smoke obscuration are changing. Figure 4 shows the membership function for the output fire confidence. This linguistic variable represents the system’s confidence in the presence of fire. For example, if the fire confidence value is higher than 80, we are more than 80% certain that there is a fire. If the fire confidence is smaller than 50, the probability that there is no fire is higher.

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

177

Fig. 3. Membership functions for the input linguistic variables

Fig. 4. Fire confidence membership function

To provide a baseline for our results, we performed crisp-value experiments with both the burning mattress and burning chair data. The temperature and smoke obscuration thresholds used in the crisp logic experiments are threshold values used in commercial smoke and heat detectors, 55 and 0.15 m−1 , respectively [35,36]. The membership functions in Figure 3 were also built according to these threshold values. The results from the crisp-value experiment are shown in Figure 5 a) and b). In these and all following figures, the origin of the graph represents the time of fire ignition. As we can see from the two figures, using crisp values resulted in a very large number of false fire detections. In the burning mattress scenario in particular, there were 40 false fire detections in the period prior to the fire ignition, which constitutes about 1.3% of the measurements. This considerable number or false positives significantly affect the efficiency and fidelity of an event detection system. Admittedly, part of these false positives can be attributed to the aggressive crisp value thresholds. However, if the thresholds are set higher, this could lead to failures in detecting actual fires. Since for a real fire detection system it is more important to decrease the number of false negatives than that

178

K. Kapitanova, S.H. Son, and K.-D. Kang

Fig. 5. Crisp value simulation: a) burning mattress b) burning chair

Fig. 6. Fuzzy value simulation: a) burning mattress b) burning chair

of false positives, we have kept the threshold values in compliance with the commercial standards. What we wanted to investigate with our next set of experiments was whether fuzzy logic can do better in terms of false positives while still reporting promptly the presence of a fire when one actually occurs. In the first couple of fuzzy logic experiments a node decides if there is a fire based only on its own readings. The readings of neighboring sensor nodes are not considered as inputs to the decision process. The values of the linguistic variables used in the decision process can be classified as Low (L), Medium (M), and High (H), as shown by Figure 3 and Figure 4. We have used heuristics to build the rule-base for our fire detection experiments. In cases where this is not possible, e.g. when more complex events are to be detected, domain experts could be consulted for the definition of the rule-bases. The results from our first couple of fuzzy logic experiments, a burning mattress and a burning chair, are presented in Figure 6 a) and b), respectively. As we can see, the fuzzy logic event detection mechanism performs very well. It detects the presence of a fire shortly after the ignition. In addition, unlike the crisp-value fire detection, there are no false positives. Both graphs show fire confidence around 0 before the ignition, except for a number of small peaks when the confidence increases to 25. At the same times when the fuzzy-value peaks occur, we can also notice crisp-value peaks but with much higher confidence. The raw sensor data revealed that the peaks were caused by a number of one-second-long reports of increased smoke values. This proves our hypothesis that fuzzy logic is able to

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

179

Fig. 7. Simulating a burning mattress: including neighbor readings in the decision. The results when only own values are used are plotted on the first y-axis. Including the neighbor values is plotted on the second y-axis.

Fig. 8. Simulating a burning chair: including neighbor readings in the decision. The results when only own values are used are plotted on the first y-axis. Including the neighbor values is plotted on the second y-axis.

accommodate the often imprecise sensor readings. Even in the cases when the nodes erroneously report the presence of smoke, the fuzzy logic mechanism keeps the fire confidence low enough so that a false alarm is not triggered. We also studied how including neighbor node values in the decision process affects the detection accuracy. The average of the neighbor values is represented with an additional linguistic variable that we include in the decision rules. The results in Figure 7 and Figure 8 show that fire is detected almost as quickly as when the decision process is based on only own sensor readings. Although the peak areas are still present, the corresponding fire confidence values are lower when the neighbor readings are included in the decision process. This shows that including the readings of neighbor nodes in the decision process positively affects the detection accuracy. 6.2

Analysis

Why does fuzzy logic perform better?. An interesting question is why fuzzy logic is more precise than crisp-value logic. From the considerable decrease in the number of false positives, it appears that fuzzy logic handles the fluctuating sensor readings much better. To see why this happens we take a closer look at the first false fire detection reported by the crisp-value logic. In the burning mattress scenario this occurs approximately 12 minutes into the experiment. The values

180

K. Kapitanova, S.H. Son, and K.-D. Kang

Fig. 9. Simulating a burning chair with a reduced rule-base. The results when the full rule-base is used are plotted on the first y-axis. Using the reduced rule-base is plotted on the second y-axis.

that caused the false alarm are; T = 25.21 , ΔT = 0 , S = 0.203 %, and ΔS= 0.109%. Since the smoke level and ΔS are both classified as High, the crisp logic concludes that there must be a fire. What does the fuzzy logic event detection do differently? According to the membership functions in Figure 3, temperature value of 25.21 is classified as 100% Low; temperature change of 0 is classified as 100% Low; smoke obscuration level of 0.203% is classified as 33% Medium and 66% High; and smoke obscuration change of 0.109% is classified as 100% High. The decision making process checks which rules from the rule-base are satisfied, and the defuzzifier reports a fire confidence value of 22.5. This value maps to fire confidence which is 55% Low and 45% Medium. Since the Low confidence is higher, the logic determines that there is no need to report a fire. This example illustrates why a fuzzy logic event detection system tends to perform better than a crisp one in the presence of short-lasting inaccurate sensor readings, which often occur in WSNs. Fuzzy logic takes into account the certainty with which an event occurs, instead of only relying on exact values, which helps improve the accuracy of event detection. Decreasing the rule-base. We applied our reduction techniques to the rulebase used in the simulation experiments. All nodes in the simulation are equipped with both a smoke and a temperature sensor which makes the first technique not applicable. Therefore, we only used the other two reduction techniques. The rule base initially had 81 rules. Combining the rules with similar outcomes reduced the number of rules to 30, which is a 63% decrease. In general, when there are more than two input linguistic variables, applying the second method decreases the rule-base by approximately two thirds. Excluding the rules that result in Low fire confidence additionally reduced the size of the rule-base to 24, which is 30% of the original rule-base. We have compared the behavior of the fire detection system when the full and the reduced rule-bases are used. Figure 9 shows the results for the burning chair scenario. The fire confidence is consistently higher when the reduced rule-base is used. However, since this confidence remains low, this does not cause false fire

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

181

Table 4. Number of incorrect classifications by a Naive Bayes classifier and a J48 Tree Naive Bayes J48 Decision Tree Fuzzy logic number percent number percent number percent Burning chair 105 1.56% 7 0.13% 0 0 Burning mattress 89 2.35% 5 0.13% 0 0

Table 5. Fire detection delay in seconds Scenario Crisp values Fuzzy values Plus neighbor readings Reduced readings Burning chair 236 236 248 236 Burning mattress 103 97 117 97

detections. For future work, we plan to perform deeper analysis of the memory requirements associated with using fuzzy logic. Detection accuracy. To further understand the behavior of our fuzzy logic approach, we have compared it to two well established classification algorithms: a Naive Bayes classifier [37] and a J48 decision tree which is an open source implementation of the C4.5 algorithm [38]. Fuzzy logic is more suitable than these two algorithms for WSN event description since, unlike Bayes classifiers and decision trees where values are considered to be nominal, it works with continuous values, which is exactly what the sensor readings are. In addition, specifying the membership functions is more intuitive and simpler than building a probability model. We ran this set of experiments using the Weka data mining tool [39]. The input values to the classification algorithms were the same as the ones used in the fuzzy logic experiments - temperature, temperature difference, smoke obscuration, and smoke obscuration difference. We performed a 10-fold cross validation for both classification algorithms. Table 4 shows the number of incorrectly classified instances for the two fire scenarios as well as what percentage of the total instances was incorrectly classified. Both algorithms produce a number of inaccurate classifications. Although the percentage of the erroneously classified instances is low, it is higher than what the fuzzy logic detection managed to achieve. Fire detection delay. Table 5 shows the delay incurred by the different fire detection mechanisms. Fire is detected just as fast, and in the burning mattress scenario even faster, when fuzzy values are used. In addition, decreasing the size of the rule-base does not delay the fire detection. We also notice that including the readings of neighbor sensors in the decision process slightly slows down the detection. This is not surprising since not all sensors are located at the same distance from the fire and therefore they start registering abnormal values at different times. Consequently, if a sensor is waiting for its neighbors to also detect the fire, and those neighbors are located further away from the fire source, the detection might be slowed down.

182

7

K. Kapitanova, S.H. Son, and K.-D. Kang

Conclusions

A disadvantage of the current event detection approaches used in WSNs is that they cannot properly handle the often imprecise sensor readings. In this paper we show that fuzzy logic is a powerful and accurate mechanism which can successfully be applied to event detection in WSNs. Compared to using crisp values, fuzzy logic allows the detection algorithm to maintain a high accuracy level despite fluctuations in the sensor values. This helps decrease the number of false positives while still providing fast and accurate event detection. Our experiments support the hypothesis that incorporating the readings of neighbor nodes in the decision process further improves the event detection accuracy. The evaluation also shows that the rule-base reduction techniques we have developed are efficient and preserve both the correctness and the timeliness of event detection. Using two of these techniques together reduced the size of our experimental rule-base by more than 70%. Further, compared to two well-established classification algorithms, fuzzy logic provides more accurate event detection.

References 1. Cornell Database Group-Cougar, http://www.cs.cornell.edu/bigreddata/cougar/ 2. Govindan, R., Hellerstein, J., Hong, W., Madden, S., Franklin, M., Shenker, S.: The sensor network as a database. Computer Science Department, University of Southern California, Technical Report 02-771 (2002) 3. Li, S., Son, S.H., Stankovic, J.: Event detection services using data service middleware in distributed sensor networks. In: Zhao, F., Guibas, L.J. (eds.) IPSN 2003. LNCS, vol. 2634, pp. 502–517. Springer, Heidelberg (2003) 4. Madden, S., Franklin, M., Hellerstein, J., Hong, W.: The design of an acquisitional query processor for sensor networks. In: SIGMOD, pp. 491–502 (2003) 5. Franklin, M.: Declarative interfaces to sensor networks. Presentation at NSF Sensor Workshop (2004) 6. Jiao, B., Son, S., Stankovic, J.: GEM: Generic event service middleware for wireless sensor networks. In: INSS (2005) 7. Kapitanova, K., Son, S.H.: MEDAL: A compact event description and analysis language for wireless sensor networks. In: INSS (2009) 8. Tapia, E., Intille, S., Larson, K.: Activity recognition in the home using simple and ubiquitous sensors. In: Pervasive Computing, pp. 158–175 (2004) 9. Wren, C., Tapia, E.: Toward scalable activity recognition for sensor networks. In: Location and Context-Awareness (LoCA), pp. 168–185 (2006) 10. Castro, P., Chiu, P., Kremenek, T., Muntz, R.R.: A probabilistic room location service for wireless networked environments. In: Abowd, G.D., Brumitt, B., Shafer, S. (eds.) UbiComp 2001. LNCS, vol. 2201, pp. 18–34. Springer, Heidelberg (2001) 11. Duarte, M., Hu, Y.-H.: Distance based decision fusion in a distributed wireless sensor network. In: Zhao, F., Guibas, L.J. (eds.) IPSN 2003. LNCS, vol. 2634, pp. 392–404. Springer, Heidelberg (2003) 12. Chen, T.M., Venkataramanan, V.: Dempster-shafer theory for intrusion detection in ad hoc networks. IEEE Internet Computing, 35–41 (2005)

Event Detection in WSNs – Can Fuzzy Values Be Accurate?

183

13. Wu, H., Siegel, M., Stiefelhagen, R., Yang, J.: Sensor fusion using dempster-shafer theory. In: Proceedings of IEEE IMTC, pp. 21–23 (2002) 14. Murphy, R.: Dempster-shafer theory for sensor fusion in autonomous mobilerobots. IEEE Transactions on Robotics and Automation, 197–206 (1998) 15. Wood, A., Virone, G., Doan, T., Cao, Q., Selavo, L., Wu, Y., Fang, L., He, Z., Lin, S., Stankovic, J.: Alarm-net: Wireless sensor networks for assisted-living and residential monitoring. University of Virginia, Technical Report CS-2006-13 (2006) 16. Lymberopoulos, D., Ogale, A., Savvides, A., Aloimonos, Y.: A sensory grammar for inferring behaviors in sensor networks. In: IPSN, pp. 251–259 (2006) 17. Ghasemzadeh, H., Barnes, J., Guenterberg, E., Jafari, R.: A phonological expression for physical movement monitoring in body sensor networks. In: MASS, pp. 58–68 (2008) 18. Amft, O., Kusserow, M., Tr¨ oster, G.: Probabilistic parsing of dietary activity events. In: BSN, pp. 242–247 (2007) 19. Gupta, I., Riordan, D., Sampalli, S.: Cluster-head election using fuzzy logic for wireless sensor networks. In: CNSR, pp. 255–260 (2005) 20. Kim, J., Park, S., Han, Y., Chung, T.: CHEF: Cluster head election mechanism using fuzzy logic in wireless sensor networks. In: ICACT, pp. 654–659 (2008) 21. Lee, H., Cho, T.: Fuzzy logic based key disseminating in ubiquitous sensor networks. In: ICACT, pp. 958–962 (2008) 22. Kim, B., Lee, H., Cho, T.: Fuzzy key dissemination limiting method for the dynamic filtering-based sensor networks. In: Huang, D.-S., Heutte, L., Loog, M. (eds.) ICIC 2007. LNCS, vol. 4681, pp. 263–272. Springer, Heidelberg (2007) 23. Lazzerini, B., Marcelloni, F., Vecchio, M., Croce, S., Monaldi, E.: A fuzzy approach to data aggregation to reduce power consumption in wireless sensor networks. In: NAFIPS, pp. 436–441 (2006) 24. Kim, J., Cho, T.: Routing path generation for reliable transmission in sensor networks using GA with fuzzy logic based fitness function. In: Gervasi, O., Gavrilova, M.L. (eds.) ICCSA 2007, Part III. LNCS, vol. 4707, pp. 637–648. Springer, Heidelberg (2007) 25. Chiang, S.-Y., Wang, J.-L.: Routing analysis using fuzzy logic systems in wireless sensor networks. In: Lovrek, I., Howlett, R.J., Jain, L.C. (eds.) KES 2008, Part II. LNCS (LNAI), vol. 5178, pp. 966–973. Springer, Heidelberg (2008) 26. Ren, Q., Liang, Q.: Fuzzy logic-optimized secure media access control (fsmac) protocol wireless sensor networks. In: CIHSPS, pp. 37–43 (2005) 27. Munir, S.A., Bin, Y.W., Biao, R., Jian, M.: Fuzzy logic based congestion estimation for qos in wireless sensor network. In: WCNC, pp. 4336–4341 (2007) 28. Xia, F., Zhao, W., Sun, Y., Tian, Y.-C.: Fuzzy Logic Control Based QoS Management in Wireless Sensor/Actuator Networks. Sensors, 3179–3191 (2007) 29. Liang, Q., Wang, L.: Event detection in wireless sensor networks using fuzzy logic system. In: CIHSPS (2005) 30. Marin-Perianu, M., Havinga, P.: D-FLER: A distributed fuzzy logic engine for rulebased wireless sensor networks. In: Ichikawa, H., Cho, W.-D., Satoh, I., Youn, H.Y. (eds.) UCS 2007. LNCS, vol. 4836, pp. 86–101. Springer, Heidelberg (2007) 31. Zadeh, L.: Outline of a new approach to the analysis of complex systems and decision processes. IEEE Transactions on Systems, Man, and Cybernetics, 28–44 (1973) 32. Klir, G.J., Yuan, B.: Fuzzy sets and fuzzy logic: theory and applications. PrenticeHall, Inc., Upper Saddle River (1995) 33. NRC FuzzyJ Toolkit, http://www.csie.ntu.edu.tw/sylee/courses/fuzzyj/docs/

184

K. Kapitanova, S.H. Son, and K.-D. Kang

34. Building and fire research laboratory, http://smokealarm.nist.gov/ 35. WS4916 Series Wireless Smoke Detector 36. Geiman, J., Gottuk, D.: Alarm thresholds for smoke detector modeling, pp. 197– 208 (2003) 37. Lewis, D.D.: Naive (Bayes) at forty: The independence assumption in information retrieval. In: N´edellec, C., Rouveirol, C. (eds.) ECML 1998. LNCS, vol. 1398, pp. 4–15. Springer, Heidelberg (1998) 38. Quinlan, J.R.: C4.5: Programs for Machine Learning (1993) 39. Hall, M., Frank, E., Holmes, G., Pfahringera, B., Reutemann, P., Witten, I.: The WEKA data mining software: An update (2009)

An Efficient Geo-Routing Aware MAC Protocol for Underwater Acoustic Networks (Invited Paper) Yibo Zhu, Robert Zhong Zhou, James Peng Zheng, and Jun-Hong Cui Comupter Science & Engineering Department, University of Connecticut, Storrs CT 06269, USA {yibo.zhu,zhongzhou,zhengpeng,jcui}@engr.uconn.edu

Abstract. In this paper, we propose an efficient geo-routing aware MAC protocol (GOAL) for underwater acoustic networks. It smoothly integrates self-adaptation based RTS/CTS, geographic cyber carrier sense and implicit ACK to do combined channel reservation and next hop selection. As a result, it possesses the advantages of both geo-routing protocol and reservation based MAC protocol. Specifically, its self-adaptation based RTS/CTS, node can dynamically find out the best nexthop with low route discovery cost. In addition, through geographic cyber carrier sense, node can map its neighbors’ time slots for sending/receiving DATA packets to its own time line, and thus the collision among data packets can be greatly reduced. With these features, GOAL outperforms geo-routing protocols. Plentiful simulation results show that GOAL provides much higher end-to-end reliability with lower energy consumptions than existing VBF routing with broadcast MAC protocol. Keywords: underwater sensor network, geo-routing, MAC, self-adaptation, geographic cyber carrier sense.

1 Introduction Underwater acoustic network is a promising technique that could connect underwater vehicles, sensor nodes and other devices working in an underwater environment via acoustic channels. It can be used to collect oceanographical data, monitor oceanic volcano activity or oil/gas field [1–3]. Although it is one class of ad hoc networks, the routing and MAC protocols for terrestrial ad hoc networks cannot serve it. This is because of its long signal propagation delay, narrow channel bandwidth, and high node mobility. Due to the same reasons, it is a big challenge to design efficient routing and MAC protocols for underwater acoustic networks [1–5]. In underwater acoustic networks, traditional routing protocols such as AODV [6] do not work here because of their intolerable costly route discovery process in longdelay underwater environments. Geo-routing protocols, such as VBF [7], VBVA [8] and DBR [9], are preferred here. These protocols do not need a dedicated route discovery and forward packets directly based on nodes’ locations. Since location information is indispensable for many aquatic applications [10–14], these protocols do not cause much extra cost and are very efficient from the routing perspective. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 185–200, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

186

Y. Zhu et al.

However, geo-routing protocols [7, 9] are usually based on the broadcast nature of the underlying acoustic channel. And it is highly possible that multiple nodes are selected as the next hop, which leads to collisions if all these next hop candidates relay the packet. Although the self-adaptation methods such as those in [7, 9] narrow down the size of the candidate set to some extent, the collision probability is still very high without proper medium access control(MAC) design and optimization. Existing MAC protocols for underwater acoustic networks, such as R-MAC [15], UWAN-MAC [16] and T-Lohi [19], are usually based on channel reservations. In these protocols, senders and receivers interact with each other to reserve channel for data communications. And the sender/receiver pair must be known before the channel reservation process, which cannot be satisfied by current geo-routing protocols since a node cannot know its next-hop node before hand in the geo-routing protocol. For example, in R-MAC, a node reserves a channel by measuring the propagation delay and mapping the slot at the sender side to the receiver side, which is not compatible with the geo-routing protocol that cannot provide the next-hop information. Thus, a new MAC protocol which can effectively suppress collisions and can be smoothly combined with geo-routing protocol is highly desirable. In this paper, we propose an efficient Geo-rOuting Aware MAC protocoL (GOAL) for underwater acoustic networks. GOAL smoothly integrates self-adaptation based RTS/CTS, geographic cyber carrier sense, and implicit ACK to find the next-hop node and to do channel reservation at the receiver side. Utilizing self-adaptation based RTS/CTS, forwarder can determine the best next-hop node with little route discovery cost. By adopting geographic cyber carrier sense, collisions among the data packets are almost eliminated. With implicit ACK strategy, control messages are significantly reduced and thus fewer collisions occur among control packets. With these techniques, GOAL is energy efficient and provides high end-to-end reliability. Another amazing feature is that GOAL can work in mobile underwater acoustic networks with localization services such as in [14]. The rest of this paper is organized as follows. Section 2 briefly discusses the related works. Then, GOAL is presented in detail in section 3. After that, simulation results and discussion are shown in section 4. At last, section 5 concludes and provides future works clues.

2 Background and Related Works In this section, we will first review related works on geo-routing protocols in underwater acoustic networks and show disadvantages in collision resolutions. And then, we will review MAC protocols for underwater acoustic networks and their differences from our work. In underwater acoustic networks, nodes communicate via acoustic channels and the propagation delay is pretty long, so it takes long time and lots of energy to do route discovery. As a result, geo-routing protocols which are based on nodes’ location draw a lot of attention. VBF appears in [7]. In this protocol, packets are forwarded along the routing pipe from the source to destination. All nodes within the routing pipe will participate in the packet forwarding process. The authors in [9] propose a protocol

An Geo-Routing Aware MAC Protocol

187

based on the depth information where packets are forwarded to nodes with less depth. A new routing protocol based on VBF with void avoidance capability was proposed in [8]. For all these protocols, A node does not explicitly choose the next-hop but cooperates with its neighbors to determine the best relay node(s) according to some self-adaptation schemes. The basic idea of self-adaptation is as follows. Whenever an eligible node gets a data packet, it starts to back off according to its location before forwarding the packet. Such a scheme can guarantee that a better relay node back off for a shorter time, so the best relay node ends back off and forwards data packet first. For example, as shown in Fig. 1, node B is closer to the vector from source node S to sink node D and also nearer to the sink node D than other neighbors of forwarder F . Thus, according to the self-adaptation scheme, the back off time of node B is shorter than that of node A after they receive the DATA packet from node F . As a result, node B first forwards the DATA packet. By overhearing the forwarding, other nodes, such as node A, cancel the back off and do not forward packet any more. In this procedure, several optimal relay nodes 1 can forward firstly and other nodes are suppressed by the overhearing the forwarding. While this procedure is repeated again and again, the packet gets closer to the sink node.

E

K A F W S

W

G

B

C

D

J

H

I

Fig. 1. Number of nodes vs. packet received ratio

Although such a self-adaptation scheme can improve the system’s performance to some extent, it cannot prevent MAC collision when there are two or more adjacent nodes forwarding packets at the same time. As shown in Fig. 1, if node J happens to forward DATA packet for node I when B relays the DATA packet, collision might occur at the common neighbors of node B and J, and thus the DATA packet might not be further forwarded. This definitely harms the end-to-end reliability of the routing protocol. Also, the dropped packets waste plenty of energy, i.e., it is not energy efficient. To further improve the performance of geo-routing protocol, effective collision resolution schemes should be employed, which are usually implemented in the MAC protocols. MAC protocols have been widely investigated for underwater acoustic networks in the last few years. In FAMA [17], RTS/CTS and carrier sense are combined to avoid collision. However, it is not energy efficient because the RTS/CTS packet is pretty long, 1

Some nodes might already forward the data packet before overhearing the forwarding by the best relay node.

188

Y. Zhu et al.

which consume lots of energy. To improve the energy efficiency, in slotted FAMA [18], a modified FAMA, both control packets and data packets are sent at the beginning of a slot. In this way, the length of RTS/CTS packet is not determined by the maximal propagation delay as that in FAMA, so the energy is much more efficient. However, the RTS/CTS handshake requires routing protocol to provide the explicit next hop, i.e., it cannot be the MAC protocol for self-adaptation based routing protocol. In T-Lohi [19], short tone message is used to reserve the channel to send data. However, even through a node does not receive any tone during a contention period, it cannot ensure that there is no collision at receiver side. In other words, it still suffers hidden terminal problem and cannot effectively avoid the collision. R-MAC appears in [15], which consists of three phases. In the first phase, each node measures the propagation delay to its neighbors. In the second phase, each node reserves receiving slot at the receiver side and then receiver confirms if the reservation is collision-free. This phase can make sure that there is no collision at receiver side for data packet. In the last phase, each node follows the reservation in the second phase to transmit data packet. Explicit receiver address is needed in phase two for the channel reservation, so it cannot work with self-adaptation based routing protocol. Unlike other reservation based MAC protocols, UWAN-MAC [16] does reservation via one way communication. Assuming the delay between neighbors does not vary, each node piggybacks the relative sending time of next packet in current packet. As a result, node knows when it will receive the next packet. However, such one way handshake cannot solve hidden terminal problem. Thus, collision is still heavy in multiple hop networks. In addition, UWAN-MAC requires node to foresee the exact sending time of next packet, which is unpractical in self-adaptation based routing protocols. Different from above works, GOAL, the new approach in this paper, smoothly integrates self-adaptation scheme and MAC reservation techniques. First, it employs the self-adaptation scheme to do handshake and find the next hop. Similar to implicitly finding the best relay in self-adaptation based geo-routing protocol, the cost of selecting next hop is pretty low in GOAL. Also, the receiving slot is reserved based on the geographic information during the handshake, and then the DATA packet can be forwarded without collision. Thus, GOAL can avoid more collision while keeping a low route cost.

3 Description of GOAL In this section, we will discuss our new geo-routing aware MAC protocol (GOAL), which is reservation based and can smoothly integrate with any known geo-routing protocols with self-adaptation. For instance, if GOAL adopts the self-adaptation scheme of VBF, it can be considered as a reservation based MAC protocol coupling with VBF. We first present the basic idea of GOAL. Then, we describe three key components of GOAL, Self-adaptation based RTS/CTS handshake, Geographic cyber carrier sense and Implicit acknowledgement. We apply the self-adaptation scheme of VBF to GOAL as a special case in the description to make it clear (Note that GOAL can be used with any self-adaptation scheme). After that, we will use one example to show the overall working process of GOAL with detailed analysis.

An Geo-Routing Aware MAC Protocol

189

3.1 Basic Idea GOAL is a reservation based MAC protocol. In GOAL, each node maintains a time schedule, which records the time slot corresponding to its neighbors’ packet sending/receiving time. Whenever a node wants to send a packet, it should make sure that the selected sending time does not overlap with any existing time slot in the time schedule line. In this way, DATA packet can be made collision free. To map the sending/receiving slot to nodes’ own time schedule line, self-adaptation based RTS/CTS handshake is employed where only a few qualified neighbors are allowed to reply a CTS packet to an RTS request, which will certainly reduce the collisions. The RTS/CTS handshaking process in GOAL is used to implement two-fold functionalities: determining the next hop and mapping neighbors’ sending and receiving slot to node’s own time schedule. S

S

R T

T1

T2

S T

S

: packet transmission

: packet reception from S

(a) collision interval

T1 N

T’-2Tprop S R T2

T’ S : packet transmission

: packet reception from S

(b) Fig. 2. (a) Mapping neighbor’s sending slot, (b) Mapping neighbor’s receiving time slot

As shown in Fig. 2(a), when sending a packet, node S piggybacks the transmission time T2 and relative sending time T of the coming packet. After receiving this packet, node R can figure out the receiving time of the packet, i.e., it maps the sending time of the packet at node S to its time schedule line. Concretely, the interval between the time when node S sends the first and second packet is T . Assuming that the propagation delay between node R and S does not change much during the coming T time period, the interval between the time when node R begins to receive these two packets is still T . Therefore, after node R totally receives the first packet, it knows that it will receive next packet during time slot [T − T1 , T − T1 + T2 ], where T1 is the transmission time of the

190

Y. Zhu et al.

first packet. Note that the time slot is expressed by relative time and can be converted to absolute time easily. Method of mapping receiving time is illustrated in Fig. 2(b). When node R sends the first packet, of which the transmission time is T1 , it notifies that it will receive the next packet T  time later. Suppose that node N knows that the propagation delay between it and node R is Tprop . After totally receiving the first packet, node R can find out that there will be a collision at node R if it emits any packet signal during [T  − T1 − 2Tprop, T  − T1 − 2Tprop + T2 ], where T2 is the transmission time of the packet which node R will receive. Therefore, to avoid collision at node R, node N should make sure that the sending interval does not overlap with time slot [T  −T1 −2Tprop, T  −T1 −2Tprop +T2 ] when it sends out any packet. Applying these two mapping schemes, node can map neighbor’s sending and receiving time period to its time schedule line so as to avoid collision when transmitting DATA packet. 3.2 The GOAL Protocol GOAL protocol consists of three parts: self-adaptation based RTS/CTS handshake, geographic cyber carrier sense, and implicit acknowledgement. As described in section 3.1, self-adaptation based RTS/CTS handshake and geographic cyber carrier sense are used to determine the optimal next hop and make channel reservation. In addition, implicit acknowledgement is imported to reduce the number of control message. The detail of these parts is provided as follows. Self-Adaptation Based RTS/CTS Handshake. When current forwarder F intends to send out a data packet, it first selects a qualified sending time to broadcast a RTS{PS , PF , PD , T , TDAT A } packet. Via the RTS packet, node F tells its neighbors that it will send the DATA packet T 2 time later and the corresponding transmission time is TDAT A . It also provides the location of the source, current forwarder, and the destination: PS , PF , PD . After receiving RTS, neighbors of the forwarder get to know that they will receive the DATA packet T − TRT S time later, where TRT S is the transmission time of RTS packet. Then, the neighbors of which the location is better than that of the forwarder start to back off according to the self-adaptation scheme of VBF. Once backoff terminates, the node sends the forwarder a CTS{Pthisnode , T  , TDAT A } packet, where Pthisnode is the location of this node and T  is the relative time it will send DATA packet. Due to the broadcast feature of acoustic medium, part of node F ’s neighbors which is still in backoff state can overhear the CTS packet. Then, they cancel backoff because the CTS shows that there is a better relay. Finally, node F decides the next hop according to received CTS packets. If it does not receive any CTS, it waits for a random time and tries to resend the RTS. Otherwise, node F must receive at least one CTS from its 2

Note that T must be bigger than two times maximum propagation delay plus maximum backoff time, which is decided by the self-adaptation scheme. Otherwise, node F cannot determine the next hop before it sends DATA packet. In this case, node F will issue a new RTS for this DATA packet later.

An Geo-Routing Aware MAC Protocol

191

neighbors. In this case, it sets next hop as the one with shortest adaptation time, which can be calculated by applying self-adaptation scheme again. Once the pre-scheduled DATA sending time comes, node F sends the DATA packet to the selected next hop. As an improvement, multiple DATA packets to the same sink node can be transmitted in one packet train [18], so RTS/CTS handshake could do reservation for multiple DATA packets in one round. Obviously, this strategy can enhance the efficiency of handshake. Geographic Cyber Carrier Sense. In underwater acoustic networks, it is difficult for node to avoid collision completely due to the long propagation delay. To address this issue, nodes in GOAL apply the two mapping methods in section 3.1 and utilize geographic information to mapping neighbors’ packet sending and receiving slot to its time schedule line. Thus, the collision at neighbors can be greatly reduced if the selected packet sending time does not overlap with any slot in its time schedule line. This is so-called geographic cyber carrier sense. Specifically, after receiving the RTS packet, node knows that it will receive the DATA packet during [T − TRT S , T − TRT S + TDAT A ] by applying method of mapping neighbor’s sending time slot, where TRT S stands for the transmission time of RTS packet. Then, this node converts the time slot to absolute time and inserts into its time schedule line. CTS packet has two-fold functionalities: responding the RTS packet and notifying neighbors to avoid collision. On one side, with the CTS packet from node R, the sender of RTS knows that node R is a potential next hop. On the other side, based on the information in CTS packet, other neighbors of node R can evaluate the propagation delay Tprop between them and node R. The evaluation method is to use propagation speed to divide the Euclidean distance. Then applying the method of mapping neighbors’ receiving time slot, this node gets that the there will be a collision at node R if it send packet during [T  − TCT S − 2Tprop, T  − TCT S − 2Tprop + T2 ], where TCT S is the transmission time of CTS packet. To avoid collision, this node should not emit any packet signal during this period. Note that propagation delay measure method might introduce error because the acoustic signal is transmitted along bent way and the nodes are mobile. To tolerate the error, guard time Tguard is used, i.e., the propagation delay is in range [Tprop − Tguard , Tprop + Tguard ]. Thus, the time period becomes [T  − TCT S − 2Tprop − 2Tguard , T  − TCT S − 2Tprop + TDAT A + 2Tguard ]. Based on geographic cyber carrier sense, nodes can obtain neighbors’ sending and receiving schedule after RTS/CTS handshake. By recording the schedules in its time schedule line, node can conveniently choose a qualified time to send packet. Implicit Acknowledgement. In terrestrial ad hoc networks, RTS/CTS/DATA/ACK can really improve the reliability of one hop transmission. In underwater acoustic networks, however, this scheme definitely causes more collisions among control packets because of low bandwidth and long propagation delay. A possible way is to adopt implicit acknowledgement scheme to reduce the number of control packets. Specifically, if the node which receives the DATA packet is not the destination, it must send RTS to determine the next hop within certain time. Because the previous hop is still within one hop range with a high probability, it can also overhear the RTS. Based on this heuristic rule,

192

Y. Zhu et al.

RTS is revised to include packet identifier (PID) of the DATA packet. As a result, the previous hop can confirm that the DATA packet is successfully forwarded. For the destination node, it explicitly acknowledges the DATA packet using an ACK packet. In addition, a node will send explicit ACK packet without backoff if it receives a RTS when both of the following conditions are met: 1) the location of this node is better than the sender of RTS; 2) this node has received the DATA packet which RTS requests for. For any node, if it does not receive an implicit acknowledgement or ACK packet within certain time after sending out the DATA packet, it will initiate a new RTS/CTS/DATA round to retransmit the DATA packet. Although retransmission can improve the transmission reliability, maximum times of retransmission should not be infinite because it introduces more delay and energy consumption. Thus, we define maximum retransmission times as a tradeoff. Specifically, one node can transmit and retransmit (both failure of receiving RTS and acknowledgement can cause retransmission) a DATA packet at most maximum retransmission times. If the maximum retransmission times is exceeded, the node should give up the DATA packet. 3.3 An Example of GOAL In the example, the network topology is shown as Fig. 3(a). Node F now tries to forward the DATA packet from source node S to destination node D. Following GOAL protocol, node F selects a qualified sending time to broadcast a RTS{PS , PF , PD , T , TDAT A , P ID} packet. Via the RTS packet, node F notifies its neighbors that it will send the DATA packet T time later and the corresponding transmission time is TDAT A . With the information in the RTS packet, node C, A, and B figure out that they will receive the DATA packet T − TRT S time later. Note that node C will not back off because its location is even worse than that of current forwarder node F , so these nodes except node C start to back off according to the self-adaptation scheme in VBF. Same as VBF, since node B first ends backoff state, it sends CTS{PB , T  , TDAT A } packet to node F . On one side, by overhearing the CTS packet from node B, node A realizes that there is a better relay, so it cancels the backoff. Also, based on the information in the CTS packet, node A can evaluate TBA , which denotes the propagation delay between node B and node A. Thus, node A will not send any packet during time interval [T  − TCT S − 2TBA − 2Tguard , T  − TCT S − 2TBA + TDAT A + 2Tguard ]. On the other side, node F finds out that the next hop could be node B after receiving the CTS packet. When the scheduled DATA packet sending time comes, node F sends DATA packet to node B since node B is the optimal one. Later on, node B tries to forward the DATA packet, so it sends RTS{PS , PB , PD , T , TDAT A , P ID} packet to do handshake. Receiving this RTS packet, node F makes sure that the forwarding is successful and then prepares to forward other DATA packets. 3.4 Analysis of GOAL In GOAL, nodes apply self-adaptation scheme in the RTS/CTS handshake to find the next hop. This procedure is similar to the normal self-adaptation based geo-routing protocol for data packet. Because RTS/CTS packet is much shorter than data packet, the

An Geo-Routing Aware MAC Protocol

193

E A B

F

D

G H

C S

(a) F

F C B F

T

F A

B

B

Tbackoff F

B

B

F

F Tbackoff T-TCTS

TCTS

: CTS transmission

: RTS transmission F

F : RTS reception from F

: DATA transmission F

: CTS reception from F

: DATA reception from F

(b) Fig. 3. (a) Network topology, (b) One hop forwarding

probability of collision among RTS/CTS packets in GOAL is much lower than that among data packets in self-adaptation based geo-routing protocol. Note that data packets in GOAL are almost collision-free owing to geographic cyber carrier sense, so the entire collision probability in GOAL is lower than that in self-adaptation based georouting protocol. As a result, GOAL could provide a higher end-to-end reliability than self-adaptation based geo-routing protocol. As analyzed above, GOAL introduces MAC collision among short RTS/CTS while avoiding collision among long data packets. As a result, the collision probability is reduced. Note that the collision among long data packets wastes more energy than the collision among short ones, so GOAL requires less energy consumption for packet delivery than self-adaptation based geo-routing protocol. To achieve above desirable features, however, GOAL incurs longer delay. As explained in section 3.2, nodes schedule the sending time of DATA packet T time later after sending the RTS packet, where T is at least two times maximum propagation delay plus maximum backoff time. Moreover, to use implicit acknowledgement, nodes also wait for more than one round trip. Furthermore, nodes in GOAL do retransmission if

194

Y. Zhu et al.

any failure occurs during the forwarding procedure, which also increase the delivery delay. Hence, the delivery delay in GOAL is higher than self-adaptation based geo-routing protocol.

4 Performance Evaluation In this section, we use simulations to evaluate the performance of GOAL. Aqua-Sim [20], a NS-2 based underwater acoustic network simulator which is developed by the UWSN lab at UCONN, has been used for our simulations. 4.1 Simulation Settings We simulate GOAL in a practical underwater scenario, which is abstracted from the application of monitoring gas/oil/oceanic volcano activity. Nodes are randomly deployed within a 300×300×500m cubic area. Whenever a node detects an event, it will send the data collected to the sink node. To simplify the simulations, we make two assumptions: 1) a node can detect the event occurring within its sensing range; 2) event lasts for a long time 3 , so nodes send data to sink node periodically as long as it can sense the event. The period is defined as sensing interval. All node can move freely in horizontal two-dimensional space, i.e., in the X-Y plane. The speed of node follows an uniform distribution between 0.2 and 1.5 m/s. The transmission range is set to 120 meters. Sink node which is the destination for all data packets is fixed at (250, 250, 0). Nodes’ sensing range is 80 meters. The size of data packet is 300 bytes. The maximum retransmission times is set to be 6. Each simulation last for 5000 seconds. The energy consumption parameters are based on a commercial underwater acoustic modem, UMW 1000, from LinkQuest [21]: the power consumption on transmission mode is 2 Watts; the power consumption on receive mode is 0.75 Watts; and the power consumption on sleep mode is 8 mW. Three metrics are used to quantify the performance: packet delivery ratio, energy consumption per byte and delivery delay. Specifically, the packet delivery ratio is the ratio of the total number of packet sent by source nodes to the number of packet received by sink. The energy consumption per byte is to divide the total network energy consumption by the number of data bytes successfully received by the sink. The delivery delay is the average end-to-end delay of each packet received by the sink. We compare the performance of GOAL with VBF coupling with broadcast MAC (we use VBF for short in the rest parts) [7]. 4.2 Simulation Results Impacts of Data Sensing Interval. In this set of simulations, the number of nodes in the network is fixed to be 100 and we change the data sensing interval of every node from 20 to 70 seconds. 3

This is practical. For example, oceanic volcano usually belches slight smoke and ashes for a long time before it finally erupts.

An Geo-Routing Aware MAC Protocol

195

As shown in Fig. 4– 6, GOAL can provide a high end-to-end reliability. Fig. 4 clearly show us that GOAL can provide much higher packet delivery ratio than VBF. This is because GOAL can greatly reduce collision by its RTS/CTS handshaking process and its channel reservation mechanism. In addition, we can see that the packet delivery ratio of GOAL increases while the sensing interval becomes larger. This is because nodes with a larger sensing interval generate less packets, which accordingly causes less collisions. Since the maximum retransmission times is fixed, the packet delivery ratio is improved when there are less collisions. We can also see that the packet delivery ratio of VBF does not vary much while the sensing interval increases. This is because VBF is best-effort and the collision probability of VBF mainly depends on the self-adaptation scheme, which is highly related to the node distribution. Note that the size of network is fixed and nodes are uniformly deployed. Hence, the node distribution is decided by node density. In this group of simulation, node density is fixed, so the packet delivery ratio nearly keeps the same value. 1

Packet Delivery Ratio

0.9 0.8 0.7 0.6 0.5

GOAL VBF

0.4 20

30

40 50 Sensing Interval(s)

60

70

Fig. 4. Sensing interval vs. packet delivery ratio

GOAL can also achieve high energy efficiency. From Fig. 5, we can clearly see that GOAL is more energy efficient than VBF, especially when the sensing interval becomes larger. This is because in GOAL, when the sensing interval is shorter, multiple packets can be sent together with just one RTS/CTS handshaking, which can improve the system’s energy efficiency. In addition, as the sensing interval becomes larger, less data packets is sent in the network. Thus, most nodes will waste its energy in the idle state with constant rate (8mw). This also increases the energy consumption when the sensing interval is larger. Considering the reliability requirement, the energy consumption in VBF is much higher than that in GOAL. For example, let us set PG as the delivery ratio of GOAL and PV as delivery ratio of VBF. And set EG and EV as energy consumption of GOAL and VBF respectively. To achieve the same packet delivery ratio, VBF should do retransmission for N times in average and thus the energy consumption is N EV , where N satisfies 1 − (1 − PV )N = PG (1)

196

Y. Zhu et al.

Hence, N can be expressed as follows N = log1−PV (1 − PG )

(2)

In Fig. 4, for example, the packet delivery ratio of GOAL and VBF is approximately 0.97 and 0.73 when sensing interval is 50 second, respectively. Applying above equation, we can get the average times that VBF should transmit each packet to reach the same packet delivery ratio as GOAL as follows. N = log1−0.73 (1 − 0.97) = 2.67

(3)

Therefore, the energy consumption of VBF should be at least doubled. In other words, the energy consumption in GOAL is less than half of that in VBF, which indicates that GOAL is more energy-efficient.

Energy Consumption(Joule/Byte)

0.8 0.7

GOAL VBF

0.6 0.5 0.4 0.3 0.2 20

30

40 50 Sensing Interval(s)

60

70

Fig. 5. Sensing interval vs. energy consumption

Fig. 6 shows us that the end-to-end delay of GOAL decreases with the increase of sensing interval. This is because collisions increase when sensing interval is shorter. With collisions, nodes have to initiate a new RTS/CTS/DATA round to do retransmission, which introduces extra delay. As the sensing interval becomes larger, less collisions and retransmission appear. Therefore, the delay decreases while the sensing interval increases. For VBF, which is a best-effort protocol, the delivery delay nearly has nothing to do with traffic rate, but is mainly decided by the backoff time in self-adaptation scheme. Thereby, the delivery delay in VBF almost does not change in Fig. 6. Impacts of Node Density. In this set of simulations, we set the sensing interval of every node to be 50 seconds and change the number of nodes in the network from 70 to 120. The impact of node density is shown in the coming three figures. In Fig. 7, we can see that the packet delivery ratio of GOAL is much higher than that of VBF. This is still because GOAL reduces more collision than VBF and VBF is best-effort. Also, we

An Geo-Routing Aware MAC Protocol

50

197

GOAL VBF

Delivery Delay (s)

40 30 20 10 0 20

30

40 50 Sensing Interval(s)

60

70

Fig. 6. Sensing interval vs. delivery delay

can see that the packet delivery ratio of both GOAL and VBF increases while there are more nodes within network. One reason is as mentioned before: GOAL largely reduces the MAC collision via doing reservation for DATA packet. The other reason it related to the self-adaptation scheme. When the node density is lower, there are fewer qualified next hops according to the self-adaptation scheme. Particularly, some forwarders do not have qualified next hop. For VBF, a best effort protocol, the DATA packet is definitely dropped in such case. In GOAL, forwarding failure can be detected by missing the implicit acknowledgement, and thus retransmission is issued. From Fig. 8, we can see that GOAL consumes less energy than VBF for transmitting every unit data from source to sink. The reason is similar to that of Fig. 5. In VBF, the collision probability is higher than that in GOAL. Moreover, each collided packet in VBF wastes much more energy than that in GOAL because the packet in VBF is much 1

Packet Delivery Ratio

0.9 0.8 0.7 0.6 0.5

GOAL VBF

0.4 70

80

90 100 Number of Nodes

110

Fig. 7. Number of nodes vs. packet delivery ratio

120

198

Y. Zhu et al.

Energy Consumption(Joule/Byte)

0.8

GOAL VBF

0.7 0.6 0.5 0.4 0.3 0.2 70

80

90

100

110

120

Number of Nodes

Fig. 8. Number of nodes vs. energy consumption 50

GOAL VBF

Delivery Delay (s)

40 30 20 10 0 70

80

90 100 Number of Nodes

110

120

Fig. 9. Number of nodes vs. delivery delay

longer. As a result, GOAL saves more energy. Like the analysis for Fig. 5, if we analyze the energy consumption with same packet delivery ratio, we can see that the energy consumption in GOAL is much less than that in VBF, especially when there are less nodes within the network. In Fig. 9, the delivery delay of GOAL is higher than VBF. The reason has been mentioned before: the handshaking and implicit acknowledgement in GOAL introduce more delay while VBF is a best-effort protocol which does not care whether the transmission to next hop is successful. Due to the same reason, the delivery delay of VBF is almost same in Fig. 9. Additionally, we can see that the delivery delay of GOAL slightly increases while the node deployment becomes dense. This is because dense deployment causes more collisions among the control messages. Thus, the retransmission times, which raises the delivery delay, are increased.

An Geo-Routing Aware MAC Protocol

199

5 Conclusion In this paper, GOAL, an efficient geo-routing aware MAC protocol is proposed for underwater sensor networks. It is a reservation based MAC protocol which can smoothly integrate with any existing geo-rouing protocols with self-adaptation capability. Selfadaptation based RTS/CTS handshaking, geographic cyber carrier sense and implicit acknowledgement are used in GOAL to improve system performance. Although the end-to-end delivery delay increases because of the hop-by-hop retransmission mechanism in GOAL, it can achieve high end-to-end delivery ratio with low energy consumption. Plentiful simulation results show that GOAL outperforms existing VBF with broadcast MAC in both end-to-end delivery ratio and energy efficiency.

References 1. Cui, J., Kong, J., Gerla, M., Zhou, S.: Challenges: Building Scalable Mobile Underwater Wireless Sensor Networks for Aquatic Applications. IEEE Network, Special Issue on Wireless Sensor Networking 20(3), 12–18 (2006) 2. Heidemann, J., Ye, W., Wills, J., Syed, A., Li, Y.: Research Challenges and Applications for Underwater Sensor Networking. In: Proceedings of IEEE Wireless Communications and Networking Conference, pp. 228–235 (2006) 3. Partan, J., Kurose, J., Levine, B.N.: A Survey of Practical Issues in Underwater Networks. In: Proceedings of ACM WUWNet (2006) 4. Liu, L., Zhou, S., Cui, J.: Prospects and problems of wireless communication for underwater sensor networks. Wireless Communications & Mobile Computing 8(8), 977–994 (2008) 5. Chitre, M., Shahabudden, S., Stojanovic, M.: Underwater acoustic communicatin and networks: Recent advances and future challenges. Marine Technology Society Journal 42(1), 103–116 (2008) 6. Perkins, C.E., Royer, E.M.: Ad hoc on-demand distance vector routing. In: Proceedings of IEEE Workshop on Mobile Computing Systems and Applications, pp. 90–100 (1999) 7. Xie, P., Cui, J., Lao, L.: VBF: Vector-Based Forwarding Protocol for Underwater Sensor Networks. In: Proceedings of IFIP Networking, pp. 228–235 (2006) 8. Xie, P., Zhou, Z., Peng, Z., Cui, J., Shi, J.: Void Avoidance in Three-Dimensional Mobile Underwater Sensor Networks. In: Liu, B., Bestavros, A., Du, D.-Z., Wang, J. (eds.) Wireless Algorithms, Systems, and Applications. LNCS, vol. 5682, pp. 305–314. Springer, Heidelberg (2009) 9. Yan, H., Shi, Z., Cui, J.: DBR: Depth-Based Routing for Underwater Sensor Networks. In: Proceedings of IFIP Networking, pp. 72–86 (2008) 10. Cheng, X., Shu, H., Liang, Q., Du, H.: Slient positioning in underwater acoustic sensor networks. IEEE Transactions on Vehicular Technology 57(3), 1756–1766 (2008) 11. Erol, M., Vierira, L.F.M., Gerla, M.: AUV-Aided localization for underwater sensor networks. In: Proceedings of International Conference on Wireless Algorithms, Systems and Applications (WASA), pp. 44–51 (2007) 12. Frampton, K.D.: Acoustic self-localization in a distributed sensor network. IEEE Sensors Journals 6(1), 166–172 (2006) 13. Zhou, Z., Cui, J., Zhou, S.: Localization for Large Scale Underwater Sensor Networks. In: Proceedings of IFIP Networking, pp. 108–119 (2007) 14. Zhou, Z., Cui, J., Bagtzoglou, A.: Scalable Localization Scheme with Mobility Prediction for Underwater Sensor Networks. In: Proceedings of IEEE INFOCOM, pp. 2198–2206 (2008)

200

Y. Zhu et al.

15. Xie, P., Zhou, Z., Cui, J., Shi, Z.: R-MAC: An Energy-Efficient MAC Protocol for Underwater Sensor Networks. In: Proceedings of International Conference on Wireless Algorithms, Systems and Applications (WASA), pp. 187–198 (2007) 16. Park, M.K., Rodoplu, V.: UWAN-MAC: An Energy-Efficient MAC Protocol for Underwater Acoustic Wireless Sensor Networks. IEEE Journal of Oceanic Engineering 32(3), 710–720 (2007) 17. Fullmer, C.L., Garcia-Luna-Aceves, J.: Floor Acquisition Multiple Access (FAMA) for Packet-Radio Networks. In: Proceedings of ACM SIGCOMM, pp. 262–273 (1995) 18. Molins, M., Stojanovic, M.: Slotted FAMA: a MAC protocol for underwater acoustic networks. In: Proceedings of IEEE OCEANS, pp. 16–19 (2006) 19. Syed, A., Ye, W., Heidemann, J.: T-Lohi: A New Class of MAC Protocols for Underwater Acoustic Sensor Networks. In: Proceedings of IEEE INFOCOM, pp. 231–235 (2008) 20. Xie, P., Zhou, Z., Peng, Z., Yan, H., Hu, T., Cui, J., Shi, Z., Fei, Y., Zhou, S.: Aqua-Sim: An NS-2 Based Simulator for Underwater Sensor Networks. In: Proceedings of IEEE/MTS Oceans, pp. 1–7 (2009) 21. LinkQuest, http://www.link-quest.com

A Decentralized Scheduling Algorithm for Time Synchronized Channel Hopping (Invited Paper) Andrew Tinka1 , Thomas Watteyne2 , and Kris Pister2 1 Electrical Engineering, University of California, Berkeley, USA [email protected] 2 Berkeley Sensor & Actuator Center University of California, Berkeley, USA {watteyne,pister}@eecs.berkeley.edu

Abstract. Time Synchronized Channel Hopping (TSCH) is an existing medium access control scheme which enables robust communication through channel hopping and high data rates through synchronization. It is based on a time-slotted architecture, and its correct functioning depends on a schedule which is typically computed by a central node. This paper presents, to our knowledge, the first scheduling algorithm for TSCH networks which both is distributed and which copes with a mobile nodes. Two scheduling algorithms are presented. Aloha-based scheduling allocates one frequency channel for broadcasting advertisements for new neighbors. Reservation-based scheduling augments Aloha-based scheduling with a dedicated slot for targeted advertisements based on gossip information. A mobile ad-hoc network with frequent connectivity changes is simulated, and the performance of the two proposed algorithms is assessed against the optimal case. Reservation-based scheduling performs significantly better than Aloha-based scheduling, suggesting that the improved network reactivity is worth the increased algorithmic complexity and resource consumption. Keywords: time-synchronized channel hopping, mobile ad-hoc networks, decentralized scheduling, simulation.

1

Introduction

The Floating Sensor Network (FSN) project[1] at UC Berkeley builds autonomous floating sensor packages for deployments in rivers and estuaries (see Fig. 1). The floating sensors (or “drifters”, in the terminology of the hydrodynamic community) are untethered; once deployed in the river, they are carried by the current. Some versions of the drifters can affect trajectory modifications using propellers. The Berkeley FSN devices carry two communication systems: a GSM module for transmissions to a central server, and an low-power low-range IEEE802.15.4-2006[2] radio for communication between nodes. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 201–216, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

202

A. Tinka, T. Watteyne, and K. Pister

Fig. 1. Prototype of a motorized drifter (left). Five drifters in a river (right).

The GSM communication channel is both expensive (both monetary and in terms of energy) and unreliable (due to variable GSM coverage). One strategy for delivering data from individual nodes to a remote server is to have one or several nodes with good GSM connection act as ad hoc sink nodes. Nodes connected by IEEE802.15.4-2006 links that do not have their own GSM connections available can send their data to one of the sinks, which retransmits it via GSM to the server. Since the nodes with GSM connectivity are not known a priori, the design objective for the IEEE802.15.4-2006 network must be to maximize point-to-point connectivity. We define the physical connectivity graph to be the ensemble of wireless links good enough to be used for communication at a given instant in time. We define the logical connectivity graph to be the set of links scheduled to be used at the same instant. The mobility of the nodes means that the physical connectivity between nodes changes significantly over time. Global connectivity is not guaranteed. Therefore, centralized schemes for determining a communication schedule are poor fits for the problem. Our goal is to develop an algorithm which schedules intermittent bi-directional links between neighboring nodes as these links become available. We assess candidate schemes by evaluating how close the logical connectivity gets to the physical connectivity; that is, how many of the possible links are actually scheduled by the protocol. Time Synchronized Channel Hopping (TSCH) is a Medium Access Control (MAC) scheme which enables robust communication through channel hopping and high data rates through synchronization. It is based on a time-slotted architecture, where a schedule indicates to the nodes on which slot and which frequency channel to transmit/receive data to/from which neighbor. TSCH is being standardized by the the IEEE802.15.4e Working Group[3], and expected to be included in the next revision of the IEEE802.15.4 standard.

A Decentralized Scheduling Algorithm for TSCH

203

TSCH only defines the mechanism, and makes no assumption on how the schedule is built. Typically, nodes report their communication needs (expressed in terms on throughput, reliability and latency) to a central scheduler which computes a schedule and injects this into the network. This technique has proven perfectly adequate for static networks such as industrial control Wireless Sensor Networks (WSNs). A distributed solution seems more appropriate for mobile networks. In those type of networks, each topological change would have to be reported to the central scheduler, which would have to recompute a schedule and inform the nodes about the change. This paper presents distributed scheduling algorithms to be used on top of a TSCH MAC protocol, and which deals with mobile nodes. It is, to our knowledge, the first of its kind. The remainder of this paper is organized as follows. Section 2 provides a comprehensive overview of MAC protocol approaches and standardization activities, and highlights the need for a distributed scheduling algorithm for TSCH. Section 3 then details the two scheduling algorithms proposed in this paper, called Aloha-Based Scheduling and Reservation-Based Scheduling. We evaluate the goodness of the connectivity achieved by both algorithms when compared to the physical connectivity in Section 4. The metric of interest are relative connectivity (a metric for goodness in the static case) and link duration (dynamic case). Finally, Section 5 concludes this paper and presents future work.

2

Time Synchronized Channel Hopping

There are two main approaches for regulating access to a shared wireless medium: contention-based and reservation-based approaches. Any derived Medium Access Control (MAC) protocol is based on one of those two approaches, or a combination thereof. Contention-based protocols are fairly simple, mainly because neither global synchronization nor topology knowledge is required. In a contention-based approach, nodes compete for the use of the wireless medium and only the winner of this competition is allowed to access to the channel and transmit. Aloha and Carrier Sense Multiple Access (CSMA) are canonical representative schemes of contention-based approaches. They do not rely on a central entity and are robust to node mobility, which makes it intuitively a good candidate for dynamic mobile networks. Preamble-sampling is a low-power version of contention-based medium access, widely popular in WSNs. All nodes in the network periodically sample the channel for a short amount of time (at most a few milliseconds) to check whether a transmission is ongoing. Nodes do not need to be synchronized, but all use the same check interval. To ensure all neighbors are listening, a sender pre-pends a preamble which is at least as long as the check interval. Upon hearing the preamble, nodes keep listening for the data that follows. The optimal check interval, which minimizes the total energy expenditure, is a function of the average

204

A. Tinka, T. Watteyne, and K. Pister

network degree and its load. Durations of 100 ms are typical. Numerous works have proposed ways to optimize the sampling[4], reducing the preamble length by packetization[5] or by synchronizing the nodes[6]. Despite its success, contention-based protocols suffer from degraded performance in terms of throughput when the traffic load increases. In addition, the distributed nature prevents them from achieving the same efficiency as ideal reservation-based protocols. Finally, frequency agility is hard to achieve by such protocols as nodes are not synchronized. Reservation-based protocols require the knowledge of the network topology to establish a schedule that allows each node to access the channel and communicate with other nodes. The schedule may have various goals such as ensuring fairness among nodes, or reducing collisions by avoiding that two interfering nodes or more access to the channel and transmit at the same time. TDMA (Time Division Multiple Access) is a representative example for such a reservation-based approach. In TDMA, time is divided into slots which are grouped into superframes which repeat over time. A schedule is used to indicate to each node when it has to transmit or receive, to/from which neighbor. Provided the schedule is correctly built, transmissions do not suffer from collisions, which guarantees finite and predictable scheduling delays and also increases the overall throughput in highly loaded networks. The reliability of a wireless link is mainly challenged by external interference and multi-path fading. [7] and [8] show how channel hopping combats both of these, respectively. If a transmission fails, the sender retransmits the packet on a different frequency channel. Because this frequency change causes the wireless environment to be different, the retransmission has a higher probability of being successful than if it were retransmitted on the same channel. Channel hopping was first applied to Wireless Sensor Networks (WSNs) in a proprietary protocol called Time Synchronized Mesh Protocol (TSMP)[9]. In TSMP, nodes in the network are synchronized on a slotted time base. An individual slot is long enough for a sender to send a data frame, and for a receiver to acknowledge correct reception (a slot of 10 ms is common). L consecutive slots form a superframe, which indefinitely repeats over time. A schedule of length L slots indicates, for each slot, whether the node is supposed to transmit or receive, to/from which neighbor and on which frequency channel. TSMP runs on IEEE802.15.4-2006[2] compliant radios, which offers 16 frequency channels in the 2.4GHz ISM band. A central scheduler is used to compute a schedule, which is then injected and used in the network. TSMP, which combines time synchronization and frequency agility, has been shown to achieve reliabilities of over 99.999%[10]. Its core idea has been standardized for industrial applications in WirelessHART[11–13] and ISA100.11a[14]. In 2009, it has been introduced in the draft standard IEEE802.15.4e under the name Time Synchronized Channel Hopping (TSCH). This draft standard will replace the current IEEE802.15.4-2006 standard in its next revision.

A Decentralized Scheduling Algorithm for TSCH

205

All of the above standards rely on a central controller to compute a schedule for the network to use. The goal of this paper is to propose a distributed alternative, targeted at mobile nodes.

3 3.1

Distributed Scheduling Algorithms Goal and Metrics

The goal of the proposed schedule is to achieve full connectivity, which is achieved when each node of the neighbor has established a bidirectional link to each of its physical neighbors. A bidirectional link is established between nodes A and B when, in the superframe, there is at least one slot scheduled from A to B, and one from B to A. The unreliability of the wireless link and the movement of the nodes are challenges the scheduling algorithm needs to cope with. If a link is present in the physical graph, it is feasible; if a link is present in the physical but not in the logical graph, it is said to be unscheduled ; a link which still appears in the logical while is disappeared from the physical graph is called stale. We use the ratio between the scheduled and feasible links as a metric for the static goodness of the scheduling algorithm. Node mobility causes links to come and go. A link therefore has a finite lifetime, or link duration. To take advantage of a link, the scheduling algorithm needs to establish a logical link as soon as the physical link appears, and unscheduled it as soon as it disappears from the physical graph. We quantify the dynamic goodness of the scheduling algorithm by comparing the link duration between the physical and logical graphs. Results presented in Section 4 are normalized against the optimal case, i.e. the physical connectivity graph. The variables to be used in this paper are listed in Table 1. Table 1. Variables used in this paper c s L S = {S0, S2 , . . . , SL−1 } C = {C0 , C2 , . . . , CL−1 } N = {N0 , N2 , . . . , NL−1 } P = {(r, c)1 , . . .} {C}

a channel a slot number number of slots in a superframe state for each slot data frequency channel for each slot neighbor for each slot (can be NULL) list of potential neighbors (contains identifier and channel) list of neighbors self has a connection to

To be able to communicate, two nodes need to schedule a slot to one another. They hence need to communicate to agree which slot in the superframe to use, and which channel. We present two variants of the proposed scheduling mechanism. Aloha-based scheduling (Section 3.2) is a simple, canonical algorithm, in which neighbor nodes opportunistically discover each other and establish links.

206

A. Tinka, T. Watteyne, and K. Pister

Reservation-based scheduling (Section 3.3) builds upon that. By adding an explicit reservation channel, nodes discover each other faster, which is desirable in the presence of mobile nodes. 3.2

Aloha-Based Scheduling

Each of the L slots in the superframe is attached a state S, a channel C, and a neighbor N . There are five states: “Aloha”, “Transmit Connection Request”, “Receive Connection Request ”, “Transmit Data”, “Receive Data”. A slot is assigned a channel C and a neighbor N only in the latter four states. The “Aloha” state is the default. When establishing a unidirectional link from A to B, the scheduling algorithm causes a slot in A’s schedule to transition from “Aloha” to “Transmit Connection Request”, to “Transmit Data”. Similarly, the same slot in B’s schedule transitions from “Aloha” to “Receive Connection Request ”, to “Receive Data”. When both A and B’s slots are in the “Transmit Data” and “Receive Data” state, resp., data packets can be transmitted from A to B, once per superframe if exactly one slot is scheduled in the superframe. If A has no data to send, it sends an empty keep-alive message. While communicating, A monitors whether its data packets are acknowledged; B monitors whether it receives data at all. If for 5 consecutive superframes no data is successfully transmitted, the slot returns to the “Aloha” state; connection is then lost. Three types of packets move through the network: – Advertisement packets contain a list of “Receive Connection Request” slots of the sender node. This can be unseed by neighbors to know where it can be reached to establish a link. Each entry is a tuple (s, c) of slot and associated channel. Advertisements are broadcast and always exchanged on channel 0; – Connection Request packets are sent in response to Advertisements; they are unicast on one of the slots announced in the Advertisement (at the announced frequency channel, always different from channel 0); – Data packets flow over the slot when a link is established. Their content is determined by the application, but their successful transmission is monitored by the scheduling algorithm to detect stale links. An empty data packet is used as a keep-alive. Data packets are always sent on a frequency channel different from channel 0. Note that there are L slots in a superframe, each of which can be used for an independent link. That is, a independent state machine is running for each slot. IEEE802.15.4-2006 compliant radios can transmit on 16 independent frequency channels. We dedicate channel 0 exclusively to Advertisements, and channels 1-15 exclusively to Connection Requests and Data packets. Algorithm 1 presents Aloha-based scheduling in pseudo-code. It is executed by every node in the network. Upon startup (lines 1-4), all the slots are set to the “Aloha” state. The main loop (lines 5-49) iterates at each slot; different actions are taken according to the state of the slot. When in an “Aloha” slot, a node listens for Advertisements 90% of the time (on channel 0, lines 15-26), 10% of the time it transmits an Advertisement (lines 8-15).

A Decentralized Scheduling Algorithm for TSCH

207

Sending Advertisement packets. The idea of sending an Advertisement is for a node to announce different rendez-vous slot/channel tuples so that interested nodes can establish a link to it. When sending an Advertisement, a node converts all of its “Aloha” slots to the “Receive Connection Request ” state, and assigns each of those a random channel other than channel 0 (lines 10-13). It puts that list in an Advertisement which it sends on channel 0. It then waits to be contacted on one of the “Receive Connection Request ” slots it just announced. Receiving Connection Request packets. When reaching a slot in the “Receive Connection Request” state (lines 27-36), a node listens to the channel it has previously randomly picked and announced in its Advertisement. If it does not receive anything (lines 33-34), it converts that slot back to “Aloha” state. If it does receive a Connection Request (lines 29-32), it converts that slot to “Receive Data” state and records the identifier of the requester. Receiving Advertisement packets. When receiving an Advertisement (lines 15-26), a node learns about the presence of a neighbor and is given the opportunity to contact it. If it has no slot scheduled to that neighbor, it picks one of the slots announced in the Advertisement where itself is in the “Aloha” state, i.e. it picks a rendez-vous slot and channel. If case there are multiple slots which satisfy these requirements, it picks one of them randomly. It changes the state of that slot in its scheulde to “Transmit Connection Request ” (line 21), records the channel announced in the Advertisement (line 22), and the sender of that packet (line 23). Transmitting Connection Request packets. When reaching a slot in the “Transmit Connection Request” state (lines 36-42) a node sends a Connection Request to the neighbor recorded in N , at the channel recorded in C (line 37). If it receives an acknowledgment, it puts that slot in the “Transmit Data” state, and the logical link is established. If the Connection Request is not established (due to e.g. a collision or nodes moving apart), the slot is reset to the “Aloha” state. 3.3

Reservation-Based Scheduling

The Reservation-Based Scheduling protocol behaves like the Aloha-based protocol, with the following additions: – Slot 0 is a permanent rendez-vous slot, i.e. only Advertisements can be exchanged. Unlike other slots, Advertisements can be exchanged on any of the 16 available channels, in slot 0. Each node picks a channel on which it listens for Advertisements. Using slot 0 as a reservation slot gives nodes more opportunities to establish links to one another. – In their Advertisements, nodes also include the list of the neighbors they are connected to, and the channel those neighbors are listening on in slot 0. This means that nodes learn about their two-hop neighbors. – Each node maintains a list P of potential neighbors and the channel they are listening on in slot 0. This information is obtained by listening to Advertisements. Each node also maintains a list C of neighbors it is currently

208

A. Tinka, T. Watteyne, and K. Pister

connected to. The scheduling algorithm tries to get as many nodes from P to C. – A node only announces the even slots in its Advertisement. When the state of even slot i becomes “Transmit data” (resp. “Receive data”), the state of odd slot i + 1 is implicitly changed to “Receive data” (resp. “Transmit data”). This means that links are scheduled in pairs, one in each direction, establishing only bidirectional links. Algorithm 2 presents Reservation-based scheduling in pseudo-code. The part of p. 210 details the execution in slot 0, while the part on p. 211 focuses on the other slots.

4

Simulation Setup and Results

We use a Python-based simulator1 to model the mobility and RF propagation characteristics for a fleet of 25 mobile nodes. The superframe size was chosen to be 17 slots. The size must be co-prime with 16 in order to gain the benefits of the channel offset scheme; a relatively small superframe size was chosen to ensure that the scheduling constraints would be significant. -20

14 12

-50 received power, dBm

Mean node degree

8 6 4

-60 -70 -80 -90 -100 -110 -120

2

-130 50

100 150 200 Density, nodes per square kilometer

250

Fig. 2. Mean node degree vs density of nodes in simulated environment

4.1

Free space Sensitivity limit

-40

10

00

Link strength model

-30

-140 0 10

101

distance, m

102

103

Fig. 3. Received power from randomly chosen locations in simulated environment

Propagation Model

The design objective for the RF propagation model is to create a deterministic model which captures the variance of the distance-to-received-power relationship observed in empirical studies of static spatial configurations, while also providing plausible spatial correlation of link strength. Approximately 30% of the simulated environment is covered with obstacles. The radiated power from a transmitting antenna is attenuated by an inverse square law as it moves through “obstaclefree” space, but is attenuated by an inverse fourth power law as it moves through 1

As an on-line addition to this paper, the source code of the simulator is made freely available at http://lagrange.ce.berkeley.edu/

A Decentralized Scheduling Algorithm for TSCH

209

Algorithm 1. Aloha-Based Scheduling (executed by each node) 1: for each slot i ∈ L do 2: Si =“Aloha” 3: Ni =NULL and Ci =NULL 4: end for 5: loop 6: Go to the next slot i 7: if Si ==“Aloha” then 8: if uniform(0,1)< 0.1 then 9: Enumerate all other slots with state S ==“Aloha”, {sj } 10: for each of these slots do 11: S=“Receive Connection Request” 12: C =uniform(1,15) 13: end for 14: Send Advertisement listing these slots and their channels {(sj , cj )}, on channel 0 15: else 16: Listen for an Advertisement on channel 0 17: if Advertisement {(sj , cj )} received then 18: Find own set of slots {sk } which are of state Si ==“Aloha” 19: if {sk } ∩ {sj } = ∅ then 20: Choose common slot sn ∈ {sk } ∩ {sj } randomly 21: Sn =“Transmit Connection Request” 22: Cn set to the receiving channel for that slot, read from Advertisement 23: Nn set to the node that sent the Advertisement 24: end if 25: end if 26: end if 27: else if Si ==“Receive Connection Request” then 28: Listen for a Connection Request to self on channel Ci 29: if valid Connection Request received then 30: Send Acknowledgment 31: Si =“Receive Data” 32: Ni set to the ID of the requesting node 33: else 34: Si =“Aloha” 35: end if 36: else if Si ==“Transmit Connection Request” then 37: Send Connection Request on channel ti to node ci 38: if Acknowledgment received then 39: Si =“Transmit Data” 40: else 41: Si =“Aloha” 42: end if 43: else if Si ==“Receive Data” or Si ==“Transmit Data” then 44: if no successful communication for n consecutive superframes then 45: Si =“Aloha” 46: Ni =NULL 47: end if 48: end if 49: end loop

210

A. Tinka, T. Watteyne, and K. Pister

Algorithm 2. Reservation-Based Scheduling (for slot 0) 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28:

for each slot i do Si =“Aloha” Ni =NULL end for P=∅ C=∅ Go to the next slot i if this is slot i == 0 then if P = ∅ and uniform(0,1)< 0.1 then Choose (id, c) (identifier, channel) randomly from neighbors of interest in P Transmit Advertisement to node id on channel c if Acknowledgment received then set state of all advertised slots to S =“Receive Connection Request” end if else Listen for an Advertisement on channel C0 if Advertisement received then Send Acknowledgment If neighbor of interest, choose common slot n (similar to Algorithm 1) Sn =“Transmit Connection Request” Nn set to the ID of the node that sent the Advertisement Cn set to the receiving channel for that slot in the Advertisement end if end if end if Continues on p. 211

“obstacle” space. This “higher power attenuation” scheme is inspired by empirical models of the effect of foliage on line-of-sight transmission[15]. The foliage model and density of obstacles is intended to represent an outdoor estuarial environment similar to that encountered by the Floating Sensor Network project. The multipath effect of the signal reflecting off the ground is modeled. The reflection is assumed to result in a 180◦ phase change and no attenuation. The size of the simulated enviroment is modified as needed to yield desired node densities. The minimum and maximum densities are 25 and 250 nodes per square kilometer. Figure 2 shows the mean node degree (number of neighbors in the physical connectivity graph) for the different simulated densities. The bars represent the 95% certainty interval for the estimate of the mean. 4.2

Co-channel Interference Model

The interfering effects of two nodes transmitting on the same channel at the same time (usually called a “collision”) is one of the main constraints on the decentralized schedule.

A Decentralized Scheduling Algorithm for TSCH

211

Algorithm 2. [cont.] Reservation-Based Scheduling (for slots other than 0) 1: Continues from p. 10 2: 3: if this is slot i! = 0 then 4: if Si ==“Aloha” then 5: if uniform(0,1)< 0.1 then 6: Enumerate all other even slots with state S ==“Aloha”, {sj } 7: for each of these slots do 8: S=“Receive Connection Request” 9: C =uniform(1,15) 10: end for 11: Send Advertisement listing these slots channels {(sj , cj )}, on channel 0; also include all (id, c) tuples present in C 12: else 13: Listen for an Advertisement on channel 0 14: if Advertisement {(sj , cj )} received then 15: Add new possible neighbors to P using the information in the Advertisement 16: Find own set of slots {sk } with state S ==“Aloha” 17: if {sk } ∩ {sj } = ∅ then 18: Choose common slot sn ∈ {sk } ∩ {sj } randomly 19: Sn =“Transmit Connection Request” 20: Nn set to the ID of the node that sent the Advertisement 21: Cn set to the receiving channel for that slot in the Advertisement 22: end if 23: end if 24: end if 25: else if Si ==“Receive Connection Request” then 26: Listen for a Connection Request for self on channel C0 27: if valid Connection Request received then 28: Send Acknowledgment 29: Si =“Receive Data”; Si+1 =“Transmit Data” 30: Ni and Ni+1 set to the ID of the requesting node 31: else 32: Si =“Aloha” 33: end if 34: else if Si ==“Transmit Connection Request” then 35: Send Connection Request on channel Ci to node Ni 36: if Acknowledgment received then 37: Si =“Transmit Data”; Si+1 =“Receive Data” 38: Ni+1 set to the ID of the requesting node 39: Ni = to C and remove it from P 40: else 41: set slot i to “Aloha” 42: end if 43: else if Si ==“Receive Data” or Si ==“Transmit Data” then 44: if no successful communication for n consecutive superframes then 45: Si =“Aloha” 46: move Ni from C to P 47: Ni =NULL 48: end if 49: end if 50: end if

212

A. Tinka, T. Watteyne, and K. Pister

The IEEE802.15.4-2006 standard specifies required jamming resistance for interference coming from an adjacent channel (1 channel away) or an alternate channel (2 channels away), but does not specify a required resistance to interference on the same channel. The Texas Instruments CC2420 2.4 GHz IEEE802.15.4-2006 compliant transceiver[16] has a specified co-channel rejection of -3 dB; in other words, if node A receives a transmission from node B with p dBm power, and a simultaneous transmission from node C with (p − 3) dBm power, the transmission for B will be received correctly and the transmission from C rejected. We use this model for our simulation. Adjacent and alternate channel interference are not modeled in this simulation. 4.3

Node Mobility Model

Each node is modeled as a mobile device moving at a constant speed in the environment described above. The speed of each node is drawn from a uniform distribution over [0.8,1.2] m/s. Each node transmits at 1 mW using an isotropic antenna. The height of the antenna from the ground (used for the multipath calculations) is drawn from a uniform distribution over [0.7,1.3] m for each node. Node motion is controlled by a random waypoint procedure: nodes select a cardinal direction randomly, then a distance to move in that direction. When they reach their destination, they repeat the selection process. The nodes are confined to a square area with dimensions determined by the desired node density. Fig. 3 shows the received power for randomly located transmitter and receiver nodes in the simulated environment. 4.4

Static Metric: Relative Connectivity

The static connectivity test proceeds as follows: 1. Simulate 25 mobile nodes for 60 seconds; 2. pick a node and a superframe at random; 3. from the physical connectivity graph, count the number of unique edges incident to that node over the superframe (that is, the number of one-hop 1.0

1.0

Reservation-based Aloha-based

0.8 Connectivity fraction

Connectivity fraction

0.8

0.6

0.4

0.2

0.0 0

Reservation-based Aloha-based

0.6

0.4

0.2

5

10 Degree

15

20

0.0 0

5

10 Degree

15

20

Fig. 4. Unidirectional (left) and bidirectional (right) mean connectivity ratio vs degree

A Decentralized Scheduling Algorithm for TSCH

213

neighbors connected for at least 1 slot during the superframe); this is the degree of the node; 4. from the logical connectivity graph, find the number of outbound edges (for the unidirectional test), or find the number of neighbors with both an outbound and inbound edge (the bidirectional test); 5. the ratio of the logical connection count to the node degree is the connectivity ratio for the node. Fig. 4 shows the mean connectivity ratio vs. the node degree for 1250 simulations, for both unidirectional and bidirectional connections. The bars represent the 95% confidence interval in the estimate of the mean. The reservation-based algorithm outperforms the Aloha-based algorithm at almost all node degrees (the confidence intervals overlap for degree 1). The reservation-based algorithm has more resources allocated to neighbor discovery, and a successful advertisement/connection request exchange results in a bidirectional connection. For both algorithms, increased node degree results in decreased relative connectivity ratios. More local nodes means more collisions between Aloha advertisements, which reduces the effectiveness of neighbor discovery, and more cases of multiple nodes responding to an advertisement, resulting in collisions and lost connectivity. The superframes also fill up when more neighbors are present; since the superframe size is 17 slots, a node cannot have bidirectional links with more than 8 neighbors. The difference between the Aloha-based and reservation-based algorithm performance at high node degrees, however, demonstrates that both collisions and saturation must be significant. 4.5

Dynamic Metric: Link Durations

The dynamic link duration test proceeds as follows: 1. Simulate 25 nodes for 60 seconds; 2. pick a node and a superframe at random; 3. pick one of the edges on the physical connectivity graph incident to that node at random; this is the link we will test; 1.0

Reservation-based Aloha-based

Link lifetime fraction

0.6

0.6

0.4

0.4

0.2 0.00

Reservation-based Aloha-based

0.8

Link lifetime fraction

0.8

1.0

0.2

50

100 150 Density, nodes per square km

200

250

0.00

50

100 150 Density, nodes per square km

200

250

Fig. 5. Unidirectional (left) and bidirectional (right) mean link lifetime ratio vs density

214

A. Tinka, T. Watteyne, and K. Pister

4. count the number of consecutive superframes (forward and backward in time) in which this link is in the physical connectivity graph; this is the physical link duration; 5. starting from the beginning of the link’s lifetime, find the first superframe in which the link exists in the logical connectivity graph, either as a unidirectional link (the original node to the destination) or as a bidirectional link; 6. count the number of consecutive superframes until the link no longer exists in the logical connectivity graph; this is the logical link duration; 7. the ratio of the logical link duration to the physical link duration is the link lifetime ratio. Fig. 5 shows the mean link lifetime ratio vs. the density of the nodes in the simulated environment for 1250 simulations. The bars represent the 95% confidence interval for the estimate of the mean. The degree of the node is not well defined over many superframes, as the physical and logical connectivity change. While the static connectivity test could use the node degree as the independent variable, for the dynamic link duration test we use the node density as a surrogate. See Fig. 2 for the relationship between the mean node degree and node density. The dynamic performance also shows that the reservation-based algorithm outperforms the Aloha-based algorithm. Again, the Aloha-based algorithm is at a disadvantage, because its advertisement/connection request transactions build unidirectional links, not bidirectional links. At low densities, the ratio between the algorithms’ performances is roughly 2, which suggests that the unidirectional/bidirectional allocation difference dominates in this regime. But at higher densities, the difference between the two algorithms widens, which means other effects must be significant as well. The saturation effects at work in the connectivity tests are also significant in the dynamic case. Links can be broken by co-channel interference, if another pair of nodes begins transmitting at the same channel/slot as an existing link. Nodes that have many active links also have less vacant slots available to form new links. Saturation effects alone cannot explain the decreased performance at high density, however, since the Aloha-based algorithm’s performance decreases significantly more than the reservation-based algorithm’s performance. The reservation-based algorithm benefits when advertisements are exchanged frequently, because information about connected neighbors is carried by the advertisement packets. The reactivity of the reservation-based algorithm therefore increases at higher densities, as nodes learn about possible new neighbors more quickly. Because the advertisements in the reservation-based algorithm carry more information than the advertisements in the Aloha-based protocol, the reservationbased algorithm gains relative performance at higher node densities.

5

Conclusions and Future Work

In this paper, we present what is, to our knowledge, the first scheduling algorithm for Time Synchronized Channel Hopping networks which both is distributed and

A Decentralized Scheduling Algorithm for TSCH

215

which copes with mobile networks. The two variant algorithms are based on an advertisement and rendez-vous scheme: nodes continuously advertise their presence which allows neighbor nodes to discover and contact one another. A inactivity threshold mechanism is used to tear down previously established links. The algorithms are tuned for a network of 25 drifter nodes randomly moving inside a lake or river. Simulation results show, under realistic propagation and mobility models, the efficiency of the algorithms. Figs. 4 and 5 demonstrate that the reservation-based algorithm outperforms the Aloha-based algorithm in practically all density conditions. This suggests that in an environment with highly dynamic connectivity, including networks of mobile nodes, devoting additional resources to neighbor discovery and coordination pays off. The goal of the scheduling algorithms presented in this article is to establish two-way connections between neighbor nodes, subject to the constraints of the superframe structure and the physical connectivity. We did not make assumptions about what kind of data is sent over the links; the latency, throughput, and reliability requirements are not specified. These scheduling algorithms could be adapted to meet either pre-determined or dynamic provisioning requirements. For example, a pair of nodes that need to exchange a large amount of data might wish to schedule more than one transmission slot per superframe. Many wireless sensor network applications are highly energy-constrained. Our scheduling algorithms, as described here, require the radios to constantly either receive or transmit. This would consume too much power for many applications. An obvious modification is to reduce the duty cycle of the Aloha coordination activities; the algorithms could be implemented exactly as written, while only performing Aloha listen/transmit actions on a subset of the idle slots. The obvious tradeoff is between the energy consumed for Aloha coordination versus the reactivity of the network to changes in the physical connectivity graph. Further work could focus on characterizing the rate of change of the connectivity graph, and determining a method for balancing power consumption and reactivity.

References 1. Tinka, A., Strub, I., Wu, Q., Bayen, A.M.: Quadratic Programming based data assimilation with passive drifting sensors for shallow water flows. International Journal of Control (to appear, 2010) 2. IEEE Standard for Information technology - Telecommunications and information exchange between systems - Local and metropolitan area networks - Specific requirements. Part 15.4: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications for Low-Rate Wireless Personal Area Networks (WPANs), IEEE Std., Rev. 2006 (September 8, 2006) 3. IEEE P802.15.4e/D0.01 Draft Standard for Information technology Telecommunications and information exchange between systems Local and metropolitan area networks Specific requirements Part 15.4: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Specifications for Low-Rate Wireless Personal Area Networks (WPANs) Amendment 1: Add MAC enhancements for industrial applications and CWPAN, IEEE P802.15 Working Group for Wireless Personal Area Networks (WPANs) Std. IEEE Std 802.15.4e, Rev. D0.01/r3 (September 13, 2009)

216

A. Tinka, T. Watteyne, and K. Pister

4. Polastre, J., Hill, J., Culler, D.: Versatile Low Power Media Access for Wireless Sensor Networks. In: Second ACM Conference on Embedded Networked Sensor Systems (SenSys), November 3-5, pp. 95–107. ACM Press, Baltimore (2004) 5. Buettner, M., Yee, G.V., Anderson, E., Han, R.: X-MAC: A Short Preamble MAC Protocol for Duty-Cycled Wireless Sensor Networks. In: 4th International Conference on Embedded Networked Sensor Systems (SenSys), October 31-November 3. ACM, Boulder (2006) 6. Ye, W., Silva, F., Heidemann, J.: Ultra-Low Duty Cycle MAC with Scheduled Channel Polling. In: 4th ACM Conference on Embedded Networked Sensor Systems (SenSys), November 1-3, pp. 321–334. ACM, Boulder (2006) 7. Watteyne, T., Mehta, A., Pister, K.: Reliability Through Frequency Diversity: Why Channel Hopping Makes Sense. In: 6th ACM International Symposium on Performance Evaluation of Wireless Ad Hoc, Sensor, and Ubiquitous Networks (PE-WASUN), Tenerife, Canary Islands, Spain, October 26-30 (2009) 8. Watteyne, T., Lanzisera, S., Mehta, A., Pister, K.: Mitigating Multipath Fading Through Channel Hopping in Wireless Sensor Networks. In: IEEE International Conference on Communications (ICC), May 23-27. IEEE, Cape Town (2010) 9. Pister, K., Doherty, L.: TSMP: Time Synchronized Mesh Protocol. In: Parallel and Distributed Computing and Systems (PDCS), Orlando, Florida, USA, November 16-18 (2008) 10. Doherty, L., Lindsay, W., Simon, J.: Channel-Specific Wireless Sensor Network Path Data. In: 16th International Conference on Computer Communications and Networks (ICCCN), August 13-16, pp. 89–94. IEEE, Turtle Bay Resort (2007) 11. HART Field Communication Protocol Specifications, Revision 7.1, DDL Specifications, HART Communication Foundation Std. (2008) 12. Gustafsson, D.: WirelessHART - implementation and evaluation on wireless sensors. Master’s thesis, Kungliga Tekniska h¨ ogskolan (April 2009) 13. Song, J., Han, S., Mok, A.K., Chen, D., Lucas, M., Nixon, M., Pratt, W.: WirelessHART: Applying wireless technology in real-time industrial process control. In: IEEE Real-Time and Embedded Technology and Applications Symposium, pp. 377–386 (2008) 14. ISA: ISA-100.11a-2009: Wireless Systems for Industrial Automation: Process Control and Related Applications, International Society of Automation Std. (September 11, 2009) 15. Oestges, C., Montenegro Vollacieros, B., Vanhoenacker-Janvier, D.: Radio Channel Characterization for Moderate Antenna Heights in Forest Areas. IEEE Transactions on Vehicular Technology 58(8), 4031–4035 (2009) 16. CC2420, 2.4 GHz IEEE 802.15.4 / ZigBee-Ready RF Transceiver (Rev. B), Texas Instruments, Inc. (March 20, 2007), data Sheet SWRS041B [available online]

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs Rodolfo de Paz and Dirk Pesch Nimbus Center for Embedded Systems Research Cork Institute of Technology Rossa Avenue, Cork Ireland {rodolfo.depaz,dirk.pesch}@cit.ie

Abstract. The current specification for IEEE 802.15.4 beacon-enabled networks does not define how active and sleep schedules should be configured in order to achieve the optimal network performance in all traffic conditions. Several algorithms exist in the literature that dynamically vary these schedules based on traffic load estimations. But it is still uncertain how these adaptive schemes perform with regard to each other as their performance has only been compared with the standard beacon mode. In this paper, we compare the current state-of-the-art schemes, and with the objective of overcoming the performance deficiencies shown by previous approaches, we introduce DCLA, an adaptive duty-cycle scheme for IEEE 802.15.4 beacon-enabled Wireless Sensor Networks (WSN) that employs a reinforcement learning technique. Simulation results show that the proposed scheme achieves the best overall network performance for a wide range of traffic conditions and performance parameters when compared with existing IEEE 802.15.4 duty-cycle adaptation schemes. Keywords: Wireless Sensor Networks (WSNs), IEEE 802.15.4, duty-cycle, energy efficiency, machine learning, reinforcement learning.

1 Introduction A Wireless Sensor Network (WSN) is a collection of wireless sensor nodes deployed to measure and report through collaboration certain parameters such as temperature, pressure, humidity, etc. These nodes are usually battery operated and, in most of the deployments, replacing or recharging their batteries is infeasible. Consequently, the power consumption is considered as a primary requirement for WSN communication protocols. Specifically, at the medium access control (MAC) layer a balance needs to be struck between achieving high quality radio resource allocation and energy expenditure. For this, the idle listening problem must be solved as it has been identified as one of the major sources of energy expenditure. This is caused when nodes do not know when the data traffic is generated from other nodes and its transceiver continuously stays in the receiving mode even when there is no data traffic for them. The IEEE 802.15.4 standard [1], which is currently the most commercially adopted MAC protocol for WSNs, specifies the beacon enabled mode for energy efficient operation. This mode is designed to support the transmission of beacon frames from J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 217–232, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

218

R. de Paz and D. Pesch

coordinator to end devices allowing node synchronization. Synchronization allows devices to sleep between coordinated transmissions avoiding idle listening, which results in prolonged network lifetimes. The beacon mode employs the superframe structure depicted in Fig. 1. Its format is based on two fundamental parameters: the Beacon Interval (BI), which defines the time between two consecutive beacon frames, and the Superframe Duration (SD), which defines the nodes’ active period in the BI. The superframe duration is composed by a contention access period (CAP) in which all devices use a slotted CSMA/CA protocol to gain access for time slots and a contention free period (CFP) for QoS demanding applications. The coordinator can introduce an inactive period to reduce energy consumption by choosing BI > SD. BI and SD are determined by two parameters, the Beacon Order (BO) and the Superframe Order (SO), respectively, as follows: BI = aBaseSuperframeDuration · 2BO ⎫⎪ ⎬ SD = aBaseSuperframeSuration · 2SO ⎪⎭

for 0 ≤ SO ≤ BO ≤ 14

(1)

where aBaseSuperframeDuration is a parameter defined by the standard that depends on the data rate and frequency employed and denotes the minimum duration of the superframe which corresponds to SO = 0.

Fig. 1. Superframe structure of the IEEE 802.15.4 beacon enabled mode

As the energy savings in the beacon-enabled mode depend on the amount of periodic sleep periods introduced, it is important to control the fraction of the time that the node is active. This time, known as duty-cycle (DC), can be computed as the ratio between the superframe duration and the beacon interval that can be related to BO, SO as follows: DC =

SD = 2 SO−BO BI

(2)

The smaller the duty-cycle the lower is the energy consumption. However, due to the inherent resource constrictions of sensor nodes, a small duty-cycle can cause buffer overflows and delays. On the other hand, although a high duty-cycle enables end devices to transmit a higher number of data frames and decrease delay it also increases the time the coordinator spends in idle listening. Consequently, duty-cycle adaptation is necessary to enhance the performance of IEEE 802.15.4 beacon-enabled networks. Several works in the literature have been conducted for adjusting the duty cycle to the traffic load using the IEEE 802.15.4 beacon-enabled MAC. For simplicity, these works adapt the duty cycle fixing SO and adapting BO to traffic [2][3] or fixing BO

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs

219

and adapting SO [4][5]. However, we observed that, with these techniques, if SO is fixed to a low value, the WSN produce lower network throughput than a higher SO value with the same duty-cycle ratio. Alternatively, algorithms that fix BO can suffer from unnecessary beacon overheads in idle WSNs. This motivated us to explore a new duty-cycle adaptation algorithm that could jointly adjust BO and SO values to find the optimal network performance. Also, as no comparison has been carried out among the different IEEE 802.15.4 adaptive duty-cycle adaptation schemes proposed by the community, there is still a doubt on how these approaches performs with regard to each other for different traffic loads. Therefore, it is also the objective of this work to carry out a performance comparison of previous IEEE 802.15.4 duty-cycle adaptation algorithms. For the main goal of finding the optimal duty cycle, we have decided to employ a reinforcement learning (RL) technique known as Q-learning. RL is a machine learning (ML) approach that finds the optimum value through trial-error iterations. Within the ML field, reinforcement learning (RL) is the most widely used method for solving WSN problems as they incur only minimal communication overhead and achieve optimal results [6]. In particular the Q-learning algorithm is well suited to this problem as it does not require any prior information about the environment. Additionally, the only memory and computation requirements for the nodes are the values of the possible actions the algorithm can take. We thus propose in this paper a duty-cycle learning adaptation algorithm (DCLA) for IEEE 802.15.4 beacon-enabled networks that employs the Q-learning technique. Our simulation results show that the proposed algorithm outperforms the existing state-of-the-art schemes in terms of average drop rate, throughput, energy efficiency and end-to-end delay. The rest of this paper is organised as follows. In Section 2, related works on dutycycle adaptation schemes for IEEE 802.15.4 beacon-enabled networks are discussed. In section 3, we introduce the basis of the reinforcement learning technique employed. Section 4 describes in detail the design of the proposed duty-cycle learning algorithm which includes traffic estimation, design of the RL agent and the algorithm for BO and SO joint adaptation. Section 5 presents simulation results that accentuate the distinct advantages of the proposed approach when compared with the existing schemes. Finally, the conclusions are drawn and future works are proposed in Section 6.

2 Related Work Recently, the performance trade-off in IEEE 802.15.4 beacon-enabled networks between power consumption, reliability and delay has produced attention within the research community. Inline with this, several MAC layer duty-cycle adaptation methods that try to balance those objectives with the ultimate purpose of increasing life-time have been proposed. We discuss in this section the four IEEE 802.15.4 dutycycle adaptation schemes found in the literature. The earliest work in 802.15.4 duty-cycle adaptation is known as the Beacon Order Adaptation Algorithm (BOAA) [2]. BOAA fixes SO to zero and adapts the beacon order using a coordinator buffer matrix B(nED,lb). The matrix records end devices’ number of received messages nED at coordinator during a number of beacon intervals lb. At the beginning of the network operation the buffer matrix is assumed to be empty

220

R. de Paz and D. Pesch

and a counter is initialized. Every beacon interval the number of messages received by every end device is stored in the matrix rows. After a lb number of beacon intervals the maximum number of messages nmax from any of the end devices is extracted from the B matrix. Then, BO is increased or decreased depending if nmax is lower or greater than a pre-fixed lower or an upper bound. The B matrix gives memory to the algorithm as the messages are collected by a lb number of beacon intervals. Although this is a nice feature, this matrix might not be scalable for large networks as the number of rows increases with the number of end devices in the topology. Also, BOAA does not work well if the traffic is not uniformly distributed among the sensor nodes as the algorithm only takes into account the device with the highest number of messages sent. Aware of this, B. Gao and C. He proposed an Individual Beacon Order Adaptation Algorithm (IBOAA) [3]. Their proposal aims to improve BOAA in networks where the traffic is not uniformly distributed. With this aim, the algorithm adapts the beacon interval of each end device individually based on a traffic queue flag embedded by nodes in one of the reserved bits of the standard MAC header. Because of the difference in the lengths of the individual beacon intervals, a minimal beacon interval BOmin is fixed in the coordinator as the reference period of any possible value of the end devices’ beacon intervals BIi. Moreover, the algorithm sets a common maximum value BOmax to avoid unbearable delays. Therefore any BIi set for a sensor device is limited in the range [BOmin, BOmax]. In most scenarios, this imposes the limitation for coordinator node to be plug powered as for low values of BOmin the coordinator cannot sleep even if the sensors become idle, while setting a high BOmin value would cause the loss of many packets in congestion situations. Finally, it is worth noticing that this scheme introduces extra control overhead in the beacon payload compared to the other approaches since the coordinator has to indicate the BO increase or decrease (1 byte) and the address list of end devices which BOs are going to be modified (variable size). Authors in [4] propose an adaptive MAC for efficient low power communications, named as AMPE for the rest of the paper. A coordinator running AMPE fixes BO to a maximum BOmax and adapts SO to the superframe occupation. Then, it estimates the superframe utilization by performing CCA measurements every 30 beacon intervals or whenever the number of received packets is halved or doubled. Then SO is increased or decreased depending on comparisons between a fixed threshold and the measured superframe utilization. CCA measurements may give a more accurate value of the superframe occupation than estimations, but they would imply a great deal of work to the coordinator in traffic variant scenarios where the measurement needs to be frequently activated. On the other hand, in scenarios where a small gradual increase of the traffic exists, the estimation updates could be too infrequent which in turn would decrease the performance of the network. Finally, J.Jeon et. al. introduced a Duty Cycle Adaptation (DCA) Algorithm for 802.15.4 beacon-enabled WSNs in [5]. DCA assumes BO constant and adapts SO to the traffic according to superframe estimations. In order to gather the end devices traffic information, the scheme modifies the reserved frame control field in the MPDU header as was also proposed by IBOA. However, in this case the algorithm also embeds the queuing delay. Based on the collected information the DCA coordinator estimates the number of packets being queued in the end devices and varies SO accordingly. This is the first scheme that takes into account delay measurements in the

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs

221

DC adaptation, however, a problem arises in idle networks as DCA does not adapt SO if no packets are received from end devices and it is limited to a fixed BOmax equal to seven due to the manner the superframe occupation is estimated in the scheme.

3 Reinforcement Learning The idea behind designing any learning system is to guarantee robust behavior without the complete knowledge, if any, of the system/environment to be controlled. In reinforcement learning (RL) an agent takes actions and learns from its environment through the rewards received. A crucial advantage of RL compared to other machine learning approaches is that it requires no information about the environment except for the reinforcement signal. This is especially appropriate to embedded systems such as wireless sensor networks were resources are scarce. The standard reinforcement-learning model is depicted in Fig. 2. An agent must learn the best behavior, formally called policy, through trial and error interactions with a random environment. At each step, the agent selects some possible action, a, and receives an immediate reward, r, from the environment for the current state s. If delayed reinforcement is employed, the action will not only affect the immediate reward but also future rewards. The agent then chooses an action a following a defined policy π(s) to generate an output. The action changes the state of the environment, and the value of this state transition is communicated again to the agent through the reinforcement signal r. The process is then repeated. The agent's objective is to choose actions that tend to increase the long-run sum of values of the reinforcement signal.

Fig. 2. The standard reinforcement learning model

Formally, the model is described by the experience tuple (S, A, T, R), where S is a discrete set of environment states s1, s2,…,sn, A is the set of possible actions from each state a1, a2,…,am, T(s,a,s’) is the transition probability from state s to a successor state s’ when taking action a and R(s,a) is the reward function. The reward function in delayed reinforcement is computed as a combination of the immediate reward r plus future discounted rewards as follows: R(s, a) =

∑γ r t

t

t

= r0 + γr1 + γ 2 r2 ...

(3)

222

R. de Paz and D. Pesch

The discount factor γ is a number in the range [0..1] and is used to weight near term reinforcement more heavily than distant future reinforcement. The closer γ is to 1 the greater the weight of future reinforcements is. The agent’s rule for selecting actions, which is the policy π(s,a), is a mapping from each state s and action a to the transition probability T of taking action a in state s. The agent’s goal is to find a policy π that maximizes the expected return by either maximizing the state-value Vπ(s) or action-value Qπ(s,a) functions.

{

V π (s) = E π R s, π

{

}

Q π (s, a) = Eπ R s, a, π

(4)

}

(5)

The state-value function Vπ(s) estimates the expected return when starting from state s and following π thereafter whereas the action-value Qπ(s,a) is the function that estimates the reward of taking action a in state s following policy π. Both functions can be related by means of the transition probability and reward functions as follows: Q(s, a) = r0 + γ

∑

s'∈S

T(s, a, s' ) ⋅ V(s' )

(6)

In the literature, several researchers have applied reinforcement schemes to design WSN communication protocols. Among them, RL-MAC [7] applies reinforcement learning to adjust the sleeping schedule of a MAC protocol in a WSN setting. The MAC protocol is similar in its idea to DCLA and it is therefore worthwhile explaining their differences. RL-MAC employs a time frame based structure similar to S-MAC [8]. Time is divided into frames while each frame is further divided into time slots and updating schedules is accomplished by sending SYNC packets similar to the beacon frame of the IEEE 802.15.4 MAC. However, as opposite to the standard, if multiple neighbors want to transmit to a node, they contend for the medium using a mechanism similar to that in IEEE 802.11, i.e., using RTS (Request To Send) and CTS (Clear To Send) packets. Although RTS/CTS can alleviate the hidden terminal problem, it incurs high overhead (40% to 75% of the channel capacity [9]) in WSNs because data packets are typically very small. Overhead has been identified as one of the sources that cause energy inefficiency in MAC protocols and for this reason the 802.15.4 MAC does not allow the transmission of RTS/CTS packets. However, this means that the current standard needs of some upper layer mechanism to avoid beacon collisions and allow multihop support. Adapting nodes’ duty-cycles while supporting scheduled multihop IEEE 802.15.4 communications is therefore a whole and complex new problem out of the scope of this paper. Also, some differences also exist in the manner the both RL agents are defined. In the RL-MAC, the agent which runs in every node, employs the number of packets queued for transmitting at the beginning of the frame as the state and the reserved active time as the action generated. The reward function depends on the number of waiting messages on the nodes and on the number of successfully transmitted messages during the reserved slot. However, in our case the agent is only present in the coordinator which has to rely on the information collected from end devices to derive

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs

223

the optimal policy. This scenario is challenging in the way that coordinator cannot employ the local information of its transmitting queue to build the reward function but on some estimation of end devices traffic status. Further, the DCLA agent employs a heterogeneous reward function that aims to improve the multiple goals present in WSNs as it is explained in the following section.

4 DCLA Protocol Design In this section, the design of the DCLA protocol is explained in detail. Specifically, we discuss the traffic estimation used by coordinator; we formulate the actions, states, reward and policy functions employed by the agent to find the optimum duty-cycle and we describe the algorithm for the selection of Beacon and Superframe Order values. 4.1 Coordinator Traffic Estimation The coordinator node needs to employ some estimation of the end devices’ traffic requirements to find the optimal duty-cycle. For this, the number of received messages during the active period is collected by the coordinator. On the other side, end devices embed their transmit queue occupation and delay values in the MAC header of sent data frames. The number of messages m received from end devices is employed by coordinator to estimate the superframe utilization SFu, which can be defined as the ratio of the superframe utilized for data communication as follows: ⎛ SD − Tbeacon SFu = ⎜⎜ ⎝ m ⋅ Ts

⎞ ⎟ ⎟ ⎠

(7)

This estimation is calculated by dividing the total time available for data transmission, which is the superframe duration SD minus the time Tbeacon spent for beacon transmission, with the time portion of the superframe m·Ts end devices employed to transmit the m messages received by coordinator. The time used by coordinator to transmit a beacon is calculated by summing the 802.15.4 MAC and physical overhead (30 symbols) plus the beacon payload. Alternatively, the superframe portion utilized by a sensor node to send one message, denoted as Ts, is estimated as follows: TS = TCCA + TDATA + TIFS + TACK

(8)

Where the time for clear assignment TCCA is two backoff slots (40 symbols); the time to transmit the data frame TDATA is the standard 802.15.4 MAC and physical headers (38 symbols) plus the payload length; the inter-frame spacing TIFS is the sum of the turnaround time (12 symbols) plus the time to find the next backoff boundary (up to 20 symbols); and the TACK is the time to receive the acknowledgement frame from coordinator (equal to 22 symbols). This estimation does not consider the time the node is in backoff because during this time the radio is in idle mode and others are free to use the superframe to transmit.

224

R. de Paz and D. Pesch

Fig. 3. The queuing occupancy and delay bits are embedded inside the reserved field of the MAC frame control field

On transmitter side, end devices embed the queue occupation and delay in the MAC header of the first frame sent within the superframe. These values represent the packet accumulation and delay caused by the last duty-cycle selection. Specifically, this information is embedded in the 3 reserved bits of the frame control field (as per Fig. 3). Two bits are used by the queue occupancy (O) to indicate 4 different levels whereas the queuing delay (D) is divided into 2 levels. The queuing delay bit indicates whether the delay is above ‘1’ or below ‘0’ an application defined delay threshold (δ). 4.2 The DCLA Agent The DCLA agent is assumed to be run by a coordinator that does not have any knowledge with regard to the traffic generated by end devices. The coordinator should therefore determine the optimal duty-cycle without any prior knowledge of the environment by employing a RL technique known as Q-Learning [10]. In our agent, we select the energy saving level A , defined as the difference between BO and SO values, as the state, and the actions are represented by the possible state transitions. Then, let L be the set of states {A 0 , A 1 ,..., A n } of our DCLA approach and let ρ be the number of possible states as follows: L{A 0 , A1,...,An } =

ρ

∑2

−A

ρ ≤ 14

(9)

A =0

The number of states ρ must be lower to 14 due to the limitation of the SO and BO values in the 802.15.4 standard specification. The energy saving level A of a sensor node can be computed at any time by employing e.q. (3) as follows: DC = 2SO − BO = 2−A → A = BO − SO

0≤A≤ρ

(10)

Thus, A represents the level of energy the sensor node is saving which is inversely proportional to the activity of the node (duty cycle). A high A value means that the duty cycle is low and the node is thus saving batteries. Alternatively, a low A value implies that the battery will be rapidly depleted. When the DCLA agent moves to a new state A , its corresponding duty cycle can be calculated using equation (10) whereas BO, SO values can be selected following the algorithm described in section 4.3. In Q-learning, the objective is to find the optimal policy π*, by means of delayed reinforcement. The policies and the value function are represented by a twodimensional lookup table, known as the Q table, indexed by state-action pairs. During

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs

225

the Q-learning process, a learned action value function Q directly approximates Q* through value iteration. Mathematically, the optimal policy and Q value are defined as: π* (s) = arg max(Q* (s, a)) a∈A(s)

Q*(s, a) = R(s, a) + γ

∑

s'∈S

T(s, a, s' ) ⋅ max(Q*(s' , a' )) a'∈A ( s')

(11) (12)

And the rule to update the Q value after each time step t that the agent takes is: Qt +1 (s, a) = Qt (s, a) + α[R(s, a) − Qt (s, a)]

(13)

As we can see, every new Q value is computed as the sum of the old value and a correction term. This term consists of the total received reward and the last Q value. The learning rate α is a number in the range [0..1] that prevents the Q values from changing too fast and thus oscillating. This value is set to 0.1 in our tests. For the DCLA agent, we employ one step Q-learning technique with discount factor γ equal to 0.5. Following the general delayed reinforcement definition of the reward function (see equation 3), the total received reward is obtained as the sum of the immediate reward rt and the discounted future rewards (Q values in this case). Due to sensor nodes’ memory constraints we only account for the immediate and next future reward as follows: R(s, a) =r t + γ max Q(s t +1 , at +1 ) a'∈A(s')

(14)

The design of the immediate reward rt is crucial as it allows the DCLA agent to learn the optimal duty-cycle. In our case, the agent objective is to improve the general overall WSN performance, which can be viewed as the sum of multiple goals such as: reduce energy consumption, increase throughput, decrease end-to-end delay and packet drop. To reflect this aim, we define a heterogeneous reward function in which the immediate reward is defined as the sum of four components: the energy re, superframe utilization ru, delay rd, and queue occupation ro rewards as shown in equation 15. The presence of positive and negative reinforcement in the continuous learning algorithm guarantees that the learner will not get stuck in local optimum. ⎧re ⎪ ⎪ru rt = re + ru + rd + ro = ⎨ ⎪rd ⎪ ⎩ro

= (1 − DC) = SFu = −D

(15)

= O(t − 1) − 2 ⋅ O(t)

The energy reward re represents the amount of energy that the network saves through the duty-cycle selection. The lower is the duty-cycle the higher is the energy savings for the nodes. This reward is thus defined as a positive function that depends on the duty-cycle value DC as per equation 15. The utilization reward ru is equal to the ratio of the superframe utilized for data communication as calculated in equation 7. This is also a positive reward as the greater the superframe utilization SFu is, the less time the coordinator is assumed to be in idle listening.

226

R. de Paz and D. Pesch

On the other hand, the delay reward rd is calculated as a negative function by averaging the delay bits D received from end devices during the last active period. An increase in the delay thus represents a penalty from the environment in the computation of the new Q-value. A final reward is introduced to avoid possible packet drops due to queues overloads. For this, we control end devices’ transmit queues occupation by averaging the received O values in the headers of the data frames sent by end devices. The resulting value is used to calculate the occupation reward ro as the difference between the occupation in the last (t-1) and current t steps of the algorithm. This reward can therefore oscillate among positive and negative values depending on previous and current queue occupation. It is seen as a penalty function if the number of queued packets has increased during the last beacon interval, whereas will be a positive reward function if lesser packets have been queued during the last interval. We next discuss how the DCLA agent selects the next action to take. As in the Q-algorithm the best possible action is never known a-priori, the agent starts with an empty Q-table and it starts trying several actions to learn the optimal value. The most straight forward rule for selecting actions would be the greedy policy, which selects the action that maximizes the Q-value for the current state. However, if this policy is employed, the algorithm could get stuck at a local optimum before finding the global optima. In order to overcome this situation, an exploration strategy is defined within the DCLA policy. The idea is thus to design a policy that can find an optimal balance between exploring and exploiting as the training progresses by employing the increasing amount of learnt knowledge. DCLA initially maximizes the exploration by selecting actions randomly. Then, the responses from the environment are observed, action probabilities are updated based on that responses and the procedure is repeated until all the states in the Q table have been visited. When all these positions in the table are filled, the exploration rate is decreased to a probability ε in which the best action is chosen following an epsilon greedy policy. In our experiments, ε is set to 0.01. Thus, the DCLA algorithm follows the next policy: π(s, a) = ε ⋅ rand[Q(s, a)] + (1 − ε) ⋅ arg max a [Q(s, a)]

(16)

The pseudo-code for the Q-learning algorithm that is run by the DCLA agent every beacon interval to decide the optimal duty-cycle is defined as follows: FOR each <s,a> DO Initialize table entry: Q(s,a)←0 Observe current energy level state st = A t WHILE (true) DO Select action at according to π(s,a) and execute it Observe the new state st+1 and receive the immediate reward rt Update table entry Q(st,at) as follows: Q(s t , at ) = Q(s t , at ) + α[rt + γ ⋅ max a Q(s t +1 , at +1 ) − Q(s t , at )]

Move to the next energy level state st+1= A t +1

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs

227

4.3 Beacon Order and Superframe Order Selection Immediately after the DCLA agent decides to move to a new energy saving state A t + 1 , the beacon order BO and superframe order SO need to be selected. The simplest way to do this, it is to fix one of both parameters and select the other according to equation 10. However, some observations lead us to take a slightly different approach. According to this equation, for a specific energy level A a total number of k = ρ − A combinations of SO, BO are possible. Simulations showed that, among them, those that produce low SO, BO values obtain lower network throughput and higher energy expenditure than combinations with higher SO, BO values and same duty-cycle. This realized that these results are basically due to two factors. First, the overhead of the beacon frame is more significant with low BO values since beacons are more frequent. Therefore, if this extra control overhead is not controlled may result in higher energy expenditure. Second, CCA deference is also more frequent with lower SO values, leading to more collisions at the start of each superframe. Let us explain this problem. Two or more nodes defer their transmissions when the remaining time in the current superframe is insufficient. All such nodes will start their CCAs immediately following the beacon frame. In the first two backoff periods, the channel will be found idle. Consequently, all the nodes will conclude that the channel is free and start their transmission in the third backoff period resulting in a collision. This contention problem is obviously more pronounced for small superframe values i.e. SO = 0 or 1 causing lower throughput and an increased number of retransmissions. However, we should be careful with increasing BO and SO values since average delays and packet drops can be proportionally increased. This is because high BO values lead to longer sleep times between beacons. We then decided to increase the BO and SO values only if the same duty-cycle is selected a consecutive number of times (denoted as B in Fig. 4) unless the average delay and/or queue occupation parameters are above some pre-defined threshold or the network is working at maximum dutycycle. Experimental tests suggested 5 as a proper value for this threshold. The exception of the maximum duty-cycle is because in this case BO is equal to SO which means there cannot be any harm in terms of delay or packet drop when increasing the beacon periodicity. Alternatively, BO and SO values are reduced while maintaining the same duty cycle, anytime delay or queue occupation is above the threshold as it means. This beacon order and superframe order selection allows devices to reduce beacon overhead and CCA deference without increasing delay or packet drop as results will show later. The flow chart for the BO/SO selection algorithm is depicted in Fig. 4.

Fig. 4. Flow chart for the Beacon Order and Superframe Order selection algorithm. B is incremented each time the DCLA agent selects the same duty cycle ( A t = A t −1 ).

228

R. de Paz and D. Pesch

5 Simulation Environment and Results Several simulations were carried out to compare DCLA performance with the state-ofthe-art IEEE 802.15.4 duty-cycle adaptation schemes described in Section 2. All the algorithms were implemented using as a base the Open-ZB IEEE802.15.4 simulation model [11] developed with OPNET Modeler [12] and following the 2.4GHz physical layer definition. The battery model was modified to more accurately estimate the radio chip consumed energy during the simulation. The computation was done by summing up the energy spent in the different power modes (transmitting, receiving, idle and power down) of the CC2420 datasheet [13]. The network topology used for our simulations is a star topology of MicaZ nodes [14]. The WSN is composed by a central coordinator surrounded by 8 equispaced sensor nodes. The coordinator node employs any of the described adaptation algorithms to determine the duty cycle and periodically broadcasts the result inside the beacon. The other nodes act as end devices without the capacity of beacon transmission. End devices periodically listen for beacons and generate data frames containing 40 bytes of payload following a Poisson distribution. Several tests, lasting 10 hours each, were carried out with different traffic conditions by varying the mean of the statistical distribution within a range from 0.1 to 10 seconds. End devices transmit the data frames during the contention access period (CAP) of the superframe using CSMA-CA channel access technique. If the transmitting queue is found empty, the end device goes into sleep mode to save energy. The contention free period of the superframe is not used for transmissions in this scenario. Every time the coordinator correctly receives a data frame an acknowledgement is sent to the end device as described by the 802.15.4 standard specification. In the proposed scenario the delay threshold δ for the WSN application running in the end devices is set to 1 second. BO and SO values are initially set to 7 and 0 respectively for all the duty-cycle adaptation schemes. The only exception for that is the individual beacon order adaptation (IBOA) in which the coordinator’s beacon order is set to 0 as it is used by end devices as the reference time. End Devices’ transmitting queues have a size of 1Kbyte which allows them to accumulate around 18 data frames. 5.1 Simulation Results Simulations results present a comparison between the state-of-the-art duty-cycle adaptation schemes and the proposed learning technique for duty cycle adaptation for different traffic conditions. The WSN performance metrics considered are: drop rate, throughput, end-to-end delay and energy efficiency. Fig. 5 shows the duty-cycle selected by a coordinator that employs the DCLA algorithm described in this paper. The resulting values are in the range 0.1-100% reflecting the different traffic loads generated. In addition, it can be clearly seen that the higher is the traffic load the higher is the duty cycle selected, as would be expected. Fig. 6 shows the average drop rate for all the different approaches. This parameter represents the average number of bits per second dropped in end devices transmitting queue. A frame in the simulator can be dropped due to the following reasons: (i) the frame cannot be inserted in the transmitting queue because it is already full; (ii) the device fails more than 4 times to acquire the channel, which corresponds to the default

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs

229

value for the macMaxCSMABackoffs attribute; (iii) the channel has been acquired but the data transfer attempt fails more than 4 times which is the default value for (1+macMaxFrameRetries). We only show in Fig. 6 the drop rate results for the case (i) as it is the situation that mostly depends on the duty-cycle selection. Inter-arrival times greater than 1 second are omitted in the figure as the number of dropped packets was nearly zero. On the other hand, when the nodes’ packet inter-arrival time is at its maximum (0.1sec) the network is overloaded. This means that end devices are not able to transmit all the generated data to coordinator even if it would be awake the 100% of the time. In this case, we have therefore high drop rates for all the schemes. But due to the proposed algorithm’s ability to control the occupation of end devices transmit queues through the optimum duty-cycle selection, the drop rate is much lower than with the other schemes. The figure also shows that the worst performance is given by AMPE and BOAA as they do not consider the end devices queues status for the dutycycle selection. Fig. 7 depicts the average throughput. DCLA also outperforms the other schemes studied here as it presents the maximum throughput for all the range of traffic loads presented. This is mainly due to two reasons, firstly the coordinator is able to find a near optimal duty-cycle thanks to the Q-learning algorithm employed, and secondly the BO, SO selection algorithm avoids low BO, SO values that could result in CCA deferences when occupation and queuing delay permits it. Fig. 8 shows the average end-to-end delay. This value is obtained as the delay in seconds since the frame was generated by the end device’s application layer until is received by the physical layer at coordinator node. As we explained in Section 4.2 , DCLA considers a reward component that accounts for application delay. The figure clearly shows how DCLA meets the 1sec delay imposed in our scenario. In this case DCA also shows good performance as it also considers queuing delays in the dutycycle decision process. Other algorithms increase their delays as the traffic in the network decreases as they try to minimize energy consumption without taking into account the delay requirement. Specifically AMPE presents the highest delay as it updates the duty-cycle decision after longer time intervals than the other schemes. We can say from the results that AMPE, IBOA and BOAA would not be suitable for WSN applications with specific delay bounds. Perhaps, the most important metric for the wireless sensor network is the energy efficiency. For its computation in the simulator, we calculate the number of bits correctly delivered to coordinator per Joule spent by WSN nodes’ batteries during each simulation run. The results for this performance parameter are shown in Fig. 9. DCLA presents the best energy efficiency when compared to the other approaches. This shows again that the learning algorithm in combination with the proposed BO/SO selection mechanism shows better performance than current state-of-the-art approaches. It is worth noticing that the energy efficiency results show the problem that IBOA presents in managing the coordinator duty-cycle. Because in this scheme the end devices can select different beacon intervals, a minimal beacon interval had to be set in the coordinator as the reference time period. This means that compared to the other approaches, the coordinator consumes much more energy as it cannot sleep even the traffic in the network is reduced being thus less energy efficient.

230

R. de Paz and D. Pesch

Fig. 5. DCLA average duty-cycle selection

Fig. 6. Average drop rate measured in bps

Fig. 7. Average throughput measured in Kbps

DCLA: A Duty-Cycle Learning Algorithm for IEEE 802.15.4 Beacon-Enabled WSNs

231

Fig. 8. Average end-to-end delay measured in seconds

Fig. 9. WSN energy efficiency measured in correctly delivered bits per mJ

6 Conclusion and Future Work This paper proposes a duty-cycle learning algorithm (DCLA) to enhance the overall network performance of IEEE802.15.4 beacon-enabled networks. We employ a reinforcement learning algorithm to solve the problem of adapting the coordinator’s dutycycle according to end devices’ traffic conditions with the objective of minimizing the energy consumption while balancing at the same time other important WSN performance parameters. Once the RL agent has reached the duty-cycle decision, beacon order and superframe order are selected trying to find a compromise among beacon overhead and queuing delay. Simulation results show that DCLA outperforms current state of the art on duty-cycle adaptations for IEEE 802.15.4 beacon-enabled networks in terms of average drop rate, throughput, energy efficiency and end-to-end delay. As a future direction of this work other scenarios with more rapid traffic fluctuations may be studied. In these situations and with the purpose of accelerating the RL adaptation process, it will be explored how the choice of different initial Q-values,

232

R. de Paz and D. Pesch

progress estimators or reward shaping could be used. We also think that an RL agent could be created at end devices to deal with collisions as this could reduce even more the energy expenditure. The new agent could control the minimum backoff exponent of the CSMA-CA according to the number of failed transmissions and the duty cycle chosen by DCLA. Finally, we also plan to expand this work to IEEE 802.15.4 multi-hop topologies by employing our solution on distributed beacon scheduling [15] with DCLA.

References 1. IEEE 802.15.4 Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications for Low-Rate Wireless Personal Area Networks, LR-WPANs (2007) 2. Neugebauer, M., Plonnigs, J., Kabitzsch, K.: A new beacon order adaptation algorithm for IEEE 802.15.4 networks. In: Proceedings of the Second European Workshop on Wireless Sensor Networks, pp. 302–311 (2005) 3. Gao, B., He, C.: An individual beacon order adaptation algorithm for IEEE 802.15.4 networks. In: Proceedings of the 11th IEEE Singapore International Conference on Communication Systems, pp. 12–16 (November 2008) 4. Barbieri, A., Chiti, F., Fantacci, R.: Proposal of an Adaptive MAC Protocol for Efficient IEEE 802.15.4 Low Power Communications. In: Proceedings of the IEEE Global Telecommunications Conference, pp. 1–5 (December 2006) 5. Jeon, J., Lee, J.W., Ha, J.Y., Kwon, W.H.: DCA: Duty-Cycle Adaptation Algorithm for IEEE 802.15.4 Beacon-Enabled Networks. In: Proceedings of the 65th IEEE Vehicular Technology Conference, pp: 110–113 (April 2007) 6. Di, M., Joo, E.M.: A survey of machine learning in wireless sensor networks. In: Proceedings of the 6th International Conference on Information, Communications and Signal Processing (ICICS), pp: 1–5 (2007) 7. Liu, Z., Elahanany, I.: RL-MAC: A reinforcement learning based MAC protocol for wireless sensor networks. International Journal on Sensor Networks, 117–124 (2006) 8. Ye, W., Heidemann, J., Estrin, D.: Medium access control with coordinated adaptive sleeping for wireless sensor networks. IEEE/ACM Transactions on Networks 12(3), 493–506 9. Polastre, J., Hill, J., Culler, D.: Versatile low power media access for wireless sensor networks. In: Proceedings of the Second ACM Conference on Embedded Networked Sensor Systems (SenSys), Baltimore, MD (November 2004) 10. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press, Cambridge (1998) 11. Jurčík, P., Koubâa, A.: The IEEE 802.15.4 OPNET simulation model: reference guide v2.0. IPP-HURRAY Technical Report, HURRAY-TR-070509 (May 2007) 12. OPNET Modeler, OPNET Technologies Inc. Version 15.0., http://www.opnet.com 13. CC2420 radiochip datasheet: 2.4Ghz IEEE 802.15.4 Zigbee Ready RF Transceiver. Texas Instruments, http://www.ti.com 14. Low Power 2.4Ghz MicaZ mote for Wireless Sensor Networks, http://xbow.com 15. Muthukumaran, P.S., de Paz, R., Špinar, R., Pesch, D.: MeshMAC: Enabling Mesh Networking over IEEE802.15.4 through distributed beacon scheduling. In: Zheng, J., et al. (eds.) ADHOCNETS 2009. LNICST, vol. 28, pp. 561–575 (2010)

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm for IEEE 802.16 Mesh Networks Yaaqob Ali. A. Qasem, Ali Z. Alhemyari, Chee Kyun Ng, Nor Kamariah Noordin, and Omar. M. Ceesay Department of Computer and Communication Systems, Faculty of Engineering, Universiti Putra Malaysia, UPM Serdang, 43400 Selangor, Malaysia {y_alrefaei,alizuhair10}@yahoo.com, {mpnck,nknordin}@eng.upm.edu.my, [email protected]

Abstract. Being an emerging wireless broadband technology with numerous prospects, Worldwide Interoperability for Microwave Access (WiMAX) not only provides high data rate but also the last mile solution for broadband wireless access (BWA). WiMAX enables both point to multi-point (PMP) and mesh topologies solution. In centrally controlled WiMAX mesh mode, all data packets are routed to the subscriber stations (SSs) via base station (BS). Thus, the BS link may be constantly congested with data traffic which impacts the performance of the system. The goal of this paper is to create an efficient multihops routing with suitable scheduling algorithm in WiMAX mesh networks (WMN). This algorithm is called energy/bit minimization routing and centralized scheduling (EbMR-CS) algorithm. Here, a routing tree is constructed based on the energy/bit minimization (EbM). This algorithm looks for a short path from the current node to BS, while the optimal path is achieved when the whole path has the lowest EbM. After the route is fixed, and the traffic demanded at each node is known, the total traffic arriving at a node is centrally scheduled such that the transmission interferences can be avoided. The proposed algorithm has considered some important design metrics such as fairness, reuse timeslot, balanced load, concurrent transmissions and hop count. The result shows that the proposed EbMR-CS algorithm has reduced the length of scheduling up to 43%. The system throughput and channel utilization ratio (CUR) have been enhanced up to 68% and 45%, respectively. Keywords: WiMAX mesh networks, fairness, CUR, EbM routing, centralized scheduling.

1 Introduction Advances in broadband wireless access (BWA) technologies over the years intensively stimulate the interest of customers [1]. The IEEE 802.16 family of technologies which is popularly known as Worldwide Interoperability for Microwave Access (WiMAX) is one such standard that renders various service types in last mile BWA solutions. Currently, this standard enables both point to multi-point (PMP) and mesh topologies solutions [2]. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 233–248, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

234

Y.A.A. Qasem et al.

In the mesh mode operaations, data packets are not necessarily routed through the base station. They may alternatively a be routed to the destination through otther intermediary nodes as messh routers. This reduces the base station (BS) link frrom becoming a network bottle neck in that mesh networks. They have self-organizatioonal capabilities in the case of link failure as shown in Figure 1. Hence, this WiMAX plementation of WiMAX mesh networks (WMN) [3], [44]. specification ushers the imp It could be recalled that routing r and link scheduling continue to pose problems to the improvement of the mesh h networks technology, which provokes the interestt of researchers. The authors in [5] have proposed a heuristic greedy scheduling model for mesh networks. They furth her gave an estimate of the achievable throughput in this scenario. Narlikar et. al. in [6] has discussed the various routing models proposedd by researchers for mesh modee applications. However, these models can only affordd an interference-free routing wiith scheduling algorithm to eliminate the interference in the network by utilizing a two-fframe period where half the links are activated in each fraame. A mathematical model has been presented for centralized scheduling, which the authors refer to as routing, channel and link scheduling (RCL) algorithm [7]. A croosslayer design for tree type routing, r level-based centralized scheduling and distribuuted power control to improve th he network throughput has been presented in [8]. Similaarly, various bandwidth allocatio on strategies have been proposed for WiMAX mesh m mode operation. One such researcch is from [9], where the authors employed matrix algeebra to construct collision avoid dance schedule of links to enhance system throughput. On the other hand, the authors in [10] differed in their approach. They argued that uppon mize constructing a routing treee, the path with the least number of hops will optim system throughput.

Fig. 1. IEEE 802.16 based WMNs

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm

235

The authors in [11] suggested that the path with the least number of blocking nodes improve system throughput. In recognition to the transmission requirements of realtime services, and attempting to improve efficiency, the authors in [12] achieved this goal by inducing adjustments to the key parameter of interest. Constructing a routing path in a short path algorithm was done by Han et. al. in [13, 14] and Peng et. al. in [15]; they select the neighbouring subscriber station (SS) with the least node ID (Identifier) number as parent node (PN). Similarly, Wang in [16] used the Breadth First Search (BFS) algorithm to construct the routing tree by firstly choosing the BS in the first level as the new routing tree’s root node and then selecting the neighbouring node by choosing the node with a small ID number. All the nodes in this routing may choose the same node with the least ID number from their neighbouring nodes, and this may cause series of interferences in the network. Moreover, in [16], the assignment of the transmission slots to the nodes is determined by the number of service tokens. It is crucial to note that a node is allocated a service token based on its packets. In the nearest-to-the BS model, which gives a better result than the farthest model in [16], the author did not consider other important parameters into the selection process such as the traffic load and the interferences. Kyasanur and Vaidya in [17] introduced a multi-radio network for the IEEE 802.11 network interface card. While one radio is kept fixed to one channel, the other radio keeps switching between the remaining channels for communication with the fixed channel of its neighbor nodes. Kodialam and Nandagopal in [18] proposed a WMN with the assumption that fast channel switching can be attained. With that in mind, the authors modeled a dynamic Channel Allocation algorithm; however, the algorithm resembles a fixed channel allocation algorithm. In this paper, an efficient multi-hops routing for WMN through suitable scheduling algorithm called energy/bit minimization routing and centralized scheduling (EbMRCS) is proposed. We focus on eliminating the collision packets and enhancing the performance of WMN system. Some important design metrics such as fairness, reuse timeslot, balanced load, concurrent transmission, relay model and hop-count are considered. The results show that the proposed EbMR-CS algorithm has reduced the length of scheduling up to 43% besides enhanced the system throughput and channel utilization ratio (CUR) up to 68% and 45% respectively. The rest of this paper is arranged as follows. Scheduling in IEEE 802.16 WMN is presented in Section 2. In Section 3, the interference in WMN is described. The proposed EbMR-CS algorithm is explained in Section 4. The system performances of the proposed EbMR-CS algorithm in terms of length of scheduling, CUR and system throughput are discussed in Section 5. This paper is concluded in Section 6.

2 Scheduling in IEEE 802.16 WMN To be precise, scheduling in WiMAX can be defined as an array of fixed length timeslots as shown in Figure 2. Each potential transmission is allocated some timeslots in a manner that higher priority is given to the SS with higher traffic demand. Thus, the SS with higher traffic demand needs a longer scheduling period and causes a transmission delay to the SS with lower traffic demand. For a mesh networks operating in the centralized mode, the BS is charged with the responsibility

236

Y.A.A. Qasem et al.

of determining the scheduling for the whole network. However, it is usually employed in the communication between BS and SS. An example of centralized scheduling in the WiMAX based mesh networks is illustrated in Figure 3.

Fig. 2. The IEEE 802.16 frame structure

The WiMAX based mesh networks supports both mesh centralised scheduling (MSH-CSH) and mesh distributed scheduling (MSH-DSCH) schemes. To allocate data (time) mini slots to the various SSs in either of these scheduling schemes, control packets are exchanged in the scheduling control sub frame [19]. However, the total number of possible transmissions per sub frame is a reconfigurable parameter as shown in Figure 2. While centrally controlled scheduling is most favoured for communication between BS and SSs which corresponds to internet traffic, distributed scheduling tends to favour SS to SS communication which corresponds to intranet traffic. When operating in centralized mode, all scheduling decisions are made at the BS to allocate the mini slots to all the SSs.

(a) Network topology

(b) Scheduling tree

Fig. 3. The network architecture; (a) network topology and (b) scheduling tree

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm

237

3 Interferences in WMN As IEEE 802.16 networks operate synchronously in a timeslot mode. It is also essential to allocate timeslots such that collision is avoided over the network to achieve the required bandwidth and high system throughput in each connection. A challenging issue in the IEEE 802.16-2004 WiMAX networks is that the routing and scheduling schemes are strongly coupled. This is particularly different from that in Wi-Fi based mesh networks, in which the medium access control (MAC) layer is contention-based. Their routing algorithms and MAC layer protocols can be separately designed and operated. To schedule two links at the same timeslot, the scheduling should be done in such a way that interference is avoided [20]. This particularly depends on the transmission and signalling mechanism between the nodes. Packet collision in the IEEE 802.16d WiMAX networks basically occurs in two distinct forms, namely primary and secondary interferences [21]. Primary interference is due to the half duplex nature of the transceiver. A node cannot transmit and receive simultaneously as shown in Figure 4(A) [13]. Besides that, a node also cannot transmit to or receive from multiple neighbour nodes at the same time as shown in Figure 4(B) [14]. Secondary interference occurs where the transmission of one link can be corrupted by the interference from a neighbouring link as shown in Figure 5 [22]. Figure 5 explains this case, when a receiver of node D tuned to a transmitter of node C is within the range of transmitter of node B which is transmitting to node A, D’s reception interferes with B’s transmission, though not intended for receiver D. Peculiar to multi-hops communication is the problem of contention for channel access by both the arriving and departing data packets. This is solved by equipping the nodes with a multiple transceivers system, with each tuned to a different channel, thus enabling simultaneous transmission and reception per node.

Fig. 4. Primary interference in WMN

Fig. 5. Secondary interference in WMN

4 Proposed EbMR-CS Algorithm The proposed EbMR-CS algorithm is presented in this section. Here, a routing tree is constructed based on EbM. This algorithm looks for a short path from the current node to BS, while the optimal path is achieved when the whole path has the lowest EbM. After the routing is fixed and the traffic demanded at each node is known, the total

238

Y.A.A. Qasem et al.

traffic arriving at a node must be determined for further transmission. This system design algorithm consists of four parts: namely network model, EbM routing algorithm, channel allocation and multi-transceivers scheduling algorithm as shown in Figure 6. In order to avoid interference and system bottleneck, a network model is constructed where both path loss (PL) and signal to noise ratio (SNR) are taken into consideration to maintain the communication link between the transmitter and receiver. Hence, the EbM routing algorithm is designed to find the optimal path from source to destination to create a balanced network. The multi-channel allocation is then implemented by using first order algorithm to increase the reuse timeslot and concurrent transmission. Finally, a multi-transceiver scheduling algorithm is added to enhance the performance of the system and avoid the collision in the network.

Fig. 6. The EbMR-CS algorithm methodology stages

4.1 Network Model The network topology is modelled as a directed graph G(V,E). An example of this topology is shown in Figure 3a, where V represents the set of nodes in the mesh cell {SS1, SS2, ........, SSn}, n is the number of nodes in the network, and E represents the set of edges between every SS and its PN that carry data. In wireless communication, the signal from transmitter suffers from PL attenuation as it traverses the network to the receiver. This PL is a function of the distance of separation, d of the nodes. Thus, we can calculate the PL using the NLOS equation [23] as follows

PL = 122.5 + 26.5 * log10 ( d )

(1)

If node p, for example, sends packets to node q, it is regarded successful if and only if the following condition holds [24].

PTx −10log10 (BW) + GTx + GRx− PL −10log10 (KTo) + NF > SNRthreshold

(2)

Where, PTx is the mean power at the antenna port, BW is the occupied bandwidth, GTx, GRx are the antenna gain for Tx and Rx respectively, NF is the receiver noise figure and KTo = -144 dBW/MHz = Equipartition Law. The SNR thresh is the minimum threshold below which the signal will not be received at node q. Table 1 shows the value of SNR threshold [2], which can be applied to calculate the SNRp,q at the receiver of every link. The SNR must always be kept above the QPSK ½ threshold for correct reception of data packets at receiver. Whenever the SNR falls below this threshold, the link disconnects immediately and

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm

239

link capacity nullified. In this way, a connectivity graph G(V,E) is derived using the links marked with its capacity, which is then used as the basis for routing-tree construction. Table 1. Receiver SNR threshold assumptions [2] Modulation

Coding Rate

Receiver SNR (dB)

BPSK

½ ½ ¾ ½ ¾ 2/3 ¾

6.4 9.4 11.2 16.4 18.2 22.7 24.4

QPSK 16-QAM 64-QAM

In the proposed algorithm, the nodes are randomly distributed. The link between any two nodes is selected. The PL and SNR of this link are measured. If the SNR is found to be larger than the SNR threshold, the link is connected. Otherwise the link fails. This procedure is repeated until all the nodes select their neighbours. In this way, the PL and SNR are used to reduce the number of link connectivity, which helps the routing algorithm to make optimal decisions for selecting their parents. 4.2 EbM Routing Algorithm The routing strategy is employed to transfer traffic from a node to the BS to determine which path is feasible. As such, only static routes are considered in this paper. Beginning with the BS, the SS nodes are added into the tree one by one. The routing tree is constructed after the connectivity graph is obtained. Since EbM model minimizes energy used per bit transmitted to the mesh BS, the overall energy consumption is kept to a minimum without any regard to the number of hops. In Wireless MAN/HIPERMAN systems, this function is handled by the Mesh Networks Configuration (MSH-NCFG) messages. The energy value eb(n) = eb(n,PN(n)) is the dissipated energy per unit data byte received by the parent node PN(n) from node n [24]. To compute the energy metric Eb(i) dissipated along the routing path from node i to Mesh BS, the following formula is used [24] Eb ( i ) =

∑

v ∈ path

e (v)

b (i)

(3)

The BS is chosen as the new routing tree’s root node. For each of the candidate subscriber node (CSN), n is the node with neighbours in the routing tree but as at now not in the tree themselves. Hence, the route to mesh BS is found by choosing the path with the smallest energy (E) and its parent node PN(n) is selected as follows

PN (n) = arg min{Eb (i ) + eb (n, i)}

(4)

This strategy typically produces a fairly high hop count to reach the BS. This results in using shorter hops with higher modulation complexities.

240

Y.A.A. Qasem et al.

After the topology is fixed, the algorithm can select only one PN to obtain the routing tree. Four metrics are of paramount significance to this algorithm instead of one used in [24]; these are minimum energy per bit (EpB), minimum collision metric, minimum hop-count and smallest ID number. However, the collision metric and hopcount metrics play important roles to maintain a balanced network. • In the algorithm, for each CSN, the parent node is selected from the upper level which has the minimum hop count to the BS. At initialization, there is only one sponsor node (SN), and hence this SN is selected to be the PN. (If the number of SN=1, then PN1=SN, where PN1 is the parent node with only one SN). • If there is more than one SN for each CSN, the PN with the minimum EpB is selected as the PN. (If the number of SN1 >1, then PNe = arg min{E b (i ) + eb (n, i )} , where SN1 is the number of SNs, and PNe is the parent node with minimum EpB). • If there are more than one SN having the same minimum EpB, the PN with the smallest collision metric (the number of neighbouring nodes) is selected as the PN. (If the number of SNe>1, then PNc = arg min{ Neigh ( SNe )} , where SNe is the number of the SNs that have equal minimum EpB, PNc is the PN having the minimum collision metric and Neigh(SNe) is the number of neighbouring nodes for each SNe). • If there are more than one SN that has the same minimum collision metric, the PN having the minimum hop-count to the BS is selected as the PN. (If the number of SNc>1, then PNh = arg min{hop( SNc)} , where SNc is the number of SNs having equal minimum collision metric, PNh is the PN with the minimum hop-count and hop(SNc) is the number of hop-count from the SNc to BS). • If there are more than one SN having equal minimum hop-count to BS, the SN with the smallest ID number is selected as the PN. (If the number of SNh>1, then PNid = arg min{ID ( SNh )} , where SNh is the number of SNs with equal minimum hop-count to BS, the PNid is the PN which has the smallest ID number, and ID (SNh) is the ID number of each SNh). 4.3 Channel Allocation The idea behind the graph G(V,E) is to join the mesh nodes. On a given set of SSs and BS, communication trees with the BS as roots need to be built. An important prerequisite to this channel allocation strategy is the important assumption that nodes are fitted with one or two radios. The goal is to model a channel assignment set for the set of radios such that interference is kept at a minimum and capacity fairly distributed. Figure 7 shows the nodes fitted with two radio interfaces, with each having a distinct function. While one maintains connectivity with other colleague nodes, the other ensures connectivity with subscriber stations. Hence, a reliable allocation mechanism should put into consideration, the service demands of each and every node in the network. Channel distribution for interference minimization is one good way to achieve this.

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm

241

Fig. 7. An example of network with multiple interfaces

The following aptly describes the channel assignment mechanism of this network. • As depicted in Figure 7, all the other nodes are subscribers to node B. As such, they must use a common channel to access B. The channel used for this purpose is dependent on B’s base station interface. Similarly, each station tunes its SS interface to match B’s BS interface. • To establish subscription to some other node, i.e. C in our scenario, a different channel is used by B. as shown in Figure 7. • Nodes transmitting on different channels do not cause any interference to one another. In this paper, the breadth first order is used. As depicted in Figure 8, a channel is first allocated to the BS, followed by assigning it to the SSs nearest to the BS, and then to all the SSs with more than one hop away from BS, and so on. However, taking into consideration the cost implication of equipping all the nodes with multi transceivers, the nodes in the edges {SS3, SS5, SS6, SS7, SS8} of the network are fitted with only one transmitter. This is because of the fact that while a multi transceiver for these edge nodes will raise their cost, there is no child node for these nodes and hence need not be fitted with multi transceivers as shown in Figure 8. Furthermore, since the BS is also fitted with two receivers operating at different channels, the proposed model uses both for transmitting because there is no PN for BS as shown in Figure 8.

Fig. 8. Breadth first algorithm

242

Y.A.A. Qasem et al.

4.4 Multi-transceiver Scheduling Algorithm This is an infrastructure based multi-hop WMN comprising of a BS and a number of SSs, where the BS also acts as the internet gateway. In such a scenario, the SSs not only transmit packets to and from BS, but also act as routers. The multiple transceivers on the SSs enable multi channel operations to reduce interferences. In the centralized scheduling algorithm, the following assumptions were made according to the WiMAX standard: • We assume that the transceiver cannot switch to other channels after the transceiver is fixed at one channel. • Any pair of nodes separated by two-hops and using different channels is considered to be none interfering. • Multi-transceiver nodes may communicate simultaneously with as many neighbours as there are transceivers without interference. • Each node’s communication radius is only enough to cover its one-hop neighbours. • Concurrent interference-free communication is possible along different channels. • The network is assumed to be constant throughout the course of any particular scheduling. Control and scheduling sub-frames are considered to be of considerable length. The aim of the centralized scheduling is to employ reuse timeslot and concurrent transmission to reduce the length of schedule and to attain optimum performance of the system. To achieve this, we must maximize concurrent transmissions while simultaneously minimizing interferences in the network. Therefore, we must take into account the traffic needs of the various SSs in the network. In 802.16-2004, two scenarios of centralized scheduling algorithm are therefore proposed. First scenario: multi-transceiver systems as shown in Figure 9, which enable transmission and reception simultaneously at all nodes. Second scenario: the nearest multi-transceiver system (only the nearest nodes to the BS have two transceivers) as shown in Figure 10. In this system, only the nodes closest to the BS can transmit and receive simultaneously.

Fig. 9. Multi-transceiver system

Fig. 10. The nearest multi-transceiver system

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm

243

In both types of centralized scheduling algorithm scenarios (i.e. the multitransceiver system and the nearest multi-transceiver system), the hop-count, rely model, node ID number, reuse timeslot, concurrent transmission and fairness (for this, the fairness constraints need to be taken into account to prevent starvation to nodes further away from the BS) are considered. The idea behind the centralized scheduling algorithm is to take note of all interfering nodes in the forward link and to allow non-interfering nodes to transmit simultaneously for optimal bandwidth use. The algorithm considers four important metrics instead of only the two used in [14]; these metrics are nearest nodes (minimum hop-count to BS) to reduce the system bottleneck, number of traffic (number of packet) to achieve the fairness among the nodes, number of interfering node to maximize the reuse timeslot and concurrent transmissions. The node ID number is finally used to break the tie between the nodes. The allocation of service tokens to the subscriber stations is a function of their traffic demand. The procedure ensures that timeslot allocation is directly proportionate to the traffic requirement of the various links, thereby ensuring fairness. For scheduling of a link to occur, its associated service token must not be zero. In short concerned SS must at least have a free non-interfering channel and its PN having buffer capacity for the incoming packets. Once these conditions are fulfilled, the link is considered available; else it is considered unavailable. In addition to fulfilling the above conditions, for a link to be available, it must satisfy the nearest to BS condition (meaning it exhibits the minimum hops to the base station BS) and is scheduled in the current timeslot. In the case of a tie in the nearest to base station condition, the first priority is accorded to the node having the maximum packets for achieving QoS and fairness. In the case of a tie in the number of packets, priority is given to the node with the minimum interfering neighbours for want of maximum reuse timeslot and concurrent transmissions. As a final step, if a tie occurs in the number of interfering nodes, priority is given to the node with the smallest ID number. When the link is finally selected, it is denoted as scheduled while its interfering neighbours are denoted as interfered. Now that the link is allocated a timeslot, the transmitter’s service token is decremented by one while the receiver’s is incremented by one at every timeslot. Then, the algorithm keeps repeating itself in this manner until the service tokens of all the SSs are decreased to 0. Hence, using the change of service token, the 802.16-2004 can be integrated into the proposed algorithm. Multi-channel multi-transceivers lead to shorter lengths of scheduling, increase system throughput, improve the CUR and provide collision free channel access. Hence, it can provide a better performance as compared to the single transceiver system. The proposed scheme has shown good performance and can be implemented in the IEEE 802.16 WiMAX based mesh networks. In addition, it has also achieved better system throughput in WMNs.

5 System Performances of Proposed EbMR-CS Algorithm In this paper, an efficient algorithm of EbMR-CS is developed with the aim of finding the optimal path of routing and scheduling, which is evaluated through simulation. In the simulation, the performance of the EbMR-CS scheme was assessed using

244

Y.A.A. Qasem et al.

MATLAB platform [25]. After A a series of 300 Monte Carlo simulation runs, the ddata is collected and averaged. The simulation configurations are set as follows: the mode simulated model consists off 100 SSs positioned around a BS operating in mesh m for uplink traffic. The SS S to BS communication was ensured through singlee or multiple hops. Nevertheleess, SS mobility is not considered. Using a step w wise increment of 5 nodes, the network n is loaded from 5 to 100. For each SS, the num mber of packets was selected forr one packet and between one to three randomly generaated packets. m consist of multi-transceiver system and The results from the multi-transceiver nearest multi-transceiver system which are intended to eliminate the prim mary interference and reduce thee scheduling length. These two schemes are dependentt on the EbM routing tree alg gorithm, multi-channel scheduling and the number of transceivers in the network k. The first multi-transceiver system is denoted as EbM MRCS1 and a construct of the EbM E routing, multi-channel scheduling; the whole netw work is equipped with two transcceivers except the nodes in the edges. In the nearest muultitransceiver system, which is denoted by EbMR-CS2, the EbM routing and muultiped only the nearest nodes to BS with multi-transceivvers. channel scheduling equipp Moreover, reuse timeslot, the hop-count, relay model, node ID number, concurrrent a also considered in both systems. Both the EbMR-C CS1 transmission and fairness are and EbMR-CS2 are compaared with the routing and scheduling schemes proposedd by Wang in [16] in terms of len ngth of scheduling, CUR and system throughput. The results from the sim mulation on the scheduling length are presented in Figuures 11 and 12. They are based d on packet generation for one packet and between 1 tto 3 randomly generated packets, as shown in Figures 11 and 12 respectively, to repressent the heterogeneous traffic deemands for all the nodes. Based on the assumption that one transceiver can transmit at most m one token (packet) at a timeslot, it is clearly seen tthat the duration of the schedu uling cycle gradually increases with network size forr all scenarios as shown in Figu ure 11. The increasing traffic demand also led to lonnger length of scheduling as sho own in Figure 12. The results from the simulation indiccate that the EbMR-CS1 achieveed a higher system performance in terms of the schedulling length compared to the otther schemes. In other words, the EbMR-CS1 maintaains shorter length of scheduling g as compared to EbMR-CS2 and Wang [16]. The EbM MRCS1 at 80, for example, where w the node has one packet to send is equivalent too 42 timeslot, whereas the EbNR R-CS2 is equivalent to 47 timeslots.

Fig. 11. Leng gth of scheduling with one packet for each node

Collision-free Routting Centralized Scheduling Using EbMR-CS Algorithm

245

Fig. 12. Length of sccheduling with random packets from 1 to 3 in each node

Using Wang [16], this is found to be 57 timeslot. Besides, both schemes proposedd in this study achieved better sy ystem performance than the scheme presented in Wang [16]. The simulation results for the CUR are presented in Figures 13 and 14, w with on for one packet and between 1 to 3 randomly generaated number of packet generatio packets respectively. After obtaining the length of scheduling, the number of timeslots required to send all the pacckets to BS has to be determined, so as to identify the tootal CUR using the following eq quation: CUR (%) =

( No . Packets * No . Hops ) ( No . Nodes * Length .of . scheduling

)

(6)

Based on this equation, it is observed that the increase in the number of nodes in the network can reduce the CUR, as the additional nodes in the network increases the odes when one node is transmitting as shown in Figure 13. number of interfering SS no This may be due to the fact that the additional nodes worsen the interference leveel in the case where only one node is sending data, thereby downgrading the possiible nsmissions. In addition, it is also observed that when the number of concurrent tran number of packet is increassed, the CUR is stabilized. It is determined by the ratioo of the reuse timeslot and the concurrent transmissions as illustrated in Figure 14. om the simulation indicate that the EbMR-CS1 achieeved However, the results fro higher CUR. For example,, the EbMR-CS1 at 60 when the node has one packet to send is equivalent to 7%, whereas w the EbNR-CS2 is equal to 6%.

Fig. 13. CUR with each node has one packet

246

Y.A.A. Qasem et al.

w each node having random packets from 1 to 3 Fig. 14. CUR with

In Wang [16], this corresponds to 4% at 60 nodes. At 10 nodes, however, both the EbMR-CS1 and EbMR-CS2 2 are equal to 26%, as compared to 18% in Wang’s scheeme. Hence, both EbNR-CS1 and d EbNR-CS2 schemes outperform Wang’s scheme. Figure 15 shows the ressults from the simulation on the throughput of the BS S in which each node has one service s token to send. As shown in Figures 15 and 16, the throughput of the EbM-CS1 algorithm is much larger than the other algorithms w when all the nodes are equippeed with two transceivers. This result indicates that the algorithm in the present study can support more users as compared to the others.

Fig. 15. Systeem throughput in which each node has one packet

Fig. 16. System throughput when each node has random packets from 1 to 3

Collision-free Routing Centralized Scheduling Using EbMR-CS Algorithm

247

Both Figures 15 and 16 show that the throughput is reduce as the number of the service tokens and nodes increase. For example, the system throughput at 80 nodes, for the EbMR-CS1 is equal to 1.9 packets/timeslot, while in the EbMR-CS2 it is equal to 1.7 packets/timeslot when each node has one packet to send. In Wang’s scheme, the system throughput is very small as compared to the other schemes which are equivalent to 1.4 packets/timeslot. Based on this example, the analyzed results show that the algorithms in the present study not only optimizes system throughput, but also enhances the overall efficiency of centrally controlled scheduling scheme.

6 Conclusions Routing and scheduling in WiMAX are active areas of research, in which many algorithms have been proposed in an attempt to enhance system throughput, reduce the length of scheduling, increase CUR and provide more robustness over the wireless channel. The related problems of routing and scheduling for IEEE 802.16-2004 are still open research challenges left unsolved. The impact of interference on IEEE 802.16-2004 based mesh networks is very strong. In this paper, the EbMR-CS1 and EbMR-CS2 have been proposed for centralized scheduling in WiMAX to improve system throughput. In particular, the EbMR-CS1 and EbMR-CS2 use the multitransceiver and multi-channel systems. Therefore, the interferences could be virtually eliminated. The results from simulation indicate that the proposed schemes achieved shorter lengths of scheduling, higher CUR and higher system throughput, as compared to Wang [16]. At the same time, it ensures fairness and better load balanced in the IEEE 802.16-2004 based mesh networks, particularly when the number of nodes is large. Moreover, the algorithms increase the reuse timeslot and concurrent transmission which allows the non-interfering links to transmit simultaneously. In addition, the proposed EbMR-CS can avoid packet loss and afford collision-free operations. The work in this paper can be extended in several directions. First, it can be further extended to distribute traffic when multiple paths are available to the destination. Second, the SS mobility can be introduced so as to compare its performance on IEEE 802.16d and IEEE 802.16e.

References 1. Lu, K., Qian, Y., Chen, H.H.: Wireless broadband access: WiMAX and beyond - a secure and service-oriented network control framework for WiMAX networks. IEEE Communications Magazine 45(5), 124–130 (2007) 2. IEEE Standard 802.16-2004: IEEE Standard for Local and metropolitan area networks Part 16: air interface for fixed broadband wireless access systems (2004) 3. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Computer Networks 47(4), 445–487 (2005) 4. Akyildiz, I.F., Wang, X.: A survey on wireless mesh networks. IEEE Communications Magazine 43(9), S23–S30 (2005) 5. Viswanathan, H., Mukherjee, S.: Throughput-range tradeoff of wireless mesh backhaul networks. IEEE Communications 24(3), 593–602 (2006) 6. Narlikar, G., Wilfong, G., Zhang, L.: Designing multihop wireless backhaul networks with delay guarantees. Wireless Networks 16(1), 237–254 (2008)

248

Y.A.A. Qasem et al.

7. Chen, J., Chi, C., Guo, Q.: A Bandwidth Allocation Model with High Concurrence Rate in IEEE802.16 Mesh Mode. In: IEEE Communications, Asia-Pacific Conference, pp. 750– 754 (2005) 8. Peng, M., Wang, Y., Wang, W.: Cross-layer design for tree-type routing and level-based centralised scheduling in IEEE 802.16 based wireless mesh networks. IET Journal 1(5), 999–1006 (2007) 9. Kim, D., Ganz, A.: Fair and efficient multihop scheduling algorithm for IEEE 802.16 BWA systems. In: IEEE, BroadNets Conference, vol. 2, pp. 833–839 (2005) 10. Tao, J., Liu, F., Zeng, Z., Lin, Z.: Throughput enhancement in WiMax mesh networks using concurrent transmission. IEEE, Wireless Communications, Networking and Mobile Computing 2, 871–874 (2005) 11. Wei, H.H., Ganguly, S., Izmailov, R., Haas, Z.J.: Interference-aware IEEE 802.16 WiMax mesh networks. In: IEEE, VTC 2005, vol. 5, pp. 3102–3106 (Spring 2005) 12. Fu, F., Cao, Z., Fan, P.: Spatial reuse in IEEE 802.16 based wireless mesh networks. In: IEEE International Symposium, Communications and Information Technology, ISCIT 2005, vol. 2, pp. 1358–1361 (2005) 13. Han, B., Tso, F.P., Ling, L., Jia, W.: Performance Evaluation of Scheduling in IEEE 802.16 Based Wireless Mesh Networks. In: IEEE Mobile Adhoc and Sensor Systems (MASS), IEEE International, pp. 789–794 (2006) 14. Han, B., Jia, W., Lin, L.: Performance evaluation of scheduling in IEEE 802.16 based wireless mesh networks. Computer Communications 30, 782–792 (2007) 15. Peng, D., Weijia, J., Liusheng, H., Wenyan, L.: Centralized Scheduling and Channel Assignment in Multi-Channel Single-Transceiver WiMax Mesh Network. In: IEEE, Wireless Communications and Networking Conference, WCNC, pp. 1734–1739 (2007) 16. Wang, J., Jia, W., Huang, L.: An efficient centralized scheduling algorithm for IEEE 802.16 multi-radio mesh networks. In: ACM, Proceedings of the 2nd International Conference on Ubiquitous Information Management and Communication Suwon, Korea, pp. 1–5 (2008) 17. Kyasanur, P., Vaidya, N.H.: Routing and interface assignment in multi-channel multiinterface wireless networks. Wireless Communications and Networking 4, 2051–2056 (2005) 18. Kodialam, M., Nandagopal, T.: Characterizing the capacity region in multi-radio multichannel wireless mesh networks. In: Mobile computing and networking, MobiCom, pp. 73–87 (2005) 19. Djukic, P., Valaee, S.: 802.16 mesh networking. In: WiMAX: Standards and Security, pp. 147–174 (2007) 20. Wang, W., Wang, Y., Li, X.Y., Song, W.Z., Frieder, O.: Efficient interference-aware TDMA link scheduling for static wireless networks. In: ACM, 12th Annual International Conference on Mobile Computing and Networking, Los Angeles, pp. 262–273 (2006) 21. Ramanathan, S., Lloyd, E.: Scheduling algorithms for multihop radio networks. IEEE/ACM Transactions on Networking (TON) 1, 166–177 (1993) 22. Kuran, M.S., Gur, G., Tugcu, T., Alagoz, F.: Cross-layer routing-scheduling in IEEE 802.16 mesh networks. In: 1st international conference on MOBILe Wireless MiddleWARE, Operating Systems, and Applications Innsbruck. ICST, Austria (2007) 23. Grńsund, P., Engelstad, P., Johnsen, T., Skeie, T.: The physical performance and path loss in a fixed WiMAX deployment. In: International Conference on Wireless Communications and Mobile Computing, IWCMC 2007, p. 444 (2007) 24. Wenhua, J., Pin, J., Ruoju, L., Ming, L.: Centralized Scheduling Tree Construction Under multi-Channel IEEE 802.16 Mesh Networks. In: IEEE, Global Telecommunications Conference, GLOBECOM 2007, pp. 4764–4768 (2007) 25. The Mathworks accelerating the pace of engineering and science, http://www.mathworks.com/access/helpdesk/help/techdoc/ learn_matlab/f0-14059.html

Energy-Efficient Target Tracking in Sensor Networks Loredana Arienzo1,2 and Maurizio Longo2 1

Institute for the Protection and Security of the Citizen, Joint Research Centre, European Commission Ispra, 21027, Varese, Italy [email protected] 2 Department of Electrical and Information Engineering, University of Salerno Fisciano, 84084, Salerno, Italy [email protected]

Abstract. In this paper, the problem of collaborative tracking of mobile nodes in wireless sensor networks is addressed. Aiming at an energy efficient solution, we propose a strategy of combining target tracking with node selection procedures in order to select informative sensors to minimize the energy consumption of the tracking task using the energy model by Heinzelman, 2000. We layout a cluster-based architecture to address the limitations in computational, battery power and communications of the sensor devices. The node selection problem is formulated as a crosslayer optimization problem that is solved using a greedy algorithm. To track mobile nodes two particle filters are used: the bootstrap particle filter and the unscented particle filter, both in the centralized and in the distributed manner. Their performance are compared with the distributed sigma-point information filter in literature, under two common channel models: the log-normal shadowing and the Rayleigh fading. Keywords: Sensor networks, tracking, particle filters, energy efficient, cross-layer optimization, cluster.

1

Introduction

The issue of tracking moving targets in wireless sensor networks (WSNs) [1] has received significant attention in recent years [2–4]. A tracking algorithm designed for sensor networks should be: (1) self-organizing, i.e. it should not depend on global infrastructure; (2) energy efficient, i.e. it should require little computation and, especially, communication; (3) robust, i.e. it should not depend on noise and movement of the target; (4) accurate, i.e. it should work with accuracy and precision in various environments, and should not depend on sensor-to-sensor connectivity in the network; (5) reliable, i.e. it should be tolerant to node failures. The minimization of energy consumption for a sensor network with target activities is complicated since target estimation involves collaborative sensing and communication between different nodes. The problem of selecting the best J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 249–264, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

250

L. Arienzo and M. Longo

nodes for tracking a target in a distributed wireless sensor network was investigated since 1999 [5]. The main idea is for a network to determine participants in a sensor collaboration by dynamically optimizing an utility function of data for a given cost of communication and computation. Previous research [6–9], has focused on information-theoretic node selection approaches, i.e. on heuristics to select an informative sensor such that the fusion of the selected sensor observation with the prior target location distribution would yield the greatest reduction in the entropy of the target location distribution. In [6], the sensor node which will result in the smallest expected posterior uncertainty of the target state is chosen as the next node to contribute to the decision. Specially, minimizing the expected posterior uncertainty is equivalent to maximizing the mutual information between the sensor node output and the target state [6]. In [7], an entropy-based sensor selection heuristic is proposed for target localization in which a sensor node is chosen at each step and the observation of that node is incorporated into the target location distribution using sequential Bayesian filtering. Many criteria influence the design of energy-efficient tracking approaches, and a wide range of schemes have been proposed. The rest of this Section provides a partial overview of such schemes. Wang et al. [10], Chen et al. [11], propose cluster-based tracking schemes. They envision a hierarchical network composed of (a) a static backbone of sparsely placed position-aware sensors which may assume the role of a cluster head (CH) upon triggered by certain signal events; and (b) moderately to densely populated low-end sensors whose function is to provide sensor information to CHs upon request. In these schemes, sensors are grouped into clusters either statically or dynamically (upon detection of the target in the proximity), and a CH collects information from its cluster members and determines the target location using either the trilateration technique [10] or the Voronoi diagram-based approach [11]. Both localization approaches aim to determine the exact location of the target at the expense of considerable computational overhead because of the potentially high number of nodes in the cluster. From a topology perspective, the tracking approaches could use a global or local knowledge about the location of every node in the network. As opposed to the tree-based schemes [12–14] that use a global information, the clusterbased schemes [10, 11] rely on local topology knowledge to limit the scope of target’s location updates. As to the signal processing, the tracking approaches can be classified as centralized or distributed. Usually the tree-based are centralized approaches, while the cluster-based are distributed schemes in which the CH is the leader node in the processing. References [13, 15] are centralized approaches, while [3, 4, 10, 14, 16] are distributed approaches. As defined in [17] the sensor management is the process of dynamically retasking sensors in response to an evolving environment. In a tracking task the sensor management addresses the problem of choosing informative sensors needed to obtain information about the target state and therefore maximize the network lifetime. Based on the collaboration, the existing approach of target tracking can be classified as information-driven or information-based. Zhao et al. [16] propose IDSQ

Energy-Efficient Target Tracking

251

(Information Driven Sensor Querying), in which the selection of the best node is based on a Mahalanobis distance that leads to a heuristic method favoring the sensors whose Euclidean distance to the target is small. In [7–9] the node selection problem has been addressed using an information-based approach. The main idea behind these approaches is to optimize an utility function, representing the location accuracy, using entropy-based metrics. Instead, the main idea underlying our approach is that the heuristics select an informative sensor such that the fusion of the selected sensor observation with the prior target location distribution would yield to optimize an utility function combining the overall energy in a cluster and the mutual information between the sensor node observations and the target state. The remainder of the paper is organized as follows: Section 2 describes the model of the overall system by introducing the network architecture and energy-based metrics. Section 3 provides the model of the dynamical system being sensed. Section 4 describes the tracking algorithms and formulates the distributed tracking problem introducing a node selection rule. In Section 5 we describe the optimization problem. Section 6 discusses the performances of proposed algorithms, while in Section 7, we draw the main conclusions.

2

System Model

We make the following assumptions about the sensor network. First, the network is composed of a single gateway (sink) node and multiple sources. Next, the network is modeled as a combination of 1) a static backbone of sensor nodes aware of their position that are candidate to assume the role of a CH and 2) randomly distributed low-end sensors which sense a moving target and report data to their CH upon request. Finally, we assume that the network is composed of dynamic clusters, depending on the predicted target trajectory (see Fig. 1). The details of the clustering algorithm are out of the aim of this paper. In the following we will limit ourselves to consider only the intra-cluster communication issues. The acting CH will predict the trajectory of a target by means of a particle filter based on the history of the target location and some recent observations communicated by a subset of active sensors in the cluster. This subset is selected to minimize the overall energy consumption. 2.1

Energy Model

To describe the energy consumption of a tracking algorithm, we use the energy model for wireless sensor networks introduced by Heinzelman et al. [18]. Hence, given l bits of data, the overall energy consumption to transmit the packet of l bits between two nodes at a distance d with a given received SNR can be expressed as: E(d, l) = (2Eelec + Eamp · dα ) · l

(1)

where Eelec [Joule/bit] is the energy needed by the transceiver circuitry to transmit or receive one bit and Eamp [Joule/(bit · mα )] is a constant which represents

252

L. Arienzo and M. Longo

Fig. 1. Sensor Network Topology

Fig. 2. Location Discovery Protocol

the energy needed to transmit one bit over a distance d to achieve an acceptable SNR at the destination. This model assumes that the energy consumption is dominated by the radio communication rather than the computation. We refer to (1) as the energy-based metric. In our analysis we adopt a slightly different metric accounting not only for the energy consumption as in the above model but also for the remaining energy at the node. Hence, if we refer to a link (i,j) with distance dij along which a lij -bit packets is transmitted, the residual energy at node i evolves in time according to:  (2Eelec + Eamp · dα (2) Er (i, k + 1) = Er (i, k) − i,j (k)) · li,j j

Energy-Efficient Target Tracking

253

where Er (i, k) is the remaining energy at node i and time k. Note that, in the scenario of a mobile node i, dij is a function of time k. We refer to (2) as the residual energy-based metric. 2.2

Network Model

As stated before, we consider a network composed of randomly deployed sensor nodes which sense a moving target and transmit their observations to a positionaware sensor which acts as CH, namely the data gathering node. The network is divided into Nc clusters each having Na neighbors, i.e. sensor nodes within the radio range of the cluster head. Each sensor is equipped with a low data rate radio interface. The CHs are equipped with two radio transmitters, i.e., a low data rate transmitter to communicate with the sensors, and a high rate wireless interface for CH-CH communication. We assume that the (static) position of the CHs is known and want to estimate the distance of each neighbor with respect to the acting CH. Many types of sensors provide measurements that are function of the relative distance between the sensor and the sensed object (e.g. acoustic sensors, sonar, etc.). We consider a common example, as in [4], where sensors measure the power of a radio signal emitted by the object: assuming the exponential decrease of received power with distance and the log-normal shadowing model for the observations we get Pr (d) = Pr (do ) · (do /d)α + Xσ

(3)

where Pr (d) is the received power at a receiver at distance d from a transmitter, Pr (do ) is the transmitted power at a reference distance do , α is the path loss exponent α ∈ [2,5], and Xσ is the shadow fading component, with Xσ Gaussian distribution N (0, σ). Hence, the distance from the i-th sensor of the cluster to the CH can be estimated as d = (Pr /Pa )−1/α , where Pr is the received signal strengthat the CH at , Pa is the strength of the signal emitted by the sensor.

3

Model for State Estimation of a Dynamical System

The problem of estimating the state vector of a dynamical system (i.e. single target tracking) can be formulated as follows. The state and the observations of the target of interest are assumed to obey the equations xk+1 = Fk xk + Ak wk zk = Hk (xk ) + Bk vk

(4) (5)

where, k is the discrete time index; xk Rn is the state vector; zk Rm is the observation vector; wk is a zero-mean Gaussian process noise with nonsingular covariance matrix Qk ; vk is a zero-mean Gaussian measurement noise independent of wk with nonsingular covariance matrix Rk . The matrices Fk , Ak , Bk are

254

L. Arienzo and M. Longo

independent of the state vector whereas Hk is a function dependent of the state vector xk . Further we assume that the initial state x0 has a known probability density function p(x0 ). In this paper we assume as dynamic model of the target the constant velocity model. Hence, denoting by xk = [αk , α˙k , βk , β˙k ]T the state vector (coordinates along x, y axes and the velocities) of a target, the state-space model is given by, ⎛ ⎞ ⎛ ⎞ 2 1 ΔT 0 0

ΔT /2

0

0 0 0 1

0

ΔT

0 ⎟ ⎜0 1 0 0 ⎟ ⎜ ΔT xk+1 = ⎝ x + w 0 0 1 ΔT ⎠ k ⎝ 0 ΔT2 /2 ⎠ k

where vk ∼ N (0, diag(σx2 , σy2 )) denotes the motion noise and ΔT the length of the measurement interval. 3.1

Observation Model for Lognormal Shadowing

As observation model of the measurements, we first use the log-normal shadowing model [19]. Hence, let {αs , β s } be the fixed position of sensor s and dk = xk − s1/2 = [(αk − αs )2 + (βk − β s )2 ](1/2) be the distance between the sensor s and the target, in logarithmic scale the measurements are modeled by Hk (xk ) = K − 10α log(dk ) zk = Hk (xk ) + Bk vk

(6) (7)

where the measurement noise vk accounts for the shadowing effects and other uncertainties. The noises vk are zero-mean Gaussian with covariances values σx2 = σy2 = σo2 , and uncorrelated w.r. to k, i.e. from sensor to sensor; K is the transmission power, and α ∈ [2,5] is the path loss exponent. 3.2

Observation Model for Rayleigh Fading

The problem of estimating the state vector of a nonlinear dynamical system over a fading channel can be formulated as follows. The state and the observations of the target of interest are assumed to obey the equations xk+1 = Fk xk + Ak wk

(8)

zk = Hk (xk )rk + Bk vk

(9)

where rk is the fading channel gain. Specifically, for a Rayleigh fading channel, the probability density of rk is: pR (rk ) =

rk r2k exp(− 2 2 ), σR 2σR

rk ≥ 0.

(10)

Energy-Efficient Target Tracking

4

255

Tracking Algorithms

As to the tracking task, we use sequential Monte Carlo (SMC) approaches, also known as particle filtering [20, 21]. A different approach by Kaplan [8, 9] estimates the target location using a Kalman filter based on the current measurement at a sensor and the past history at other sensors. The main idea of particle filtering (PF) is to represent the required posterior distribution density by a set of random samples with associated weights and to compute estimates based on these samples and weights, updating them recursively in time using sequential importance sampling (SIS). A proposal or importance distribution is introduced, which is an approximation of the required posterior distribution of states given the value of previous step π(xk |xk−1 , z1 , ..., zk−1 ). As the number of samples becomes very large, the SIS filter approaches the optimal Bayesian estimate. A common problem with the SIS PF is the degeneracy phenomenon, since after a few iterations all the particles, except one, will have negligible weight. Because of this phenomenon, resampling techniques are used to eliminate particles that have small weights and to concentrate on particles with large weights. The PF using sequential importance resampling (SIR) techniques is known as bootstrap f ilter or SIR PF. // Indeed, unfortunately even when resampling schemes are used, degeneracy may still be a problem. Using the prior distribution as importance distribution could lead to the degeneracy problem of the particles because the most recent observations are ignored. Samples may eventually collapse to a single point if, during the resampling stage, samples with high importance weights are duplicated an extremely large number of times. There have been numerous proposals to mitigate the degeneracy problem. Notable techniques include local linearization using the extended Kalman filter (EKF) or the unscented Kalman filter (UKF) to estimate the importance distribution. A particle filter which uses UKF to generate the importance distribution is referred as unscented particle f ilter (UPF) or sigma-point particle filter [22]. In particular, the unscented transformation is used to generate and propagate a Gaussian proposal distribution for each particle. The unscented transformation still approximates the proposal distribution by a Gaussian distribution, but it is specified using a minimal set of deterministically chosen sample points or sigma points. These points complectly capture the true mean and covariance of the Gaussian distribution, and when propagated through the true nonlinear system, they accurately capture the posterior mean and covariance to the third order for any nonlinearity. In this paper, we implement both the SIR PF and the UPF and we compare thire performances in terms of estimation accuracy and energy efficiency when integrated into collaborative and distributed schemes for tracking a moving target. 4.1

Node Selection by Information

According to SMC approach, each new sensor measurement zk is combined with the current estimate p(xk |z1 , ..., zk−1 ), hereafter called belief state, to form a

256

L. Arienzo and M. Longo

new belief state p(xk |z1 , ..., zk ). We consider now the problem of selecting a new sensor in order to provide the greatest improvement of accuracy at the lowest cost. Let Zk represent all measurements that have already been used at time k in the inference of the current belief state, which is maintained at the CH. The objective function for this optimization problem can be defined as a mixture of both information gain and cost. In the remainder of this Section we consider the information gain, while the energy consumption issue is discussed in next Section. The information gain stemming from selection of sensor s can be defined as Φ˜s (p(xk |Zk )) = ΦUtility (p(xk+1 |Zk , zk+1,s )), where zk+1,s is the new measurement from sensor s at time k + 1. The utility function can be defined as the uncertainty of the target state reduced by the additional measurements zk+1,s [6], i.e. Φ˜s (p(xk |Zk )) = Htarget (Zk ) − Htarget (Zk , zk+1,s ); or, equivalently as the mutual information ΦUtility (p(xk+1 , zk+1,s |Zk )) = I(xk+1 , zk+1,s |Zk ) conveyed on xk+1 by the new measurements zk+1,s [15]. The utility function based on the entropy is difficult to compute in practice since we need to have the measurement before deciding how useful it is. Instead of the true aposteriori distribution, a more practical alternative is to compute the entropy based on the expected posterior distribution. In the ideal case when a real new measurement zk+1,s is available, the new belief or posterior is evaluated using sequential Bayesian filtering p(xk+1 |zk+1,s , Zk ) ∝ p(zk+1,s |xk+1 )p(xk+1 |Zk ). Since zk+1,s is not available, we may compute the expected posterior distribution Ezk+1,s (p(xk+1 |zk+1,s , Zk )). We can estimate the measurement ¯ zk+1,s from the predicted belief and compute the expected likelihood function p ˆ (¯ zk+1,s |xk+1 ) = p(zk+1,s (νk+1 )|xk+1 ) × p(νk+1 |Zk )dνk+1 . Then, the expected posterior belief can be defined as follows: pˆ(xk+1 |¯ zk+1,s , Zk ) = pˆ(¯ zk+1,s |xk+1 )p(xk+1 |Zk ).

(11)

The entropy of expected posterior distribution can be computed based on the disj crete belief state {xjk , wkj }L j=1 [23], where wk is the importance sampling weight in the resampling step of the particle filters and L represents the number of weights, i.e. the number of particles. According to (11), the expected posterior j ˜k+1,s }L belief for sensor s can be represented by the discrete belief state {xjk , w j=1 j with the weights w ˜k+1,s given by j j w ˜k+1,s = pˆ(¯ zk+1,s |xjk+1 )wk+1 .

(12)

j ˜k+1,s }L Then the entropy of the discrete belief state {xjk+1 , w j=1 can be computed as

H=−

L 

j j w ˜k+1,s log w ˜k+1,s .

(13)

j=1

This expected posterior entropy can be used as a criteria to select the best among the sensor candidates to maximize the information gain. The objective function expected to improve the estimation of the target is given by

Energy-Efficient Target Tracking

Ns = arg max Φ˜s (p(xk |Zk )) sNa

257

(14)

= arg max(Htarget (Zk ) − Htarget (Zk , zk+1,s )) sNa

where Na indicates the set, with cardinality Na , of active neighbor nodes in the cluster that receive a signal exceeding a predetermined RSS threshold.

5

Energy Efficient Tracking

Let us consider the following location discovery protocol for a given snapshot. Let Ns , as above, be the set of active neighbor nodes maximizing the utility function as in (14). We now want to select, inside Ns , a subset Nd of Nd active nodes collaborating in the localization task with an aim to minimize the energy consumption. Then, each node i ∈ Ns transmits the sensing information (the distance of the node i from the target) to the CH which processes the data and updates the current target location. According to metric (1), and assuming that target T emits discovery signal with period TM , the energy cost associated with sensor i ∈ Ns is given by: l TM + [Eelec (Nd + 1) + Eamp · riα ] · b

Ei (ri , di ) = [Eelec (Nd + 2) + Eamp · dα i ]·

(15)

where b represents the bit rate [bit/s] between the CH and the neighbor i, ri and di are, respectively, the distance of the node i from the CH and the target, Eelec Nd represents the energy needed at the neighbors to receive one bit. In the energy cost we have omitted the energy consumption in the path between the target and the CH due to the calibration phase. The total energy consumption by all nodes in Ns , is hence given by  E T OT (Ns ) = Ei (ri , di ) (16) i∈Ns

As our objective is to select the optimal subset Nd ⊂ Ns , i.e. which minimizes the total energy cost in Eq. (16) subject to a constraint of the cardinality Nd of said subset, the objective function can be formally cast as follows: Nd = arg min E T OT (N ) N ⊆Ns

(17)

subject to Nd ≥ 2 A unique solution to this problem exists, since the objective function is strictly concave and the feasible set is convex. Following [24], the solutions of this optimization problem depend on the scenario, static or dynamic, as illustrated in the following Subsections.

258

L. Arienzo and M. Longo

Algorithm 1. Energy Efficient Tracking Algorithm Synopsis: [Nd , C, Etot ] =DynamicSelection(Na ,Nd ). Output: Set of desired nodes Nd , new set of candidate nodes in the cluster C,total energy of the desired set Etot. 1. The initial leader node does the following step: (a) draw initial samples {xj0 , w0j = 1}L j=1 of the target from the prior information; (b) update the belief state {xj1 , w1j }L j=1 by the sensor fusion algorithm based on the new measurement z1 at the leader node in the set Na ; (c) compute the expected j j posterior belief state {xj2 , w2,i }L ˜2,i j=1 for each neighboor node i with the weights w computed by (12); (d) compute the entropy of the expected posterior belief state j {xj2 , w2,i }L j=1 for each neighboor node i by (13) and determine the next best sensor, say b in the set Ns . 2. [Nd , C, Etot , node, Eb , k] =Greedy(Ns ,Nd ) 3. Loop until time runs out: 4. Prediction and update steps of particle filtering to estimate the target’s trajectory. 5. Update the candidate set C during the dynamic of the target. 6. [Nd , C  , Etot] =BranchBound(Nd , C, Eb , node)

5.1

The Solution in the Static Scenario

If we formulate our combinatorial optimization problem as an integer linear programming problem, the computational complexity consists of enumerating all the Nd -node subsets, O(NaNd ), and adding the computational complexity of the assignment problem, O(Nd3 ). In such cases, heuristic methods are usually employed to find good, but not necessarily guaranteed optimal solutions. Here we adopt the meta-heuristic GRASP: each iteration consisting of two phases, a construction phase, in which a feasible solution is produced, and a local search phase, in which a local optimum in the neighborhood of the constructed solution is sought. The best overall solution is kept as the result. The implementation of the optimal Greedy node selection procedure is described in [24]. It provides the set of desired nodes Nd with the related total energy Etot , the last node selected node with the related specific energy Eb , the set of nodes in the cluster which are candidated for the next snapshot C. 5.2

The Solution in the Dynamic Scenario

In this section we extend the greedy node selection procedure over multiple snapshots, so that we can select active nodes for the next measurement intervals. In a dynamic scenario, due to the target mobility, the distance di in Eq. (15) varies with the time and hence the energy consumption varies with time k. For this dynamic version of the optimization problem we use the dynamic programming, that is based on the idea of breaking down the problem into stages at which the decisions take place and finding a recurrence relation that takes us backward from one stage to the previous stage. For this purpose, a branchand-bound method is developed, in which the branch refers to the partitioning

Energy-Efficient Target Tracking

259

process into stages, that are repeatedly decomposed until a solution is found or infeasibility is proved, and the bound refers to lower bounds that are used to construct a proof of optimality without exhaustive search. We introduce an energy bound as in follows: the energy bound is the maximum energy consumption, namely the energy consumed by the nodes selected in the previous snapshot. The pseudo-code of an efficient implementation of our branch-and-bound approach is described in [24]. It provides the set of desired nodes Nd , the new set of candidate nodes in the cluster C  and the total energy of the desired set Etot . The overall tracking algorithm which combines the node’s selection procedures with the particle filtering algorithm is finally outlined as Algorithm 1.

6

Performance Evaluation

In this section, we investigate the performance of the overall target tracking system looking first at the node selection algorithm, second at the estimation bound, then at the tracking algorithm and finally at the energy consumption. We present simulation results for the scenarios illustrated in Table 1. In the experiment we use the target trajectories introduced in [24] for two different velocities, equal to 0.1 and 0.3 m/s. Table 1. Parameters of the model used for simulations Parameters Scenario1 size 20m2 velocity 0.1m/s K 9 dB α 3 Eelec 50 nJ/bit Eamp 100 pJ/bit/m2 Er 100 mJ TM 2 sec b 10 bit/sec l 8 bits ΔT 1 sec σp 1.0 σo 0.3 L 100, 500

6.1

Scenario2 200m2 0.3m/s 9 dB 3 50 nJ/bit 100 pJ/bit/m2 100 mJ 2 sec 10 bit/sec 8 bits 1 sec 1.8 0.3 100, 500

Optimal Node Selection

First we compare from a computational point of view our node selection algorithm with the Kaplan algorithm in [8] and [9], the difference between the last two being that in [8] a global topology knowledge is assumed and every active node reaches the entire network, while in [9] only the knowledge of the relative position to the target and the active nodes from the previous snapshot is

260

L. Arienzo and M. Longo

required. For each iteration, using Nd as number of selected node, the computational complexity of greedy algorithm is O(Ns − Nd ) while the computational complexity of Kaplan algorithm is O((Ns − Nd )2 ). As stated above, the node selection procedure is combined with the maximization of the utility function. Hence in our analysis the computational complexity of the problem in (14) needs to be considered, namely O(Na L log L) where L is the number of weights. Finally, the computational complexity for all iterations

Na to maximize the utility function in (14) is given by i=1 iL log L. In Table 2 the results of a runtime measurement are illustrated, conducted on a system with AMD opteron XP processor 250, 2400 MHz frequency and 4,00 GB RAM. Table 2 provides the execution time of the node’s selection algorithm versus the number of desired nodes using Ns equal to 10 and L equal to 100. Table 2. Time to process greedy and Kaplan algorithms Number of desired nodes Greedy Time 2 6.6639e-5 3 7.8829e-5 4 1.0095e-4 5 1.2099e-4 6 1.4041e-4

6.2

Kaplan Time 13.9816e-4 32.5238e-4 10.0543e-3 23.1352e-3 52.2764e-3

Tracking Accuracy

We implemented the node selection algorithms and the particle filters in a Matlab simulator. Fig. 3 shows the root-mean-squared error (RMSE) on the position of the target of different filters versus the number of desired nodes using 100 runs and the log-normal shadowing model. The bootstrap and the unscented particle filters have been implemented both in centralized and distributed manner using the node selection rules. The performance of the distributed PF (DPF) and distributed UPF (DUPF) are compared with the performance of the distributed sigma-point information filter (DSPIF) from [4]. Confidence intervals are not shown for the sake of clarity. In Scenario1, nodes are randomly deployed on an area of 20m×20m and the target speed is 0.1 m/s. Clearly, the centralized filter PF outperforms the distributed filter DPF in tracking accuracy as it may be expected since in the distributed case nodes only have local knowledge. Also the RMSE of DUPF is always larger than that for UPF. On the other hand, as we will show, the energy consumption is higher for the centralized approach. Finally, DPF and DUPF outperform DSPIF. In Scenario2, nodes are randomly deployed on an area of 200m×200m and the target speed is 0.3 m/s. Fig. 3 (b) illustrates RMSE on the position of the target, depending on the number of active nodes in the network. From Fig. 3 (b), the unscented particle filter with 100 particles gives best results than the bootstrap particle filter using 100 particles. Not shown in Fig. 3 (b), the error when Nd = 2

Energy-Efficient Target Tracking 0.28

1

PF DPF DSPIF UPF DUPF

0.24 0.22 0.2

PF DPF DSPIF UPF DUPF

0.9

RMSE (meters)

0.26

RMSE (meters)

261

0.18 0.16

0.8 0.7 0.6 0.5

0.14 0.4

0.12 2

3

4

5

6

3

Number of Selected Nodes

4

5

6

Number of Selected Nodes

(a)

(b)

Fig. 3. Performance accuracy comparison of particle filters (a) target velocity = 0.1 m/s, σo = 0.3. and σp = 1.0, (b) target velocity = 0.3 m/s, σo = 0.3. and σp = 1.8.

0.48

0.9

PF with Rayleigh fading UPF with Rayleigh fading

0.85 RMSE (meters)

RMSE (meters)

0.47

0.95

PF with Rayleigh fading UPF with Rayleigh fading

0.46 0.45

0.8 0.75 0.7 0.65

0.44

0.6 0.43 2

3 4 5 Number of Selected Nodes

(a)

6

0.55 3

4 5 Number of Selected Nodes

6

(b)

Fig. 4. Performance accuracy comparison of particle filters with fading σR =1.35x10−4 mV (a) target velocity = 0.1 m/s, (b) target velocity = 0.3 m/s

diverges. Simulation results indicate a decrease in tracking performance with increase of noise and fast target movement. Figs. 4 and 5 show the RMSE versus Nd using 100 runs under the Rayleigh fading model. Like for the lognormal shadowing model, simulation results indicate a decrease in tracking performance as noise and target velocity increase. Note that, using 100 particles the bootstrap PF performs better than the UPF while using the lognormal shadowing model the opposite is true. This behavior could be due to the approximation of the proposal distribution that UPF uses. Indeed, while PF uses the prior distribution as proposal distribution, UPF uses a distribution depending also on observations which, in a fading channel, are characterized by fluctuations of RSS from the mean value. For each value of Nd , 10000 different random configurations and a maximum range between node and cluster head equal to 10 meters have been considered. The DPF and DUPF computational complexity is given by O(L3 ), while the Kaplan computational complexity is given by O(L2 ). Specifically, the time to process the bootstrap particle filter with 100 and 500 particles is equal to 1,605

262

L. Arienzo and M. Longo

0.48

1

PF with Rayleigh fading UPF with Rayleigh fading

0.46 0.45 0.44 0.43 2

PF with Rayleigh fading UPF with Rayleigh fading

0.9 RMSE (meters)

RMSE (meters)

0.47

0.8 0.7 0.6

3 4 5 Number of Selected Nodes

6

0.5 3

4 5 Number of Selected Nodes

(a)

6

(b)

Fig. 5. Performance accuracy comparison of particle filters with σR =4.5x10−4 mV (a) target velocity = 0.1 m/s, (b) target velocity = 0.3 m/s

fading

0.3

PF UPF DPF DUPF

0.9 0.8

RMSE (meters)

Energy Consumption (J)

1

0.7 0.6 0.5 0.4

0.1

Greedy Kaplan

0.3 0.2 2

0.2

3

4 5 Number of selected nodes

6

0.2

0.3

0.4

0.5

0.6

Energy Consumption (J)

(a) Energy consumption of Kaplan and (b) RMSE vs. Energy consumption for difGreedy algorithms vs. number of nodes ferent tracking algorithms using Greedy selection Fig. 6. Energy consumption comparison

sec and 8,052 sec respectively, with three active nodes, while the time to process the unscented particle filter with 100 particles is equal to 3,281 sec. In conclusion, the UPF is less computational efficient than the PF but performs a more accurate estimation of the target’s position compared to the UPF. 6.3

Energy Consumption

Fig. 6 illustrates the energy consumption of node selection algorithms using the residual energy-based metrics defined in (2). In Fig. 6(a), the energy consumption of the greedy algorithm is compared to that of Kaplan algorithm as the number of selected nodes increases. It is very clear that the greedy selection algorithm outperforms the Kaplan selection algorithm in terms of energy efficiency. In Fig. 6(b), RMSE vs. energy consumption of PF, DPF, UPF, DUPF algorithms using greedy selection is shown. Results indicate an increase of the energy consumption with growing number of nodes.

Energy-Efficient Target Tracking

263

In conclusion, the energy consumption increases with the number of active nodes; on the other hand the tracking error decreases as the the number of active nodes increases. A tradeoff between the performance and the number of nodes is needed to save energy.

7

Conclusions

The focus of the article was the energy-efficient and collaborative target tracking in wireless sensor networks. The tracking problem is formulated as a crosslayer optimization with an aim to minimize the total energy consumption in the cluster. The node selection procedures were integrated into several tracking algorithms: the bootstrap particle filter, the unscented particle filter (both centralized and distributed schemes) and the distributed sigma-point information filter. They have been implemented and tested on simulated data, to evaluate the tracking performances for linear dynamic models and for both a Gaussian and a Rayleigh fading observation model. Extensive simulations showed that, overall, the target tracking system yields better accuracy for lower velocity of the target while performances get worse as the noise and the target velocity increase. Simulations also indicate that the algorithms here proposed compare favorably with previously proposed ones in terms of accuracy and computational complexity and energy consumption.

References 1. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless Sensor Networks: A Survey. Computer Networks (Elsevier), 393–422 (March 2002) 2. Ihler, A.T., et al.: Nonparametric belief propagation for self-localization of sensor networks. IEEE Journal on Selected Areas in Communications 23(4) (April 2005) 3. Lee, J., Cho, K., Lee, S., Kwon, T., Choi, Y.: Distributed and energy-efficient target localization and tracking in wireless sensor netwoks. Computer Communications 29, 2494–2505 (2006) 4. Vercauteren, T., Wang, X.: Decentralized Sigma-Point Information Filters for Target Tracking in Collaborative Sensor Networks. IEEE Transactions on Signal Processing 53(8) (August 2005) 5. Oshman, Y., Davidson, P.: Optimization of observer trajectories for bearings-only target localization. IEEE Transactions on Aerospace and Electronic Systems 35, 892–902 (1999) 6. Ertin, E., Fisher, J.W., Potter, L.C.: Maximum Mutual Information Principle for Dynamic Sensor Query Problems. In: Zhao, F., Guibas, L.J. (eds.) IPSN 2003. LNCS, vol. 2634, pp. 405–416. Springer, Heidelberg (2003) 7. Wang, H., Pottie, G., Yao, K., Estrin, D.: Entropy-based Sensor Selection Heuristic for Target Localization. In: International Workshop on Information Processing in Sensor Networks (IPSN), Berkeley, California, April 26-27 (2004) 8. Kaplan, L.M.: Global node selection for localization in a distributed sensor network. IEEE Transactions on Aerospace and Electronics Systems 42(1), 113–135 (2006) 9. Kaplan, L.M.: Local node selection for localization in a distributed sensor network. IEEE Transactions on Aerospace and Electronics Systems 42(1), 136–146 (2006)

264

L. Arienzo and M. Longo

10. Wang, Q., Chen, W.-P., Zheng, R., Lee, K., Sha, L.: Acoustic Target Tracking Using Tiny Wireless Sensor Devices. In: Zhao, F., Guibas, L.J. (eds.) IPSN 2003. LNCS, vol. 2634, pp. 642–657. Springer, Heidelberg (2003) 11. Chen, W.-P., Hou, J.C., Sha, L.: Dynamic Clustering for Acoustic Target Tracking in Wireless Sensor Networks. IEEE Trans. on Mobile Computing 3(3), 258–271 (2004) 12. Kung, H.T., Vlah, D.: Efficient Location Tracking Using Sensor Networks. In: WCNC (March 2003) 13. Lin, C.-Y., Tseng, Y.-C.: Structures for In-Network Moving Object Tracking in Wireless Sensor Networks. In: BROADNETS (2004) 14. Zhang, W., Cao, G.: DCTC: Dynamic Convoy Tree-Based Collaboration for Target Tracking in Sensor Networks. IEEE Trans. on Wireless Communications (2004) 15. Liu, J., Reich, J., Zhao, F.: Collaborative in-network processing for target tracking. EURASIP, J. Appl. Signal Processing, 378–391 (2003) 16. Zhao, F., Shin, J., Reich, J.: Information-driven dynamic sensor collaboration for tracking applications. IEEE Signal Processing Magazine 19(2), 61–72 (2002) 17. Kreucher, C.M., Hero, A.O., Kastella, K.D., Morelande, M.R.: An InformationBased Approach to sensor Management in Large Dynamic Networks. Proceeding of IEEE 95, 978–999 (2007) 18. Heinzelman, W.B., Chandrakasan, A.P., Balakrishnan, H.: Energy-Efficient Communication Protocol for Wireless Microsensor Networks. In: Proceedings of 33rd Hawaii International Conference on System Sciences (HICSS 2000), Maui, Hawaii (January 2000) 19. Rappaport, T.S.: Wireless Communications. Principles and Practice, 2nd edn. Prentice Hall, Englewood Cliffs 20. Doucet, A., de Freitas, N., Gordon, N.: Sequential monte carlo methods in practice. Springer, Heidelberg (2001) 21. Gordon, N.J., Salmond, D.J., Smith, A.F.M.: Novel approach to nonlinear/nonlinear gaussian bayesian state estimation. IEEE Proceedings-F 140(2), 107– 113 (1993) 22. Van der Merwe, R., Doucet, A., de Freitas, N., Wan, E.: The Unscented Particle Filter. In: NIPS, pp. 584–590 (2000) 23. Guo, D., Wang, X.: Dynamic sensor collaboration via sequential monte carlo. IEEE JSAC 22(6), 1037–1047 (2004) 24. Arienzo, L., Longo, M.: Energy-Efficient Tracking Strategy for Wireless Sensor Networks. In: Proceedings of IEEE MASS 2008, Workshop on Localized Communication and Topology Protocols for Ad hoc Networks, Atlanta, Georgia (September 2008)

QoS for Wireless Sensor Networks: Service Differentiation at the MAC Sub-Layer Bilel Nefzi and Ye-Qiong Song LORIA - Nancy University INPL Campus Scientifique - BP 239 - 54506 Vandoeuvre-Les-Nancy Cedex, France {Bilel.Nefzi,Ye-Qiong.Song}@loria.fr

Abstract. Providing service differentiation in wireless sensor networks while proposing simple and highly scalable solution is a challenging problem. We retain the use of CSMA/CA as access protocol because of its simplicity, versatility and good scalability properties. We developed CoSenS on the top of it to address its weaknesses while facilitating the implementation of scheduling policies. In this article, we propose a simple and scalable service differentiation solution; we implement fixed priority and earliest deadline first on the top of CoSenS. The simulation analysis shows that our solution greatly enhances end-to-end delay, reliability and deadline meet ratio for urgent traffic while not degrading best effort traffic compared to IEEE 802.15.4 original protocol and IEEE 802.15.4 implementing these scheduling policies. Keywords: wireless sensor networks, QoS, scheduling, csma/ca, MAC protocols.

1 Introduction Many sensor applications require quality of service (QoS) support in the communication stack they use in order to work properly. These requirements include but are not limited to reliability, end to end delay and service differentiation. Most of the proposed solutions are TDMA based [7]. However, the use of TDMA scheme needs careful network configuration for efficient time-slot allocation and it is not scalable in general for large scale WSN and not robust for automatically adapting to the changes in the network (nodes mobility and death, environment changes, etc). CSMA/CA is a well known access protocol. It is simple to implement and does not need any synchronization between nodes. This makes it a suitable choice for WSN. CSMA/CA performs well in light traffic mode but the performance quickly degrades in heavy traffic (lack of reliability, high end-to-end delays and low throughput). This makes it not suitable for QoS. To address QoS offer, we propose firstly CoSenS, a simple MAC layer protocol, which is implemented on the top of CSMA/CA to overcome its weaknesses. The idea of CoSenS is that a router does not retransmit packets as they arrive. Instead, it collects data from its children and other neighbor routers and then sends them into a burst during a period of time that we call transmission period. The performance analysis shows that CoSenS greatly enhances throughput, end to end delay and reliability. In addition, J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 265–280, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

266

B. Nefzi and Y.-Q. Song

since packets are queued at the router, CoSenS allows us to design efficient scheduling algorithms for providing differentiated QoS. Finally, for service differentiation, we propose two scheduling policies that we implemented on the top of CoSenS. The first one is a fixed priority scheduler. We consider two types of traffic; periodic traffic and event driven traffic which has higher priority. The second one is Earliest Deadline First (EDF). We suppose that some deployed applications provide a deadline for their data in order to work properly. The organization of this paper is as following. The related work is given next. CoSenS is described in Sec. 3. Section 4 describes the proposed scheduling policies. In Sec. 5 we present the simulation results. Section 6 concludes the paper.

2 Related Work An implicit prioritized access protocol (I-EDF) [1] is designed especially for hard real time. A cellular backbone network is adopted and different frequency channels are assigned. In a cell, time is divided into frames and all nodes are frame synchronized and follow earliest deadline first (EDF) schedule for packet transmission to guarantee bounded delay. A capable router node is required at the center of each cell and equipped with two transceivers for separate transmission and reception. Inter-cell communication is supported by a globally synchronized TDMA scheme and the messages are ordered by their earliest deadlines too. The mixed FDMA-TDMA scheme offers a collisionfree solution. Simulations show that I-EDF can provide high throughput and low latency even in heavy loads. However, the system architecture and requirements appear impractical for WSNs. Nodes are assumed synchronized. Routers need to be deployed specifically following the cellular structure and topology knowledge is required. PEDAMACS [2] is a TDMA-based MAC protocol that aims to achieve both energy efficiency and delay guarantee. It considers a special class of sensor networks with high powered access point (AP) which can reach all nodes in one hop and with nodes periodically generating packets. Topology information is gathered by AP and a scheduling algorithm is then adopted to determine when a node should transmit and receive data. PEDAMACS guarantees bounded delay and eliminates network congestion. However, the requirement of powerful AP has restricted the protocol to only few applications and reduced its attractiveness. Munir et al. [3] addressed link burstiness and proposed a scheduling algorithm which bounds the latency of a set of periodic streams while providing 100 % of data reliability over bursty links. The protocol used a TDMA like mechanism for data transmissions. GTS mechanism of IEEE 802.15.4 [4] and its various enhancements like i-GAME [5] are also possible solutions for providing QoS support in WSN. In the beacon-enabled synchronized mode, the PAN coordinator may allocate portions of the active superframe to form guaranteed time slots (GTSs). The major drawback of this mechanism is that only seven nodes can request GTS allocations. i-GAME solves this problem by letting multiple nodes share the same GTS. An admission control algorithm is used then to accept time slot requests if the requirements do not exceed available resources.

QoS for Wireless Sensor Networks

267

Although the use of TDMA schemes ensure deterministic performance, it still needs careful network configuration for efficient time-slot allocation and it is not scalable in general for large scale WSN and not robust for automatically adapting to the changes in the network. Our aim is to develop much simpler solution which does not require any configuration effort and can self adapt to the dynamic network evolutions. Koubaa et al. [6] proposed a differentiated MAC protocol in which traffics are categorized into high and low priority queues which employ different CSMA/CA settings. The result offers a heuristic solution to provide different QoS for messages of different priorities. This solution is only suitable for light traffic environment because the performance of CSMA/CA on which this solution is based quickly degrades in heavy traffic load. A full survey about Real time QoS support in WSN can be found here [7]. We retain the widely spread CSMA/CA protocol for its simplicity and good scalability properties and implement CoSenS on the top of it. CoSenS addresses at the same time CSMA/CA’s drawbacks and enables the implementation of scheduling policies such as EDF or fixed priority. Thus our scheme is scalable because it does not require any synchronization between nodes. In addition, CoSenS performs well in medium and heavy network load where delivering critical data reliably and rapidly to the sink is more important.

3 CoSenS 3.1 General Description We consider a 2-tiers architecture network composed of routers to which are associated simple nodes (typically FFD and RFD of ZigBee [8]). A simple node is a device that generally has measurement sensors built in and has limited routing decisions capabilities. A router is a device that implements full routing and network management protocols. However, it can also act as a simple device when its routing and management capabilities are disabled. A simple node is associated to a router. In that case, the association forms a childfather relation where the child is the simple node and the father is the router. We consider the case where every simple node has only one father. A router can be associated with more than one child and all routers form an ad-hoc network. In this case, any routing protocol, like AODV or ”Hierarchical Tree Routing” (HTR) of ZigBee for instance, can be used to establish routes between routers1. So, if a simple node wants to send data to a receiver node, it simply forwards them to its father. Besides, all requests like obtaining information about a destination are also addressed to the father. The basic medium access protocol for all nodes (simple nodes and routers) is CSMA/ CA but with different parameters for each type of node. We used the unslotted version proposed by IEEE 802.15.4 standard. CoSenS is implemented on the top of CSMA/CA protocol. It is enabled for routers and disabled for simple nodes by default (in the rest of the section, routers are used to describe CoSenS operating mode). The MAC layer operates on the top of the ”2450 MHz DSSS IEEE 805.15.4” physical layer. The data rate is equal to 250 kb/s. 1

In the simulations, we used HTR.

268

B. Nefzi and Y.-Q. Song

3.2 Energy In this work, we suppose that routers do not have severe energy constraints. This is a reasonable assumption in many real-world deployment scenarios, like building monitoring systems, monitoring of industrial environments or patient monitoring in hospitals. In the latter scenario for example, the patient carry simple nodes to monitor its vitals. These simple nodes transmit data to a router located in the room. All routers in the hospital form an ad-hoc network and transmit data to the sink. Simple nodes, are battery supplied. Hence, energy efficiency have to be taken into account for these nodes. Since no synchronization is required for simple nodes using CoSenS, they are always in sleep unless they have a message to send. However, a first extension of CoSenS has been already done to tackle energy and throughput tradeoff [9]. 3.3 Basic Rules CoSenS has four basic rules. First, routers have the priority to access the medium over simple nodes. This is done by setting lower backoff exponent value (which is a parameter of CSMA/CA). Second, a router does not transmit packets one by one upon their arrival. Instead, it waits for a period of time, WP, and collects data either from its children or from its neighbor routers. Third, after the end of the WP, the router starts transmitting all packets queued in its buffer in a single burst during the TP. Finally, at the end of the TP, the router starts another cycle and goes again to the waiting for reception state. WP and TP are illustrated in the example given by the Fig. 1.

Fig. 1. An example of how CoSenS works

Note that since simple nodes use CSMA/CA, they transmit data during the WP of their respective fathers because the channel can be free only during that period (CSMA/ CA checks the channel status before transmitting). 3.4 Transmission Period The collected data during the waiting period are transmitted in a burst. CSMA/CA protocol is used to transmit only the first packet. Then, the remaining packets are directly transmitted upon the reception of the acknowledgement (ack)2 . If an ack is not received, the transmitted packet is retransmitted using CSMA/CA protocol again. Burst 2

Acknowledgments are enabled in CoSenS.

QoS for Wireless Sensor Networks

269

transmission resumes if the packet is successfully transmitted; The remaining packets will be sent directly after the reception of the ack. TP has a variable duration which depends on the number of collected packets during WP and the eventual retransmissions. We note that all queued packets are transmitted during the current TP. So, the next WP will start with an empty queue. The use of CSMA/CA for the first packet to transmit ensures also that the TPs of neighbor routers do not collide. In fact, if a router does not win the access to the medium at the beginning of its TP, its WP is extended until the end of the TP of the transmitting router. 3.5 Waiting Period The length W P depends on the amount of incoming traffic; the higher the traffic volume is, the longer the WP is. Consider the example of a waiting and transmission cycle given by the Fig. 2. Intuitively, CoSenS performs better than CSMA/CA if the sum of the idle delays in the WP plus one Backoff Period (BP) (of the first transmitted packet in the TP) is lesser than the sum of the BPs which will be generated at the transmission of the received packets if CSMA/CA is used (1). Hence WP must satisfy the condition given by (2). In (1), Npkts is the number of received packets during the WP and pktSvcTi is the service time of packet i. N pkts

∑

N pkts

BPi ≥ W P −

i=1

∑

pktSvcTi + BPT P

(1)

i=1 N pkts −1

WP ≤

∑ ∑

BPi + sumSvcT

i=1 N pkts −1

=

N pkts

BPi +

∑

pktSvcTi

(2)

i=1

(3)

i=1

WP is adapted at runtime as follows. N pkts and sumSvcT are estimated at each cycle. Then, WP is set according to (4). BP is given by 5 where aUnitBacko f f Period = 0, 32 ms (a constant defined in IEEE 802.15.4 standard) and macMinBE = 2. W P = (Npkts − 1) · BP + sumSvcT

(4)

BP = 2macMinBE−1 · aUnitBacko f f Period

(5)

The Estimation Algorithm: Let’s define the set Ω ∈ IN as the set of W Ps during which the router receives data and W Pk , k ∈ Ω these W Ps. Let us denote by sumSvcT and Npkts the exponential moving average of sumSvcT k and N pkts k , k ∈ Ω, respectively. sumSvcT k = (1 − α) · sumSvcT k−1 + α · sumSvcT k−1

(6)

Npkts k = (1 − α) · N pkts k−1 + α · N pkts k−1

(7)

270

B. Nefzi and Y.-Q. Song

Fig. 2. Comparison between CoSenS and CSMA/CA. ”Backoff expires” corresponds to the expiration of the backoff timer for the transmitting node. In CoSenS, it is included in the idle time.

α is the smoothing factor. We used a non linear filter where α is bigger when sumSvcT k−1 ≥ sumSvcT k−1 in (6) and Npkts k−1 ≥ Npkts k−1 in (7) allowing these two values to adapt more swiftly to traffic increase. N pkts k and sumSvcT k are evaluated before the beginning of the kth W P and then used in the Algorithm 1 to determine the duration of the kth W P. Algorithm 1. Estimation algorithm /* Update N pkts k and sumSvcT k */ /* α is set according the the kth value */ sumSvcT k = (1 − α) · sumSvcT k−1 + α · sumSvcT k−1 ; Npkts k = (1 − α) · N pkts k−1 + α · Npkts k−1 ; /* Update WP */ W Pk = (Npkts k − 1) · BP + sumSvcT k W Pk = max(W Pmin , min(W P,W Pmax ))

4 Packet Scheduling Policies 4.1 Fixed Priority In this packet scheduling policy, we consider two types of traffic; periodic and eventdriven. Hence, we have 2 priorities. We consider that event driven traffic has higher priority. Generally, an event corresponds to a measure of a physical phenomenon which is different from the normal value. Thus, these events must be transmitted as quickly as possible and as reliably as possible. Periodic data corresponds to periodic updates which report no or minor fluctuations of the measured phenomenon. We consider that this type of data is delay tolerant. Event-driven data are transmitted first. The transmission of packets which are classified as periodic occurs after the transmission of all event-driven data. For each traffic class, packets are served according to their order of arrival.

QoS for Wireless Sensor Networks

271

4.2 Earliest Deadline First Some applications may require a deadline to packet’s end-to-end delay. In this case, EDF is a suitable solution. We define t prem as the remaining time until the deadline of packet p expires. It is included in the packet header. The application at the source node initializes this variable with the deadline. Then, it is updated at each hop to account for queuing, contention, and transmission delays. We define prioRp as the priority of packet p calculated by router R (8). prioRp =

1 t prem

(8)

Upon packet’s reception, the router calculates its priority and inserts it to the queue according to this priority; higher priorities are transmitted first. Hence, the waiting time of a packet at the receiving router is not taken into account immediately for priority calculation. We note that the packets that do not provide or have missed their deadlines are scheduled at the tail of the transmission queue according to their order of arrival. In the original version of EDF [10], packets which miss their deadlines can still be scheduled near the queue head. This delays further the transmission of queued packets which increases the missed deadline ratio especially in an overloaded network. t prem is updated at each router when it acquires the channel and is about to transmit, as follows. First, it time-stamps the packet when it is received and when it becomes qdel the head of the queue to account for queuing delay, t p . The router calculates then the cont contention delay (caused by CSMA/CA), t p , and the transmission delay, t tr p . In case of retransmission due to packet or acknowledgment loss, the router shall update t pcont and t tr p to account for these additional delays. In this case, the acknowledgment delay has to be taken into account also (added to t pcont ). Propagation delays are ignored as their values are negligible in comparison with queuing and transmission delays. Finally t prem is updated as follows: t prem = t prem − t pqdel − t pcont − t tr p

(9)

We note that t prem is computed locally at each node and does not require any global time synchronization.

5 Simulation Results We present a comparative study of CoSenS with IEEE 802.15.4 using the unslotted version of CSMA/CA (denoted IEEE 802.15.4 in the rest of the document) . FP and EDF scheduling policies (in addition to FIFO) are implemented on both protocols. The metrics used are end to end delay, successful transmission rate (STR) and deadline meet ratio (DMR). The end to end delay is equal to the average end to end delays of all successfully received packets by the receivers application layer. The STR is equal to the total number of successfully received bits by the receivers application layer divided by the total number of generated bits by all nodes application layer. The DMR is defined

272

B. Nefzi and Y.-Q. Song

Table 1. General simulation parameters. CSMA/CA parameters are used by CoSenS and IEEE 802.15.4. Variable

value

data rate PHY simple node Coverage Router node APL Packet length MAC simple node CSMA/CA macMinBE router node α1 α α2 CoSenS W Pmin WP W Pmax Cm Zigbee Tree Routing parameters Rm Lm Simulated duration

√250 kb/s 2 · 100 m 200 m 402 bits 500 bits 3 2 0.08 0.1 0.001 s 0.07 s 12 5 6 6 min

as the number of packets which meet their deadlines divided by the number of received packets. The general simulation parameters are summarized in Table 1. The simulator used is OPNET[11] (Modeler v15.0.A PL3). 5.1 Network and Scenarios Description Network: We consider a multi-hop network composed of 25 routers and 125 simple nodes deployed in a 1000x1000m2 sensor area. The nodes are organized as shown in Fig. 3. The sensor area is divided in 25 equal squares. For each square, one router is located at the center and five nodes are randomly deployed. The routing protocol used is the ZigBee HTR protocol. The simulation duration is equal to six minutes; the first minute is reserved for node association and the remaining five minutes for data generation and transmission. This network represent a patient monitoring system or an equivalent system as suggested in Sec. 3.2. Scenarios: The first set of simulations compare the intrinsic performance of CoSenS and IEEE 802.15.4 using only Periodic or Poisson traffic. In the second set of simulations three scenarios are simulated. In each one, a group of 10%, 25%, and 50% of the total number of nodes is randomly selected and generates only Poisson traffic with a inter-arrival time equal to 1s. The rest of nodes generate Periodic traffic. Poisson traffic is considered as event traffic and has higher priority. For each scenario, five simulations are conducted and a different group of nodes is randomly selected each time. The results are then averaged. In all scenarios, we varied the intensity of periodic traffic and studied the performance of both protocols using FIFO, FP and EDF. In both sets, all nodes send data to the root node located in the center of the network. The third set of simulation compares the performance of CoSenS and IEEE 802.15.4 using EDF in the following scenario. For each event generated by a simple node, the Root node issues a response event to it. These response messages represent the actions the monitor of the network takes in response to an event.

QoS for Wireless Sensor Networks

273

Fig. 3. Simulation network. Blue discs represent simple nodes. White lines correspond to the formed tree between routers (router-simple nodes association is not drawn). The root node is located in the center of the network.

5.2 Results Hereafter, we present the results of the three sets of simulations. Intrinsic Performance: Figure 4(a) shows the end-to-end delay performance as a function of the total load generated by the application layer (sum of event and periodic traffic load). We note that in the rest of the figures, we used this definition for the load. The performance of CoSenS protocol is better for both types of traffic. In the case of a Poisson traffic, IEEE 802.15.4 and CoSenS obtains similar end to end delays for light traffic. After that, CoSenS outperforms IEEE 802.15.4. As far as the load increases, the difference in terms of end to end delay becomes more and more obvious. In fact, as the load increases, the router collects many packets and then transmits them in burst during the T P. This minimizes the transmission time (CoSenS uses CSMA/CA only for the

274

B. Nefzi and Y.-Q. Song 0.5

0.4

CoSenS − FIFO − periodic traffic IEEE 802.15.4 − FIFO − periodic traffic CoSenS − FIFO − Poisson traffic IEEE 802.15.4 − FIFO − Poisson traffic

1

0.9 successful transmission rate

0.45

end−to−end delay (s)

0.35 0.3 0.25 0.2 0.15

0.8

0.7

0.6

0.5

0.1 0.4

0.05 0 0

1

2

3 4 load (b/s)

5

(a) End-to-end delay.

6

7 4

x 10

0

CoSenS − FIFO − periodic traffic IEEE 802.15.4 − FIFO − periodic traffic CoSenS − FIFO − Poisson traffic IEEE 802.15.4 − FIFO − Poisson traffic 2

4

6 load (b/s)

8

10 4

x 10

(b) Successful transmission rate.

Fig. 4. End-to-end delay and successful transmission rate comparison between CoSenS and IEEE 802.15.4 using Periodic and Poisson traffic

first packet). In addition, the W P is adapted according to the incoming traffic. Hence, the idle time (where the router is not receiving data during the WP) and the transmission time are minimized. This explains the good performance of CoSenS. For Periodic traffic, CoSenS always outperforms IEEE 802.15.4. In fact, the 125 nodes generate periodic traffic with an initial offset randomly selected between [0,1). For example, if the period T is greater than 1 second, a packet is generated each 0.008s in average between [T,T+1). Hence, the network is momentarily overload at each period. Since CoSenS quickly adapts the WP (due to the non linear filter that we used), it takes advantage of the situation; many packets are collected which ameliorates the end-to-end delay like explained for Poisson traffic. The saturation load for CoSenS and IEEE 802.15.4 are reached around 50 kb/s and 75 kb/s, respectively. The performance of both protocols falls after that. For STR (Fig. 4(b)) both protocols obtain 100% before their respective saturation point. This is due to the acknowledgment mechanism. The performance fall after that but CoSenS obtains better results. CoSenS and IEEE 802.15.4 with Scheduling Performance: Figures 5(a), 5(b) and 5(c) show the end-to-end delays of event data for the three scenarios. The performance of event traffic is greatly enhanced using FP for both protocols. IEEE 802.15.4 with FP performs well in case of low event traffic load (10% of nodes generating events). In fact, as the load increases, the routers in IEEE 802.15.4 become the network bottlenecks because the packets can easily be transmitted to them (the maximum MAC throughput of CSMA/CA is around half the MAC data rate, 125 kb/s). Since the event traffic is relatively small compared to periodic traffic and is always inserted at the queue head, its end-to-end delay is small. However, the performance decreases as the event load increases (25% and 50% of nodes generating events). CoSenS with FP almost always obtains better results than IEEE 802.15.4. Also, it efficiently schedules events in low event traffic load scenario leading to a better end-toend delay compared to FIFO version. In addition, the performance increases as the event

QoS for Wireless Sensor Networks

0.5

0.5 CoSenS − FIFO − 10% − event CoSenS − FP − 10% − event IEEE 802.15.4 − FIFO − 10% − event IEEE 802.15.4 − FP − 10% − event

0.45 0.4

CoSenS − FIFO − 25% − event CoSenS − FP − 25% − event IEEE 802.15.4 − FIFO − 25% − event IEEE 802.15.4 − FP − 25% − event

0.45 0.4 0.35 end−to−end delay (s)

0.35 end−to−end delay (s)

275

0.3 0.25 0.2 0.15

0.3 0.25 0.2 0.15

0.1

0.1

0.05

0.05

0 1

2

3

4

5 load (b/s)

6

7

8

0 2

9

3

4

5 load (b/s)

4

x 10

(a) Event traffic delay - 10% case.

6

7

8 4

x 10

(b) Event traffic delay - 25% case.

0.5 CoSenS − FIFO − 50% − event CoSenS − FP − 50% − event IEEE 802.15.4 − FIFO − 50% − event IEEE 802.15.4 − FP − 50% − event

0.45 0.4

1.2

CoSenS − FIFO − 10% − periodic CoSenS − FP − 10% − periodic IEEE 802.15.4 − FIFO − 10% − periodic IEEE 802.15.4 − FP − 10% − periodic

1 end−to−end delay (s)

end−to−end delay (s)

0.35 0.3 0.25 0.2

0.8

0.6

0.4

0.15 0.1

0.2 0.05 0 3

3.5

4

4.5

5 5.5 load (b/s)

6

6.5

0 1

7

2

3

4 load (b/s)

4

x 10

(c) Event traffic delay - 50% case.

5

6

7 4

x 10

(d) Periodic traffic delay - 10% case. 0.5

1.2

CoSenS − FIFO − 25% − periodic CoSenS − FP − 25% − periodic IEEE 802.15.4 − FIFO − 25% − periodic IEEE 802.15.4 − FP − 25% − periodic

CoSenS − FIFO − 50% − periodic CoSenS − FP − 50% − periodic IEEE 802.15.4 − FIFO − 50% − periodic IEEE 802.15.4 − FP − 50% − periodic

0.45 0.4

1 end−to−end delay (s)

end−to−end delay (s)

0.35 0.8

0.6

0.4

0.3 0.25 0.2 0.15 0.1

0.2 0.05 0 2

3

4

5 load (b/s)

6

7

8 4

x 10

(e) Periodic traffic delay - 25% case.

0 3

3.5

4

4.5

5 5.5 load (b/s)

6

6.5

7 4

x 10

(f) Periodic traffic delay - 50% case.

Fig. 5. Event and Periodic end-to-end delays comparison between CoSenS and IEEE 802.15.4 using FIFO and FP scheduling policies

276

B. Nefzi and Y.-Q. Song

1.05

1.05

1

1 0.95 successful transmission rate

successful transmission rate

0.95 0.9 0.85 0.8 0.75 0.7

0.9 0.85 0.8 0.75 0.7

0.65 0.6 0.55 1

CoSenS − FIFO − 10% − event IEEE 802.15.4 − FIFO − 10% − event CoSenS/IEEE 802.15.4 − FP/EDF − 10% − event 2

3

4

5 load (b/s)

6

7

0.65 CoSenS − FIFO & CoSenS/IEEE 802.15.4 − FP/EDF − 25% − event IEEE 802.15.4 − FIFO − 25 % − event

8

9

2

3

4

5 load (b/s)

4

x 10

(a) Event traffic STR - 10% case.

6

7

8 4

x 10

(b) Event traffic STR - 25% case.

1.05 1

1

successful transmission rate

successful transmission rate

0.95 0.9 0.85 0.8 0.75

0.9

0.8

0.7

0.7 0.6 0.65 CoSenS − FIFO & CoSenS/IEEE 802.15.4 − FP/EDF − 50% − event IEEE 802.15.4 − FIFO − 50% − event

3

3.5

4

4.5

5 5.5 load (b/s)

6

6.5

0.5 1

7

0.9

0.9 successful transmission rate

successful transmission rate

1

0.8

0.7

0.6

4

5 load (b/s)

4

6

7

8 4

x 10

(e) Periodic traffic STR - 25% case.

5 load (b/s)

6

7

8

9 4

x 10

0.8

0.7

0.6

0.5

CoSenS − FIFO/FP/EDF − 25% − periodic IEEE 802.15.4 − FIFO − 25% − periodic IEEE 802.15.4 − FP/EDF − 25% − periodic 3

3

(d) Periodic traffic STR - 10% case.

1

0.4 2

2

4

x 10

(c) Event traffic STR - 50% case.

0.5

CoSenS − FIFO − 10% − periodic CoSenS − FP/EDF − 10% − periodic IEEE 802.15.4 − FIFO − 10% − periodic IEEE 802.15.4 − FP/EDF − 10% − periodic

0.4 3

CoSenS − FIFO/FP/EDF − 50% − periodic IEEE 802.15.4 − FIFO − 50% − periodic IEEE 802.15.4 − FP/EDF − 50% − periodic 3.5

4

4.5

5 5.5 load (b/s)

6

6.5

7 4

x 10

(f) Periodic traffic STR - 50% case.

Fig. 6. Successful transmission rate comparison between CoSenS and IEEE 802.15.4 using FIFO, FP and EDF scheduling policies. Here, points which are greater than 0.99 are ceiled.

QoS for Wireless Sensor Networks

277

traffic load increases. This is due to two factors. On one hand, CoSenS performs well below its saturation load. On the other hand, the total load decreases below the saturation point of CoSenS as the event load increases. This is because each selected node generate a traffic with a constant inter-arrival time. Figures 5(d), 5(e) and 5(f) show the end-to-end delays of periodic data for the three scenarios. The results show that the FP scheduling policy does not degrade too much the end-to-end delay of periodic data for both protocols. The results in Fig. 5(d) are similar to the intrinsic performance results since the event data is low. CoSenS obtains better results than IEEE 802.15.4. Thus, CoSenS can handle high event loads more efficiently in terms of end-to-end delay while not degrading much the performance of periodic data. This makes it a good choice for QoS support in WSN. The results of STR for event data are illustrated in Figures 6(a), 6(b) and 6(c). Using FP and EDF greatly enhances the performance of both protocols in comparison with FIFO; they obtain 100% in all event load cases. However, for IEEE 802.15.4, this performance is reached at the cost of an additional degradation of the performance of periodic data. In fact, we observe in figures 6(d), 6(e) and 6(f) that increasing the event load decreases the STR of periodic data after the saturation load when using FP or EDF. Before the saturation load, the acknowledgment mechanism ensures a STR of 100%. CoSenS saturation point is higher than IEEE 802.15.4. Thus, the STR of event loads is enhanced with a minimum degradation of the STR of periodic data. In order to study the DMR, we set the deadline of event data to 0.08s. This time corresponds to average delay obtained by IEEE 802.15.4 for the event data in the high event load case (50%) and the minimum generated periodic traffic. Using EDF scheduling policy enhances the DMR for both protocols in all event load cases. For IEEE 802.15.4 using FIFO, we observe that for high loads, the DMR increases. This is because the number of received packets decreases significantly (STR decreases). So there is higher chances that the arrived packets meet their deadlines. We remember that the DMR is calculated based on the received packets and not all generated packets. We observe also that the performance of CoSenS with FIFO is better than IEEE 802.15.4 one except after the saturation load. This is because the STR of CoSenS is largely greater than IEEE 802.15.4 one even after it reaches its saturation point for both types of traffic. Nonetheless, the fact that CoSenS provides good STR for periodic traffic after the saturation load may harm the DMR of event data at these loads which is not desirable. Hence, a congestion control mechanism have to be implemented to prevent such a situation. CoSenS with EDF largely outperforms IEEE 802.15.4 with EDF. We observe that below the saturation load of CoSenS, the difference between both protocols increases as the total load increases. This is due to the amelioration of end-to-end delay for medium and high loads. Moreover, CoSenS ameliorates its DMR while it falls for IEEE 802.15.4 as the event load increases (three scenarios). CoSenS and IEEE 802.15.4 with Response Events Performance: In this set of simulations we set the deadline of event data (both types) to 0.08s. We consider the case where a group of 10% of the total number of nodes is randomly selected and generates Poisson traffic. The results (Fig. 8) are similar to the DMR study of the first scenario (10%) of the previous set of simulations. Before its saturation load, CoSenS using EDF

B. Nefzi and Y.-Q. Song

1 CoSenS − FIFO − 10% − event CoSenS − EDF − 10% − event IEEE 802.15.4 − FIFO − 10% − event IEEE 802.15.4 − EDF − 10% − event

0.9 0.8

deadline meet ratio

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1

2

3

4

5 load (b/s)

6

7

8

9 4

x 10

(a) 10% case. 1 CoSenS − FIFO − 25% − event CoSenS − EDF − 25% − event IEEE 802.15.4 − FIFO − 25% − event IEEE 802.15.4 − EDF − 25% − event

0.9

deadline meet ratio

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 2

3

4

5 load (b/s)

6

7

8 4

x 10

(b) 25% case. 0.9 CoSenS − FIFO − 50% − event CoSenS − EDF − 50% − event IEEE 802.15.4 − FIFO − 50% − event IEEE 802.15.4 − EDF − 50% − event

0.8 0.7 deadline meet ratio

278

0.6

0.5 0.4

0.3 0.2 0.1 3

3.5

4

4.5

5 5.5 load (b/s)

6

6.5

7 4

x 10

(c) 50% case. Fig. 7. Deadline meet ratio of CoSenS and IEEE 802.15.4 using EDF

QoS for Wireless Sensor Networks

279

1 CoSenS − events reported to Root IEEE 802.15.4 − events reported to Root CoSenS − Response events gen by Root IEEE 802.15.4 − Response events gen by Root

0.9

1

0.9 successful transmission rate

deadline meet ratio

0.8 0.7 0.6 0.5 0.4

0.8

0.7

0.6

0.3 0.5 0.2 0.1 1

2

3

4

5 load (b/s)

6

7

8

9 x 10

4

(a) DMR of reported and generated to/by the Root events.

0.4 1

CoSenS − Periodic data IEEE 802.15.4 − Periodic data CoSenS/IEEE 802.15.4 − all event data 2

3

4

5 load (b/s)

6

7

8

9 x 10

4

(b) STR of all event and periodic data.

Fig. 8. Deadline meet ratio and successful transmission rate comparison between CoSenS and IEEE 802.15.4 using EDF for response events performance

outperforms IEEE 802.15.4 using EDF for both types of events. However, its performance falls after that. Again, this is because CoSenS keeps providing better STR for periodic traffic which harms the DMR of event data. The comparison between the DMR of both types of events shows the limit of EDF scheduling policy; as the load increases, the DMR of response events become lower than simple nodes events. This is because the remaining time before deadline expiration tend to zero, and more rapidly as the load increases, as packets converge to the Root. Hence, compared to the remaining time of response event packets, they will be scheduled first. However, simple node events have fewer remaining hops than response events.

6 Conclusion CoSenS is a simple but efficient scheme which improves the performance of the widely used CSMA/CA protocol. It efficiently handles periodic and Poisson traffic, has higher STR and lower end-to-end delay. Using scheduling policies greatly enhances the performance of the original protocols for event traffic. The performance analysis shows also that CoSenS overall performance is better than IEEE 802.15.4. First, it increases as the event traffic load increases. Finally, CoSenS with FP or EDF does not degrade much the performance of periodic data. Certainly, this may harm the event data when the load exceeds the saturation. However, since the router in CoSenS has a complete knowledge about the collected data, it has the possibility to aggregate them which improves throughput and perform congestion control more easily; it can discard packets more efficiently. Along with an improved delay, throughput and reliability, CoSenS is a simple and elegant MAC solution for QoS support in WSNs. Future work aim at the development of admission and congestion control mechanisms.

280

B. Nefzi and Y.-Q. Song

References 1. Caccamo, M., Zhang, L.Y., Sha, L., Buttazzo, G.: An implicit prioritized access protocol for wireless sensor networks. In: 23rd IEEE Real-Time Systems Symposium, RTSS 2002, pp. 39–48 (2002) 2. Ergen, S.C., Varaiya, P.: PEDAMACS: power efficient and delay aware medium access protocol for sensor networks. IEEE Transactions on Mobile Computing 5(7), 920–930 (2006) 3. Munir, S., Lin, S., Hoque, E., Nirjon, S.M., Stankovic, J.A., Whitehouse, K.: Addressing Burstiness for Reliable Communication and Latency Bound Generation in Wireless Sensor Networks. In: 9th ACM/IEEE Conference on Information Processing in Sensor Networks (IPSN 2010) (2010) 4. IEEE-TG15.4. Part 15.4: Wireless Medium Access Control (MAC) and Physical Layer (PHY) specifications for low-rate Wireless Personal Area Networks (LR-WPANs). IEEE standard for information technology (2006) 5. Koubˆa, A., Alves, M., Tovar, E., Cunha, A.: An implicit GTS allocation mechanism in IEEE 802.15.4 for time-sensitive wireless sensor networks: theory and practice. Real-Time Syst. 39, 1–3 (2008) 6. Koubaa, A., Alves, M., Nefzi, B., Song, Y.Q.: Improving the IEEE 802.15.4 Slotted CSMA/CA MAC for Time-Critical Events in Wireless Sensor Networks. In: Workshop on Real Time Networks RTN (July 2006) 7. Li, Y., Chen, C.S., Song, Y.-Q., Wang, Z.: Real-time QoS support in wireless sensor networks: a survey. In: 7th IFAC International Conference on Fieldbuses & Networks in Industrial & Embedded Systems - FeT 2007, Toulouse (2007) 8. ZigBee Specification Document 053474r17, http://www.zigbee.org 9. Nefzi, B., Cruz-Sanchez, H., Song, Y.-Q.: SCSP: An energy efficient network-MAC crosslayer design for wireless sensor networks. In: IEEE 34th Conference on Local Computer Networks, LCN 2009, pp. 1061–1068, October 20-23 (2009) 10. Liu, C.L., Layland, J.W.: Scheduling Algorithms for Multiprogramming in a Hard-Real-Time Environment. J. ACM 20(1), 46–61 (1973) 11. OPNET: OPNET Simulator, v 15.0, http://www.opnet.com

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes (CBR-Mobile) Protocol in Wireless Sensor Networks Samer A.B. Awwad, Chee Kyun Ng, Nor K. Noordin, Mohd. Fadlee A. Rasid, and A.R.H. Alhawari Department of Computer and Communication Systems, Faculty of Engineering, Universiti Putra Malaysia, UPM Serdang, 43400 Selangor, Malaysia [email protected], {mpnck,nknordin,fadlee}@eng.upm.edu.my

Abstract. The technological advances in wireless communication, microelectro-mechanical system (MEMS) technologies and digital electronics over the past few years have enabled the development of wireless sensor networks (WSN). In some applications, the WSN nodes are dedicated to be mobile rather than static. This requirement poses new and interesting challenges for both medium access control (MAC) and routing protocols design. In this paper, we propose mobility and traffic adapted cluster based routing for mobile nodes (CBR-Mobile) protocol in WSN to support mobility of sensor nodes in an energy-efficient manner, while maintaining maximum delivery ratio and minimum average delay. The mobility and traffic adapted scheduling based MAC design enables the cluster heads to reuse the free or unused timeslots to support the mobility of sensor nodes. Each cluster head maintains two simple database tables for mobility and traffic to achieve this adaptation. The designed CBR-Mobile protocol enables mobile sensor nodes that disconnected with their cluster heads to rejoin the network through other cluster heads within a short time. The proposed protocol can achieve around 43% improvement on packet delivery ratio while simultaneously offering lower delay and energy consumption compare to LEACH-Mobile protocol. Keywords: WSN, cluster head, mobility, traffic, LEACH, LEACH-Mobile.

1 Introduction Recent advances in wireless communication and micro-electro-mechanical system (MEMS) technologies led to emergence of low-power, small-scale, low-cost and multifunctional devices called sensors nodes. These devices are endowed with sensing, signal processing and wireless communication capabilities. Sensor nodes coupled with wireless networking to form powerful and collaborative information gathering system that is known as wireless sensor networks (WSN). J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 281–296, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

282

S.A.B. Awwad et al.

The WSN consists of hundreds or thousands of wireless sensor nodes that are densely deployed, either inside the phenomenon or very close to it, to collect information about the events in the surrounding environment. It becomes the remote "eyes" and "ears" to monitor a variety of applications ranging from environmental monitoring applications, like habitat monitoring [1] - [3], to infrastructure monitoring [4], emergency health care and medical response [5] - [7], military and battlefield surveillance applications [8] - [10]. A general conception of sensor nodes is static and remains fixed in their position once they have been deployed for a long period of time [11] - [13]. As a consequence research interests are mainly focusing on energy consumption. However, some applications like habitat monitoring, wildlife (animal) tracking, search and rescue [14], RoboMote [15], Parasitic-Mobility [16], and medical care and disaster response applications [17], [18], have enabled mobility in WSN components [11] - [14]. In this paper, we proposed mobility and traffic adapted cluster based routing for mobile nodes (CBR-Mobile) protocol in WSN. The proposed CBR-Mobile protocol collaborates with mobility and traffic adapted medium access control (MAC) layer protocol to support sensor nodes mobility and improve packet delivery ratio. The cluster heads in CBR-Mobile support sensor nodes mobility by adaptively reassigning the timeslots according to mobility and traffic environments. The proposed protocol creates forward and backward schedules to adapt time scheduling according to mobility and traffic environments. It assigns two owners for each timeslot; original owner and alternative owner, such that it can work adaptively to mobility and traffic environments. The protocol keeps the new mobile sensor nodes in the simple database tables and serves these nodes whenever the unused timeslot is available. It can achieve around 43% improvement on packet delivery ratio while simultaneously offering lower delay and energy consumption compare to LEACHMobile protocol. The rest of this paper is arranged as follows. Related work is presented in Section 2. The proposed mobility and traffic adapted scheduling in CBR-Mobile is discussed in Section 3. In Section 4, the mobility and traffic adapted techniques in CBR-Mobile are presented. Performance evaluation of proposed CBR-Mobile is discussed in Section 5. This paper is concluded in Section 6.

2 Related Work Few protocols that handle mobility at MAC layer and routing have been proposed for WSN [13], [19]. Mobility-aware MAC (MS-MAC) protocol is a contention-based protocol designed to allow mobile sensor nodes to make quicker connections, while crossing the boundary of a virtual cluster [20]. A sensor node detects its neighbour’s mobility based on a change in its received signal level from the neighbour, or based on a loss of connection with this neighbour after a timeout period. The protocol maintains the connectivity by having the neighbours of a mobile sensor node stay awake for a longer time. When there is a mobile sensor node crossing cluster borders, the mobile sensor node and surrounding nodes form an active zone, and start working with a higher duty cycle. The MS-MAC protocol handles mobility in WSN by

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes

283

keeping the border sensor nodes between the clusters awake for a longer time, leading to very high energy consumption. A scheduled-based protocol, called mobility adaptive collision free MAC (MMAC) protocol has been designed to support sensor nodes mobility by adapting the frame size, transmission slots, and random-access slots according to mobility environments [21]. The MMAC protocol uses mobility estimation algorithm to predict the position of the nodes in next frame if large number of nodes are expected to enter or leave. However, the MMAC protocol has the disadvantages of high complex scheduling algorithm and a synchronization problem as the two hop neighbours need to be synchronized simultaneously. Mobility adaptive hybrid MAC (MH-MAC) protocol has been proposed to handle sensor nodes mobility by adjusting the ratio of contention timeslots and scheduled timeslots as well as the frame time [19]. MH-MAC is a hybrid protocol that uses both contention and schedule-based channel access mechanism. The protocol uses contention timeslots for mobile sensor nodes while uses the schedule timeslots for static sensor nodes. MH-MAC uses mobility estimation algorithm to determine the mobility type of sensor nodes. When the sensor node estimates its mobility type, it broadcasts its mobility type to the neighbourhood, and hence they can estimate their mobility types and so on. The protocol responds to different levels of mobility; it dynamically adapts the ratio between static and mobile timeslots and the frame size according to the mobility types of sensor nodes. In addition, the protocol is trafficadaptive; if the sensor node which owns the current timeslot do not has data to send, the timeslot can be assigned to another sensor node. However, the MH-MAC protocol has mobility information beacon message overhead at the beginning of each frame to exchange the mobility information. Each node is assumed to have mobility information about all sensor nodes in the network which is invalid assumption in the large WSNs. All the sensor nodes have to be synchronized when the frame time is changed, and the mobile timeslots are adjusted. Similar to MAC layer protocols, a number of routing protocols have been designed to support mobility in mobile WSN [13]. Cluster based energy efficient scheme (CES) for mobile WSN [22] was designed based on low energy adaptive clustering hierarchy (LEACH) protocol [23]. The CES relies on weighing k-density, residual energy and mobility parameters for cluster head election. The protocol re-elects new cluster head when the cluster head moves to another cluster. The LEACH-Mobile protocol [24] supports sensor nodes mobility in WSN by adding membership declaration to LEACH protocol. The LEACH-Mobile outperforms LEACH in terms of packet loss in mobility centric environment. Cluster head election in LEACH-Mobile has been improved by LEACH-Mobile enhanced (LEACH-ME) [13] whereby sensor node with minimum mobility factor is elected as cluster head. However, the LEACH-Mobile protocol has the following drawbacks: • •

Packet loss in LEACH-Mobile is relatively high and increases rapidly when the number or speed of mobile sensor nodes is increased in the network. Mobile sensor nodes in LEACH-Mobile have to wait for two consecutive failures in receiving 'data request' message from the cluster head before they attempt to join the new cluster.

284

• •

S.A.B. Awwad et al.

The LEACH-Mobile protocol is neither traffic-adaptive nor mobility-adaptive protocol. The assigned timeslots to the sensor nodes that do not have data to send or move out of the cluster are remained wastes. The wasted energy in idle listening and overhearing of LEACH-Mobile protocol is high. After mobile sensor nodes lose the connection with their cluster heads, they will keep their receivers ON to receive the cluster heads’ announcement messages. Hence, they will receive any message sent in the vicinity even the messages are not destined to them.

3 Mobility and Traffic Adapted Scheduling in CBR-Mobile In mobility environment WSN, scheduling has to be done while taking into account supporting mobility and maintaining the low duty cycle of sensor nodes through keeping sensor nodes in sleep mode as long as possible. In CBR-Mobile, scheduling should address the following characteristics: • •

•

Mobility-adaptive: the scheduling algorithm should be adaptive to the mobility of sensor nodes. This enables avoidance of waste timeslots or bandwidth utilization in the old cluster and fast inclusion of mobile sensor nodes in the new cluster. Traffic-adaptive: the scheduling algorithm should be adaptive to the traffic in the sensor nodes. This enables sensor nodes to release their timeslots when there are no data to send. Hence, it increases the bandwidth utilization and decreases the latency in the network. The scheduling should maintain the low duty cycle of sensor nodes. This keeps the sleep schedule for the sensor nodes such that it can decrease the energy consumption of the small battery.

In order to enable time division multiple access (TDMA) scheduling to support mobility and traffic adaptation, the cluster head should update all sensor nodes members by all the changes in the schedule which require the sensor nodes to be in active mode most of the time or wake up periodically to receive the TDMA schedule updates. This increases sensor nodes duty cycle which depletes the energy of these sensor nodes. In our scheduling technique, the assigned timeslots to the mobile sensor nodes that moved out of the cluster or have not data to send will be reassigned to the mobile sensor nodes that join the cluster or to the sensor nodes that have data to send. In CBR-Mobile, each cluster head maintains two simple database tables to enable mobility and traffic adapted scheduling while maintaining low duty cycle to conserve the energy in the small battery. These simple databases will be discussed in the following subsections. 3.1 New Membership Requesters (NEW_MEM_REQs) Database New membership requesters (NEW_MEM_REQs) database is a simple database introduced to keep information about all mobile sensor nodes that joined the cluster.

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes

285

When mobile sensor node loses the connection with its cluster, it will broadcast Join Message (JOIN_MSG) to all nearby cluster heads. The cluster heads that receive this JOIN_MSG will add the mobile sensor node to its database. The cluster head keeps and queues the mobile sensor nodes in NEW_MEM_REQs database as first-in firstout (FIFO) scheduling order. Figure 1 illustrates how the new mobile sensor node is added to the NEW_MEM_REQs database.

Fig. 1. Registration of mobile sensor node in the new cluster

3.2 Alternative Schedule Database (ALT_SCH) Whenever the cluster head creates TDMA time schedule, it creates alternative schedule (ALT_SCH) database as well. This database determines which sensor nodes can replace the sensor nodes in the original schedule. The ALT_SCH is created as the reverse schedule to original TDMA. This will overcome the cumulative free timeslots problem which is necessary for waking up the sensor nodes to announce the schedule updates for the cluster members. The cluster head assigns two owners sensor nodes for each timeslot; the original owner from the original schedule and the alternative owner from the ALT_SCH. Figure 2 shows the frame timeslots and the scheduled original and alternative owners.

286

S.A.B. Awwad et al.

Fig. 2. Frame structure of the original and ALT_SCH time schedules

4 Mobility and Traffic Adapted Techniques in CBR-Mobile The mobility and traffic adapted specifications, besides supporting sensor nodes mobility, have the added values of high bandwidth utilization by eliminating wasted timeslots. These specifications enable the cluster head to assign unused timeslots to other sensor nodes. When the cluster head assures that the original owner of the current timeslot does not want to use it (either move out of the cluster or has not data to send), the cluster head will reassign the timeslot to other sensor or mobile sensor nodes that enter the cluster border. CBR-Mobile acts as query based protocol. At the beginning of the timeslot, both original and alternative owners are wakeup. The cluster head sends a data request message (DATA_REQ_MSG) to the original owner sensor node that owns the current timeslot. If the original owner has data, it switches on its radio transmitter and sends its data back to the cluster head as shown in Figure 3. At the end of the transmission, the node turns off its radio and goes back to sleep mode, thus minimizing energy dissipation. The alternative owner will wait for timeout period called timeout second data request (TOUT2_DATA_REQ), then go back to sleep mode since the cluster head will never offer timeslot for replacement. This conserves energy consumption in the alternative owner. The TOUT2_DATA_REQ is calculated as ⎛ ⎡ Transmit t ime(DATA_R EQ_MSG) ⎜ + TOUT2_DATA _REQ > MAX ⎜ ⎢⎢ ⎜ ⎢ Transmit t ime(HASN' T DATA_MSG) ⎝⎣

⎤ ⎥, TOUT_DATA ⎥ ⎥⎦

⎞ ⎟ ⎟ ⎟ ⎠

(1)

If the original owner did not receive the DATA_REQ_MSG within timeout called timeout data request (TOUT_DATA_REQ), it assumes that it has moved out of the cluster. Then, it broadcasts JOIN_MSG to nearby cluster heads to join the cluster. The TOUT_DATA_REQ is given by

0 < TOUT_DATA_REQ ≤ Transmit time(DATA_REQ_MSG)

(2)

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes

287

Fig. 3. Data transfer in the original and alternative owners’ time schedule

CBR-Mobile works adaptively with the traffic. If the original owner sensor node has no data to send, the original owner will respond by sending hasn’t data message (HASN’T_DATA_MSG) and go back to sleep. Then, the cluster head will query NEW_MEM_REQs database to retrieve the mobile sensor node that has the highest priority. If the NEW_MEM_REQs database has some mobile sensor nodes, the cluster head assigns the timeslot to the retrieved mobile sensor node and removes it from NEW_MEM_REQs database. The mobile sensor node stays awake during the timeslot to send the data to the cluster head. The alternative owner waits for TOUT2_DATA_REQ and goes back to sleep when it assures that the timeslot is assigned to mobile sensor node. If the NEW_MEM_REQs database is empty, the cluster head will assign the timeslot to the alternative owner in the ALT_SCH. However if the alternative owner has data, it stays awake and sends the data to the cluster head as shown in Figure 4. Besides that, the alternative owner will release its timeslot during this frame. As shown in Figure 4(a), the timeslot is reassigned to the mobile sensor node in NEW_MEM_REQs. In Figure 4(b), the timeslot is reassigned to the alternative owner sensor in ALT_SCH database. The CBR-Mobile works adaptively with the mobility of sensor nodes. If the original owner sensor node did not respond during the timeout period, called timeout data (TOUT_DATA), the cluster head assumes that the sensor node has been moved out of the cluster. The TOUT_DATA is derived as

TOUT_DATA > Transmit t ime(DATA_R EQ_MSG) Hence, the cluster head can reuse this timeslot to another sensor node.

(3)

288

S.A.B. Awwad et al.

(a)

(b) Fig. 4. Traffic-adaptive. (a) reassigned the timeslot to the mobile sensor node, and (b) reassigned timeslot to the alternative owner.

Then, the cluster head will query NEW_MEM_REQs database to retrieve the mobile sensor node that has the highest priority. If the NEW_MEM_REQs database has some mobile sensor nodes, the cluster head assigns the timeslot to the retrieved mobile sensor node. The alternative owner waits for TOUT2_DATA_REQ and goes back to sleep when it assures that the timeslot is assigned to mobile sensor node. If the NEW_MEM_REQs database is empty, the cluster head will assign the timeslot to the alternative owner in the ALT_SCH as shown in Figure 5. In Figure 5(a), the timeslot is reassigned to the mobile sensor node retrieved from the NEW_MEM_REQs database, while in Figure 5(b), the timeslot is reassigned to the alternative owner sensor node retrieved from the ALT_SCH database.

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes

289

(a)

(b) Fig. 5. Mobility-adaptive (a) reassigned the timeslot to the mobile sensor node, and (b) reassigned timeslot to the alternative owner.

At the end of the frame, the cluster head checks NEW_MEM_REQs database. If it contains some mobile sensor nodes, they will be allowed to join the cluster. The cluster head sends announcement message (ANN_MSG) to mobile sensor nodes. Upon receiving this message, the mobile sensor nodes will respond by sending JOIN_MSG. At this time, the frame time is finished and the cluster head prepares and goes for next round. The cluster head will remove non-responding sensor nodes, add the new mobile sensor nodes to the schedule and broadcast the new schedule to all members. This mobility and traffic-adaptive technique enable the mobile sensor nodes to join the cluster within a short time and hence achieve high packet delivery ratio. It maintains low duty cycle and decrease overhearing so that it conserves the sensor

290

S.A.B. Awwad et al.

nodes’ energy. It decreases the latency of the protocol since the sensor nodes that have data will not suffer from the latency caused by unused timeslots. The integrated pattern of mobility and traffic-adaptive scheduling in CBR-Mobile gives the protocol special capability to support mobility and high traffic of sensor nodes. As a result, mobility and traffic adapted scheduling support one the other.

5 Performance Evaluation of Proposed CBR-Mobile In order to evaluate the performance of the CBR-Mobile, a simulator is developed using MATLAB as compared with the LEACH-Mobile. The measured performance metrics in demonstrating the improvement and strength features of our design are packet delivery ratio, average energy consumption and average packet delay. One hundred sensor nodes are deployed randomly in the 50 x 50 m2 field. The network is divided into five clusters with one sensor node acting as cluster head. All cluster heads are assumed to be static. These sensor nodes are moving in the network according to the random waypoint model [25]. Each mobile node picks its random direction from (0, 2π] and moves from its current position in that direction toward a new position for a distance d with a speed v between [Vmin, Vmax], where d is exponentially distributed. When mobile sensor node reaches the destination, it pauses for a period of time. When the node hits the boundary, it will be reflected at the boundary. Each sensor node communicates with its cluster head directly. At the beginning of the simulation, 5% of the sensor nodes are randomly elected to be cluster head. These cluster heads broadcast advertisement messages (ADV_MSGs) to the rest of the sensor nodes in the network as in LEACH and LEACH-Mobile protocols. All cluster heads use the same transmit energy when they broadcast ADV_MSGs. Each sensor nodes may receive one or more ADV_MSGs from nearby cluster heads. After noncluster head sensor nodes have received ADV_MSGs from one or more cluster heads, the sensor nodes will compare the received signal strength of these messages and decide to which cluster it belongs. By assuming symmetric propagation channels, the sensor node selects cluster head to which the minimum amount of transmitted energy is needed for communication. In the case when two cluster heads need the same minimum amount of transmitted energy, a random cluster head is chosen. After choosing the cluster head, sensor node sends registration message (REG_MSG) to inform the cluster head that it wants to be one of the cluster’s members. After the cluster head receives REG_MSGs from sensor nodes that willing to join the cluster, the cluster head creates both original schedule and ALT_SCH to achieve traffic adaptive scheduling. The main simulation parameters are shown in Table 1. Figure 6 depicts that the CBR-Mobile achieves a higher and more stable packet delivery ratios when the percentage of mobile nodes is increased. The mobility and traffic adapted techniques enable the mobile sensor nodes that lost connection with their cluster heads to join the new cluster within a short time. Hence, these mobile sensor nodes will not suffer from high packet loss. The CBR-Mobile achieves 43% higher packet delivery ratio compared to LEACH-Mobile when 90% of sensor nodes are mobile.

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes Table 1. Main Simulation Parameters Parameters and Models Network (field) Size (L x W) Number of Sensor Nodes (N) Location of the Sink Node Sensor Nodes Deployment Sensor ID Maximum Transmission Range Percentage of Cluster Head Percentage of Mobile Sensor Nodes Data Size Mobility Model Speed Radio Model NEW_MEM_REQs Database ALT_SCH Battery Traffic Model Queuing Model Idle Power Rx Power Tx Power Sleep Power

Value 50 x 50 m2 100

(25,25) Random Deployment 1-100 19 m 5% 0 – 90 % 2000 bits Random Way Point Model 1-10 m/s Two-Ray Ground model Initially is empty Reverse order of original schedule Initial capacity is constant Constant Bit Rate and Poisson FIFO with Drop Tail Queue Mechanism 2.4 mW 67.2 mW 76.8 mW 0.0048 mW

Fig. 6. Delivery ratio versus percentage of mobile sensor nodes

291

292

S.A.B. Awwad et al.

Figure 7 depicts that increasing the number of mobile sensor nodes in the network will increase the average energy consumption for both protocols. This is because when the number of mobile sensor nodes is increased, the mobile sensor nodes that leave their clusters will be increased as well. These mobile sensor nodes remain temporarily disconnected from the network for a period of time. During this period of disconnection, mobile sensor nodes keep their radio on to receive the registration messages. Hence, they will overhear any message sent in the vicinity and waste the energy. When the disconnection period is long, the mobile sensor node will overhear more messages and waste more energy. In addition, the figure indicates that when the percentage of mobile sensor nodes is increased, the CBR-Mobile protocol decreases the average energy consumption from 4 - 16% in high mobility environment compared with LEACH-Mobile. In CBR-Mobile, the disconnection periods for mobile sensor nodes are short. Hence, the mobile sensor nodes avoid overhearing many messages from the vicinity and this keeps their energy. On the contrary, the disconnection periods for mobile sensor nodes in LEACH-Mobile are long such that the mobile sensor nodes will overhear many data and control messages from the neighbourhoods. Hence, more energy is wasted.

Fig. 7. Average energy consumption versus percentage of mobile sensor nodes

Figure 8 shows the packet delivery ratio as related to traffic generation rate for static and mobile networks. In mobility environment, the figure indicates that CBRMobile achieves high and stable packet delivery ratio when packet generation rate in sensor nodes is increased. It offers 9 - 25% higher delivery ratio than LEACHMobile. Mobility and traffic adapted techniques enable the protocol to achieve this ratio since the mobile sensor nodes can rejoin the new cluster within a short time.

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes

293

Fig. 8. Packet delivery ratio versus traffic generation rate for static and mobile networks

In static environment, the LEACH-Mobile maintains the steady state situation until 0.27 packets per second while CBR-Mobile maintains it until 0.25 packets per second. In comparison, the figure shows that, in mobility environment, the LEACH-Mobile maintains the steady state situation until 0.19 packets per second, while CBR-Mobile maintains it until 0.25 packets per second. LEACH-Mobile shortens the life of steady state situation by 30% in mobility as compared with static environment. Packet delivery ratio for LEACH-Mobile in mobility environment drops by 30% relative to the static environment, while the packet delivery ratio for CBR-Mobile in mobility environment drops by 13% relative to the static environment. In the WSN, the generated traffic load heavily depends on the application. This generated traffic can be classified into two categories; event-driven reporting and periodic data collection. In event-driven reporting, sensor nodes are responsible for detecting and reporting events after the occurrence. This type of traffic is modelled by Poisson traffic [26]. While in periodic data collection, sensor nodes report their samples in specific time intervals. This type of traffic is modelled by constant bit rate traffic [27]. To investigate the effect of different generated traffic on the average delay of CBRMobile, the average delay between CBR-Mobile and LEACH-Mobile are compared under constant bit rate and Poisson traffic. Figure 9 shows that CBR-Mobile achieves less average delay for both Poisson and constant bit rate traffics. It has the advantage of 30 - 39% less average delay for constant bit rate traffic and 13 - 47% for the Poisson traffic as shown in Figures 9(a) and 9(b), respectively. This figure shows that the presented CBR-Mobile protocol is adaptive to any type of traffics and can achieve lower average delay compared to the LEACH-Mobile protocol.

294

S.A.B. Awwad et al.

(a)

(b) Fig. 9. Average delay versus traffic generation rate, (a) constant bit rate and (b) Poisson

Mobility and Traffic Adapted Cluster Based Routing for Mobile Nodes

295

6 Conclusions Supporting mobility in sensor nodes becomes increasingly useful in various applications. Mobility and traffic adapted specifications result in substantial support for mobility of sensor nodes. They enable disconnected mobile sensor nodes to rejoin the network within a short time and avoid the accumulative data loss. The timeslots originally assigned to the outgoing mobile sensor nodes and those assigned to sensor nodes that do not have traffic to send can be reassigned to the new mobile sensor nodes when they enter into any of these clusters. The simulation results demonstrate that our proposed CBR-Mobile protocol has significantly increased the packet delivery ratio compared to the LEACH-Mobile protocol. It also decreases the energy consumption and the average delay at the same time.

References 1. Mainwaring, A., Polastre, J., Szewczyk, R., Culler, D., Anderson, J.: Wireless Sensor Networks for Habitat Monitoring. In: Proceedings of the First ACM Wireless Sensor Networks and Applications Workshop, Atlanta, Georgia, USA, pp. 88–97 (2002) 2. Cerpa, A., Elson, J., Estrin, D., Girod, L., Hamilton, M., Zhao, J.: Habitat Monitoring: Application Driver for Wireless Communications Technology. In: Proceeding of the First ACM SIGCOMM Workshop on Data Communications in Latin America and the Caribbean, San Jose, Costa Rica, pp. 20–41 (2001) 3. Szewczyk, R., Mainwaring, A., Polastre, J., Anderson, J., Culler, D.: An Analysis of a Large Scale Habitat Monitoring Application. In: Proceeding of the Second ACM Conference on Embedded Networked Sensor Systems (SenSys), Baltimore, MD, USA, pp. 214–226 (2004) 4. Ye, W., Heidemann, J., Estrin, D.: An Energy-Efficient MAC Protocol for Wireless Sensor Networks. In: Proceeding of the Twenty-First Annual Joint Conferences of the IEEE Computer and Communications Societies (INFOCOM 2002), vol. 3, pp. 1567–1576 (2002) 5. Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: A Survey on Sensor Networks. IEEE Communications Magazine 40(8), 102–114 (2002) 6. Raviraj, P., Sharif, H., Hempel, M., Ci, S.: An Energy Efficient MAC Approach for Mobile Wireless Sensor Networks. In: Proceeding of IEEE International Conference on Computer Systems and Applications, pp. 565–570 (2006) 7. Senner, T., Karnapke, R., Lagemann, A., Nolte, J.: A Combined Routing Layer for Wireless Sensor Networks and Mobile Ad-Hoc Networks. In: Proceeding of the Second International Conference on Sensor Technologies and Applications (SENSORCOMM 2008), pp. 147–153 (2008) 8. Arora, A., Dutta, P., Bapat, S., Kulathumani, V., Zhang, H., Naik, V., Mittal, V., Cao, H., Demirbas, M., Gouda, M., Choi, Y., Herman, T., Kulkarni, S., Arumugam, U., Nesterenko, M., Vora, A., Miyashita, M.: A Line in the Sand: A Wireless Sensor Network for Target Detection, Classification, and Tracking. Elsevier Computer Networks, Special Issue on Military Communications Systems and Technologies 46(5), 605–634 (2004) 9. Arora, A., Ramnath, R., Ertin, E., Sinha, P., Bapat, S., Naik, V., Kulathumani, V., Zhang, H., Cao, H., Sridharan, M., Kumar, S., Seddon, N., Anderson, C., Herman, T., Trivedi, N., Nesterenko, M., Shah, R., Kulkami, S., Aramugam, M., Wang, L., Gouda, M., Choi, Y.R., Culler, D., Dutta, P., Sharp, C., Tolle, G., Grimmer, M., Ferriera, B., Parker, K.: Exscal: Elements of an extreme scale wireless sensor network. In: Proceedings of the 11th IEEE International Conference on Embedded and Real-Time Computing Systems and Applications, pp. 102–108 (2005)

296

S.A.B. Awwad et al.

10. Ren, B., Ma, J., Chen, C.: The Hybrid Mobile Wireless Sensor Networks for Data Gathering. In: Proceedings of Sensor and Wireless Resource Management Symposium, Vancouver, British Columbia, Canada, pp. 1085–1090 (2006) 11. Ali, M., Saif, U., Dunkels, A., Voigt, T., Romer, K., Langendoen, K., Polastre, J., Afzal Uzmi, Z.: Medium Access Control Issues in Sensor Networks. ACM SIGCOMM Computer Communication Review 36(2), 33–36 (2006) 12. Akkaya, K., Younis, M.: A Survey on Routing Protocols for Wireless Sensor Networks. Elsevier Ad Hoc Networks 3(3), 325–349 (2005) 13. Santhosh Kumar, G., Vinu Paul, M.V., Poulose Jacob, K.: Mobility Metric based LEACHMobile Protocol. In: Proceeding of the 16th International Conference on Advanced Computing and Communications (ADCOM 2008), pp. 248–253 (2008) 14. Liliana, M., Arboleda, C., Nasser, N.: Cluster-Based Routing Protocol for Mobile Sensor Networks. In: Proceedings of the 3rd International Conference on Quality of Service in Heterogeneous Wired/Wireless Networks. ACM International Conference Proceeding Series, vol. 191 (2006) 15. Dantu, K., Rahimi, M., Shah, H., Babel, S., Dhariwal, A., Sukhatme, G.S.: Robomote: Enabling Mobility in Sensor Networks. In: Proceeding of the Fourth International Conference on Information Processing in Sensor Networks, pp. 404–409 (2005) 16. Laibowitz, M., Paradiso, J.A.: Parasitic Mobility for Pervasive Sensor Networks. In: Gellersen, H.W., et al. (eds.) PERVASIVE 2005. LNCS, vol. 3468, pp. 255–278. Springer, Heidelberg (2005) 17. Shnayder, V., Rong Chen, B., Lorincz, K., Fulford-Jones, T.R.F., Welsh, M.: Sensor Networks for Medical Care. In: Technical Report TR-08-05, Harvard University (2005) 18. Lorincz, K., Malan, D.J., Fulford-Jones, T.R.F., Nawoj, A., Clavel, A., Shnayder, V., Mainland, G., Welsh, M., Moulton, S.: Sensor Networks for Emergency Response: Challenges and Opportunities. IEEE Pervasive Computing, Special Issue on Pervasive Computing for First Response 3(4), 16–23 (2004) 19. Raja, A., Su, X.: A Mobility Adaptive Hybrid Protocol for Wireless Sensor Networks. In: Proceeding of the 5th IEEE Consumer Communications and Networking Conference, pp. 692–696 (2008) 20. Pham, H., Jha, S.: An Adaptive Mobility-Aware MAC Protocol for Sensor Networks (MSMAC). In: Proceeding of IEEE International Conference on Mobile Ad-hoc and Sensor Systems, pp. 558–560 (2004) 21. Ali, M., Afzal Uzmi, Z.: Medium Access Control with Mobility-Adaptive Mechanisms for Wireless Sensor Networks. International Journal of Sensor Networks 1(3/4), 134–142 (2006) 22. Lehsaini, M., Guyennet, H., Feham, M.: CES: Cluster-Based Energy-Efficient Scheme for Mobile Wireless Sensor Networks. In: IFIP International Federation for Information Processing, Wireless Sensor and Actor Networks II, vol. 264, pp. 13–24. Springer, Boston (2008) 23. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: Energy-Efficient Communication Protocol for Wireless Microsensor Networks. In: Proceedings of the 33rd Annual Hawaii International Conference on System Sciences (HICSS 2000), vol. 2, pp. 1–10 (2000) 24. Kim, D.S., Chung, Y.J.: Self-Organization Routing Protocol Supporting Mobile Nodes for Wireless Sensor Network. In: Proceedings of the First International Multi-Symposiums on Computer and Computational Sciences (IMSCCS 2006), vol. 2, pp. 622–626 (2006) 25. Bansal, N., Liu, Z.: Capacity, Delay, and Mobility in Wireless Ad Hoc Networks. In: Proceedings of the Twenty-Second Annual Joint Conference of the IEEE Computer and Communications (INFOCOM 2003), vol. 2, pp. 1553–1563 (2003) 26. Demirkol, I., Alagoz, F., Deliç, H., Ersoy, C.: Wireless Sensor Networks for Intrusion Detection: Packet Traffic Modeling. IEEE Communication Letters 10(1), 22–24 (2006) 27. Krishnamurthy, V., Sazonov, E.: Bandwidth Optimization in 802.15.4 Networks Through Evolutionary Slot Assignment. International Journal of Communications, Network and Systems Sciences (IJCNS) 6, 518–527 (2009)

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy on Geo-Routing in Urban Vehicular Environments Aisling O’ Driscoll and Dirk Pesch Nimbus Centre for Embedded Systems Research, Cork Institute of Technology Rossa Avenue, Cork, Ireland {aisling.odriscoll,dirk.pesch}@cit.ie

Abstract. Vehicular routing has been an extremely active research field in recent years with geo-routing protocols typically favoured over conventional topology based routing protocols due to their advantages in terms of scalability and lower overhead. Before a geo-routing protocol can transmit a packet, it must be aware of the position of the target node and is reliant upon a location service to supply this information. Therefore the correct and efficient operation of the routing protocol is entirely dependant on the accuracy of this information. In this paper, a simulation based analysis is conducted to determine the tolerance of a geo-routing protocol to position inaccuracy as reported by a location service. As the inherent mobility of a vehicular network may also have a negative impact on protocol performance, we also evaluate characteristics such as vehicular density, transmission range and query range. Keywords: Vehicular Ad-hoc Networks (VANETs), Location Service, GeoRouting Protocol.

1 Introduction Over the past number of years, all major automobile manufacturers, often supported by government research initiatives, have invested in vehicular research, with vehicular communication networks poised to have a large societal impact. While the primary impetus in this space has been towards traffic information and safety systems, many platforms have been proposed promising a plethora of next generation vehicular applications such as distributed gaming, the evolution of vehicular communities for file sharing and social networking, infotainment and P2P content distribution applications amongst others. Given the infancy of this field of study, it still remains unclear which exact applications will prevail, but it can be envisaged that future vehicular applications will become highly customizable and will be based on a combination of broadcast, multicast and unicast communications. A commonality of Vehicle-to-Vehicle (V2V) unicast applications is the requirement to discover the location of a destination vehicle towards which traffic can be routed. This is a necessary requirement for geo-routing protocols, which are the typical method employed to enable multi-hop data communication in vehicular networks. In recent years many VANET-specific geo-routing derivatives have been proposed to overcome J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 297–313, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

298

A. O’ Driscoll and D. Pesch

the drawbacks of traditional geo-based routing protocols. All of these need to discover the destination location prior to forwarding. Intermediate nodes subsequently extract the destination position from the packet to make a decision regarding the next hop according to the geo-routing algorithm employed. To discover the destination location, the source vehicle firstly references a location service which stores a mapping for each vehicle ID to its current location – it is the responsibility of each individual vehicle within the network to update the location service index at regular intervals to ensure its accuracy. Since the correct and efficient operation of any geo-routing protocol is entirely reliant on the accuracy of the destination information returned by the location service, this is a vital field of study if vehicular unicast communications are to succeed. Interestingly, while geo-routing has been a very topical area of research in vehicular networks, the location service algorithms utilized are typically treated as an orthogonal issue and are largely studied as mutually exclusive research areas. While many of the recently proposed VANET specific geo routing protocols exhibit excellent performance, they either assume the existence of a perfect location service or neglect to specify how the destination’s position would be determined. Others simply employ location algorithms developed for generic MANETs that have been shown in literature to exhibit poor robustness and location accuracy over vehicular environments. Given the significant impact that a location service has on geo-routing protocol performance and the difficulty in maintaining either an available or accurate service in highly dynamic networks, this is not a valid assumption. Specifically a location service algorithm must specify methods for efficient location update and maintenance, techniques for requesting and retrieving location information and ensuring the availability of the location service. The overall design objective is to limit the location inaccuracy within the algorithm while maintaining availability of the location service to each vehicle. Thus the challenge of a scalable distributed location service algorithm is a difficult one, particularly given a high level of dynamism. The challenges can be categorised as follows: o Availability of the location service (as unavailability will render the geo-routing protocol defunct with the exception of localized communications) and the overhead associated with maintaining availability. o Accuracy of location Information and the impact on the packet delivery rate of the respective geo-routing protocol when inaccurate location information is provided. Inability of a location server to receive location updates will also lead to stale location information when the service is queried. o Scalability of the location service update mechanism and efficient resolution of location queries between vehicles at geographically disparate locations with minimum overhead. Given the inherent importance of a location service, the focus of this paper is to evaluate the impact of location inaccuracy on geo-routing protocol performance and the effect of topological conditions on the success rate. Location inaccuracy can occur due to stale location entries as a result of delayed or lost position update packets. Factors leading to reported position inaccuracy include packet collisions, lack of route availability as a result of a partitioned/sparse ad-hoc network or temporary unavailability of the location server to receive a position update (dependant on the location service algorithm employed). Furthermore location inaccuracy can occur as a

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

299

result of mobility induced errors, as the optimal update interval largely depends on the vehicular speed, road topology and density of vehicles. To clearly distinguish between location service inaccuracy and the negative impact of vehicular characteristics such as mobility, transmission range and vehicular density, a detailed study of geo-routing performance over varied vehicular topological conditions is also performed. The remainder of the paper is structured as follows: Section 2 discusses related work in this field, with Section 3 outlining the negative impact that stale location information has on geo-routing as well as the occurrence of mobility induced errors. Section 4 discusses the simulation environment and quantitive performance analysis performed. Finally Section 5 concludes the paper and outlines future work.

2 Related Work A location service has a large impact on the geo-routing protocol as its performance is entirely dependant on the locations service accuracy and scalability. As acknowledged recently in [1, 2], a major hurdle for vehicular ad-hoc networks is the complexity associated with maintaining location services, often suppressing the potential gains. Given the non trivial nature of maintaining an accurate and available vehicular location service, it is surprising that much of the research published in recent years devoted to vehicular-specific geo-routing protocols does not focus on the location management method used. Geo-routing protocols such as VADD[3], GOSR[4], MDDV[5], ASTAR[6], ACAR[7] and GPCR[8] assume a perfect or idealised location service whilst others such as GPSR [9], RBVT-P[10], GSR[11], GyTAR[12] and SARC[13] utilise MANET-developed location services that have been shown to suffer from robustness and accuracy issues when applied to vehicular networks. However while the drawbacks are known, this is very rarely included as part of the geo-routing protocol performance analysis even though it represents a necessary component for the correct functioning of the routing protocol. In some cases if their performance were to be considered in conjunction with the use of MANET-developed location service, they can often be out performed by traditional routing protocols, negating the purpose of their original inception. To highlight the negative performance effects that a location service can have on a geo routing protocol, it is shown in [14] that the routing overhead for GPSR is significantly higher than topology based routing protocols like AODV and DSR, predominantly because of the overhead incurred as a result of the location service. Furthermore, GSR uses a location service that is dependant on global flooding thus scalability is not guaranteed and GSR only achieves similar results to AODV. Proactive location services such as the Hierarchical Location Service (HLS) [15] and the Geographical Location Service (GLS) [16] do not consider street layout, contribute significantly to the routing overhead incurred in the network and also the very frequent message exchanges may interfere with data transmissions leading to packet drops. RLS [17], used by GSR, relies on initial global flooding so does not scale well. [18] presents PLS, where instead of using the last known location of a node for the location service reply, the location service tries to extract the mobility pattern of the node and predicts its actual location. The Dead Reckoning-Based Location Service [19] adjusts the periodic dissemination of geographic information

300

A. O’ Driscoll and D. Pesch

based on a first order deterministic mobility prediction model. Mobility Prediction based GLS [20] improves GLS by adapting the periodic location maintenance with two prediction models deterministic first order and history-based first order Markovien. Thus, many location services have been developed based on generic MANET characteristics but may not perform efficiently in vehicular environments. Subsequently, vehicular-specific location services have been proposed. Gerla et al propose a solution, V-Grid [21], a dual location service with functionality to exploit the intermittent existence of Road-Side Unit (RSU) infrastructure and another location service to operate in strictly ad-hoc vehicular networks. The performance of the purely ad-hoc location service can however quickly deteriorate and requires the use of a dedicated truck to record vehicle registrations in given areas which is not always a feasible approach. Chang et al described the Intersection Location Service (ILS) in [22] where vehicles near intersections represent location servers and a Chord ring is used as a fault recovery mechanism when the corresponding location services go out of service. This approach is again limited by the excessive overhead associated with maintaining mobile location servers and overlay “stretch” between the Chord overlay and the MANET underlay hops. Inconsistency of the Chord overlay, possibly leading to stale or unavailable location servers, was also not investigated. PHLS [23] presents a similar approach to PLS but is hierarchical exhibiting the limitations of traditional MANET client-server location services and does not account for changes in vehicle direction. VLS [24] does not describe the performance overhead implications associated with location server maintenance or provide a fault-tolerance mechanism to ensure location server availability in temporary void areas. MALM [25] and PLM [26], present a localised approach to maintaining location information by exchanging historical location information with its neighbours. Whilst this passive approach yields performance results exhibiting low overhead, it depends on vehicles opportunistically encountering one another or encountering a vehicle that can inform the source about the destination. This can introduce considerable, possibly indefinite, delays as well as rendering this service unusable in terms of success rates i.e. successful packet delivery can’t be quantified or guaranteed as it’s based on opportunistic vehicle encounters. The authors overcome this by proposing an on demand query mechanism but only query one hop neighbours limiting the scalability of the system. Therefore the challenge of providing an accurate, available and scalable location service for vehicular networks is a difficult one, as acknowledged in many publications including most recently in [27]. Before a solution can be devised, the impact of the location inaccuracy on a geo-routing protocol should firstly be examined to gain an understanding into what inaccuracy can be endured and to investigate the effects of inaccuracy on the routing protocol. Some attempts have already been made in literature. Helmy et al have published a number of papers in this space [28, 29, 30, 31] but are mainly concerned with the specific failures of the face changing algorithm in GPSR. More recently a study over sensor networks was conducted [32] but only considered a small number of nodes with no mobility and very short transmission ranges. These studies do not consider the inherent characteristics of vehicular networks and as such do not consider the road layout, possible distance or the unique impact of the speeds associated with vehicles.

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

301

3 Mobility Induced Errors and Inaccurate Location Service Positions It is important to distinguish between mobility induced errors and errors caused as a result of location service inaccuracy though they may impact on each other. The possible effects are now described. Geo-routing protocols typically rely on greedy heuristics in order to route a packet. The next hop is chosen based on the neighbour node is geographically closest to the intended destination. Due to temporary partitions in a network and the non uniform distribution of nodes, packets can reach a point that is still not within radio range of the destination and where greedy routing is no longer possible. This is referred to as the “local maxima” or a “void” in the network. A recovery mechanism is used to circumvent this problem, the specifics of which differ depending on the routing protocol employed. In the case of GPSR, a method known as perimeter mode routing is employed to circumvent void areas based on planarized graphs such as the Relative Neighbourhood Graph (RNG) or the Gabriel Graph (GG). Since no greedy neighbour node is available, a next hop node is chosen based on the Right Hand Rule. However, the edge between sending node and receiving next hop neighbor should not cross the edge between origin node and destination. However node mobility can easily induce routing loops for face routing if the graph is not planar. This can easily cause a routing loop with a packet eventually being dropped when the TTL of 64 has decremented to 0. An example of how this occurs is described in [11]. Packets can also be dropped as a result of the TTL being reached when the intended destination vehicle has left the network before a packet can be delivered or because of exceptionally long perimeter routes. Furthermore vehicular mobility in opposite directions can contribute to excessively long routes as the packet may not always be routed progressively towards the destination as interpreted by the routing protocol. Furthermore a packet may be discarded as a result of a transmission error if the WLAN retry threshold is exceeded. This may be the result of collisions or because of the intended next hop node moving beyond the radio range of the transmitting node. Whilst an increase in density can improve the packet delivery success rate as it become increasingly less likely that a source or interim node will become isolated, the increased neighbour density for next hop routing may also result in an increase in the number of neighbours on the radio range perimeter that may be chosen as a result of the greedy algorithm. Furthermore vehicular mobility i.e. source and destination vehicles moving in opposite direction may exacerbate this issue. An example of how this can occur is seen in Figure 1a. Vehicle A chooses vehicle B as the next hop vehicle according to the greedy heuristics employed in the geo-routing protocol algorithm and routes to the position stored in its neighbour table, B(old). However B has actually moved beyond the radio range of vehicle A and is now at position B(new) but the entry has not been phased out of A’s neighbour table. In this case, A will attempt to transmit and the packet will be eventually be dropped as the WLAN retry threshold will be exceeded. Some routing protocols have suggested link breakage modifications to the 802.11 MAC in order to choose an alternative neighbour, if such a neighbour exists. This scenario could also occur if a destination vehicle has reached its destination thus leaving the network but has not yet been phased out of the source vehicles routing table.

302

A. O’ Driscoll and D. Pesch

Vehicular mobility can cause frequent temporary partitions in the ad-hoc network causing source or interim next-hop vehicles to become disconnected from their neighbours. Packets can be discarded if a source vehicle has become isolated, temporarily having an empty neighbour table, thus this node can cannot query a destination vehicle’s location. This is illustrated in Figure 1b.

Fig. 1. WLAN Retry Threshold Exceeded and Source Node Isolation

Fig. 2. Mobility Induced Error – Interim Node Isolation

Furthermore an interim vehicle can become isolated such as in Figure 2. In Figure 2(a) vehicle A chooses vehicle B, a vehicle on the periphery of its radio range as the next hop vehicle. Vehicle B successfully receives the packet but subsequently moves outside the radio range and becomes temporarily isolated from the ad-hoc network. This results in a packet drop. While the packet dropped may represent an unsuccessfully delivered data packet, it may also represent a dropped location update packet leading to inaccuracy if the location service were to be subsequently queried. It is suggested in [27] that vanet routing protocols should exhibit four properties, one of which is the implementation of a store and carry paradigm for delay tolerant applications to cope with temporary network disconnections. Such a strategy could minimise the problems outlined in Figure 2 but would have limited usefulness regarding a location update packet as the time passed may directly related to the accuracy of the position. As stated, location service inaccuracy can lead to inefficient and incorrect routing decisions. If the target destination position received by the geo-routing protocol is

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

303

inaccurate, the algorithm cannot function correctly. Therefore greedy neighbour nodes are chosen incorrectly and perimeter mode decisions are made with the incorrect location in mind. An illustration of how a packet can be routed incorrectly can be seen in Figure 3. Vehicle A wishes to route to vehicle B. The location service returns vehicle B’s last registered position in the location search index, B(old). This situation could occur because the location update interval is too large or more likely because vehicle B’s location update message failed to reached a location server and hence its location was not updated. Vehicle A routes in the direction of the stale position with the packet eventually dropped when the TTL is reached or because of failure of perimeter routing to find the destination i.e. no edge to route.

Fig. 3. Geo-routing with Stale Location Service Information

Therefore a number of factors contribute to routing failure and packet drops and are categorized as follows in next section: Source Node Isolated: Packets are discarded as a result of a vehicle having no neighbour table entries. This can be caused by a temporarily partitioned network, especially in a sparsely connected network. No Edge to Route: A packet has traversed the perimeter and has arrived back at the node at which it entered perimeter mode without finding the destination. This can occur because no route exists, because the destination vehicle may have left the network or because of inaccurately reported location information. WLAN Retry Threshold Exceeded: Packets are dropped by the WLAN MAC because of consistently failing retransmissions, exacerbated by routing loops and stale neighbour tables. TTL Reached: This represents the packets dropped due to the expiration of a packet’s Time-To-Live (TTL) field (64 hops) without ever reaching the intended destination. Interim Node Isolated: A node receives a packet just as it becomes isolated in the network.

304

A. O’ Driscoll and D. Pesch

4 Performance Evaluation 4.1 Simulation Models and Scenario Establishment While many VANET-specific routing protocols have recently been proposed, no clear “one size fits all” solution routing protocol for unicast vehicular communications has yet emerged as a prevalent standard. Greedy Perimeter Stateless Routing (GPSR) acts as the base geo-routing protocol for this simulation study. We acknowledge that GPSR has recognised drawbacks in vehicular networks which can affect the packet delivery success rate however this is immaterial to the content of this study as we are not concerned with the specifics of the geo-routing protocol but rather the negative effects that mobility and location inaccuracy has on base-line routing protocol performance. VANET specific geo-routing protocols that significantly improve on GPSR routing performance were listed in Section 2. Despite the individual merits that a particular routing protocol exhibits over another, they all require a location service to accurately report the location of the destination node. If the destination reported is inaccurate this will ultimately lead to their respective routing algorithms making incorrect routing decisions. A GPSR model has been developed in OPNET, as seen in Figure 4a. This is the first publicly available stateless geo-routing model that the authors are aware of that is available in OPNET. A previous model was referenced in [33] but the model is not freely available. Our model has been implemented according to the specification in [9] using the RNG for planarization.

Fig. 4. (a) OPNET GPSR Model (b) SUMO Extraction of 1500m2 Area of Cork City

Vehicular movement is modeled in OPNET via a custom mobility model based on trace files generated by SUMO [34], a microscopic road traffic simulation package. The vehicular spatial environment is based on an urban road topology of Cork City in Ireland, by importing detailed road layouts from OpenStreetMap.org and inserting

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

305

random traffic flows. SUMO simulates vehicle behaviour according to the Krauss vehicle following model. This model regulates vehicle speed and behaviour based on the movements of the preceding vehicles. Figure 4b illustrates the SUMO road topology generated for a 1500m2 area. We have limited the maximum speed of the vehicles to 50 km/h, the speed limit restriction within Cork City, as the default speed restrictions associated with the road types imported by the OSM map were too high for safe motoring within an urban environment. The speed profiles for two random vehicles are shown in Figure 5a. It can be seen that since that in the congested network, the vehicle movement is constrained by traffic congestion where as the other vehicle speed profile is relatively free flowing. We also consider a broad spectrum of traffic densities with SUMO injecting cars into the network at a specified Vehicle Arrival Rate (VAR). A low density network is represented with 2 vehicles per second injected into the network (VAR=2). A busy but somewhat free flowing density is considered with VAR=3 v/s and a densely populated scenario with a VAR=4.5 v/s is also considered. These arrival rates are dependant on the area size and were chosen by examining the GUISIM output while ensuring no collisions occurred for the dense scenario. The number of vehicles in the metropolitan area over the duration of the simulation is depicted in Figure 5b. Vehicular presence times are shown in Figure 5c for a VAR of 2 v/s. Vehicles exist in the simulation for approximately 170 seconds on average.

Random Vehicle, High - VAR=4.5 v/s Random Vehicle, Low - VAR=2 v/s 16 14

S p e e d (K P H )

12 10 8 6 4 2 0 1

51

101

151

Vehicle Duration in Network (s)

1200

V e h ic l e D e n s i ty

V e h ic le P re s e n c e T im e ( s )

High - VAR=4.5 v/s Low - VAR=2 v/s Medium - VAR=3 v/s

1000 800 600 400 200 0 0

200

400

600 Simulation Time (s)

800

1000

1200

450 400

VAR=2 v/s

350 300 250 200 150 100 50 0 1

101

201

301

401

501

601

701

801

901

Number of Cars Generated

Fig. 5. (a) Random Vehicle Speed Profiles (b) Vehicular Density (c) Vehicle Presence Times

306

A. O’ Driscoll and D. Pesch

Within OPNET, a square grid of 1500m2 is employed. All vehicles use the IEEE 802.11 MAC with a 2.4GHz radio interface and a transmission speed of 11Mbps. Radio propagation for these urban simulations follows the Two Ray Ground model incorporating the effects of path-loss with an exponent of 4dB, a log-normal shadowing component and antenna heights of 1.5m when determining the transmission range. Each simulation represents a time period of 1200 seconds. Vehicles are injected into the network for 600 seconds with the remaining time providing an opportunity for vehicles to complete their journeys. A vehicle generates a location query message approximately 10 seconds (varies slightly with jitter) after it enters the network, with subsequent requests made at 100s intervals. Every vehicle represents both a vehicle that will periodically query for a location and a target destination vehicle. Two custom OPNET models are used to generate these data queries as well as to provide the location service to the geo-routing protocol respectively. 4.2 Experimental Results It is our objective to obtain insight into the negative impact of mobility and location service inaccuracy on a geo-routing protocol. Given this, we must carefully design the simulation to distinguish between mobility induced and network based parameters that can impact on the successful delivery of a packet vs packets drops that occur as a direct result of inaccurate reported locations. Thus we first examine the impact of several key influences such as vehicular density, transmission range as well as the distance between source and destination vehicles i.e. range of data queries, on the geo-routing protocol performance. For this reason we assume the availability of the location service at no additional overhead. Unless specified otherwise, source and destination vehicle pairs are randomly chosen. It is assumed every vehicle knows its own position. A summary of the parameters used can be seen in Table 1. Table 1. Simulation Parameters Wireless Standard Propagation Model

802.11b Two-Ray Ground + Log-Normal Shadowing

Antenna Heights Radio Range Area Size Vehicle Injection Rate

1.5 m 50m-200m 1500 m2 Low: 2 v/s, Medium: 3 v/s High: 4.5 v/s 0-50 km/h uniform (500, 600) ms constant (1) s

Vehicle Speed Beaconing Interval Nbr Expiry Interval

4.2.1 Vehicle Transmission Range We next evaluate the impact of a varied transmission range on the Query Success Rate (QSR), with a query defined as any source to destination communication. Data messages are generated for random target vehicles within the urban topology. Experimental analysis on vehicular geo-routing protocols [7] and location services [24, 35, 36] often opt for overly optimistic radio ranges such as 250m. In contrast, the

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

307

simulations we describe assume a transmission range of 100m and in this section we evaluate the effects of a radio range as low as 50m. It can be seen in Figure 6a, that significant deterioration in the QSR can be noted as the vehicle radio range is reduced. A decrease from 62% to 6% is observed as the transmission range is reduced from 200m to 50m for the scenario with a VAR = 2 v/s. Similar decreases in QSR can be noted for the medium and highly dense scenarios. A short radio range will result in smaller neighbour tables thus limiting the number of potential next hop neighbours available to the vehicle for greedy routing. This can lead to excessively longer packet routes as perimeter routing will be more frequently employed. It also reduces the likelihood that the position of the destination can be resolved directly from the source vehicle’s neighbour table, increasing reliance on the location service as shown in Table 2 where 13.1% of queries are directly resolved for a 200m transmission range in comparison to 3.1% for a 50m transmission range. Similar results were noted for the medium and highly dense scenarios. Furthermore, the possibility of a partitioned network increases as the radio range is reduced along with the high probability that an end to end route does not exist between the source and destination. This can be observed in Figure 6b. As the transmission is reduced it can be observed that the number of packets dropped as there’s no edge to route on increases. This is as a result of perimeter mode routing being invoked more frequently and the completion of the perimeter failing to locate the destination or another node that can route greedily to the destination. Furthermore the number of packets dropped as a result of source nodes becoming isolated increases. The numbers of packets discarded due to the TTL being reached maintains steady values for all transmission ranges evaluated. It could be anticipated that the number of hops taken to resolve a query would increase as the radio range is reduced but we observed that the average number of hops and latency remained in the range of 2-4 hops. This is because this is representative of those packets successfully received, with transmitted packets being dropped for higher hop counts. This can be attributed to the higher risk of partitioning and isolated nodes. Comparable values were noted for the medium and highly dense scenarios. 1

1 Low - VAR=2 v/s Medium - VAR=3 v/s

Q u ery S u ccess Ratio (% )

0.8

High - VAR=4.5 v/s

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0.9 P ackets Drop p ed Catego riz atio n (% )

0.9

0.8

WLAN Retry Threshold Exceeded No Edge to Route Interim Node Isolated

TTL Reached Source Node Isolated

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0 200m

175m

150m

125m

100m

Transmission Range (m)

75m

50m

0 200m

175m

150m

125m

100m

75m

50m

Transmission Range (m)

Fig. 6. (a) Query Success Rate vs Transmission Range (b) Packet Drop Categorization vs Transmission Range VAR = 2 v/s

308

A. O’ Driscoll and D. Pesch

Table 2. Neighbour Resolved Locations, Hop Count, Query Latency vs Transmission Range Radio Range

200m 175m 150m 125m 100m 75m Locations Resolved Via Neighbour Table (Packets Transmitted)

50m

VAR 2 v/s

13.1% 10.7% 8.7% 7.3% 5.6% Average No of Hops to Resolve Query (Packets Received)

4.5%

3.1%

VAR 2 v/s

3.1668 4.278 4.315 3.924 3.575 Average ETE Query Latency (ms) (Packets Received)

2.828

2.173

VAR 2 v/s

3.339

2.104

1.48

3.721

3.684

3.284

2.98

4.2.2 Query Range We next examine the impact of query range, the distance between a source and destination vehicle and its impact on the success rate of a packet. Given an area size of 1500m2, the maximum distance that is possible between source and destination vehicles is 2121m when they are located in diagonally opposing corners of the area border. Up until this point, target destinations have been generated randomly, anywhere within the boundary of the network topology. In Figure 7 it can be seen that an increase in the distance between the querying vehicle and the target destination has a dramatic effect on the QSR, most notably when the query range extends from 0-100, i.e. direct neighbours of the source, to 100200 i.e. reliant on geo-routing. A steady decrease in QSR is noted as the distance increases though it can be observed that density improves the QSR slightly as more routes are available to forward the packet. The decrease in QSR can be attributed to the increasing likelihood of the network becoming partitioned as the distance increases, that the TTL may be reached or that a neighbour may have moved out of range of an interim hop. No packets were successfully received in the low density network past the range of 700-800m. While packets were received in the medium and higher density networks, a steady decline in the QSR is noted. It can also be observed in Table 3 that as expected there is an increase in the end-to-end delay for received packets as the number of hops increases, with 1 hop approximately corresponding to 1ms in delay for all vehicular densities. These results highlight the inadequacies of multi-hop routing over very large distances and the possible benefits if some infrastructure were available to support V2V packet routing. An outline architecture proposing use of Road-Side Units to aid V2V communications has previously been proposed by Gerla et al [21]. 4.2.3 Location Service Inaccuracy – Time-Related and Absolute Position Errors Now that we have studied the vehicular environment under the influence of a number of topological and mobility based parameters, we next examine the impact of time related position errors on the location service accuracy. Some location service protocols rely on static location update intervals to update the vehicle mappings within the location service index. This location update interval directly affects the accuracy of the position information available. It can be expected that a static update interval of t could lead to the return of inaccurate information if information is queried at time t-1, t-2 and so on. The location information stored for a vehicle could, by this time, be stale. While this might not cause a major drop in geo routing success

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

309

given a sufficiently low update interval the most likely reason for the return of inaccurate location information by the location service resulting in packet drop is as a result of a periodic update messages being delayed or dropped either as a result of a collision or lack of existence of a route to location service index. Thus the last registered position is returned by the location server. If update packets were to be frequently dropped, highly inaccurate location information could be returned. Deterioration in the packet delivery rate of GPSR can be seen when there is an increase in the time-related position error as shown in Figure 8a. The baseline QSR performance is already poor with 18.1%, 24.6% and 35% observed for the low, medium and highly dense scenarios respectively when no error is introduced.

Q u ery S u cce ss Ra tio (% )

1

Low - VAR = 2 v/s Medium - VAR = 3 v/s

0.9 0.8

High - VAR = 4.5 v/s

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0-100

100-200 200-300 300-400 400-500 500-600 600-700 700-800 Query Range (m)

Fig. 7. Query Success Rate vs Query Range Table 3. Hop Count & Query Latency vs Query Range Query Range

VAR 2 v/s VAR 3 v/s VAR 4.5 v/s VAR 2 v/s VAR 3 v/s VAR 4.5 v/s

700600500400300200800 700 600 500 400 300 Average No of Hops to Resolve a Query (Packets Received) 9 8.08 8 5.4 4.43 13.3 11.3 9.43 7.35 5.52 4.02 14.9 11.6 9.73 7.9 5.5 4.13 Average ETE Query Latency (ms) (Packets Received) 8.9 7.33 7.3 4.68 3.79 14.5 12.2 9.45 7.1 5 3.5 15.1 11.9 9.62 7.56 5.8 3.67

100200

0 - 100

2.4 2.44 2.71

1.02 1.01 1

1.16 1.82 2.1

0.3 0.3 0.3

It can be seen that the QSR decreases from 18.1% to 3.9% as the error increases from 0-40s for a VAR of 2 v/s. A very small amount of packets are still received up to the 25s error even though a vehicle could have theoretically travelled a distance of 325m from the report position. This is as a result of small periods of congestion in certain areas of the network such as intersections where the network is not free flowing. Furthermore, the QSR never decreases to 0% because the introduced error is only applicable to those destination queries resolved via the location service and not directly via the neighbour table. The inaccurate location information results in the incorrect operation of the routing protocol with the packet forwarded towards an old

310

A. O’ Driscoll and D. Pesch

location. Thus packets will be routed “greedily” to incorrect neighbours and will traverse incorrect faces when in perimeter mode. While there is a slim possibility that an interim node encountered on the path has a record of the destination in its neighbour table and that the packet may be successfully delivered – the more likely result is that the packet will continue to loop until the TTL has fully decremented or until a full perimeter is traversed and no viable next edge is found as shown in Figure 8b. The geo-routing cannot make correct routing decisions with inaccurate location information. We found that the QSR for the highly congested topology is impacted less negatively. This is a result of the high VAR of 4.5 v/s. The network becomes densely populated and this heavy congestion results in little or no vehicle movement i.e. vehicles are stuck in heavy traffic. As a result their reported position remains relevant for a longer time period, regardless of the level of time error being purposefully introduced. In contrast for a VAR of 3 v/s, considered to be a busy yet moving traffic flow, it can be seen that introduced time error has an impact all the way through the simulation. 0.4

1

0.3

Packets Dropped Categorization (%)

Query Success Ratio (%)

0.9

Low - VAR=2 v/s Medium - VAR=3 v/s High - VAR=4.5 v/s

0.35

0.25 0.2 0.15 0.1 0.05

WLAN Retry Threshold Exceeded No Edge to Route Interim Node Isolated

0.8

TTL Reached Source Node Isolated

0.7 0.6 0.5 0.4 0.3 0.2 0.1

0 40s

35s

30s

25s

20s

15s

10s

5s

0

0s

40s

35s

30s

25s

Time-Related Error (s)

20s

15s

10s

5s

0s

Time-Related Error (s)

0.4 Low - VAR=2 v/s Medium - VAR=3 v/s High - VAR=4.5 v/s

Query Success Ratio (%)

0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 250

225

200

175

150

125

100

75

50

0

Distance Error (m)

Fig. 8. (a) Query Success Rate vs Time-Related Distance Error (b) Packet Drop Categorization vs Time-Related Distance Error (c) Query Success Rate vs Location Service Distance Error

Recent location service algorithms typically require a vehicle to update its location after a pre-defined distance has been travelled in order to overcome the drawbacks of static time based update intervals. This distance is usually 100m but this typically assumes a transmission range of 250m. We introduced a distance error by referencing

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

311

a vehicle’s past positions from its trajectory list, determining the distance travelled in each increment until a position is returned that is equal to or slightly greater than the specified distance error required. As previously noted the baseline QSR is already poor with no location error. It can be seen in Figure 8c, that once the transmission range of a node is exceeded the QSR deteriorates quickly with a few occasional queries resolved from long perimeter routes with entries for the destination in their neighbour tables. Furthermore even though a vehicle has travelled a distance, they may not be necessarily be out of range of their old reported position depending on the road layout i.e. they could have travelled around a U-shaped or curved bend in the road and have only travelled a short Euclidean distance from their old reported position, thus can still be found via the unit disk graph transmission assumption. It can be seen however that like the time related errors, QSR quickly decreases such that only direct neighbour queries can be solved from neighbour table entries. It can be seen that for the highly dense scenario the QSR did not drop such that all location service queries failed but this is because the vehicles may not have travelled the distance required to introduce a sufficiently large error and hence we are the error introduced is limited. The QSR shown for the medium and low density scenarios is more indicative of the effect of the error.

5 Conclusions and Future Work This paper describes the impact on a geo routing protocol of vehicular mobility characteristics such as transmission range, query range, vehicular density as well as quantifying the impact of explicit location errors. Paper findings provide guidelines for what query success rate can be expected, given particular vehicle mobility scenarios and absolute location inaccuracy. Future work will consider the availability and fault tolerance of the location service, a particularly challenging topic in multihop vehicular networks with high vehicle speeds. Location server availability for update will have a direct impact on accuracy and availability for query is vital if communication is even to be possible. We will also examine the possibility of exploiting vehicular characteristics such as the intermittent existence of roadside unit infrastructure to improve connectivity and aid location service scalability.

References 1. Chen, W., Guha, R., Kwon, T., Lee, J., Hsu, Y.: A Survey and Challenges in Routing and Data Dissemination in Vehicular Ad Hoc Networks. Wiley Interscience Journal on Wireless Communications and Mobile Computing, WCMC (2009), doi:10.1002/wcm.862 2. Ros, F.J., Ruiz, P.M., Sanchez, J.A.: Mobile Ad Hoc Routing in the Context of Vehicular Networks. In: Vehicular Networks From Theory to Practice. Chapman & Hall/CRC Computer and Information Science Series (2009) 3. Zhao, J., Cao, G.: VADD: Vehicle-Assisted Data Delivery in Vehicular Ad Hoc Networks. IEEE Transactions on Vehicular Technology 57(3) (May 2008) 4. Zhongyi, L., Tong, Z., Wei, Y., Xiaoming, L.: GOSR: Geographical Opportunistic Source Routing or VANETs. ACM SIGMOBILE Mobile Computing and Communications Review 13(1) (January 2009)

312

A. O’ Driscoll and D. Pesch

5. Wu, H., Fujimoto, R., Guensler, R., Hunter, M.: MDDV: A Mobility-Centric Data Dissemination Algorithm for Vehicular Networks. In: ACM 1st International Workshop on Vehicular Ad hoc Networks (2004) 6. Liu, G., Lee, B., Seet, B., et al.: A Routing Strategy for Metropolis Vehicular Communications. In: Kahng, H.-K., Goto, S. (eds.) ICOIN 2004. LNCS, vol. 3090, pp. 134–143. Springer, Heidelberg (2004) 7. Yang, Q., Lim, A., Shung, L., Fang, J., Agrawal, P.: ACAR: Adaptive Connectivity Aware Routing Protocol for Vehicular Ad Hoc Networks. In: IEEE 17th International Conference on Computer Communications and Networks (ICCN) (August 2008) 8. Lochert, C., Mauve, M., Fubler, H., Hartenstein, H.: Geographic Routing in City Scenarios. ACM SIGMOBILE Mobile and Computing and Communications Review 9(1), 69–72 (2005) 9. Karp, B., Kung, H.T.: GPSR: Greedy Perimeter Stateless Routing for Wireless Networks. In: ACM/IEEE 6th International Conference on Mobile Computing and Networking (MOBICOM) (August 2000) 10. Nzouonta, J., Rajgure, N., Wang, G.: VANET Routing on City Roads Using Real-Time Vehicular Traffic Information. IEEE Transaction on Vehicular Technology 58(7) (September 2009) 11. Lochert, C., Hartenstein, H., Tian, J., Fubler, H., Hermann, D., Mauve, M.: A Routing Strategy for Vehicular Ad hoc Networks in City Environments. In: Proceedings of the IEEE Intelligent Vehicles Symposium, Columbus (June 2003) 12. Jerbi, M., Senouci, S., Meraihi, R., Ghamri-Doudane, Y.: An Improved Vehicular Ad Hoc Routing Protocol for City Environments. In: IEEE International Conference on Communications (ICC) (2007) 13. Kim, H., Paik, J., Lee, B., Lee, D.: SARC: A Street-Based Anonymous Vehicular Ad Hoc Routing Protocol for City Environments. In: IEEE/IFIP International Conference on Embedded and Ubiquitous Computing (EUC) (2008) 14. HsiaoKuang Wu, E., Kumar Sahu, P., Sahoo, J.: Destination Discovery Oriented Position Based Routing in VANET. In: APSCC 2008 (2008) 15. Kiess, W., Fubler, H., Widmer, J., Mauve, M.: Hierarchical Location Service for Mobile Ad-Hoc Networks. ACM SIGMOBILE Mobile Computing and Communications Review (October 2004) 16. Jannotti, J., De Couto, D.S.J., Karger, D.R., Morris, R.: A Scalable Location Service for Geographic Adhoc Routing. In: Proceedings of 6th Annual ACM/IEEE MobiCom, Boston, USA (2000) 17. Kaesemann, M., Fuessler, H., Hartenstein, H., Mauve, M.: A Reactive Location Service for Mobile Ad Hoc Networks, Technical Report TR-14-2002, University of Mannheim, Germany (2002) 18. Luo, X., Camp, T., Navidi, W.: Predictive Methods for Location Services in Mobile Ad Hoc Networks. In: Proceedings of the 19th IEEE Parallel and Distributed Processing Symposium (2005) 19. Kumar, V., Das, S.: Performance of a Dead Reckoning Based Location Service for Mobile Ad Hoc Network. ACM Wireless Communications and Mobile Computing Journal 4(2) (March 2004) 20. Sharma, S., et al.: A comparative study of mobility prediction schemes for gls location service. In: Proceedings of the IEEE Vehicular Technology Conference, VTC 2004 (2004) 21. Gerla, M., Zhou, B., Lee, Y.-Z., Soldo, F., Lee, U., Marfia, G.: Vehicular Grid Communications: The Role of the Internet Infrastructure. In: Proceedings of the 2nd International Workshop on Wireless Internet, ACM WICON, Boston, USA (2006) 22. Chang, Y.-J., Shih, T.L.: Intersection Location Service and Performance Comparison of Three Location Service Algorithms for Vehicular Ad Hoc Networks in City Environments.

Quantifying the Negative Impact of Mobility and Location Service Inaccuracy

23.

24.

25.

26.

27.

28.

29.

30.

31.

32.

33.

34. 35.

36.

313

In: IEEE International Symposium on Wireless Pervasive Computing (ISWPC), Greece (May 2008) Amar, E., Boumerdassi, S., Renault, E.: Hierarchical Location Service with Prediction in Mobile Ad-Hoc Networks. International Journal of Computer Networks & Communications 2(2) (March 2010) Bai, X., Ye, X., Li, J., Jiang, H.: VLS: A Map-Based Vehicle Location Service for City Environments. In: IEEE 44th International Conference on Communications (ICC) (June 2009) Mo, Z., Zhu, H., Makki, K., Pissinou, N.: Mobility Assisted Location Management (MALM) for Vehicular Ad Hoc Networks. In: IEEE Wireless Communications and Networking Conference (April 2008) Mo, Z., Zhu, H., Makki, K., Pissinou, N., Karimi, , M.: On Peer-to-Peer Location Management in Vehicular Ad Hoc Networks. International Journal on Interdisciplinary Telecommunications and Networking (to appear), Draft Copy available at: http://users.cis.fiu.edu/~zhuha/PLM_Draft.pdf Cabrera, V., Ros, F.J., Ruiz, P.M.: Simulation-based Study of Common Issues in VANET Routing Protocols. In: IEEE 69th International Vehicular Technology Conference (VTC) (April 2009) Kim, Y., Lee, J., Helmy, A.: Impact of Location Inconsistencies on Geographic Routing in Wireless Networks. In: ACM International Workshop on Modeling Analysis and Simulation of Wireless and Mobile Systems (MSWIM) (2003) Seada, K., Helmy, A., Govindan, R.: On the Effects of Location Inaccuracy on Geographic Face Routing in Wireless Networks. ACM Mobile Computer and Communications Review (2003) Son, D., Helmy, A., Krishnamachari, B.: Mobility Induced Location-Errors and its Effect on Geographic Routing in Ad Hoc Networks: Analysis and Improvement using Mobility Prediction. IEEE Transactions on Mobile Computing 3(3) (July 2004) Seada, K., Helmy, A., Govindan, R.: Modeling and Analyzing the Correctness of Geographic Face Routing under Realistic Conditions. ScienceDirect Ad-Hoc Networks 5(6) (August 2007) Peng, B., Mautz, R., Kemp, A., Ochieng, W., Zeng, Q.: On the Effect of Localization Errors on Geographic Routing in Sensor Networks. In: IEEE International Conference on Communications (2008) Hyland, M., Mullins, B., Baldwin, R., Temple, M.: Simulation-Based Performance Evaluation of Mobile Ad Hoc Routing Protocols in a Swarm of Unmanned Aerial Vehicles. In: IEEE 21st International Conference on Advanced Information Networking and Applications Workshops (AINAW) (2007) Bonert, M., Krajzewicz, D., Wagner, P.: The Open Source Traffic Simulation Package SUMO (2006) Chang, Y.-J., Shih, T.L.: Intersection Location Service and Performance Comparison of Three Location Service Algorithms for Vehicular Ad Hoc Networks in City Environments. In: IEEE International Symposium on Wireless Pervasive Computing (ISWPC), Greece (May 2008) Amar, E., Boumerdassi, S.: Enhancing Location services with Prediction. In: ACM 5th International Wireless Communications and Mobile Computing Conference (IWCMC) (June 2009)

Establishing Trust on VANET Safety Messages (Invited Paper) Subir Biswas1 and Jelena Miˇsi´c2 1

Dept. of Computer Science, University of Manitoba Winnipeg MB, Canada R3T 2N2 [email protected] 2 Dept. of Computer Science, Ryerson University Toronto ON, Canada M5B 2K3 [email protected]

Abstract. We introduce a new scheme for safety message authentication in VANETs. For a practical implementation of VANET, we anticipate that road side units (RSUs) are not physically protected and are prone to several different attacks including node compromise attacks. Thus, an RSU should not be automatically trusted by on road vehicles. In our proposed scheme, a road side controller (RSC) is responsible for controlling all the RSUs, and delivering messages through RSUs to vehicles in a given area, where each RSU uses a proxy signature mechanism based on Elliptic Curve Cryptography (ECC), which is a variation of known ECDS-based proxy signature schemes and modified according to the VANET’s criteria and security requirements. The underlying network constraints and properties from VANET standards have been taken into consideration along with the security, reliability and other related issues. We also discuss the potential forgery and attack scenarios on our proposed scheme. The security analysis and simulation results prove the strength and adaptability of our proposed scheme in future VANETs. Keywords: VANET, ECDSA, proxy signature, replay attack.

1 Introduction Vehicular Ad hoc Networks (VANETs) open up a new horizon to researchers, developers, and entrepreneurs with several multipurpose applications for driving safety, navigation, and entertainment. A VANET consists of three fundamental components, namely on board unit (OBU), road side unit (RSU), and an appropriate infrastructure to coordinate with the whole system as well as the connectivity to the Internet. Generally, an RSU is responsible for providing vehicles on road with safety information like traffic collision warning, accident notification, potential rules violation warning, changed road condition etc. It can also be used as an advertisement agent for commercial benefits. On the other hand, an OBU can communicate, exchange messages regarding traffic/road condition, destinations etc. with other vehicles on road. Apart from that, an 

This research has been partially funded by the Canadian Govt’s AUTO21 project.

J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 314–327, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Establishing Trust on VANET Safety Messages

315

OBU periodically disseminates the vehicle’s GPS location data, brake, acceleration information etc. to the other vehicles in its vicinity, and to the closest RSU. For all these anticipated ideas, trust is a vital aspect for the elements of VANETs since a user (e.g. a driver) will be comfortable with any VANET application when he can reasonably trust the network components, and at the same time, a VANET is useful only when all its entities are trustworthy. If an adversary deliberately broadcasts false messages for traffic safety, or sends an old and expired notification to all the vehicles on road, it may misdirect the entire traffic or even impair the transportation safety. Similarly, if roadside electronic toll collection systems are not trustworthy, it may bring the whole system into a massive disarray. As required by VANET protocols, RSUs are normally installed on road side locations where they are usually without any good physical protection and enough surveillance. Hence, an RSU is always on the brink of being compromised by an adversary for sending false, expired, and misleading messages that might bring some dangerous consequences for an accepting OBU. To overcome this risk factor, vehicular communication must have the capability of verifying the identity of the message sender as well as maintaining the integrity of the delivered message, both of which can be accomplished through a suitable signature protocol. However, identity verification in a realtime VANET environment is not a simple task due to high and variable velocity of nodes, varying node density, and the need to operate on roads with nonuniform characteristics. Thus, the issue of scalability is of prime importance as, under heavy traffic, a single controller might need to attend to hundreds, perhaps even thousands of vehicles in a given segment of the transportation network. Given that future VANETs will be implemented using IEEE 802.11p WAVE (Wireless Access in Vehicular Environments) protocol family [1], it is important to limit the wireless traffic intensity in order to avoid packet collisions at the Medium Access Control (MAC) layer. The presence of many messages from vehicles and RSUs on a particular road may increase the message collision rate and thus impair the performance of on-road vehicular communication. Eichler et al. [2] have shown that the WAVE standard can’t deal with many high priority messages in a dense network scenario. Hence, we need a trust scheme that has low computational complexity, high scalability, as well as a reliable and quick verification mechanism. To cope with the above requirements, we look into the issue of safety message delivery in vehicular networks with a perspective of authenticating the RSU as a valid member of the corresponding RSU group to on road OBUs, and delivering the signed safety messages to OBUs by the RSU on behalf of a road side controller (RSC). As mentioned before, RSUs may deliver several messages including routine messages on road-safety issues, accident notifications, traffic congestion alerts, and commercial messages etc. An RSU may require to keep broadcasting each message repeatedly for a particular time span. We propose to exploit the features of proxy signature [3] for the RSU message delivery. The term proxy signature refers to a variation of digital signature that designates an entity (called a proxy signer) to sign a message on behalf of the original signer. We

316

S. Biswas and J. Miˇsi´c

Fig. 1. The framework for trust in safety message delivery in VANETs

considered a number of signature schemes and found the Elliptic Curve Digital Signature Algorithm (ECDSA) [4] most suitable as the original signature protocol for fast and efficient signing of safety messages over the VANET. The current VANET standards for security services [5] are also suggesting the ECDSA scheme for signing the secure messages in VANETs. In our scheme, RSUs will be the proxy signers signing safety and other application messages to the OBU recipients on behalf of the original creator of the messages- a road side controller (RSC). A recipient of the messages (OBU) can verify the identity of the original signer, and it can also verify the integrity of the contents of the received message. Our scheme is derived from the modification of some of the ECDSA-based proxy signature protocols which are described in [6,7,8]; using which a corresponding RSU can neither replay an expired message, nor alter an original message. Therefore, the control of the message broadcast is kept with the message originator. A partial delegation mechanism [3] produces a new secret key from the original signer’s secret, and the new secret is then used as the key for proxy signing. We organize the paper in the following manner. Section 2 discusses the related work on VANET authentication and trust management together with evolution of some proxy signature schemes. A brief account of ECDSA preliminaries is given in Section 3. Section 4 illustrates the proposed mechanism for VANET safety message delivery. The security analysis is provided in Section 5, while the simulation results are presented in Section 6. Finally, Section 7 concludes the paper. all subsequent paragraphs are.

2 Related Work Our study of related work includes the existing VANET message authentication literature, as well as the digital proxy signature approaches which we considered while proposing our scheme.

Establishing Trust on VANET Safety Messages

317

2.1 On Message Authentication in VANETs Number of papers have been published in recent years addressing the issue of VANET authentication (e.g. [9], [10], [11], [12] and [13]), where researchers mainly focused on authenticating OBU messages to RSUs and to other OBUs in the light of vehicular anonymity and other VANET requirements while the safety messages from RSUs are mostly taken for granted, and considered trustable. Lin et al. [14] proposed in 2007, an ID-based signature [15] scheme where the RSU location is used as the public key for the message signature. Each message sent from the RSU contains the physical location information so that once the message is received by an OBU, it can be verified based on the location information. This is how it prevents the RSU-OBU communications from a potential replication attack as messages from the RSU will be discarded if the RSU is dislocated. In 2009, Studer et al. [16] proposed VANET Authentication using Signatures and TESLA++ (VAST). A combination of ECDSA [4] and a modified TESLA protocol [17] have been used to carry out VANET’s requirement of message verification in an efficient way. Upon receiving a message from a VANET entity, the receiver would perform 2 types of verification: 1. TESLA++ verification, and 2. an ECDSA verification is used normally when non-repudiation is necessary. To prevent the major computational and memory-based DoS attacks, the received payload is verified using TESLA++ before the ECDSA verification is performed. If the initial TESLA++ verification fails due to some reasons, CPU utilization, as well as the size of the message queue are taken into consideration before switching to ECDSA mode of authentication. However, RSUs are still assumed to be trusted and the approach doesn’t explicitly address the issue of replay attack and false message injection attack by a supposedly legitimate RSU. There is no clear indication on how a VANET should react upon detection of a node compromise. Moreover, Hass et al. in [18] remarked that ECDSA performs better than a TESLA implementation at longer communication distances since TESLA requires a second packet delivery for the message verification purpose. Wen et al. [19] exploits the spatial and temporal properties of physical layer channel responses for securing each communication pair in VANETs. The basic idea is to distinguish one sending transmitter from another based on the physical layer measurements for a series of messages. The sending entity attaches the authenticating signals which is a unique channel response along with the information payload transmission. A receiving entity can verify the legitimate transmission by comparing the authenticator signal with the estimated channel response. This approach is efficient and scalable for a physical layer approach, but again, comes with some limitations. For instance each vehicle has to be preloaded with public keys of other vehicles to be able to access the network. This will obviously pose a huge task of updating and maintaining a huge number of keys as the size of the VANET grows. Authors also haven’t addressed the node compromise attack which can have a deadly consequence in the form of false message broadcast, or replaying expired safety messages. Among the few other recent work, ASIC scheme [20] introduces a faster and efficient way of aggregated verification of signatures and certificates for VANETs using Bilinear pairing technique. This approach can verify a large number of signatures and certificates

318

S. Biswas and J. Miˇsi´c

to ensure a reliable and optimized operation of VANETs, nevertheless, establishment of trust on VANET safety messages has not been covered by this mechanism. 2.2 On Proxy Signatures The original proxy signature scheme, proposed by Mambo et al. [3], was further extended by Kim et al. [21] who proposed two additional features – proxy signature by partial delegation with warrant and the threshold delegation based proxy signature. Further enhancements include blind proxy signature schemes [8,22,23], and [24] by which a proxy signer is made unable to manipulate the message contents (and, replay the expired messages). However, blind proxy signatures are not practical for VANET applications, since they require a new proxy tuple to be generated and delivered to the proxy signer every time there is a new message to be posted. ECDSA-based proxy signatures are comparatively new in research and [6,7,25] are among the most recent publications.

3 Preliminaries In this section, we give a quick review on the Elliptic Curve facts that would be necessary for understanding the cryptographic materials used in ECDSA. We are going to use an elliptic curve over a finite field for our scheme. A Finite Field. A finite field F p is a finite set of of p elements along with addition and multiplication operations on F. The number of elements is denoted as the order of the finite field. There exists a finite field of order q if and only if q is a prime power, and on the other hand, if q is a prime power, then there exists only one finite field of order q denoted by Fq . Elliptic Curve. An Elliptic Curve E over a finite field F p is defined in the form of the following equation: y2 = x3 + ax + b, (1) where a prime p > 3; a, b ∈ Fp , and 4a3 + 27b2  0(mod p). The set of elements of the Elliptic Curve E(F p ) consists of the points (x, y) where x ∈ F p and y ∈ F p . A point at infinity O together with the set of points E(F p ) identifies an elliptic curve. Note that the addition, multiplication, and inversion operations on an elliptic curve points are different from ordinary binary operations. Please refer to [26] for the detailed description of the above mentioned operations. ECDSA. The domain parameters of Elliptic Curve Digital Signature Algorithm (ECDSA) require an Elliptic Curve E over a finite field of size q, and a base point G ∈ (Fq ). Value q is chosen as a prime power pt where p is a prime number, and t is a positive integer. In our case of elliptic curve, t = 1, thus p = q. Also, as indicated in (1), two field elements a and b are chosen, where, a, b ∈ (Fq ). All these parameters could be shared by the entities or by some specific user depending upon the ECDSA configuration.

Establishing Trust on VANET Safety Messages

319

4 ECDSA-Based Proxy Signature for VANETs The ECDSA-based proxy signature mechanism requires a proxy key which should be preprocessed and generated by the trusted authority. This preprocessing is done at the trusted authority and RSC during the network deployment phase, and should be kept secret from other entities. Being equipped with the proxy key, an RSU can sign on the safety message on behalf of the corresponding RSC using the proxy key and other session components. We assume that RSUs are independent of each other while a predefined group of RSUs in a geographical area are working under the same RSC. The RSC (and not the RSU) is physically protected, trusted entity, and responsible for issuing safety messages for OBUs. Then, the safety messages are delivered through individual RSUs in a same RSU group. Fig 1 outlines our network design. An OBU is preloaded with the public key of the trusted authority, which can verify the safety message with the signature components of the received payload. Following are the necessary steps for our scheme. 4.1 Key Material Preprocessing Consider a large prime q for a finite field F∗q and let G be a generator over an Elliptic Curve Eq . Also, x be a random number, generated by the trusted authority (TA), and would be used as the private key. The public key for x, V = xG is calculated and delivered to all concerning entities along with the certificate from the TA. TA generates a unique and distinct random secret ko where (1 < ko < q). It then computes: g = ko Gmod q

(2)

For each member RSUi in VANET, TA generates a random number ki where (1 < ki < q). Ri = k i G

(3)

(x1 , y1 ) = ki g

(4)

The proxy key for the RSUi is calculated at TA using the following equation. Si = (xx1 + ki )ko−1 mod q

(5)

This proxy key will be used by the RSU to sign a message payload when issued by the corresponding RSC. Hence, each RSUi is pre-loaded with the proxy key Si , along with g, Ri , x1 at the end of this key material preprocessing phase. 4.2 Payload Preprocessing Each message M includes a message payload m, and a message expiry information tx (i.e. M = m||tx ). When there is a safety message M to be released, the RSC determines a new session parameter k p using the hash over ko , m and tx . This session parameter is unique for each session of safety message delivery, i.e. a different message session will

320

S. Biswas and J. Miˇsi´c

have a different k p value than the current message. We choose SHA-1 (160 bit output blocks) for the hash operation. k p = h(ko ||m||tx )

(6)

The used ko is the same ko of eqn.(2). The purpose of generating this session parameter is to prevent the vehicles from the replay attacks as described in the security analysis section. The RSC further calculates: (x p , y p ) = k p g

(7)

Then, the safety message M, session parameter k p , and x p are delivered to all the legitimate RSUs through a secure channel. The next step is producing the proxy signature, carried out by individual RSUs operating under the RSC. 4.3 RSU Proxy Signature The RSUi receives the tuple (M||k p ||x p ) intact from the corresponding RSC, signs the message content on behalf of the RSC, and keeps broadcasting the signed message contents to the on road vehicles until the message is expired. Sr,i = k−1 p (h(M) + Si x p )mod q

(8)

The combination (M||g||Ri ||x1 ||x p ||Sr,i ) is used as the broadcast payload for the safety message and transmitted to all the OBUs in the communication range of the RSUi . We assume that the length of the prime q is 160 bits. Table 1 gives the account of the lengths of the payload components. Table 1. Size of the safety message overheads. q is chosen of 160 bits.

Component Size (in bits) g 160 Ri 160 x1 80 xi 80 Sr,i 160 Total 640 (= 80 Bytes)

Message Verification by OBU. The received contents are utilized in the following verification mechanism. −1 (x j , y j ) = (h(M)g + x p(Ri + x1 Q))Sr,i mod q

(9)

If (x j = x p ) the received message is accepted, else rejected. Note that the inverse operations used in 8 and 9 (and all throughout the paper) are modular inverse operations. Refer to [26] for the details on modular inverse operation.

Establishing Trust on VANET Safety Messages

321

4.4 Enhancement Table 2 lists the components of time complexity of our scheme. The most expensive operations here are the point multiplication operations since the mechanism of point multiplication involves both ordinary multiplications as well as modular inversion operations. In the above description, workload for signature generation in the RSU is kept low by having no point multiplications in that phase. However, we can transfer a couple of point multiplications from the verification phase to the signature generation phase. This change would be significant as verifying parties are OBUs which need to save on battery power and more importantly, an OBU is attached with a fast moving vehicle. Table 2. Time complexity of the proposed approach. t pt , tm , tmi , and th f are to mean the time taken by a point multiplication, ordinary multiplication, modular inverse, and hashing operations respectively.

Type Point Multiplication Multiplication Modular Inverse Hash function

Preprocessing 4 t pt 2 tm 1 tmi 1 th f

Signature 0 t pt 2 tm 1 tmi 1 th f

Verification 3 t pt 3 tm 1 tmi 0 th f

As the extension of our scheme, we shift the load of computation from the OBU to the RSU. The verification operation defined for an OBU (eqn. 9) is replaced by the following set of calculations to take place in the RSU: −1 u1 = h(M)Sr,i mod q −1 u2 = x p Sr,i mod q u3 = x1 u2 mod q

(10)

d = u 1 g + u 2 Ri

(11)

An alternative message payload (M||d||u3 ||x p ) is then sent to the OBU. If we consider q a 160 bit prime number, d and u3 will be of 160 bits each while the other component xi is of 80 bits. The total size of the safety message overhead is only 50 Bytes which is much lighter than the one prior to the extension (80 Bytes). Our scheme overheads just 10 Bytes extra for proxy signature on a safety message compared to the ordinary ECDSA signature (40 Bytes) which is suggested by the VANET standards for message delivery. Therefore, our approach is not much extravagant on communication bandwidth for the safety message broadcast. The verification process at the OBU would now compute: (x j , y j ) = d + u3V

(12)

Again, if the relation (x j = x p ) holds, the message is a valid one and therefore accepted by the receiving OBU, else it’s discarded. Fig. 2 summarizes the scheme with the help of a flow diagram.

322

S. Biswas and J. Miˇsi´c

Fig. 2. Flow diagram of the ECDSA-based Proxy Signature for VANETs

5 Security Analysis The security of our scheme depends mainly on the size of the prime q, and the hardness of solving the elliptic curve problem. We assume here that the RSC is a trusted authority, and secured against any kind of physical compromise, whereas an RSU is not trusted by the on road vehicles. An unprotected RSU can be compromised by an adversary attempting to inject false messages, or modifying the messages in order to harming the traffic system. A replay attack is possible be launched using a valid signature on an expired message. A malicious RSU might attempt to forge other RSU’s signature to falsify an innocent RSU by sending a valid signature on a malicious message. This property is referred to as exculpability. Another potential misbehavior is repudiation of signed messages, which refers to a situation when a malicious RSU would deny its

Establishing Trust on VANET Safety Messages

323

involvement in producing a particular signature which has been actually generated by that RSU. In the light of the above discussion, we derive the following lemmas and proofs on our scheme. Lemma 1. An RSUi can not generate a valid proxy key Si . Proof. In order to generate a valid proxy key Si , an adversary would require the Trusted Authority’s secret key (x), and two other random secrets ki , ko as indicated in eqn.(5). The secret x is irreversible from the knowledge of the public key V , since that involves point multiplications of an elliptic curve Eq . The other two random numbers ki and ko are generated and stored at the TA. Assuming that both of them are 160 bit random numbers, the joined probability of a successful brute force guessing of ki and ko is 1/(2160 − 1)2 . Lemma 2. An RSUi can not launch a false message attack and a replay attack. Proof. While signing a message payload M on behalf of the RSC, RSUi uses k p (in eqn.(6)) which has been derived by hashing the ko value and the payload M. The value ko is a secret kept at the TA, and any changes on M would result into a different k p value for which the signature would be different. Note that the payload M contains the main message m, as well as the message expiry information tx (i.e. M = m||tx ). Therefore, for a modified tx , k p would be changed with a consequence of an invalid signature from the RSUi . If we deploy SHA-1 for hashing in eqn.(6), the probability that an adversary or a malicious RSUi would be successful in brute forcing the hash function is 1/(2160 − 1). Lemma 3. A proxy signature by an RSUi is non-repudiable and exculpable. Proof. The random number ki is a non-overlapping one which makes a proxy key Si unique for a particular RSUi (refer to eqn.(5)). We assume that the number of RSUs working under a given RSC is much less than q. Thus, an adversary can not reproduce the Si value with an acceptable probability. Hence, a proxy signature for a given message Sr,i can only be created by the RSUi , and therefore, non-repudiable. RSC stores all individual proxy keys (Si ) along with the corresponding ki , k p , and x p values. If there is any dispute, the RSC can reproduce the signature using the credentials of the disputed RSU. Therefore, an RSUi can not sign a message that would appear as signed by a different RSU. That preserves the exculpability property. Note that the RSC can always reproduce d and u3 for a given payload M. Since reproducing them would require Sr,i and Ri , which are private parameters and only accessible to the RSC and RSUi , we can argue that the enhancement of our protocol is non-repudiable and exculpable.

6 Simulation We evaluate our scheme for VANET in NS2 simulation over a 8 lane straight highway (4 lanes in either direction) of length 1.5 km. Vehicles with varying traffic densities and speeds are moving in both directions. We assume that all the vehicles in the simulation scenario are always in the communication range of a RSU group. A safety message

324

S. Biswas and J. Miˇsi´c

is signed using our scheme by an RSU and broadcast to OBUs. We don’t simulate RSCs here assuming that the RSC has already delivered the safety message and other credentials for proxy signature using some secure means. Also, we don’t simulate the initial wireless association and certificate verification process to keep things simple and straight. We configure each RSU and OBU with communication range set to 300m while all the devices are equipped with an omni-directional antenna. The Two-Ray Ground propagation model is used on top of the Nakagami model at the physical layer. All other PHY and MAC layer parameters are set as suggested by NS2 Mac/802 11Ext and Phy/WirelessPhyExt for IEEE802.11p. The bandwidth of the channel is fixed at 6Mbps. The interface queue length for each device is set to 20 for OBUs and RSUs in the simulation. Each OBU periodically (every 100 ms) sends a heartbeat message that contains its current address, location, velocity information along with the size of the packet, broadcast address, and authentication data. The size of the packet delivered by the RSU is considered 500 Bytes including 50 Bytes of signature overhead from our scheme. An RSU broadcasts the signed safety message periodically (every 100 ms) until the message gets expired.

Fig. 3. Percentage of on-road vehicles verifying the safety message for different message expiry time (tx )

Fig. 3 shows the impact of safety messages’ expiry time on the proportion of vehicles that would be able to receive and verify them for a traffic scenario of 200 vehicles moving at three different speed levels. For a faster moving traffic (speeding at least 70Km/h), it takes between 10 to 15 seconds to get the message verified by all the vehicles in the communication range. Slower traffic on the other hand, would require a longer broadcast session in order for an OBU to receive and verify the same message. Fig. 4 reveals that the percentage of vehicles verifying the signed safety messages is almost constant with the different size of the traffic, while tx > 10 sec enables 100% of vehicles in the communication range, verifying safety messages within the given time period.

Establishing Trust on VANET Safety Messages

325

Fig. 4. Percentage of OBUs verifying a safety message for different number of vehicles in the communication range

Fig. 5. Percentage of safety messages verified by one OBU for different number of vehicles in the communication range

As indicated in Fig. 4, for tx = 10 to 15, almost all the vehicles receive and verifies a safety message, while for a shorter expiry time, the percentage of vehicles is significantly less. The proportion of safety messages being verified by one OBU depends on the total number of vehicles in the communication range of the RSU group as indicated in Fig. 5. It shows for three different speed levels of the traffic, only a portion of total delivered safety messages are received and verified by an OBU. Since our scheme considers rebroadcasting of a safety message until its expiry (tx ), an OBU doesn’t require to receive and verify all the safety messages in a particular session.

326

S. Biswas and J. Miˇsi´c

7 Conclusion and Future Direction We presented a new trust scheme for VANETs’ safety message delivery system, using ECDSA-based proxy signature. This mechanism would help a VANET user to authenticate the safety message, as well as the message originator along with the transmitting RSUs. In our scheme, an adversary can not forge an RSU, or a compromised RSU can not broadcast false, altered, and expired safety message. Thus, it makes the VANET safety message more reliable, trustworthy and acceptable to a user. The security analysis proved the strength of the scheme by analyzing the potential attack scenarios, while the simulation results support the usability of the scheme in the real-world VANETs. The approach has low communication overhead, which is compliant to the IEEE802.11p/WAVE standards as it uses the basic ECDSA signature scheme for the proxy signature in RSUs. Our future work includes study and research on security and trust management for vehicle to vehicle (V2V), and vehicle to infrastructure (V2I) safety message forwarding mechanisms with vehicle privacy. In order to make it energy efficient and bandwidth friendly, our research will have a focus on low power cryptographic primitives. For the experimental evaluation of our scheme, and comparison with its counterparts, we’ll be implementing a more realistic and dynamic VANET simulation that’ll provide a number of real-world traffic scenarios.

References 1. Draft amendment for wireless access in vehicular environments (WAVE). IEEE, New York, IEEE Draft 802.11p (July 2007) 2. Eichler, S.: Performance evaluation of the ieee 802.11p wave communication standard. In: IEEE 66th Vehicular Technology Conference, VTC-2007, 30 2007-October 3, pp. 2199– 2203 (Fall 2007) 3. Mambo, M., Usuda, K., Okamoto, E.: Proxy signatures for delegating signing operation. In: CCS 1996: Proceedings of the 3rd ACM Conference on Computer and Communications Security, pp. 48–57. ACM, New York (1996) 4. Johnson, D.B., Menezes, A.J.: Elliptic curve dsa (ecsda): an enhanced dsa. In: SSYM 1998: Proceedings of the 7th Conference on USENIX Security Symposium, p. 13. USENIX Association, Berkeley (1998) 5. IEEE trial-use standard for wireless access in vehicular environments (wave)- security services for applications and management messages. IEEE, New York (July 2006) IEEE Std 1609.2 6. Chang, M.-H., Chen, I.-T., Chen, M.-T.: Design of proxy signature in ecdsa. In: ISDA 2008: Proceedings of the 2008 Eighth International Conference on Intelligent Systems Design and Applications, pp. 17–22. IEEE Computer Society, Washington (2008) 7. Sun, X., Xia, M.: An improved proxy signature scheme based on elliptic curve cryptography. In: International Conference on Computer and Communications Security, pp. 88–91 (2009) 8. Qi, C., Wang, Y.: An improved proxy blind signature scheme based on factoring and ecdlp. In: International Conference on Computational Intelligence and Software Engineering, CiSE 2009, Wuhan, China, pp. 1–4 (2009) 9. Raya, M., Papadimitratos, P., Hubaux, J.-P.: Securing vehicular communications. Wireless Communications, IEEE 13(5), 8–15 (2006) 10. Guo, J., Baugh, J., Wang, S.: A group signature based secure and privacy-preserving vehicular communication framework, pp. 103–108 (May 2007)

Establishing Trust on VANET Safety Messages

327

11. Sha, K., Xi, Y., Shi, W., Schwiebert, L., Zhang, T.: Adaptive privacy-preserving authentication in vehicular networks. In: Proceedings of the First International Conference on Communications and Networking in China, ChinaCom 2006, pp. 1–8 (October 2006) 12. Calandriello, G., Papadimitratos, P., Hubaux, J.-P., Lioy, A.: Efficient and robust pseudonymous authentication in vanet. In: VANET 2007: Proceedings of the Fourth ACM International Workshop on Vehicular Ad hoc Networks, pp. 19–28. ACM, New York (2007) 13. Xi, Y., Sha, K., Shi, W., Schwiebert, L., Zhang, T.: Enforcing privacy using symmetric random key-set in vehicular networks, pp. 344–351 (March 2007) 14. Lin, X., Sun, X., Ho, P.-H., Shen, X.: Gsis: A secure and privacy-preserving protocol for vehicular communications. IEEE Transactions on Vehicular Technology 56(6), 3442–3456 (2007) 15. Barreto, P.S.L.M., Libert, B., McCullagh, N., Quisquater, J.-J.: Efficient and provably-secure identity-based signatures and signcryption from bilinear maps. In: Roy, B. (ed.) ASIACRYPT 2005. LNCS, vol. 3788, pp. 515–532. Springer, Heidelberg (2005) 16. Studer, A., Bai, F., Bellur, B., Perrig, A.: Flexible, extensible, and efficient vanet authentication. Journal of Communications and Networks 11(6), 574–588 (2009) 17. Perrig, A., Canetti, R., Tygar, J.D., Song, D.: The tesla broadcast authentication protocol. RSA CryptoBytes 5(2), 2–13 (2002) 18. Haas, J.J., Hu, Y.-C., Laberteaux, K.P.: Real-world vanet security protocol performance. In: IEEE GLOBECOM, pp. 1–7 (2009) 19. Wen, H., Ho, P.-H., Gong, G.: A novel framework for message authentication in vehicular communication networks. In: Global Telecommunications Conference, IEEE GLOBECOM 2009, pp. 1–6 (2009) 20. Wasef, A., Shen, X.: ASIC: Aggregate signatures and certificates verification scheme for vehicular networks. Engine (2009), https://129.97.58.88/ojs-2.2/index.php/pptvt/article/view/487 21. Kim, S., Park, S., Won, D.: Proxy signatures, revisited. In: ICICS 1997: Proceedings of the First International Conference on Information and Communication Security, pp. 223–232. Springer, London (1997) 22. Park, J.-H., Kim, Y.-S., Chang, J.H.: A proxy blind signature scheme with proxy revocation. In: International Conference on Computational Intelligence and Security Workshops, pp. 761–764 (2007) 23. Wei-min, L., Zong-kai, Y., Wen-qing, C.: A new id-based proxy blind signature scheme. Wuhan University Journal of Natural Sciences 10(3), 555–558 (2005) 24. Cai, M., Kang, L., Jia, J.: A multiple grade blind proxy signature scheme. In: International Conference on Intelligent Information Hiding and Multimedia Signal Processing, vol. 2, pp. 130–133 (2007) 25. Xue, Q., Li, F., Zhou, Y., Zhang, J., Cao, Z., Qian, H.: An ecdlp-based threshold proxy signature scheme using self-certified public key system. In: MobiSec 2009, pp. 58–70 (2009) 26. Johnson, D., Menezes, A.: The elliptic curve digital signature algorithm (ecdsa). Certicom Research, Canada; and Dept. of Combinatorics and Optimization, University of Waterloo, Canada, Tech. Rep. (1999)

Accelerating Signature-Based Broadcast Authentication for Wireless Sensor Networks Xinxin Fan and Guang Gong Department of Electrical and Computer Engineering University of Waterloo Waterloo, Ontario, N2L 3G1, Canada {x5fan@engmail,ggong@calliope}.uwaterloo.ca

Abstract. In wireless sensor networks (WSNs), broadcast authentication is a crucial security mechanism that allows a multitude of legitimate users to join in and disseminate messages into the networks in a dynamic and authenticated way. During the past few years, several public-key based multi-user broadcast authentication schemes have been proposed to achieve immediate authentication and to address the security vulnerability intrinsic to μTESLA-like schemes. Unfortunately, the relatively slow signature verification in signature-based broadcast authentication has also incurred a series of problems such as high energy consumption and long verification delay. In this contribution, we propose an efficient technique to accelerate the signature verification in WSNs through the cooperation among sensor nodes. By allowing some sensor nodes to release the intermediate computation results to their neighbors during the signature verification, a large number of sensor nodes can accelerate their signature verification process significantly. When applying our faster signature verification technique to the broadcast authentication in a 4 × 4 grid-based WSN, a quantitative performance analysis shows that our scheme needs 17.7% ∼ 34.5% less energy and runs about 50% faster than the traditional signature verification method. Keywords: Wireless Sensor Networks, Security, Broadcast Authentication, Elliptic Curve Cryptosystems.

1

Introduction

Wireless senor networks (WSNs) are typically composed of a few powerful base stations and a large number of resource-constrained sensor nodes [1]. In WSNs, multi-user broadcast is an efficient and common communication paradigm, in which a host of network users will join in WSNs and disseminate messages (i.e., queries or commands) into the networks dramatically for obtaining the information of their interest [12,18,19]. Unfortunately, due to the nature of wireless communication in WSNs, adversaries can easily eavesdrop the traffic, impersonate other users, inject bogus data or alter the contents of legitimate messages during the multi-hop forwarding. Hence, authentication mechanisms need to be implemented to protect broadcast messages from various malicious attacks. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 328–343, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Accelerating Broadcast Authentication for Wireless Sensor Networks

329

According to the cryptographic primitives employed, three categories of solutions have been proposed in literature for addressing broadcast authentication in WSNs. Earlier research mainly focused on designing symmetric-key based broadcast authentication schemes. Typical examples are μTESLA [17] and its variants [9,11,12], which provide source authentication and message integrity by utilizing one-way hash chains and delayed disclosure of authentication keys. μTESLA-like schemes provide efficient broadcast authentication mechanisms for WSNs in terms of computational overhead and energy consumption. However, the inherent features of μTESLA-like schemes, such as the need for (loose) time synchronization and the delayed authentication, have made them vulnerable to a variety of attacks [15,18,19]. Moreover, scalability is another concern for symmetric-key based solutions [18]. The second category of solutions achieve broadcast authentication through the use of one-time signatures [3,16]. Unlike μTESLA, one-time signature based solutions do not need the time synchronization and authentication is also immediate. Unfortunately, such schemes have some undesirable features such as large key sizes and a limited number of usages per key, which make them only suitable for applications with infrequent messages at unpredictable times [14]. Considering the security and scalability of symmetric-key based broadcast authentication schemes, a couple of public-key based solutions have been proposed during the past few years [4,18,19]. The main impetus for using public key cryptosystems in WSNs comes from the advances in the manufacturing technology of wireless sensor nodes as well as the efficient implementation of public key cryptographic algorithms on sensor platforms [10,20,21,22]. Employing publickey cryptography for implementing broadcast authentication in WSNs provides simple solutions, strong security resilience, good scalability and immediate message authentication, when compared to symmetric-key based solutions. However, public-key based broadcast authentication schemes have a common shortcoming: the signature verification is much slower than the message authentication code verification used in symmetric-key based solutions. As a result, a large number of packages might wait in a message queue of a senor node for signature verifications when many users broadcast messages. In this paper we address the issue of speeding up the signature verification for public-key based multi-user broadcast authentication schemes in WSNs by exploiting the cooperation among sensor nodes. The basic idea is that some sensor nodes randomly release their intermediate computation results to their neighbors during the signature verification. Then many sensor nodes can use the received intermediate computation results to accelerate their signature verifications. To demonstrate the performance of the proposed technique, we conduct a case study for broadcast authentication in a 4 × 4 grid-based WSN. The detailed quantitative analysis shows our scheme is greatly superior to the traditional signature verification method in terms of energy consumption of the entire network. The remainder of this paper is organized as follows: Section 2 gives a brief introduction about the elliptic curve digital signature algorithm (ECDSA). Section 3 presents the system model and adversary model. In Section 4, we

330

X. Fan and G. Gong

describe the acceleration technique for signature verification in WSNs and discuss the selection of system parameters. Section 5 analyzes the performance of the proposed scheme by a case study. Finally, Section 6 concludes this paper.

2

Elliptic Curve Digital Signature Algorithm (ECDSA)

Let Fq be a finite field with q = pm elements, where p > 3 is a prime and m is a positive integer. An elliptic curve E(Fq ) is the set of solutions (x, y) over Fq satisfying an equation of the form E : y 2 = x3 + ax + b, where a, b ∈ Fq and 4a3 + 27b2 ∈ F∗q , together with an additional point at infinity, denoted by O. The points on an elliptic curve form an (additive) Abelian group, where O is the identity element and the group operation is given by the well known chordand-tangent rule. The ECDSA is a widely standardized variant of the ElGamal signature scheme, which is described as follows: 1. System-wide parameters. Let G be a cyclic subgroup of E(Fq ) generated by the point P with prime order n and an identity element O. Let H : {0, 1}∗ → Z∗n be a collision-resistant hash function. 2. Initial set-up. Singer A randomly selects an integer d ∈ [1, n−1] and publishes its public key Q = dP . The parameter d is kept secret to A. 3. Signature generation. Singer A uses his/her private key d to generate a signature (r, s) for a message M ∈ {0, 1}∗. (a) Select a random integer k ∈ [1, n − 1], compute R = kP and set r to be the x-coordinate of R. (b) Compute s = k −1(e + dr) mod n, where e = H(M ). (c) If r, s ∈ [1, n − 1], return (r, s); otherwise, go to Step (3-a). 4. Signature verification. Upon receiving the message M ∈ {0, 1}∗ and the signature (r, s) from A, a verifier B verifies the signature using A’s public key Q. (a) Check that r, s ∈ [1, n − 1]. If any verification fails, return ‘reject signature’. (b) Compute R = s−1 (eP + rQ), where e = H(M ). (c) Check that the x-coordinate of R is equal to r. If verification succeeds, return ‘accept signature’; otherwise, return ‘reject signature’. Note that like most ElGamal signature schemes, the signature verification of ECDSA is about twice as slow as signature generation, which is an undesirable property when using ECDSA for multi-user broadcast authentication in WSNs.

3

System and Adversary Models

System Model : We consider a large-scale WSN consisting of a base station and a large number of sensor nodes. While the base station is powerful enough to execute various complicated operations, the sensor nodes are usually have constrained resources in terms of computational capabilities, memory, bandwidth,

Accelerating Broadcast Authentication for Wireless Sensor Networks

331

and power supply. We assume that users need to register and obtain necessary credentials for using the broadcast service in the WSN. We also assume that the base station is always trustworthy and the sensor nodes can be completely captured and manipulated by adversaries. Furthermore, both the base station and users may broadcast messages into the network. We further assume that a single-chip 2.4GHz IEEE 802.15.4 compliant RF transceiver [23] is used as the wireless transmission module in sensor nodes, which supports up to 102-byte payload and thus provides enough space to contain both the broadcast message and its digital signature in one package. Finally, we assume that the loose clock synchronization is available in the WSN. Adversary Model : We assume that an adversary can launch a wide range of attacks against the signature-based broadcast authentication schemes. For example, the attacker can simply mount the DoS attacks by injecting bogus messages into the network, aiming at exhausting the limited storage and energy of sensor nodes. We assume that some efficient pre-authentication techniques like those in [7,15] have been employed in the network to mitigate DoS attacks. Moreover, it is also possible that an adversary attempts to impersonate other legitimate sensor nodes and obtains valuable information through eavesdropping, modifying, deleting, replaying, forging or blocking any network traffic. We also assume that a small fraction of user devices and sensor nodes can be compromised by an adversary and therefore the attacker can manipulate compromised devices to disseminate messages into the network. Some efficient private-key protection mechanism and user revocation scheme [4,19] can be used to thwart the potential node compromise attack in WSNs.

4

Faster Signature Verification in WSNs

In this section, we first describe broadcast authentication in WSNs and then we use the ECDSA as an example to show how to accelerate the signature verification through the cooperation among sensor nodes. 4.1

Problem Statement

When WSNs are deployed in hostile environments, broadcast authentication (i.e., verify the authenticity of broadcast packages) is a crucial security mechanism to ensure the trustworthiness of network applications. After registration a user first contacts with several sensor nodes in the vicinity and sends a request for the broadcast service. Then the user and the sensor nodes conduct a mutual authentication procedure, which grants the access to the WSN only to a legitimate user and, at the same time, guarantees the user of the trustworthiness of the sensor nodes. As illustrated in Figure 1, once the user and the sensor nodes establish an authenticated channel, the user will sign a query or command and forward it to the sensor nodes (e.g., nodes A, B and C). Nodes A, B and C then verify the signature of the user, respectively. If the verification succeeds, they will locally broadcast the user’s query/command (within their communication range).

332

X. Fan and G. Gong

A

Qu ery EC /Co DS mm AS an ign d atu re

User

Qu ery /C om Sig ma na nd tur e

When some node, say D, receives the broadcast package for the first time, it will execute the same signature verification and determine whether the received package should be forwarded to other nodes (e.g., node E). This broadcast and authentication procedure continues until all reachable nodes receive the user’s broadcast package. If any verification fails during the broadcast, sensor nodes will drop the package and report to the base station.

Release l1 P

User B

C D

D E

A Release l 2Q B

E

C F G

Base Station

Fig. 1. User broadcast in wireless sensor networks. A broadcast package is usually forwarded multiple times through multi-hop communication.

4.2

Fig. 2. Faster ECDSA digital signature verification. Nodes A and B release l1 P and l2 Q, respectively, which will significantly accelerate the signature verification of nodes D to G.

A Faster Signature Verification Scheme

In Figure 1, all sensor nodes execute the same signature verification after receiving a broadcast package. Assuming that the ECDSA is employed in WSNs, we show how to speed up the ECDSA signature verification below. Note that the proposed acceleration technique is also applicable to other public-key based multi-user broadcast authentication schemes such as those in [4,18,19]. For verifying an ECDSA signature, each node needs to calculate R = s−1 (eP + rQ) = l1 P + l2 Q, where e = H(M ), l1 = s−1 e and l2 = s−1 r (see Section 2). In other words, two scalar multiplications l1 P and l2 Q have to be computed by each node. Our acceleration technique comes from the following key observation: all sensor nodes independently execute the same signature verification procedure during the broadcast authentication. Therefore, if some sensor nodes would like to consume their energy to release some intermediate results, the signature verification of their neighboring nodes can be accelerated significantly. Moreover, the energy consumption of the entire network will be decreased as well. Our idea is clearly illustrated in Figure 2. In Figure 2, a user’s broadcast package M, r, s will be received by nodes A, B and C, where M denotes a user’s query or command and (r, s) is the corresponding ECDSA signature of M . When all these three nodes finish the signature verification successfully, nodes A (the green node) and B (the yellow node) decide to release (i.e., locally broadcast) their intermediate computation results l1 P and l2 Q, respectively, whereas node C would like to keep silence.

Accelerating Broadcast Authentication for Wireless Sensor Networks

333

By this means, nodes D and E (the orange nodes), which are the neighbors of node A, can fast verify the digital signature by performing an elliptic curve point addition l1 P + l2 Q, where l2 Q is computed by nodes D and E themselves and l1 P comes from the contribution of node A. Moreover, nodes F and G can also perform fast signature verification in a similar way. Hence, if some node in WSNs releases its intermediate computation result, all its neighbors can fast verify the digital signature by calculating one scalar multiplication and one elliptic curve point addition, which can achieve about 50% performance improvement as compared to the traditional signature verification procedure. For the scenario described in Figure 2, the acceleration of signature verification on nodes D to G benefits from the release of the intermediate computation results from nodes A and B. Note that some sensor nodes might receive both intermediate computation results l1 P and l2 Q from their neighboring nodes. However, the sensor nodes cannot use both received l1 P and l2 Q to fast verify the signature with one elliptic curve point addition. The reason is that an adversary can capture a sensor node and easily launch the following attack: ; Step 1. The attacker generates a bogus message M Step 2. The attacker randomly chooses two integers l1 , l2 ∈ [1, n − 1] and calcu = l1 P + l2 Q; lates R  together with some random Step 3. The attacker takes the x-coordinate rˆ of R integer sˆ ∈ [1, n − 1] to form the fake signature pair (ˆ r , sˆ); , rˆ, sˆ, Step 4. The attacker successively releases two bogus broadcast packages M    l1 P  and M , rˆ, sˆ, l2 Q to its neighbors; Step 5. The victims compute l1 P + l2 Q and compare the x-coordinate of the  as a valid message. result with rˆ. As a result, the victims accept M Hence, if sensor nodes use two received intermediate computation results to verify a signature, they might accept any bogus broadcast messages from an attacker. To avoid the above attack, we only allow sensor nodes to use at most one intermediate result (i.e., l1 P or l2 Q) from their neighboring nodes for signature verification. Moreover, for the sake of simplicity of presentation, we further assume that if some sensor nodes release their intermediate computation results they will release l2 Q in the rest of this paper. We first illustrate a basic scheme of our faster ECDSA signature verification procedure in Figure 3. Let SCA and ADD denote the elliptic curve scalar multiplication and the elliptic curve point addition, respectively. In the basic scheme, a sensor node may receive a data package M, r, s or M, r, s, l2 Q. If a fresh package M, r, s is received, the sensor node will first compute l1 P and then wait for a very short time period τ to see whether it can obtain useful information from its neighbors for accelerating the signature verification. If it is, the node can finish the signature verification with 1 SCA + 1 ADD. Otherwise, the node will complete the verification itself with 2 SCA + 1 ADD after the time period τ . On the other hand, if a fresh package M, r, s, l2 Q is received, the sensor node will first calculate l1 P and then perform a fast signature verification with 1 SCA + 1 ADD. For the above two cases, if the signature is verified successfully, the sensor

334

X. Fan and G. Gong A Data Package

M, r, s

New?

M, r, s, l2 Q

Type?

No

No

Random Data

New?

Discard the package

Yes

Yes

Compute l1 P & Wait for τ seconds

Compute l1 P

Fast Signature Verification Path Does the package including l2 Q be received?

No

Yes Fast Signature Verification Path

Compute l2 Q

Compute l1 P + l2 Q and verifiy the signature

Successful?

No

Discard the package and report to the base station

Yes Release l2 Q with probability prel and forward the signed broadcast package

Fig. 3. Faster ECDSA Signature Verification for WSNs (Basic Scheme)

node will continue forwarding the broadcast package to its neighbors. Moreover, the intermediate computation result l2 Q will also be released with probability prel . Otherwise, if the signature verification is failed, the sensor node will send a signed report to the base station. Once the base station receives enough reports from the network, it will perform appropriate security mechanisms (see [24] for an example) to identify compromised nodes in WSNs. Although the basic scheme is simple and efficient, it is still vulnerable to the following attack: ; Step 1. The attacker generates a bogus message M Step 2. The attacker randomly chooses an integer k  ∈ [1, n − 1], computes R = k  P and sets r to be the x-coordinate of R ; Step 3. The attacker randomly chooses an integer s ∈ [1, n − 1] and computes ); l2 Q = R − s−1 e P , where e = H(M  and releases Step 4. The attacker uses (r , s ) as the signature of the message M     the bogus broadcast package M , r , s , l2 Q to its neighbors; Step 5. The victims compute l1 P + l2 Q and compare the x-coordinate of the  as a valid message. result with rˆ. As a result, the victims accept M The above attack works because the attacker knows that all its neighbors will first calculate l1 P = s−1 eP and then use the released value l2 Q to accelerate their signature verifications. Consequently, the attacker chooses random r , s and , and forges the broadcast package M , r , s , l Q that can the bogus message M 2 pass the signature verification. Note that if sensor nodes use traditional ECDSA signature verification procedure (see Section 2) the bogus broadcast package , r , s , l2 Q cannot pass the verification since the released l2 Q is equal to M

Accelerating Broadcast Authentication for Wireless Sensor Networks

335

Data Packages

Valid Data Package

Type?

Random Data Discard the package

No

New? Yes Wait for τ seconds & cache σ packages

Do σ packages have identical M , r, s and l2 Q?

Discard σ packages and report to the base station

No

Yes

Do σ packages include M, r, s, l2 Q?

No

Compute l1 P and l2 Q

Yes Compute l1 P Yes Fast Signature Verification Path

Compute l1 P + l2 Q and verifiy the signature

Successful?

No

Discard the package and report to the base station

Yes Release l2 Q with probability prel and forward the signed broadcast package

Fig. 4. Faster ECDSA Signature Verification for WSNs (Enhanced Scheme)

s−1 r Q with negligible probability. To thwart the attack in the basic scheme, we propose an enhanced scheme as shown in Figure 4, which takes advantage of the redundance of broadcast packages in WSNs. In the enhanced scheme, each sensor node first waits for τ seconds and caches σ data packages (i.e., M, r, s or M, r, s, l2 Q) received from its neighbors, where τ and σ are selected such that the sensor node can receive at least one data package from an honest neighbor (see Section 4.3). The sensor node then checks whether the cached σ data packages have identical M, r, s and l2 Q. Note that the main goal of adversaries is to broadcast fake data packages into WSNs. While all honest nodes forward a correct data package M, r, s or M, r, s, l2 Q to their neighboring nodes, adversaries try to deceive their neighbors by broadcasting a bogus data package M  , r , s , l2 Q as described in the basic scheme. Hence, if the senor node finds that the received data packages have different M, r, s or l2 Q, it will report the potential attack to the base station immediately. On the other hand, if all the cached σ data packages have identical M, r, s and l2 Q, the sensor node will further check whether it has received useful data packages M, r, s, l2 Q for accelerating signature verification. If it is, the sensor node will calculate l1 P and then complete the signature verification with 1 SCA + 1 ADD. Otherwise, the sensor node will perform the traditional ECDSA signature verification with

336

X. Fan and G. Gong

2 SCA + 1 ADD. The remaining steps after the signature verification are the same as those in the basic scheme. 4.3

Selection of System Parameters

In this subsection, we provide the guidance for selecting the system-wide parameters τ, σ and prel . Selection of the Delay τ and the Threshold σ. In the enhanced scheme, a sensor node needs to wait for τ seconds and cache σ data packages. We assume that on average a sensor node A has λ neighbors and half of them will broadcast data packages to A at a certain communication round1 . We further assume that among A’s λ/2 neighbors μ nodes can be compromised by adversaries and each of them can send at most ν bogus data packages to A during that communication round. Note that all compromised nodes must collude to send identical bogus data packages to A. Otherwise, A will discard all cached data packages and report to the base station. To thwart the collusive attacks from adversaries, the threshold σ should satisfy the following condition: λ/2 ≥ σ ≥ μ · ν + 1. The above condition guarantees that the sensor node A can receive at least one correct broadcast package from an honest neighbor. As a result, A will not accept the bogus messages from collusive adversaries since all the catched data packages are not identical. After the threshold σ is determined, we can choose the delay τ such that σ data packages can be received by the sensor node A. If a CC2420 IEEE 802.15.4 radio transceiver from Texas Instruments [23] is used in sensor nodes, the data transmission rate can achieve 250Kbps. We also assume that the signed broadcast package fits in the maximum allowable transmission limit (i.e., 128 bytes) of the CC2420 radio transceiver. So a 128-byte broadcast package will be transmitted in the physical layer and the transmission delay is about 4.096ms. Moreover, we also need to consider the CC2420 radio backoff that is a period of time where the radio pauses before attempting to transmit. Two backoff periods, namely initial backoff and congestion backoff, can be chosen for the CC2420 radio [23]. The initial backoff is 1 ∼ 32 backoff units2 (i.e., 300μs - 10ms), whereas the congestion backoff is 1 ∼ 8 backoff units (i.e., 300μs - 2.5ms). Therefore, in the worst case a 128-byte broadcast package can be received by the sensor node A within around 17ms. Taking into account all these factors, the delay τ should satisfy the following condition: τ ≥ 17 · 10−3 · σ. 1

2

The procedure of broadcast authentication can be divided into a series of communication rounds. Although a sensor node A has λ neighbors, only half of them will broadcast packages to A on average at certain communication round and the other half will receive the broadcast package from A in the following communication round. The units of backoff are 10 jiffies (32KHz ticks).

Accelerating Broadcast Authentication for Wireless Sensor Networks

337

Selection of the Release Probability prel . In the enhanced scheme, prel is the probability that a node will release its intermediate computation result, which is a predefined system parameter characterizing the trade-off between the verification speed of digital signature and the energy consumption of WSNs. Generally speaking, if a large value of prel is used, more sensor nodes will consume additional energy to broadcast their intermediate computation results and the signature verification for a large group of sensor nodes will be accelerated as a result, and vice versa. The selection of the release probability prel is closely related to the topology and the deployment of a WSN. Once the WSN is deployed, a network administrator first needs to analyze the network topology and estimates the average number of sensor nodes that might work on the signature verification during the transmission of a broadcast package. The administrator then determines a release probability prel that can achieve the optimal trade-off between the energy consumption of the network and the efficiency of the signature verification. More specifically, assuming that on average N sensor nodes work on the signature verification at each communication round, the probability that T sensor nodes will release their intermediate computation results is   N pT = · pTrel · (1 − prel )(N −T ) . T Let Es , Er , and Esca be the energy consumption of sending and receiving one packet, and calculating one elliptic curve scalar multiplication on sensor nodes, respectively. Recall that we assume that each node has λ neighbors on average. Then we can roughly estimate the additional energy consumption/saving due to the use of our fast signature verification technique as follows: – T sensor nodes will locally broadcast their intermediate computation results, the energy consumption of which is T · Es ; – About λT sensor nodes will receive the intermediate computation results, 2 the energy consumption of which is λT · Er ; 2 λT – About 2 sensor nodes will accelerate their signature verifications using the intermediate computation results, the energy saving of which is λT 2 · Esca . Therefore, the expected additional energy consumption/saving will be E=

N  T =1

  λT λT pT · T · Es + · Er − · Esca . 2 2

(1)

Note that the value in the round bracket of Equation (1) might be positive or negative, which depends on the energy consumption of the microcontroller and the radio component on different sensor motes. If the above value is positive, we need to choose a release probability prel that can minimize the energy consumption E. Otherwise, we select a prel that can maximize the energy saving E. Here, we only provide the guidance about the selection of the release probability prel and omit the details of specific applications.

338

5

X. Fan and G. Gong

Security and Performance Analysis

In this section, we analyze the performance of the proposed acceleration technique for signature verification with respect to communication and computation overheads (in terms of energy consumption). The performance analysis focuses on a simple 4 × 4 grid-based WSN. We also compare our scheme with the traditional ECDSA signature verification when applied to the broadcast authentication in the 4 × 4 grid-based WSN. 5.1

Case Study

User

Qu ery EC /Co DS mm AS ign and atu re

Note that the performance of our faster signature verification technique is closely related to the deployment of WSNs and the distribution of attackers in the network. To analyze the performance of our scheme, we conduct a case study for the broadcast authentication problem in a 4 × 4 grid-based WSN, as illustrated in Figure 5. In the grid-based sensor network, each node only can directly communicate with its one-hop neighbors. A user sends its signed broadcast package to the Node 1 at Round 0. After six communication rounds, the broadcast package will be received and verified by all sensor nodes. Furthermore, in our faster signature verification scheme we assume that one sensor node will release the intermediate computation result l2 Q in each communication round (see the green nodes 1, 2, 6, 7, 11 and 12 in Figure 5).

Round 0

1

Round 1

2

Round 2

Round 3

3

4 Round 4

5

6

7

8 Round 5

9

10

11

12 Round 6

13

14

15

16

Fig. 5. Broadcast Authentication in a 4 × 4 grid-based WSN

To give a detailed quantitative analysis, we further assume that MICAz motes are used in the WSN. Under a typical configuration such as a 3V supply voltage and a 7.37MHz clock frequency, the MICAz mote draws a current of 12mA in an active mode (i.e., CPU is operating) [6]. Based on the formulae of calculating the energy consumption on MICAz motes [8], we obtain the following basic facts:

Accelerating Broadcast Authentication for Wireless Sensor Networks

339

– A Chipcon CC2420 radio used in MICAz motes consumes Es = 83.6μJ and Er = 90.4μJ to transmit and receive l2 Q with 40 bytes3 , respectively; – An Atmega128L microcontroller used in MICAz motes consumes about Ever = 22.68mJ and Esca = 11.52mJ to verify an ECDSA signature and compute a scalar multiplication on a 163-bit elliptic curve. Note that we will not count the energy consumption of sensor nodes when sending and receiving the broadcast package throughout the whole network since it is the same for both faster and traditional signature verification schemes. We only compare the difference of both schemes in terms of communication and computation overhead in the next subsections. 5.2

Performance in the Ideal Case

In this subsection we analyze the performance of our faster signature verification technique in the ideal case (i.e., no adversaries exist), which can serve as the upper bound of performance improvement when applying our approach to broadcast authentication in the 4 × 4 grid-based WSN. In our scheme, some sensor nodes need to release their intermediate computation results in order to accelerate the signature verification for their neighboring nodes. Hence, our scheme consume more energy for transmitting the intermediate computation result l2 Q, when compared to using the traditional ECDSA signature verification in WSNs. In the 4 × 4 grid-based WSN (see Figure 5), the six green nodes (i.e., Nodes 1, 2, 6, 7, 11, 12) will locally broadcast their intermediate computation results to their one-hop neighbors, which causes an extra energy consumption of 6 × 83.6μJ = 501.6μJ in the network. Note that although some sensor nodes (e.g., Node 6) has four one-hop neighbors, only two of them (i.e., Nodes 7 and 10) will receive the intermediate computation results since the other two (i.e., Nodes 2 and 5) have finished the signature verification in the previous round (i.e., Round 1) and gone into the power-saving sleep mode. Therefore, there are totally 11 sensor nodes receiving the intermediate computation results, which causes an extra energy consumption of 11 × 90.4μJ = 994.4μJ in the WSN. In brief, our faster signature verification incurs an extra energy consumption of 501.6 + 994.4 = 1496μJ ≈ 1.5mJ for transmitting (i.e., sending and receiving) the intermediate computation results for the WSN in question. With respect to the computation aspect, it is not difficult to find that the signature verification on Nodes 2, 3, 5, 6, 7, 8, 10, 11, 12, 15 and 16 will be accelerated by 50% (i.e., saving one elliptic curve scalar multiplication) due to the use of the intermediate computation results from their neighboring nodes, which leads to a significant energy saving of 11 × 11.52mJ = 126.72mJ in the WSN, as compared to the traditional ECDSA signature verification technique. Moreover, the signed broadcast package will be forwarded to other sensor nodes only if the signature verification is successful. Therefore, the performance improvement on the signature verification will also reduce the transmission delay of the broadcast 3

Assuming that a 160-bit elliptic curve cryptosystem is employed in WSNs, the size of l2 Q is around 40 bytes.

340

X. Fan and G. Gong

package by 50% in each round accordingly. Consequently, the service quality of the broadcast authentication in WSNs has been improved remarkably by using our faster signature verification technique. To sum up, for the broadcast authentication in the target 4 × 4 grid-based WSN, our faster signature verification can save the energy consumption of 126.72mJ − 1.5mJ = 125.22mJ in total, considering both the communication and computation overheads. Therefore, using the proposed signature verification technique, one can save up to 125.22mJ × 100% = 34.5% 16 × 22.68mJ energy consumption for the grid-based WSN in question. 5.3

Security and Performance under Attacks from Independent Adversaries

In this subsection, we analyze the security and performance of our scheme when there exist independent adversaries in the 4 × 4 grid-based WSN. To this end, we assume that two independent adversaries (i.e., 12.5% nodes are compromised and become malicious nodes.) are deployed in the WSN and each of them will broadcast a bogus package to its neighbors. Note that in the grid-based WSN each bogus package will be received by two sensor nodes in the following communication round. To maximize the influence of independent adversaries, we select Node 3 and Node 9 as two independent adversaries in the 4 × 4 grid-based WSN. Moreover, for those nodes (i.e., Nodes 2, 3, 4, 5, 9, 11) that can only receive one data package from its neighbors in certain communication rounds, they will verify the signature themselves in order to avoid the attack descried in the basic scheme. Other nodes (i.e., Nodes 6, 7, 8, 10, 11, 12, 15, 16) will cache two received data packages from their neighbors and decide whether they can perform faster signature verifications. Like the ideal case, we also assume that the six green nodes (i.e., Nodes 1, 2, 6, 7, 11, 12) will locally broadcast the intermediate computation results l2 Q to their neighbors. Under the above settings, our faster signature verification still incurs an extra energy consumption of 1.5mJ for transmitting the intermediate computation results l2 Q like the ideal case. However, due to the existence of independent adversaries Node 3 and Node 9, Node 7 and Node 10 will receive two different data packages from their neighbors. Therefore, Node 7 and Node 10 have to verify the signature themselves instead of using the released l2 Q from Node 6 for faster verification. Consequently, the signature verification will be accelerated only for Nodes 6, 8, 11, 12, 15, 16 in this case, which can save the energy of 6 × 11.52mJ = 69.12mJ in the WSN. Combining both communication and computation overheads, one can obtain that in the case of two independent adversaries the total energy saving is 69.12 − 1.5mJ × 100% = 21.3% 14 × 22.68mJ

Accelerating Broadcast Authentication for Wireless Sensor Networks

341

for the 4 × 4 grid-based WSN. In addition, attacks from two independent adversaries have no effect on the security of our scheme and all the bogus data packages from independent adversaries are also discarded by legitimate nodes. 5.4

Security and Performance under Attacks from Collusive Adversaries

In this subsection, we analyze the security and performance of our scheme under the existence of collusive adversaries in the 4 × 4 grid-based WSN. Again, we assume that two collusive adversaries are deployed in the WSN and they will broadcast identical bogus data packages to their neighbors. To maximize the influence of collusive adversaries, we choose Node 7 and Node 10 as two collusive adversaries in the 4 × 4 grid-based WSN. Furthermore, we use the same assumptions as the case of independent adversaries for other nodes. Note that Node 11 will be cheated by collusive adversaries because of receiving two identical bogus data packages from Node 7 and Node 10. Although Node 11 continues broadcasting the bogus data package, Node 12 and Node 15 will discard it and verify the signature themselves because they receive two different data packages. As a result, bogus data packages from two collusive adversaries cannot be injected into the WSN successfully. Moreover, if Node 11 listens to the channel for one more communication round after broadcasting the bogus package, it is also possible for Node 11 to find the potential attacks itself. More specifically, after Node 12 and Node 15 verify the signature successfully at Round 5, they will broadcast the correct data package to their neighbors. Node 11 will find that all the data packages received from its neighbors (i.e., Nodes 7, 10, 12, 15) are different and therefore some attacks have happened. Like other cases, our faster signature verification needs to consume an extra energy of 1.5mJ for transmitting the intermediate computation results l2 Q. However, due to the existence of collusive adversaries Node 7 and Node 10, Node 11 will be fooled and then broadcast bogus data packages to Node 12 and Node 15. Note that both Node 12 and Node 15 will receive two different data packages from their neighbors and therefore verify the signature themselves. As a result, the signature verification will be accelerated only for Nodes 2, 3, 5, 6, 16 in this case, which can reduce the energy consumption of the WSN by 5×11.52mJ = 57.6mJ. Taking both communication and computation overheads into consideration, one can save around 57.6 − 1.5mJ × 100% = 17.7% 14 × 22.68mJ energy consumption for the 4 × 4 grid-based WSN. In particular, attacks from two collusive adversaries have very limited effect on the security of our scheme and those attacks can also be detected in the certain communication round.

6

Conclusions

Signature-based broadcast authentication schemes for WSNs have attracted a lot of attention in recent years due to their desirable features such as strong

342

X. Fan and G. Gong

security resilience, good scalability and immediate message authentication. However, the relatively slow signature verification in public-key cryptosystems causes high energy consumption and long verification delay for broadcast authentication in WSNs. In this chapter, we propose a novel and efficient acceleration technique for signature verification in WSNs. Our scheme fully exploits the cooperation among sensor nodes and enables the significant energy consumption saving for the whole network. As a case study, we apply our technique to the broadcast authentication in a 4 × 4 grid-based WSN and analyze the security and performance of our scheme under the existence of independent and collusive adversaries. While independent adversaries do not have any influence on the security of our scheme, collusive adversaries have very limited effect. Particularly, bogus data packages from adversaries cannot be disseminated successfully through the entire network in both cases. Moreover, a quantitative performance analysis shows that our scheme can save about 17.7% ∼ 34.5% energy consumption and run 50% faster than traditional signature verification method.

References 1. Akyildiz, L., Su, W., Sankarasubramaniam, Y., Cayirci, E.: A Survey on Sensor Networks. IEEE Communications Magazine 40(8), 102–116 (2002) 2. Bellare, M., Namprempre, C., Neven, G.: Security Proofs for Identity-Based Identification and Signature Schemes. In: Cachin, C., Camenisch, J.L. (eds.) EUROCRYPT 2004. LNCS, vol. 3027, pp. 268–286. Springer, Heidelberg (2004) 3. Chang, S., Shieh, S., Lin, W., Hsieh, C.-M.: An Efficient Broadcast Authentication Scheme. In: Proceedings of the 2006 ACM Symposium on Information, Computer and Communications Security (ASIACCS 2006), pp. 311–320 (2006) 4. Cao, X., Kou, W., Dang, L., Zhao, B.: IMBAS: Identity-Based Multi-User Broadcast Authentication in Wireless Sensor Networks. Computer Communications 31(4), 659–667 (2008) 5. Courtois, N., Finiasz, M., Sendrier, N.: How to Achieve a McEliece-Based Digital Signature Scheme. In: Boyd, C. (ed.) ASIACRYPT 2001. LNCS, vol. 2248, pp. 157–174. Springer, Heidelberg (2001) 6. Crossbow Technology Inc.: MICAz – Wireless Measurement System, http://www. xbow.com/Products/Product_pdf_files/Wireless_pdf/MICAz_Datasheet.pdf 7. Dong, Q., Liu, D., Ning, P.: Pre-Authentication Filters: Providing DoS Resistance for Signature-Based Broadcast Authentication in Wireless Sensor Networks. In: Proceedings of the First ACM Conference on Wireless Network Security (WiSec 2008), pp. 2–12 (2008) 8. Driessen, B., Poschmann, A., Paar, C.: Comparison of Innovative Signature Algorithms for WSNs. In: Proceedings of the First ACM Conference on Wireless Network Security (WiSec 2008), pp. 30–35 (2008) 9. Drissi, J., Gu, Q.: Localized Broadcast Authentication in Large Sensor Networks. In: Proceedings of the International Conference on Networking and Services (ICNS 2006), pp. 25–31 (2006) 10. Liu, A., Ning, P.: TinyECC: A Configurable Library for Elliptic Curve Cryptography in Wireless Sensor Networks. In: Proceedings of the 7th International Conference on Information Processing in Sensor Networks (IPSN 2008), SPOTS Track, pp. 245–256 (2008)

Accelerating Broadcast Authentication for Wireless Sensor Networks

343

11. Liu, D., Ning, P.: Multi-Level μTESLA: Broadcast Authentication for Distributed Sensor Networks. ACM Transactions on Embedded Computing Systems 3(4), 800– 836 (2004) 12. Liu, D., Ning, P., Zhu, S., Jajodia, S.: Practical Broadcast Authentication in Sensor Networks. In: Proceedings of the Second Annual International Conference on Mobile and Ubiquitous Systems: Networking and Services (MobiQuitous 2005), pp. 118–129 (2005) 13. L´ opez, J., Aranha, D., Cˆ amara, D., Dahab, R., Oliveira, L., Lopes, C.: Fast Implementation of Elliptic Curve Cryptography and Pairing Computation for Sensor Networks. In: The 13th Workshop on Elliptic Curve Cryptography (ECC 2009), http://ecc.math.ucalgary.ca/sites/ecc.math.ucalgary.ca/files/u5/ Lopez_ECC2009.pdf 14. Luk, M., Perrig, A., Whillock, B.: Seven Cardinal Properties of Sensor Network Broadcast Authentication. In: Proceedings of the Fourth ACM workshop on Security of Ad Hoc and Sensor Networks (SASN 2006), pp. 147–156 (2006) 15. Ning, P., Liu, A., Du, W.: Mitigate DOS Attacks Against Broadcast Authentication in Wireless Sensor Networks. ACM Transactions on Sensor Networks 4(1), 1:1–1:35 (2008) 16. Perrig, A.: The BiBa One-Time Signature and Broadcast Authentication Protocol. In: Proceedings of the 8th ACM conference on Computer and Communications Security (CCS 2001), pp. 28–37 (2001) 17. Perrig, A., Szewczyk, R., Tygar, J.D., Wen, V., Culler, D.E.: SPINS: Security Protocols for Sensor Networks. ACM Wireless Networks 8(5), 521–534 (2002) 18. Ren, K., Lou, W., Zeng, K., Moran, P.J.: On Broadcast Authentication in Wireless Sensor Networks. IEEE Transactions on Wireless Communications 6(11), 4136– 4144 (2007) 19. Ren, K., Yu, S., Lou, W., Zhang, Y.: Multi-user Broadcast Authentication in Wireless Sensor Networks. To appear in IEEE Transactions on Vehicular Technology (2009) 20. Seo, S.C., Han, D.-G., Song, S.: TinyECCK: Efficient Elliptic Curve Cryptography Implemenation over GF (2m ) on 8-bit Micaz Mote. IEICE Transactions on Information and Systems E91-D(5), 1338–1347 (2008) 21. Shirase, M., Miyazaki, Y., Takagi, T., Han, D.-G., Choi, D.: Efficient Implementation of Pairing Based Cryptography on a Sensor Node. IEICE Transactions on Information and Systems E92-D(5), 909–917 (2009) 22. Szczechowiak, P., Kargl, A., Scott, M., Collier, M.: On the Application of Pairing Based Cryptography to Wireless Sensor Networks. In: Proceedings of the Second ACM Conference on Wireless Network Security (WiSec 2009), pp. 1–12 (2009) 23. Texas Instrument Inc. 2.4 GHz IEEE 802.15.4/ZigBee-ready RF Transceiver, http://focus.ti.com/lit/ds/symlink/cc2420.pdf 24. Zhang, Q., Yu, T., Ning, P.: A Framework for Identifying Compromised Nodes in Wireless Sensor Networks. ACM Transactions in Information and Systems Security (TISSEC) 11(3), 1–37 (2008)

Secure Data Aggregation in Wireless Sensor Networks: Homomorphism versus Watermarking Approach Jacques M. Bahi, Christophe Guyeux, and Abdallah Makhoul Computer Science Laboratory (LIFC), University of Franche-Comt´e Rue Engel-Gros, BP 527, 90016 Belfort Cedex, France {jacques.bahi,christophe.guyeux,abdallah.makhoul}@univ-fcomte.fr

Abstract. Wireless sensor networks are now in widespread use to monitor regions, detect events and acquire information. Since the deployed nodes are separated, they need to cooperatively communicate sensed data to the base station. Hence, transmissions are a very energy-consuming operation. To reduce the amount of sending data, an aggregation approach can be applied along the path from sensors to the sink. However, usually the carried information contains confidential data. Therefore, an end-to-end secure aggregation approach is required to ensure a healthy data reception. End-to-end encryption schemes that support operations over cypher-text have been proved important for private party sensor network implementations. These schemes offer two main advantages: end-to-end concealment of data and ability to operate on cipher text, then no more decryption is required for aggregation. Unfortunately, nowadays these methods are very complex and not suitable for sensor nodes having limited resources. In this paper, we propose a secure end-to-end encrypted-data aggregation scheme. It is based on elliptic curve cryptography that exploits a smaller key size. Additionally, it allows the use of higher number of operations on cypher-texts and prevents the distinction between two identical texts from their cryptograms. These properties permit to our approach to achieve higher security levels than existing cryptosystems in sensor networks. Our experiments show that our proposed secure aggregation method significantly reduces computation and communication overhead and can be practically implemented in on-the-shelf sensor platforms. By using homomorphic encryption on elliptic curves, we thus have realized an efficient and secure data aggregation in sensor networks. Lastly, to enlarge the aggregation functions that can be used in a secure wireless sensor network, a watermarking-based authentication scheme is finally proposed. Keywords: wireless sensor networks; secure data aggregation; authentication; homomorphic encryption; elliptic curves; watermarking.

1 Introduction Wireless sensor networks have received enormous attention over past few years, due to a wide range of potential applications (environmental, ecological, military, etc.). A typical sensor network is expected to consist of a large number of sensor nodes deployed randomly in a large scale. Usually, these nodes have limited power, storage, communication, and processing capabilities, making energy consumption an issue. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 344–358, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Secure Data Aggregation in Wireless Sensor Networks

345

A major functionality of a sensor node is to measure environmental values using embedded sensors, and transmit it to a base station called “sink”. The sensed data needs to be analyzed, which eventually serves to initiate some action. Almost this analysis presumes computation of the maximum, minimum, average, etc. It can be either done at the base station or by the nodes themselves, in a hierarchical scenario. In order to reduce the amount of data to be transmitted to the sink, it is beneficial that this analysis can be done over the network itself. To save the overall energy resources of the network, it is agreed that the sensed data needs to be aggregated on the way to its final destination. Sensor nodes send their values to certain special nodes, i.e., aggregators. Each aggregator then condenses the data prior to sending it on. In terms of bandwidth and energy consumption, aggregation is beneficial as long as the aggregation process is not too CPU-intensive. The aggregators can either be special (more powerful) nodes or regular sensors nodes. At the same time, sensor networks are often deployed in public or otherwise untrusted and even hostile environments, which prompts a number of security issues (e.g., key management, privacy, access control, authentication, etc.). Then, if security is a necessary in other (e.g., wired or MANET) types of networks, it is much more so in sensor networks. Actually, it is one of the more popular research topic and many advances have been reported on in recent years. From the above observations, we can notice the importance of a cooperative secure data aggregation in sensor networks. In other terms, after the data gathering and during transmissions to the base station, each node along the routing path cooperatively integrates and secures the fragments messages. In this paper, we focus on security data aggregation and we propose a simple secure homomorphic cypher-system that allows efficient aggregation of encrypted data. Data encryption becomes necessary in sensor networks when this type of sensors can be subject of many types of attacks [1]. Without encryption, adversaries can monitor and inject false data into the network. In a general manner the encryption process is done as follows: sensor nodes must encrypt data on a hop-by-hop basis. An intermediate node (i.e., aggregator) possessing the keys of all sending nodes, decrypts the received encrypted value, aggregates all received values, and encrypts the result for transmission to the base station. Though viable, this approach is fairly expensive and complicated, due to the fact of decrypting each received value before aggregation, which generates an overhead imposed by key management. Encryption can solve the security problem, but how can we aggregate over encrypted data [1]? Some privacy homomorphism based works have been proposed recently [2,3,4] that, without participating in checking, the aggregators can directly aggregate the encrypted data. However, such schemes, for the moment, need high and complex computations to encrypt data and aggregate it, which leads to large cypher-texts. Sensor nodes cannot provide sufficient CPU, memory and bandwidth to address such complex operations. For instance, Rivest Shamir Adleman (RSA) cryptosystems [5,6] are used, which requires high CPU and memory capabilities to perform exponential operations. Therefore, in our study we adopt an elliptic curve encryption [7] that allows nodes to generate a smaller key size while providing the same security level of existing complex schemes. The cypher-system we exploit permits N additions and one product, thus it is not

346

J.M. Bahi, C. Guyeux, and A. Makhoul

limited to a single basic function. A major advantage of our method is the fact that it has been proved safe, and until now it has not been cryptanalized. To assess the practicality of our technique, we evaluate it and compare it to existing cypher-system. The obtained results show that we significantly reduce computation and communication overhead and that our secure aggregation method can be practically implemented in on-the-shelf sensor platforms. The rest of this paper is organized as follows: in the next section we present a review of some previous related work. Section 3 presents our security model. Sections 4 and 5 discuss the details of the proposed aggregation scheme for sensor networks. In Section 6, we describe simulation and results of simulation experiments. In Section 7 is proposed a new authentication scheme based on a watermarking approach, to improve the variety of aggregation functions through the secure wireless sensor network. Finally, we end the paper by a conclusion.

2 Related Work The benefit and vulnerability, as well as the need to secure in-network aggregation, have been identified by a number of schemes in the literature. One approach [8] proposed a secure information aggregation protocol to answer queries over the data acquired by the sensors. Even though their method provided data authentication to provide secrecy, the data still sent in plain text format, which removes the privacy during transmission. Another one [9] proposed a secure energy efficient data aggregation (ESPDA) to prevent redundant data transmission in data aggregation. Unlike conventional techniques, their scheme prevents the redundant transmission from sensor motes to the aggregator. Before transmitting sensed data, each sensor transmits a secure pattern to the aggregator. Only sensors with different data are allowed to transmit their data to the cluster-head. However, since each sensor at least needs to transmit a packet containing a pattern once, power cannot be significantly saved. In addition, each sensor mote uses a fixed encryption key to encrypt data; data privacy cannot be maintained in their scheme. In [10], the authors presented a secure encrypted-data aggregation scheme for wireless sensor networks. The idea is based on eliminating redundant sensor readings without using encryption and maintains data secrecy and privacy during transmission. This scheme saves energy on sensor nodes but still do not guarantee the privacy of sent data. The problem of aggregating encrypted data in sensor networks was introduced in [3] and further refined in [2]. The authors propose to use homomorphic encryption schemes to enable arithmetic operations over cypher-texts that need to be transmitted in a multihop manner. However, these approaches provide a higher level of system security, since nodes would not be equipped with private keys, which would limit the advantage gained by an attacker compromising some of the nodes. Unfortunately, existing privacy homomorphisms used for data aggregation in sensor networks have exponential bound in computation. It is too computationally expensive to implement in sensor nodes. Moreover, the expansion in bit size during the transformation of plain text to cypher-text introduces costly communication overhead, which directly translates to a faster depletion of the sensors energy. On the other hand and from security viewpoint, the cryptosystems [11]

Secure Data Aggregation in Wireless Sensor Networks

347

used in these approaches were cryptanalized [12,13], which means they can’t guarantee anymore high security levels. In this paper we try to relax the statements above by investigating elliptic curve cryptography that allows feasible and suitable data aggregation in sensor networks beside the security of homomorphisms schemes. First, our proposed scheme for secure data aggregation in sensor networks is based on a cryptosystem, which has been proved safe and has not been cryptanalyzed. Another property that enforces the security level of such approach is coming from the fact that, as it is the case in ElGamal cryptosystem, for two identical messages it generates two different cryptograms. This property suggested fundamental for security in sensor networks [7,10,14], to the best of our knowledge, was not addressed in previous homomorphism-based security data aggregation works. Beside all these properties and due to the use of elliptic curves, our approach saves energy by allowing nodes to encrypt and aggregate data without the need of high computations. Lastly, the scheme we use allows more aggregations types over cypher data than the homomorphic cryptosystem used until now.

3 Security Model In this work, we are primarily concerned with data privacy in sensor networks. Our goal is to prevent attackers from gaining any information about sensor data. Therefore, ensuring an end-to-end privacy between sensor nodes and the sink becomes problematic. This is largely because popular and existing cyphers are not additively homomorphic. In other words, the summation of encrypted data does not allow for the retrieval of the sum of the plain text values. Moreover, privacy existing homomorphisms have usually exponential bound in computation. To overcome this problem, in our model we propose a security scheme for sensor networks using elliptic curve based cryptosystem. We show that our model permits many operations on crypted data and does not demand high sensor capabilities and computation. 3.1 Operations over Elliptic Curves In this section, we give a brief introduction to elliptic curve cryptography. The reader is referred to [15] for more details. Addition and Multiplication. Elliptic curve cryptography (ECC) is an approach to public-key cryptography based on the algebraic structure of elliptic curves over finite fields [15]. Elliptic curves used in cryptography are typically defined over two types of finite fields: prime fields Fp , where p is a large prime number, and binary extension fields F2m [16]. In our paper, we focus on elliptic curves over Fp . Let p > 3, then an elliptic curve over Fp is defined by a cubic equation y2 = x3 + ax + b as the set   E = (x, y) ∈ Fp × Fp , y2 ≡ x3 + ax + b (mod p) where a, b ∈ Fp are constants such that 4a3 + 27b2  0 (mod p). An elliptic curve over Fp consists of the set of all pairs of affine coordinates (x, y) for x, y ∈ Fp that satisfy an equation of the above form and an infinity point O.

348

J.M. Bahi, C. Guyeux, and A. Makhoul

The point addition and its special case, point doubling over E is defined as follows (the arithmetic operations are defined in Fp ) [15] : Let P = (x1 , y1 ) and Q = (x2 , y2 ) be two points of E. Then:  O if x2 = x1 and y2 = −y1 , P+Q= (x3 , y3 ) otherwise. where: – x3 = λ2 − x1 − x2 , – y3 = λ × (x1 − x3 ) − y1 ,  (y2 − y1 ) × (x2 − x1 )−1 λ= (3x21 + a) × (2y1 )−1

if if

PQ, P = Q.

Finally, we define P + O = O + P = P, ∀P ∈ E, which leads to an abelian group (E, +). On the other hand the multiplication n × P means P + P + .... + P n times and −P is the symmetric of P for the group law + defined above for all P ∈ E. Public/Private Keys Generation with ECC. In this section we show how we can generate the public and private keys for encryption, following the cryptosystem proposed by Boneh et al. [7]. The analysis of the complexity will be treated in a later section. Let τ > 0 be an integer called “security parameter”. To generate public and private keys, first of all, two τ-bits prime numbers must be computed. Therefore, a cryptographic pseudo-random generator can be used to obtain two vectors of τ bits, q1 and q2 . Then, a Miller-Rabin test can be applied for testing the primality or not of q1 and q2 . We denote by n the product of q1 and q2 , n = q1 q2 , and by l the smallest positive integer such that p = l × n − 1. l is a prime number while p = 2 (mod 3). In order to find the private and public keys, we define a group H, which presents the points of the super-singular elliptic curve y2 = x3 + 1 defined over Fp . It consists of p + 1 = n × l points, and thus has a subgroup of order n, we call it G. In another step, we compute g and u as two generators of G and h = q2 × u. Then, following [7], the public key will be presented by (n, G, g, h) and the private key by q1 . Encryption and Decryption. After the private/public keys generation, we proceed now to the two encryption and decryption phases: – Encryption: Assuming that our messages space consists of integers in the set {0, 1, ..., T}, where T < q2 , and m the (integer) message to encrypt. First, a random positive integer is picked from the interval [0, n − 1]. Then, the cypher-text is defined by C = m × g + r × h ∈ G, in which + and × refer to the additive and multiplication laws defined previously. – Decryption: Once the message C arrived to destination, to decrypt it, we use the private key q1 and the discrete logarithm of (q1 × C) base q1 × g as follows: m = logq1 ×g q1 × C.

Secure Data Aggregation in Wireless Sensor Networks

349

√ This takes expected time T using Pollard’s lambda method. Moreover, this decryption can be speed-up by precomputing a table of powers of q1 × g. 3.2 Homomorphic Properties As we mentioned before, our approach ensures easy encryption/decryption without any need of extra resources. This will be proved in the next section. Moreover, our approach supports homomorphic properties, which gives us the ability to execute operations on values even though they have been encrypted. Indeed, it allows N additions and one multiplication directly on cryptograms, which prevents the decryption phase at the aggregators level and saves nodes energy, which is crucial in sensor networks. Additions over cypher-texts are done as follows: let m1 and m2 be two messages and C1 , C2 their cypher-texts respectively. Then the sum of C1 and C2 , let call it C, is represented by C = C1 + C2 + r × h where r is an integer randomly chosen in [0, n − 1] and h = q2 × u as presented in the previous section. This sum operation guarantees that the decryption value of C is the sum m1 + m2 . The addition operation can be done several times, which means we can do sums of encrypted sums. The multiplication of two encrypted values and its decryption are done as follows: let e be the modified Weil pairing on the curve and g, h the points of G as defined previously. Let us recall that this modified Weil pairing e is obtained from the Weil pairing E [7], [17] by the formula: e(P, Q) = E(x × P, Q), where x is a root of X3 − 1 on Fp2 . Then, the result of the multiplication of two encrypted messages C1 , C2 is given by [Cm = e(C1 , C2 ) + r × h1 ], where h1 = e(g, h) and r is a random integer pick in [1, n]. The decryption of Cm is equal to the discrete logarithm of q1 ×Cm to the base q1 × g1 : m1 m2 = logq1 ∗g1 (q1 × Cm .) where g1 = e(g, g).

4 Our Secure Data Aggregation for Sensor Networks 4.1 Presentation Data aggregation schemes aim to combine and summarize data packets of several sensor nodes so that amount of data transmission is reduced. An example data aggregation scheme is presented in Figure 1 where sensor nodes collect information from a region of interest. When the user (sink) queries the network, instead of sending each sensor node’s data to the base station, aggregators collect the information from its neighboring nodes, aggregates them, and send the aggregated data to the base station over a multihop path. As the majority of wireless sensor network applications require a certain level of security, encryption of the sensed data before its transmission becomes necessary and it is preferable to decrypt the data only at the base station level (c.f. previous sections). In our work, we adopt the following scenario as shown in Figure 1: after collecting information, each sensor node encrypts its data according to elliptic curve encryption (c.f. Section 3.1) and sends it to the nearest aggregator. Then, aggregators aggregate the received encrypted data (without decryption) and send it to the base station, which in

350

J.M. Bahi, C. Guyeux, and A. Makhoul Decryption & Aggregation Sink (base station)

Aggregation over cypher−text Aggregators

Aggregation over cypher−text Aggregators

Collecting data & Encryption

Normal Sensors

Fig. 1. Secure data aggregation in sensor networks

his turn decrypts the data and aggregates it. We notice that all aggregators can done N additions and the final layer of aggregators can done one multiplication on encrypted data. 4.2 Example of Use Computing the Arithmetic Mean. The arithmetic mean is the “standard” average, often simply called the “mean”, defined for n values x1 , . . . , xn by x¯ =

n 1  · xi . n i=1

To compute the average of nodes measurements, aggregators can calculate the sum of the encrypted measurements and the number of nodes took these measurements and send it to the base station. More precisely, when using our scheme, each sensor encrypts its data xi to obtain cxi. The sensor then forwards cxi to its parent, who aggregates all the cx j’s of its k children by simply adding  them up. The resulting value is then forwarded. The sink ends up with value Cx = ni=1 cxi . It can then decrypt Cx, and divide the result by n to derive the average. Computing the Variance. Another common aggregation is to estimate the variance of the sensed values. Let us recall that the variance of n values x1 , ..., xn is defined by: ⎛ n ⎞ n  ⎜⎜ 1  ⎟⎟ 1 2 2 (xi − x) = ⎜⎜⎜⎝ s2n = x2i ⎟⎟⎟⎠ − x . n n i=1

i=1

Our scheme can also be used to derive the variance of the measured and encrypted data, by the same method as in [18]. In this case, each sensor i must compute yi = x2i , where xi is the measured sample, and encrypts yi to obtain cyi . xi must also be encrypted, as explained in the previous section. The sensor forwards cyi , together with cxi , to its parent. The parent aggregates all the cy j of its k children by simply adding them up.

Secure Data Aggregation in Wireless Sensor Networks

351

It also aggregates, separately, the cx j, as explained in the previous section. n The two resulting values are then forwarded. The sink ends up with values Cx = i=1 cxi and  Cy = ni=1 cyi . Cx is used to compute the average Av, when Cy is used to compute the Vy variance as follows: Var = n − Av2 , where Vy is the decryption of Cy. Computing the Weighted Mean. The weighted mean of a non-empty set of data x1 , x2 , . . . , xn with non-negative weights w1 , w2 , . . . , wn , is the quantity x¯ =

w1 x 1 + w2 x 2 + · · · + wn x n . w1 + w2 + · · · + wn

We suppose now that each aggregator i of the first aggregation layer has computed the mean xi of the encrypted values received from its sensor node. Additionally, we suppose that these aggregators are weighted, depending on their importance. For security reasons, this weight is also encrypted and the cypher value is denoted by wi . This wi can be proportional to the number of aggregated sensors. This weight can also illustrate the fact that two given regions have not the same relevance. To achieve weighted mean, each aggregator multiplies its encrypted mean xi with encrypted weight wi as it has been explained previously. The resulting value is then forwarded to the sink, which can decrypt wi × xi and sum all these decrypted values, to obtain the weighted mean defined above.

5 Security Study Due to hostile environments and unique characteristics of sensor networks, it is a challenging task to protect sensitive information transmitted by nodes to the end user. In addition, this type of networks has security problems that traditional networks do not face. In this section, we present a security study dedicated to wireless sensor networks. First we introduce the principal attacks that sensor networks can face and how our approach can support them, then we present some practical issues that improve the network security. 5.1 Related Attacks and Results In a sensor network environment adversaries can commonly use the following attacks: Known-plain text attack: They can use common key encryption to see when two readings are identical. By using nearby sensors under control, attackers can conduct a known-plain text attack. Chosen-plain text attack: Attackers can tamper with sensors to force them to predeterminated values. Man-in-the-middle: They can inject false readings or resend logged readings from legitimate sensor motes to manipulate the data aggregation process. In Tables 1, 2 and similar to [16], we present a comparison between different encryption policies and possible attacks. In our method, as data are encrypted by public keys, and these public keys are sent by the sink to the sole authenticated motes, the wireless sensor

352

J.M. Bahi, C. Guyeux, and A. Makhoul

Table 1. Encryption polices and vulnerabilities Encryption Policy

Possible attacks

Sensors transmit readings without encryption Man-in-the-middle Sensors transmit encrypted readings

Known-plain text attack

with permanent keys

Chosen-plain text attack

Sensors transmit encrypted readings

None of above

Man-in-the-middle with dynamic keys Our scheme

None of above Table 2. Encryption polices and aggregation

Encryption Policy Sensors transmit readings without encryption Sensors transmit encrypted readings with permanent keys

Data aggregation Generating wrong aggregated results Data aggregation is impossible, unless the aggregator has encryption keys

Sensors transmit encrypted readings with dynamic keys Our scheme

Data aggregation cannot be achieved unless the aggregator has encryption keys Data aggregation can be achieved

network is then not vulnerable to a Man-in-the-middle attacks. On the other hand, our approach guarantees that for two similar texts gives two different cryptograms, which prevents the Chosen-plain text attacks and the Man-in-the-middle attacks. Finally, as the proposed scheme possesses the homomorphic property, data aggregation is done without decryption, and no private key is used in the network. 5.2 Practical Issues In this section we present some practical issues to our security model. First we study the sizes of the encryption keys and we compare it to existing approaches. Then, we present how we can optimize the sizes of cryptograms in order to save more sensors energy. Sizes of the Keys. Cryptograms are points of the elliptic curve E. They are constituted by couples of integer coordinates lesser than or equal to p = lq1 q2 − 1. It is commonly accepted [19], [20] that for being secure until 2020, a cryptosystem: – must have p ≈ 2161 , for EC systems over Fp , – must satisfy p ≈ 21881 for classical asymmetric systems, such as RSA or ElGamal on Fp . Thus, for the same level of security, using elliptic curve cryptography does not demand high keys sizes, contrary to the case of RSA or ElGamal on Fp . The use of small keys leads to small cryptograms and fast operations for encryption.

Secure Data Aggregation in Wireless Sensor Networks

353

Reducing the Size of Cryptograms. In this section we show how we can reduce the size of cryptograms while using ECC. This is benefit for sensor nodes in terms of reducing energy consumption by sending data with smaller size. The messages are encrypted with q2 bits, which leads to cryptograms with a mean of 160 bits long. Let us suppose that p ≡ 3 ( mod 4). As the cryptogram is an element (x, y) of E, which is defined by y2 = x3 + 1, we can compress this cryptogram (x, y) to (x, y mod 2)) before sending it to the aggregator (as the value of y2 is known). In this situation, we obtain cryptograms with a mean of 81 bits long for messages between 20 and 40 bits long. To decompress the cryptogram (x, i), the aggregator must compute z = x3 + 1 mod p √ and y = z mod p, which can be written as y = z(p+1)/4 mod p, then: – if y ≡ i(mod 2), then the decompression of (x, i) is (x, y). – else the decompression is (x, p − y).

6 Experimental Results To show the effectiveness of our approach we conducted a series of simulations comparing our method to another existing one based on RSA cryptosystem. We considered a network formed of 500 sensor nodes, each one is equipped by a battery of 100 units capacity. We consider that the energy consumption  E of a node is proportional to the computational time t, i.e., E = kt. The same coefficient of proportionality k is taken while comparing the two encryption scenarii. Sensor nodes are then connected to 50 aggregators chosen randomly. Each sensor node choose the nearest aggregator. The running of each simulation is as follows: each sensor node takes a random value, encrypts it using one of the encryption methods then sends it to its aggregator. Aggregators compute the sum of the encrypted received data and send it to the sink. We compared our approach to the known RSA public-key cryptographic algorithms, and we evaluated the energy consumption of the network while varying the sizes of the keys and obviously the security levels. The energy consumption is the units of the battery used to do the encryption. Tables 3 and 4 show the energy consumption of sensor nodes to do the encryption operations using our encryption method and the RSA one respectively. We varied the keys sizes and obviously the security levels. We notice that for the same level of security in our approach we used small keys while saving more energy. For instance, for high security levels (4 for example) a node using our approach needs to use a key of 167 bits instead of 1891 in the case of RSA and consumes 0.1 % of the battery power instead of 3.63 %. Table 3. Our approach Security level Size p of the key E (battery units) 1 46 0.02 2 85 0.05% 3 125 0.07 4 167 0.10

354

J.M. Bahi, C. Guyeux, and A. Makhoul

Table 4. RSA encryption Security level Size of the key E (battery units) 1 472 0.08 2 945 0.53 3 1416 1.63 4 1891 3.63 Table 5. Our approach Security level Size p of the key E (battery units) 1 46 0.02 2 85 0.04 3 125 0.07 4 167 0.10 Table 6. RSA encryption Security level Size of the key E (battery units) 1 472 1.13 2 945 8.09 3 1416 24.74 4 1891 56.27

Tables 5 and 6 give the energy consumption E at the aggregation stage. The same hypothesis that above have been made, the sole difference is that aggregator nodes have a battery of 1000 units of energy. It can be seen that the energy needed by aggregators are between 50 and 500 times more important in the RSA-based scheme, for the same level of security. Figure 2 gives the comparison between RSA and elliptic curve based encryption, concerning the average energy consumption of an aggregating wireless sensor network. We can notice that our approach saves the energy largely greater than the case of RSA, where its depletion is so fast. Finally let us notice that, in addition of reducing the amount of energy units needed for encryption and aggregation, the sink receives many more values per second in EC-based networks than in RSA-based one.

7 Enlarging the Number of Allowing Authentication Functions In the previous sections, we have proposed to use a homomorphism encryption scheme to support in-network processing while preserving privacy. Compared to existing secure aggregation schemes based on homomorphism encryption, our method has not been cryptanalysed. Moreover, due to the possibility to realize n additions and one product over the cypher values, this scheme enlarges the variety of allowing aggregation operations through cyphertexts. However, all of the homomorphism encryption schemes only allow some specific query-based aggregation functions, e.g., sum, average, etc.

Secure Data Aggregation in Wireless Sensor Networks

355

Agregator's energy evolution

Energy

EC 46 EC 85 RSA 472 RSA 945

Time Fig. 2. Comparison of energy consumption

Another way to achieve secure data aggregation in wireless sensor networks is to authenticate sensing values. In-network processing presents a critical challenge for data authentication in wireless sensor networks. Current schemes relying on Message Authentication Code (MAC) cannot provide natural support for this operation, because a MAC computation is a very energy-consuming operation. Additionally, even a slight modification to the data invalidates the MAC. In [21] a new way to achieve authentication through wireless sensor networks is introduced. It is based on digital watermarking and proposes an end-to-end, statistical approach for data authentication that provides inherent support for in-network processing. In this scheme, authentication information is modulated as watermark and superposed on the sensory data at the sensor nodes. The key idea formerly presented in [21] is to visualize the sensory data at a certain time snapshot as an image. Each sensor node is viewed as a pixel and its value corresponds to the gray level of the pixel. Due to this equivalency, information hiding techniques can be used to authenticate a wireless sensor network. The watermarked data can be aggregated by the intermediate nodes without incurring any en route checking. Upon reception of the sensory data, the sink is able to authenticate the data by validating the watermark, thereby detecting whether the data has been illegitimately altered. In this way, the aggregation-survivable authentication information is only added at the sources and checked by the data sink, without any involvement of intermediate nodes. In [21] the authors proposes to use a data hiding scheme based on spread spectrum techniques to achieve authentication. In their proposal, “each sensor node embeds part of the whole watermark into its sensory data, while leaving the heavy computational load of watermark detection at the sink”. Moreover, as stated before, their scheme supports

356

J.M. Bahi, C. Guyeux, and A. Makhoul

in-network aggregation. However spread spectrum is known to be not robust: even if their scheme survives to a certain degree of distortion, spread-spectrum cannot face to elementary blind attack. Furthermore, spread-spectrum data hiding techniques are only stego-secure in the “Natural Watermarking” situation [22]. The spread-spectrum subclass used in [21] is related to classical SS, i.e. with BPSK modulation [22]. This subclass is neither stego-secure [22], nor chaos-secure [23]. Among the consequences of these lack of security is the fact that an attacker how observes the network can access to the secret embedding key in all of the following situations: – Watermarked Only Attack (WOA): the attacker has access only to watermarked contents. – Known Message Attack (KMA): the attacker has access to pairs of watermarked contents and corresponding hidden messages. – Known Original Attack (KOA): occurs when an attacker has access to several pairs of watermarked contents and their corresponding original versions. – Constant-Message Attack (CMA): the attacker observes several watermarked contents and only knows that the unknown hidden message is the same in all contents. To improve the security of the network in WOA setup, the use of Natural Watermarking instead of BPSK modulation is required [22]. Indeed, this subclass of spread-spectrum techniques, recalled in [22], is stego-secure and so can face WOA attacks. However, Natural Watermarking is less chaos-secure than the data hiding algorithm presented in [24]. This algorithm, based on chaotic iterations, is able to withstand attacks in KMA, KOA and CMA setups [25]. Moreover, this technique is more robust than spreadspectrum, as it is stated in [26]. To sum up, the use of the scheme proposed in [24] improves the security and robustness of the scheme presented in [21]. Finally, an hybrid approach of secure data aggregation in wireless sensor networks can be obtained by combining homomorphic encryption and watermarking-based authentication, as it is summed up in Figure 3.

Decryption & Aggregation Sink (base station)

Aggregation over cypher−text Aggregators

Authentication over clear−text Authentication

Collecting data & Encryption

Normal Sensors

Fig. 3. Secure data authentication and aggregation in sensor networks

Secure Data Aggregation in Wireless Sensor Networks

357

8 Conclusion In this paper, we presented an elliptic curve based approach for secure data aggregation in sensor networks. It is based on data encryption with homomorphic properties that provide the possibility to operate on cypher-text. It prevents the decryption phase at the aggregators layers and saves nodes energy. Existing works have exponential bound in computation and are not suitable for sensor networks, which we tried to relax in our approach. The proposed scheme permits the generation of shorter encryption asymmetric keys, which is so important in the case of sensor networks. The experimental results show that our method significantly reduces computation and communication overhead compared to other works, and can be practically implemented in on-the-shelf sensor platforms.

References 1. Chandramouli, R., Bapatla, S., Subbalakshmi, K.P.: Battery power-aware encryption. ACM Transactions on Information and System Security, 162–180 (2006) 2. Castelluccia, C., Mykletun, E., Tsudik, G.: Efficient aggregation of encrypted data in wireless sensor networks. In: Proc. of the 2nd Annual MobiQuitous, pp. 119–117 (2005) 3. Girao, J., Schneider, M., Westhoff, D.: Cda: Concealed data aggregation in wireless sensor networks. In: Proceedings of the ACM Workshop on Wireless Security (2004) 4. Acharya, M., Girao, J., Westhoff, D.: Secure comparison of encrypted data in wireless sensor networks. In: Third International Symposium WiOpt 2005, pp. 47–53 (2005) 5. Haodong, W., Bo, S., Qun, L.: Elliptic curve cryptography-based access control in sensor networks. International Journal of Security and Networks 1(3-4), 127–137 (2006) 6. Liu, A., Ning, P.: Tinyecc: A configurable library for elliptic curve cryptography in wireless sensor networks. In: Proceedings of IPSN 2008, pp. 245–256 (2008) 7. Boneh, D., Goh, E.-J., Nissim, K.: Evaluating 2-dnf formulas on ciphertexts. In: Kilian, J. (ed.) TCC 2005. LNCS, vol. 3378, pp. 325–341. Springer, Heidelberg (2005) 8. Przydatek, B., Song, D., Perrig, A.: Sia: Secure information aggregation in sensor networks. In: Proceedings of ACM SenSys Conference, pp. 255–265 (2003) 9. Cam, H., Ozdemir, S., Nair, P., Muthuavinashinappan, D., Sanli, H.O.: Espda: Energyefficient secure pattern based data aggregation for wireless sensor networks. Computer Communication Journal (29), 446–455 (2006) 10. Huang, S.-I., Shieh, S., Tygar, J.D.: Secure encrypted-data aggregation for wireless sensor networks. Wireless Networks Journal, 1022–1038 (2009) 11. Domingo-Ferrer, J.: A provably secure additive and multiplicative privacy homomorphism. In: Boyd, C., Mao, W. (eds.) ISC 2003. LNCS, vol. 2851, pp. 471–483. Springer, Heidelberg (2003) 12. Cheon, J., Kim, W.-H., Nam, H.: Known-plaintext cryptanalysis of the domingo ferrer algebraic privacy homomorphism scheme. Inf. Processing Letters 97(3), 118–123 (2006) 13. Wagner, D.: Cryptanalysis of an algebraic privacy homomorphism. In: Boyd, C., Mao, W. (eds.) ISC 2003. LNCS, vol. 2851, pp. 234–239. Springer, Heidelberg (2003) 14. Lin, H.-Y., Chiang, T.-C.: Cooperative secure data aggregation in sensor networks using elliptic curve based cryptosystems. In: Luo, Y. (ed.) Cooperative Design, Visualization, and Engineering. LNCS, vol. 5738, pp. 384–387. Springer, Heidelberg (2009) 15. Hankerson, D., Menezes, A., Vanstone, S.: Guide to elliptic curve cryptography. Springer, Heidelberg (2004)

358

J.M. Bahi, C. Guyeux, and A. Makhoul

16. Cheung, R.C.C., Telle, N.J., Luk, W., Cheung, P.Y.K.: Secure encrypted-data aggregation for wireless sensor networks. IEEE Trans. on Very Large Scale Integration Systems 13(9), 1048–1059 (2005) 17. Boneh, D., Franklin, M.: Identity-based encryption from the weil pairing. In: Kilian, J. (ed.) CRYPTO 2001. LNCS, vol. 2139, pp. 213–229. Springer, Heidelberg (2001) 18. Castelluccia, C., Chan, A., Mykletun, E., Tsudik, G.: Efficient and provably secure aggregation of encrypted data in wireless sensor networks. ACM Trans. Sen. Netw. 5(3), 1–36 (2009) 19. Barker, E., Roginsky, A.: Draft nist special publication 800-131 recommendation for the transitioning of cryptographic algorithms and key sizes (2010) 20. Lenstra, A.K., Verheul, E.R.: Selecting cryptographic key sizes. Jour. of the International Association for Cryptologic Research 14(4), 255–293 (2001) 21. Zhang, W., Liu, Y., Das, S.K., De. Secure, P.: data aggregation in wireless sensor networks: A watermark based authentication supportive approach. Pervasive and Mobile Computing 4(5), 658–680 (2008) 22. Cayre, F., Bas, P.: Kerckhoffs-based embedding security classes for woa data hiding. IEEE Transactions on Information Forensics and Security 3(1), 1–15 (2008) 23. Bahi, J.M., Guyeux, C.: A chaos-based approach for information hiding security. ArXiv eprints (May 2010) 24. Bahi, J.M., Guyeux, C.: Hash functions using chaotic iterations. Journal of Algorithms & Computational Technology 4(2), 167–181 (2010) (accepted manuscript) (to appear) 25. Guyeux, C., Friot, N., Bahi, J.M.: Chaotic iterations versus Spread-spectrum: chaos and stego security. ArXiv e-prints (May 2010) 26. Bahi, J.M., Guyeux, C.: A new chaos-based watermarking algorithm. In: SECRYPT 2010, International conference on security and cryptography, Athens, Greece (to appear, 2010)

Guaranteeing Reliable Communications in Mesh Beacon-Enabled IEEE802.15.4 WSN for Industrial Monitoring Applications Berta Carballido Villaverde, Susan Rea, and Dirk Pesch Nimbus Centre for Embedded Systems Research Cork Institute of Technology Rossa Avenue, Cork, Ireland {berta.carballido,susan.rea,dirk.pesch}@cit.ie

Abstract. Wireless Sensor Networks (WSN) are a very promising solution for industrial monitoring applications in terms of safety, costs, efficiency and productivity. However, in order to move from the adopted manual/wired designs to wireless designs, certain guarantees must be assured, especially in terms of reliability, bandwidth and message transmission delays. Although quality of service (QoS) requirements can be satisfied in star based topologies using the Guaranteed Time Slots (GTS) feature of the IEEE802.15.4 standard, GTS communications in multihop scenarios are currently limited by IEEE802.15.4 beacon scheduling designs and peer-to-peer GTS allocation methods. This restricts applications to run over sensors that are within radio range of the network coordinator. In this work, we propose a distributed IEEE802.15.4 MAC modification to improve GTS usability and scalability in mesh networks. Also, we propose a reactive multihop GTS allocation technique based on our MAC modification to ensure reliable and latency aware end-toend communications. Results show that our techniques improve greatly the reliability of multihop GTS communications. Keywords: Wireless Sensor Networks, IEEE 802.15.4, Mesh, GTS, Beacon Scheduling.

1 Introduction Condition monitoring is recognized to benefit engineers by providing an evaluation of current equipment health, enabling optimized maintenance schedules to be made, and ensuring plant uptime [1]. This can be very critical because shutting down a factory to repair defective equipment can be very costly, in addition to the possibility of losing some of the manufactured product in the process. In this context, the benefits of using a wireless sensor network can be numerous in terms of safety, cost, efficiency and productivity. Wireless sensor networks are gaining in popularity for industrial monitoring applications due to their relatively low cost and simplicity for retrofitting into existing infrastructure. They are particularly suited to this type of applications as they do not require cabling, which will lead to shorter outages during installation and a lower capital outlay than their wired equivalent. However, some of the desired J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 359–370, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

360

B.C. Villaverde, S. Rea, and D. Pesch

measurements for industrial monitoring applications, such as vibration and acceleration, may generate tens of kilobits of data with bandwidth, reliability and soft latency requirements (of a couple of seconds) for long sampling periods. In order to satisfy these demands and to fully exploit the benefits of a WSN, a QoS aware multihop solution must be developed. The most used WSN communication protocol to date is the IEEE 802.15.4 [2] protocol stack as it has been designed specifically for low power, low cost wireless communications. Its specification defines the physical (PHY) and medium access control (MAC) layers. The IEEE 802.15.4 MAC protocol supports two operational modes: the beaconless mode, in which nodes stay active all the time, and the beacon mode, in which beacon frames are periodically sent by coordinators to synchronize sensor nodes. The advantage of this synchronization scheme is that all nodes can wake up and sleep at the same time allowing very low duty cycles and hence save energy. In addition, when the beacon mode is used, nodes can use Guaranteed Time Slots specifically designed to fulfill application’s QoS requirements. In order to use the beacon mode for multihop networks, beacon scheduling mechanisms have to be utilized to avoid direct and indirect collisions of the beacon frames. Most designs for beacon scheduling, if not all, are based on two main techniques initially proposed by the Task Group (TG) 15.4b [3]: Superframe Duration Scheduling (SDS) and Beacon Only Period (BOP). These designs, like in star based networks, allow the use of CSMA/CA transmissions and GTS transmissions. CSMA/CA based media access is not a reliable communications mechanism due to the likelihood of packet collisions and the inability to reliably estimate and guarantee communication delays. This leaves the GTS feature as the best option for reliable communications where each guaranteed time slot is allocated by the coordinator for the sole use of a single node. Although GTS communications are possible with the beacon scheduling designs alluded to above, both schemes present problems in terms of bandwidth usability, delay and scalability when the GTSs are used (see section 3.1.). Also, if the standard way of allocating GTSs is used (peer-to-peer) a complete GTS allocation from source to destination of the information cannot be guaranteed. To overcome these problems, in this paper we propose two solutions: First, to solve the scalability problems of the current beacon scheduling techniques when using GTS communications, we propose a IEEE802.15.4 MAC modification based on the BOP beacon scheduling technique. Second, we propose a multihop GTS allocation technique to ensure reliable and latency-aware mesh communications. Our technique will reactively search for a reliable end-to-end GTS route that fulfils the specific application requirements using MAC and Routing layer information. Results show that our techniques improve greatly the reliability of multihop GTS communications. The rest of this paper is organised as follows. In Section 2, the IEEE 802.15.4 MAC is described. In section 3, the challenges presented when trying to use GTS communications in multihop networks and the solutions adopted to overcome those problems are presented. Section 4 describes the simulation scenarios and results that accentuate the distinct advantages of the proposed approaches when compared with other schemes. Finally, conclusions are drawn and future work is described in Section 5.

Guaranteeing Reliable Communications in Mesh Beacon-Enabled IEEE802.15.4 WSN

361

2 IEEE 802.15.4 MAC The IEEE 802.15.4 standard describes the physical layer and the MAC sub-layer for Low-Rate Wireless Personal Area Networks (LR-WPANs). The MAC sub-layer has two operational modes: beacon-enabled and non beacon-enabled. Medium access can be contention based (slotted or unslotted CSMA/CA) or contention free based (only when the beacon-enabled mode is active). When the beacon-enabled mode is active, the Personal Area Network Coordinator (PANC) sends beacon frames at the start of every Beacon Interval (BI). The beacon frames are used to identify the PAN, to allow the synchronisation of associated devices and to inform the nodes of the superframe structure -consisting of an active period and, optionally, an inactive period. The active period of the Superframe Duration (SD) is divided into 16 equally sized time slots, during which data transmission is allowed. Each active period of the SD can be further divided into a Contention Access Period (CAP) and an optional Contention Free Period (CFP) – composed of GTSs. Slotted CSMA/CA is used during the CAP. The superframe structure is characterized by two parameters, the Superframe Order (SO) and the Beacon Order (BO), which establish the active period (Superframe Duration -SD) and the length of the superframe (Beacon Interval -BI) respectively. When establishing the values of both parameters, the following relationship must be satisfied: 0 ≤ SO ≤ BO ≤ 14. BI and SD are defined as follows: (1) BI = aBaseSuperframeDuration × 2^BO SD = aBaseSuperframeDuration × 2^SO

(2)

The aBaseSuperframeDuration constant represents the minimum length of the superframe when BO is equal to 0. The PANC allocates the GTSs. When receiving a GTS allocation request, the PANC verifies whether there are sufficient resources and, if possible, allocates the requested GTS on a first come first served basis. The PANC can allocate up to 7 GTSs in each superframe and each allocation can be composed of one or more time slots. When allocating GTSs, the PANC must reserve a minimum length for the CAP. Any device with an allocated GTS can also transmit during the CAP.

3 GTS Use in Beacon-Enabled IEEE802.15.4 Multihop Networks This section describes the challenges associated to using GTS communications in mesh networks and the solutions proposed to overcome the described problems. 3.1 Problem Specification As stated before, the Task Group (TG) 15.4b, a group formed to enhance the 2003 version of the IEEE 802.15.4 standard, proposed two beacon scheduling techniques to avoid beacon collisions in the beacon mode and hence facilitate multihop beaconenabled communications in IEEE 802.15.4 WSN. In the first approach, namely superframe duration scheduling (SDS) each coordinator transmits its superframe

362

B.C. Villaverde, S. Rea, and D. Pesch

during the inactive period of its neighbours and its neighbours’ neighbours to avoid direct and indirect beacon collisions (Fig. 1 a)). In the second approach, a beacon-onlyperiod (BOP) is created at the start of the superframe where every coordinator selects a free time-slot to transmit its own beacon and thus avoid collisions (Fig. 1 b)). However, the two beacon scheduling approaches considered by the TG4b were not included in the revision of the standard in 2006 [4]. While research has continued in this field, it is unclear if these mechanisms will be included in future releases of the standard, i.e 802.14.4e, as some beacon scheduling proposals have appeared within the group [5]. Inline with this, several researchers have employed SDS as the beacon scheduling mechanism to enable multihop topologies over the beacon-enabled mode [6][7][8]. Each and every one of these techniques allows the full use of the GTS slots without any modification and due to the way the scheduling is performed, reliability is assured since direct and indirect collisions among transmitting nodes in the CFP are eliminated. However, the scheduling design has a drawback: the delay introduced in each hop makes it unsuitable and not scalable for delay bounded communications since, on average, a node would have to wait BI/2 sec. to transmit to a neighbor of the mesh network (Fig. 2 a)). On the other hand, research has been conducted studying the BOP approach to enable mesh networking [9][10][11]. The problem with BOP based approaches and the use of the GTS lies in the fact that since nodes share the same superframe duration, once a node occupies a GTS, its neighbors and neighbors’ neighbors can not reuse it to avoid collisions. Additionally, in order to make neighbours aware of blocked GTS, all GTS command transactions have to be broadcasted which makes the command frames vulnerable to the hidden terminal problem. Although this approach is more efficient in terms of delay than the SDS approach, the blocking and collision problems make it impractical and not scalable (Fig. 2 b)). Therefore, given the fact that the available beacon scheduling techniques are unsuitable for performing mesh GTS communications with reliability and latency demands, a different approach must be taken for this part of the superframe. In this work, we propose a modification of the CFP of the BOP approach so reliable and scalable GTS communications can be performed in mesh beacon-enabled IEEE 802.15.4 WSN. However, even if the GTS usage problem is solved, there still exists one unanswered issue: how to allocate the necessary GTS resources along a route from source to destination in the mesh network while fulfilling all the application requirements. Current trends in allocating GTSs in multihop networks are based on peer-to-peer allocation mechanisms such as proposed by TG4e [12], as proposed in the IEEE802.15.5 standard [13] or as per the proposal in [14]. However, due to the fact that the number of slots is limited, a multihop peer-to-peer allocation can be unsuccessful if a node in the multihop path has all GTSs slots in use, which may likely happen near sinks or clusterhead nodes. In addition, end-to-end delay from source to destination cannot be estimated until all peer-to-peer connections have been established and hence it is not possible to guarantee that the application’s end-to-end demand will be met. One could argue that a reactive routing protocol could be used to find a suitable GTS route between two nodes of the mesh network. This way, once a route is found and communicated to the source node, the source node could start

Guaranteeing Reliable Communications in Mesh Beacon-Enabled IEEE802.15.4 WSN

363

requesting peer-to-peer GTSs along that route. We think however that once the best route or a suitable route that fulfills the application requirements is known, the nodes along the path to the source can directly reserve the GTS resources when receiving confirmation of the route formation (in a destination-origin order) and hence reduce the route establishment time and overhead. Therefore, we think that a cross-layer reactive GTS allocation technique is needed to guarantee that reliability, bandwidth (GTS slot is guaranteed from origin to destination) and delay constraints are met. Additionally, we propose a GTS path recovery mechanism not considered by peer-to-peer allocation methods. It is worth highlighting that the ZigBee specification [15] for mesh IEEE802.15.4 WSNs only considers the non beacon-enabled mode (which does not support GTS) and hence does not contemplate this situation.

a) Superframe Duration Scheduling (SDS)

b) Beacon-only-period Scheduling (BOP)

Fig. 1. Task Group 15.4b beacon scheduling approaches

a) Delay problem with SDS

b) GTS blockage problem with BOP

Fig. 2. GTS usage problems with different Beacon Scheduling techniques

3.2 Solution to the Problem Here we describe the two solutions we propose to improve GTS communications in multihop beacon-enabled IEEE802.15.4 WSNs. 3.2.1 Contention Free Period Modification for BOP The BOP approach for scheduling beacons and superframes has the inconvenience that the use of the GTS suffers from blockages that reduce the bandwidth utilization and limit the scalability. However, a BOP design has an advantage in that it introduces less delay than a SDS design. Here we propose a modification of the GTS part of the

364

B.C. Villaverde, S. Rea, and D. Pesch

superframe to be used with BOP scheduling to improve the bandwidth usage and increase scalability. It is conceivable that dividing the inactive period into virtual time slots, the same approach adopted by the IEEE802.15.5 standard [13], would solve the BOP scalability problem (Fig. 3 a)). This is true in the sense that by doing so the number of usable slots is increased. However, this solution still presents one difficulty, in order to use a slot for GTS frame reception, neighbors must be notified so they do not use the same slot to receive GTS frames or transmit to other neighbors. In order to notify other nodes, notification packets are broadcasted whenever a slot is allocated. Considering the collision prone nature of CSMA/CA communications and the fact that broadcast packets can not be acknowledged or combined with RTS/CTS techniques, this approach is not reliable –a failure in receiving a blocked slot warning translates into jeopardized GTS communications in neighboring nodes. Therefore, a different approach must be taken. Here, we propose dividing the sleep time into Virtual Contention Free Periods (VCFP) composed of Virtual GTS slots (VGTS), and then assign these in a distributed and periodic fashion to different nodes (Fig. 3 b)). With this design, each node has a different set of VCFP within its two hop neighborhood so data frame collisions are avoided. This also eliminates the need for broadcasting control packets so acknowledgements between the interested nodes can be used. The distributed allocation of the VCFPs is performed using a similar algorithm used by MeshMAC approach for distributed beacon scheduling [8] and it is done at the time of the association. The term virtual is used because the coordinator does not need to stay active in the VGTS if the slot is not allocated to any node which in turn saves energy. All VCFPs are composed of 5 VGTS (each VGTS lasts approx 1ms) – 5 is the minimum size to transmit the maximum IEEE802.15.4 frame length (127 bytes plus headers). The periodicity of the VCFPs depends on the expected number of two hop neighbors per node, i.e., if a node is expected to have a maximum five neighbors in the two hop vicinity, the periodicity of its own VCFPs will be every five VCFP.

a) BOP modification based on IEEE802.15.5

b) Proposed BOP modification of the CFP

Fig. 3. Contention Free Period modifications for BOP Beacon Scheduling

3.2.2 Multihop Cross-Layer Reactive GTS Allocation Mechanism In this work we propose a reactive cross-layer protocol that seeks a route with enough GTSs resources that fulfils the bandwidth and delay demands of the application and after that reserves the GTS slots along that route on a destination-source order. Unlike peer-to-peer allocation methods that ask for resources on a hop-by-hop basis, our

Guaranteeing Reliable Communications in Mesh Beacon-Enabled IEEE802.15.4 WSN

365

mechanism ensures that all nodes in a selected route are able to provide the requested GTS resources. The proposed mechanism works as follows: Nodes start the mesh GTS route search by broadcasting the Mesh GTS Allocation Request (MGA-Req) in the CAP (Fig. 4). Nodes that have free GTS slots that match the GTS Characteristics requirements rebroadcast the packet until the destination is reached and reserve the selected slots until the Mesh GTS Allocation Confirm (MGA-Con) is received or a timeout expires. On receiving more than one MGA-Req for the same route, the MGA-Req packet with the lowest accumulated delay will be re-broadcasted to reduce flooding of requests, given that the maximum tolerated delay requirement of the application is fulfilled. The accumulated delay per hop is calculated as the time that covers a GTS frame reception in the node’s own reserved slot to the transmission of the frame in the previous hop node. To calculate this delay, the reserved slots information is passed from one node to the next locally. A node might occupy more than one VCFP if necessary. Also, in order to decrease request flooding, packets will expire when a predefined hop count is reached. The best path information is stored in the Mesh Cross-Layer GTS Allocation Table while the node waits for confirmation. Request packets have a unique ID that together with the delay information avoid loop formation. The destination, upon receiving the reservation requests, will estimate which paths fulfil the application delay requirements by checking the “Accumulated Delay” and “Tolerated delay” fields of the MGA-Req packet. Once the path is selected, it will start the allocation and confirm process by unicasting a MGA-Con packet to its selected neighbour (Fig. 5) containing the allocated GTS information. A node receiving a MGACon will acknowledge the packet and it will unicast it to the following neighbour. In addition, the node will store the GTS allocation information in the allocation table. The process stops when the origin node is reached and, at this time, the node can start sending packets through the newly established multihop GTS connection.

Fig. 4. MGA-Req packet format

Fig. 5. MGA-Con packet format

3.2.2.1 Multihop GTS Allocation Maintenance Another drawback of the peer-to-peer GTS allocation mechanisms available in the literature is that they do not provide recovery mechanisms for multihop GTS allocation. In a WSN, a route may suddenly become unavailable because of changes in the wireless channel, node failure or due to node mobility. The link failure might happen once the route is established or even in the route establishment process. Therefore, route maintenance mechanisms must be available. If a failure occurs in the GTS allocation process, for example if an MGA-Con is not acknowledged, the failure must be communicated to the nodes that have already

366

B.C. Villaverde, S. Rea, and D. Pesch

reserved resources so they can free them. Also, if a failure occurs after the route is established, detected when nodes cannot hear beacon frames of their data packet destinations or because nodes do not receive data packets in the allocated GTSs, the link failure must also be communicated. In all these cases, a deallocation packet (Mesh GTS Deallocation Request MGD-Req) will be sent indicating the direction of the deallocation (towards the destination or towards the source) and if the deallocation is performed due to a link failure - to differentiate it from a normal deallocation (Fig. 6). A node receiving a failure deallocation packet towards the origin will start a new allocation process towards the destination to find an alternative route.

Fig. 6. MGD-Req packet format

4 Simulation Scenario and Results In designing the experimental environment we rely on the multihop IEEE802.15.4 OPNET simulation model [16] developed with OPNET Modeler [17] as a basis for implementing and testing our BOP modification and multihop GTS allocation mechanism. In addition, we use a distributed version of BOP named DBOP presented in [16] to perform our simulations. Typical industrial scenarios may have multiple sinks with the number of sinks being far smaller than the total number of nodes. Also, they may be composed by between 10 to 200 field devices and usually, they have a maximum number of hops of 20 [18]. We use a random network topology composed of 30 MicaZ nodes [19]. The maximum distance between two nodes is 12 hops and there are between 1 to 5 neighbours per node. The network has 4 sinks and nodes select a destination sink randomly. All nodes send data frames of 127 bytes (the maximum possible size). The buffer size for all nodes is limited to 1250 bytes. Nodes detect tool monitoring events such as vibrations or acceleration randomly and report the generated data to the corresponding sink. The time a node is detecting an event is referred to as a session and their duration is modeled with an exponential distribution of mean 300 seconds. The interarrival times between sessions follow a Poisson distribution with means ranging from 300s to 1800s (in steps of 300s). Varying the session interarrival times has the same effect as fixing them and varying the number of nodes. In our simulation rounds, we fix SO to 2 and BO to 6. Finally, data generation rates per session are also varied (1kbps and 3.5kbps) to examine how requesting more or less resources affects the overall performance. Finally, To test the benefits of our MAC layer modification and multihop GTS allocation algorithm, we compare first with the BOP modification based on IEEE802.15.5 MAC (Fig. 3 a)), referred to as 15.5BOP with our own MAC modification (Fig. 3 b)), referred to as Distributed CFP BOP (DCBOP) using a peerto-peer GTS allocation method. Afterwards, using our MAC modification, we compare the peer-to-peer GTS allocation method with our reactive allocation method. Every time the peer-to-peer GTS allocation is used, AODV [20] is used first to find a route if

Guaranteeing Reliable Communications in Mesh Beacon-Enabled IEEE802.15.4 WSN

367

the route is not already known. We select a typical value of Active Route Time-Out (ART) of 300 seconds [21]. 4.1 Simulation Results Fig. 7 depicts the number of successfully established multihop GTS connections for 15.5BOP and DCBOP MAC modifications with a peer-to-peer allocation method and DCBOP with the reactive allocation method for data generation rates per session of 3.5kbps. When both MACs are compared with the same allocation method (red-circle and black-triangle lines) we can see the clear benefit produced when our MAC modification is used. Since allocation requests do not have to be broadcasted to make other neighbors aware of GTS slots usage when our MAC modification is used (because the CFP is not shared among neighboring nodes), the problems caused by the hidden terminal problem are eliminated and therefore the number of successfully established multihop GTS connections is increased. If our reactive allocation method is added to our modified MAC (blue line), the gains are even greater. It has to be noted that for the selected data generation rate (3.5kbps) and BO/SO parameters the network is congested and therefore it is not possible to satisfy all GTS connections (successful connections are below 60% at all cases). Fig. 8 depicts the successfully established multihop GTS connections for our DCBOP MAC modification combined with a peer-to-peer allocation method and our reactive allocation method for data generation rates per session of 1kbps and 3.5kbps. As can be seen, our reactive allocation method always outperforms the peer-to-peer method because routes are established through nodes that have available resources at the request time. Peer-to-peer allocation methods cannot check if the nodes in the route (obtained using AODV) have available resources at the request time and therefore it is not possible to ensure that the connection will be successful beforehand even though the path to a given destination is known. If the data generation rate per session is decreased from 3.5kbps to 1kbps the occupancy of resources is decreased and the performance of the allocation methods is obviously improved. Fig. 9 shows the overhead caused by the peer-to-peer and reactive allocation methods for session interarrival times of 600s and 1800s and for 1 and 3.5 kbps data generation rates. As can be observed, the control bits / data bits ratio is quite low in both cases. This is due to the fact that sessions are quite long (condition monitoring events such as vibrations in industrial machinery can last for minutes generating tens of kilobits of data) and hence the number of control packets is low compared to the number of data packets. The difference between the peer-to-peer allocation method and our reactive method is also minimal since the peer-to-peer allocation method uses AODV reactive routing first to find the routes if a new route is needed or the previous one has expired. Therefore, the use of our reactive approach is justified. Finally, Fig. 10 depicts the Cumulative Distribution Function of the end-to-end delay for all received packets at all network sinks. As can be seen, both allocation mechanisms are able to maintain the end-to-end delay below 1.5 seconds for 90% of received packets which is sufficient for industrial monitoring applications that can tolerate up to a couple of seconds of delay. Our reactive scheme introduces slightly more delay than the peer-topeer allocation method because it does not always choose the optimal path in terms of delay since that path may have all of its GTS resources occupied.

368

B.C. Villaverde, S. Rea, and D. Pesch

Fig. 7. Successfully established multihop GTS connections for different BOP CFP modifications

Fig. 8. Successfully established multihop GTS connections for different allocation methods

Fig. 9. Overhead introduced by the different multihop GTS allocation methods

Guaranteeing Reliable Communications in Mesh Beacon-Enabled IEEE802.15.4 WSN

369

Fig. 10. Received packets end-to-end delay

5 Conclusion and Future Work In this work, we have proposed a distributed IEEE802.15.4 MAC modification to improve GTS usability and scalability in mesh networks. Also, we have proposed a reactive multihop GTS allocation technique based on our MAC modification to guarantee reliable and latency aware communications from source to destination in mesh beacon-enabled IEEE802.15.4 WSNs. As shown by our simulation results, both techniques increase the reliability of the network communications since more GTS connections are established successfully which permits more hidden-terminal-free communications in mesh networks. We also show that the cost of using our reactive allocation method is minimal in terms of overhead for our targeted industrial monitoring applications. Future work will be based on combining service differentiation techniques with our allocation method to distribute the available resources when the network is congested.

References 1. Baker, P.C., Catterson, V.M., McArthur, S.D.J.: Integrating an Agent-based Wireless Sensor Network within an Existing Multi-agent Condition Monitoring System. In: International Conference on Intelligent System Applications to Power Systems (ISAP 2009), November 8-12 (2009) 2. IEEE 802.15.4 Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications for Low-Rate Wireless Personal Area Networks, LR-WPANs (2003) 3. Lee, M., Zheng, J., Zhang, J., Liu, Y., Dai, H., Shao, H.R.: Combined Beacon Scheduling Proposal to 802.15.4, TG4b (September 16, 2004) 4. IEEE 802.15.4 Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications for Low-Rate Wireless Personal Area Networks, LR-WPANs (2006) 5. Shin, C., Jeong, W., Hwang, S., Lee, A., Joo, S.: Beacon collision avoidance mechanism for TG4e MAC, TG4e (September 2008)

370

B.C. Villaverde, S. Rea, and D. Pesch

6. Koubâa, A., Cunha, A., Alves, M., Tovar, E.: TDBS: a time division beacon scheduling mechanism for ZigBee cluster-tree wireless sensor networks. Real-Time Systems 40(3), 321–354H.-I (2008) 7. Burda, R., Wietfeld, C.: A Distributed and Autonomous Beacon Scheduling Algorithm for IEEE802.15.4/ZigBee Networks. In: Proc. of IEEE MASS 2007, Pisa, Italy (October 2007) 8. Muthukumaran, P.S., de Paz, R., Špinar, R., Pesch, D.: MeshMAC: Enabling Mesh Networking over IEEE802.15.4 through distributed beacon scheduling. In: Zheng, J., et al. (eds.) ADHOCNETS 2009. LNICST, vol. 28, pp. 561–575 (2009) 9. Jeon, H., Kim, Y.: BOP (Beacon-Only Period) and Beacon Scheduling for MEU (MeshEnabled USN) Devices. In: The 9th International Conference on Advanced Communication Technology, vol. 2, pp. 1139–1142 (February 2007) 10. Kim, E., Choi, H.: EBBS: Energy-efficient BOP based Beacon Transmission Scheduling for WSNs. In: Proceedings of PIMRC 2008 (2008) 11. Ferrari, M., Pizziniaco, L.: An Adaptive Scheme for Active Periods Schedule in IEEE 802.15.4 Wireless Networks. In: 3rd International Symposium on Wireless Communication Systems (2006) 12. Shen, J., Yao, D., Xing, T., Zhao, Z.F., Liu, H.: Supporting Peer to Peer Network and Improving throughput by enhanced GTS, TG4e (September 2008) 13. IEEE 802.15.5 Mesh Topology Capability in Wireless Personal Area Networks (WPANs) (2009) 14. Khayyat, A., Safwat, A.: The Synchronized Peer-to-Peer Framework and Distributed Contention-Free Medium Access for Multihop Wireless Sensor Networks. Journal of Sensors 2008, Article ID 728415, 28 pages (2008) 15. ZigBee Alliance Document 053474r17, ZigBee Specification, v. 1.0 r17 (2007) 16. Carballido Villaverde, B., De Paz Alberola, R., Rea, S., Pesch, D.: Experimental Evaluation of Beacon Scheduling Mechanisms for Multihop IEEE 802.15.4 Wireless Sensor Networks. In: The Fourth International Conference on Sensor Technologies and Applications, Sensorcomm 2010 (2010) 17. Opnet Tech. Inc., Opnet Modeler - version 14.5, http://www.opnet.com/ (Last accessed 23/06/10) 18. Industrial Routing Requirements in Low Power and Lossy Networks. Work In Progress Internet Draft. Networking Working Group, Internet Engineering Task Force, http://tools.ietf.org/html/ draft-ietf-roll-indus-routing-reqs-05 (Last accessed 23/06/10) 19. MicaZ motes datasheet from Crossbow Technology, http://www.xbow.com/Products/productdetails.aspx?sid=164 (Last accessed 23/06/10) 20. Perkins, C.E., Royer, E.M.: Ad hoc On-Demand Distance Vector Routing. In: Proceedings of the 2nd IEEE Workshop on Mobile Computing Systems and Applications, New Orleans, LA, pp. 90–100 (1999) 21. Richard, C., Perkins, C.E., Westphal, C.: Defining an Optimal Active Route Timeout for the AODV Routing Protocol. In: Second Annual IEEE Communications Society Conference on Sensor and Ad-Hoc Communications and Networks, IEEE SECON 2005, California, USA, September 26-29 (2005)

A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks Yingjun Jiang1, Chung-Horng Lung1, and Nishith Goel2 1

Department of Systems and Computer Engineering Carleton University, Ottawa, Ontario, Canada {yjiang6,chlung}@sce.carleton.ca 2 Cistel Technology, Ottawa, Ontario, Canada [email protected]

Abstract. This paper introduces a static Tree-based Multiple-Hop Distributed Hierarchical Agglomerative Clustering (TMH-DHAC) approach for wireless sensor networks (WSNs). The proposed TMH-DHAC is derived from the Hierarchical Agglomerative Clustering (HAC) and the distributed HAC (DHAC) methods. TMH-DHAC adopts an energy-aware cluster-head election policy to balance the energy consumption and workload among sensor nodes in the network. The multi-hop tree structure provides the near-optimal routes for intracluster data transmissions. The proposed TMH-DHAC method for the time-slot allocation enables simultaneous conflict-free communications between different pairs of sensors and increases the maximum transmission throughput. The simulation results show that TMH-DHAC performs better than previous approaches: LEACH, LEACH-C and the DHAC-RSS (Received Signal Strength) protocols, in terms of the network lifetime, total data amount, energy efficiency, and average transmission distance. Keywords: wireless sensor networks, clustering, multi-hop, minimum spanning tree, performance evaluation.

1 Introduction A wireless sensor network (WSN) may consist of a number of sensor nodes which work in a collaborative way. One of the important issues in WSNs is how to manage and organize sensor nodes effectively to perform the data-gathering in the network. In general, sensors consume a considerable amount of energy on wireless communications. Devising an energy-efficient communications scheme is a crucial concern for sensor networks. Generally, depending on the basic network structures, the routing protocols for WNSs are classified into two categories: flat and hierarchical [9]. Hierarchical protocols have the advantages of scalability, communication efficiency and network lifetime. LEACH [6] is one of the first clustering approaches for WSNs. It has been widely used as a base model for extension and/or performance comparison. Most clustering approaches adopt the same or similar model used in LEACH. One main feature of the LEACH model is the assumption of star or star-like topology within J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 371–383, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

372

Y. Jiang, C.-H. Lung, and N. Goel

clusters. In other words, cluster member (CM) nodes communicate directly with the cluster head (CH). This communication model is not energy efficient as the distances between CM nodes and the CH in a cluster could be large and the energy consumption has a direct relationship with the communication range; see Transmission Energy Dissipation in Section 4. The motivation of this paper is to devise an energy-efficient and distributed hierarchical clustering approach for WSN applications to reduce the transmission distance for sensor nodes and hence the corresponding energy cost. It is well known that the minimum spanning tree (MST) is a mathematical structure connecting all nodes in a network with the minimum sum of weights of its branches/links. Applying the MST concept to a hierarchical clustering protocol can reduce the total transmission distance and achieve saving of energy consumption in communications. Another direction to improve the energy-efficiency is to exploit the immobility of sensor nodes using static clustering. Sensors can be grouped into an unchanged cluster formation in early phase and avoid the frequent and often less-optimal re-clustering process. The hierarchical agglomerative clustering (HAC) [14] algorithm is a simple but effective centralized clustering approach and has been successfully applied to many disciplines. [11] and [22] adapted HAC and developed a bottom-up method, Distributed HAC (DHAC) algorithm, that forms clusters in a WSN without the global knowledge of the network. This paper proposed a Tree-based Multiple-Hop DHAC (TMHDHAC) algorithm which is derived from DHAC but further modifies the clustering method by adopting a multi-hop transmission tree structure to avoid most of the longdistance direct transmissions from CMs to the CH to reduce energy consumption. The rest of the paper is organized as follows: Section 2 describes related work. Section 3 discusses the TMH-DHAC algorithm. Section 4 presents the simulation and results. Finally, Section 5 is the conclusions.

2 Related Work Hierarchical routing provides a better solution to organize and utilize the resourceconstrained sensors to work collaboratively. Hierarchical protocols can be further classified as distributed and centralized approaches. Some representative protocols in these two categories are introduced here. 2.1 Distributed Hierarchical Routing Protocols LEACH [6] is self-adaptive routing protocol. The LEACH protocol randomly chooses some nodes as CHs and organizes the sensor nodes into clusters. CHs collect and aggregate data from regular nodes and send it to the base station (BS). The role of CH is rotated among regular nodes to avoid depleting the batteries of certain sensors. One main drawback of LEACH is the randomness in the clustering procedure. In the unequal clustering mechanism of LEACH [3] and energy efficient CH selection algorithm [15], new CH election methods were proposed, which factor in the energy status of sensors to improve the balance of the energy consumption among sensors and prolong the network lifetime.

A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks

373

HEED [19] improves the cluster selection process in LEACH and also generates well-distributed clusters. In HEED, the energy status of a node is considered to determine the CH and a high-energy node has a better chance to be a CH. HEED also avoids the irregular cluster formation issue in LEACH. PEGASIS [10] is a chain-based protocol proposed to reduce the transmission distance between nodes. All of the nodes are organized into one logical chain with the greedy algorithm. Every node only communicates with its two direct neighbors in the chain and data are transmitted along the chain. ERA [2] adopts the same CH election process as LEACH but improves the cluster formation by factoring energy status into the protocol. In LEACH, a non-CH node simply chooses the closest cluster to join. Such design can cause low-energy CHs exhausted early. In ERA, non-CH nodes join the cluster with a path that has the most residual energy to the BS. DHAC [11, 22] is a novel static clustering protocol inspired by the HAC algorithm. DHAC exploits the immobility attribute in most WSNs and proposes a bottom-up approach in clustering. In DHAC, the clustering structure is formed before the CH selection. The static clustering process is performed when sensor network is initialized and is not required to be repeated later. DHAC groups the similar nodes together by the location or signal strength information exchanged by sensor nodes. DHAC also adopts solutions to reduce the routing overhead and balance the energy consumption of nodes such as the automatic CH rotation and rescheduling. 2.2 Centralized Hierarchical Routing Protocol LEACH-C [6] utilizes a centralized algorithm for clustering but adopts the same steady-state scheme as LEACH. In the set-up phase of LEACH-C, every node in the network sends its location and residual energy information to BS directly. The BS performs the simulated annealing algorithm with the information to calculate the optimal cluster formation in that round. The BS broadcasts the cluster formation message in the network after it is generated. Every node receives the message and knows its role either as a CH or a CM. Dynamic/Static Clustering (DSC) [1] is an extension of LEACH-C. Each node gets its location using GPS and sends the location information and energy status to BS. The BS will then determine the number of CHs based on the collected information and broadcast the clustering result to each node. BSDCP [12] adopts an iterative cluster splitting method to achieve an evenly scattered cluster formation. BSDCP also utilizes a CH-to-CH multi-hop routing scheme to reduce the energy cost. The inter-cluster routing paths are a minimum-spanning-tree structure generated by the BS in a centralized manner. 2.3 Multi-hop Routing Multi-hop routing is an energy-efficient approach for data aggregation in WSNs. Multi-hop routing for WSNs has been studied by various researchers, e.g., [4], [5], [12] (see BSDCP in Section 2.2), [16], [18], [20], [21]. The objectives of those papers range from studies of energy-aware broadcasting/optimal number of hops, overlapping

374

Y. Jiang, C.-H. Lung, and N. Goel

clusters, cooperative MIMO techniques, to identification of thresholds that are more suitable for different coding, modulation, and fading models. [5] compares energy efficiency of one-hop vs. multi-hop routing in WSNs using the chain topology in a small network. Both [7] and [16] also uses spanning tree for data aggregation. But clustering is not used in either one and both MST algorithms are centralized for the entire network. [17] discusses how to determine the optimal hop number for energy efficiency. Clustering is not considered in the study and network performance results, such as lifetime, packet delivery ratio, and etc. are beyond its scope. [18] assumes that a network has been hierarchically clustered and then devises a multi-hop scheme. [21] deals with communications of a node to another node in a different cluster. Our approach, on the other hand, takes into account both clustering and multi-hop at the same time, including the setup, scheduling of data transmissions, rotation of CHs and maintenance of clusters, and the establishment of an MST within a cluster. In addition, interference-avoidance is considered in our multi-hop communications.

3 Tree-Based Multiple-Hop Distributed Hierarchical Agglomerative Clustering Protocol In this paper, the TMH-DHAC algorithm is proposed to improve the performance of the DHAC algorithm [11, 22]. TMH-DHAC adopts a tree-like intra-cluster structure to shorten data transmission distances within a cluster and simplifies the complexity in the clustering procedure. TMH-DHAC forms a minimum spanning tree (MST) within a cluster. The CH acts as the root of the tree and every other node has its parent node. In TMH-DHAC, a multi-hop approach is adopted within a cluster to shorten transmission distance and hence energy cost. Data transmissions start from the leaf nodes to their parent nodes which aggregate their own data with those received from their children and send the aggregated data to its upper-level node. The process will be repeated from leaf nodes all the way to the root node in a cluster. By this means, every node merely communicates with its parent node and child node(s). The average transmission distance is much shorter in TMH-DHAC than star-like topology as a result of the MST formation. 3.1 Clustering This section presents the proposed clustering process. The first two steps are common to well-known HAC methods [14]. The third step is a modification of DHAC to build an MST instead of a star-like topology. Step 1: Input data set An input data set for HAC is a component-attribute data matrix. Components are the nodes that we want to group based on the attributes, e.g., their locations, in WSNs. Every node knows its one-hop neighbors by exchanging a topology discovery message. Nodes detect the distance by measuring the signal strength and update their initial matrix. Figure 1 shows an example of an 8-node network.

A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks 10

5

Y-axis

2

5

1

0 0

3

4

5 X-axis

375

8 7 6 10

Fig. 1. A Simple 8-node Network

Step 2: Compute the resemblance coefficients A resemblance coefficient in this context represents the distance between two components or sensor nodes. We can calculate Euclidean distance based on the location information or from the strength of the received signal. Step 3: Execute the cluster-merging operation Initially, each node is treated as a singleton cluster and is the CH. Table 1 depicts the matrices for nodes 1 and 4 for illustration. Each cluster finds the smallest coefficient (shown in bold in Table 1) and send a CONNECT message to the corresponding cluster if they have not been merged. The CONNECT message contains the sender and receiver cluster ID and the shortest link to connect the two clusters. Cluster ID is the tie-breaker if both nodes send CONNECT to each other. Table 1. Initial resemblance matrices for Node 1 and Node 4: an example

Cluster {1} 1-hop Neighbor Coefficient {3} 2.90 {4} 3.40 -

Cluster {4} 1-hop Neighbor Coefficient {1} 3.40 {3} 1.70 {6} 3.83

The cluster receiving the CONNECT message performs the merging operation and combines the nodes originally in both of the sender and receiver clusters to form a new larger cluster. Further, the sender and receiver nodes build a link between them. The link will become part of the tree structure used for data transmission in a later phase. The CH in the receiver cluster automatically becomes the CH of the new cluster and updates the resemblance matrix of its cluster. The CH in the sender cluster becomes a regular cluster member. In this step, every node builds a link with its closest neighbor and all of the links together form a minimum spanning tree within the cluster. The number of clusters reduces quickly in the merging process. The initial 8 singleton clusters in the example are merged into 3 clusters in the first iteration of clustering. Nodes 3, 7, 8 which receive the merging requests, become the cluster heads of the new clusters {1,2,3,4}, {5,8}, and {6,7}, respectively. Table 2 shows the updated matrices for two clusters after all clusters exchange their information. SLINK [14] method is used in TMH-DHAC since it uses the shortest distance to other clusters for resemblance matrix recalculation after merging. The links are necessary to build the MST when clusters are merged.

376

Y. Jiang, C.-H. Lung, and N. Goel Table 2. Resemblance matrices of clusters: an illustration

Cluster {1,2,3,4} Neighbor Coefficient Link {6,7} 3.83 {4}-{6} -

Cluster {6,7} Coefficient 3.83 3.80

Neighbor {1,2,3,4} {5,8}

Link {6}-{4} {7}-{8}

Next, every CH broadcasts an INFORM message containing its cluster members to its neighbor clusters. The clusters receiving the message update their matrices. Then clusters repeat the operation to find the shortest distance in the matrix and send a CONNECT message to the corresponding cluster associated with it. Each cluster repeats the merging step until a certain stop condition or threshold, e.g., maximum cluster size, is met. The merging of two clusters comes with building a shortest link between nodes from the two clusters; thus the MST structure can be maintained in the newly formed cluster. For this example, if the maximum threshold is 4 nodes per cluster, then clusters {6,7} and {5,8} will exchange information due to its shortest distance in the matrix and will merge with each other. The shortest link between them, {7}-{8}, is also captured. Figure 2 demonstrates the MST generated by the TMHDHAC from the network shown in Figure 1. 3.2 Cluster Scheduling In a tree structure with the CH as the root, the data-gathering should start from the leaf node(s) to their parent nodes and all the way to CH eventually. The procedure of the transmission time-slot allocation takes the reverse order and starts from the CH assigning available slots for its child nodes. Then every child node repeats the action for its own child nodes. The iteration stops at leaf node. 10

8

9

5

8

2

7

Y-axis

6

7

5 4

3

3

6

1

4

2 1 0 0

2

4

6

8

10

X-axis

Fig. 2. Cluster formation and intra-cluster MST structure. Dashed circles represents clusters, lines between nodes represent links of MST.

Cluster scheduling is vital to avoid interference in the data transmission phase. Usually in practice, the interference range of a radio signal is considered twice as large as its transmission range [13]. Any other node located in the interference area should not utilize the same time-slot that a pair of nodes use simultaneously. Starting

A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks

377

from the CH, when one node chooses the transmission time-slots for its child nodes, it declares the choice with high power level to cover the interference area. The child node receiving the message also responds with a declaration notifying nodes in its interference range with its assigned time-slot. Nodes locating within the interference range must avoid using the same time-slot for their transmissions. The allocation of time-slot follows the rule of “the one to claim it first is the one to use it”. Nodes do not reuse the same time-slots which it knows that they have been chosen by its neighbor(s) from the time-slot declaration. But nodes are not always able to receive the declarations from the neighbors. For example, it is possible that node A is outside of the interference range of node B but node B is within the interference range of A when node A has a longer transmission distance than node B does. If node B declares its time-slot choices first, but node A does not receive it. Then node A may choose a same time-slot and declare it to neighbors including node B. In this case, node B must inform node A to invalidate A’s time-slot choice and node A will choose a new slot that has not been used by B. Detailed scheduling and maintenance algorithms can be found in [8].

4 Simulation and Results The simulation experiments of several WSN protocols were carried out on NS-2 using randomly generated network topologies. Each WSN contains 100 sensor nodes. A number of experiments have been conducted. The simulation results demonstrated in this section are the average based on the simulation tests of 10 random topologies. Table 3 summarizes the simulation parameters and their values used in our experiments, mostly were obtained from LEACH [6]. Note that the initial energy level (Einitial) is only 0.5 J/node instead of 1 J/node used in LEACH and many other papers due to the longer simulation time needed for TMH-DHAC for each run. With 1J/node, the results for TMH-DHAC will be even better. Table 3. Simulation parameters

Parameter Number of sensor nodes Channel bandwidth Crossover Distance (dcrossover) Transceiver Electronics Energy Dissipation Rate (Eelec)

Value 100 1 Mbps 87 m

Transceiver Amplifier Energy Dissipation Rate (Єfs, Єmp)

10 pJ/bit/m2 0.0013 pJ/bit/m4

Data Fusion Energy Dissipation Rate (Efusion) Data rate Data Packet Size Einitial Spreading Factor of CDMA

50 nJ/bit

5 nJ/bit/signal 3 TDMA frames per 10s 500 Bytes 0.5 J/node 8

378

Y. Jiang, C.-H. Lung, and N. Goel

The following energy dissipation models from LEACH are used for the simulation [6]. Though many other clustering approaches have been reported since LEACH in the literature, LEACH can still be used for indirect comparisons with those approaches that have been compared with LEACH. Receiving energy dissipation. The transceiver energy dissipation rate, Eelec, depends on coding and modulation. The energy consumed by receiving an L bit message is given by E Rx = L × Eelec . Transmission energy dissipation. Depending on if the transmission distance (D) is above a certain threshold, (dcrossover), the multi-path fading model, ETx = L× Eelec + L×ε fs × D2 , or else the free space model, ETx = L× Eelec + L ×ε mp × D4 , is used to calculate the signal attenuation compensation. Computation energy dissipation. Data aggregation and resemblance matrix updating also consumes energy. Ecom = E fusion × SizeSignal × NumberSignal defines the computational al costs of performing data calculation. The performance of a sensor network can be measured by several metrics, e.g., network lifetime, energy efficiency. These two metrics are used for evaluating LEACH, LEACH-C, DHAC-RSS (Received Signal Strength) and TMH-DHAC. 4.1 Different Base Station Locations Figures 3 and 4 illustrate that the network lifetime of the TMH-DHAC algorithm is the longest, while LEACH-C and DHAC-RSS have smaller lifetimes but both are better than LEACH. Let Tn denote the time when nth sensor nodes dies. For instance, Figure 3 depicts T80 for TMH-DHAC is around 600 sec compared with 400 sec for LEACH. At the time T80, a certain amount of nodes are still functioning in the network. Further, TMH-DHAC prolongs the time T20 27.4% from LEACH, 5.5% from LEACH-C, and 6.0% from DHAC-RSS. TMH-DHAC also prolongs T80 42.8% from LEACH, 23% from LEACH-C, 19.9% from DHAC-RSS. Figure 4 shows similar results if the BS is farther away at (0, 150), except the offsets become larger. Specifically, TMH-DHAC prolongs the time T20 by 26.5% from LEACH, 10.8% from LEACH-C, 10.9% from DHAC-RSS. TMH-DHAC also prolongs T80 by 47.0% from LEACH, 29.0% from LEACH-C, 16.8% from DHAC-RSS. Figure 5 compares the proportion of data amount (in packets)/energy dissipation of four protocols at the time 80% of sensor nodes die, T80. In three cases when the base station’s location varies, TMH-DHAC has best energy efficiency and LEACH protocol has the lowest efficiency. When the BS moves from the center (50, 50) to outside of the network (0, 150), energy efficiency reduces in all protocols. Among the four protocols, efficiency in TMH-DHAC reduces the least. The reason is that TMHDHAC adopts a tree-like structure for each cluster. As a result, the distance is reduced and hence TMH-DHAC becomes more energy efficient.

Number of Nodes alive

A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks

120 100 80 60 40 20 0

379

Time(s)

0

200

400 600 LEACH

800 1000 LEACH-C

DHAC-RSS

1200

TMH-DHAC

Number of Nodes alive

Fig. 3. Network lifetime, base station at (50, 50)

120 100 80 60 40 20 0

Time(s)

0

500 LEACH

1000

DHAC

LEACH-C

1500

TMH-DAHC

Fig. 4. Network lifetimes with base station at (0, 150)

4.2 Different Node Densities Table 4 shows the results of three protocols that run in three different sizes of network areas, 100m×100m, 200m×200m, and 300m×300m. Each network still consists of 100 nodes. The BS is at the center of the network for all difference sizes. With the expansion of network area size, the density of nodes becomes lower and the distance between sensors increases generally. It causes higher energy cost of data transmissions regardless of algorithm adopted. Table 4 reveals that all of the algorithms’ total transmission packets drop when the network area enlarges. 400

LEACH

200

LEACH-C DHAC-RSS

0 BS(50,50) BS(0,100) BS(0,150)

TMH-DHAC

Fig. 5. Total data amount (packets)/energy dissipation when 80% nodes die

380

Y. Jiang, C.-H. Lung, and N. Goel Table 4. Total transmitted data amount (packets) with varying network sizes

Sensing Field

Protocol

T1

T25

(0,0)(100,100)

LEACH LEACH-C TMH-DHAC LEACH LEACH-C TMH-DHAC LEACH LEACH-C TMH-DHAC

5792 9608 9640 5396 8200 8549 3571 6996 7660

8836 12019 12278 8357 11462 12178 6568 10211 11962

(0,0)(200,200) (0,0)(300,300)

T50

T75

T100

9767 12775 13746 9018 12152 13684 7388 10876 13471

10043 13118 13889 9463 12556 13753 7781 11276 13553

10219 13402 15660 9651 12855 15379 7952 11525 14580

When sensor density is 100 nodes in a 100m×100m area, TMH-DHAC is 53.2% more than LEACH and 16.8% higher than LEACH-C. When the sensor density becomes 100 nodes in a 200m×200m area, TMH-DHAC still performs better than the other two protocols in all recorded time. In this setting, the performance of total transmitted data amount for TMH-DHAC is 59.4% better than that of LEACH and 19.6% higher than that of LEACH-C. When the network area is even larger at 300m×300m, TMH-DHAC protocol still has the highest transmitted data amount from time T1 to T100 and the offsets become larger compared with those of smaller areas. It exceeds LEACH and LEACH-C by 83.4% and 26.5%, respectively. With respect to the network life, it can be observed from Figures 6 to 8 that among 3 protocols, in all area sizes of 100m×100m to 300m×300m, TMH-DHAC performs the best. Its performance advantage becomes more obvious when sensor density is lower in a larger network. 150 LEACH

100

LEACH-C

50

TMH-DHAC

0 0

500

1000

1500

Fig. 6. Network lifetime for 200m×200m, base station (100,100)

150 LEACH

100

LEACH-C

50

TMH-DHAC

0 0

500

1000

1500

Fig. 7. Network lifetime for 300m×300m. Base Station at (150, 150)

A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks

LEACH

30 0x 30 0

LEACH-C

20 0x 20 0

10 0x 10 0

400 300 200 100 0

381

TMH-DHAC

Fig. 8. Total data amount (packets)/energy dissipation at time T80

4.3 Transmission Distance Generally, tree structure has shorter links between sensor nodes than star structure. As shown in the transmission energy equation in section 3, the energy cost increases when the transmission distance increases. The comparison of the average distance in Figure 9 for different protocols can give us a rough idea as to how these protocols perform.

Fig. 9. Average Transmission Distance (100m×100m)

Figure 9 shows the difference in average transmission distance for four WSN protocols. The LEACH protocol has the largest value, about 23.5m, and LEACH-C has a smaller one, 18.0m, which is 23.4% less than LEACH. DHAC-RSS performs slightly better than LEACH-C and has 16.7m, which is 28.9% less than LEACH. The tree-like structure generates the smallest distance, 6.8m, which is 71% less than LEACH. On the other hand, the quadratic or biquadratic of the distance is used in energy calculation. The comparison of those values can present how the transmission distance affects the transmission cost and the protocols performance. In the simulation, the sensing field is a limited area and most distance between sensor nodes is less than dcrossover(86.2m), so the comparison of quadratic distance is more typical and shown in Figure 10. In Figure 10, the order of the average squared distance of four protocols is the same as that of the distance in Figure 9, but the difference between them is much larger. The average squared link length in LEACH is 663m2, while it is 42.3% smaller in LEACH-C (382.5m2), 48.6% smaller in DHAC-RSS (340.7m2), and 91.3% smaller in TMHDHAC (57.7m2).

382

Y. Jiang, C.-H. Lung, and N. Goel

Fig. 10. Average of Squared transmissions Distance (100m×100m)

Both figures show that tree-like structure has an advantage over LEACH-style cluster structure in terms of transmission distance. The shortened transmission distance in sensor network means energy cost for transmission is reduced and network lifetime is prolonged.

5 Conclusions This paper presented an effective distributed TMH-DHAC clustering algorithm for WSNs. The approach also offers distributed time-slot allocation and scheduling method that enables simultaneous transmissions without interference between different pairs of sensors and reduces the transmission delay of WSNs. Many clustering approaches use the star-like topology comprised of the CH and member nodes. Its long communication distance could be a major factor that limits the energy efficiency in WSNs. TMH-DHAC adopts the multi-hop relay scheme in the intra-cluster data-gathering process which reduces the transmission distance and hence reduces energy consumption. Based on our simulation experiments, TMHDHAC shows much improvement in terms of network lifetime, energy efficiency and total data amount due to its effective multiple-hop transmission tree structure.

Acknowledgements This research was supported in part by MITACS, Canada and Cistel.

References [1] Bajaber, F., Awan, I.: Dynamic/Static Clustering Protocol for Wireless Sensor Network. In: Proc. of the 2nd European Symposium on Computer Modeling and Simulation, pp. 524–529 (2008) [2] Chen, H., Wu, C.S., Chu, Y.S., Cheng, C.C., Tsai, L.K.: Energy Residue Aware (ERA) Clustering Algorithm for Leach-Based Wireless Sensor Networks. In: Proc. of Int. Conf. Sys. & Networks Commu., p. 40 (2007) [3] Chen, X., Yang, Z., Cheng, H.: Unequal Clustering Mechanism of LEACH Protocol for Wireless Sensor Networks. In: Proc. of Comp. Sci. & Info. Eng., WRI World Congress, pp. 258–262 (2009)

A Tree-Based Multiple-Hop Clustering Protocol for Wireless Sensor Networks

383

[4] Ergen, S.C., Varaiya, P.: On Multi-Hop Routing for Energy Efficiency. IEEE Communications Letter 9(10), 880–881 (2005) [5] Fedor, S., Collier, M.: On the Problem of Energy Efficiency for Multi-hop vs. One-hop Routing in Wireless Sensor Networks. In: Proc. of 21st Int’l. Conf. on Advanced Info. Networking and Applications Workshops (2007) [6] Heinzelman, W.B.: An Application-specific Protocol Architecture for Wireless Microsensor Networks. Ph.D. Dissertation, MIT, MA, USA (2000) [7] Hussain, S., Islam, O.: An Energy Efficient Spanning Tree Based Multi-hop Routing in Wireless Sensor Networks. In: Proc. of Wireless Commu. and Networking Conf. (2007) [8] Jiang, Y.: A Tree-based Multiple-Hop Distributed Hierarchical Agglomerative Clustering Protocol for Wireless Sensor Networks. Masters Thesis, Dept of SCE, Carleton University (2009) [9] Al-Karaki, J.N., Kamal, A.E.: Routing Techniques in Wireless Sensor Networks: A Survey. IEEE Wireless Commun. Mag. 11(6) (2004) [10] Lindsey, S., Raghavendra, C.S.: PEGASIS: Power-Efficient Gathering in Sensor Information Systems. In: Proc. Aerospace Conf., pp. 1125–1130 (2002) [11] Lung, C.-H., Zhou, C.: Using Hierarchical Agglomerative Clustering in Wireless Sensor Networks: An Energy-Efficient and Flexible Approach. In: Proc. of IEEE GLOBECOM, pp. 1–5 (2008) [12] Muruganathan, S.D., Ma, D.C.F., Bhasin, R.I., Fapojuwo, A.O.: A Centralized EnergyEfficient Routing Protocol for Wireless Sensor Networks. IEEE Communications Magazine 43(3), 8–13 (2005) [13] Padhye, J., Agarwal, S., Padmanabhan, V.N.: Estimation of Link Interference in Static Multi-hop Wireless Networks. In: Proc. of the Internet Measurement Conf., pp. 305–310 (2005) [14] Romesburg, H.C.: Cluster Analysis for Researchers. Krieger Publishing Company, Malabar (1990) [15] Sim, I., Choi, K., Kwon, K., Lee, J.: Energy Efficient Cluster Header Selection Algorithm in WSN. In: Proc. of Int’l. Conf. on Complex, Intelligent and Software Intensive Systems, pp. 584–587 (2009) [16] Tan, H.O., Korpeoglu, I.: Power Efficient Data Gthering and Aggregation in Wireless Sensor Networks. SIGMOD Rec. 32(4) (2003) [17] Wang, J., Lee, Y.-K.: Determination of Optimal Hop Number forWireless Sensor Networks. In: Gervasi, O., Taniar, D., Murgante, B., Laganà, A., Mun, Y., Gavrilova, M.L. (eds.) Computational Science and Its Applications – ICCSA 2009. LNCS, vol. 5593, pp. 408–418. Springer, Heidelberg (2009) [18] Wu, Y., Fahmy, S., Shroff, N.B.: Energy Efficient Sleep/Wake Scheduling for Multi-hop Sensor Networks: Non-convexity and Approximation Algorithm. In: Proc. of INFOCOM (2007) [19] Younis, O., Fahmy, S.: Distributed Clustering in Ad-hoc Sensor Networks: A Hybrid, Energy-efficient Approach. In: Proc. INFOCOM (2004) [20] Yuan, Y., He, Z., Chen, M.: Virtual MIMO-Based Cross-Layer Design for Wireless Sensor Networks. IEEE Trans. on Vehicular Tech. 55(3) (2006) [21] Zhang, J., Jeong, C.K., Lee, G.Y., Kim, H.J.: Cluster-based Multi-path Routing Algorithm for Multi-hop Wireless Network. International Journal of Future Generation Communications and Networking 1(1), 67–75 (2007) [22] Zhou, C., Lung, C.-H.: Application and Evaluation of Hierarchical Agglomerative Clustering in Wireless Sensor Networks. In: Makki, et al. (eds.) Sensor and Ad-Hoc Networks: Theoretical and Algorithmic Aspects. LNEE, vol. 7, pp. 255–276. Springer, Heidelberg (2008)

Evaluation of Wireless Body Area Sensor Placement for Mobility Support in Healthcare Monitoring Systems Sergio González-Valenzuela1, Min Chen2, and Victor C.M. Leung1 1

Department of Electrical and Computer Engineering The University of British Columbia 2332 Main Mall, Vancouver BC, V6Z1T4, Canada {sergiog,vleung}@ece.ubc.ca 2 School of Computer Science and Engineering Seoul National University Korea, 151-742 [email protected]

Abstract. We present a 2-tier wireless system for healthcare monitoring of convalescing patients in non-critical condition. A network of sensors adhered to the patient’s body used for vital signs collection, and a portable coordinator device forms one tier, whereas a point-to-point link between the coordinator and a fixed access point forms the other tier. We implemented a simple but effective handoff protocol to support uninterrupted monitoring of mobile patients by employing sensor devices featuring limited radio range and low power usage that are amenable for home use. Our experiments reveal that the wrist location is the most favoured for data relaying at walking speeds, as compared to the shoulder, hip, and ankle locations. Additionally, we observed that using sensor nodes as temporary relays may reduce the packet loss rate down to 20% of the value measured when employing a single hop delivery scheme between the coordinator and the access points. Keywords: Wireless Sensor Networks, Wireless Body Area Sensor Networks, Handoff, Performance Evaluation, Healthcare Monitoring.

1 Introduction Wireless Body Area Sensor Networks (WBASNs) continue to draw significant interest as a subcategory of Wireless Sensor Networks (WSNs) that enable untethered vital signs monitoring of people with healthcare problems and/or some form of disability [1]. From the perspective of data communications, the majority of WBASN research deals with Physical Layer and Medium Access Control (MAC) particularities of IEEE 802.15.4 radio technology [2]. This IEEE standard has been widely adopted both in the industry and in the academia, as it provides an excellent scheme for implementing low-cost, low-power sensor devices. On the other hand, miscellaneous issues at the Network Layer and at the Control Plane observe only moderate research activity. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 384–399, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

Evaluation of Wireless Body Area Sensor Placement

385

Advances in the area of WBASN have a special significance, primarily because of their potential for improving the lives of convalescing patients by enabling health monitoring at home, instead of at a hospital [3], [4]. In addition, the economic incentives for making WBASNs practicable for both public and private sectors are also clear. In fact, public healthcare expenses around the world are expected to increase significantly due in good part to the rising population of the elderly, in contrast to a smaller population of younger, working-age people to cover these costs through taxation. This circumstance has been a matter of concern for many years now [5]. Therefore, researching novel technologies that help offset healthcare costs and elevate patients’ quality of life becomes of paramount importance. In this paper, we focus our attention on two particular aspects of a WBASN-based healthcare monitoring system. The first aspect stems from the fact that we consider WBASNs formed by devices that employ IEEE 802.15.4 radio technology in their communications interface, including a coordinator node designated to arbitrate network traffic. Given the nature of the application, this coordinator node forwards the collected sensor data to an external device (using the same radio technology) for subsequent assessment. A problem arises here in that the transmitter’s radio range in sensor nodes and the coordinator is deliberately restrained to conserve battery power. Therefore, multiple, fixed Access Points (APs) are needed in a home setting to ensure that the patient(s) will always be in range of a device through which vital signs information can be forwarded. This circumstance requires incorporating a handoff process that enables uninterrupted communications when a person moves from the coverage area of one AP into another. However, existing research considers only WSNs, whereas WBASN-related provisions are lacking [6], [7]. A second issue arises also because of the same limited transmission power in the WBASN nodes, which has a direct impact on the received signal strength (RSS) at the APs when a person moves about his/her residence (i.e., similar to what occurs when a mobile phone user travels from one site to another). This circumstance can further aggravate if the location of the coordinator device on the patient’s body (e.g., waist, ankle) is not beneficial for relaying data to any given AP. One simple solution is to have the coordinator employ one of the WBASN nodes as a temporary data relay in an attempt to improve the RSS at the AP. Therefore, it becomes important to investigate if and when this approach is suitable, especially since it may leverage the coverage area of APs around the house, translating into fewer hardware devices needed, and the corresponding money savings. We summarize the contributions of this paper as follows: 1. We describe the protocol design for the WBASN handoff process by employing distinct radio channels to leverage system capacity in a multi-user setting. 2. We describe the protocol design used by a WBASN coordinator for finding a sensor that can relay data onto a fixed AP. 3. We present the results in terms of packet loss and coverage area of a 2-tier system implementation with actual sensor devices, including variations in the coordinator and sensors’ positions on a person’s body, and their use as relays. 4. We discuss practical experiences observed during performance evaluations, which can serve as preliminary indicators that can be referenced during the subsequent design of systems with a similar architecture.

386

S. González-Valenzu uela, M. Chen, and V.C.M. Leung

The rest of this paper is i structured as follows. Section 2 lays down the dessign foundations and rationale of o the system that we evaluate. Section 3 explains in deetail the protocol design for WB BASN handoff and data, as well as the signal smoothhing techniques tested. Section n 4 presents the experiments’ setup and performaance evaluations results. Section n 5 provides a discussion of the evaluations and of our practical experiences and ob bservations, and Section 6 concludes this paper.

2 Foundations and Design D Rationale In this section, we describee the working assumptions and foundations for the wirelless elements of an end-to-end d health monitoring system. We adhere to the WBA ASN concept of having one or more m convalescing patients individually wearing a num mber of sensors on their bodies in n order to collect relevant information needed to determ mine their current health status. In I its simplest form, a portable coordinator device colleects vital signs readings from the t WBASN, and forwards them through a fixed AP tto a data processor where it is pre-analyzed (i.e., in a multi-tier fashion [8]). Detection oof a health-related anomaly according a to the embedded algorithms triggers a communications session with a remote monitoring station, whereby qualiffied personnel make a more acccurate assessment, as depicted in Fig. 1. The propoosed system can be deployed in n a regular home or in a nursing home for the eldeerly, allowing a number of patieents to be concurrently monitored in real-time at a sinngle station. In that case, a cond dition assessment station could be placed locally in ordeer to respond quickly to emergen ncy situations (shown as a shadowed element in Fig. 1.) Home mon nitoring

Healthcare provider Data processor

Channel X

Access A points

Remote monitoring

Channel Z

Remote conditio on assessment

WBASN WBASN

Porrtable coord dinator

Local monitoring

Fig. 1. Depiction of a healthcaare monitoring system of convalescing patients at home basedd on WBASNs and multiple channeel usage for improving system capacity.

The main advantage of incorporating i a WBASN into the monitoring system is tthat it allows patients to move freely in their living quarters without wearing wires, as otherwise required by ex xisting commercial products (e.g. [9]). However, this improved, wireless approacch requires several considerations to make it practicabble. From a data communicatio ons perspective, it is clear that data link utilization coould become a problem if multtiple users require that their Electrocardiographic (EC CG) signal be transmitted. Thou ugh the IEEE 802.15.4 standard stipulates a baseline ddata rate of 250 Kbits/sec, in prractice this value degrades significantly. This can be eaasily inferred by the following analysis for a non-beacon enabled MAC scheme.

Evaluation of Wireless Body Area Sensor Placement

387

Considering that the maximum packet size (MAC-level protocol data units or MPDU) is limited to 127 bytes (of which 114 bytes at most are user-defined, as per a 13-byte minimum overhead at the MAC layer), and adding 6 extra bytes of overhead (due to the Start of Frame Delimiter [SHR], and Frame Length [PHR]), the transmission delay of a full frame at the Physical Layer yields: [(127 + 5 + 1) × 8] / (250 × 103) = 4.256 mS. It is also desirable that packets be acknowledged between WBASN devices. Since ACK frames occupy 11 bytes, at 250 Kbits/sec, an ACK frame transmission takes an additional 0.352 mS. Adding the two previous values to a 0.192 mS turnaround delay reserved for the radio to switch from receiving to transmitting mode, and to a predefined 2.368 mS delay ascribed to the CSMA/CA channel access time (with a default backoff exponent of 3) equals to 7.168 mS. Dividing the 114 bytes defined for user payload by this number yields a maximum data rate of 127.2 Kbits/sec, although a much higher packet error rate can be observed for a non-beacon enabled MAC scheme. Moreover, in a multiple user, star-topology scenario, the actual channel utilization degrades significantly as seen in a typical Aloha network, even though more optimistic values obtained through computer simulations have been reported [10]. Given these circumstances, a WBASN-based, ECG monitoring system operating at a minimum sampling rate of 250 Hz [11] in a room with just a handful of patients may become altogether ineffective. In addition to the previous analysis, it is also evident that using a WBASN to continuously transmit vital signs data would most certainly drain the battery of its forming devices, thus defeating the very purpose of the low-power, low-data-rate IEEE 802.14.5 radio scheme. As a result, we argue that a WBASN should be used for healthcare monitoring of patients in non-critical condition that predominantly require follow-up treatment, or for people that otherwise require some form of limited monitoring. Under this assumption, patients’ health can still be permanently monitored by the WBASN hardware, whereas a vital signs digest data can be sporadically forwarded by the corresponding coordinator in order to save bandwidth and battery power. To this effect, a data digest implies that only statistics and overall trends need to be transmitted in compact packets, though counter-arguments exist [8]. We also note that the limited radio range constraint of IEEE 802.15.4 radios can be used to the system’s advantage, so that multiple users can be concurrently monitored at different sites without interfering with each other if each AP is assigned one of the 16 channels available in the 2.4 GHz band, as depicted in Fig. 1. In fact, assuming reduced co-channel interference, and using a strategic channel allocation scheme, a single cell can employ one of the 16 channels for communications between the WBASN coordinator and the AP, whereas the remaining 15 channels can be used for intra-WBASN communications (i.e., one channel per WBASN). Finally, it is evident that the Received Signal Strength Indicator (RSSI) at the APs will be highly variable when users move around. Therefore, we implement and test the performance of three types of filters in order to smooth out the RSSI values acquired through the radio interface of the APs and reduce handoff decision uncertainty. It is evident that the algorithms that implement these filters need to be highly efficient due to the severe hardware constraints found in WBASN coordinators and in cost-efficient APs. We investigate how this RSSI value estimation can be employed to initiate the process of finding a data relay for a WBASN, or to initiate a handoff process.

388

S. González-Valenzuela, M. Chen, and V.C.M. Leung

3 Protocol Design for WBASN Handoff and Data Relay Setup In this section, we explain the signalling process that enables a WBASN coordinator to associate with any given AP within the system’s deployment setting. We describe the corresponding protocols for a coordinator’s association process to an AP, a coordinator/WBASN handoff process from one AP to another, and a relay assessment process that may help reduce the number of lost packets, as explained before. 3.1 Device Association and Handoff Protocol We implemented a 4-way protocol that a WBASN coordinator device follows when first associating with an AP immediately after powering-up, or after a handoff command is received. However, a preliminary step is required to enable a coordinator discover APs within radio range. Since each AP in a deployment setting operates in a different channel, a coordinator scans all the 16 available by issuing a PING_MSG packet to the device with address 0 (pre-assigned to all APs) in each channel. In addition, all devices in the system are programmed to immediately return an ACK packet for every packet received at their radio interface. Therefore, when a coordinator receives the corresponding ACK for the PING_MSG packet issued, it first records its instantaneous RSSI value, and then tunes into the next channel n to conduct the same AP discovery process, as illustrated in Fig. 2. It follows that once the initial discovery phase completes, the coordinator might have registered more than one ACK reply, and so it chooses to associate with the AP that yielded the highest instantaneous RSSI value. At this point, the coordinator issues an ASOC_RQST packet to the respective AP, and waits for the corresponding ACK signal. APs maintain a simple registry for each of the coordinator devices being managed, including a state variable that describes their current association level, which is updated as needed. After an association request is received and an ACK signal is automatically issued, the AP continues the process by sending an ASOC_RPLY packet that conveys the operating parameters that the coordinator will employ (e.g., its WBASN channel, check-in period, etc.). The reason for issuing an ASOC_RPLY packet separate from the ACK packet (instead of merging them) obeys to predefined directives of the IEEE 802.15.4 standard. Finally, the coordinator enters the CONNECTED state immediately after issuing the corresponding ACK for the ASOC_RPLY packet received. At this point, periodic DATA_MSG packet transmissions of the patient’s vital signs information take place. The RSSI from a WBASN coordinator’s transmissions remains relatively stable at the AP as long as the patient remains stationary. However, this value decays rapidly as soon as s/he begins to move away from the coverage area of its current AP. When the RSSI value decays to a certain threshold, the AP issues a HNDF_CMD packet to the WBASN coordinator to initiate the handoff process, as depicted in the middle section of Fig. 2. At this point, it can be seen that the coordinator repeats the initial association process followed during power-up, leading to its association to a new AP located within proximity. Once the WBASN coordinator associates with a new AP, it switches back to its previous operating frequency and issues a FOLLOW_ME command instructing the sensors to retune their radios to the new AP’s channel.

Evaluation of Wireless Body Area Sensor Placement

389

AP discovery phase

Association phase

Data forwarding phase

Handoff phase

Handoff phase

Fig. 2. Signalling protocol for the (re)association and handoff process between a WBASN Coordinator and Access Points, including sensors’ retuning their radios into a new channel.

390

S. González-Valenzuela, M. Chen, and V.C.M. Leung

We refer to the WBASN’s channel retuning as “sensor herding”, whereby the coordinator momentarily switches to the old (but sill in-use) channel to communicate the new WBASN and AP-coordinator link channels, respectively. It can be inferred that sensors associate with their coordinator at power-up, and they operate in the same channel until the coordinator actively herds them into a new one. This process is necessary because the new AP might instruct the use of a new channel for intraWBASN communications in its managed sector, as mentioned in Section 2. 3.2 Relay Assessment Process As mentioned before, a patient’s vital signs digest is continuously forwarded to the AP for as long as the RSSI value of the corresponding DATA_MSG packet remains above a predefined threshold. However, the system may attempt to leverage the RSSI value observed at the AP by instructing the coordinator to forward the DATA_MSG packet through a WBASN node that yields a higher RSSI due to its placement in the patient’s body and its current orientation. Once again, if the RSSI value decays to a warning threshold value, the relay assessment protocol depicted in Fig. 3 initiates.

Fig. 3. Signalling protocol employed to determine whether there is a sensor node with a better RSSI value reading that can relay packets to the Access Point

The process begins by an AP’s sending a RELAY_MSG packet to the WBASN coordinator from which the decaying RSSI signal was received. The coordinator might be unable to get the WBASN devices involved in the process immediately since

Evaluation of Wireless Body Area Sensor Placement

391

they may be operating in low-power mode, which often includes turning off their radios for certain periods of time. When the coordinator deems that the sensors have (temporarily) re-enabled their radios, it forwards the RELAY_MSG packet onto them. At this point, each sensor individually issues a PING_MSG packet to the AP in order to obtain an instantaneous RSSI reading from the corresponding ACK signal. Then, each sensor replies with a RELAY_RPLY packet containing this value back to the WBASN coordinator, which keeps the identity of the sensor that observed a RSSI value over the warning threshold. A successful outcome indicates that at least one of the sensors yielded a better RSSI reading that can be employed to relay communications to the AP. When the next DATA_MSG packet transmission is due, the coordinator simply forwards it to the sensor node that was chosen as data relay, whereas the rest of the sensors can go back to low-power state. The relayed data operation mode can be maintained so long as it provides the best means to maintain a more reliable communication session between the AP and the coordinator (since a low RSSI is a good indicator of a higher probability of having packet errors). It is inferred that the relay node’s low-power duty cycle would need to be modified, so as to participate in both WBASN communications, and as a relay. If the patient is not moving, then the respective sensor will only enter the relay mode occasionally. However, when a patient moves, the RSSI value at the AP might fall once again below a warning threshold. When this happens, the relay assessment process is carried out again, thus providing a “fall-back” mechanism whereby the WBASN coordinator’s transmissions might yield a better RSSI reading than before, so that the relay sensor’s role is relinquished, and direct coordinator-to-AP communications resume. This process can be repeated as necessary until a patient moves far enough from the current AP’s, and the received RSSI value (either with the coordinator or with any sensor nodes acting as relay) no longer stays above the warning threshold level. At this point, the direct coordinator-AP link is kept until a handoff command is received and the corresponding process executes. 3.3 RSSI Estimation at the Access Point We have described the protocol design for association, handoff, and relaying of WBASN devices and APs. As explained before, all of these processes are triggered by the AP when the RSSI value of a coordinator or a relay device fall below either a warning, or a handoff threshold. However, it is a well known fact that RSSI values can vary unpredictably and significantly from one reading to another, especially when a person carrying a radio device moves. IEEE 802.15.4 radio is not immune to this problem [12], and the propagation effects introduced by a person’s body are currently being explored [13]. In our case, the handoff and relay assessment processes can be visibly affected by highly-variable RSSI values at the device’s receivers. Nonetheless, by applying one of several existing filtering techniques, we can obtain a smoothed sequence of RSSI values that can be separately referenced at distinct time periods for making both relay-use assessment, and handoff decisions [14]. In particular, we look at filtering techniques whose algorithmic complexity is amenable to resourceconstrained devices, and ignore other salient techniques that impose a prohibitivelyhigh computation cost (e.g., particle filtering).

392

S. González-Valenzuela, M. Chen, and V.C.M. Leung

Simple Moving Average. This is the simplest technique considered in our approach, whereby the summation of a sequence of values is divided over the number of samples k to obtain an averaged value, as depicted in the next expression: ∑

(1)

The RSSI values v employed in the equation are time dependent, as indicated by the subscript t, and the oldest value is always replaced by the newest one in a shifting fashion. Evidently, a small k value yields an irregular output sequence that is more susceptible to large variations in either of its inputs. Conversely, a larger k value is more immune to large value variations whose effect is shifted a number time steps into future readings. Discrete Kalman. The Kalman filter is one of the most popular and studied filters in signal processing that can yield remarkably accurate results. Its performance has already been evaluated for the case of device handoff in IEEE 802.11 networks [15]. One of the different forms commonly employed to represent the Kalman filter is: x P

u

(2)

P

(3) (4)

x

x

K

z

x

(5) (6)

The Kalman Filter comprises two time-update equations (2), (3), and three measurement update equations (4) - (6). These values are iteratively computed at each time step during RSSI packet measurement, where A, B and H represent the n × n state transition matrix, the 1 × n control input matrix for a particular state x, and the m × n matrix relating to the state of the measurement, respectively. Additionally, Q and R represent the process and measurement noise covariance matrices, respectively. Finally, K represents the Kalman gain, the state estimate, P the covariance estimate, and u the value for the control input. Assuming a stationary RSSI value when a patient is static, the matrices A, H are set to 1, B = 0 (because there is no control signal), and the subscript k is dropped, which greatly simplifies these equations. Exponential Smoothing. This filtering technique is conceptually much simpler than the Kalman filter, and yields values that become weighted averages of future computations. Unlike the Kalman Filter that sports a time-varying gain that adjusts the respective weighs applied to new values, the exponential smoothing approach applies the same weigh to each new value, and does not rely on matrices, making it computationally undemanding. Similarly, the weigh value α used in the corresponding exponential smoothing equation determines the degree of autocorrelation, and thus the effect of large input value fluctuations, as well as the time shifting observed for the output signal: 1

×

×

(7)

Evaluation of Wireless Body Area Sensor Placement

393

4 Practical Evaluatio on Setup and Results 4.1 Experiment Setup Here, we describe the setup for the practical evaluation of our system with actual sennsor devices. Though some pracctical implementations and WBASN experimentations hhave been reported before (e.g., [16]), outcomes of handoff, sensor placement, and ddata relaying tests are yet to be reported. We coded the corresponding protocols using the TinyOS (ver. 2.1) platform [17], [ and employed TelosB sensor nodes [18] to emulatee the WBASN operation. We dep ployed 2 APs: one using a MIB600 board, and another ussing a MIB510 board, both of wh hich had a Micaz node as their radio interface. A total of ffour sensors were secured around d the body of test subject as shown on the left of Fig. 4.

Front view

Front Waist

Wrist

Shoulder

Back 8 mts.

Top view

Ankle

Deployment setting

AP1 6 mts. AP2

Fig. 4. View of the sensor’s placement p on a subject for the experimentation of our propoosed system. The figure on the left shows s a full-front body perspective. The figure on the upper-rright shows a view from top that deepicts the anticipated direction from the sensor radio beams. The figure on the lower-right show ws the walking cycle followed during the testing of our systtem, and the Access Points’ placem ment. The dot indicates the start, and cycle turn-around point.

The placement of the sen nsors obeys to the following rationale. The shoulder areea is a highly plausible spot fo or placing a sensor on patients having their ECG siggnal monitored. Wrist placemen nt is a natural choice because of the number of people w who already wear a battery-pow wered timepiece. The hip/waist area can be consideredd an area in which a WBASN coordinator with larger batteries can be placed withhout causing significant discomffort to a person. Finally, the ankle area has been propoosed by other WBASN researchers as a plausible spot to place an accelerometer sensoor to detect a patient’s movemen nt or current activity. It is evident that this sensor placem ment scheme has a good potential to beaming radio waves particularly to the front andd to the sides of an individual, as seen in the upper-right sketch of Fig. 4. (The wrist and ankle sensors appear shadow wed since they are hidden from a top view perspective).. Because our system is intended for healthcare monitoring at home, we deciided to run our experiments in an actual department dwelling, instead of at a computer llab.

394

S. González-Valenzuela, M. Chen, and V.C.M. Leung

A sketch of the setting appears in the lower-right section of Fig. 4. Experiments were run by having our tests’ subject walk at two different speeds: 0.5 m/s, and 1.0 m/s, which we deem as reasonable walking paces for convalescing people at home. In order to gauge the number of lost packets and the utilization of the sensor nodes when assuming the role of data relays, the placement of the WBASN coordinator was cycled through the available positions (i.e., shoulder, wrist, hip, and ankle). Measurements were taken by having our test subject complete the walking cycle shown in the deployment setting sketch of Fig. 4 20 times – 10 times in each direction to eliminate bias in the results by exposing the WBASN devices to the inner and outer planes of the deployment setting. Otherwise, sensors would be exposed to the effects of indoors multi-path fading observed by walking in a single direction, but not from the corresponding effects observed by walking in the reverse direction. At 0.5 m/s, the walking cycle completes in approximately 1 minute, and at 1.0 m/s, in ½-minute. All nodes transmit packets with the full power available at their radio interfaces (0 dbm). For simplicity, the WBASN coordinator is always set to communicate with the respective AP once per second. All performance measurements collected by the APs are sent to a PC for subsequent processing. 4.2 Experiment Results Our first tests were run to verify the correct system’s operation, and the behaviour of the RSSI filters implemented. Fig. 5 shows sample plots for each the two walking speeds considered. Both raw and filtered signals are depicted for every one of the 3 algorithms described in Section 4 after a single walking cycle when only AP1 was active and no sensors were used, other than the coordinator node placed at the hip level. Fig. 5 (a) and (b) shows sample plots after using the Simple Moving Average algorithm that implements equation (1), which computes the outcome from 10 raw RSSI values collected. The number of chosen samples significantly reduces the magnitude of the variations seen in the RSSI readings, but the filtered sequence is displaced in time to some extent. At 0.5 m/s, AP1 is able to sample a larger number of RSSI values depicting magnitude variations of up to ±20 dBm from one reading to the next, evidencing the presence of a fast-fading channel as observed by the IEEE 802.15.4 radio interface. Conversely, the coarser granularity of a RSSI sample/meter at 1.0 m/s yields readings that omit smaller variations from one value to the next. Fig. 5 (c) and (d) shows the corresponding plots for the Discrete Kalman algorithm with arbitrarily chosen parameters Q = 0.5, and R = 5. Whereas the sample plot for the 0.5 m/s walking speed shows no evident superiority of this algorithm over the Simple Moving Average, a detailed visual inspection of the 1.0 m/s case shows that Discrete Kalman filtering shows better performance in terms of a smaller response to larger variations in the raw RSSI readings, an output value that is less affected by previous variations, and thus higher stability during small variations from one sample to another. However, it is also evident that the effectiveness of the Discrete Kalman algorithm is not being fully exploited here because of the simplification in the values for the state transition matrices A and H, as explained before. Otherwise, distinct transition matrices would be needed for each motion case.

Evaluation of Wireless Body Area Sensor Placement RSSI sample number

RSSI sample number

1 6 11 16 21 26 31 36 41 46 51 56

1

-30

dBm

-50

Raw Filtered (moving average)

-60

-90

31

-60

-80

(b) Moving average - 1.0 m/s

RSSI sample number

RSSI sample number 1

6

11

16

21

26

31

-30 Raw Filtered (discrete Kalman)

-60

Raw Filtered (discrete Kalman)

-40

dBm

dBm

26

-50

1 6 11 16 21 26 31 36 41 46 51 56

-70

-50 -60 -70

-80 -90

-80

(c) Discrete Kalman - 0.5 m/s

(d) Discrete Kalman - 1.0 m/s

RSSI sample number

RSSI sample number

1 6 11 16 21 26 31 36 41 46 51 56

1

Raw Filtered (exponential smoothing)

dBm

dBm

21

(a) Moving average - 0.5 m/s

-30

-20 -30 -40 -50 -60 -70 -80 -90

16

-70

-80

-50

11

Raw Filtered (moving average)

-40

-70

-40

6

-30

dBm

-40

395

(e) Exponential smoothing - 0.5 m/s

-20 -30 -40 -50 -60 -70 -80 -90

6

11

16

21

26

31

Raw Filtered (exponential smoothing)

(f) Exponential smoothing - 1.0 m/s

Fig. 5. Sample performance of three RSSI filters at an Access Point: (a), (b) – Moving average with 10 sampled values; (c), (d) – Discrete Kalman with Q = 0.5, R = 5; (e), (f) – Exponential smoothing with α = 0.5. Raw RSSI plots show markers.

Fig. 5 (c) and (d) shows the corresponding plots for the Exponential Smoothing filtering algorithm with α = 0.5. This value was purposefully chosen to obtain a more irregular signal. Here, the resulting sequence shows negligible time lag and more susceptibility to larger variations in the raw RSSI values. However, the filter is still able to discriminate large variations. Additional test runs with a smaller value for α (e.g., 0.1) yielded a filtered output very similar to the Discrete Kalman case, except that the Exponential Smoothing algorithm is computationally more efficient. This is an important consideration because the division operation in the Microcontroller Unit chip of the TelosB sensor nodes is implemented in software, which has longer execution times, and increased power consumption implications.

S. González-Valenzuela, M. Chen, and V.C.M. Leung

40 35 30 25 20 15 10 5 0

(a) Lost packets @ 0.5 m/s Relayed

Shoulder

Bare

Wrist

Lost packets

Lost packets

396

Hip

Ankle

WBASN coordinator position

40 35 30 25 20 15 10 5 0

(b) Lost packets @ 1.0 m/s Relayed

Shoulder

Wrist

Bare

Hip

Ankle

WBASN coordinator position

Fig. 6. Lost packet performance at: (a) 0.5 m/s, and (b) 1.0 m/s using relayed and bare (coordinator only) WBASN transmissions. The Exponential Smoothing filtering technique was used to estimate RSSI values.

Fig. 6 (a) and (b) depict the number of lost packets as measured by the two APs for the respective walking speeds after the full 20-round walking cycle is completed, as explained in Section 5 (10 cycles in each direction), and all sensor nodes are used. We decided to employ Exponential Smoothing filtering at the APs with α = 0.5. We explicitly chose this parameter’s value for two reasons. First, we wanted that the system to purposefully experience moderate variations in the magnitude of filtered RSSI values, and force it into entering the relay assessment process relatively often once the filtered values hovered around the warning threshold set at -60 dbm. However, the filtered signal should not respond to large variations in raw RSSI values so as to avoid triggering an unnecessarily high number handoff processes that occur when the -70 dbm threshold is crossed. We can see from the 0.5 m/s walking speed tests that, except for case when the WBASN coordinator is placed at the hip/waist level, using sensors as temporary data relays yielded fewer packets lost. In this regard, packet loss was reduced from a 2-fold value in the coordinator-at-the-shoulder case, up to a 5-fold value in the coordinator-at-the-ankle case. The latter case is not surprising, since radio transmissions from a coordinator placed at the ankle (i.e., almost at ground level) can be expected to be decidedly more susceptible to a poor signal reception. However, we consider surprising that fact that using sensors as data relays at a 1.0 m/s walking speed yielded no benefits. Since the actual number of lost packets as observed in both cases barely differs by 1 or 2 (i.e., using sensors vs. using no sensors for data relaying), we deem that there is no significant advantage from one case to another, and so their performance is equivalent. Fig. 7 and Fig. 8 illustrate the data relaying utilization of each sensor as a function of their placement for each of the 4 positions being considered for walking speeds of 0.5 and 1.0 m/s, respectively. For example, after completing the 20-cycle walk at 0.5 m/s and with the WBASN coordinator sensor placed at the shoulder, its utilization as the sole data transmitter spans 61% of the time, whereas the wrist, hip and ankle sensors are utilized 28%, 6% and 5% of the time, respectively, as shown in Fig. 7 (a), as per the combined results reported by both AP1 and AP2. We can also see that the WBASN coordinator always yields the larger utilization of the 4 sensors. On the other hand, the ankle sensor always has the lowest utilization among them all.

100 90 80 70 60 50 40 30 20 10 0

(a) Relay utilization - Shoulder coord.

Utilization

Utilization

Evaluation of Wireless Body Area Sensor Placement

61 28

Shoulder [C]

Wrist

6

5

Hip

Ankle

100 90 80 70 60 50 40 30 20 10 0

(b) Relay utilization - Wrist coord. 64

17

Shoulder Wrist [C]

(c) Relay utilization - Hip coord. 62

21

14

3 Shoulder

Wrist

Hip [C]

WBASN sensor position

12

7

Hip

Ankle

WBASN sensor position

Utilization

Utilization

WBASN sensor position

100 90 80 70 60 50 40 30 20 10 0

397

Ankle

100 90 80 70 60 50 40 30 20 10 0

(d) Relay utilization - Ankle coord.

56

19

15

Shoulder

Wrist

10

Hip

Ankle [C]

WBASN sensor position

Fig. 7. Relay utilization at 0.5 m/s for each respective sensor and the corresponding coordinator position as described

Fig. 8. Relay utilization at 1.0 m/s for each respective sensor and the corresponding coordinator position, as described

398

S. González-Valenzuela, M. Chen, and V.C.M. Leung

5 Discussion We briefly elaborate on important aspects observed in the previous results. First, we expected to see a higher shoulder sensor utilization as a relay, since its position gave it an advantage over the rest of the sensors in situations where the latter would be blocked by different objects (e.g., furniture), whereas the former should benefit from a direct line-of-sight with either AP, except for a few cases. However, this was not the case, which brings us to the second observation: on average, the wrist sensor was the second-most utilized device as a relay, regardless of the coordinator’s placement, the subject’s walking speed, and the fact that the wrist experiences a 2-degree freedom of movement inherited from the patient’s own displacement and his/her arm’s inertial angular motion (though minor). Finally, the utilization of the coordinator in the ankle position as a percentage of the rest surpassed the one observed for the coordinator in the hip and wrist locations at 1.0 m/s walking speed, as evidenced in Fig. 8. The previous observations have various implications. For instance, given that the wrist location favours the coordinator role of a device when walking at 0.5 m/s, placing a WBASN coordinator inside a timepiece would drain its battery quicker than the user would hope for, thus causing an inconvenience. However, this is not the case for a walking speed of 1.0 m/s. Additionally, our experiments depict results of a patient in motion from one home location to another only. Additional experiments (not reported in this paper) indicated that the coordinator’s transmissions can degrade significantly even when a patient is static (e.g., while seating at a sofa watching TV). To deal with this type of issues, it can be reasonably argued that the coordinator might need to relay data through a different, off-WBASN device located near him/her, which motivates further research.

6 Conclusions We designed and tested the performance of a handoff and data relay system that enables patient mobility for healthcare monitoring in home environments. Preliminary results indicate that using sensors in a WBASN on a temporary basis can help reduce the number of lost packets in certain circumstances. We learned that the actual placement of the WBASN coordinator has a direct effect on the system’s performance, and that the wrist position was amply favoured for using as a temporary relay regardless of the patient’s speed. Conversely, the sensor favoured with a line-ofsight to the access points does not necessarily yield the best reception. We believe that our investigations provide sufficient motivation for further research in this area.

Acknowledgements This project was supported by the National Sciences and Engineering Research Council of the Canadian Government under grants STPGP 322208-05 and 365208-08. In addition, this work was supported in part by NAP of Korea Research Council of Fundamental Science & Technology.

Evaluation of Wireless Body Area Sensor Placement

399

References [1] Cao, H., Leung, V.C.M., Chow, C., Chan, H.: Enabling Technologies for Wireless Body Area Networks: A Survey and Outlook. IEEE Communications Magazine 47(12), 84–93 (2009) [2] The IEEE 802.15.4 Radio Standard, http://www.ieee802.org/15/pub/TG4.html [3] Kulkarni, P., Öztürk, Y.: Requirements and Design Spaces of Mobile Medical Care. ACM SIGMOBILE Mobile Computing and Communications Review 11(3), 12–30 (2007) [4] Varshney, U.: Pervasive Healthcare and Wireless Health Monitoring. Mobile Networks and Applications (MONET) 12(2-3), 113–127 (2007) [5] Kinsella, K.G., Velkof, V.A.: An Aging World: 2001. International Population Reports, U.S. Department of Health and Human Services, U.S. Department of Commerce, United States Government (2001) [6] Walsh, M., Mahdi-Alavi, S.M., Hayes, M.: A Modified Bumpless Transfer Technique For Seamless Handoff In Mobile Ad-Hoc 802.15.4 Wireless Sensor Networks. In: Proceedings of the IET Irish Signals and Systems Conference (ISSC), Galway, Ireland, June 18-19 (2008) [7] Boudriga, N., Baghdadi, M., Obaidat, M.S.: A New Scheme for Mobility, Sensing, and Security Management in Wireless Ad Hoc Sensor Networks. In: Proceedings of 39th Annual Simulation Symposium (ANSS), Huntsville, Alabama, USA, April 2-6 (2006) [8] Misic, J., Misic, V.: Bridge Performance in a Multitier Wireless Network for Healthcare Monitoring. IEEE Wireless Communications 17(1), 90–95 (2010) [9] The MHM 100 Heart Monitor by Medick, http://www.medick.com [10] Li, C., Li, H.-B., Kohno, R.: Performance Evaluation of IEEE 802.15.4 for Wireless Body Area Network (WBAN). In: Proceedings of the First International Workshop on Medical Applications Networking (IEEE ICC Workshops), Dresden, Germany (June 14, 2009) [11] Pizzuti, G.P., Cifaldi, S., Nolfe, G.: Digital Sampling Rate and ECG Analysis. Journal of Biomedical Engineering 7(3), 247–250 (1985) [12] Ilyas, M.U., Radha, H.: Measuring Packet-level Memory Length of 802.15.4 Wireless Channels with Relative Mutual Information. In: Proceedings of the IEEE International Conference on Communications (ICC), Beijing, China, May 19-23 (2008) [13] Miluzzo, E., Zheng, X., Fodor, K., Campbell, A.T.: Radio Characterization of 802.15.4 and its Impact on the Design of Mobile Sensor Networks. In: Verdone, R. (ed.) EWSN 2008. LNCS, vol. 4913, pp. 171–188. Springer, Heidelberg (2008) [14] Srinivasan, K., Levis, P.: RSSI is Under Appreciated. In: Proceedings of the 3rdWorkshop on Embedded Networked Sensors, Cambridge, USA, May 30-31 (2006) [15] Bellavista, P., Corradi, A., Giannelli, C.: Evaluating Filtering Strategies for Decentralized Handover Prediction in the Wireless Internet. In: Proceedings of the 11th IEEE Symposium on Computers and Communications, Cagliari, Italy, June 26-29 (2006) [16] Jovanov, E., Milenkovic, A., Otto, C., De Groen, P., Johnson, B., Warren, S., Taibi, G.: A WBAN System for Ambulatory Monitoring of Physical Activity and Health Status: Applications and Challenges. In: Proceedings of the 27th Annual IEEE Conference on Engineering in Medicine and Biology, Shanghai, China, September 1-4 (2005) [17] TinyOS for Deeply Embedded Sensor Programming, http://www.tinyos.net [18] Crossbow Technology, http://www.xbow.com

Optimal Relay Node Placement and Trajectory Computation in Sensor Networks with Mobile Data Collector Ataul Bari, Fangyun Luo, Will Froese, and Arunita Jaekel University of Windsor, Windsor, ON, N9B 3P4, Canada {bari1,luoh,froesew,arunita}@uwindsor.ca

Abstract. Most sensor network architectures typically assume that nodes are stationary after deployment. However, a number of recent papers have shown that the use of mobile nodes or mobile data collectors (MDC) can significantly improve the performance of a network. In this model, the network can be viewed as a three-tier architecture, where the lowest-tier consists of a set of sensor nodes. The middle-tier contains a number of higher powered relay nodes, each acting as a cluster head for a number of sensor nodes in the tier below, and one or more mobile data collector(s), constitute the uppertier. For such hierarchical architectures, there are a number of important design problems such as determining the number of relay nodes that are needed and their locations, determining the appropriate buffer capacities in the relay nodes to ensure there is no data loss due to buffer overflow and calculating a suitable trajectory for each MDC. In this paper, we first propose an integrated integer linear program (ILP) formulation that calculates the optimal number and positions of the relay nodes in the middle-tier, along with the requisite buffer sizes. We then present an algorithm for calculating the trajectory of the MDC, based on the relay node locations and the load on each individual relay node, in a way that minimizes the maximum energy dissipation of the relay nodes. Experimental results demonstrate that our approach is feasible for networks with hundreds of sensor nodes.

1

Introduction

The performance benefits of hierarchical sensor networks [1], [6] has been well investigated in the literature. In such networks, sensor nodes are grouped into clusters and transmit their data to their respective cluster heads. The cluster heads collect the data from all nodes in their own cluster and transmit their data to the base station(s), using an appropriate routing scheme. Since the cluster heads are required to transmit large amounts of data over longer distances, compared to individual sensor nodes, the use of specialized nodes for cluster heads has gained considerable support in recent years [1], [3], [5], [9]. These specialized nodes, often called relay nodes, are typically equipped with enhanced capabilities in terms of energy provisioning, buffer capacities and transmission 

A. Jaekel has been supported by discovery grants from the Natural Sciences and Engineering Research Council of Canada.

J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 400–415, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Optimal Relay Node Placement and Trajectory Computation

401

ranges. The resulting architecture, where a large number of low-power, sensor nodes with limited capabilities form the lower-tier and relatively fewer relay nodes with enhanced capabilities form the upper-tier, has been shown to improve network performance in a number of areas including network lifetime, loadbalanced routing and fault-tolerance [1] – [7]. For the above two-tier architecture, the lifetime of the network is primarily determined by the lifetime of the upper-tier relay node network. Each relay node is responsible for receiving (and possibly aggregating) the data from all sensor nodes in its cluster and then transmitting the data to (or towards) the base station, using either single-hop or multi-hop paths [6], [8], [9]. The energy dissipation of the relay nodes increases rapidly with the distance between the sender and the receiver (either another relay node or the base station), and has a significant impact on the lifetime of the network. A number of energy-aware routing strategies have been proposed to extend the lifetime of the relay node network [6], [8], [9]. However, such strategies are of limited use for relay nodes that are far away from other nodes and must therefore transmit over a large distance, or for nodes near the base station that must transmit data from many other nodes, in case of multi-hop routing. In fact, for sparse networks it is even possible that no feasible routing scheme exists, since the distance to the nearest neighbor may be greater than the transmission distance of a node. A number of recent papers have shown that the use of some mobile nodes or mobile data collectors (MDC) can significantly improve the performance of a network in terms of lifetime, coverage, and connectivity [15], and techniques for effectively utilizing the unique capabilities of mobile nodes have been attracting increasing research attention in the past few years [19] – [25]. In this paper, we consider a new network model that extends the traditional two-tier architecture, by adding a third tier consisting of one (or possibly more) mobile data collector(s) (MDC), above the relay node network (which now constitutes the middle-tier). The MDC, which is not power constrained, visits all relay nodes in the middletier, following a fixed trajectory [15], collects data from them, and delivers the collected data to the base station. Thus, the relay nodes are relieved from the burden of “routing” data towards the base station, possibly over long distances, resulting in considerable energy savings at these nodes. In such a three-tier architecture, we assume that the lower-tier sensor nodes have already been deployed, and the number and locations of these sensor nodes have been determined by the monitoring needs of the specific application. Therefore, our goal is to design the middle and the upper tiers of the network. In this context, we address the following design problems in this paper: i) Find the locations of the relay nodes constituting the middle-tier, such that each sensor node is covered by at least one relay node (i.e. there is at least one relay node within the transmission range of each sensor node), and the number of the relay nodes is minimized. This is the relay node placement problem. ii) Determine the set of sensor nodes assigned to the cluster of each relay node, such that the overall buffer requirements for the relay nodes is minimized.

402

A. Bari et al.

iii) Determine a suitable trajectory for the MDC(s), such that the energy dissipation rate of the relay nodes is minimized (or is below a specified level). Restricting the energy dissipation allows the network to remain functional for a specified minimum period of time. In this paper we first propose an integer linear program (ILP) formulation that, given a set of potential locations of relay nodes in a network, optimally solves the relay node placement problem. Our formulation also computes the buffer requirements for the relay nodes, so that data generated by the lower-tier sensor nodes can be stored at the relay nodes and delivered to the MDC without any loss of data (i.e. without buffer overflow). We also provide a modification of our ILP that, given the maximum number of relay nodes to be used in the middle-tier, finds the locations of the relay nodes and minimizes their buffer requirements. Once the locations of the relay nodes has been determined, we present an algorithm for calculating the trajectory of the MDC, such that the energy dissipation of the relay nodes due to data transmission is minimized. The remainder of the paper is organized as follows. In section 2, we review the relevant work. In sections 3 and 4, we present our ILP formulations and algorithm for trajectory computation respectively. We discuss our experimental results in section 5 and conclude in section 6.

2 2.1

Review Relay Nodes Placement, Clustering and Routing

The relay node placement problem deals with placing a minimum number of relay node such that each sensor node can send its data to at least one relay node. This is an NP-hard problem [3], and has been widely investigated in the literature [3], [7], [10], [11], [13] and [14]. In [3], the authors solve the placement problem by dividing the entire region into cells, finding an optimal solution for each cell, and then combining the solutions. They consider single and double connectivity problem, both at the sensor level, as well as, at the relay node network level. An approximation algorithm to achieve single and double connected network is proposed in [10]. In [11], the authors propose a two-step approximation algorithm to obtain 1-connected (in the first step) and 2-connected (in the second step, by adding extra back-up nodes to the result of the first step) sensor and relay node network. A mixed-integer non-linear program, and a heuristic algorithm in proposed in [7], that focuses on prolonging the lifetime of sensor networks by deploying relay nodes within the networks. The network connectivity and the fault-tolerance is also addressed in [13], where the authors propose an ILP-based approach for the appropriate placement of the relay nodes in a sensor network. This approach optimally selects the minimum number of relay nodes from a given set of “potential positions” of the relay nodes, obtained from the area of the network deployment. Noting that the number of potential positions in a real plane can be infinite, the authors use a heuristic to limit such locations to a level so that the ILP becomes

Optimal Relay Node Placement and Trajectory Computation

403

computationally tractable. This heuristic views the entire networking area as an imaginary grid, and the potential relay node locations are taken at the center and the corner positions of each cell. The authors shows that such an approach provides good solutions, especially when the network area is small and the sensor nodes are densely deployed. They also shows that a finer grid leads to a better solution. The clustering and the routing schemes, for two-tiered networks, are addressed in a number of recent papers, including those in [1], [2], [6], [9], [12]. In [1], the authors focus on load balanced clustering and propose a heuristic solution for the problem. Fault tolerant clustering is addressed in [2], and in [12], the authors investigate the problem of maximizing network lifetime by appropriately placing nodes, which are not energy constrained (e.g., connected to a wall outlet). A Tenet architecture for tiered sensor networks is proposed in [4] that simplifies application development and reuses mote-tier software. In [5], [6] and [9], the authors focus on clustering and routing schemes that maximize the network lifetime. These approaches do not consider the problem of minimizing the number of relay nodes and finding their locations. 2.2

Mobility in Sensor Networks

Improving the performance of sensor networks, exploiting some mobile entities, has been investigated in [19] – [25]. In [19], multiple mobile base stations, which might be required to move periodically, is considered to prolong the network lifetime. The authors propose an ILP formulation to compute the new locations of the mobile base stations. In [20], multiple mobile base stations, which move in parallel straight paths, are considered, and is shown that such mobility can achieve scalability and load-balancing. In [26], the authors focus on optimal data collection, and propose a protocol that extends the lifetime of the network. Their approach takes into account both the base station mobility and the multi-hop routing. In [22], the authors focus on fault-tolerance. They propose a scheme that uses K-means clustering and TSP-based shortest path approaches to achieve resiliency. In [23], a three-tiered network is considered where mobile data collectors, lie in the middle-tier, move randomly within the network and pick-up data from the sensor nodes. A sensor node transmits data only when a MDC enters the direct communication range of the node. In [24], the authors propose a queuing theory based mathematical model that analyzes the performance and trade-offs of the three-tier architecture. Partitioning Based Scheduling (PBS) heuristic that computes the trajectory of MDC is used in [25]. The focus here is to reduce sensor buffer overflow. They propose a solution that has two parts. In the first part, nodes are partitioned into groups based on their locations and the data generation rates. Then, a node visiting schedule is generated within a group. Finally, the solutions of these groups are combined to obtain the final path of the MDC. The problem of delivering data to multiple mobile sinks is addressed in [26], [27] and [28]. A hierarchical sensor network architecture, using higher powered relay nodes as cluster heads, which utilizes a MDC to collect data from the cluster heads has

404

A. Bari et al.

been discussed in [17], and [18]. The objective, in both works, is to extend the overall lifetime of the network. In [17], an ILP formulation is proposed, which optimally selects the relay nodes, from a set of potential relay nodes positions, such that i) the number of relay nodes are minimal, and ii) the length of the trajectory of the MDC is as short as possible. In [18], the authors focus on reducing the length of the trajectory of the MDC, by allowing the MDC to visit the neighborhood of each relay nodes, instead of visiting their exact locations. The authors propose two heuristic solutions for the problem. 2.3

Network Power Model

The power needed for data communication is the dominant factor in power consumption in sensor networks. We have considered the first-order radio model [8] to account for the energy consumption due to communication where receive (transmit) circuitry consumes α1 nJ/bit (α2 nJ/bit) of energy. The total energy to receive b bits is given by, ERx (b) = α1 b while the total energy needed to transmit b bits over a distance d is given by ETx (b, d) = α2 b + βbdq , where q is the path loss exponent, 2 ≤ q ≤ 4 [8] and β is the amplifier energy to transmit unit bit of data over unit distance. In our experiments, we have used α1 = α2 = 50nJ/bit, β = 100pJ/bit/m2 and the value of path-loss exponent, q = 2.

3

Network Design Exploiting Node Mobility

3.1

Network Model

We consider a three-tiered wireless sensor network, where the lower-tier consists of a set S of n sensor nodes, i.e., |S| = n. We assume that the deployment of the sensor nodes ensures the appropriate coverage of the sensing area. We consider a set R of m potential locations of relay nodes, i.e., |R| = m. A subset R of R will constitute the middle-tier network. Each element of R would act as a cluster head. We also consider a MDC, lying in the upper-tier of the network. The MDC visits all relay nodes in R, collects their data, and delivers the data to the base station. Each relay node has a buffer size B. We assign, to each node, a unique label as follows: i) for each sensor node, a label i, 1 ≤ i ≤ n, ii) for each possible location of the relay nodes, a label j, n + 1 ≤ j ≤ n + m, and iii) for the MDC, a label n + m + 1. If a sensor node i ∈ S is covered by a relay node at location j (we shall refer to such relay node as relay node j), then i can transmit its data directly to j. A sensor node i may be covered by more than one relay node, however, our objective is to design the relay node network such that each sensor node belongs to exactly one cluster, C j , corresponding to a relay node j. Our proposed formulation

Optimal Relay Node Placement and Trajectory Computation

405

i) determines the minimum number and the positions of the relay nodes, which will act as the cluster heads, and form the middle-tier network, and ii) assign sensor nodes to clusters such that the relay nodes buffer requirements are minimized. We assume that the positions of the sensor nodes are known (or can be determined, e.g. using GPS), and the relay nodes can be placed at the computed locations. This approach is feasible for many applications (e.g., monitoring industrial environments, road condition, and habitat). We have used a grid based approach [13] to generate R. However, our ILP formulation does not depend on how R is generated and other approaches such as approach given in [3] can easily be used. We consider that a sensor node i ∈ S generates data at a rate of bi bits per unit time, and transmits to the corresponding cluster head. The value of bi , ∀i ∈ S can either be the same, or may vary. In our model, each relay node j, receives data from the sensor nodes belonging to its own cluster C j , and buffers them until j can transmit buffered data to the MDC, while it is visiting j. Data buffering is essential for applications where it is important not to lose any data generated by the sensor nodes. In our model, the MDC visits each relay node j, periodically, at fixed time intervals. Once transmitted, the buffer of j is cleared so that the buffer can be reused to store data until the next visit by the MDC. A relay node j transmits its buffered data only when the MDC is closest to j, in its trajectory. The MDC traverses at a constant speed following a predetermined trajectory, and it needs Tr time units to complete the trajectory. That is, the time interval between any two successive visits by a MDC to a relay node j is known and is equal to Tr . 3.2

Notation Used

In our formulation we are given the following data as input: • n: The total number of sensor nodes, with each sensor node having a unique label i, 1 ≤ i ≤ n. • m: The total number of possible positions of relay nodes, each position having a unique label j, n + 1 ≤ j ≤ n + m. • j: The relay node at location j, n + 1 ≤ j ≤ n + m. • n + m + 1: The label of the MDC. • rmax : The transmission range of each sensor node. • di,j : The Euclidean distance between node i and node j. • bi : Number of bits generated by sensor node i in unit time. • C j : The set of sensor nodes belonging to the cluster of relay node j. • W1 , W2 : Positive constants. • Tr : Time required by the MDC between two successive visits at any relay node j. • ymax : Maximum allowable number of relay nodes.

406

A. Bari et al.

We also define the following variables: variable defined as follows: • Xi,j : Binary ⎧ ⎨ 1 if the sensor node i selects relay node j as its cluster head, Xi,j = ⎩ 0 otherwise. variable defined as follows: • Yj : Binary ⎧ ⎨ 1 if relay node at location j is included in the middle-tier network, Yj = ⎩ 0 otherwise. • Rj : Continuous variable indicating the total number of bits generated (during the period Tr ) by the sensor nodes belonging to the cluster C j . of the relay node j. • Bmax : A Continuous variable such that Rj ≤ Bmax , ∀j, n+1 ≤ j ≤ n+m. 3.3

ILP Formulation for Minimizing the Number of Relay Nodes

In this section, we propose our ILP formulation. Our formulation i) ensures that each sensor node is covered by at least one relay node, ii) minimizes the number of relay nodes, and iii) minimizes the maximum buffer capacity of any relay node j, which is included in the middle-tier, as a secondary objective. Using the notations discussed in section 3.2, we present our formulation as follows: Minimize W1 ·

n+m 

Yj + W2 · Bmax

(1)

j=n+1

Subject to: a) A sensor node i can transmit data to a relay node j, only if the distance between i and j is less than the transmission range rmax of the sensor node i. Xi,j · di,j ≤ rmax

∀i, 1 ≤ i ≤ n, ∀j, n + 1 ≤ j ≤ n + m

(2)

b) The relay node at location j must be included in the middle-tier network, if it is selected as a possible cluster head by at least one sensor node i. Yj ≥ Xi,j

∀i, 1 ≤ i ≤ n,

(3)

∀j, n + 1 ≤ j ≤ n + m c) A sensor node transmits data to exactly one relay node. n+m  j=n+1

Xi,j = 1

∀i, 1 ≤ i ≤ n

(4)

Optimal Relay Node Placement and Trajectory Computation

407

d) Compute the total number of bits buffered at the relay node j during the interval Tr . Rj = T r ·

n 

bi · Xi,j

i=1

∀i, 1 ≤ i ≤ n,

(5)

∀j, n + 1 ≤ j ≤ n + m

e) The total amount of data to be buffered by any relay node in one round of data gathering cannot exceed Bmax . Rj ≤ Bmax 3.4

∀j, n + 1 ≤ j ≤ n + m

(6)

Justification of the ILP Equations

Equation (1) is the objective function for the formulation, and consists of two terms. The primary goal (represented by the first term) is to minimize the total number of relay nodes used to form the middle-tier network. As mentioned earlier, a relay node j is included in the middle-tier network (i.e., Yj = 1), only if j selected as a cluster head by at least one sensor node i. Therefore, by counting the number of relay nodes selected to be the cluster heads, we can determine the number of relay nodes needed in the middle-tier network. This is exactly the value calculated by the first term in the objective function. The second term is used to minimize the maximum buffer capacity of the relay nodes, which is the secondary objectives. By choosing appropriate values for W1 and W2 , we can select the relative importance of the two objectives being minimized. For example, if we set W2 = 0, then the only parameter we are interested in minimizing is the number of relay nodes. a) Constraint (2) specifies that a sensor node can communicate with a relay node j, only if j is within the transmission range of the sensor node. b) Constraint (3) ensures that if a relay node j is chosen as a cluster head by at least one sensor node, then j must be included in the middle-tier network. If a relay node j is not selected to be a cluster head by any sensor node, normally it should not be selected. This is not specifically enforced by any constraint, but is taken care of by the objective function, which will set Yj = 0, if this does not violate any other constraints. c) Constraint (4) requires that each sensor node belongs to exactly one cluster and transmits data to the corresponding cluster head. d) If a relay node j is selected to be included in the middle-tier and a sensor node i belongs to its cluster C j , then Xi,j = 1. Constraint (5) thus calculates the total number of bits, Rj , buffered in j during the interval Tr , by summing the data transmitted to it from all the sensor nodes belonging to the cluster C j and then multiplying this value by the interval Tr . e) Constraint (6) ensures that if a relay node j is selected, then the total bits buffered at j during the interval Tr do not exceed the buffer size Bmax . The left hand side of constraint (6) is the total amount of bits buffered at j during the interval Tr . The right hand side Bmax , of constraint (6), must be greater

408

A. Bari et al.

than or equal to the maximum of the amount of bits buffered by the relay nodes. Since the objective function is to minimize Bmax , constraint (6) forces Bmax to be the maximum buffer capacity required by any relay node. 3.5

ILP Formulation for Minimizing the Buffer Size of Relay Nodes

The ILP formulation given in Section 3.3 minimizes the total number of relay nodes required to form the middle-tier network, and considers the buffer size as a secondary objective. This gives the lower bound for the number of relay nodes. However, it some cases, the total number of relay nodes may be given, and the problem is to find out the placement of the given number of relay nodes such that the maximum size of the buffer is minimized. This can be easily achieved by the following modification of the ILP formulation, given in Section 3.3. Minimize Bmax

(7)

Subject to: a) - e) Constraint (2) - Constraint (6). b) Maximum number of relay nodes cannot exceeds ymax . n+m 

Yj ≤ ymax

∀j, n + 1 ≤ j ≤ n + m

(8)

j=n+1

Equation (7) is the objective function that minimizes the maximum buffer capacity requirement of any relay node. Constraint (8) enforces the limit on the maximum number of relay nodes.

4

Computation of Trajectory

In this section, we present a heuristic approach to compute a trajectory for the MDC, such that the maximum energy dissipated by any relay node is minimized. A number of papers have considered the use of complex trajectories, where the MDC visits each node individually [25], [17], [18]. However, in this paper the goal is to use a very simple trajectory that can be easily traversed by the MDC. So, we consider the case where the MDC travels back and forth along a straight line. A straight line trajectory has been shown to be a practical and useful option for mobile nodes in [20]. Given the positions and the expected loads of the relay nodes, our approach finds a straight line, to be used by the MDC as the path to collect the data from the relay nodes, such that the energy dissipation by the relay nodes are minimized. We note that the energy minimization can be accounted as the sum of the energy dissipated by the relay nodes. In such case, any standard weightedregression analysis method, using the communication cost as discussed in Section 2.3, can be applied to compute the best fitting trajectory. However, in our network model, if any relay node depletes power, then all sensor nodes belong to the relay node become inaccessible, and the network may fail to meet the reliability standard. In this case, to extend the lifetime, it is important to minimize the

Optimal Relay Node Placement and Trajectory Computation

409

y

y' p

p dp

y = c'

Initial trajectory dq

x

puted

Recom

y = c'

q

ry

trajecto

x'

q N

N E

E

(a)

(b)

Fig. 1. Computation of the (a) initial trajectory, (b) improved trajectory by rotating the orientation of the axis

maximum energy dissipated by any relay node in the network. We propose a simple approach that can be used to compute a suitable straight line trajectory. Let R be the set of relay nodes, where each relay node is given an unique label j, 1 ≤ j ≤ |R|. Also, let the coordinate (jx , jy ) specify the position of relay node j, ∀j ∈ R. We start by assuming that, given the area of the network, the trajectory is a horizontal line. Let the line be given by the equation y = c, where c is a constant. We set the initial value of c as the midpoint of the ycoordinate values of the uppermost and the lowermost relay nodes in the network. Since a relay node j transmits to the MDC when the MDC is closest to j, the transmission distance of j is the vertical distance (i.e., y-axis distance) of j, projected on the trajectory line y = c. Using this initial trajectory, we find the relay node p (q) that dissipates the maximum amount of energy, among all nodes located in the above (below) the initial line. To minimize the maximum energy of the relay nodes, we need to find a new value c , qy ≤ c ≤ py , for the constant c, so that the energy dissipation of nodes p and q is balanced (Fig. 1(a)). We achieve this by setting the energy dissipation of nodes p and q (computed using the model discussed in Section 2.3), corresponding to the trajectory y = c , as equal. Let the vertical distance of the node p and the node q, from the new trajectory be dp and dq , respectively. Also, let the vertical distance between nodes p and q be λ. Then, using the equation given in Section 2.3, and the notation given in Section 3.2, we have: d2p − γd2q = ξ (9) dp + dq = λ R

(10)

Where γ = Rpq , and ξ = β1 (α1 + α2 )(γ − 1). The values of α1 , α2 and β are obtained from network power model discussed in Section 2.3. We obtain the new value of c (= qy + dq ) by solving the above two equations. Using the new line, we recompute p and q, and apply the process, in an iterative manner, until the difference between the energy dissipation by p and q becomes less than a small preset value.

410

A. Bari et al.

Based on the actual layout of the networking area and the distribution of the sensor nodes, a different orientation, rather than strictly horizontal, for the trajectory may be beneficial. Once we obtain the initial trajectory, we compute the best orientation of the trajectory by rotating it in the range 0◦ − 180◦ . We rotate the line by a small angle ψ, at a time, and get a new orientation, as shown in Fig. 1(b). At each orientation, we recompute the value of Emax , using the approach described above. After the rotation is complete, we select the orientation that gives the minimum among all orientations. As shown in the Section 5.2, this rotation can substantially improve the solution, based on the actual layout of the network.

5 5.1

Results Simulation of ILP Formulation

In this section, we present the simulation results of our formulation for selecting the relay nodes in the middle-tier of the network. We have conducted different sets of experiments by setting different values for the parameters in our formulation. In the first set of experiments, our objective is to jointly optimize the number of relay nodes required to form the middle-tier relay node network, and the maximum buffer requirement of the relay nodes. The relative importance of each term is determined by the value of the constant W1 and W2 , used in equation (1). Since our primary goal is to minimize the number of relay nodes, while the secondary objective is to reduce the buffer requirement of each node, we set W1 = 8000 and W2 = 0.1 for our simulations. We have used an experimental setup, where the sensor nodes are randomly distributed over a 200 × 280m2 area. We have assumed that rmax = 40m. The results are obtained by CPLEX 9.1 solver. We have simulated our scheme with different number of sensor nodes, ranging from 100—600. For each size of the sensor node network, we randomly generate five different sets for the locations of the sensor nodes in the network, and compute the results using each set. The results reported in the tables and figures in this section reflect the averages of all the different runs for each network size. As in [13], we have used a grid based approach to compute the initial potential positions of the relay nodes. The number of potential relay node locations were set to 48 (for coarse grid) and to 165 (for fine grid), indicated as 48-Grid and 165-Grid respectively in the following discussions of our results. We have also assumed that each sensor node generates data at a rate of 100 bits/unit-time, i.e., bi = 100, ∀i, 1 ≤ i ≤ n and Tr = 5. Fig. 2 shows the number of relay nodes needed in the middle-tier for different number of sensor nodes distributions. We note that for the same distribution, using 165-grid (fine grid) consistently leads to better solutions compared to 48grid. It is also interesting to note that, although the number of relay nodes required increases with the number of sensor nodes, the rate of increase is not very high. For example using the 165-grid only a few additional relay nodes are required to cover 600 sensor nodes, as compared to 100 sensor nodes.

Optimal Relay Node Placement and Trajectory Computation

411

30

Number of relay nodes

25

20

48-Grid

15

165-Grid

10

5

0 100

200

300

400

500

600

Number of sensor nodes

Fig. 2. The number of relay nodes required to form the middle-tier network for different number of sensor nodes

Next we consider the buffer requirements of the relay nodes selected for the middle-tier. Fig. 3 shows the value of the maximum buffer size (Bmax ) calculated by our ILP using 48-grid and 165-grid configurations. At first glance, the figure seems to indicate that 48-grid produces better results (i.e. lower buffer size) compared to 165-grid. However, we must remember that the 48-grid configuration requires a higher number of relay nodes. This means that the same amount of data is distributed over more relay nodes, resulting in a lower buffer requirement per node. But, when we compare the total buffer requirements (as shown in Fig. 4), we see that the 165-grid generates better results, both in terms of the number of relay nodes and the total buffer size. 20 18

Maximum buffer size, Kb

16 14 12 48-Grid

10

165-Grid

8 6 4 2 0 100

200

300

400

500

600

Number of sensor nodes

Fig. 3. The maximum buffer capacity per node required for different number of sensor nodes

412

A. Bari et al.

450 400 350

Total buffer, Kb

300 250

48-Grid 165-Grid

200 150 100 50 0 100

200

300

400

500

600

Number of sensor nodes

Fig. 4. The total buffer capacity of the relay nodes required for different number of sensor nodes

5.2

Simulation of Trajectory Computation Algorithm

The goal of our trajectory computation algorithm is to calculate the trajectory of the MDC (along a straight line), such that the maximum energy dissipation (Emax ) of any relay node is minimized. Fig. 5 shows the average value of Emax , for different size of sensor node networks, corresponding to the trajectory that minimizes the value of Emax , for each configuration. As before, we note that although the value of Emax appears to be lower for 48-grid, this is because it requires more relay nodes resulting in lower energy dissipation per node. As expected, the value of Emax increases steadily with the number of sensor nodes 1.40E+07

1.20E+07

Min-max energy, nJ

1.00E+07

8.00E+06 48-Grid 165-Grid 6.00E+06

4.00E+06

2.00E+06

0.00E+00 100

200

300

400

500

600

Number of sensor nodes

Fig. 5. The minimum of the maximum energy dissipation by the relay nodes in the networks with different number of sensor nodes

Optimal Relay Node Placement and Trajectory Computation

413

4.00E+07 3.50E+07

Min-max enetgy, nJ

3.00E+07 n = 100

2.50E+07

n = 200 n = 300

2.00E+07

n = 400 n = 500

1.50E+07

n = 600

1.00E+07 5.00E+06

90 10 5 12 0 13 5 15 0 16 5

60

75

45

30

0

15

0.00E+00

Rotation angle

Fig. 6. Variation of the minimum of the maximum energy dissipation by the relay nodes with the rotation of the axis, in the networks with different number of sensor nodes with grid setting 165-Grid

for 165-grid case. However, for the 48-grid case, we notice an anomalous case, where the value of Emax for 300 sensor nodes is actually higher than that for both 400 and 500 sensor node distributions. This is because, the performance of coarse grid configurations is not always reliable and may sometimes fail to find a good solution (e.g. for the 48-grid and 300 sensor node case). On the other hand, when we use finer grids (e.g. 165-grid), the computational complexity increases but the we get more consistent and reliable solutions. Finally, Fig. 6 shows how the value of Emax varies with the angle of the straight line trajectory for the MDC. In general, the angle at which the value of Emax is minimized will depend on the distribution of the sensor node and the shape of sensing area. In our experiments the sensing area was a rectangular shape (200m along x-axis and 280m along y-axis), and the sensor nodes were randomly distributed in the sensing field. Therefore, we can expect that the best trajectory will be a (nearly) vertical line. This is exactly what we find in Fig. 6, where the minimum value of Emax is obtained at an angle of about 90◦ for each sensor node distribution. We also note that the value of Emax varies widely with the angle for higher values of n, but as n decreases, these variations are greatly reduced.

6

Conclusions

In this paper, we have proposed a new formulation that, given a set of potential locations of relay nodes, optimally determines the minimum number of relay nodes, along with their locations, in a hierarchical sensor network, which includes a MDC that travels along a fixed trajectory. The placement is done in such a

414

A. Bari et al.

way that i) each sensor node is covered by at least 1 relay node, ii) no relay node suffers from the buffer overflow, and iii) maximum buffer requirement of the relay nodes is minimized. Our ILP is able to generate optimal solutions for networks with hundreds of sensor nodes. We have also proposed a new heuristic for calculating a straight line trajectory for the MDC that minimizes the maximum energy dissipation of the relay nodes. Although, we have focussed on straight line trajectories in this paper, our approach can be adapted to consider other simple trajectories, such as a rectangular or circular path. We are currently extending our approach to consider such trajectories and to determine the most suitable path, based on different layouts of sensing areas and different node distributions.

References 1. Gupta, G., Younis, M.: Load-balanced clustering of wireless sensor networks. In: IEEE ICC, vol. 3, pp. 1848–1852 (2003) 2. Gupta, G., Younis, M.: Fault-tolerant clustering of wireless sensor networks. In: IEEE WCNC, pp. 1579–1584 (2003) 3. Tang, J., Hao, B., Sen, A.: Relay node placement in large scale wireless sensor networks. Computer Communication 29(4), 490–501 (2006) 4. Gnawali, O., Greenstein, B., Jang, K.-Y., Joki, A., Paek, J., Vieira, M., Estrin, D., Govindan, R., Kohler, E.: The Tenet Architecture for Tiered Sensor Networks. In: The Proceeding of SenSys (2006) 5. Bari, A., Jaekel, A., Bandyopadhyay, S.: Clustering Strategies for Improving the Lifetime of Two-Tiered Sensor Networks. Computer Communications 31(14), 3451– 3459 (2008) 6. Bari, A., Jaekel, A., Bandyopadhyay, S.: Integrated Clustering and Routing Strategies for Large Scale Sensor Networks. In: Akyildiz, I.F., Sivakumar, R., Ekici, E., de Oliveira, J.C., McNair, J. (eds.) NETWORKING 2007. LNCS, vol. 4479, pp. 143–154. Springer, Heidelberg (2007) 7. Hou, Y.T., Shi, Y., Sherali, H., Midkiff, S.F.: On Energy Provisioning and Relay Node Placement for Wireless Sensor Networks. IEEE Transactions on Wireless Communications 4(5), 2579–2590 (2005) 8. Heinzelman, W., Chandrakasan, A., Balakrishnan, H.: Energy effcient communication protocol for wireless micro-sensor networks. In: HICSS, pp. 3005–3014 (2000) 9. Bari, A., Jaekel, A., Bandyopadhyay, S.: Optimal placement and routing strategies for resilient two-tiered sensor networks. In: Wireless Communications and Mobile Computing. Wiley, Chichester (2008), doi:10.1002/wcm.639 10. Hao, B., Tang, J., Xue, G.: Fault-tolerant relay node placement in wireless sensor networks: formulation and approximation. In: HPSR, pp. 246–250 (2004) 11. Liu, H., Wan, P., Jia, W.: Fault-Tolerant Relay Node Placement in Wireless Sensor Networks. In: Wang, L. (ed.) COCOON 2005. LNCS, vol. 3595, pp. 230–239. Springer, Heidelberg (2005) 12. Yarvis, M., Kushalnagar, N., Singh, H., Rangarajan, A., Liu, Y., Singh, S.: Exploiting heterogeneity in sensor networks. In: INFOCOM 2005, vol. 2, pp. 878–890 (2005) 13. Bari, A., Jaekel, A., Bandyopadhyay, S.: Optimal Placement of Relay Nodes in Two-Tiered, Fault Tolerant Sensor Networks. In: IEEE ISCC (2007)

Optimal Relay Node Placement and Trajectory Computation

415

14. Bari, A., Xu, Y., Jaekel, A.: Integrated Placement and Routing of Relay Nodes for Fault-Tolerant Hierarchical Sensor Networks. In: IEEE ICCCN - SN, pp. 1–6 (2008) 15. Bari, A., Jaekel, A.: Techniques for Exploiting Mobility in Wireless Sensor Networks, Handbook of Research on Mobile Business: Technical, Methodological and Social perspective, 2nd edn. IGI Global, Information Science Reference, pp. 445– 455 (2008) ISBN: 978-1-60566 16. Pazzi, R.W.N., Boukerche, A.: Mobile data collector strategy for delay-sensitive applications over wireless sensor networks. Computer Communications 31(5), 1028– 1039 (2008) 17. Bari, A., Chen, Y., Jaekel, A., Bandyopadhyay, S.: A New Architecture for Hierarchical Sensor Networks with Mobile Data Collectors. In: ICDCN, Networking track (2010) 18. Bari, A., Chen, Y., Roy, D., Jaekel, A., Bandyopadhyay, S.: Energy Aware Trajectory Computation of Mobile Data Collectors in Hierarchical Sensor Networks. In: IEEE ICC (to appear, 2010) 19. Gandham, S.R., Dawande, M., Prakash, R., Venkatesan, S.: Energy efficient schemes for wireless sensor networks with multiple mobile base stations. In: IEEE Global Telecommunication Conference, GLOBECOM 2003, vol. 1, pp. 377–381 (2003) 20. Jea, D., Somasundara, A., Srivastava, M.: Multiple controlled mobile elements (Data Mules) for data collection in sensor networks. In: Prasanna, V.K., Iyengar, S.S., Spirakis, P.G., Welsh, M. (eds.) DCOSS 2005. LNCS, vol. 3560, pp. 244–257. Springer, Heidelberg (2005) 21. Luo, J., Hubaux, J.-P.: Joint mobility and routing for lifetime elongation in wireless sensor networks. In: INFOCOM, vol. 3, pp. 1735–1746 (2005) 22. Nakayamaa, H., Ansarib, N., Jamalipourc, A., Katoa, N.: Fault-resilient sensing in wireless sensor networks. Computer Communications 30(11-12), 2375–2384 (2007) 23. Shah, R.C., Roy, S., Jain, S., Brunette, W.: Data mules: Modeling a Three-Tier Architecture For Sparse Sensor Networks. In: IEEE Workshop on Sensor Network Protocols and Applications (2003) 24. Jain, S., Shah, R., Brunette, W., Borriello, G., Roy, S.: Exploiting Mobility for Energy Efficient Data Collection in Sensor Networks. Mobile Networks and Applications 11(3), 327–339 (2006) 25. Gu, Y., Bozdag, D., Ekici, E., Ozguner, F., Lee, C.-G.: Partitioning-Based Mobile Element Scheduling in Wireless Sensor Networks. In: IEEE Conf. Sensor and Ad Hoc Communication and Network (2005) 26. Luo, H., Ye, F., Cheng, J., Lu, S., Zhang, L.: Ttdd: Two-tier data dissemination in large-scale wireless sensor networks. Wireless Networks 11, 161–175 (2005) 27. Kim, H.S., Abdelzaher, T.F., Kwon, W.H.: Minimum-energy asynchronous dissemination to mobile sinks in wireless sensor networks. In: Proc. of the 1st Int. Conf. on Embedded Networked Sensor Systems (SENSYS), pp. 193–204 (2003) 28. Hwang, K., In, J., Eom, D.: Distributed dynamic shared tree for minimum energy data aggregation of multiple mobile sinks in wireless sensor networks. In: R¨ omer, K., Karl, H., Mattern, F. (eds.) EWSN 2006. LNCS, vol. 3868, pp. 132– 147. Springer, Heidelberg (2006) 29. Pan, J., Hou, Y.T., Cai, L., Shi, Y., Shen, S.X.: Topology Control for Wireless Sensor Networks. In: International Conference on Mobile Computing and Networking, pp. 286–299 (2003)

Balanced Itinerary Planning for Multiple Mobile Agents in Wireless Sensor Networks Min Chen1 , Wei Cai1 , Sergio Gonzalez2 , and Victor C.M. Leung2 1

2

Department of Computer Science and Engineering, Seoul National University, Korea [email protected], [email protected] Department of Electrical and Computer Engineering, University of British Columbia, V6T 1Z4, Canada {sergiog,vleung}@ece.ubc.ca

Abstract. In this paper, we consider the use of multiple mobile software agents to perform different tasks in wireless sensor networks (WSNs). To this regard, determining the number of mobile agents in the WSN remains an open issue in solving multi-agent itinerary planning (MIP) problem. We propose a novel scheme entitled MST-MIP based on minimum spanning tree, where each branch stemmed from the sink corresponds to a group of source nodes assigned for a mobile agent to visit. Furthermore, a balancing factor α is introduced to achieve a flexible trade-off control between energy cost and task duration, and the balancing MST-MIP algorithm is named BST-MIP. Extensive experiments show that MST-MIP has lower energy consumption than previous MIP proposals, while BST-MIP decreases the task duration up to 50%. Keywords: mobile agent, itinerary planning, minimum spanning tree, wireless sensor networks.

1 Introduction Compared to conventional data fusion in wireless sensor networks, the mobile agent (MA) system is better due to its intrinsic flexibility. In addition, it has been shown that data compression and fusion using MAs achieve better energy efficiency. However, using MAs also introduces larger task latency across a network with a large number of source nodes. To address this issue, a multi-agent system is proposed to achieve a balanced trade-off between energy cost and task latency. In a multiple mobile software agents system, serval MAs roam in the network simultaneously. Each MA visits a subset of source nodes to retrieve information for the sink. In contrast to single MA itinerary planning (SIP), it is more challenging to determine the number of MAs and their corresponding subsets of source nodes for multi-agent itinerary planning (MIP), which is also called source grouping problem in this paper. To address this issue, we propose the use of minimum spanning tree (MST) to solve the MA grouping problem. We model the network topology as a totally connected graph (TCG). In order to simplify the TCG, only source nodes and the sink node constitute its vertices, while the weight of each arc can be basically estimated by the hop count among source nodes or the sink. According to such a hop-count-oriented TCG (H-TCG), we J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 416–428, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Balanced Itinerary Planning for Multiple Mobile Agents in WSNs

417

can further generate a MST, where each branch stemmed from the sink corresponds to a group of sources included in the branch. Then, the number of branches in the MST is equivalent to the number of MAs used in the network, and each agent will collect sensory data by traversing the sources in its corresponding branch before returning to the sink. In this paper, we refer to the proposed solution for the MIP problem as MST-based MIP, which is denoted by MST-MIP. In MST-MIP, the critical issue is how to define the weight of each arc in the H-TCG. Intuitively, we can use the hop count of two vertices1 in the H-TCG as the arc weight. Additionally, we further introduce a balancing factor α for the calculation of the weight. By adjusting α to suitable value, a balanced MST can be generated for the H-TCG. In this paper, the balanced MST-MIP algorithm is named BST-MIP, which can achieve a flexible trade-off control between energy cost and task duration. The importance of balanced source grouping in BST-MIP can be addressed in the following aspects: – Task duration: Since multiple agents work in parallel, the task duration is mainly dependent on the delay incurred when an agent traverses along the branch containing the largest number of source nodes. Balancing the source grouping to eliminate the “bottleneck” branch will be the key to the reduction of task duration. – Lifetime: From the perspective of the effective operation time before the first node depletes its energy, the nodes along a longer branch in the H-TCG will consume their energy more quickly than other nodes. Thus, a more balanced source grouping increases the lifetime by spreading the traffic load generated by multi-agent immigration more evenly in the whole network. Extensive OPNET simulations are performed to show that the novel scheme outperforms the existing works. The reminder of the paper is organized as follows: Related work and a problem statement are introduced in Section 2 and Section 3. Then we describe the novel minimum spanning tree based source nodes grouping algorithm and its enhanced version in Section 4 and Section 5. The performance of our simulations will be analyzed in Section 6. Finally, Section 7 concludes this paper.

2 Related Work Devising itinerary planning solutions for single MA is an essential part of MA research, which has been investigated by a number of researchers. Work in [2] proposes the simplest SIP solution: Local Closest First (LCF) and Global Closest First (GCF), whereas focus on energy efficiency is achieved by means of a genetic algorithm based solution presented in [3]. However, these approaches are not energy efficient as indicated in [4], in which the authors propose a better scheme named IEMF. In particular, IMEF denotes the importance of choosing the first visiting node. Based on this conclusion, it estimates energy costs of different choices of the first node and adopts the best solution to achieve energy efficiency. However, the migration period of a single MA introduces larger end-to-end delay. Because of this, a multiple MAs system provides an alternative solution to achieve energy efficiency and reasonable task duration simultaneously. In CL-MIP [5], the authors 1

Two vertices can be a pair of two sources or a pair of a source and the sink.

418

M. Chen et al.

divide the MIP problem into two parts: source node grouping and a source node visiting sequence for each MA. They discover dense centers of source nodes and group them within a specific radius as one subset, each of which is assigned to a corresponding MA. Then, they determine the source node visiting sequence for each MA by employing IEMF. However, this paper still leaves an open research issue on choosing the optimal radius with the purpose of minimizing the total communication cost. Researchers in [6] also applied the genetic algorithm approach for the multiple MAs itinerary planning problem. Their work considers both the grouping and visiting sequence together to produce an energy-aware solution. The drawback of this is that the procedure of genetic evolution is complicated and hard to be realized in practice. In this paper, we focus on the trade-off between energy cost and task duration in MIP solutions and provide a novel grouping scheme base on minimum spanning tree theory.

3 Problem Statement 3.1 Grouping Problem As shown in Fig.1, the data sink is denoted by the blue star. We also assume that the data sink is the information center with infinite power and sufficient computational capacity. Source nodes (denoted by green circles) are uniformly distributed in the deployment area. As seen in [2] [4] [5] [6], we assume geographical information of all source nodes stored in the sink, and that they remain static during MA migration. The dashed lines between the source nodes and data sink denote the distances, which will be utilized for the calculation of the estimated hop count.

7

1

S

2

4 3

5

6

Fig. 1. A Typical Network Topology with Multiple Sources and Single Sink Node in WSNs

Balanced Itinerary Planning for Multiple Mobile Agents in WSNs

419

The key issue of solving the MIP problem is source-grouping, which includes two challenges: 1) determining the optimal number of MAs, and 2) assigning the corresponding subset of source nodes to each of them. The goal of source-grouping algorithm is to minimize task duration while decreasing total energy cost as possible. 3.2 The Metric to Evaluate the MA Migration Cost In order to solve the grouping problem, we need to identify the key metric to evaluate the cost for MA’s migration between two sources in terms of energy consumption and delay. In order to analyze the delay, let’s consider the latency between two neighboring nodes first, which is the summation over the queuing, processing, propagation, and transmission delays: – Queuing delay: since WSNs are normally assumed to support a low packet rate, communication traffic is considered to be rather low, thus queuing delay can be ignored. – Processing delay: With respect to the processing delay, we assume that each node incurs similar delay to handle one MA. – Propagation delay: This parameter can be neglected when compared to the other delays. – Transmission delay: Because the size of an MA does not change between two source node, its transmission delay remains constant between any pair of intermediate sensor nodes. Therefore, generally speaking, the delay taking place between any pair of intermediate nodes between two sources is similar. Consequently, the delay between two source nodes is proportional to the hop count between the two sources. On the other hand, the energy consumption for the migration of an MA between two sources is proportional to the number of transmissions, which is also proportional to the hop count. Thus, in order to estimate energy cost and/or delay, the metric to evaluate the weight between two sources can be simply the hop count. However, most of previous works [2, 3, 7, 8] ignore this issue, since they use distance between two sources as the weight between two sources. 3.3 Hop Count Estimation Formula Compared to the distance, the estimated hop count can describe the energy cost more accurately, as discussed in Section 3.2. Thus, the issue of estimating hop count needs to be addressed at first. Assuming that there are two source nodes (i.e., i and j), let Dji denote the distance between the two sources, and let Let R represent the maximum transmission range for each hop. Since the actual hop distance between two nodes is smaller than R, we introduce a factor of ξ, 0 < ξ ≤ 1 and let R × ξ represent the expected hop distance. Finally, let Hji denote the estimated hop count between i and j. Then, Hji can be estimated as follows: Dji Hji = (1) R×ξ According to this equation, we can calculate estimated hop count between each pair of nodes.

420

M. Chen et al.

4 Minimum Spanning Tree Based Source Grouping for MIP In the section, we first introduce the construction of a minimum spanning tree (MST) based on the estimated hop count information, and propose an efficient source-grouping algorithm for multi-agent itinerary planning (MIP) in WSNs. 4.1 The Hop Count Based Minimum Spanning Tree Total Connected Graph. We model the network topology as a totally connected graph (TCG). In order to simplify the TCG, only source nodes and the sink node constitute its vertices, while the weight of each arc can be basically estimated by the hop count among source nodes or the sink. Table.1 gives an example corresponding to Fig.1. Each element in the table represents the estimated hop count between two source nodes. The information from the table can be easily transformed into a hop count based total connected graph (H-TCG). In H-TCG, the vertices represent the source and sink nodes in the network, while the weight of each edge can be expressed at the corresponding estimated hop count. Table 1. The hop count between nodes

s 1 2 3 4 5 6 7

s ×

1 2 ×

2 3 6 ×

3 2 3 4 ×

4 3 2 7 1 ×

5 4 5 4 2 3 ×

6 2 7 2 3 6 5 ×

7 4 2 6 7 6 11 9 ×

Definition of MST. Given a total connected graph G = (V, E), we denote (u, v) as the edge connected to the vertex v and u; thus, (u, v) ∈ E. The weight of edge (u, v) is denoted as w(u, v). If there is a subset T , which includes all of the vertices and has the feature that w(T ) = (u,v)∈T w(u, v) is minimum, then T is the minimum spanning tree of G. Calculation of MST. There are a number of well-known algorithms that calculate the MST of a total connected graph. In our approach, we adopt one of the simplest approaches, known as the Prim Algorithm. Its pseudo-code is given as follows: MST Based Node Grouping. In MST, there are several branches stemming from the sink. For each branch, a MA is dispatched to traverse the source nodes contained in the branch and return to the sink. Thus, in the proposed MST-MIP scheme, the number of MAs is equal to the number of direct vertices connected with the sink node. For each MA, the group of sources is determined by its corresponding branch. As observed in the example shown in Fig.2, (s, 1), (s, 2) and (s, 3) are three trunks originating at the sink node. (s, 1) and (1, 7) form the first branch; (s, 2) and (2, 6)

Balanced Itinerary Planning for Multiple Mobile Agents in WSNs

421

Algorithm 1. Prim Algorithm for Minimum Spanning Tree T ⇐φ V ⇐ {Sink} while ∃(u ∈ V, v ∈ / V ) do find (u, v) which has the minimum w(u, v) T ⇐ T ∪ (u, v) V ⇐V ∪v end while return T

7

1

S

2

4 3

5

6

Fig. 2. An Illustrative Example of MST based Source-grouping

represent the second branch; (s, 3) (3, 4) and (3, 5) constitute the third branch. Thus, three MAs are dispatched. One MA visits the source nodes 1, 7, the second MA visits the source nodes 2, 6, and the third MA will visit the source nodes 3, 4 and 5. After source-grouping, the visiting sequence in each subset of sources can be obtained by solving the SIP problem, which has been widely studied in previous work [2] [3] [4].

5 Balanced MST-MIP Algorithm 5.1 The Residual Problem In the basic MST-MIP algorithm, the hop count of two vertices is directly set to the arc weight in the H-TCG, and the Prim algorithm with greedy feature is used to construct the MST. When the source nodes are close to each other, the relatively small hop count between two adjacent sources easily becomes the shortest edge during the selection of Prim algorithm. Consequently, the longer the branch is, the higher that the possibility

422

M. Chen et al.

2

3

3

4

3

5

2 6 3

4 4

6

1

4

S (a)

w = α × h ij + (1 − α ) × ( h si + h sj ) e.g. α = 0.6 hij = 3 hsi = 6 hsj = 4 ⇒ w = 5 .8 ( Weight is Updated )

6

5.8

3

4

5

5

2 6 5.8 1

4 4

6 4

S (b)

Fig. 3. MST-based MIP: (a) illustrative example of the problem in the basic MST-MIP; (b) an improved solution for source-grouping with α = 0.6

to connect more source nodes is, which causes the reduction of the number of branches stemmed from the sink. Given the example shown in Fig.3(a), the MST contains a single branch, which means that only one MA is sent to the network, and that the task duration will be as high as in the SIP algorithm. Intuitively, the basic MST-MIP approach does not partition source nodes intelligently without considering the distribution of source nodes. Thus, we need to find a much better solution to achieve a balance between the energy cost and task duration.

Balanced Itinerary Planning for Multiple Mobile Agents in WSNs

423

5.2 Balancing Factor: α Considering two source nodes i and j in the network, let us denote the estimated hop count between them as hji , and denote their estimated hop count to the sink node as his and hjs , respectively. In order to address the unbalance issue existing in the basic MST-MIP algorithm, we introduce a balancing factor α to calculate the weights in the TCG as follows: w = α × hij + (1 − α) × (his + hjs )

(2)

where α ∈ [0, 1]. By adjusting α to a suitable value, a balanced MST can be generated for the H-TCG. In this paper, the balanced MST-MIP algorithm is denoted by BST-MIP, which can achieve a flexible trade-off control between energy cost and task duration. Given the example show in Fig. 3(b), if we set α to 0.6, the weight will be updated according to the relative distance between the source nodes and the sink, which produces a different minimum spanning tree. As the result, three MAs will be dispatched along three branches stemming from the sink, with the subsets of source nodes of {1, 2}, {3, 4} and {5}. Compared to the MST generated in Fig. 3(a), the updated MST is more balanced.

6 Simulation and Analysis 6.1 Simulation Setup We implemented the basic MST-MIP and BST-MIP schemes in the OPNET environment, and compare them to CL-MIP as presented in [5]. In the implementation, we define a WSN deployment area of 1000m × 500m and allocate the sink node in the center of the network. Wireless sensors with a 802.11b/g network interface are uniformly distributed in the network. Random seeds are used to determine the position for the source nodes. For each MA, the parameters are set as shown in Table.2. 6.2 Evaluation Metrics In order to evaluate the energy efficiency, task duration, and their overall performance from our simulation results, we consider the following three performance metrics, as reported previously [4] [5] [6]: – Average Communication Energy: Used to indicate the total communication energy consumption in the network, including transmitting, receiving, retransmissions, overhearing and collision, to obtain each sensory data from all the target sources. – Task Duration: Used for calculating the period for one particular task. For the case of the SIP algorithm, it is equivalent to the average end-to-end reported delay, which is the average delay from the time when a MA is dispatched by the sink to the time when it returns to the sink. For the case of the MIP algorithm, since multiple agents work in parallel, there must be one agent that returns to the sink at the end. Then, the task duration of the MIP algorithm is the delay of that agent.

424

M. Chen et al. Table 2. Simulation Parameters for MA Raw Data Reduction Ratio Aggregation Ratio Raw Data Size MA Code Size MA Accessing Delay Data Processing Rate

0.8 0.9 2048 bits 1024 bits 10 ms 50 Mbps

– Energy-Delay Product (EDP): Used for representing the overall performance from both the energy efficiency and task duration aspects. For time-sensitive applications over energy constrained WSNs, EDP (calculated by EDP = energy × delay) gives us a unified view. The smaller this value is, the better the unified performance will be. 6.3 The Selection of a Balancing Factor α in BST-MIP In this section, we study the impact of the factor α on energy cost, task duration and EDP, in order to discover its optimal value. Fig.4 shows the impact of α on energy cost. When the factor α is smaller than 0.5, the network energy cost is stable at around 0.65 Joules/Task, which corresponds to the extreme case when all of the source nodes are connected to the sink directly in the MST, i.e., each source node is visited by an individual MA, and the benefits of the MA system in terms of data reduction and fusion are not utilized. Therefore, the energy cost is high. When the value of α is larger than 0.5, the impact of the distance between two sources on the arc weight increases, and a source node is easier to be included in an existing branch stemmed from the sink. Fig.5 shows the impact of α on task duration. When the value of α is lower than 0.5, the impact of the distance between source and the sink on the arc weight increases, and a new branch stemming from the sink is more likely to be generated. Contrary to the

Task Energy (Joules/Task)

0.7

MST−MIP

0.6 0.5 0.4 0.3 0.2

0

0.2

0.4

Value of α

0.6

Fig. 4. The impact of α on energy cost

0.8

1

Balanced Itinerary Planning for Multiple Mobile Agents in WSNs

425

0.25

Task Duration (s)

MST−MIP 0.2

0.15

0.1

0.05

0

0.2

0.4

Value of α

0.6

0.8

1

Fig. 5. The impact of α on task duration

energy cost, the task duration is relatively lower when the value of α is below 0.5. This is because the system delay is reduced for the task operation simultaneously performed by multiple MAs in parallel. When α increases from 0.5 to 1.0, the number of MAs decreases, which leads to a longer itinerary for each MA, and thus increasing the task duration as shown in Fig.5. Similar to how Fig.4 and Fig.5 show the trade-off between energy cost and task duration, Fig.6 describes the impact of α on EDP, which is the metric that we employ to depict the overall performance. We observe that when α is below 0.5, EDP is stable according to the constant value of both energy cost and task duration in this interval. When α increases, the value of EDP decreases until α = 0.6, where EDP reaches its lowest value, indicating the best EDP performance. However, if we keep raising α from 0.6 to 1.0, the EDP raises back to a relative high value. From this study, we can conclude that 0.6 is the best value regarding to the EDP performance. 6.4 Performance Comparison In this section, we compare the proposed MST-MIP and BST-MIP to two typical existing approaches, i.e., IEMF [4] and CL-MIP [5]. As a latest proposed solution for SIP, IEMF has the best performance in terms of energy cost and task duration, while CLMIP is the first solution for the MIP problem. We changed the number of source nodes from 10 to 40 with the step of 5, and perform a series of simulations for each scheme. For a single data point, various random seeds are adopted, each of which corresponds to a scenario with different deployment of source nodes. As shown in Fig.7, all of the three MIP schemes have lower task duration than IEMF, which verifies the effectiveness of using a multi-agent approach for reducing the system delay. The delay of CL-MIP and MST-MIP are comparable. However, BST-MIP achieves up to 50% reduction on the delay performance compared to CL-MIP and MST-MIP, which shows that BST-MIP allocates source nodes to multiple MAs in a more balanced fashion.

426

M. Chen et al.

0.047

MST−MIP

0.046

EDP

0.045 0.044 0.043 0.042 0.041

0

0.2

0.4

Value of α

0.6

0.8

1

Fig. 6. The impact of α on EDP 1.4

IEMF CL−MIP MST−MIP BST−MIP

Task Duration (s)

1.2

1

0.8

0.6

0.4

0.2

0 10

15

20

25

30

35

40

Number of Source Node

Fig. 7. The impact of the number of source nodes on task duration

However, when our focus is moved to the energy performance, it is observed that the overwhelming delay performance of BST-MIP is achieved by comprising some energy performance, and thus increasing about 15% energy cost compared to IEMF and CLMIP. When compared to MST-MIP, BST-MIP requires 30% more energy cost if the number of source nodes is 40. This is because the fewer number of MAs used in MSTMIP, thus saving the communication overhead of delivering processing codes carried by a lager number of MAs. Since MST-MIP and BST-MIP achieve the best performance in terms of energy cost and task duration, respectively, we need to further compare them through EDP

Balanced Itinerary Planning for Multiple Mobile Agents in WSNs

427

0.65

Task Energy (Joules/Task)

0.6

IEMF CL−MIP MST−MIP BST−MIP

0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2

10

15

20

25

30

35

40

Number of Source Node

Fig. 8. The impact of the number of source nodes on energy cost

0.7

0.6

0.5

IEMF CL−MIP MST−MIP BST−MIP

EDP

0.4

0.3

0.2

0.1

0 10

15

20

25

30

35

40

Number of Source Node

Fig. 9. The impact of the number of source nodes on EDP

performance. As shown in Fig. 9, BST-MIP has the least EDP, which illustrates that BST-MIP achieves efficient trade-off between energy and delay. When there are 40 source nodes, BST-MIP decreases EDP up to 70% compared to IEMF, and achieves a reduction of EDP up to 50% compared to CL-MIP and MST-MIP.

428

M. Chen et al.

7 Conclusion Compared to a single mobile agent system, the source-grouping problem is a key issue in planning itineraries for a multiple mobile agents system. In this paper, we first present a minimum spanning tree based source-grouping algorithm, and further propose the introduction of a balancing factor to achieve flexible trade-off control between energy cost and task duration. By adjusting the balancing factor, the QoS requirements in terms of delay can be satisfied for a large range of applications while reducing the energy cost to a maximum capacity. As part of our future work, we need to investigate a more efficient function to evaluate the arc weight for the construction of minimum spanning trees.

References 1. Chen, M., Gonzalez, S., Leung, V.: Applications and Design Issues of Mobile Agents in Wireless Sensor Networks. IEEE Wireless Communications Magazine (WCM), Special Issue on Wireless Sensor Networking 14(6), 20–26 (2007) 2. Qi, H., Wang, F.: Optimal itinerary analysis for mobile agents in ad hoc wireless sensor networks. In: Proceedings of the IEEE 2001 International Conference on Communications (ICC 2001), Helsinki, Finland (2001) 3. Wu, Q., Rao, N.S.V., Barhen, J., Iyengar, S.S., Vaishnavi, V.K., Qi, H., Chakrabarty, K.: On computing mobile agent routes for data fusion in distributed sensor networks. IEEE Trans. Knowledge and Data Engineering 16, 740–753 (2004) 4. Chen, M., Leung, V., Mao, S., Kwon, T., Li, M.: Energy-Efficient itinerary planning for mobile agents in wireless sensor networks. In: Proceedings of the IEEE 2009 International Conference on Communications (ICC 2009), Bresden, Germany, pp. 1–5 (2009) 5. Chen, M., Gonzlez, S., Zhang, Y., Leung, V.: Multi-agent itinerary planning for sensor networks. In: Proceedings of the IEEE 2009 International Conference on Heterogeneous Networking for Quality, Reliability, Security and Robustness (QShine 2009), Las Palmas de Gran Canaria, Spain (2009) 6. Cai W., Chen M., Hara T., Shu L.: GA-MIP: Genetic Algorithm based Multiple Mobile Agents Itinerary Planning in Wireless Sensor Network. In: Proceeding of the 5th Annual International Wireless Internet Conference (WICON 2010), Singapore (March 2010) 7. Gavalas, D.: A Heuristic Algorithm for Designing Near-Optimal Mobile Agent Itineraries. Journal of Communications and Networks, IEEE/KICS 8(1), 123–131 (2006) 8. Gavalas, D., Mpitziopoulos, A., Pantziou, G., Konstantopoulos, C.: An Approach for NearOptimal Distributed Data Fusion in Wireless Sensor Networks. In: Wireless Networks (2010), doi:10.1007/s11276-009-0211-0

Analytical Modeling of Address Allocation Protocols in Wireless Ad Hoc Networks (Invited Paper) Ahmad Radaideh and John N. Daigle University of Mississippi, University, MS, USA 38677 [email protected], [email protected]

Abstract. A detailed description of the Internet Protocol Address Assignment (IPAA) is presented and state diagrams for its behavior constructed. Formulae for the expected latency and communication overhead of the IPAA protocol are derived, with the results being given as functions of the number of nodes in the network with message loss rate, contention window size, coverage ratio, and the counter threshold as parameters. The IPAA and MANET Configuration (MANETconf) are compared in detail. The results show that the latency and communication overhead for MANETconf are significantly higher than for the IPAA protocol. Results of extensive sensitivity analyses for the IPAA protocol are also presented. Keywords: ad hoc network, address assignment latency, address assignment overhead, probabilistic modeling.

1

Introduction

In this paper, a detailed description of the Internet Protocol Address Assignment (IPAA) protocol is presented, and analytical derivations for the expectations of latency and communication overhead are given. Expected values of the performance measures as a function of the number of network nodes at the time the new node joins the network are presented. Analytical and simulation results are presented to show the effect of changes in message loss rate, coverage ratio, and the contention window size on the performance measures. Performance of this protocol is also compared to that of the MANET Configuration (MANETconf). Baccelli [1] discusses design of address configuration protocols for MANETs, where uniqueness of the assigned address is a major issue. Some proposed solutions are based on duplicate address detection (DAD) mechanism while others use the binary-split idea [2] so that each node has a disjoint subset of network addresses. Most mechanisms are classified in [3]. Perkins et al. [4] configures by first choosing a random address then performing a DAD procedure within the MANET. The protocol performs DAD only when assigning an IP address to a new node, the proposed protocol lacks support for partitioning and merging in MANET. Jeong et al. [5] proposes a protocol J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 429–446, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

430

A. Radaideh and J.N. Daigle

with two address detection mechanisms. A strong DAD, based on the protocol proposed in [4], is performed in the initial phase to verify the uniqueness of the randomly selected address and a weak DAD, based on [6], which is always executed in order to prevent address conflicts with existing nodes. The weak DAD uses a virtual address, which is a combination of an address and a key. The key, which is assumed to be unique in the network, is appended to the address in the routing messages and the routing table. The weak DAD also identifies duplicate addresses, but his protocol monitors and changes the routing messages and therefore is considered routing protocol dependent. A passive DAD approach [7] has been adopted in some proposed solutions for dynamic address configuration, such as in [8] and [9]. Passive DAD enables nodes to detect duplicate addresses in the network by analyzing received routing protocol messages. One way to detect address conflicts is based on the sequence number of a link-state routing messages. Other ways to detect address conflicts are based on locality and neighborhood, and they are all based on analyzing routing information, which makes these solutions routing-protocol dependent. Zhou et al. [10] proposed a mechanism based on a stateful function, f (n), to derive addresses with low probability of address duplication and therefore it avoids the use of DAD. The initial state of f (n) is a seed that generates a sequence of unique numbers that can be used as network addresses. The function has to be designed carefully such that the interval between two occurrences of the same number in the generated sequence is extremely long and the probability of generating the same number in finite number of sequences initiated by different seeds is extremely low, which is considered a hard mathematical problem. Additional approaches based on genetic algorithms and a quadratic residue approach are presented in [11] and [12], respectively. In MANETconf [13], each configured node maintains state information of the currently assigned addresses so it can choose an available address for a new node and verify its uniqueness throughout the network. In case the newly joining node is unable to find a neighbor, joining node concludes that is the first node in the network and performs address configuration for itself; otherwise it asks one of its neighbors to process its address allocation. The selected neighbor chooses an available address and performs the DAD procedure to verify the uniqueness of the chosen address. Mohsin et al. [14] proposed the IPAA protocol which is based on a dynamic configuration of addresses using the concept of binary split. Each node can independently assign a unique address to a new node without consulting any other node in the MANET. Each node in the network has a disjoint subset of the address space. When a new node joins the network, it tries to find a neighbor node that can perform an address configuration on its behalf. If a neighbor is found, the neighbor splits its available address space and sends one half to the new node. The new node then assigns itself the first IP address in the received address space and keeps the rest to configure other nodes in the future. Both MANETconf and IPAA handle network partitioning and merging as well as address recovery due to node departures. A very similar approach to IPAA, which is based on the binary split idea, is proposed in Tayal et al. [15].

Address Allocation Protocols

431

The main contributions in this work are building state diagrams for MANETconf [13] and IPAA [14], deriving analytical formulations for the expectations of latency and communication overhead for IPAA [14], and presenting analytical and simulation results for the performance measures of that protocol for different number of existing nodes in the network.1 The paper is organized as follows. In Section 2, a detailed description of the proposed address allocation protocol in [14] is presented; a detailed description of the protocol presented in [13] is omitted due to lack of space. For each protocol, two state diagrams are derived based on the protocol specifications. The state diagrams give a whole picture of the protocol behavior, messages, timers and handshakes. One diagram shows the states of the new node during the address allocation process while the other one shows the states of an existing node that performs address allocation for the new node. In Section 3, analytical formulas for latency and communication overhead for the protocol in [14] are derived. The analytical formulas represent the expected values of the performance measures as a function of the number of nodes in the network and with message loss rate, contention window size, coverage ratio, and counter threshold as parameters. Section 4 presents the analytical as well as the simulation results for different values of the message loss rate, contention window size, and the coverage ratio. Section 5 concludes the paper work and suggests future research.

2

Detailed Descriptions of the IPAA Protocol

The IPAA protocol (and also MANETconf) is designed with the following network operating characteristics in mind. Nodes are free to move in the network and to join and leave at any time. Address allocation and maintenance have to be performed whenever the topology changes. The MANET is configured as a private IPv4 network in which the participating nodes are configured in advance to use a specific private address block. At any given time a group of connected nodes forms a network partition that has a universal unique identifier (UUID). As time evolves, a partition could either split or merge with another partition. The nodes in MANET communicate with each other using IP datagrams. Communications between distant nodes of a partition are carried over intermediate nodes running an ad hoc routing protocol. IPAA presents a distributed dynamic address allocation protocol for a standalone MANET. The protocol avoids the DAD process by employing a proactive approach using the binary-split idea. The binary-split idea is that each node has a disjoint subset of the address block and it can independently allocate a unique address and hand half of its address space to a newly joining node without getting an agreement from every other node in its partition. When a new partition is formed, the only node in that partition reserves the whole address block and assigns itself the first address in that block. A newly joining node, a requester (or client), asks an existing neighbor node, an initiator 1

The protocol proposed in [16] is compared to MANETconf via simulation with latency reported at about one half of that of MANETconf.

432

A. Radaideh and J.N. Daigle

(or server), for a network address. The server divides its address space in two and gives one half to the requester. The requester assigns itself the first address from the received address space and keeps the rest of it to serve other nodes in the future. If the initiator has no space left, it borrows an address space from an existing node and forwards it to the requester. This procedure avoids flooding the whole partition to verify the uniqueness of the selected address as DAD based protocols do. Maintaining the whole Address Block is the key issue in this protocol. The protocol performs maintenance procedures to avoid address leak and conflict problems. The address leak problem occurs when a node abruptly departs the network or moves out its partition without returning its address space. Without cleaning up departed nodes addresses, these addresses will be considered allocated to existing nodes and can not be allocated to newly joining nodes. Nodes keep track of allocated address blocks to resolve the address leak problem. Graceful departure is provided where nodes that want to leave the network return their address blocks and confirm their departure. Address conflict problem could happen when two partitions merge together since each partition reserve the whole address block for itself. Nodes with the same address in both partitions are allocated the same address space when configured. The conflicting node that has a bigger address space should give up its allocated space and ask other nodes for new allocation. New Node Address Allocation. Figures 1 and 2 show the state diagrams of a client and a server nodes during address allocation process, respectively. When a client joins the network it broadcasts a one-hop REQUEST message. A neighbor replies with a REPLY message to the client. The client selects one of its existing neighbors who replied to its request message to be its server. The client sends an ACK message to the selected server asking for a unique address. When receiving the ACK message, the server starts the allocation process for the requested client. Neighbors are expected to reply to the client request within REPLY TIMEOUT amount of time. If the timer expires without receiving any reply, the client repeats searching for neighbors for NEIGHBOR REQUEST THRESHOLD number of times. If all trials have failed, the client reserves the whole address block for itself, allocates itself the first address in that block and sets a UUID for that partition. If reply messages have been received, the client selects one of its neighbors to perform the address allocation process by sending an ACK message and starting a timer with a timeout value ADDRESS BLOCK TIMEOUT. When a node receives the ACK message, it divides its available address space into two disjoint subsets and sends one subset to the client in the ADDRESS BLOCK message. The client receives the ADDRESS BLOCK and the partition UUID, configures itself the first address in that block, and keeps the rest of its address space to configure newly joining nodes in the future. The client confirms a successful address allocation by sending CONFIRM message back to the server. If the client timer expires before receiving the ADDRESS BLOCK message, the client considers that the server is no longer

Address Allocation Protocols

433

FIND NEIGHBOR

Broadcast REQUEST Start Timer1 Retries ++ Timer1 = NEIGHBOR_REPLY_TIMEOUT Retries != Timer2 = ADDRESS_BLOCK_TIMEOUT NEIGHBOR_REQUEST_THRESHOLD and Retries != NEIGHBOR_REQUEST_THRESHOLD

WAIT NEIGBOR Timer1 = REPLY REPLY_TIMEOUT and Retries = NEIGHBOR_REQUEST_THRESHOLD

Receive REPLY

NEIGHBOR REPLIED

SELF CONFIGURED

Send ACK Start Timer2

Timer2 = ADDRESS_BLOCK_TIMEOUT and Retries = NEIGHBOR_REQUEST_THRESHOLD WAIT SERVER REPLY

Receive DENY

Receive ADDRESS_BLOCK

ADDRESS BLOCK RECEIVED

DENIED

Send CONFIRM CONFIGURED BY SERVER

Fig. 1. State diagram of a client node during the address allocation process of IPAA protocol

exist and searches for another server to perform the address allocation process again up to NEIGHBOR REQUEST THRESHOLD number of times. If ADDRESS BLOCK message has not been received in all trials, the client performs self allocation to configure itself an address as described above. In case that the selected server has no available addresses to serve the client, the server searches for an existing node in the partition that has an available addresses. For this purpose, each node maintains state information about the allocated address blocks in Allocated Blocks data set. The Allocated Blocks set lists the configured nodes in the partition with their available Address Blocks. The server selects the node with the largest available Address Block and sends it

434

A. Radaideh and J.N. Daigle

WAIT REQUEST

Receive REQUET

REQUEST RECEIVED

Send DENY

Timer3 = CONFIRM_TIMEOUT

Send REPLY Start Timer1 Timer1 = ACK_TIMEOUT

Receive CONFIRM Update Allocated_Blocks set

WAIT CONFIRM

Address space available Send ADDRESS_BLOCK

WAIT ACK

ADDRESS BLOCK NOT AVAILABLE

Receive ACK Address space not available CHECK AVAILABLE ADDRESS BLOCK Timer2 = BORROW_TIMEOUT and Retries != BORROW_THRESHOLD

Timer2 = BORROW_TIMEOUT and Retries = BORROW_THRESHOLD

Forward ADDRESS_BLOCK to the client ADDRESS BLOCK RECEIVED

Address space available in other nodes Send BORROW Start Timer2 Receive ADDRESS_BLOCK

WAIT ADDRESS BLOCK

Fig. 2. State diagram of a server node during the address allocation process of IPAA protocol

a BORROW message requesting half of its available space. Once the ADDRESS BLOCK message is received from the requested node, the server forwards the message to the client. If all nodes in the server Allocated Blocks set have no available address space, the server sends a DENY message to the client indicating that addresses are not available in this partition. Nodes could depart the network or move out their partitions at any time. If a node does not respond to the BORROW message within a timeout period of BORROW TIMEOUT, the server sends the borrow message to the node who has the second largest address space. The borrow process is repeated up to BORROW THRESHOLD number of times. If all trials have failed or no address space

Address Allocation Protocols

435

is available in the rest of the partition nodes, the server sends DENY message to the client indicating that addresses are not available in this partition. Network Partitioning and Merging. Nodes may depart the network or move out their partitions at any time. Departure of a node leads to address leak problem where the nodes address block will not be used by other existing nodes. Each node is responsible for cleaning up the Address Block of its missing buddy node. To achieve this, the Allocated Blocks set in each node has to be updated regularly. Each node in the partition broadcasts its Allocated Blocks set to every other node in the partition. A node updates its Allocated Blocks set when receiving other nodes sets to keep its information up to date. Since the state information a node has is assumed up to date, each node looks up its Allocated Blocks set from time to time to check the existence of its buddy node. If the buddy node is missing from the set, a node claims that its buddy has departed the network. Therefore, it merges its buddy node Address Block with its block. Each partition has a UUID to be identified from other existing partitions. Partitioning is detected if the node with the lowest address is missing. When detecting the partitioning event, each node sets the partition UUID to the lowest address currently allocated to a node in their partition. It is possible that a partition merges with another partition in the network. When two configured nodes get close to each other, they exchange their partition UUIDs. If the nodes UUID are different, a merging event is detected. The nodes that detect the merging event exchange their Allocated Blocks sets. Each node broadcasts the other node’s set to all nodes in its partition. When a node receives The Allocated Blocks set, it searches the set to determine whether a node with the same address exists in the other partition. If an address conflict is detected, the node with the larger address space gives up its address space and asks existing nodes for a new allocation. Merging of two partitions ends when all address conflicts are resolved. The partition UUID maintained by each node is updated to the lowest address allocated in the resulting partition.

3

Analytical Modeling of Address Allocation Protocols

In this Section, we derive analytical expressions of the expected values of latency and communication overhead that are used to evaluate the performance of IPAA [14]. The main objective of such analytical derivations is to obtain mathematical formulations that can clarify the impact of network characteristics and the number of existing nodes in the network on the performance measures of the address allocation process under consideration. The network characteristics that have an impact on the performance measures are the network area, the node’s coverage area, collisions and message loss rate. The derived formulas also show the impact of the protocol parameters which are the timeout values and the counter values on the performance measures of the selected protocol. The derivations of the expected latency and communication overhead are first carried out for small number of existing nodes in the network and then

436

A. Radaideh and J.N. Daigle

generalized for an arbitrary number of nodes. Section 3.1 presents the network model under consideration. It defines the network boundary, the node’s coverage area, the message loss as well as collision for the incoming traffic at the new node. In Section 3.2, derivations for the expected latency and communication overhead are conducted. 3.1

The Network Model

Let A represent the total area of the network. Each node in the network is located at a random position and nodes are assumed to be uniformly distributed over the network area. Nodes are assumed to have a coverage area of Ax . Let n ˜ denote the number of existing nodes in the network at the time a new node wishes to join. Let x ˜i indicate the presence of node i, i = 1, 2, · · ·, n, within the transmission range of the new node; that is, x ˜i = 1 is a neighbor of the newly entering node, and x ˜i = 0 otherwise. Since the existing nodes are uniformly distributed over the network area independent of the placement of all other nodes, x ˜i , i = 1, 2, · · ·, n is set of identical, independent, Bernoulli trials with success probability AAx . Thus, the number of existing nodes that are within the transmission rangeof the new node, denoted as x ˜, is a binomial random variable with parameters n, Ax and A E[˜ x|˜ n = n] = n P {˜ xi = 1|˜ n = n} = n

Ax . A

(1)

The new node communicates with a neighbor node over a wireless channel. A bad channel condition results in a low signal to interference plus noise ratio (SINR) for the received message and the message is considered lost if its SINR is below a given threshold. Assume that the wireless channel between the new node and a neighbor node has a message loss rate equals to . To avoid collision, each node in the network has a contention window of size W . A node that has data to send chooses a random slot number uniformly from {1, 2, · · ·, W } and sends its data in that selected slot. Collision in a given slot could happen if two or more nodes transmit in the same slot despite the SINR value of their messages. A message is received successfully if it does not collide with other messages in the transmission slot and has a good SINR value. 3.2

Expectations of the Performance Measures

The amount of time for the new node to be configured with a network address is denoted as ˜ and the number of messages sent during the address allocation process is denoted as c˜. The derivations for the expected values of latency and ˜ n = n and c˜|˜ communication overhead for a given number of nodes, |˜ n = n, respectively, are first constructed for the case n ˜ = 0 and then generalized to an arbitrary number of nodes n ˜ = n. Due to space limitations, only a sketch of the developments are presented here, and the interest reader is referred to [17].

Address Allocation Protocols

3.3

437

The n ˜ = 0 Case

In this case the new node is the only node in the network. The new node starts by broadcasting a REQUEST message. Since there are no nodes in the neighborhood, the timer for receiving the REPLY message will reach the timeout value TNR without receiving any REPLY message. Since no REPLY messages have been received during the first timeout period, the new node sends another REQUEST message and waits for timeout. The new node will repeat sending the REQUEST message and waits for timeout for a maximum of KNR trials. After the last timeout period, the new node considers itself the first node in the network and performs address configuration itself. It allocates the whole address space for itself and assigns itself the first address in that space. The expected latency for a new node to be allocated a network address, de  ˜ n = 0 , and the expected communication overhead, denoted as noted as E |˜ E [˜ c|˜ n = 0], for this case are easily found to be   E ˜ | n ˜ = 0 = KNR TNR and E [˜ c|n ˜ = 0] = KNR . (2) 3.4

The General Case n ˜=n

In this section we derive formulas for the expectations of latency and communication overhead for the IPAA protocol in the general case where the number of nodes in the network is n ˜ = n. Since this protocol performs address allocation through local communications with the new node neighbors, the expectations are derived by conditioning first on the number of nodes that are within the transmission range of the new node. The general formulas for the expected latency and communication overhead are then given by n      E ˜ | n ˜=n = E ˜ | x ˜ = x P {˜ x=x|n ˜ = n} ,

(3)

x=0

and E [˜ c|n ˜ = n] =

n 

E [˜ c | x˜ = x] P {˜ x=x|n ˜ = n} ,

(4)

x=0

  where x ˜ is a binomial random variable with parameters n, AAx . For the special case when there is no node within the transmission range of the new node, that is {˜ x = 0}, the expected latency and communication overhead are obtained from (2) and are given by E[˜ | x ˜ = 0] = KNR TNR

and E[˜ c|x ˜ = 0] = KNR .

(5)

For all other cases, where x ˜ > 0, the expectations for the latency and communication overhead for a given number of neighbors are calculated by conditioning

438

A. Radaideh and J.N. Daigle

on the number of trials, denoted by κ ˜ , required to successfully allocate a network address to the new node. Conditioning on the value of κ ˜ yields KNR

E[˜ | x ˜ = x] =



E[˜ | x ˜ = x, κ ˜ = κ] P {˜ κ =κ|x ˜ = x} +

κ=1

E[˜ | x ˜ = x, κ ˜ > KNR ] P {˜ κ > KNR | x ˜ = x} ,

(6)

and KNR

E[˜ c|x ˜ = x] =



E[˜ c|x ˜ = x, κ ˜ = κ] P {˜ κ=κ|x ˜ = x} +

κ=1

E[˜ c|x ˜ = x, κ ˜ > KNR ] P {˜ κ > KNR | x ˜ = x} .

(7)

The κ-th trial is successful if an ADDRESS BLOCK message is received from a neighbor node in response to the ACK message. The ADDRESS BLOCK message contains a disjoint subset of the address space where the new node selects the first address from that space for itself and keeps the rest for serving other joining nodes in future. Assuming that successive trials are independent and identical, the probability that a successful address allocation occurs on the κ-th trial is given by P {˜ κ=κ|x ˜ = x} = (P {˜ a= 0|x ˜ = x})κ−1 P {˜ a=1|x ˜ = x} ,

(8)

where a ˜ = 0 is the indicator random variable for the event of a successful allocation. As discussed earlier in previous sections, the trial fails if the ADDRESS BLOCK message is not received when at least one REPLY message is received or when no REPLY message is received. The probability of a failed trial in presence of x neighbors is given by P {˜ a = 0|˜ x = x} = P {˜ a = 0|˜ x = x, s˜ = 0} P {˜ s = 0|˜ x = x} + P {˜ a = 0|˜ x = x, s˜ = 1} P {˜ s = 1|˜ x = x} = P {˜ s = 0|˜ x = x} + ( + (1 − ) ) P {˜ s = 1|˜ x = x} ,

(9)

where s˜ is the indicator random variable for the event of the reception of at least one reply message from a neighbor node and that message is transmitted with no other messages in a slot and has a good SINR value. Define r˜ to be the number of responders who received a REQUEST message and r˜i to be the number of responders who responded in slot number i. i ∈ {1, 2, · · ·, W }. Then the probability of receiving at least one REPLY message successfully given that there are x neighbors is given by W x   P {˜ s = 1|˜ x = x} = P {˜ ri = 1}|˜ r = r P {˜ r = r|˜ x = x} . (10) r=1

i=1

Address Allocation Protocols

439

The first term of the equation represents the probability that at least one transmission slot has exactly one REPLY message and that message has a good SINR value. The second term represents the probability that r˜ neighbors received the REQUEST message out of the total number of neighbors. Since message transmissions are assume to have identical failure probabilities of  and messages are transmitted independently, r˜ is a binomial random variable with parameters (x, (1 − )). The probability that at least one slot has exactly one reply message given that there are r responders is equal to [18] W    P {˜ ri = 1} | r˜ = r = P {˜ ri = 1 | r˜ = r}− P {˜ ri = 1, r˜j = 1 | r˜ = r} + i=1

i



i<j

P {˜ ri = 1, r˜j = 1, r˜k = 1 | r˜ = r} − · · · +

i<j
(−1)W +1 P {˜ ri = 1, r˜j = 1,˜ rk = 1, · · ·, r˜W = 1 | r˜ = r} . (11) The first term of (11) represents the probability that slot i has exactly one message, in other words, exactly one responder transmitted the message in slot number i and that message when received had a good SINR value. That is



r−1 r 1 1 P {˜ ri = 1|˜ r = r} = 1− (1 − ). (12) 1 W W The other terms of (11) represent joint probabilities of having more than one slot with exactly one message. In general, the joint probability of having exactly one reply in each of the slots i, j, · · ·, z where the r˜i + r˜j + · · · + r˜z = h, is equal to zero for h > r and for the h ≤ r case it is calculated as

 r (W − h)r−h P {˜ ri1 = 1, · · ·, r˜zh = 1|˜ r = r} = h! (1 − )h . (13) h (W )r Substituting this result in (11) and performing some algebra yields W min{r,W }

    r (W − h)r−h h+1 W P {˜ ri = 1} | r˜ = r = (−1) h! (1 − )h . r h h W i=1 h=1 (14) Now, we consider the conditional expectations of the latency and communication overhead on the number of trials required for the new node to be allocated a network address. Recall that κ ˜ represents the number of trials that has been performed, the expected latency when the address allocation succeeded on the κth trial, where κ ≤ KNR is the sum of of the timeouts for the first (κ − 1) failed trials plus the timeout of receiving a REPLY message and the timeout of receiving the ADDRESS BLOCK message in the last successful trial. The latency for a failed trial is the sum of the timeout to receive the REPLY message plus the timeout to

440

A. Radaideh and J.N. Daigle

receive the ADDRESS BLOCK message if at least one reply is successfully received. That is ˜ x = x, κ E[|˜ ˜ = κ] = (κ − 1) (TNR + P {˜ s = 1|˜ x = x} TAR ) + TAR + TAR = κ TNR + TAR (1 + (κ − 1) P {˜ s = 1|˜ x = x}), (15) where P {˜ s = 1|˜ x = x} is presented in (10). The expected communication overhead for a successful address allocation on the κ-th trial is the sum of messages sent during the first (κ − 1) trials and they are one REQUEST message, the expected number of REPLY messages and one ACK message that is sent if at least one reply is received correctly. In the last trial, the new node sent a REQUEST and received at least one REPLY so it sent out an ACK message to one of the responders and received an ADDRESS BLOCK message that contains a portion of the address space. Therefore, the expected communication overhead is equal to E[˜ c|˜ x = x, κ ˜ = κ] = (κ − 1) (1 + E[˜ r |˜ x = x] + P {˜ s = 1|˜ x = x}) + (1 + E[˜ r|˜ x = x] + 2) = κ (1 + E[˜ r |˜ x = x]) + (κ − 1) P {˜ s = 1|˜ x = x} + 2. (16) In the case where all trials have failed, that is κ > KNR , the expected latency and communication overhead are given by ˜ x = x, κ E[|˜ ˜ > KNR ] = KNR (TNR + P {˜ s = 1|˜ x = x} TAR ),

(17)

E[˜ c|˜ x = x, κ ˜ > KNR ] = KNR (1 + E[˜ r |˜ x = x] + P {˜ s = 1|˜ x = x}).

(18)

and

Substituting (8), (15), (16), (17) and (18) into (6) and (7) gives E[˜ | x ˜ = x] =

KNR



κ TNR + TAR (1 + (κ − 1) P {˜ s = 1|˜ x = x})

κ=1

(P {˜ a = 0|x ˜ = x})κ−1 P {˜ a = 1|x ˜ = x} + KNR (TNR + P {˜ s = 1|˜ x = x} TAR ) (P {˜ a= 0|x ˜ = x})KNR , and E[˜ c|x ˜ = x] =

KNR



(κ (1 + E[˜ r|˜ x = x]) + (κ − 1) P {˜ s = 1|˜ x = x} + 2)

κ=1

(P {˜ a= 0|x ˜ = x})κ−1 P {˜ a = 1|x ˜ = x} + KNR (1 + E[˜ r |˜ x = x] + P {˜ s = 1|˜ x = x}) (P {˜ a=0|x ˜ = x})KNR , (19) where the probabilities P {˜ a= 0|x ˜ = x} and P {˜ s = 1|˜ x = x} are derived in (9) and (10), respectively.

Address Allocation Protocols

4

441

Numerical Results

In this section, we present numerical results for latency and communication overhead based on analysis and simulation, and in addition, we consider the sensitivity of these measures to message loss rate, contention window size and coverage ratio. in Subsection 4.1, a description of the simulation carried out for the IPAA and MANETconf protocols is presented. Then, Subsection 4.2 presents a comparison of the IPAA and MANETconf protocols. Subsection 4.3 presents the analytical and simulation results, which show good agreement, for the IPAA protocol as functions of network size with different parameter values. A summary of results is deferred to the conclusion section. 4.1

Simulation Description

The network dimensions are set to 100m × 100m. Each node is placed at a uniformly distributed location within the network area. Results are collected for node populations from 0 to 50. For a given number of nodes in the network, the number of nodes within the transmission range of the new node depends on its coverage area. Results are collected for coverage ratios 10%, 15%, and 20%. Each message sent from one node to another is subject to loss with rate ; loss rates examined were 0, 0.1, and 0.2. Contention window sizes, W , considered were 10, 20, and 30. For the IPAA protocol, the timeout value to find a neighbor TNR was set to 0.2 seconds and the timeout to receive the address block message TAR was set to 0.02 seconds, and KNR was varied from 3 to 5 to show the effect of the counter threshold on the performance measures of the protocol. For MANETconf protocol TNR was set to 0.2 seconds and TAR was set to 2 seconds, and KNR was set to 3 and KAR was set to 5. The simulation results presented are the average values of 10000 simulation runs. For large number of simulation runs, the sample mean of a performance measure for any network size follows the normal distribution N (μ, √sn ), where μ is the true population mean, s is the sample standard deviation, and n is the sample size. Throughout simulation, we have found that the maximum value of the sample standard deviation is 0.002 for latency samples and it is 0.1 for communication overhead samples for all values of network size. The 95% confidence interval, that is likely to include the true population mean of a performance measure, is equal to (¯ a ± 1.96 √sn ), where a ¯ represents the sample mean for the performance measure. Therefore, the 95% confidence interval for latency samples is (¯ x − 3.92 × 10−5 , x ¯ + 3.92 × 10−5 ) and for communication overhead it is (¯ y − 0.002, y¯ + 0.002). 4.2

IPAA vs. MANETconf Protocol

Figure 3 compares the latency of IPAA protocol to the MANETconf protocol latency. In MANETconf, the initiator node selects an address and performs DAD process throughout the network. Therefore, the latency of the allocation process

442

A. Radaideh and J.N. Daigle

Fig. 3. Latency of the IPAA and MANETconf protocols as a function of network size

increases as the number of node in the network increases. In IPAA protocol, the address allocation process is carried out by an existing neighbor or it may be carried out by the new node itself in case no neighbor exists. For large number of nodes, the probability of finding a node within the transmission range of the new node is high so that address allocation can be carried out by a neighbor node and therefore it decreases as the number of nodes increases. Figure 4 presents the communication overhead for IPAA and MANETconf protocols. Since MANETconf performs the DAD process throughout the network, the number of messages sent increases as the number of nodes increases. The IPAA protocol performs address allocation through local communications with neighbor nodes only. Therefore, the number of messages sent is less than that for the MANETconf protocol. 4.3

Analytical and Simulation Results for the Performance Measures of the IPAA Protocol

Figures 5 shows the latency at loss rates  = {0, 0.1, 0.2}. The contention window size is set to W = 30 and the coverage ratio is set to AAx = 10%. The results show that the latency increases as the loss rate increases since a loss of the protocol messages may results in a failed trial and force the new node to start the process again. Communications overhead was also analyzed. For relatively small number of nodes in the network, the communication overhead increases as the loss rate increases since there exists a small number of neighbor nodes within the transmission range of the new node so the probability of loss of the protocol message is high. For large number of neighbors, the probability of loss for all messages is lower and the number of messages becomes closer to the case

Address Allocation Protocols

443

Fig. 4. Communication overhead of the IPAA and MANETconf protocols as a function of network size

Fig. 5. Latency of the IPAA protocol as a function of network size with message loss rate as a parameter

where no loss is assumed. The graph showing these results is omitted because it is so similar to the results shown in Figure 4. Figures 6 and 7 show the effect of the contention window size W on the latency and the communication overhead, respectively when  = 0.1, AAx = 10% and KNR = 3. The window size has a greater effect on latency and communication

444

A. Radaideh and J.N. Daigle

Fig. 6. Latency of the IPAA protocol as a function of network size with contention window size as a parameter

Fig. 7. Communication overhead of the IPAA protocol as a function of network size with contention window size as a parameter

overhead for a large number of nodes since the collision probability increases as the number of nodes increases. For a given large number of nodes, decreasing the window size results in more collisions for the incoming reply messages at the new node which may results in failed allocation trial. More failed trials results in more latency and more communication overhead as shown in the figures.

Address Allocation Protocols

445

Numerous additional analytical and simulation results were carried out and are given in [17], but are omitted here due to lack of space. Also analyzed were the the effect of the coverage ratio AAx on the latency and communication overhead, effect of the protocol parameter KNR on the latency and communication overhead, and effect of the loss ratio on the number of neighbor allocations. In all cases analytical and simulation results showed close agreement.

5

Conclusions

Two address allocation protocols in wireless ad hoc network were discussed. The MANETconf protocol [13], represents the DAD-based address allocation scheme while the IPAA protocol [14], represents the neighbor-based scheme. A state machines that shows the behavior was shown for the IPAA protocol while that for the MANETconf was deferred to [17]. The state machine gives a complete picture of the states, handshakes, timers, and types of messages for a protocol. For IPAA, analytical formulas for the expectation of latency and communication overhead are carried out as a function of the number of nodes in the network with message loss rate, contention window size, coverage ratio, and the counter threshold as parameters. Extensive numerical results were collected. The results show that the latency and communication overhead for MANETconf are higher than the measures of the IPAA protocol and that the increase results from performing the DAD process in MANETconf. For IPAA, it has been shown that the latency and communication overhead of the allocation process increase as the message loss rate increases. The contention window size has an effect on the performance measures for a relatively large number of nodes in the network. Extensive additional numerical results and conclusions are given in [17]. As an example, numerical results not presented here show that for loss rate of 0.1, contention window size of 30, and coverage ratio of 10%, address allocation is carried out by neighbor nodes for almost 95% of the time when the number of nodes in the network more than 30. The address allocation problem for ad hoc networks is still unsolved. The work done in this paper provides understanding of the nature of the problem and the way to evaluate the performance of an allocation protocol. What is needed is a protocol that handles all types of exceptions such as message losses, node departures, network partitioning, and merging. The performance of the protocol should be analyzed and compared to the performance of existing protocols.

References 1. Baccelli, E.: Address autoconfiguration for MANET: Terminology and problem statement, IETF draft-ietf-autoconf-statement-02 (September 2007) 2. Knowlton, K.C.: A fast storage allocator. Commun. ACM 8(10), 623–624 (1965) 3. Bernardos, C., Calderon, M., Moustafa, H.: Survey of IP address autoconfiguration mechanisms for MANETS, IETF draft-bernardos-manet-autoconf-survey-03 (April 2008)

446

A. Radaideh and J.N. Daigle

4. Perkins, C.: IP address autoconfiguration for ad hoc networks, IETF draft-perkinsmanet-autoconf-01 (November 2001) 5. Jeong, J.: Ad hoc IP address autoconfiguration, IETF draft-jeong-adhoc-ip-addrautoconf-06 (January 2006) 6. Vaidya, N.: Weak duplicate address detection in mobile ad hoc networks. In: Proc. of ACM MOBIHOC, Lausanne (2002) 7. Weniger, K.: Passive duplicate address detection in mobile ad hoc networks. In: Proc. of IEEE WCNC (March 2003) 8. Mase, K., Adjih, C.: No overhead autoconfiguration OLSR, IETF draft-masemanet-autoconf-noaolsr-01 (April 2006) 9. Weniger, K.: PACMAN: Passive autoconfiguration for mobile ad hoc networks. IEEE Journal on Selected Areas in Communications 23(3), 507–519 (2005) 10. Zhou, H., Ni, L., Mutka, M.: Prophet address allocation for large scale MANETs. In: Proceedings of INFOCOM (2003) 11. Yong, L., Ping, Z., JiaXiong, L.: Dynamic address allocation protocols for mobile ad hoc networks based on genetic algorithm. In: Proc. of 5th International Conference on Wireless Communications, Networking and Mobile Computing (2009) 12. Chu, X., Sun, Y., Xu, K., Sakander, Z., Liu, J.: Quadratic residue based address allocation for mobile ad hoc networks. In: Proc. IEEE ICC (2008) 13. Nesargi, S., Parakash, R.: MANETconf: Configuration of hosts in a mobile ad hoc network. In: Proc. IEEE INFOCOM. IEEE Press, New York (2002) 14. Mohsin, M., Parakash, R.: IP address assignment in a mobile ad hoc network. In: Proc. IEEE MILCOM, IEEE Press, New York (2002) 15. Tayal, A.P., Patnaik, L.M.: An address assignment for the automatic configuration of mobile ad hoc networks. Personal Ubiquitous Comput. 8(1), 47–54 (2004) 16. Xu, T., Wu, J.: Quorum based ip address autoconfiguration in mobile ad hoc networks. In: ICDCSW 2007: Proceedings of the 27th International Conference on Distributed Computing Systems Workshops, p. 1. IEEE Computer Society, Washington (2007) 17. Radaideh, A.M.: Analytical modeling of address allocation protocols in wireless ad hoc networks, Master’s thesis, The University of Mississippi (December 2008) 18. Ross, S.M.: Introduction to Probability Models, 9th edn. Academic Press, London (2008)

Analysis of One-Hop Packet Delay in MANETs over IEEE 802.11 DCF Jun Li1 , Yifeng Zhou1 , Louise Lamont1 , and Camille-Alain Rabbath2 1

Communications Research Centre, Ottawa, ON, Canada K2H 8S2 {jun.li,yifeng.zhou,louise.lamont}@crc.gc.ca 2 DRDC-Valcartier, Quebec, QC, Canada G3J 1X5 [email protected]

Abstract. In mobile ad hoc networks (MANETs), the estimation of packet end-to-end delay depends on that of one-hop packet delay. In this paper, we conduct an analysis of the one-hop packet delay in MANETs, where the medium access control (MAC) layer uses the IEEE 802.11 distributed coordination function (DCF) to share the medium. In the MANET, each node runs the IEEE 802.11 DCF and a routing protocol. It is assumed that all nodes are one-hop neighbors, and that any pair of nodes can send data over the wireless channel with a fixed data rate. The light traffic condition is used, i.e., each node generates packets at the network layer according to a Poisson process. By modeling each wireless node as an M/M/1 queueing system, we derive the mean onehop packet delay analytically under the light traffic condition. Simulation analysis is carried out to verify the derived results. Results show that the mean one-hop packet delay increases with either the network size or the packet generation rate in networks subject to the light traffic condition. The mean one-hop packet delay derived in this paper is analytical and exact for networks under the light traffic condition. Results that can be found in the literature are usually based on the heavy traffic condition, and they tend to overestimate by a large amount the mean one-hop delay for networks with light traffic. Keywords: Mobile ad hoc networks (MANETs), medium access control (MAC), IEEE 802.11, distributed coordination function (DCF), modeling and analysis, M/M/1 queue, packet sojourn time, one-hop delay.

1

Introduction

A mobile ad hoc network (MANET) consists of a collection of wireless mobile nodes that are dynamically connected. The mobile nodes can communicate with each other without the assistance of any pre-existing or centralized infrastructure. Because MANETs can be deployed rapidly and operated with no single point of failure in a whole network, the MANET technology plays a crucial role in military networks. 

This work was supported by Defence Research and Development Canada (DRDC).

J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 447–456, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

448

J. Li et al.

An envisaged military application of a mobile network is a team of cooperative unmanned aerial vehicles, or UAVs, carrying out intelligence, surveillance, and reconnaissance (ISR). Cooperative control entails planning, coordination, and execution of a mission by two or more UAVs. Ideally, humans issue high-level commands to the robot team, such as searching an area for a certain amount of time. The team then responds to such orders, in the most efficient way. To act as an effective team, the UAVs will possess a certain level of autonomy in the execution of the mission enabled by their onboard systems and the sharing of information. The information passed from one UAV to another may pertain to vehicle state (position, velocity), vehicle and systems operational status (for example, current condition and quantity of energy), flight mode (for example, cruising or maneuvering), and mission-specific data (such as sensor information passed on to the commander). The probability of delays and packet loss increases with the amount of information communicated from one UAV to another. Cooperative control systems can usually tolerate relatively small delays. However, if the delays are too long, a team of UAVs may be destabilized, and safety of nearby humans is at risk. Therefore, it is of high importance to mitigate the effect of the delays on the multi-UAV cooperative control systems. In military network applications, information delay (or packet delay) is an important design issue, since it has a significant impact on the performance of a system of systems (e.g., networked control systems) that are connected with MANETs [8]. In MANETs, since nodes usually communicate over a same wireless channel, a medium access control (MAC) protocol is required to regulate the access of multiple nodes to the shared medium during packet transmission. The distributed coordination function (DCF), a MAC protocol in the IEEE 802.11 standard, has been widely used in MANETs [14]. It is a random access scheme based on a distributed channel access scheme denoted by carrier sense multiple access with collision avoidance (CSMA/CA) [1]. In DCF, a node having data to transmit contends for the shared medium using CSMA/CA. In addition, the request-to-send/clear-to-send (RTS/CTS) handshaking mechanism is used to tackle the hidden terminal problem, which can occur in all types of wireless networks, including MANETs [9]. In a delay-guaranteed MANET application, packet end-to-end delay is one of the most important metrics. It follows that accurately estimating packet end-to-end delay becomes an essential but challenging part of network analysis. Since packets often traverse multiple hops before reaching their destinations in MANETs, the end-to-end delay of a packet accumulates the amount of time consumed on each hop that the packet passes through. As a result, precisely estimating packet end-to-end delay is directly related to how accurately the one-hop packet delay is computed or approximated. In this paper, we conduct an analysis of one-hop packet delay in an ad hoc network with IEEE 802.11 DCF being the underlying MAC protocol. This is the first phase of our study on the packet delay in DCF-based MANETs. More specifically, we consider a fixed number of wireless nodes, and each node runs the IEEE 802.11 DCF and a routing protocol (e.g., the optimized link state routing

Analysis of One-Hop Packet Delay

449

(OLSR) protocol [5]). It is assumed that all nodes are one-hop neighbors, and a pair of nodes can communicate data over the wireless channel with a fixed data rate. In addition, it is assumed that each wireless node generates packets (referred to as IP (internet protocol) packets in this paper) at the network layer according to a Poisson process. The generated packets are sent down to the MAC layer for transmission to the shared medium. In this paper, the Poisson process assumption for traffic is referred to as the light traffic condition. We define onehop delay of an IP packet as the duration between the time of creation of the IP packet at a node and the time of its reception at a neighbor node. In the literature, several studies on packet service time at the MAC layer have been reported when IEEE 802.11 DCF was assumed. In [2], an analytic and accurate model was developed to compute the throughput of IEEE 802.11 DCF. Based on the original or a modified version of the analytic model, the mean packet service time at the MAC layer was analyzed in [4], [6], [7], [13], [16]. All these studies were conducted under the heavy traffic condition, i.e., packets are always available for transmission at each node. In [15], the packet service time at the MAC layer was analyzed under a non-heavy traffic condition for IEEE 802.11 DCF. It was claimed that the packet service time at the MAC layer can be better approximated by an exponentially distributed random variable. However, due to the complexity involved in computing the mean packet service time, the numerical analysis was only performed for certain values of the packet collision probability. The packet collision probability is expressed as a function of the mean packet service time together with other parameters. In [11], the mean onehop packet delay was derived analytically for IEEE 802.11 DCF-based ad hoc networks under the light traffic condition. However, the expression for the mean packet service time at the MAC layer assumed a heavy traffic condition. This paper focuses on providing an expression for the mean one-hop packet delay under the light traffic condition. It also provides an expression for the mean packet service time at the MAC layer under the same condition. In doing so, we consider each node to be an M/M/1 queueing system, in which the packet sojourn time corresponds to the one-hop packet delay. The packet sojourn time is the sum of the packet service time at the MAC and the network layer. Although a standard result of the queuing theory can apply to the mean packet sojourn time (i.e., the mean one-hop packet delay), it requires the knowledge of the packet service time at the MAC layer. To calculate the mean packet service time at the MAC layer, we derive a non-linear system of equations, from which it can be numerically computed. Simulation analysis is carried out to verify the analytic result for the mean one-hop packet delay. Numerical and simulation results show that an estimate of the mean one-hop packet delay developed in [4] under the heavy traffic condition overestimates by a large amount the mean one-hop delay for networks with light traffic. The rest of the paper is organized as follows: Section 2 describes ad hoc networks operating using IEEE 802.11 DCF as the MAC layer protocol. In Section 3, a systematic method is developed to estimate the mean one-hop packet delay. Numerical and simulation examples are presented to verify the resulting mean

450

J. Li et al.

one-hop delay expression in Section 4. Finally, concluding remarks are given in Section 5.

2

MANETs over IEEE 802.11 DCF

Assume that N wireless nodes are distributed in an area. Each node runs the IEEE 802.11 DCF and a MANET routing protocol (e.g., OLSR or the dynamic source routing (DSR) protocol [10]). It is assumed that IP packets, which could be either data or routing protocol control messages, are generated in each node according to a Poisson process with a rate of λd packets per second. In this study, we assume that all IP packets have the same length. Any two wireless nodes are one-hop neighbors and they can communicate with each other over a shared wireless channel with a same data rate. The wireless channel is assumed to be error-free for transmitting data, and thus errors are only caused by collisions due to simultaneous packet transmission to the shared medium. A node having an IP packet to transmit accesses the wireless channel based on the IEEE 802.11 DCF with the RTS/CTS mechanism. The parameters in the IEEE 802.11 DCF are defined below. The reader is referred to [15] for a demonstration of the transmission process and to [1] for details of the protocol. – σ is the slot time size that equals the time needed by a node to detect transmission of a packet from any other node; – Trts is the time to transmit a RTS packet (including the physical layer header); – Tcts is the time to transmit a CTS packet (including the physical layer header); – Tack is the time to transmit an acknowledgement (including the physical layer header); – Th is the time to transmit the header (including the MAC and the physical layer header) of an IP packet; – Td is the time to transmit an IP packet (either a data or a control packet); – SIF S represents the short inter-frame space; – DIF S represents the distributed inter-frame space; – CWmin is the size of the initial contention window; – CWmax is the size of the maximum contention window; – R represents the retry limit of transmitting MAC frames, including RTS, CTS, and acknowledgement packets. We define the one-hop delay of an IP packet as the time interval between the instant the packet is generated at the node and the instant it is received at a destined neighbor node.

3

Analysis of One-Hop Packet Delay

In this section, a non-linear system of equations is derived for solving the mean packet service time at the MAC layer. The mean one-hop packet delay is obtained from the mean packet service time at the MAC layer.

Analysis of One-Hop Packet Delay

3.1

451

Mean Packet Service Time at MAC Layer

In [15], it is observed that the packet service time at the MAC layer approximately follows an exponential distribution. We define the following parameters. 1. CWi is the contention window size of a packet for its ith retransmission with the exponential back-off policy, for i = 0, 1, · · · , R, i.e., CWi = 2i (CWmin + 1) − 1,

i = 0, 1, · · · , R;

(1)

2. Ts is the time period during which the medium is sensed busy due to successful transmission, i.e., Ts = Trts + Tcts + Th + Td + Tack + 3SIF S + DIF S;

(2)

3. Tc is the time period during which the medium is sensed busy due to a collision, i.e., Tc = Trts + SIF S + Tack + DIF S.

(3)

In the network steady state, the probability τ that a station transmits during a generic time slot, given that the station has one or more packets to transmit, is given by [7], τ=

2(1 − 2pc )(1 − pR+1 ) c . (CWmin + 1)(1 − pc )(1 − (2pc )R+1 ) + pc (1 − 2pc )(1 − pR c )

(4)

In (4), pc is the probability that a packet collides with another packet when transmitted during a generic time slot, and it is given by, pc = 1 − [1 − (1 − p0 )τ ] = 1 − (1 − E[S]λd τ )

N −1

N−1

,

(5)

where, p0 is the probability that a station doesn’t have a packet to transmit, p0 = 1 − E[S]λd for an M/M/1 queue (e.g., see (8.6) in [12]), and S is the packet service time at the MAC layer. Equations (4) and (5) form a system of two equations with three unknowns, τ , pc , and the mean packet service time E[S]. A third equation in terms of these unknowns is required to solve the equation system. In the following, we derive an expression of E[S] in terms of τ and pc , which will be combined with (4) and (5) to solve for the three unknowns. Consider an arbitrary packet to be transmitted by a station and let E[S] be its mean service time at the MAC layer. As defined in [2], let Ptr denote the probability that there is at least one transmission during a generic time slot, and PS be the probability that the transmission occurring in one station at a generic time slot is successful given that at least one station transmits during the time slot. Since, during a generic time slot, the packet is transmitted with probability τ and each of the remaining N − 1 stations transmits with probability E[S]λd τ , then, Ptr = 1 − (1 − τ ) (1 − E[S]λd τ )

N −1

.

(6)

452

J. Li et al.

There are two disjoint events that contribute to the probability that exactly one station transmits during a time slot. The first event is that the packet is transmitted while the remaining N −1 stations do not transmit; the second event is that the packet is not transmitted and exactly one of the remaining N − 1 stations transmits. Therefore, PS =

τ (1 − E[S]λd τ )

N −1

+ (N − 1)E[S]λd τ (1 − τ ) (1 − E[S]λd τ ) Ptr

N −2

.

(7)

By [2] (or [4]), E[SLOT ] = (1 − Ptr )σ + Ptr PS Ts + Ptr (1 − PS )Tc . By conditioning on the event that the packet is not dropped, we have,   R    R+1 R+1 E[S] = E[SLOT ] pc E Ui + E[X](1 − pc ) ,

(8)

(9)

i=0

where Ui ∼ U nif orm(0, CWi ),

i = 0, 1, · · · , R,

(10)

and X=

R  pic − pR+1 c Ui . R+1 1 − p c i=0

(11)

Moreover, E[Ui ] = and

CWi , 2

i = 0, 1, · · · , R,

 R   R R i R+1  pi − pR+1 c c i=0 pc E[Ui ] − pc i=0 E[Ui ] E[X] = E Ui = . R+1 R+1 1 − pc 1 − pc i=0

(12)

(13)

By substituting (12) and (13) into (9),   R R R    R+1 i R+1 E[S] = E[SLOT ] pc E [Ui ] + pc E[Ui ] − pc E[Ui ] i=0

= E[SLOT ] 

R 

i=0

i=0

pic E[Ui ]

i=0

(CWmin + 1)(1 − pc )(1 − (2pc )R+1 ) − (1 − pR+1 )(1 − 2pc ) E[SLOT ] c = . 2(1 − pc )(1 − 2pc ) (14) We now obtain a system consisting of equations (4), (5), and (14) with three unknowns τ , pc , and E[S]. The mean packet service time E[S] can be numerically computed from these three equations.

Analysis of One-Hop Packet Delay

3.2

453

Mean One-Hop Packet Delay

When a node is modeled as an M/M/1 queueing system, the one-hop delay of a packet corresponds to the sojourn time of the packet in the system. Using a standard queueing result (e.g., the mean one-hop packet delay E[W ] under the light traffic condition is given by, E[W ] =

4

1 E[S]

1 E[S] = . 1 − E[S]λd − λd

(15)

Numerical and Simulation Results

In this section, simulations are performed to verify the mean one-hop packet delay result obtained in Section 3. An ad hoc network with the underlying distributed coordination function of IEEE 802.11 is simulated using the OPNET simulator. The routing protocol used is either OLSR or DSR. The system consisting of (4), (5), and (14) is coded in MATLAB to numerically compute the mean packet service time E[S], and then the mean one-hop packet delay E[W ] from (15). For the purpose of comparison, the mean packet service time developed in [4] based on the heavy traffic assumption is also computed. The wireless channel data rate is set to be 2 M bps, and the values of the parameters used in the network and MAC layers are given in Table 1. The network size N and packet arrival rate λd vary in the analysis. The varying arrival rate enables a higher degree of realism in the simulations for many practical applications such as multiple UAV networking. In multiple UAV operations, the arrival rate changes between a pair of UAVs due for example to changes in their inter-vehicle separations, and cluttered conditions in urban environments. Table 1. Parameters in Network and MAC Layers Parameter Value PHY header 192 bits MAC header+PHY header (Th ) 224 bits+192 bits (208 μs) RTS packet (Trts ) 160 bits+PHY header (176 μs) CTS packet (Tcts ) 112 bits+PHY header (152 μs) ACK packet (Tack ) 112 bits+PHY header (152 μs) IP packet (Td ) 8184 bits (4092 μs) σ 20 μs SIF S 10 μs DIF S 50 μs CWmin 31 R 6

In Fig. 1, we plot the mean one-hop packet delay results and those obtained through simulation. In the plot, the rate λd of generated IP packets at each node is set to be 8 packets per second, which is only applicable under the light

454

J. Li et al.

65 60

Analytic under Heavy Traffic Analytic under Light Traffic Simulation with DSR Simulation with OLSR

Mean one−hop packet delay E[W]

55 50 45 40 35 30 25 20 15 10 5 0

4

6

8 10 Network size N

12

14

Fig. 1. Mean One-Hop Packet Delay E[W ] vs. Network Size N 55

Mean one−hop packet delay E[W]

50 45 40 35

Analytic under Heavy Traffic Analytic under Light Traffic Simulation with DSR Simulation with OLSR

30 25 20 15 10 5 0

1

2

3

4

5 6 Rate λ

7

8

9

10

d

Fig. 2. Mean One-Hop Packet Delay E[W ] vs. Rate λd

traffic condition, and the network size N varies from 4 to 14. When the network is under the light traffic condition, we use the mean packet service time at the MAC layer developed in [4] under the heavy traffic condition to see if it is suitable for

Analysis of One-Hop Packet Delay

455

approximating E[W ]. (In [4] the heavy traffic condition was simulated as opposed to our M/M/1 system modeled for light traffic.) From Fig. 1, we observe that the mean one-hop packet delay increases with the network size N . Moreover, the mean service time at the MAC layer computed under the heavy traffic condition overestimates by a large amount the value obtained by simulation when the network is under the light traffic condition. For a fixed rate λd , the discrepancy between the heavy traffic results and the simulation results increases as the network size N increases. For a fixed network size N , the mean one-hop packet delay for the two routing protocols is close to each other with DSR providing slightly larger mean packet delays than OLSR. More importantly, the numerical and simulation results in Fig. 1 show that the analytic result (15), which is derived based on the light traffic condition, is able to accurately estimate the mean one-hop packet delay with negligible differences between the analytic and simulation results. The mean one-hop packet delay is plotted in Fig. 2 for N = 12 and λd varying from 1 to 10 (applicable only under the light traffic condition). As shown in Fig. 2, the mean one-hop packet delay calculated with the heavy traffic assumption is independent of the rate λd but overestimates the mean one-hop delay when networks are in fact under the light traffic condition. From Fig. 2, it is observed that the mean one-hop delay increases with the rate λd . Similar to the observation in Fig. 1, the mean one-hop delay is almost the same for the two routing protocols. Again, the numerical and simulation results show that our result (15) can accurately estimate the mean one-hop packet delay for a network with light traffic.

5

Conclusion

In this paper, an analysis of the mean one-hop packet delay in ad hoc networks using the IEEE 802.11 DCF as a MAC protocol was carried out. The light traffic condition was assumed. By considering a network node as an M/M/1 queueing system, an analytic result was obtained for accurately estimating the mean onehop packet delay. Simulation analysis was performed to verify the accuracy of the analytic result. The numerical and simulation examples showed that the mean one-hop packet delay in a network of light traffic increases with either the network size or the packet generation rate. It was shown that the result analytically derived based on the heavy traffic condition tends to overestimate by a large amount the mean one-hop delay when the network is under the light traffic condition.

References 1. Part 11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications: Amendment 4, further higher data rate extension in the 2.4 GHz band. IEEE Standard 802.11g (2003) 2. Bianchi, G.: Performance analysis of the IEEE 802.11 distributed coordination function. IEEE Journal on Selected Area in Communications 18, 535–547 (2000)

456

J. Li et al.

3. Bisnik, N., Abouzeid, A.A.: Queueing network models for delay analysis of multihop wireless ad hoc networks. In: Proc. of the International Conference on Wireless Communications and Mobile Computing (IWCMC 2006), Vancouver, Canada, pp. 773–778 (2006) 4. Chatzimisios, P., Boucouvalas, A.C., Vitsas, V.: IEEE 802.11 packet delay - a finite retry limit analysis. In: Proc. of the IEEE Global Telecommunications Conference (Globecom 2003), San Francisco, USA, pp. 950–954 (2003) 5. Clausen, T., Jacquet, P. (eds.): Optimized link state routing protocol (OLSR). The Internet Engineering Task Force (IETF) Network Working Group, RFC 3626, Experimental (2003) 6. Elarbaoui, I., Refai, H.H.: Enhancement of IEEE 802.11 DCF backoff algorithm under heavy traffic. In: Proc. of 2008 IEEE/ACS International Conference on Computer Systems and Applications, Doha, Qatar, pp. 1082–1087 (2008) 7. Kang, K., Lin, X., Hu, H.: An accurate MAC delay model for IEEE 802.11 DCF. In: Proc. of the 2007 IEEE International Conference on Telecommunications and Malaysia International Conference on Communications, Penang, Malaysia, pp. 654–657 (2007) 8. Kopp, C.: NCW101: an introduction to network centric warfare, Part 4 - Ad Hoc Networking, Air Power Australia (2008) 9. Kumar, S., Raghavan, V.S., Deng, J.: Medium access control protocols for ad hoc wireless networks: A survey. Elsevier Ad Hoc Networks 4, 326–358 (2006) 10. Johnson, D., Maltz, D., Hu, Y.-C.: The dynamic source routing protocol for mobile ad hoc networks (DSR). The Internet Engineering Task Force (IETF) Network Working Group, RFC 4728, Experimental (2007) 11. Naimi, A.M., Jacquet, P.: One-hop delay estimation in 802.11 ad hoc networks using the OLSR protocol. INRIA research report no. 5327 (2004) 12. Ross, S.M.: Introduction to Probability Models, 8th edn. Academic Press, San Diego (2003) 13. Xiao, Y.: Performance analysis of IEEE 802.11e EDCF under saturation condition. In: Proc. of IEEE International Conference on Communications, Paris, France, pp. 170–174 (2004) 14. Xu, K., Gerla, M., Bae, S.: Effectiveness of RTS/CTS handshake in IEEE 802.11 based ad hoc networks. Elsevier Ad Hoc Networks 1, 107–123 (2003) 15. Zhai, H., Kwon, Y., Fang, Y.: Performance analysis of IEEE 802.11 MAC protocols in wireless LANs. Wiley Wireless Communications and Mobile Computing 4, 917– 931 (2004) 16. Ziouva, E., Antonakopoulos, T.: CSMA/CA performance under high traffic conditions: Throughput and delay analysis. Elsevier Computer Communications 25, 313–321 (2002)

Performance of Packet-Based Frequency-Hopping Spread Spectrum Radio Control Systems Abdallah Ismail1, Ioannis Lambadaris1, Chung-Horng Lung1, and Nishith Goel2 1

Department of Systems and Computer Engineering Carleton University, Ottawa, Ontario, Canada [email protected], {ioannis,chlung}@sce.carleton.ca 2 Cistel Technology, Ottawa, Ontario, Canada [email protected]

Abstract. Real-time Radio Control (RC) systems require instantaneous response in the controlled device. RC systems have wide applications, including ad hoc networks. Imperfections in the wireless channel (noise and interference) result in randomly fluctuating latency in the response of the system. A lag occurs when the system latency exceeds a specified real-time threshold. System Lag Occurrence Probability (SLOP) is the probability of lag occurrence and is derived as the performance metric to characterize user experience in real-time radio control systems. Frequency hopping is used to mitigate interference effects. Uniform serial acquisition and N-state lock detection are used to simplify the derivation of SLOP. Simulation results are presented to verify the derivation of SLOP. Keywords: radio control, frequency hopping, human response time.

1 Introduction In remote control applications, a user controls a number of inputs which trigger a proportional movement in servomechanisms mounted in the controlled device. The servomechanisms in turn move the control surfaces of the device thus causing it to maneuver in its medium. Such applications are generally referred to as real-time Radio Control (RC) applications. Examples of real-time RC applications range from a simple model car Radio Control (shown in Figure 1) to highly mobile Unmanned Aerial Vehicles (UAVs) [1]. The controlled devices could form an ad hoc network. For instance, in a military scenario, UAVs that form a tactical ad hoc network could provide a backbone for ground communications when obstacles exist and provide effective direct communications between two ground devices. UAVs could also be used to actively sense environmental data for civilian applications [21]. J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 457–470, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

458

A. Ismail et al.

Fig. 1. An example of a radio control application

Real-time radio control requires that the system responds to user inputs with as little latency as possible. The time it takes the system to respond to a user input is here referred to as the system latency. System latency is mainly dependant on the interarrival time of control information from the transmitter to the receiver. Imperfections in the wireless channel (i.e. noise and interference) result in randomly fluctuating system latency. Therefore, the robustness of the wireless link established between the transmitter and the receiver has a major effect on the system latency. A lag occurs when the system latency exceeds a specified real-time limit. The user experience can be quantified by the probability of lag occurrence during a radio control session. This paper derives a performance metric for real-time radio control applications based on the probability of lag occurrence. Specifically, this paper studies the effect of using Frequency-Hopping Spread Spectrum (FHSS) technology on the user-experience. To the best of our knowledge, no previous work has been made to characterize packet-based FHSS RC applications with a performance metric such as the lag occurrence probability. The rest of the paper consists of two main parts: First, we derive the lag-occurrence probability of a packet-based frequency hopping radio control system. Second, we present the simulation results.

2 System Lag Occurrence Probability In control applications where the stimulus is an input from a user, the specified realtime limit becomes the Human Response Time (HRT) [3]. In such systems, a lag occurs when the system latency exceeds HRT. The HRT is estimated to be around 100ms in real-time applications such as Voice-over-IP (VoIP). We define the System Lag Occurrence Probability (SLOP), as the probability of a lag occurring in the output of a real-time radio control system. This section derives SLOP for a frequency hopping radio control system. Frequency hopping consists of two modes of operation: acquisition and tracking. To begin the derivation of SLOP, we’ll first look at the interaction between acquisition and tracking modes. Initially the transmitter-receiver pair is out-of-lock and therefore acquisition is first initiated. The process responsible for performing acquisition is here referred to as the acquisition engine. After some time Tacq the transmitter-receiver pair acquire lock on each other, acquisition is terminated, and tracking is initiated. The process responsible for performing tracking is here referred to as the tracking engine.

Performance of Packet-Based Frequency-Hopping Spread Spectrum RC Systems

459

During tracking, the transmitter-receiver pair will try to remain in-lock with each other for as long as possible. After some time Tlock, various factors such as interference and time/frequency drift cause the transmitter-receiver pair to loose lock of each other and re-initiate acquisition. The time it takes the tracking engine to conclude that it has lost lock is called TLA (i.e. Lock Æ Acquisition). In such as system, SLOP becomes the probability of two events happening: 1) the event that the system enters acquisition mode and 2) the event that the system fails to acquire in the time remaining for an HRT to elapse: SLOP = P Loose lock T Fail to acquire in tim e b

c

Assuming that the tracking and acquisition processes are statistically independent, SLOP becomes: S L O P = P LA P

d

b

T > H RT @ T

LA

ce

where PLA is the probability of loosing lock, TLA is the time it takes the tracking engine to conclude that it has lost lock, and T is the time it takes the acquisition engine to reacquire lock. PLA and TLA depend on the algorithm used for tracking, while the probability of successful acquisition depends on the algorithm used for acquisition. Therefore, to further derive SLOP, we must first analyze the acquisition and tracking engines in finer detail. The communication system dealt with in this study is packet-based. As will be shown later, the main decision criterion used in the tracking and acquisition engines is based on the state of packet reception (correct or corrupt). The derivation of SLOP in the next two sub-sections therefore places emphasis on the average Packet-Error Rate (PER) parameter. The use of PER abstracts away the details of the underlying wireless channel and keeps the focus on the application layer design instead, as intended from the beginning. The performance of frequency hopping systems under the presence of different types of interference is studied extensively in the literature. For example, [4] and [5] look at wideband and partial-band noise jamming. [6] looks at follower partial-band jamming, while [7] looks at partial-band multi-tone jamming. In this paper, however, the wireless channel model, the modulation type used, and the type of interferers present in the band are all abstracted in the average PER parameter. The average PER of a frequency hopping system is here calculated as: q

1f f f X PERx PER = f q x=1

where PERx is the PER of channel x and q is the total number of channels. This has the effect of averaging out any interference present in the band across all the channels. 2.1 Tracking Strategies Various tracking strategies have been studied in literature, the most common being Tau-Dither [8], Delay-Lock [9], and Split-Bit Tracking loops. For packet-based

460

A. Ismail et al.

systems, Modified Transmitted Reference [10] techniques seem to be more adequate as they are simpler to implement in firmware. A Modified Transmitted-Reference tracking engine is more commonly referred to as an N-state lock detector [2]. This section derives PLA and TLA for an N-state lock detector. The basic principle of operation is shown in Figure 2. The N-state lock detector contains N states, named Lock_0 to Lock_(N-1). Operation in one of those N states represents operation in the tracking mode. State Lock_0 represents correct packet reception during tracking mode, while states from Lock_1 to Lock_(N-1) represent corrupted packet reception during tracking mode. An additional state, called Acq, represents operation in the acquisition mode and is shown to the right of Figure 2.

Fig. 2. State transitions in an N-State lock detector. N = 3.

The N-state lock detector begins operation in state Lock_0, where the transmitterreceiver pair is in lock and is exchanging packets correctly. During each hop, the transmitter-receiver pair exchanges a new packet. If the packet is received correctly, the lock detector generates a positive decision and moves to (or remains in) state Lock_0. If the packet is received corrupted, the lock detector generates a negative decision and moves one state to the right, towards state Acq. This process is repeated at each of the states of the lock detector. If N consecutive negative decisions are generated, the lock detector terminates tracking and initiates acquisition. Positive decisions occur with probability p. Negative decisions occur with probability n. To simplify the analysis of PLA and TLA, we assume that tracking is maintained when one complete packet is exchanged correctly. This has the following effects:

n = PER

p = 1 @ PER

and

As mentioned previously, PLA is the probability of the tracking engine concluding that it has lost lock. In other words, PLA is the probability of terminating tracking and initiating acquisition. With an N-state lock detector, PLA becomes the probability of exchanging N consecutive corrupted packets. Assuming that errors between packets are statistically independent, PLA becomes: P LA = PE R

N

Performance of Packet-Based Frequency-Hopping Spread Spectrum RC Systems

461

TLA is the time it takes the tracking engine to conclude that it has lost lock. With an N-state lock detector, TLA becomes the time it takes for the exchange of N consecutive corrupted packets: T LA = N T dwell

where Tdwell is the channel dwell time (hopping period). The above derives PLA and TLA for an N-state lock detector. To complete the derivation of SLOP, we will derive the probability of successful acquisition. The probability of successful acquisition depends on the algorithm used in the acquisition engine. 2.2 Acquisition Strategies Acquisition strategies have been studied extensively in literature, the most common being serial [11] and parallel [12]. Other acquisition techniques include message passing [15] and adaptive-antenna array [16]. Serial acquisition techniques try to acquire lock between the transmitter and the receiver by searching the channels serially (i.e. one after the other) until the correct channel is found [11]. Uniform serial acquisition treats all channels as equally possible candidates for successful acquisition. Non-uniform serial acquisition favors some channels over others during its search. Parallel acquisition techniques acquire lock by examining all channels simultaneously [12]. In practice, this is achieved by employing a bank of correlators, each tuned to one of the channels. Serial acquisition is simple to implement but is relatively slow in acquiring lock. Parallel acquisition is complex to implement but is fast in acquiring lock. In practice, a hybrid of both techniques is usually employed to reach a relatively simple solution with relatively fast acquisition times. To simplify the analysis, this section will derive the required parameters using a uniform serial acquisition engine. The mean acquisition time and its standard deviation for a uniform serial acquisition engine were derived in [13] and [14]: T acq = T dwell

σ =T 2 acq

H b

2 L dwellJ

c2

1+KPFA

h

i

2f + 2f @ P qf @ 1f 1f + KP D FA f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f b

cb

2P D

cb

c

H

II

c b cb c 1@P ` a b 1f 1f 1 f 2 2 f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f f fM D k K q @f +f +6q K K+1 PFA PD @PD + 1+KPFA 4@2PD @PD +J f K 2 12 PD P2D PD f f f f f f 2j f

D

E

where Tacq is the mean acquisition time, σacq is its standard deviation, Tdwell is the channel dwell time (hopping period), and q is the number of channels. Minimum Tdwell is equal to the time it takes the transmitter to hop to a new channel plus the time it takes to transmit a packet. The probability of detection (PD) is the probability of correctly terminating acquisition when the correct channel is being probed. The probability of a false alarm (PFA) is the probability of erroneously terminating acquisition when an incorrect

462

A. Ismail et al.

channel is being probed. Such an event is called a false alarm, and has a time-penalty K associated with it. K is the number of hops it takes the tracking engine to conclude that lock was falsely acquired and therefore re-initiate acquisition. Low Signal-toNoise Ratio (SNR) and high channel distortion (fading and Inter-Symbol Interference ISI) drive PD to 0 and PFA to 1. To simplify the analysis, we assume that acquisition ends when one complete packet is received correctly, which is a fair assumption in a packet-based communication system. This has the following effects: P D = 1 @ PER

P FA = 0

and

PFA becomes negligible since it will be unlikely that an entire correct packet is falsely detected. Furthermore, assuming that q >> 1 and K << q, which are fair assumptions in practice, Tacq and σacq simplify to the following: T acq = qT dwell

H

I 2f @ P f f f f f f f f f f f f f f f f f f f f D Jf K

2P D

w w w w w w w w w w w w w w w w w w w w w w w w w w w w v w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w w I uH u 1 1 1 u f f f f f f f f f f f f f f f f f f f f f f K σ acq = qT dwell tJ f @ f + f 2

PD

12

PD

It was shown that if we assume that the tracking and acquisition processes are statistically independent, SLOP becomes: d

b

SLOP = P LA P T > HRT @ T LA

ce

where PLA is the probability of loosing lock, TLA is the time it takes the tracking engine to conclude that it has lost lock, and T is the time it takes the acquisition engine to reacquire lock. Attempts to derive the Cumulative Distribution Function (cdf) and the Probability Density Function (pdf) of T were made in [17] and [18]. The model of the acquisition engine studied in the literature is complex however. False alarms, as well as other variables in the system tend to complicate the derivation of a proper pdf or cdf. In this paper, we consider packet-based system operation, in which all decisions made in the acquisition engine are based on the state of packet reception during every hop interval: either correct packet reception or corrupt packet reception. In such a model, the probability of a false alarm becomes negligible. Essentially, every hop interval becomes a new Bernoulli trial. If a correct packet is received, the trial is declared a success and the acquisition process is terminated. Otherwise the trial is declared a failure and the acquisition process continues. This reduces the acquisition time random variable to be defined simply as the number of consecutive failed Bernoulli trials made until a successful trial is made. By definition, T therefore becomes a geometric random variable [19]. The limiting form of T as the number of trials goes to infinity is the exponential random variable.

Performance of Packet-Based Frequency-Hopping Spread Spectrum RC Systems

463

Therefore, if we model the acquisition time T as an exponential random variable, SLOP becomes: d

SLOP = P LA 1 @ F T HRT @ T LA b

ce

where FX(x) is the cumulative distribution function of the exponential random variable 1f f f f f f f f f f f f f f f f f @ A . FX(x) is defined as [19]: X with parameter λ = f E X

F X x = 1 @ e@ λx

x≥ 0

` a

Where in this case, λ=

1f f f f f f f f f f f f T acq

x = HRT @ T LA

and

Therefore, SLOP = PER

N

d

b

1 @ F T HRT @ NT dwell

ce

where H

I

2f @ Pf 1f f f f f f f f f f f f f f f f f f f f f f f f D K T acq = = qT dwellJ 2P D λ

3 Simulation and Results This section presents simulation and results. The purpose is to verify the derivations described in Section 2 and study the effect of various design variables on the performance of a frequency hopping real-time radio control system. 3.1 Packet-Based Frequency Hopping Simulator

The simulator is built using Java. It consists of a transmitter, a receiver, and a range of 20 to 80 frequency channels. Each channel is characterized with a Bit Error Rate (BER). Using its BER, each channel computes the corresponding Packet Error Rate (PER) according to the following formula [20]: PER = 1 @ 1 @ BER `

ab

where b is the length of a packet in bits. Each channel contains a random number generator which corrupts packets at a rate equal to its computed PER. There are two types of channels in the simulator: Blocked and nonblocked. Blocked channels are ones with PER equal to 100%. Blocked channels are used to simulate the effect of powerful co-channel (same-channel) interferers. Non-blocked channels are ones with PER equal to 15%. A PER of 15%

464

A. Ismail et al.

corresponds to a BER of 1e-3 and a packet length of 160 bits. These are typical values used in packet-based wireless communication systems with Additive White Gaussian Noise (AWGN) channels. The transmitter hops pseudo-randomly across all the channels and generates a single packet for each hop. The hopping pattern is uniform. Each generated packet is passed to the channel at which the transmitter currently resides. The channel in turn either corrupts the packet or passes it to the receiver uncorrupted, depending on the PER of the channel. The receiver implements uniform serial acquisition and N-State lock detection. The Finite State Machine (FSM) of the receiver is shown in Figure 3. The acquisition engine consists of a single state called ACQ. The lock detector consists of two states called LOCK_0 and LOCK_N. The receiver FSM defines the interaction between these three states. The simulator runs by generating a programmed number of timer ticks. At each tick, the transmitter hops to a new channel and transmits a packet at that channel. The channel corrupts the packet or passes it to the receiver un-corrupted depending on the PER of the channel. Two new inputs are therefore available to the receiver with each new tick: 1. The newly transmitted packet (either corrupted or un-corrupted), and 2. The channel at which the new transmission occurred.

Fig. 3. Receiver FSM

Performance of Packet-Based Frequency-Hopping Spread Spectrum RC Systems

465

The receiver generates one of three events depending on the two inputs from the simulator: 1

2

3

Non-Corrupted-Packet-Received Event is the event that the receiver’s current channel is equal to the transmitter’s current channel and the received packet was not corrupted by the channel. Corrupted-Packet-Received Event is the event that the receiver’s current channel is equal to the transmitter’s current channel but the received packet was corrupted by the channel. Missed-Packet Event is the event that the receiver’s current channel is not equal to the transmitter’s current channel and therefore the transmitted packet was entirely missed, regardless of the state of the packet itself (corrupted or uncorrupted). This event can only occur during the acquisition state (ACQ) since this is the only state during which the transmitter and the receiver can be out-of-lock.

The inter-arrival time of correctly received packets is measured in ticks (time normalized to the channel dwell time) and is logged with every new tick. NonCorrupted-Packet-Received events reset the inter-arrival time, while Corrupted Packet Received events and Missed Packet events increment the inter-arrival time. The inter-arrival time of each packet during a control session is plotted and analyzed against different design variables for performance evaluation. A lag is registered every time the inter-arrival time exceeds HRT. The reception of a correct packet represents an opportunity for a lag to occur. SLOP is computed at the end of a simulation session as the ratio of the number of lags to the number of correct packets received. SLOP is plotted and analyzed against different design variables for performance evaluation, which will be presented in the next sub-section. 3.2 Simulation Results

Two sets of simulations were carried out. The first simulation plots SLOP against the probability of detection (PD). The purpose of the first set is to verify the theoretical results derived in Section 2. The second set plots the average inter-arrival time of correctly received packets against different design variables. The purpose of the second set is to study the effect of different design variables on the performance of a packet-based frequency hopping system. Results for the first simulation set are shown in Figures 4 and 5. At low probability of detection, the system latency increases and therefore the probability of lag occurrence increases, and vice versa. Results are better for larger N (the number of states in an N-state lock detector) since a larger N improves tracking and therefore decreases the overall system latency. Results show that the simulation is in perfect agreement with theory. This validates the derivations made in the previous section and serves as evidence for correctly modeling the acquisition time as an exponential random variable.

466

A. Ismail et al.

Figure 5 demonstrates the simulation plots SLOP against N. As N increases, tracking is improved and therefore overall system latency decreases. This in-turn decreases SLOP. Naturally, performance is better at lower higher probability of detection than it is at lower probability of detection. As can be seen in the figure, results show that the simulation is in perfect agreement with theory. This validates the derivations made in the previous section and serves as evidence for correctly modeling the acquisition time as an exponential random variable.

Fig. 4. SLOP vs. PD (q = 40, HRT = 100ms)

Fig. 5. SLOP vs. N (q = 40, HRT = 100ms)

Performance of Packet-Based Frequency-Hopping Spread Spectrum RC Systems

467

The next set of simulations present the effect of different design variables on the performance of a frequency hopping radio control system. The design variables investigated include: • • •

q: The number of channels used for hopping J: The number of blockers (co-channel interferers) present in the band N: The number of states in an N-state lock detector.

The performance metric considered here is the average inter-arrival time of correctly received packets, which represents the average system latency (response). The first simulation studies the effect of blocking. Figure 6 shows the results. The y-axis represents the normalized inter-arrival time, while the x-axis represents the partial band blocking ratio. This is the ratio of the number of blocked channels to the total number of channels. Simulation is carried out with a uniform serial-search acquisition engine and a 10-State lock detector. Results in Figure 6 depict that performance begins to dramatically degrade when 40% or more of the band is being blocked, regardless of the number of channels being used in hopping. This is expected since a high partial blocking ratio causes the system to frequently loose lock and enter acquisition mode. Acquisition is worse for higher number of channels, therefore performance degradation is worse for 80 channels than it is for 40 and 20.

Fig. 6. Effect of blocking. N = 10.

The second simulation investigates the effect of N-State lock detection on performance. Figure 7 shows the results. The y-axis represents the normalized interarrival time, while the x-axis represents the number of states in an N-State lock detector. The simulation is carried out at a high partial band blocking ratio (J/q = 75%). Uniform serial search acquisition is used.

468

A. Ismail et al.

Fig. 7. Effect of Tracking. J/q = 0.75.

Results in Figure 7 show the importance of maximizing the number of states in the N-state lock detector. As N increases, performance improves, to the point where performance of systems operating over different number of channels converges at approximately N > 40. This is expected since higher N leads to a lower number of acquisition attempts and therefore smaller average inter-arrival time. The third simulation studies the effect of varying the number of hopping channels on interference. Figure 8 illustrates the results. Interference resistance in this paper is calculated as the ratio of the number of blocked channels (J) to the total number of channels (q). This is used to indicate how much interference resistance there is in a

Fig. 8. Interference resistance due to increasing number of channels, N = 10

Performance of Packet-Based Frequency-Hopping Spread Spectrum RC Systems

469

frequency hopping system. Lower ratio means more free channels and therefore better communication. The y-axis represents the normalized inter-arrival time, while the xaxis represents the number of channels used in hopping. The simulation is carried out with uniform serial search acquisition engine and a 10-state lock detector. Results show that increasing the number of channels can be used to effectively combat interference. For a given number of blocked channels, performance improves as the number of channels increase. This happens because the average PER of the system decreases as q increases. Another way to explain the results in the Figure 8 is in terms of acquisition and tracking. As the number of free channels in the system increases, the chances of finding a free channel for proper communication increases and therefore the system remains in tracking mode longer. The number of acquisition attempts therefore decrease and the overall performance of the system improves.

4 Conclusions and Future Research Real-time radio control systems have wide applications, including mobile ad hoc networks. This paper used Frequency Hopping Spread Spectrum technology to mitigate the interference effects for real-time radio control systems. System Lag Occurrence Probability (SLOP) was used as the performance metric to characterize user experience. Simulations have been conducted and the results validated the derivation of SLOP. In general, a frequency hopping system switches periodically between two modes: acquisition and tracking. Many acquisition and tracking engines are found in the literature. Uniform serial acquisition and N-state lock detection were adopted, because they simplify analysis and satisfy the conditions required for the derivation of SLOP. In practice, other acquisition and tracking engines such as those employed in Adaptive Frequency Hopping systems can be used to improve performance. SLOP can be derived based on the new proposed schemes in a similar manner to the derivation shown in Section 2. The use of more complex acquisition and tracking engines and the analysis of their effects on the performance of radio control systems is left for future work.

Acknowledgements The authors thank MITACS, Canada and Cistel Technology for their support of the research.

References [1] Abatti, J.M.: Small Power: The Role of Micro and Small UAVs in the Future. Technical Report AU/ACSC/6697/2005-04, Centre for Strategy and Technology, Air War College, Air University (2005) [2] Glisic, S., Vucetic, B.: Spread Spectrum CDMA Systems for Wireless Communications. Artech House Inc., Boston (1997)

470

A. Ismail et al.

[3] Mowbry, G.H., Gebbard, J.W.: Man’s Senses as Informational Channels. In: Sinaiko, H.W. (ed.) Human Factors in the Design and Use of Control System, pp. 115–149. Dover, New York (1961) [4] McGuffin, B.F.: Jammed FH-FSK Performance in Rayleigh and Nakagami-M Fading. In: Proc. of IEEE Millitary Communications Conf., pp. 1077–1082 (2003) [5] Teh, K.C., Kot, A.C., Li, K.H.: Partial Band Jammer Suppression in FFH SpreadSpectrum System Using FFT. IEEE Trans. on Vehicular Technology 48(2), 478–486 (1999) [6] Hassan, A.A., Start, W.E., Hershey, J.E.: Frequency-Hopped Spread Spectrum in the Presence of a Follower Partial-Band Jammer. IEEE Trans. on Communications 41(7), 1125–1131 (1993) [7] Miller, L.E., Lee, J.S., French, R.H., Torrieri, D.J.: Analysis of an Antijam FH Acquisition Scheme. IEEE Trans. on Communications 40(1), 160–170 (1992) [8] Peterson, R.L., Ziemer, R.E., Borth, D.E.: Introduction to Spread Spectrum Communications. Prentice Hall, New Jersey (1995) [9] Wilde, A.: Extended Tracking Range Delay Lock Loop. German Aerospace Research Establishment, Germany. In: Proc. of IEEE ICC, pp. 1051–1054 (1995) [10] Dominique, F., Reed, J.H.: Robust Frequency Hop Synchronization Algorithm. Electronics Letters Online 32(16), 1450–1451 (1996) [11] Polydoros, A., Weber, C.: A Unified Approach to Serial Search Spread Spectrum Code Acquisition – Part I: General Theory. IEEE Trans. on Communications 32(5), 550–560 (1984) [12] Chawla, K., Sarwate, D.: Parallel Acquisition of PN Sequences in DS/SS Systems. IEEE Transactions on Communications 42(5), 2155–2164 (1994) [13] Holmes, J.K., Chen, C.C.: Acquisition Time Performance of PN Spread-Spectrum Systems. IEEE Trans. on Communications COM-25(8), 778–783 (1977) [14] Dicarlo, D.M., Webber, C.L.: Multiple Dwell Serial Search: Performance and Application to Direct Sequence Code Acquisition. IEEE Trans. on Communications 31(5) (1983) [15] Zhu, M., Chugg, K.: Iterative Message passing Techniques for Rapid Code Acquisition. In: Proc. of IEEE Millitary Communictions Conf., pp. 434–439 (2003) [16] Dlugos, D., Scholtz, R.: Acquisition of Spread Spectrum Signals by an Adaptive Array. IEEE Trans. on Acoustics, Speech, and Signal Processing 137(8), 1253–1270 (1989) [17] Dicarlo, D.M., Weber, C.L.: Statistical Performance of Single Dwell Serial Synchronization Systems. IEEE Trans. on Communications Com-28(8), 1382–1388 (1980) [18] Jovanovic, V.M.: On the Distribution Function of the Spread-Spectrum Code Acquisition Time. IEEE Journal on Selected Areas in Communications 10(4), 760–769 (1992) [19] Garcia, A.L.: Probability and Random Processes for Electrical Engineering, 2nd edn. Addison-Wesley Publishing Company, Reading (May 1994) [20] Han, J., Lanzinger, D., Sklar, D.: Assessing the Performance of Packet Retransmission Schemes Over Satellite Link. In: Proc. of IEEE Aerospace Conf. (2006) [21] Frew, E.W., Dixon, C., Elston, J., Stachura, M.: Active Sensing by Unmanned Aircraft Systems in Realistic Communication Environments Networked Robotics. In: Proc. of IFAC Workshop on Networked Robotics (2009)

Performance Analysis of UWB Body Sensor Networks for Medical Applications Abdellah Chehri and Hussein Mouftah School of Information Technology and Engineering (SITE) University of Ottawa, 800 King Edward Avenue, Ottawa, Ontario, Canada, K1N 6N5 {achehri,mouftah}@uottawa.ca

Abstract. Wireless sensor networks can be employed in medical healthcare in many tasks such as, monitoring vital signs, controlling medical equipment, patient positioning, and in addition for non-medical service such as entertainment, psychophysiological detection of deception1 . UltraWideband (UWB) radio is a revolutionary, power-limited, and rapidly evolving technology, which employs short pulses with ultra low power for communication and ranging. Compared to narrow systems, UWB systems have several advantages, such as fading robustness, low power consumption and low cost transceiver implementation. In this paper we evaluate an UWB-based body sensor networks communication with respect to signal propagation around a human body. The performance such as, node location, AWGN (Additive white Gaussian noise), ISI (Inter-Symbol Interference) effects on the BER (bit error rate) were evaluated. Keywords: ultra-wideband, body sensor networks, wban, IEEE 802.15.6.

1

Introduction

As the population ages and the risk of chronic disease increases, the cost of healthcare will rise. The employment of new technologies for medical healthcare could reduce the cost and improve the efficiency of treatment. By the summer of 2005, the initiative of marrying information technology to medicine seemed clear, and the media was heralding what some called “the e-health” revolution. Wireless technology capabilities are growing at a fantastic rate. There appears to be no limit to what technology might accomplish, given infinite resources [1]. In order to improve the efficiency, a strong demands for introducing wireless technology in medical healthcare. This proposal comes from various parties such as medical societies as well as communications technology (ICT) industries. This caused the emergence of body sensor networks (BSN), where a set of communicating devices are located around the human body. BSN is a collection of low powered biosensor devices known as “motes” (or “nodes”). In principle, each node is an integration of embedded microprocessors, 1

Popularly referred to as a lie detector.

J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 471–481, 2010. Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

472

A. Chehri and H. Mouftah

a radio transceiver with limited amount of data storage. The recent development of high performance microprocessor and novel sensing materials has stimulated great interest in the development of smart sensors physical, chemical or biological sensors combined with integrated circuits [2], [3]. These sensors could be located on the body as tiny intelligent patches, integrated into clothing, or implanted below the skin or muscles. The introduction of wireless connections to exchanges sensor’s data could provide a great flexibility for both, patient and medical staff. This propriety will contribute to allow more mobility for the patient and more facility for the doctors during his intervention (i.e., surgical operations). So, this will be more comfortable to the patients as well as medical personnel in comparison to conventional wired sensors. The collected data could also be stored for further analysis. It could be utilized directly in a more effective way by the medical personnel. By using a computer or personal digital assistant (PDA), although outside hospital, the medical staff could be capable to monitor the patient regardless of his position as long as the he was connected to the network. To harmonize with the strong demands from both medical healthcare societies and ICT industries, a standardization committee referred to as IEEE 802.15.6 was formally set up in December 2007 [4]. The objective of IEEE 802.15.6 is to define new physical (PHY) and media access control (MAC) layers for WBAN (and even BSN). This could be used to develop a low cost, ultra low power and highly reliable wireless network. These functionalities are controlled primarily by PHY and MAC layers in conjunction with the application layers. Ultra wideband (UWB) technology has emerged as a solution for the wireless interface between medical sensors in future healthcare systems [5]. Therefore, UWB has been proposed to be used into WBAN in the IEEE 802.15.6 task group [4]. In this paper, we focus on the UWB signalization at BSN’s physical layer. We’ll assume that once the vital signs data are collected, they will be modulated and transmitted through multipath channel. At the receiver node a non-coherent detection with low complexity scheme has been used. By the link budget calculations, we show that although the performance is not pleasing, the transmission requirement of the low-rate BSN can be satisfied which means that the proposed scheme can be used to realize low-rate BSN with very low complexity. The rest of the paper is organized as follows. In Section II, we give an overview the wireless healthcare system. The system architecture is formally described in Section III. The system description is presented in Section IV. Simulation results are included in Section V. Finally, we conclude the paper in Section VI.

2

Related Work

The advances in wireless sensor networking have opened up new opportunities in healthcare systems [6]. The future will see the integration of the abundance of

Perf. Analysis of UWB-BSN for Uses in Medical Applications

473

existing specialized medical technology with pervasive and ubiquitous wireless networks. They will coexist with the installed infrastructure, augmenting data collections with real-time response. Thus far, increasing number of research groups around the world has been created. For example, the performance of an IEEE802.15.4/Zigbee MAC based WBAN operating in different patient monitoring environment has been analyzed [7]. The authors in [8] have been presented an UWB transmitter implementation for WBAN application. The pulse generator has been implemented in a digital 0.18 μm CMOS technology, showing the potential of UWB in the realization of a low-cost radio interface for the WBAN sensor nodes. While the authors in [9] summarized the design issues for the physical layer proposals with the category of narrowband and UWB signals for IEEE 802.15.6. In [10] Keong et al. have been presented a single channel real-time wireless electrocardiograph monitoring system, which has been implemented using low data rate UWB impulse radio method, aligned with the direction of IEEE802.15.6. In [11], the authors proposed a system design and realization of a wireless EEG (electro-encephalograms) and ECG (electrocardiograms) sensor network focusing on issues such as time synchronization, bandwidth, and power constraints constituent of WBANs. The authors have been used and evaluated via simulation a WSN composed of three transmitting nodes. However, only few works have been seriously done around the physical layer limitation for UWB-based medical sensor networks. For example, the authors in [12] presents architecture of a healthcare wireless network that exploits the capabilities of ultra wideband technology for medical sensing and in-body tracking and imaging. However, the authors don’t investigate thoroughly on sensor node transmission capability. In our knowledge, only one study has been done on the implementation of a UWB transmission system around the body. A simple performance evaluation of coherent RAKE receiver in a BAN has been done in [13]. However, the authors don’t take into account the real transmission system, which including body propagation channel and ISI effects. In fact, the physical layer performance and how should UWB be exploited to provide a good transmission is crucial task for the design of global transmission system. Hence, it will be useful to see the physical layer performance. Based on these results, we can exploit the UWB characteristics to design more optimum transmission system.

3

Proposed Wireless Biomedical Sensor Networks Architecture

The main goal of this paper is to investigate on UWB transmission. We investigate on BSN architecture for smart healthcare that possesses the following proprieties:

474

A. Chehri and H. Mouftah

– Real-time and long-term remote monitoring; – Tiny sensor with very low complexity; – Can be integrated with existing medical practices and technology; For the first query, the large bandwidth of UWB signal is the low electromagnetic radiation. The Federal Communications Commission (FCC) has authorized UWB communications between 3.1 GHz and 10.6 GHz. Although the regulations on UWB radiation define a power spectral density (PSD) limit of -41dBm/MHz, the low radiation has little influence on the environment and safe for human body, even in the short distance [14]. Among the most important advantages of UWB technology are the low system complexity and the low cost. UWB systems can be made nearly “all-digital”, with minimal radio frequency (RF) or micro wave electronics. The low component count leads to reduced cost, and smaller chip sizes invariably lead to low-cost systems. The simplest UWB-based node could be assumed to be a pulse generator, a timing circuit, and an antenna [15]. Other important advantage of UWB resides in the possibility to significantly reduce the power consumption of the radio frontend by switching off the transmitter during the relatively long silence periods between UWB pulses. Therefore, low UWB transmission allows to the node a longer battery life. It has been shown that the effect of multiuser interference (MUI) on system performance is generally less detrimental in UWB networks than in narrow band networks [16]. Also, the UWB transmission offers good penetrating properties that could be applied to medical applications. In addition, by using UWB signalization, the nodes could be able to allow transmission even under very bad channel conditions. One patient is equipped with several sensors monitoring different parameters. A Body Sensor Network is made up of one or more body area networks and a base station. When the information has been gathered in the sensor network it is forwarded to this base station. The information is then received at a relay station and passed on through a backbone network. In the end, the information can be viewed at terminals or monitoring stations that are connected to the network. This system has the potential of making remote monitoring and immediate diagnostics a reality [2], [17], [18]. Sensors are heterogeneous, and all integrate into the human body. The number and the type of biosensors vary from one patient to another depending on the state of the patient. The most common types of biosensors are EEG “Electroencephalography” to measure the electrical activity produced by the brain, ECG “Electrocardiogram” to record the electrical activity of the heart over time, EMG “Electromyography” to evaluate physiologic properties of muscles, Blood pressure, heart rate, glucose monitor, SpO2 “Oxymeter” to measure of oxygen saturation in blood, and to measure temperature of the body [19], [20]. As shown in the Table 1, according to the characteristics of physiological measurements or type of application services which can be real-time or non real-time with high or low rate.

Perf. Analysis of UWB-BSN for Uses in Medical Applications

475

Table 1. Service Classification of Physiological Measurements Type of Service ECG EEG, EOG, EMG Blood pressure, body temperature, heart rate, glucose monitor Medical image, X-ray, MRI

4

Data rate High Low Low

Latency Low Low High

Class of Service Real-time high rate Real-time low rate Non real-time low rate

High

High

Non real-time high rate

System Description

In order to evaluate the characteristics of UWB communication for use in a BSN, we have planned scenario where a patient has been equipped with four sensor nodes. These nodes are mounted around his body (Fig. 1). Each node is implemented with different medical sensor ECG, SpO2, body temperature, and blood pressure. The task and the position of each sensor are summarized in Table 2. Table 2. Sensor information and location

Node Node Node Node

Sensor Distance (m) Position 1 ECG 0.25 Front 2 Blood Pressure 0.35 Side 3 SpO2 0.42 Side 4 Temperature 0.50 Back

The collected data will be transmitted to a base station which has been mounted on the front of the patient. This base station can either store, or transfer the data to the remote hospital server and the related services. This could be accomplished by using a mobile phone, a PDA or Internet. The global schema of patient’s monitoring is presented shown in Figure 1. 4.1

Body Channel Model

A basic step required for a communication system simulation is to get precise models of all the elements involved in the system. This includes, of course, the radio channel, as the physical mean of transport for the wireless signal. Several WBAN channel model has been measured and analyzed, ones of them has been described in [21]. The work done by Fort and al. [21] takes into account both, the propagation of the signal around the body, and the reflections at the nearest scatters in the room. Because this model has been recommended as an improvement of previous BAN models, and has been recognized by the committee for the emerging 802.15.6 standard [22]. So, this channel has been used to evaluate the performance of UWB systems around body. The channel impulse response has been represented by h(t) = X

L  K  l=0 k=0

αk,l δ (t − Tl − τk,l )

(1)

476

A. Chehri and H. Mouftah

Fig. 1. The global schema of patient’s monitoring. The illustration of the nodes and the central node (receiver) locations on the body.

– αk,l are the multipath gain coefficients; – Tl is the delay of the lth cluster; – τk,l is the delay of the k th multipath component relative to the lth cluster arrival time Tl ; – X represents the log-normal shadowing. In calculating the path loss, the distance between the transmitter and receiver has been measured along the surface of the body rather than along a strict straight line as the waves travel along the body rather than passing through it. The P L in dB as function of the distance is given by: P L(d) = P L0 + 10.n log10

d d0

(2)

where – P L(d) represents the received power at a distance d, computed relative to a reference distance d0 . – P L0 is the interception point and is usually calculated based on the midband frequency. – n is path loss exponent. The typical value of n varies between 5 and 6. This means that the pulse propagation around the body is very demanding due to the high attenuation of the signal. Due to space limit, the channel parameters have not presented in this paper, however they can be found in a [21]. 4.2

Transmitter and Receiver Architectures

Channels encountered by UWB communication systems are highly dispersive in nature and so the channel estimation is a very challenging task. Designing a

Perf. Analysis of UWB-BSN for Uses in Medical Applications

477

•ሺ–ሻ

—Ž•‡‰‡‡”ƒ–‘” ሺ‡˜‡”›ܶ௙ •ሻ 

‡Žƒ›ܶௗ 

 

‫݌‬ሺ‫ ݐ‬െ ܶௗ ሻ ‘†—Žƒ–‹‘ ‹ˆ‘”ƒ–‹‘ ሺ൅ͳǡǦͳሻ

        

்೑ା்೘

න ்೑

௞ୀ‫ہ‬௞Ȁேೞ ‫ۂ‬ேೞ ାேೞ ିଵ

݀‫ݐ‬

෍ ௞ୀ‫ہ‬௞Ȁேೞ ‫ۂ‬ேೞ

‡…‹•‹‘ ‡˜‡”› ܰ௦ ˆ”ƒ‡•

‡Žƒ›ܶௗ

Fig. 2. Illustration of the non-coherent TR system, transmitter (a), receiver (b)

receiver that generates reference locally at the receiver, estimates the channel, and captures enough energy for data detection is a difficult and costly process. But instead of locally generating the reference signal, it can be transmitted along with the information data. Such a system is known as Transmitted Reference (TR) system. TR is a correlation receiver system; thus a TR system does not require channel estimation and has weak dependence on distortion. Compared to coherent receiver (i.e. RAKE), the TR receiver is very attractive. The block diagram of a TR transmitter and receiver are presented in Figure 2. As shown in the Figure 2 (a); the transmitter of a TR system comprises of a pulse generator, a delay line, and an antenna unit. Figure 2 (b) shows the simple receiver structure. The receiver comprises of a delay line and a correlator to demodulate the signal, and an addition unit to add over Ns pluses so that enough energy is captured to estimate the information bit. Assuming a single-user UWB system with antipodal modulation (binary pulse amplitude modulation), a typical transmitted reference frame is given by For TR the transmitted signal is modeled as:  str (t) = p(t − kTf ) + b k/Ns  p(t − kTf − Td ) (3) k

where – k is the frame index; – Tf is the frame time; – Td is the delay between reference and modulated pulse;

478

A. Chehri and H. Mouftah

– Ns is the number of successive times the frame is repeated to achieve adequate bit energy required for detection; – bl are the channel symbols with values ±1, generated randomly having equal probability of occurrence. For a single user case when operating in a multi path channel with multi path delay spread time Tm ,to avoid inter symbol interference (ISI) problem a TR frame is designed such that Tf ≥ 2Td ≥ 2Tm .

5

Results and Discussion

To evaluate the performance of the UWB-TR receiver for use in medical application, a bit error rate (BER) calculation has been calculated. Since the transmitted data bit was known in advance, using the UWB receiver, the information bit was detected. The detected and the transmitted bits were compared. 5.1

Performance of the TR Receiver Node in the Absence of ISI

Figure 3 shows the variation BER by increasing the energy bit per noise (Eb /N0 ). The black, green, red, and blue curves correspond to the BER of the Node 1 to 4, respectively. When only one node was transmitting, at the receive node, the signal was effected by multi path channel and corrupted by the AWGN. The performance of the four nodes when ISI is avoided are shown in Figure 3. As was expected, the performance for the Node 4 is about 6 dB worse than the Node 1. The corresponding Eb /N0 requirement is about 24 dB for the Node 4 if the target BER performance is set as 10−2. This is due mainly to the NLOS configuration of the Node 4. 5.2

Performance of the TR Receiver Node in the Presence of ISI

Figure 4 shows the performance of a the receiver for all nodes when Tf ≤ 2(Td + Tp ) so that ISI occurs. In this case each pulse will overlap and interfere with other pulses. As is known, a main drawback of a TR system is the noisy template used for detection. In the presence of ISI, the template that was already noisy suffers from the overlapping of the earlier transmitted pulses via multipath, thereby limiting the performance of the system. As compared to the others nodes, the impact of the ISI for the Node 1 is not that severe. This is because the bit information was concentrated in the first multipath component. So, the ISI does not pose any problem in cases where the pulse was placed in one of the first clusters. However, the received signal detection for the Node 4 is problematic when the effect of ISI was considered. For example, when the SNR = 15 dB, the BER is 0.05 for node 1, while this value increase to 0.35 for Node 4. To keep a good performance against ISI noise, and in order to be able to achieve a satisfactory performance for Node 4, it is necessary to use multiple pulses per bit. It has been seen that the increase the number of pulses can reduce the BER without increasing transmitter power.

Perf. Analysis of UWB-BSN for Uses in Medical Applications

10

BER

10

10

10

10

479

0

−1

−2

−3

Node 4 Node 3 Node 2 Node 1

−4

0

5

10 15 Ep/N0 (dB)

20

25

Fig. 3. Performance of the TR receiver node in absence of ISI vs. Ep /N0

10

−1

BER

10

0

10

10

−2

Node 4 Node 3 Node 2 Node 1

−3

0

5

10 Ep/N0 (dB)

15

20

Fig. 4. Performance of TR receiver in presence of ISI

6

Conclusion

The UWB technology has the potential to enable low-power consumption, high data rate communications within short distance and other characteristics that make it an ideal candidate for wireless body area networks. The objective of this

480

A. Chehri and H. Mouftah

paper was to analyze the performance of using UWB signalization for medical application. We evaluate the feasibility of BSN architecture based on sub-optimal communication. Based on non-coherent Transmitted Reference receiver, and for peer-to-peer communication, BER has been evaluated for two scenarios. Single user case has been addressed in this paper. Hence, for the future works, the performance of the receivers in multiuser scenario, when all nodes are transmitting in the same time, should be studied.

References 1. Mahar, M.: Money-Driven Medicine: The Real Reason Health Care Costs So Much, Harpercollins Trade Sales Dept., 1st edn. (July 7, 2009) 2. Hongliang, R., Meng, M., Chen, X.: Physiological Information Acquisition through Wireless Biomedical Sensor Networks. In: Proceedings of the 2005 IEEE International Conference on Information Acquisition (June 27-July 3, 2005) 3. Lymberis, A.: Progress In R&D On Wearable And Implantable Biomedical Sensors For Better Health care And Medicine. In: Proceedings of the 3rd Annual International IEEE EMBS Special Topic Conference on Microtechnologies in Medicine and Biology Kahuku, Oahu, Hawaii (May 12-15, 2005) 4. Kohno, R., Hamaguchi, K., Li, H., Takizawa, K.: R&D and standardization of body area network (BAN) for medical healthcare. In: Proc. IEEE Intl. Conf. on Ultra-Wideband (ICUWB 2008), Hannover, Germany, September 10-12, vol. 3, pp. 5–8 (2008) 5. Gandolfo, P., Radoviu, D., Saviu, M., Simiu, D.: IEEE 802.15.4a UWB-IR radio system for telemedicine. In: Proc. IEEE Intl. Conf. on Ultra-Wideband (ICUWB 2008), Hannover, Germany, September 10-12, vol. 3, pp. 11–14 (2008) 6. Liebert, M.A.: State-of-the-Art Telemedicine/Telehealth: An International Perspective,Inc., 2 Madison Avenue, Larchmont, NY 10538 7. Khan, J.Y., Yuce, M.R., Karami, F.: Performance Evaluation of a Wireless Body Area Sensor Network for Remote Patient Monitoring. In: 30 Annual International IEEE EMBS Conference, Vancouver, Canada (August 20-24, 2008) 8. Ryckaert, J., et al.: Ultra-wideband transmitter for low-power wireless body area networks: Design and Evaluation. IEEE Trans. Circuits and Syst. I: Regular Papers 52(12) (December 2005) 9. Lee, C., Kim, J., Lee, H., Kim, J.: Physical Layer Designs for WBAN Systems in IEEE 802.15.6 Proposals. In: 9th International Symposium on Communication and Information Technology, Incheon (September 28-30, 2009) 10. Keonge, H.C., Yuce, M.B.: Low data rate ultra wideband ECG monitoring system. In: Proceedings of the 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Vancouver, BC (2008) 11. Tseng, S.Y., Tsai, C.H., Lai, Y.S., Fang, W.C.: A wireless biomedical sensor network using IEEE802.15.4. In: IEEE/NIH Life Science Systems and Applications Workshop, LISSA 2009, pp. 183–186 (2009) 12. Ch´ avez-Santiago, R., Khaleghi, A., Balasingham, I., Ramstad, T.A.: Architecture of an ultra wideband wireless body area network for medical applications. In: Proc. 2nd IEEE Intl. Symp. on Applied Sciences in Biomed. and Commun. Technol. (ISABEL 2009), Bratislava, Slovakia (November 24-27, 2009)

Perf. Analysis of UWB-BSN for Uses in Medical Applications

481

13. Wambacq, C., VanBiesen, P., Fort, L., Desset, A.: Body area UWB rake receiver communication. In: IEEE International Conference on Communications, vol. 10, pp. 4682–4687 (2006) 14. Jauchem, J.R., Seaman, R.L., Lehnert, H.M., Mathur, S.P., Ryan, K.L., Frei, M.R., Hurt, W.D.: Ultra-wideband electromagnetic pulses: Lack of effects on heart rate and blood pressure during two-minute exposures of rats. Bioelectromagnetics 19(5), 330–333 (1998) 15. Ghavami, M., Michael, L.B., Kohno, R.: Ultra Wideband Signals and Systems in Communication Engineering. John Wiley & Sons, New York (2004) 16. Di Benedetto, M.-G., Giancola, G.: Understanding Ultra Wide Band Radio Fundamentals. Prentice Hall PTR, Englewood Cliffs (2004) 17. Blount, M.: Remote health-care monitoring using Personal Care Connect. IBM Systems Journal 46(1) (2007) 18. Baker, R., et al.: Wireless Sensor Networks for Home Health Care. In: 21st International Conference on Advanced Information Networking and Applications Workshops (AINAW 2007). IEEE, Los Alamitos (2007) 19. Ben Slimane, J., Song, Y.Q., Koubaa, A., Frikha, M.: A Three-Tiered Architecture for Large-Scale Wireless Hospital Sensor Networks. In: MobiHealthInf 2009, pp. 20–31 (2009) 20. Gyselinckx, B., Van Hoof, C., Donnay, S.: Body area networks: the ascent of autonomous wireless microsystems, pp. 73–83. Springer, Heidelberg (2006) 21. Fort, A., Desset, C., Ryckaert, J., Doncker, P.D., Biesen, L.V., Donnay, S.: Ultra wide-band body area channel model. In: IEEE International Conference on Communications, Seoul, Korea (2005) 22. http://grouper.ieee.org/groups/802/15/pub/04/

A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification of Urban Traffic Conditions Øyvind Risan1 and Evtim Peytchev2 1

Norwegian University of Science and Technology, Åbyfaret 141, N-1392 Vettre, Norway [email protected] 2 Nottingham Trent University, Computing and Informatics Building, Clifton Lane, Nottingham, NG11 8NS, United Kingdom [email protected] Abstract. This paper proposes a vehicle-to-vehicle (V2V) communication protocol which makes it possible to discover and share traffic status information in a novel, efficient and comprehensive way. The protocol is specifically designed to work in an environment without infrastructure where all the vehicles (nodes) can talk to each other (ad-hoc network) and collaboratively generate new knowledge relevant to the traffic conditions existing at that moment in an urban environment. The nature of such a network demands self-configuration and autonomous behaviour. The protocol adheres to these principles and makes it possible for the nodes to initiate discovery and determine the location of areas where specific traffic conditions apply. The proposed “Single Ripple” algorithm determines these areas by only involving vehicles with the desired conditions and their neighbours. The algorithm imposes only a minimal load onto the wireless network. Keywords: Traffic information, ad-hoc networks, area discovery, ubiquitous networking, traffic conditions communication protocol, vehicle-to-vehicle communication, V2V, collaborative (cooperative) wireless traffic information systems.

1 Introduction 1.1 Background Global transportation problems are becoming more difficult to solve year after year as the complexity of the traffic network and the number of vehicles on the roads increase. Research in the area of demand-responsive traffic control already deals with many complex problems [1] and vehicles in their own right are becoming more and more complex – for example, in a modern car, several hundred sensors are used to keep the car working properly. It is also widely accepted that the majority of the modern vehicles are well equipped with computation and communication hardware. The big question now is how to harness the potential benefits of this new in-car resource to the full benefit of the driver, the transportation system and society [2], [7], [8]. So far, system-wide solutions to urban transportation problems have relied on centralised traffic control. These solutions have worked well in the context of previous generation telemetry systems but are bound to be overtaken by a new generation of solutions enabled through a new generation of protocols based on the principles of ad-hoc wireless communication networking. Existing algorithms that provide routing information throughout the whole network are not applicable, as these involve every J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 482–494, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification

483

node within the network. Similar algorithms that are based on point-to- point would fail to discover the boundaries of a condition efficiently with the desired flexibility. After equipping a vehicle with communications technology it will no longer be an isolated node but become a part of a bigger system. Exchange of information between vehicles might prevent accidents and increase safety & efficiency [3], [6], [9], [10]. The information exchange, however, is on a peer-to-peer basis as opposed to sending all the information through a central point (server). By gathering the available information in such manner it is possible to provide new services and discover new functionalities in the wireless systems. Some countries have introduced a so called “Zero-Road-Fatality-Vision” policy [1]. The aim of these policies is to reduce the number of traffic related-fatalities to a minimum. The implementation of technology enabling vehicle-to-vehicle communication can provide a valuable contribution to achieving this vision. 1.2 Problem Description The research questions we are trying to answer in this paper are: • Given the vast amount of vehicles in our cities, can we utilise their communication and computing power to collect and distribute traffic information in an efficient manner? • What traffic conditions can we discover using vehicle-to-vehicle communication and collaboration? • Can we use information available in a single vehicle, combined with the information in other single vehicles, to detect traffic conditions? • How big is the area where these traffic conditions exist? Such traffic conditions might be ice on the road, fog, rain, snow, grid lock and so on.

2 The Ad-Hoc (Peer-to-Peer) Networking Approach 2.1 Basic Description of the System Vehicle-to-vehicle communication can be considered a form of communication in a mobile ad-hoc network, often with multiple hops. The fact that the sender and the receiver are placed in a vehicle, that can reach speeds far beyond any pedestrian, presents some challenges. At low speeds there are few location changes per second, but as speed increases the number of location changes per second rises dramatically. This fact demands that the system must be robust and self configuring. A system might consist of just two vehicles, and cover a very small geographical area. The system might also consist of hundreds of vehicles and cover a huge area. The density of vehicles and the location of each vehicle will determine the size of the covered area. Due to limitations on the transmitted power, and signal propagation, there will be a limit as to how far two adjacent vehicles can be apart before the system must be considered to be two autonomous systems, or sub-systems. The system will only exist as long as there are vehicles that have information to exchange, and as long as there are vehicles available to keep the information alive. If there are no vehicles in the system, there is no information of interest to be exchanged, and the need for the system disappears. Figure 1 describes how vehicles (both cars and buses) might communicate with each other, and with infrastructure.

484

Ø. Risan and E. Peytchev

Fig. 1. Vehicles and infrastructure working together

2.2 Message Exchange Assumptions A transmission of messages relevant for the “collaborative traffic condition area discovery” can be conducted in two ways: through retransmission either to specific addresses or to everyone. The latter method of communication will in this paper be referred to as “pure flooding”. Transmission to specific addresses is most effective if the environment and the receivers are static, or motionless. Once the discovering and determination of neighbours are finished, each message needs only be sent once, which saves capacity and time. This solution might demand special hardware such as directive antennas and controllers.

3 Protocol Description 3.1 Protocol Description – Area Discovery In order to discover how many vehicles have a specific condition, or the size of the condition the protocol utilises a special algorithm called “Single Ripple”. The essence of this algorithm is to have a single vehicle act as a trigger for the area discovery process, which normally would be the first vehicle to identify the existence of the traffic condition (e.g. slippery surface). It issues a request for area discovery which is retransmitted as a “lake wave” or a “ripple” geographically directed outwards across all vehicles. As soon as the message reaches a vehicle without the specified condition, the message (the “ripple”) is sent back (bounced) to the originator of the message. Analysing the GPS positioning of all vehicles that bounced the message provides the boundaries of the traffic condition area that is to be discovered. The minimal format of the discovery message can be constructed as illustrated in Figure 2. The field “Direction” indicates in which direction1 the message is being transmitted. It is sufficient to identify whether the message is “outwards” (away from the originator), or if it is a “return” (heading back to the originator as a reply). This field should only be changed by an originator, or a replier (bouncer). 1

Not in a geographical sense, but outwards from, or returning to the originator.

A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification

485

The content of the “Reply-sender + Status” field is dependent on the direction in which the message is travelling. In an outwards message the originator uses its own address and conditions, while in a return message the replier places its address and status in this field. This way, the intermediate vehicles can update their information tables with as much information as possible.

Fig. 2. Structure of a discovery-message

There are two ways to determine the area of a condition. One can find all the vehicles that have the condition, and determine the area by comparing their positions. This will give a very accurate description about where the conditions apply, except for the fact that it fails to discover the borders. The result is an underestimate of the size, and would be on the form: “The area is at least this big”. This information is interesting, but not as useful as a result on the form “The area is smaller than this”. By constructing the discovering algorithm carefully we can reduce the amount of uncertainty regarding the result. The information we want is the position of vehicles that do not have the condition, but which have a neighbour that does. The border of the condition then has to be between these two vehicles. A rain cloud can be used as an example, since it will illustrate the area determination quite clearly. The function of area determination is of course applicable to many other fields and conditions. This paper only focuses on getting information from the vehicles; it does not explore how this information could be utilized. Figure 3 illustrates 5 vehicles2 inside a rain cloud. These vehicles communicate with each other, and with the vehicles3 just outside the cloud. The vehicles on the outside of the cloud respond by returning their positions, and information that they do not have rain, to the vehicles inside the cloud. The vehicles inside the cloud update their information tables, and pass this information to the other vehicles inside the cloud. Based on the data in the information tables the vehicles can calculate the area that the cloud covers.

Fig. 3. Vehicles determining the size of a raincloud 2 3

Green circles. Red circles.

486

Ø. Risan and E. Peytchev

Start / Idle

Do nothing

Yes

Recieve Message

Originate area discovery

Message received before? (1)

Send “Outwards” message

No

Add message to list of received messages (2)

Return

No

Direction? (4)

Outward

Do I have condition? (3)

Split/order conditions (5) Yes

Forward the message containing all the conditions I have (6)

Forward message unaltered (7)

No

Conditions left in the list? (8)

Yes

Return message with own coordinates & condition (9) Remove condition from list

Fig. 4. Algorithm for discovering areas

A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification

487

4 Implementation 4.1 Area Discovery Figure 4 illustrates the “Single Ripple” algorithm for message handling when a condition area is discovered. Each vehicle starts off in the “start/idle” state before it migrates through the various states and actions, until it ends up back in “start/idle” again. The algorithm presented in Figure 4 illustrates the originator deciding to initiate an area discovery, and how every other vehicle reacts when it receives such a message. This algorithm allows the originator to inquire about multiple conditions at the same time. This will make it possible to discover the size of several areas with one message. A scenario that can be used as an example is the one presented in Figure 5. This scenario consists of four vehicles and two conditions (rain and ice). In this scenario the connections are as follow: 0->(1), 1->(0,2), 2->(1,3), 3->(2).

Rain & Ice

Rain 0 3

1 2

Fig. 5. Example of two conditions scenario

The area discovery is initiated by Vehicle 0. The aim is to obtain information about the respective areas. In this example, the number of vehicles (0 to 3) is kept at a bare minimum to keep the amount of messages manageable. The vehicles are positioned so as to illustrate as many different combinations of behaviour as possible. The flow of messages is presented in Figure 6. In this example, no duplicate messages are shown, and neither are messages that are redundant. The red vehicle (Vehicle 0) starts off the discovery by sending out a query about Rain- and Ice-area. A vehicle should discard any message that has either been seen before or that is without relevance to that vehicle. There are two reasons for a message to be of no relevance to the vehicle: The vehicle has already seen another version of the same message, or the vehicle is “outside” the area of interest, and will correctly discard any “return” messages. In accordance with the algorithm given in Figure 4 the “Timestamp” and “Originator+Condition.” remain unaltered at all times. These fields identify messages related to the query originated by “Vehicle 0”. The notation (!Ice) means (“not Ice”), and represents the fact that the vehicle is not within the Ice area.

488

Ø. Risan and E. Peytchev Outwards

Return

Car0 has both “Rain” and “ICE”, originates an area discovery Car0 + Car0 + 10.00 x0,y0 Out ”Rain+ICE” ”Rain+ICE”

0 x0,y0

10.00

Car0 + ”Rain+ICE”

x2,y2

Ret

Car2 + ”!ICE”

10.00

Car0 + ”Rain+ICE”

x3,y3

Ret

Car3 + ”!Rain”

Car1 forwards the messages received from Car2 Car1 has both “Rain” and “ICE”, forwards the outwards message Car0 + Car0 + 10.00 x0,y0 Out ”Rain+ICE” ”Rain+ICE”

Car2 has “Rain”, but not “ICE”, forwards the part of the outwards message that contains the conditions it has. 10.00

Car0 + ”Rain+ICE”

x0,y0

Out

1 x1,y1

2 x2,y2

10.00

Car0 + ”Rain+ICE”

x2,y2

Ret

Car2 + ”!ICE”

10.00

Car0 + ”Rain+ICE”

x3,y3

Ret

Car3 + ”!Rain”

Car2 replies that it does not have “ICE” in a new return message, and forwards the message received from Car3

Car0 + ”Rain”

10.00 Car3 has no “Rain”, does not forward the outwards message

3

Car0 + ”Rain+ICE”

x3,y3

Ret

Car3 + ”!Rain”

Car3 replies that it does not have “Rain” in a new return message

x3,y3

Fig. 6. Example of a Discovery-message flow

5 Simulation The hypothesis is: The “Single Ripple” algorithm will perform faster, and with a smaller load on the system than would “pure flooding”. The goal is to test how the algorithm in Figure 4 would perform against a standard “pure flooding” algorithm. By recording the performance of each algorithm it is possible to compare how different numbers of vehicles affect the algorithms. Relevant measurements: • • • •

How many packages were sent. How many packages were received. Time between first and last package received. How many packets where dropped due to interference.

The scenario used in the simulation can be seen in Figure 7. It is similar to the examples presented earlier in this paper, but it contains 16 vehicles and two semioverlapping conditions. The vehicles are given conditions based on their location. Vehicle 0 and 1 have two conditions, vehicle 2 and 5 have one condition (but not the same); all other vehicles do not have any conditions.

A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification

489

The placement, and numbering, of the vehicles allow testing with 4, 8, 12 and 16 vehicles without having to change their positions. To test with 4 vehicles simply remove vehicles 4 – 15. To test with 8 vehicles simply remove vehicles 8 – 15. To test with 12 vehicles simply remove vehicles 12 – 15. To test with 16 there is no need to remove any vehicles. 6

14 15 13

12

5

11

4 7

1000 m

0

10

1

3 2

9 8

1000 m

Fig. 7. Scenario for the simulator

The distances between the vehicles are rather large4; this is done to make the simulator work under as difficult conditions as possible. The simulation is based on standard 802.11 communications, as this is the most widespread technology, and a very likely candidate for the communication between the vehicles. The simulation was performed with immovable vehicles. The main concern about simulating with static vehicles is that this ignores the difference in interference patterns you would expect with any change in inter-vehicular positions. However one can assume that the distance each vehicle moves during the time it is actually involved in the communication is so short that its movement will not significantly influence the pattern of interference. From the simulation result one can see that the longest average time any one vehicle was involved in the communication is 44.1 ms, at “pure flooding” and “16 vehicles” (Table 1). The average speed in some major cities in the UK is just 17.8 mph (= 28.2 km/h). An average speed of 30 km/h means that a vehicle will travel 8.3 meters each second. In 44.1 ms the vehicle would then have moved 36.6 cm. This movement is so small that it can be neglected in this simulation. At a speed of 90 km/h the vehicles will move 1.1 meter in 44.1 ms. 4

100m < Distance between vehicles > 250m.

490

Ø. Risan and E. Peytchev

6 Results The results gathered from the simulator are displayed in Table 1. The table has four main parts, one for each group of vehicles (16, 12, 8 and 4). Each of these groups is presented with results for two scenarios: “Single Ripple” algorithm and “pure flooding”. Table 1. Results from the simulator with 16, 12, 8 and 4 vehicles 16 Vehicles

Avg time TX RX Dropped Total D/Total D/Vehicle D/TX D/RX

12 Vehicles

8 Vehicles

4 Vehicles

"Ripple"

Pure

"Ripple"

Pure

"Ripple"

Pure

"Ripple"

Pure

4.00 16 26 16 42 38.10 % 1.0000 4.0000 0.6154

44.10 143 282 143 425 33.65 % 8.9375 3.2426 0.5071

5.50 18 37 9 46 19.57 % 0.7500 1.6364 0.2432

31.20 89 157 74 231 32.03 % 6.1667 2.3718 0.4713

6.90 20 41 7 48 14.58 % 0.8750 1.0145 0.1707

17.30 44 80 16 96 16.67 % 2.0000 0.9249 0.2000

4.00 10 16 0 16 0.00 % 0.0000 0.0000 0.0000

7.80 16 24 0 24 0.00 % 0.0000 0.0000 0.0000

• “Avg time” 5 refers to the time between the first and last message at each vehicle. The average time is based on vehicles that are actually involved, leaving out any vehicles that have received 1 packet or less. • “TX” refers to the number of messages sent by each vehicle. • “RX” refers to the number of received messages. • “Dropped” refers to the number of packets lost due to interference. • “D/Total” is the number of dropped packets in relation to the total number of packets received, and dropped, in the system. This is a measure of the interference between the communicating vehicles. • “D/Vehicle” is the number of dropped packets pr vehicle. This gives an average of how many packets each vehicle has been unable to receive. • “D/TX” is the number of dropped packets pr sent packet. In a wireless system, each transmission has potentially several receivers, and it is the receiver that drops the packet. • “D/RX” is the number of dropped packets pr successfully received packet. This gives an average of packet loss in the system.

7 Discussion The results in Table 1 are more clearly illustrated with the help of some graphs. Graph 1 compares how the “Single Ripple” algorithm and the “pure flooding” algorithm performed in terms of how many packets were sent, received and dropped. The shape of the curves belonging to “pure flooding” is as expected, while “sent” and “received” from “Single Ripple” has a somewhat unexpected shape. 5

-3

In 10 seconds (ms).

A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification

491

Sent, received and dropped packets 300 275 250 225 200

Pure RX

Packets

175

Pure TX

150

Pure Dropped

125

Ripple RX Ripple TX

100

Ripple Dropped 75 50 25 0 16 Vehicles

12 Vehicles

8 Vehicles

4 Vehicles

Graph 1. Complete presentation of sent, received and dropped packets

It is evident that the number of sent and received messages peaks at “8 Vehicles” when using the “Single Ripple” algorithm. The fact that the number of received messages drops with increasing numbers of vehicles is due to increased interference in the system. This limits the throughput in the system. In “pure flooding” the increased interference is masked by the sheer number of packets being transmitted. As seen by Graph 1, “Single Ripple” reduced the number of messages being received by 91% compared to “pure flooding” (16 vehicles). Furthermore, “Single Ripple” sends 89% fewer packets than “pure flooding” with 16 vehicles. This means that the load on the scarce resources in the system is decreased. “Pure flooding” involves more vehicles than is strictly necessary to get the requested information, but offers alternative routes of communication, making up for some of the lost packets. “Single Ripple” uses the resources in a small number of vehicles, and no resources in the others; “pure flooding” uses the same amount of resources in all the vehicles. Based on the values from Table 1 it is evident that the vehicles involved in “Single Ripple” spend far less time processing messages; the time is reduced by 91% compared to “pure flooding” (16 vehicles). As can be seen in Graph 2, “Single Ripple” reduces the amount of interference at 12 vehicles, but it increases again at 16 vehicles compared with “pure flooding”. “Pure flooding” has an unexpected slow rise from 12 to 16 vehicles, but this can be attributed to the fact that most of the new vehicles from 12 to 16 are placed at the

492

Ø. Risan and E. Peytchev

edge of the simulation area. The position of the new vehicles means that they contribute to the total number of messages without increasing the interference to the same effect, and thus “pure flooding” seems less prone to interference within the area than “Single Ripple” when more vehicles are involved. Each sent and received message in “Single Ripple” contributes more to the result than does each message in “pure flooding” since the information is specific to the conditions we want to explore. This fact also makes each dropped packet more valuable. The communication protocol used in the system is UDP, which was chosen because TCP introduces an extra load. No retransmission of lost or damaged packets is a problem when using “Single Ripple”, as the algorithm aims at involving as few vehicles as possible. The possibilities of receiving the information via alternate routes are slim. In “pure flooding” there is more redundancy, and therefore it is more likely that the information will be received through several different routes. The presence of several identical messages increases the potential for the receivers to actually receive the message.

Interference 0.40

% of total packets

0.35 0.30 0.25 0.20

Pure D/Total

0.15

Ripple D/Total

0.10 0.05 0.00 16 Vehicles 12 Vehicles 8 Vehicles

4 Vehicles

Graph 2. Dropped packets out of the total number of packets

At least two solutions to the problem of lost messages in “Single Ripple” can be suggested: • Let the vehicles update each other at regular intervals. This would, over time, cover the holes left by lost messages. • Use the fact that the vehicles move. After the first run, let some time pass before the same discovery is initiated again. After 1 second, with an average of 30 km/h, the vehicles will have moved 8 meters, and after 10 seconds they will have moved 80 meters. This might be enough to change the pattern of packet loss, and thereby gain new information.

A Vehicle-to-Vehicle Communication Protocol for Collaborative Identification

493

“Pure flooding” is known to be problematic when it comes to efficiency, but it is the only alternative algorithm that was suitable for comparison against “Single Ripple” when it comes to discovering the border. The simulations have been carried out on a small selection of vehicles to show the principles behind the algorithm. With larger numbers of vehicles the amount of interference would increase, especially with “pure flooding” when the entire network is involved, and the damaging effect of lost messages in “Single Ripple” would be reduced. Security related to “Single Ripple” is not considered, as this is more related to how the messages are processed before and after transmission. The reliability of “Single Ripple” is dependent upon the density of the vehicles around the conditions in question, and will increase with higher densities, as there will be more retransmissions and more details available.

8 Conclusions This paper presents new vehicle-to-vehicle communication protocol capable of discovering areas with a specific traffic condition – e.g. congestion, slippery area, rain etc. The protocol is highly efficient – it shows appr. 10 times better results than “pure flooding” protocol. It is infrastructure-less, but this is not a restriction – it can make use of any additional nodes alongside the road. The protocol is intended to be used in all collaborative schemes for cooperative vehicle information generation and traffic information distribution.

References 1. Cooperative Vehicle Infrastructure Systems – CVIS, EU FP6 Project Reference: IST2004-027293, Contract Type: Integrated Project (IP), Project Cost: €€41.155.203, EC project funding: €€21.905.795 2. Choffnes, D.R., Bustamante, F.E.: An integrated mobility and traffic model for vehicular wireless networks. In: Proceedings of the 2nd ACM International Workshop on Vehicular ad hoc networks, Cologne, Germany (September 02-02, 2005) 3. Korkmaz, G., Ekici, E., Özgüner, F., Özgüner, Ü.: Urban multi-hop broadcast protocol for inter-vehicle communication systems. In: VANET 2004, Proceedings of the 1st ACM International Workshop on Vehicular ad hoc networks ©2004 table of contents, ISBN:158113-922-5, doi:10.1145/1023875.1023887 4. Fax, J.A., Murray, R.M.: Information flow and cooperative control of vehicle formations. IEEE Transactions on Automatic Control (2004) 5. Park, J.-S., Lee, U., Oh, S.Y., Gerla, M., Lun, D.S., Ro, W.W., Park, J.: Delay Analysis of Car-to-Car Reliable Data Delivery Strategies Based on Data Mulling with Network Coding. IEICE - Transactions on Information and Systems E91-D(10), 2524–2527 (2008) 6. Molisch, A., Turfvesson, F., Karedal, J., Mecklenbrauker, C.: Propagation Aspects of Vehicle-to-Vehicle Communications - An Overview. In: IEEE Radio & Wireless Symposium (January 2009) 7. Thomas, M., Peytchev, E., Al-Dabass, D.: Auto-sensing and distribution of traffic information in vehicular ad hoc networks. International Journal of Simulation 5(3), 59–63 (2004)

494

Ø. Risan and E. Peytchev

8. Chen, W., Cai, S.: Ad hoc peer-to-peer network architecture for vehicle safety communications, vol. 43(4), pp. 100–107 (April 2005); ISSN: 0163-6804, INSPEC Accession Number: 8375775, Digital Object Identifier: 10.1109/MCOM.2005.1421912, Current Version Published: 2005-04-18 9. Yang, X., Liu, J., Zhao, F., Vaidya, N.: A Vehicle-to-Vehicle Communication Protocol for Improving Road Safety. In: The 1st International Conference on Mobile and Ubiquitous Systems: Networking and Services (Mobiquitous 2004), Boston, MA (August 22-26, 2004) 10. Mylonas, Y., Lestas, M., Pitsillides, A.: Speed adaptive probabilistic flooding in cooperative emergency warning. In: Proceedings of the 4th Annual International Conference on Wireless Internet, Maui, Hawaii (November 17-19, 2008)

A Practical Evaluation of ZigBee Sensor Networks for Temperature Measurement Abdellah Chehri and Hussein Mouftah School of Information Technology and Engineering (SITE) University of Ottawa 800 King Edward Avenue Ottawa, Ontario, Canada, K1N 6N5 {achehri,mouftah}@uottawa.ca

Abstract. Wireless Sensor Networks (WSNs) offer numerous advantages over traditional networks, such as elimination of costly wires, enhanced monitoring precision and larger area coverage. This paper presents the design and implementation of a mine temperature monitoring based on sensor networks. For the proposed monitoring schema, we evaluated the performance and the interoperability of sensor network with various network such as IEEE 802.11g (WiFi), IEEE 802.11s (wireless mesh network) and Internet. So the ambient temperature of a mine gallery can be measured and displayed in real time no matter where we are. In addition we describe some initial results of link characteristics. We discuss on the sensor wireless link performance in terms of the received signal strength. Keywords: sensor networks, zigbee, underground mines.

1 Introduction A wireless sensor network (WSN) in its simplest form can be defined as a network of low-size, low-complex and locally powered sensor nodes that can sense the environment and communicate the information gathered from the monitored field through wireless links. The data is forwarded via multiple hops relaying to a central node (or sink) that can use it locally, or is connected to other networks through a gateway. The claim of wireless sensor network proponents is that this technological vision will facilitate many existing application areas and bring in to existence entirely new ones [1], [2]. Detection of world's physical parameters makes sensors most suitable technology for monitoring. Sensors though are not just limited to environment sensing. Any application involving sensing of physical parameters like sound, humidity, pressure, temperature, etc, might use sensor network. Considering the importance of this technology, a deployment of WSN in mining industry can be considered as original application. For example, an operator could remotely supervise different physical phenomena in the mine from his computer and thus provide safe air/oxygen for the miners underground by monitoring the level of methane and other noxious gases, dust and particulates from sources such as diesel J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 495–506, 2010. © Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010

496

A. Chehri and H. Mouftah

vehicles. He could also monitor the temperature, detect any anomaly, or locate workers and objects in the galleries of the mine in real time. These applications can be seen as a first step toward the concept of “smart mine”. Usually, we verify performance of any network by using simulations or experiments. In simulations, we cannot control precise packet timing, radio range transmission, an unlimited memory and processing resources, and real PHY/MAC layer events. In fact, not all simulation results are equal to the real experiments. In real experiments, we have complex environment settings and resource sharing problems. In this paper, we discuss the challenges and requirements of developing efficient WSN for temperature monitoring for mining industry. For this reason, we have set up a testbed at CANMET1 that provides a heterogeneous platform to support realistic environmental data measurement application. The remainder of this document is organized as follows. Section 2 provides a review and motivation of using wireless sensor network in mining industry. Then, in Section 3, we will present the characteristics of the commercial sensor used in our experiments. Section 4, describes the measurements, and gives the results and the most important challenges. Finally, we will conclude with Section 5.

2 Motivation for Using WSN Underground mining is a hazardous industrial activity. The harsh physical environment and distinct topology that make mining dangerous act as a hindrance or constraint to the very techniques and technologies that could improve safety and productivity. Working conditions in underground mining are associated with a considerable number of health risk factors, such as a high physical workload, fire and radiation exposure, high temperature and humidity conditions and exposure to dust and gas phase hazardous substances. Generally, the motivation for using wireless sensors in mining industry is twofold: economy and safety. From a system operation perspective, wireless sensors give an opportunity to safely and cost efficiently increase measurement coverage of the network, including locations where wiring is impossible. In fact, this emerging technology, unconstrained by expensive wiring, has the potential to provide operating efficiencies. These efficiencies are made possible through reduced installation costs, lower operating costs, installation flexibility, and scalability. Wireless sensor networks allow faster deployment and installation of various types of sensors because many of these networks provide self-diagnosing, self-configuring self-organizing and self-healing capabilities to the sensor nodes. Some of them also allow flexible extension of the network. Another advantage of wireless sensors is their mobility. These sensors can be placed in transporting vehicles to monitor the environment. They also can be attached on rotating equipment, such as a shaft to measure critical parameters. However, despite all its advantages the deployment of wireless sensors in underground mines is still at the beginning stage. Three majors application scenarios of wireless sensor networks can be identified [3]. 1

Canadian Center for Minerals and Energy Technology (CANMET) experimental mine.

A Practical Evaluation of ZigBee Sensor Networks for Temperature Measurement

497

2.1 Environmental Data Measurement In this scenario, the nodes are located at fixed position in the mine. Some node may be equipped with a sensor, while others are routing the data. The important parameters that could be measured in underground mines are: temperature, oxygen concentration, humidity. The global network is characterized by a great number of nodes which collects and measures in a continual way of the data towards a central unit. The management of the networks is treated in the high level of the networks. A sensor management protocols determine which nodes are needed to provide data (sensing mode selection) and which are needed to ensure a connected topology and the data routing (topology control). This last aims to rotate active nodes; therefore the energy drain of performing routing task is distributed evenly among all the nodes in the networks [4]. Other configuration process can be used to find the upper bound of attainable lifetime using sensor management algorithms that determine sensing, routing and data fusing roles for each sensor in the networks [5]. The data are periodically transmitted from child node to the parent node. For much of scenarios the typical periods of measurements are about few minutes because the supervised parameters of environment, like the temperature, the intensity of the light, do not change rather quickly to require measurements with very close intervals [6]. 2.2 Security Monitoring For this type of application, the nodes are placed at fixed locations in order to supervise continuously some parameters. The goal is to detect possible anomalies like fires, gas explosions, premature explosions of charges, toxic gases (carbon monoxides, carbon dioxide), or even a roof failure using microseismic and rock deformation sensors. 2.3 Localization Monitoring the precise location of mobile assets in underground mines is valuable information not only for safety but also as an enabler of business process optimization such as ventilation-on-demand, automated logging of LHD mucking cycles, or traffic light control in ramps. Since GPS does not work underground, an alternative method of mobile asset tracking must be implemented. Localization and tracking of moving objects is an essential capability for a sensor network. For this kind of application, we suppose that any agent (such as a miner or a vehicle) entering in underground mine can accurately located. We assume that a large number of anchors (sensor node with known location) are deployed in underground mines. The agent is equipped with small sensor, this sensor can estimate it distance from these anchors (ToA), and then use the positioning algorithm (triangulation) to determine its own position. This location can be sending over the networks and collected by central node to indicate the position of this agent. Therefore, wireless sensor networks can provide an ideal solution to track and locate the miners. It can be considered as an end-to-end rescue communication network for miners during an incident [7].

498

A. Chehri and H. Mouftah

3 Wireless Sensor Networks Testbed When choosing deployment of WSN in underground mine, it should be necessary to make a compromise between conflicting requirements. The priority is to insure a robust global network with battery-operated nodes. Therefore, the network was developed with the following goals in mind. Firstly, the node must be able to communicate with other notes via a highly reliable radio module compatible with the IEEE 802.15.4 standard. Secondly, the network should be robust to monitor temperature for long time. We deployed a wireless sensor network in April 2009. The network contains all elements of the architecture (will be described in the next subsection). To withstand the temperature conditions, dust, and the strong humidity present in underground mines, we designed environmental protective packaging that minimally obstructs sensing functionality. The selected motes by their design are fairly robust mechanically, with the battery case firmly integrated with the main processing and sensor boards. Wireless communication is achieved with a transceiver compliant with the IEEE 802.15.4/ZigBeeTM standard. ZigBeeTM is a global standard for wireless network technology that addresses remote monitoring, environmental data measurements and control applications. ZigBeeTM is an open specification that enables low power consumption, low cost and low data rate for short-range wireless connections between various electronic devices. In this section, an overview on the hardware implementation and the software protocol are given. First, a customized wireless communication test platform for evaluating wireless networking protocols is presented. A detailed description of capabilities and limitations of the test platform is discussed. The testbed consists of the following components: • • • • • •

Hardware Description; Software Description; Network Architecture; Networks Topology; Node Deployment; WSN to Internet communication.

3.1 Hardware Description The Silicon Laboratories 2.4 GHz 802.15.4 Development Board (DB) provides a hardware platform for the development of 802.15.4/ ZigBeeTM networks. The DB includes a Silicon Labs 8051-based MCU, a Chipcon CC2420 RF Transceiver, a JTAG (Joint Test Action Group or IEEE 1149.1 standard) connector for in-circuit programming, an assortment of programmable buttons and LEDs and a USB interface for connecting to the host computer. Figure 1. (a) shows a block diagram of the DB. The DB card has been developed with a minimal number of components. This is due in part to the low powerconsumption requirement and in part to the need to keep the mote size and manufacturing costs to a minimum. The core of the platform is a Silicon Labs C8051F121 (MCU) ultra-low power microcontroller. The device is quite powerful

A Practical Evaluation of ZigBee Sensor Networks for Temperature Measurement

499

with an 8051 CPU (100 MIPS). This microcontroller can typically operate at clock frequencies up to 8 MHz with 128 kB of flash memory and 8448 bytes of RAM. Wireless communication is provided by the Chipcon CC2420 radio transceiver. This circuit combines low power and efficient operation with support for IEEE 802.15.4. It operates in the 2.4 GHz Industrial-Scientific-Medical (ISM) unlicensed radio frequency band, with 16 channels. It uses an automatic PCA (Parallel Channel Adapter) and address filtering. The consumption of CC240 is estimated at 19.7 mA for Rx and 17.4 mA for Tx. Automatic acknowledgment transmission is used, and a CRC criterion (Cyclic Redundancy Check) is employed to decide whether a packet was received correctly or not. The radio module is connected via an SMA connector to an omnidirectional antenna. The DB has a total of eleven LEDs. The LEDs are used to show the state of the mote (after reset, sending a message, etc.) and two of them are used for power status indicators. An internal temperature sensor is included in the board with a measuring range of (-40 °C to +85 °C). The DB is powered with a 9 V battery. Some basic parameters are summarized in Table 1.

Fig. 1. The Silicon Laboratories 2.4GHz 802.15.4 mote (a) Development Board, (b) software interface

3.2 Software Description The 2.4 GHz ZigBeeTM development kit contains all necessary files to write, compile, download, and debug a simple IEEE 802.15.4/ ZigBeeTM -based application. The development environment includes an IDE, evaluation C compiler, software libraries, and a several code example. The software library includes the 802.15.4 MAC and PHY layers. The ZigBeeTM demonstration provides a quick and convenient graphical PC-based application. The kit also includes an adapter for programming and debugging from the IDE environment as shown in Figure 1 (b). A Network Application Programming Interface (API) contains all necessary network primitives to build a 802.15.4 network from a user-defined application. A software example illustrates the MAC API. This example builds an ad-hoc 802.15.4 network using the included MAC API software library.

500

A. Chehri and H. Mouftah Table 1. System Performance Summary Parameter Standard Basis Frequency (MHz) RF Bandwidth (KHz) Nb of channel RF Channel Spacing Data Rate Burst Data Rate Burst Modulation Type Spreading Technique PN Code Tx Output Power Rx Sensitivity Typical Supply Current (Tx) Typical Supply Current (Rx normal, frame reception)

System Specifications IEEE Std. 802.15.4™ 2400 - 2483.5 GHz 1200 KHz 16 3 MHz 250 kbit/s QPSK DSSS 15-chip m-sequence 0 dBm (to 50 Ω) -98 dBm 17.4 mA 19.7 mA

3.3 Networks Topology The geographical nature of a mine galley (narrow and elongated corridors) has a direct impact on the design of sensor network applications. In its simplest form, a sensor network is single-hop, allowing every sensor node to communicate directly with every other node.

Fig. 2. The neighbor table manager

A Practical Evaluation of ZigBee Sensor Networks for Temperature Measurement

501

The Silicon Laboratories 2.4 GHz development kit contains several preconfigured network topologies. These topologies are predefined and downloaded first to each node via a USB connector. For our measurements, cluster tree, star and linear topologies were separately adopted. For example, Figure 2 contains the blacklist table for central node of the cluster tree topology. Node A is the designated master (central node) in this topology. Other nodes are Full-Function Device (FFD) routing nodes or Reduced Function Device (RFD) terminal nodes, depending on the network topology selected. 3.4 Network Architecture Wireless sensor network is used to transfer the sensor data frames from the sensor unit over a radio interface to the central node. If a radio link can be established between these modules for peer-to-peer communication, the radio modules put each sensor data frame into a radio message, send the message over the radio link, and extract the sensor data frame from the received radio message. Figure 3 shows that the sensor data are transmitted directly from the sensor node to the central node, which then transmits them to the base station. The network organizes itself and is self-healing, i.e. network nodes automatically establish and maintain connectivity among them.

Fig. 3. Block diagram of the heterogeneous wireless network’s deployment

3.5 Node Deployment The deployment of sensor nodes in the physical environment may take several forms [8]. In the case of an underground mine, the deployment may be random (unexplored part of mine), at deliberately chosen spots on the top of the gallery or at a fixed

502

A. Chehri and H. Mouftah

position on the gallery walls. In manual deployment, the sensors are manually placed and the data are routed through predetermined path. The deployment operation may be a one-time activity, where the installation and the use of a sensor network are strictly separate activities. However, deployment may also be a continuous process, with more nodes being deployed at any time during the use of the network, for example, to replace failed nodes or to improve coverage of the network. 3.6 WSN to Internet Communication For practical deployment, a sensor network only concerned with itself is insufficient. The network rather has to be able to interact with other information devices, for example, a miner equipped with a PDA moving will be able to read the temperature sensors even this node is located in different mine galley. To this end, the WSN first of all has to be able to exchange data with such a mobile device. This schema can be generalized to other important security parameter (carbon monoxides, or smoke concentration, for example). Therefore, for the proposed WSN monitoring system, we evaluated the performance and interoperability of sensor network with various network such as 802.11g (WiFi) and IEEE 802.11s (wireless mesh network). In this schema, the nodes communicate with the central node, which is connected with a laptop on site. This last one has the capability of communicating wirelessly with other computers located in a monitoring room via IEEE 802.11 networks (or wireless mesh network). The number of access points of both WiFi and wireless mesh network are sufficient to ensure a total coverage of mine gallery. The system is connected to Internet through a gateway. Gateways play the role of communication between WSNs and Internet access. We use a single board computer with public IP address as gateway in a WSN. So the ambient temperature of a mine gallery can be measured and displayed in real time no matter where we are. The global schema of WSN mine gallery temperature monitoring is shown in Figure 3.

Fig. 4. Gallery mine (CANMET)

A Practical Evaluation of ZigBee Sensor Networks for Temperature Measurement

503

4 Measurement Setup and Results 4.1 Measurement Setup The measurements were carried out in an underground gallery of the MMSLCANMET laboratory mine located 540 km north of Montreal, QC, Canada. We have performed the measurements at the 70 m level. Figure 4 show an example the node placement in the mine gallery for LOS (line-of-sight). In this measurement campaign, the central node remained at a fixed position whereas the salve node was moved at different locations in mine gallery. The measurements were taken for both static and moving node. 4.2 Link Characteristics In this section, we describe some preliminary results of measured link characteristics from the testbed. Specifically, we discuss some statistics of the wireless link performance in terms of delay, received signal loss, link quality indicator and throughput. 4.2.1 Received Signal Strength Figure 5 plots the received signal strength versus the distance. One can observe two regions of path loss. In the first region (1 to 40 m), signal attenuation is about 40 dB between 1 m and 40 m, which is significant considering that the transmitter and the receiver, in this case, are in line-of-sight. However, the second region (from 40 m to 105 m) is characterized by small signal attenuation. This small attenuation is due to (a)

RSS [dBm]

-40 -60 -80 -100 0

10

20

30

40

50 60 Distance [m]

70

80

90

100

RSS [dBm]

-40 -60 -80 -100 0

10

20

30 40 Distance [m]

50

60

Fig. 5. The received signal of a node vs. distance (a) LOS, (b) NLOS

504

A. Chehri and H. Mouftah

the topology of the gallery. In fact, this region of the gallery is represented as a narrow corridor in which the multipath adds; therefore the signal can travel a long distance with a small attenuation. This is known as the “waveguide propagation phenomenon”. 4.3 Temperature Measurement The ultimate goal of wireless sensor network research is to enable novel applications that change the way we interact with the world around us. Wireless sensor networks allow information to be collected with more monitoring points, providing awareness for the environmental conditions (for fire detection for example) that affect overall uptime, safety, or compliance in mining environments and enabling agile and flexible monitoring and control systems. In the literature, several works deal with WSNs for temperature monitoring [9], [10]. As an illustrative example, we have used the nodes to monitor the ambient temperature of the mine gallery. Note that the nodes could easily be adapted to monitor other types of temperatures (the temperature of a machine). To test how the entire system deal with the gathering of data from multiple sensor networks five nodes were used during measurement. The first wireless sensor network (A) involved a base-station (connected to laptop) and four slave node placed in different positions. The trial was started at approximately 9:00 and finished at approximately 15:00. The data is shown in Figure 6. This figure shows that the temperature in both locations differed by at least 2 °C. The transfer of data across the Wi-Fi and mesh network was largely perfect. 4.4 Some Challenges Real applications of WSN are being explored and some of them are yet to come. While the potential of sensor networks in underground mine is only beginning to be realized, several challenges still remain. One of theme is the complex nature of wave propagation in underground mine. This is due to scattering and rough surfaces diffraction in mine tunnel. So, the RF link budget should include a safety margin of several dBs to ensure reliable communications between nodes. Also, in underground mine environment, a daisy chain of nodes installed along a tunnel must with stand the failure of a node. One effective way add to redundancy is to ensure each node in the chain is in range of at least two other nodes on each side. Networks topology designs should also take into consideration potential attenuation caused by rubble if ever part of mine tunnel collapses. Here we expose some limiting challenges for using WSN in underground mine applications: • Power consumption always an issue. • Topology change due to human activity (system should scale well on a large number of topologies). • It is necessarily to take into account the characteristics of the mine, like gallery size and shape, and the desired coverage. • Radio connectivity varies over time and is very sensitive to position. • Temperature measurements can be affected by changes in surrounding conditions (ventilation system for example).

A Practical Evaluation of ZigBee Sensor Networks for Temperature Measurement

505

12

11

10

Temperature

9

8

7

6

5

Node Node Node Node Node

4

3 8:30

9:00

A B C D E

9:30 10:00 10:30 11:00 11:30 12:00 12:30 13:00 13:30 14:00 14:30 15:00 15:30 Time

Fig. 6. Temperature data recorded simultaneously from five sensor node

5 Conclusion Wireless sensor networks became a key technology and are used in more and more industrial and environmental applications. In this paper we have evaluated a WSN testbed to conduct an integrated monitoring system. We have tried to demonstrate that the proposed architecture can be used to measure the temperature of underground mine gallery in real-time. This scenario can be easily generalized to others environmental parameters (toxic gas detection, humidity, etc). Challenges remain in conducting experiments with larger number of sensor nodes in more complicated scenarios. We highlight all the problems identified at this initial phase, which will influence the final deployment of the complete WSN. This work is being developed in the context of an ongoing project on underground mines environmental monitoring using WSNs.

References 1. Holger, K., Willig, A.: Protocols and Architecture for Wireless Sensor Networks. John Wiley and Sons, Chichester (2005) 2. Lewis, F.L.: Wireless sensor networks. In: Cook, D.J., Das, S.K. (eds.) Smart Environments: Technologies, Protocols, and Applications, New York. John Wiley, Chichester (2004) 3. Chehri, A., Fortier, P., Tardif, P.-M.: Deployment of Ad-Hoc Sensor Networks in Underground Mines. In: Sixth International Conference on Wireless Sensor Networks, WSN 2006, Banff, Alberta, Canada (July 3-5, 2006)

506

A. Chehri and H. Mouftah

4. Perillo, M., Heizelman, W.: Sensor management. In: Wireless sensor networks, ch. 16. Kluwer Academic Publishers, Dordrecht (2004) 5. Bhardwaj, M., Chandrakasan, A.: Bounding the lifetime of sensor networks via optimal role assignments. In: Proceeding of the 21st International Annual Joint Conference of the IEEE Computer and Communications Societies, INFOCOM (2002) 6. Hill, J.: System architecture for wireless sensor networks. PhD thesis, University of California, Berkeley (2003) 7. Chehri, A., Fortier, P., Tardif, P.-M.: UWB-based sensor networks for localization in mining environments. Ad-Hoc Networks 7(5), 179–188 (2008) 8. Romer, K., Mattern, F.: The design space of wireless sensor networks. Wireless Communications 11(6), 54–61 (2004) 9. Flammini, A., Marioli, D., Sisinni, E., Taroni, A.: A real-time wireless sensor network for temperature monitoring. In: Proc. IEEE Int. Symp. Ind. Electron., June 4-7, pp. 1916–1920 (2007) 10. Lee, A., Angeles, C., Talampas, M., Sison, L., Soriano, N.: MotesArt: Wireless sensor network for monitoring relative humidity and temperature in an art gallery. In: Proc. IEEEICNSC, April 6-8, pp. 1263–1268 (2008)

Minimum Total Node Interference in Wireless Sensor Networks Nhat X. Lam, Trac N. Nguyen, and D.T. Huynh Department of Computer Science University of Texas at Dallas, Richardson Texas 75083-0688 {nxl081000,nguyentn,huynh}@utdallas.edu

Abstract. The approach of using topology control to reduce interference in wireless sensor networks has attracted attention of many researchers. There are several definitions of interference in the literature. In a wireless sensor network, the interference at a node may be caused by an edge that is transmitting data [16], or it occurs because the node itself is within the transmission range of another [2], [4], [7]. The interference load of a node is either the number of nodes in the broadcasting disk defined by this node or the number of nodes whose disks cover it [2], [4], [7]. In this paper we show that the problem of assigning power level to a set of nodes in the plane to yield a connected geometric graph whose total node interference is bounded is NP-complete under both definitions. We also introduce some heuristics as well as a simplified version of an O(logn) approximation algorithm in [10] and study their performance through simulation. Keywords: wireless sensor networks, interference, NP-completeness, geometric graphs.

1

Introduction

Wireless sensor networks (WSNs) have been extensively used in military and civilian applications in the last two decades. A primary issue concerning WSNs is interference, which occurs when communication between a pair of nodes is affected by another node that is transmitting data. One well known approach to minimize interference is to use topology control by reducing the power usage of certain nodes thereby establishing a simple connected network with low interference. The approach of using topology control to reduce interference was first discussed in a number of papers including [4], [12]. The authors in [4] defined the notion of interference load of an edge in a network, and showed an interesting result that certain sparse networks may not have low interference. Following the work in [4], the authors in [12] introduced a notion of (receiver-based) node interference that is caused by surrounding nodes whose transmission range includes the given node. They analyzed the special case of the exponential 1-dimensional node chain which is also called the highway model. They showed that this sparse J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 507–523, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

508

N.X. Lam, T.N. Nguyen, and D.T. Huynh

√ network has Ω( n) node interference, where n is the number of nodes in the √ 4 network. The authors described an algorithm that provides an O( δ) approximation of the optimal connectivity preserving topology in the highway model where δ is the maximum node degree. Similarly, the papers [16], [9], and [8] focused on the notion of interference that is based on edges of the network. [8] gave a distributed algorithm called Average Path Interference that tries to preserve the spanner property of the original graph while reducing the interference in the network. [9] showed that the relative neighborhood graph and local spanning tree algorithms have a constant bounded average interference ratio. The NP completeness (NPC) of both types of node interference were discussed in [16], [2], [3], [11] and [14]. [16] extended the work of [4] with an NPcompleteness proof for the problem of minimizing edge interference for general graphs along with a couple of heuristics, and [3] provided an NP-completeness proof for finding a spanning tree with minimum node interference for grid graphs. For the receiver-based interference model, in [2] the authors showed among other results that the problem of minimizing the maximum node interference is hard to approximate. On the other √ hand, [7] showed that for a set of n points in the plane a network with O( Δ) interference can be constructed using computational geometric tools (Δ is the maximum interference in the uniform-radius network). They left open the question whether this problem is NP-hard including the 1-dimensional case. [11] was able to provide an answer to some of the questions raised in [7] by showing that minimizing receiver-based node interference is NPC for the 2-dimensional case. However, the sender-based model this problem turned out to be solvable in polynomial time as shown in [2]. In this paper, we are concerned with total node interference. (Notice that this problem is equivalent to the average node interference problem.) We study the problem of assigning power to nodes in the plane to form a connected graph in which the total interference load is bounded. Specifically, we prove that the problem of assigning power to a set of nodes in the plane to yield a connected geometric graph whose total node interference is bounded is NP-complete. Our result is significant in view of the result in [10] where NP completeness was proved for the (general) ad hoc metric model only. Note that in our work as well as most of the works in WSNs, two nodes are connected by an edge if they are within the transmission range of one another. This definition is not strictly followed in a number of papers including [1] and [13]. We also simplify an O(logn) approximation algorithm reported in [10] for this problem. The performance of this algorithm is compared against a number of heuristics through simulation. The result shows that this algorithm outperforms others. The rest of this paper is organized as follows. Section II provides the definitions and explanations used in this paper. Section III is devoted to the NP completeness result, and Section IV discusses the O(logn) approximation algorithm as well as a number of heuristics. Section V contains some concluding remarks.

Minimum Total Node Interference in Wireless Sensor Networks

2 2.1

509

Preliminaries Network Model

Consider a set V of transceivers (nodes) in the plane. Each node u is assigned a power level denoted by p(u). The signal transmitted by node u can only be received by a node v if the distance between u and v, denoted by d(u, v), is ≤ p(u). We only consider the bidirectional case in which a communication edge (u, v) exists between two nodes u and v, if both power levels p(u) ≥ d(u, v) and p(v) ≥ d(u, v). Thus, the set V of nodes in the plane together with the power levels assigned to the nodes define a geometric (also known as intersection) graph G = (V, E). A geometric graph is said to be planar if no edge crosses another. 2.2

Interference Model

There are several definitions of the notion of interference in the literature. In this paper we consider two definitions of interference described in [2], [4] and [8]. Assuming that there is no obstacle blocking the broadcasting range, the two definitions of the interference load are as follows. Let D(u, ru ) be the broadcasting disk of node u with radius ru . For interference load, the definition from [2] is formally defined for the receiver and sender-based model as follow: For receiver based interference load: RE(x) := |{w ∈ V |w = x and x ∈ D(w, rw )}| For sender based interference load: SE(x) := |{w ∈ V |w = x and w ∈ D(x, rx )}| For the total interference load of a graph G = (V, E)   T N I(G(V, E)) := RE(x) = SE(x) x∈V

x∈V

In the interference load model based on Euclidean distance, the radius rv of node v is defined to be rv := max(v,w)∈E {d(v, w)}, whereas in the interference load model based on power usage, the radius rv of node v is defined to be rv := p(v).

Fig. 1. Gadgets used in representing a variable x in a PL1-EX-3SAT instance

510

3

N.X. Lam, T.N. Nguyen, and D.T. Huynh

The NP-Completeness of Minimum Total Node Interference in Geometric Graphs

In this section, we prove the NP-completeness of the problem of assigning power levels to nodes in the plane to produce a connected geometric graph G with bounded total node interference load T N I(G) based on Euclidean distance. As a corollary, the same problem for the power level based model is also NP-complete. MINIMUM TOTAL NODE INTERFERENCE IN CONNECTED GEOMETRIC GRAPHS Instance: Given a set of N nodes V = {v1 , v2 , ..., vN } in the plane, a set of M power levels P = {p1 , p2 , .., pM } that a node can transmit, and a positive number R. Question: Is there a power assignment to all nodes which induces a connected geometric graph G(V, E) such that T N I(G(V, E)) ≤ R? In the following we show that Minimum Total Node Interference in Geometric Graphs based on Euclidean distance is NP-complete. Theorem 1: Minimum Total Node Interference is NP-complete for connected geometric graphs. Proof. Minimum Total Node Interference for Connected Geometric Graphs (MTNICG) is obviously in NP. Given a set V of nodes in the plane, a set P of power levels and a positive integer R, we can nondeterministically assign power levels to the nodes, and verify in polynomial time that (1) the power assignment yields a connected geometric graph G(V, E), (2) the total interference of all nodes is ≤ R. To prove the NP-hardness of MTNICG, we construct a polynomial time reduction from the planar 1-Exact-3SAT problem (PL1-EX-3SAT) which was proven NP-complete in [6]. Consider an instance φ of PL1-EX-3SAT where each clause has exact 3 literals, and the planar instance graph G of φ, where G = (X ∪ C, E ∪ E  ) with edge sets E = {{x, c}|x ∈ C ∨ ¬x ∈ C} and E  = {{xi , xi+1 }|1 ≤ i ≤ n − 1}. φ satisfiable iff it has a Boolean assignment such that exactly 1 literal per clause has the value TRUE.

Fig. 2. A series of gadgets representing a variable x of degree 2 in PL1-EX-3SAT

Minimum Total Node Interference in Wireless Sensor Networks

511

Fig. 3. Nodes added on the line segments replacing two curved edges (left and right) and a straight edge (down) at variable x with degree 3

Fig. 4. Nodes added on the line segments replacing two curved edges (left and right), a straight edge (up) and a straight bridge (down) at variable x with degree 4

To construct an instance < V, P, R > of MTNICG, we first create a pair of gadgets for each variable x of φ. As shown in Fig. 2, each gadget for a variable x contains 14 nodes: 12 of these nodes form 6 pairs of x and ¬x, and the remaining two nodes are center nodes o and p. Every pair of variable and its negation node (x and ¬x) is connected by an edge called curved edge. There are 12 curved edges for each gadget. 6 pairs of a variable node and its negation (x,¬x) are connected by straight edges such that every node in a gadget is adjacent with exactly one straight edge. In one gadget, one set of nodes (all x nodes or all ¬x nodes) are connected to the center nodes (o and p) so that every center node is connected with exactly three x (or ¬x) nodes. The nodes in the second gadget are connected similarly. However, if nodes x are used to connect to center nodes

512

N.X. Lam, T.N. Nguyen, and D.T. Huynh

Fig. 5. Nodes added on the line segments replacing two curved edges (left and right), a straight edge (down) and a curved bridge (up) at variable node x with degree 4

Fig. 6. Nodes added on the line segments replacing two curved edges (left and right), a straight edge (down) and the edge (up) connecting to a clause node C2 at variable node x with degree 4

Minimum Total Node Interference in Wireless Sensor Networks

513

Fig. 7. Nodes added on the line segments replacing two curved edges (left and right), a straight edge (down) and a variable link (x,y) at variable node x with degree 4

Fig. 8. Nodes added on the line segments replacing two curved edges (left and right), a straight edge (x,¬x) and a straight edge (x,o) to center node o at variable node x with degree 4

514

N.X. Lam, T.N. Nguyen, and D.T. Huynh

in the first gadget, then nodes ¬x are used in the second gadget, or vice versa. Notice that each gadget has 6 nodes of degree 3 to connect to other gadgets and clause nodes. Inside each gadget, the two center nodes (o and p) are connected by an edge. To connect every two gadgets, two edges are used: a straight bridge and a curved bridge. These bridges connect degree 3 nodes of each gadget. Thus, each gadget uses up to 4 degree 3 nodes for bridges and has at least 2 degree 3 nodes to connect with clause nodes (see Fig. 2.) If the degree of node x is dx , then the number of x’s pair of gadgets in this chain is also dx (for a total of 2dx gadgets). Thus, the new graph (also denoted G for convenience) obtained from the original graph has a maximum degree of 4 while the planarity of the graph is still preserved. Next, we use Valiant’s result [15] to embed the graph G (with maximum degree 4) into the Euclidean plane: A planar graph with maximum degree 4 can be embedded in the plane using O(|V |) area in such a way that its vertices are at integer coordinates and its edges are drawn so that they are made up of line segments of form x = i or y = j, for integers i and j. Moreover, this embedding process can easily be designed to satisfy the additional requirement that each line segment drawn to connect two original vertices of the graph G must be of length at least 2 and every two parallel segments must be at least 2 units apart. Let’s call the units of the embedded graph G original units. Each original unit is divided further into 12 smaller “pieces” of equal length. Let the original unit be δ. We define the radii ri , 1 ≤ i ≤ 8 as follows: ri := (i/12) ∗ δ. For the sake of convenience let us call the variable, clause and center nodes of G embedded in the plane the variable, clause and center nodes, respectively. We can further modify each line segment by placing additional new nodes to create the instance < V , P, R > of MTNICG as follows: 1. On every line segments, add a node on every grid point. These nodes are called grid nodes. 2. On the line segments representing a connection of a variable node and a center node, or between two center nodes, add 3 nodes on each unit to divide it into 4 pieces of length r3 each. 3. On the line segments representing a curved edge or a curved bridge, add one node at distance r8 from the variable node on each of the two adjacent units. These interfacing nodes are called i nodes. (See nodes i s in Fig. 3, 4, 5, 6, 7,). Then add another node at distance r2 between the previously added i node and the grid node. On other unit segments in between, add 3 nodes to divide each unit into 4 equal pieces of length r3 each. 4. On the line segments representing a straight bridge, add 3 nodes to each unit to divide it into 4 equal pieces of length r3 each except the two end units adjacent with variable nodes. For these two units, add 5 nodes to divide it into 6 equal pieces of length r2 each (see Fig. 4.)

Minimum Total Node Interference in Wireless Sensor Networks

515

5. On the line segments representing a straight edge, add three nodes on each unit to divide it into 4 equal pieces of length r3 each, except the unit that is adjacent with a variable node. For this unit, there are four cases: – If this unit is adjacent with a degree 3 variable node, add 5 nodes on the unit to divide it into 6 pieces of length r2 each (see Fig. 3). – If this unit is adjacent with a degree 4 variable node having an adjacent edge which is a curved bridge or a link to a clause node, add 2 pairs of nodes. The first pair is added at the distance of r3 from the variable node and from each other. The second pair is added at the distance of r2 from the grid node and from each other (see down units in Fig. 5 and 6.) – If this unit is adjacent with a degree 4 variable node having a straight bridge edge, add 5 nodes on the unit to divide it into 6 equal pieces of length r2 each. (See “up” unit in Fig. 4.) – If this unit is adjacent with a degree 4 variable node different than those in the previous cases, add 3 nodes to divide this unit into 4 equal pieces of length r3 each. 6. On the line segments representing a link from a variable node to a clause node, add one node at distance r8 to the variable node denoted n on the unit adjacent with the variable node. (See “up” unit in Fig. 6.) Add a second node at distance r2 to the previously added node (n) and the grid node on the same unit. On other unit segments leading to the clause node, add 3 nodes on each segment to divide each unit into 4 equal pieces of length r3 each. 7. On all other line segments, add 3 nodes to each unit to divide it into 4 equal pieces of length r3 each. (Figures 3, 4, 5, 6, and 7 show variable nodes x have power level r8 to connect with nodes i’s while its negation nodes ¬x have power level r3 and are not connected with its neighboring i nodes.) Before defining the interference bound R, we note there are 5 categories of nodes in < V  , P, R >: the set of variable nodes also denoted X, where each x has degree dx , the set M of center nodes o s and p’s, the set I of i nodes, the set C of clause nodes, and the set T of remaining nodes. For each set of nodes we define a corresponding interference quantity as follows. For a set U of nodes  let SE(U ) := u∈U SE(u). For a variable x ∈ X of degree dx , there are dx pairs of gadgets, and each gadget has 12 variable nodes. Each of exactly 11dx variable nodes is allowed to have the interference load of 6 when assigned the power level r8 corresponding to TRUE (see Fig. 3, 5, 6); and each of the other dx variable nodes is allowed to have the interference load of 10 (Fig. 4) to connect through a straight bridge. Each of the other 8 and 4 variable nodes (when assigned FALSE corresponding to power level r3 ) is allowed to have the interference load of 2 (Fig. 7, 8) and 1 (Fig. 3, 5, 6), respectively: SE(X) :=

 x∈X

(6 ∗ 11dx + 10 ∗ dx + 2 ∗ 8dx + 4 ∗ dx ) =

 x∈X

96 ∗ dx

516

N.X. Lam, T.N. Nguyen, and D.T. Huynh

For the center nodes in M we define:   (4 ∗ 4dx ) = 16 ∗ dx SE(M ) := x∈X

x∈X

This allows each center node to have an interference load of 4 to connect to its 4 neighbors. (Note that there are 2dx nodes os and 2dx nodes ps for a variable node x of degree dx .) The interference quantity for nodes is is so defined to accommodate the interference load of the variable nodes x: From the 24dx curved edges, 24dx (half) of nodes is are allowed to have the interference load of 4 each, while the other 24dx (half) of nodes i s to have the interference load of 1. From the 2dx − 1 curved bridges, 2dx − 1 (half) of nodes is are allowed to have interference load of 4 while the other half have interference load of 1.   ((4 ∗ 24 + 1 ∗ 24)dx + (2 ∗ dx − 1) ∗ 5) = (130 ∗ dv − 5) SE(I) := x∈X

v∈V

For the clause nodes c ∈ C, the interference quantity is defined as follows. Each clause node has degree 3 and is allowed to cover its 3 neighbors. Each clause node connects to its literal via the 3 nodes ns: exactly one of these 3 nodes is allowed to cover 4 neighbors, the other 2 are allowed to cover only 1 neighbor each. Thus,   SE(C) := (4 ∗ c + 3 ∗ c + 2 ∗ c) = 9∗c c∈C

c∈C

For the set T of remaining nodes which is the set of all nodes minus nodes i s, variable nodes, center nodes o s and p s, nodes n s as well as nodes c s, we define the interference quantity SE(T ) as SE(T ) := 2|T |. Let P := {0, r1 , r2 , r3 , ..., r8 } and R := SE(X) + SE(O) + SE(M ) + SE(I) + SE(C)+SE(T ). The correctness of the above polynomial-time reduction follows from the following claim: Claim. The instance φ of PL1-EX-3SAT is satisfiable if and only if the instance < V  , P, R > of MTNICG has a power assignment that yields a connected geometric graph G (V  , E  ) such that the total interference load T N I(G (V  , E  )) ≤ R. The proof of the Claim can be found in the Appendix. From the Claim, Theorem 1 follows.

4 4.1

Heuristics and Their Performance Heuristics

In this section we present three simple heuristics and an approximation algorithm which is a simplified version of a greedy algorithm reported in [10].

Minimum Total Node Interference in Wireless Sensor Networks

517

Node interference approach: Our first approach is built on the sender-based node interference problem, where the interference load of a node v is the number of nodes within the broadcasting disk of v and the problem is to minimize the maximum node interference. This problem has been shown to be solved in polynomial time in [2] through a simple greedy heuristic called OPT-MINMAXSIP. Based on this result, we propose the Node Power Level Search heuristic (NPLS) which first computes the minimum sender-based node interference, and then reduce the power level for each node while preserving connectivity. Although this step does not necessarily decrease the maximum node interference of the graph, it may improve the interference load of other nodes, and hence the total node interference load. The pseudo code of this algorithm is depicted in Figure 9. Node-Power-Level-Search(V ) Input: a set V of nodes in the plane Output: A power assignment P that yields a connected geometric graph whose total node interference is minimized 1 Using the greedy technique, find a power assignment P that yields a minimum node interference graph // Reduce power levels of nodes // while maintaining connectivity 2 for each node v ∈ V 3 Using binary search find the least power level for P [v] while P yields a connected graph 4 return the power assignment P Fig. 9. Node-Power-Level-Search

Minimum spanning tree approach: The basic idea of this approach is to find a minimum spanning tree for a given complete graph. We use this approach in two different algorithms, the distance-based minimum spanning tree (DMST) algorithm and the interference-based minimum spanning tree (IMST) algorithm. The main difference between these two algorithms is the definition of the edge weight. The former uses the Euclidean distance as edge weight while the latter uses the DEI edge interference load. In implementing these two heuristics, we use Kruskals algorithm [5] to find the minimum spanning tree. This algorithm uses two well-known subroutines find-set and union. find-set(u) is to find the root of tree-based connected component containing node u whereas union(u, v) is to connect two connected components of u and v. The details of these two subroutines can be found in [5]. Figure 10 contain the pseudo code of the IMST algorithm. The pseudo code of the DMST algorithm is similar. Greedy approach: In [10], the authors suggest a greedy algorithm for the ad hoc interference model with O(log n) approximation ratio. This algorithm selects the most cost-efficient star at each iteration to join some current connected

518

N.X. Lam, T.N. Nguyen, and D.T. Huynh

Inference-based-Minimum-Spanning-Tree(V ) Input: a set V of nodes in the plane Output: A power assignment P that yields a connected geometric graph whose total node interference is minimized 1 E = set of edges between all nodes in V 2 A=∅ 3 w(e) = interference of e for all e ∈ E 4 Sort edges of E into non-decreasing order by w 5 for each edge (u, v) ∈ E 6 if find-set(u) = find-set(v) 7 add edge (u, v) to A 8 union(u, v) 9 B = edges  created by adding (u, v) 10 A= A B 11 E = E\B 12 Recalculate w and sort edges in E 13 Compute the power assignment P from A 14 return P Fig. 10. Interference-based-Minimum-Spanning-Tree

components by increasing the transmission range of each node in this star until a connected graph results. This algorithm does not seem to work as claimed in [10] since there is a discrepancy between the algorithm and its proof. In the proof, the authors consider the size of each star as the number of connected components to be connected. However, the suggested method to find the best star does not guarantee that this requirement is satisfied. In the following, we provide a simpler version of the greedy algorithm in [10]. This simplified algorithm is in polynomial time and also has an O(log n) approximation ratio. The idea of our modification is based on Kruskal’s minimum spanning tree approach. The algorithm proceeds in a greedy fashion and selects in each step a proper pair of nodes to connect two connected components by increasing the transmission ranges of these nodes. The algorithm will connect pairs of nodes until there is only one connected component in the resulting graph. The major issue is the greedy property used to determine the proper pair of nodes at each step. First, considering an edge e(u, v) we defined the induced power assignment of this edge as follows. ⎧ ⎨ fu (v) if a = u P e (a) = fv (u) if a = v ⎩ 0 otherwise where fu (v) denotes the smallest power level such that v is covered by D(u, fu (v)). Let tak denote the number of nodes covered by D(a, k). Given a graph Gi

Minimum Total Node Interference in Wireless Sensor Networks

519

Greedy(V ) Input: a set V of nodes in the plane Output: A power assignment P that yields a connected geometric graph whose total node interference is minimized 1 for each a ∈ V P0 (a) = 0 2 i=0 3 while G induced from Pi is not connected 4 E = set of edges connecting two different connected components 5 e(u, v) = argmine∈E (cost(e)/|CPi (e)|) 6 for all a ∈ V Pi+1 (a) = max(Pi (a), P e(u,v) (a)) 7 i = i+1 8 return Pi Fig. 11. Greedy Algorithm

induced from a power assignment Pi at iteration i, we define the cost of an edge e(u, v) as cost(e) = cost(u) + cost(v), where taP e (a) if Pe (a) > Pi (a) cost(a) = 0 if Pe (a) ≤ Pi (a) Intuitively, the cost of an edge measures the amount of additional interference created by connecting the nodes of this edge based on the current power assignment Pi . In addition, while connecting edge e(u, v), by increasing the transmission range of u and v, we might also connect u (or v) with other nodes, which are covering u (or v), if these nodes will be covered with the new transmission range of u (or v). Let CP (e) be the set of connected components which will be connected by selecting edge e based on the current power assignment P . At each iteration i of the algorithm, we will select such an edge ei such that cost( ei )/|CPi−1 ( ei )| is minimized. Figure 11 gives a pseudo-code of this algorithm. Lemma 2: Let P be the optimal solution and λi denote the number of connected components in the graph induced by Pi . Then for every iteration of this ei ).λi−1 algorithm, it holds that cost( ≤ T N I(P ), where ei is the edge selected by |CPi−1 ( ei )| the algorithm at the i-th iteration. Proof. The proof of this lemma is similar to the proof in [10], which makes use of an integer linear program formulation of the interference problem, and is therefore omitted. The O(log n) approximation ratio follows from the following theorem whose proof can be found in the Appendix. Theorem 3: Let P Greedy denote the solution obtained from the greedy algorithm. It holds that T N I(P Greedy ) ≤ T N I(P).O(log n).

520

N.X. Lam, T.N. Nguyen, and D.T. Huynh

DMST DMST IMST Greedy NPLS

IMST Greedy NPLS 39% 34% 100% 70% 54% 100% 82% 71% 100% 0% 0% 0%

(a) Performance comparison of four heuristics

IMST NPLS DMST Aver 0.39% 144.87% 1.71% Min 3.76% 61.29% -4.03% Max 5.04% 326.51% 11.57%

(b) Average performance of Greedy Algorithm vs. others

Fig. 12. Experimental Results

4.2

Experimental Results

In our experiments, we only consider geometric graphs. We randomly generate 50 nodes on an area of size of 1000x1000. For each graph, we run all four algorithms. We compare the performance of one algorithm against another. The results are shown in Figure 12(a). As seen in this table, the Greedy algorithm performs better than the others. In fact, in 82 cases (71 cases and 100 cases) out of all 100 test cases, its results are equal or better than the results obtained by DMST (IMST and NPLS, respectively). Moreover, Figure 12(b) shows that the results of the Greedy algorithm is on average 1.83% (3.18% and 148.39%) better than the results obtained by IMST (DMST and NPLS, respectively).

5

Conclusions

In this paper we have studied the total node interference problem in WSNs. We have shown that assigning power levels to a set of nodes in the plane to yield a geometric graph with bounded total node interference is NP-complete, complementing a result in [10] by T. Moscibroda and R. Wattenhofer who proved that the problem is NP-complete for the more general ad hoc metric model. We have also provided a simplified version of their O(logn) approximation algorithm and compared its performance against other heuristics through simulation. The result shows that this algorithm is better than others. An interesting future research topic is how to construct network topologies that have low interference and are fault-tolerant at the same time. Results obtained along this line will be of practical relevance.

References 1. Benkert, M., Gudmundsson, J., Haverkort, H., Wolff, A.: Constructing Interferenceˇ Minimal Networks. In: Wiedermann, J., Tel, G., Pokorn´ y, J., Bielikov´ a, M., Stuller, J. (eds.) SOFSEM 2006. LNCS, vol. 3831, pp. 166–176. Springer, Heidelberg (2006) 2. Bil` o, D., Proietti, G.: On the Complexity of Minimizing Interference in Ad-Hoc and Sensor Networks. Theorectical Computer Science 402, 43–55 (2008)

Minimum Total Node Interference in Wireless Sensor Networks

521

3. Buchin, K.: Minimizing the Maximum Interference is Hard (2008), http://arxiv.org/abs/0802.2134 4. Burkhart, M., von Rickenbach, P., Wattenhofer, R., Zollinger, A.: Does Topology Control Reduce Interference? In: MOBIHOC 2004, pp. 9–19 (2004) 5. Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C.: Introduction to Algorithms. Section 23.2: The algorithms of Kruskal and Prim, pp. 567–574. Press and McGrawHill (2001) ISBN 0-262-03293-7 6. Dyer, M., Frieze, A.: Planar 2DM is NP-complete. Journal of Algorithms (7), 174– 184 (1986) 7. Halld´ orsson, M.M., Tokuyama, T.: Minimizing Interference of a Wireless Ad-Hoc Network in a Plane. In: Nikoletseas, S.E., Rolim, J.D.P. (eds.) ALGOSENSORS 2006. LNCS, vol. 4240, pp. 71–82. Springer, Heidelberg (2006) 8. Johansson, T., Carr-Motyckova´ a, L.: Reducing Interference in Ad Hoc Networks Through Topology Control. In: DIALM-POMC 2005, pp. 17–23 (2005) 9. Moaveni-Nejad, K., Li, X.-Y.: Low-Interference Topology Control for Wireless Ad Hoc Networks. In: Ad Hoc and Wireless Sensor Networks, vol. 1, pp. 41–64 (2005) 10. Moscibroda, T., Wattenhofer, R.: Minimizing Interference in Ad Hoc and Sensor Networks. In: DIALM-POMC 2005 (2005) 11. Nguyen, T.N., Huynh, D.T.: Minimum Interference Planar Geometric Topology in Wireless Sensor Networks. In: Liu, B., Bestavros, A., Du, D.-Z., Wang, J. (eds.) Wireless Algorithms, Systems, and Applications. LNCS, vol. 5682, pp. 149–158. Springer, Heidelberg (2009) 12. Rickenbach, P.V., Schmid, S., Wattenhofer, R., Zollinger, A.: A Roburst Interference Model for Wireless Ad-Hoc Networks. In: Proc. 19th IEEE Int. Par. and Dist. (2005) 13. Sharma, A., Thakral, N., Udgata, S., Pujari, A.: Heuristics for Minimizing Interference in Sensor Networks. In: Garg, V., Wattenhofer, R., Kothapalli, K. (eds.) ICDCN 2009. LNCS, vol. 5408, pp. 49–54. Springer, Heidelberg (2009) 14. Nguyen, T., Lam, N., Huynh, D.: Minimum Edge Interference in Wireless Sensor Networks. To appear in Proc. of Intern. Conf. on Wireless Algorithms, Systems and Applications (2010) 15. Valiant, L.: Universality Considerations in VLSI Circuits. IEEE Trans. on Compupters C-30, 135–140 (1981) 16. Wu, K.-D., Liao, W.: On Constructing Low Interference Topology in Multihop Wireless Networks. Int. J. of Sensor Networks 2, 321–330 (2007)

APPENDIX Proof of Claim in Theorem 1 For the “only-if” direction, suppose that φ has a satisfying Boolean assignment. We assign power levels to nodes in V as follows: 1. Assign power level r8 to variable node x ∈ V  if variable x has value T RU E; otherwise, assign power level r3 . 2. Assign power level r8 to nodes i s that are neighbors of variable nodes that were assigned power r8 . Other nodes i s are assigned power level r3 . 3. Assign power level r8 to all nodes n s that are neighbors of the variable nodes that were assigned power level r8 . 4. Assign the power level r3 to all remaining nodes.

522

N.X. Lam, T.N. Nguyen, and D.T. Huynh

Figures 3, 4, 5, 6, 7 and 8 show the variable node x with power level r8 and ¬x with power level r3 . The total interference load of the graph is calculated based on the interference load of the sets of nodes X, O, M, I, C and T . First, note that half of the total variable nodes have power level r8 . If such a variable node is a degree 3 or degree 4 node, it covers 6 neighbors except for the one that connects with a straight bridge. This node covers 10 neighbors. Next, observe that half of nodes i s have power level r8 to connect to the variable nodes with power level r8 . Each of these nodes i s cover 4 nodes while the other half of nodes i s only cover 1 neighbor each. All center nodes o and p cover 4 neighbors. Only 1/3 of nodes n s on the segment connecting to a clause node cover 3 nodes, while the other 2/3 of nodes n s cover only 1 neighbor. Each clause node cover 3 neighbors and the rest of nodes cover 2 neighbors each. It is straightforward to verify the total of node interference load is exactly R and the graph is a connected graph. For the “If” direction, suppose the instance < V  , P, R > has a power as   signment that yields  a connected geometric graph G (V , E ) of which the total interference load v∈V SE(v) ≤ R = SE(X) + SE(O) + SE(M ) + SE(I) + SE(C) + SE(T ), we construct a satisfying Boolean assignment for the PL1-EX3SAT instance φ based on the following observations: 1. Exactly 12 nodes i s in each gadget and 1 node i of each curved bridge must connect to 6 variable nodes for the graph to be connected and the total interference load in each gadget to be minimum. 2. Exactly 6 variable nodes of one of the two gadgets (all x variable nodes or all (¬x) variable nodes) must connect to nodes i s for the graph to be connected. If any variable node from the other set also connects to a node is, then the total interference load is > R, a contradiction. 3. The curved bridge between two gadgets requires at least one of the variable nodes at both ends to have power level r8 for connection. 4. The straight bridge connecting two gadgets allows only one variable node of either end to have power level r8 . Otherwise, if both variable nodes at both ends have power level r8 , then the total interference load is > R. 5. At least one node n has power level r8 to connect to a variable node for the graph to be connected. Furthermore, if there are more than 1 node n’s of a clause node connected to variable nodes, then the total interference load is > R. From the above observations, we can construct a Boolean assignment for the PL1-EX-3SAT instance φ using the following rules: – If a variable node in G has the power level r8 , then assign the value T RU E to that variable in φ; otherwise, assign value F ALSE. From Observation 5 it follows that each clause in φ is satisfied by exactly 1 literal having value TRUE. Moreover, from Observations 1, 2, 3 and 4 the Boolean assignment for φ is consistent. This concludes the proof of Theorem 1.

Minimum Total Node Interference in Wireless Sensor Networks

523

Proof of Theorem 3 The proof is similar to the one in [10], and the details are as follows. By the definition of CPi (e), the number of connected components λi can be computed by λi

= Lemma2

≤ ≤

λi−1 − |CPi−1 ( ei )| + 1 cost( ei ).λi−1 λi−1 − +1 T N I(P) λi−1 .(1 −

cost( ei )

2.T N I(P)

)

Let q be the number of iterations required by the Greedy algorithm. Since λ0 = |V | and λq = 1, it follows that 1 = λq ≤ |V |

q

(1 −

i=1

cost( ei )

2.T N I(P )

)

By taking the logarithm on both sides and rearranging the inequality, we obtain the following. ln 1 ≤ ln |V | +

q 

ln(1 −

i=1 q

≤ ln |V | −

cost( ei )

2.T N I(P )

)

 cost( ei ) 2.T N I(P ) i=1

q 

cost( ei ) ≤ 2.T N I(P) ln |V |

i=1

Since whenever the transmission range of a node is increased, cost( ei )) accounts Greedy for the entire cost of this increase. Thus, it holds that T N I(P ) is at most q cost( e ). The theorem therefore follows. i i=1

Reproducing Consistent Wireless Protocol Performance across Environments Taewoo Kwon1 , Emre Ertin2 , and Anish Arora1 1

The Ohio State University Columbus, OH 43210, USA {kwonta,anish}@cse.ohio-state.edu 2 [email protected]

Abstract. Full scale experimentation with wireless networks in deployment environments is difficult, so a common validation technique is to test a prototype network in a convenient environment prior to deployment. In this paper, we consider the problem of obtaining comparable protocol performance when the test and deployment environments differ in RF propagation environment and/or inter-node spacing. To achieve comparable protocol behavior in the two settings, we propose the concept of “link usage spectrum”. Based on the hypothesis that the link usage spectrum is a gross predictor for network performance, we show how to replicate in the test setting the link usage spectrum of the protocol that is expected in the deployment setting. We illustrate our technique for achieving comparable protocol behavior via experiments and simulations in multiple indoor and outdoor propagation environments. The link usage spectrum is protocol specific; we illustrate for a family of protocols how the link usage spectrum is calculated analytically, from the protocol metric for choosing forwarding links in the network, and how power scaling can be used to match the link usage spectrum across networks. Keywords: wireless sensor network, testbed design, wireless protocol performance.

1

Introduction

Experiences in deploying low-power wireless networks during this decade have yielded a number of surprises, wherein network behavior in the field diverged substantially from that seen in laboratory tests. A combination of factors has contributed to these surprises. One key factor is that the effective topology of the laboratory tests is different from that of the field deployment: Not only is inter-node spatial (separation) scale different in the two networks, but the environment signal propagation characteristics also tend to be different, and as a result the link selections and the intra-node traffic interference diverge. Differences in externally induced communication interference are another factor. Other scale differences in the field deployment, i.e., increasing the number of the nodes fielded, and consequent phase transition or instability issues are yet another factor. Moreover, network protocol behaviors can themselves exhibit J. Zheng, D. Simplot-Ryl, V.C.M. Leung (Eds.): ADHOCNETS 2010, LNICST 49, pp. 524–540, 2010. c Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2010 

Reproducing Performance

525

nontrivial variability, and this variability may only be inadequately understood in the testing phase. The multi-faceted difficulty with ensuring desired protocol behavior in the field, coupled with the high cost of testing and tuning the performance in the field, motivates the scientific study of tools and techniques for reproducing network behavior across test and deployment environments. What do we mean by reproducing performance? Even if the test and deployment environments are the same and we only displace the network in space, the realized network performance will not be identical. One can only achieve a probabilistic equivalence between two such networks. By probabilistic equivalence, we informally mean that the set of links exercised by the two networks are sampled from the same probability distribution. For two networks at different spatial scales in the same environment, positive results for achieving probabilistic equivalence have been presented earlier, i.e., by using transmission power control [1]. (Power control was realized via attenuator hardware and/or software control.) That study also experimentally compared the performance metrics of two wireless sensor network (WSN) protocols — Sprinkler [2], a protocol that provides a bulk data transmission service, and LOF [3], a protocol that provides a beacon-free routing service— at different spatial scales in Kansei, an indoor WSN testbed [4], [5], to illustrate how to select the transmission power to achieve probabilistic equivalent behavior when scaling all inter-node distances by some constant. In contrast, for two networks at different spatial scales in different environments —in particular, with different path loss exponents— it is straightforward to show that it is impossible to achieve probabilistic equivalence using only transmission power control. This necessitates consideration of alternative techniques. Link Usage Spectrum. In this paper, towards achieving comparable performance in networks in potentially different environments, we adopt the concept of realizing the same (or measurably close to the same) “link usage spectrum” in the networks. Informally speaking, the link usage spectrum of a network is the probability distribution with which the network protocol selects links of different length from among all the available links in the network at hand. The hypothesis of this paper is that the link usage spectrum is a gross predictor of the performance of (a rich class of) network protocols. With this hypothesis, a network protocol will perform comparably in two network settings if the respective link usage spectrums of the protocol match closely in these settings. Said another way, even if the “available” link spectrum in the settings is different but the probability distribution of the chosen links is comparable, the protocol behavior in the settings will be comparable. The link usage spectrum can therefore be used to achieve predictable network behavior in the deployment setting, as follows. Consider the case where a prototype network is tested somewhere, say in an indoor environment, before it is fielded elsewhere, say in an outdoor environment, with potentially different inter-node spacing. Since the indoor environment is persistent and easily instrumented for tests, it is relatively easy to collect fine-grain, long running protocol

526

T. Kwon, E. Ertin, and A. Arora

behavior information in it. Thus, if one could easily calculate the link usage spectrum data for the outdoor environment (by analysis, simulation, or experiment), then one could (analytically or experimentally) design the link usage spectrum for the indoor tests (say by choosing the indoor network spatial scaling facto and transmission power level) to be close to that of the field network. The resulting observed indoor protocol behavior would be predictive of the behavior to be observed in the field. There are two major factors that affect the link usage spectrum: the metric chosen by the network protocol for selecting parents to who nodes forward their traffic and, more generically, the signal to noise ratio (SNR), or RSSI, values of links. It is often the case that the chosen metric itself involves a function of the SNR (or RSSI) as well as the distance traversed by the link. In these cases, the link usage spectrum can be reformulated as a function of relative preference based on SNR and the forwarding distance. We emphasize that this is only one of the ways to formulate the link usage spectrum, and that the analysis we perform subsequently in the paper is readily adapted to several other routing metrics. Related Work. There has been little previous work on the link usage spectrum. There is significant literature however, e.g. [6], that models the bit error rate for radio channels and thus calculates performance metrics such as signal to noise ratio (SNR), packet reliability rate (PRR), expected number of transmissions (ETX), PRR × d (the forwarding distance) [7], and expected latency per unit 1 distance (ELD= ) [3]. A related work that implicitly exploits the link PRR×d usage spectrum idea is [7], although its role is different: the spectrum is used as a tool for calculating average network metrics that are in turn used for choosing between protocols and optimizing a protocol realization with respect to its intended forwarding metric. [7] also uses numerical simulations for calculating the spectrum; by way of contrast, we provide a closed form equation for expressing the link usage spectrum in the context of the forwarding metric at hand. Contributions. Our primary contribution includes evidence in support of the hypothesis that link usage spectrum is a gross predictor of network performance. Our evidence consists of experimental results that confirm that the network performance in different settings is most similar when the l1 distance between the corresponding link usage spectrums is minimum. Specifically, these involve multiple indoor and outdoor experiments with the Collection Tree Protocol (CTP) [8], which is a popular messaging protocol distributed with the TinyOS 2.0 release. By the same token, i.e., by minimizing the l1 distance between the two settings, we show a general technique for achieving comparable performance of a network protocol in test and deployment settings. A third contribution is to show how the link usage spectrum is analytically derived for network protocols. Specifically, we consider the case of protocols whose forwarding metric depends on Packet Reliability Rate (PRR), distance, and other variables based on SNR; this case spans a large fraction of the route selection protocols in use today. Further, we show that the analytically obtained

Reproducing Performance

527

results are corroborated by experimental measurements and simulation of the link usage spectrum in such contexts. Roadmap. In Section 2, we define link usage spectrum and formulate our method for achieving comparable performance based on the XPlantError metric. In Section 3, we provide an experimental study of CTP protocol transplantation in the context of a simple chain topology in multiple indoor and outdoor environments. In Section 4, we show how to analytically calculate the link usage spectrum for a class of network protocols, and validate the analysis via simulation based results in 1-dimensional and 2-dimensional networks. Finally, we conclude in Section 5, with a discussion of variations of the definition of XPlantError for achieving potentially higher fidelity in predicting network behavior.

2

Link Usage Spectrum and Network Transplant Error

Wireless network behavior is substantially influenced by the performance of the wireless links between the nodes of the network. Performance of a wireless link between a transmitter and its receivers is determined by the RF channel between the terminals (environment model) and the bit-error-performance of their wireless transceivers (radio model). RF channel models describe the probabilistic relation between link distance and path loss. Specifically, in any given network, links that have the same length experience different channel realizations, due to spatial variations in obstructions and reflectors in the scene. As a result, the received signal strength experienced on links of length d is a random variable R(d). The RF channel model induces a distribution on R(d). For instance, the log-normal shadowing model, a large scale fading model employed commonly in indoor and outdoor link studies, describes the received signal strength as: R(d) = Pt − P L(d0 ) − 10η log(d/d0 ) + Nσ

(1)

where η is the path loss exponent, Pt is the transmitter power, and P L(d0 ) is the path loss observed at distance d0 in dB and Nσ is a zero-mean Gaussian random variable with standard deviation σ, representing spatial variations in the RF environment. The received signal to noise ratio (SNR) at the receiver y(d) is given by the received signal power R(d) reduced by the noise power P0 : y(d) = R(d) − P0 (in dB)

(2)

The radio receiver performance can be characterized by representing the packet reception rate PRR(y) as a function of the received the receiver SNR, y. PRR(y) gives the probability that a packet received with SNR of y will be decoded correctly by the receiver. The relation between packet-reception-rate and SNR depends on the modulation scheme and the packet encoding scheme employed by the radio transceiver. The function PRR(y) is a monotonically increasing function with range [0, 1] and acts as a soft limiter. The combination of the environment and radio model completely describes the link properties observed in a wireless network for low-rate/time division access

528

T. Kwon, E. Ertin, and A. Arora

where the interference is not significant. Experimental [9] and analytical [6] studies of low power wireless links have shown that in such settings a high percentage of network links will be either good or bad, > 90% and < 10% respectively PRR. 1 In:C (tx=-13dBm) Out:C (tx=-6dBm) Out:C (tx=-3dBm) Out:C (tx=-1dBm) Out:C (tx=0dBm) In:A (tx=-13dBm) Out:A (tx=-6dBm) Out:A (tx=-3dBm) Out:A (tx=-1dBm) Out:A (tx=0dBm)

Packet Reception Rate

0.8

0.6

0.4

0.2

0 0

5

10

15 Distance (3ft)

20

25

30

Fig. 1. PRR distribution of links for indoor and outdoor propagation environments at various transmit power levels (C:Chosen, A:All)

We note that the link reliability statistics reported by previous studies are based on the a priori distribution of the link realizations. If we consider the posterior distribution of the links that are selected by a given network protocol, the distribution will be skewed towards high PRR values. Figure 1 shows the expected PRR of links of various length based on whether or not they were chosen by the forwarding protocol. We observe that the expected PRR of the links chosen by the protocol is uniformly high and markedly different from the expected PRR of all links at a given distance, especially in the case of long links. As a result, network forwarding performance is grossly determined by the link lengths that are being utilized and less so by the small variations in link qualities. Therefore, in this paper, we focus on a particular network statistic called the link usage spectrum, which captures the probability distribution with which a given network protocol selects links of different length from among all available links in the network at hand. Example 0. To illustrate the definition of link usage spectrum, consider a wireless network W = ({lj }, η, σ) with link set {lj }M j=1 and the RF environment (η, σ) employing a network protocol P. We note W is a probabilistic object, referring to the ensemble of link set realizations. For each realization of the wireless network W, network protocol, P chooses a subset of the link set for forwarding of data. For a one dimensional linear networks with uniform node spacing, where the link lengths dj ≡ d(lj ) are constrained to the finite set {τ, 2τ, 3τ, . . . , N τ }, where τ is the minimum node spacing, the link usage spectrum L(W, P, i) is the discrete

Reproducing Performance

529

probability distribution over the length of links induced by the network protocol P. L(W, P, i) = Prob[d(l) = iτ ] (3) Example 1. We extend the previous example from a one to a two dimensional two dimensional grid network with uniform node spacing. In this case, the link lengths dj ≡ d(lj ) and the progress to destination pj ≡ p(lj ) are constrained to the finite set {(τ, τ), (τ, 2τ ), (τ, 3τ ), . . . , (τ, N τ ), (2τ, τ ), (2τ, 2τ ), (2τ, 3τ ), . . . , (2τ, N τ ), . . . , (N τ, τ), (N τ, 2τ ), (N τ, 3τ ), . . . , (N τ, N τ )}, where τ is the minimum node spacing. We index elements of the above finite set from 1 to M. The link usage spectrum L(W, P, i), then, is the discrete probability distribution over the length of links and the progress of links to the destination induced by the network protocol P. L(W, P, i) = Prob[d(l) = di , p(l) = pi ], (4)  n −1 where i is the index of (mτ, nτ ), di = (mτ )2 + (nτ )2 , pi = di ∗cos(|tan ( m )− θ|), θ is the angle between source and destination. Note that the link usage spectrum is distinct from the connectivity graph. The links used by the forwarding protocol will be in general only a small subset of the connectivity graph found by maximizing the routing metric over the set of all valid links. The fundamental importance of link usage spectrum stems from the fact that many network wide metrics can be calculated as averages over link realizations weighted by the usage spectrum, as discussed in Section 5. As a result, we propose to use the link usage spectrum to match protocol behavior across scales and environments. 2.1

Comparing Protocol Performance in Two Settings

Next, we propose a procedure transplanting a network protocol to a different RF environment and/or a different inter-node separation scale. The basic idea is to minimize the distance between the link usage spectrums of the two networks; one way of realizing this is by optimizing the selection of the transmit power. Definition 1. Consider a wireless network W with inter-node distances {dj }m j=1 ˜ with inter-node distances {d˜j = αdj }m in RF enviand its scaled version W, j=1 ronments characterized by log-normal scale model parameters (n, σ) and (˜ n, σ ˜) respectively. We define the Transplant Error of a protocol P across the two netn  works as:    ˜ ˜ , P, i) XPlantError(W, W, P) = L(W, P, i) − L(W i=1

XPlantError is essentially the l1 distance between the link usage spectrums for ˜ which differ in scale and RF environment. Theorems the two networks W and W 1 and 2 show the relation between the link usage spectrum and performance metric. If we can control the XPlantError within a threshold value, we conjecture that the protocol performance will be similar across scale and environment. We ˜ is variable through P˜t = P˜0 + β assume that transmit power in the network W where β is the power attenuation or amplification in the scaled network. Since

530

T. Kwon, E. Ertin, and A. Arora

the scaled vector in general reduces the node distances for convenient testing, β in general is a negative value indicating power attenuation. As a result the ˜ scaled network realizations W(β) depends on β 1 . The optimal power attenuation is then chosen to minimize the XPlantError metric: ˜ βopt = arg min XPlantError(W, W(β), P) β

(5)

As shown in [1], if the two networks are in the same environment (i.e., η = η˜ and σ = σ ˜ )) then β can be chosen such that the link SNR realizations yj and y˜j are samples from the same multivariate Gaussian probability distribution. As a corollary, the optimal β in that case would result in identical link usage spectrums. For networks in different RF environments, the distribution of link SNR realizations cannot be matched for all link lengths simultaneously and alternative techniques as shown above are required to achieve comparable network behavior. In the next sections, we validate the proposed measure of similarity through experimental and simulation studies using linear and grid networks of wireless nodes employing 802.15.4 radios.

3

Experimental Study of Protocol Transplantation

In this section, we present an experimental study of reproducing protocol performance across different WSN environments. We focus our experiments on a well known tree-based convergecast protocol, called the Collection Tree Protocol (CTP) [8]. In the following, we first give a detailed description of our experimental setup and physical, link and messaging layer assumptions. Then, we study a scaling exercise with a linear network topology in multiple indoor and outdoor settings using data from field experiments. We show that comparable protocol performance is achieved for various metrics (end-to-end delay, mean hop length) if the transmit powers are chosen to minimize the distance between link usage spectrums for test and deployment environments. 3.1

Experimental Setup

We set up a one dimensional linear topology with a total of 20 TelosB [10] sensor nodes. Each node is separated by 3 ft and elevated about 4 inches from the ground. The TelosB mote platform is equipped with a CC2420 [11] radio and provides eight different transmission power levels: 31(0dBm), 27(-1dBm), 23(-3dBm), 19(-5dBm), 15 (-7dBm), 11(-10dBm), 7(-15dBm), 3(-25dBm) [10]. Using a -3dB attenuator in conjunction with software power level settings, 15 distinct transmit power levels can be realized. We performed experiments in three settings, to capture different environment properties. We used a single output level of -13dBm for an indoor network in 1

We could also introduce spatial variations in β across the nodes to influence the width of the resulting spectrum, for a better match.

Reproducing Performance

531

a long office corridor; three output levels (-6 dBm, -3 dBm and 0 dBm) for an outdoor network in a parking lot, corresponding to power scaling coefficient β of (7dB, 10dB, 13dB); and three output levels (-18 dBm, -13 dBm and -10 dBm) for an indoor network in an open warehouse. 3.2

Physical and Link Layer

The CC2420 [11] radio is compatible with the 2.4GHz 802.15.4 standard. The 802.15.4 physical layer employs block direct-sequence spread spectrum code with 2MChip/s chip rate and 250 kbps data rate to achieve processing and coding gain. The transmitter modulates the carrier using offset quadrature phase shift keying (O-QPSK) with half-sine shaping which is equivalent to minimum shift keying (MSK) modulation, which has the following Bit Error Rate:   1 BER = Q( 2y/PG/CG) = erfc( y/PG/CG), 2

(6)

where y gives the SNR. The processing gain (PG) for 802.15.4 is given by 10 log(2/0.25)=9 dB. The coding gain (CG) depends on the increased Hamming distance between the codes and is a function of the SNR itself. For a low packet error rate region, the coding gain is approximated as 2 dB [12]. Thus, the Packet Reception Rate equation is calculated to be  8∗packet  1 P RR = 1 − erfc( (x − 11 − P0 dB)) 2

size

(7)

where x denotes the SNR in dBm. We note that the bit-error-rate approximation given in Equation 6 assumes coherent demodulation using carrier phase information. Practical transceiver designs use non-zero IF and noncoherent demodulation. The non-ideal receiver structures can be approximated with SNR reduction or equivalently increase in the noise floor (P0 ). We use a radio sensitivity of -94 dBm to adjust for the noise power P0 . To complete our analytical model, we require RF environment parameters: Path Loss Exponents (PLE) of indoor and outdoor and standard deviation of RSSI(dB). We performed RSSI measurements in a corridor in the second floor of the Dreese Lab Building and outdoor in a parking lot. For the indoor tests, we measured RSSI at 20 different distances (1 ∼ 20 unit (1 unit= 3ft)) within maximum communication range with the highest transmission power level (0dBm). For the outdoor test, 10 measurements were taken from the distance of 1 ∼ 10 unit distances where 30 ft seems to be the maximum communication range with the same transmission power, 0 dBm. Figure 2 shows the observed received signal values and the associated log-normal fit. Table 1 presents the summary. 3.3

Messaging Layer

CTP [9] is a tree-based collection protocol. Nodes generate routes to the sink using a routing gradient. CTP uses ETX [13] as the default routing metric.

532

T. Kwon, E. Ertin, and A. Arora

Corridor Fitted

100

100 90

90 80 0

Outdoor Fitted

110 y = − 2.2776*x + 106.31

RSSI

RSSI

110 y = − 1.7555*x + 107.86

10 5 10log10(Distance)

RSSI

110

15

80 0

2

4

6

10log10(distance)

8

10

Warehouse Fitted

y = − 1.3545*x + 104.14

100 90 80 0

5

10

15

10log10(distance)

Fig. 2. Indoor Corridor, Outdoor & Indoor Warehouse (RSSI vs. Distance) Table 1. Log normal model variables for Indoor Corridor, Outdoor, and Indoor Warehouse RF environments Metrics

Indoor Corridor Outdoor Indoor Warehouse

Path Loss Exponent

1.7555

2.2776

1.3545

RSSI Standard Dev.

5.0 dB

4.2 dB

5.0 dB

ETX implicitly favor long links over short links because each node selects the path with the minimum number of expected transmissions. Therefore, we expect ETX works similar to other metrics which give preference to long links, such as PRR×d [7]. For the linear network with 20 nodes, Node 1 is designated as the source, and Node 20 is set as the sink. All other intermediate nodes act as multihop relays. A source packet is generated every two seconds. The low data rate ensures no interference from previous packets sent trough the network. For each power level we use 1000 source packets and log all the paths that each packet have gone through, and only exclude the first and lost hop to avoid edge effects. Using the information embedded into the packets we compute (i) median link length, (ii) end-to-end latency, and (iii) link usage spectrum, for each power level. 3.4

Results

We first present link usage spectrum for indoor and outdoor environments at various transmit power levels using data from the experiments in Figure 3. We see that the XPlantError is minimized for the corridor and the outdoor experiments when the latter uses a transmit power level between -3dBm and 0dBm, and for the corridor and warehouse experiments when the latter uses a transmit power level between -18dBm and -13dBm. Table 2 shows that the outdoor XPlantError is minimized for the 0dBm case, and the warehouse XPlantError is minimized for the -18dBm case.

Reproducing Performance 1 Cumulative Usage Weights

Cumulative Usage Weights

1

533

0.8 0.6 0.4 indoor (tx=-13dBm) outdoor (tx=-6dBm) outdoor (tx=-3dBm) outdoor (tx=0dBm)

0.2

0.8 0.6 0.4 Corridor (tx=-13dBm) Warehouse (tx=-10dBm) Warehouse (tx=-13dBm) Warehouse (tx=-18dBm)

0.2

0

0 0

5

10

15

20

0

5

Distance (3 ft)

10

15

20

Distance (3 ft)

Fig. 3. Cumulative Link Usage Spectrum (Left: Corridor vs Outdoor, Right: Corridor vs Warehouse) Table 2. Comparison of l1 distance between Corridor network with tx=-13dBm and various Outdoor/Warehouse networks l1 distance Outdoor

Outdoor

Outdoor Warehouse Warehouse Warehouse

tx=-6dBm tx=-3dmB tx=0dBm tx=-10dBm tx=-13dBm tx=-18dBm Anlytical

2.7556

1.8767

1.203

3.2463

1.6976

1.5986

Next, we show that matching link usage spectrums results in consistent protocol performance as measured by commonly adopted metrics of mean-link-length and end-to-end latency in these experiments. Table 3 and 4 shows summary statistics for link length and end-to-end delay for the indoor and outdoor environments, for each specified transmission power level. We see that the performance of the corridor environment network is best matched by the outdoor environment network between when the latter uses a transmit power level between -3dBm and 0dBm, and by the warehouse environment network when the latter uses a transmit power level between -18dBm and -13dBm, which is consistent with our main hypothesis.

Table 3. Link Length Statistics for Corridor, Outdoor, and Warehouse Experiments Link

Indoor

Outdoor

Outdoor

Outdoor Warehouse Warehouse Warehouse

Length tx=-13dBm tx=-6dBm tx=-3dBm tx=0dBm tx=-10dBm tx=-13dBm tx=-18dBm Average

6.9329

4.1842

5.0584

7.7749

8.6506

7.6289

5.4536

Median

7

4

5

8

8

7

5

Table 4. End-to-End Delay for Corridor, Outdoor, and Warehouse Experiments e-to-e delay

Indoor

Outdoor

Outdoor

Outdoor Warehouse Warehouse Warehouse

tx=-13dBm tx=-6dBm tx=-3dBm tx=0dBm tx=-10dBm tx=-13dBm tx=-18dBm

Average

3.712

5.2481

4.4512

3.2049

2.2291

2.6072

4.2544

Median

4

5

4

3

2

3

4

534

4

T. Kwon, E. Ertin, and A. Arora

Analytic Methods for Predicting Link Usage Spectrum

In this section, towards showing one way in which the link usage spectrum in a test environment can be calculated, we derive analytical expressions that are accurate approximations of the link usage spectrum. Our derivation is specific to forwarding protocol that maximizes PRR×d protocol metric under lognormal shadowing model; this protocol metric is sometimes referred to as ELD. With such an analytical model, we can choose the attenuation level for matching link usage spectrum of two different deployment settings, so that comparable network performance may be achieved. We also present simulation (specifically, Monte Carlo and TOSSIM) results that corroborate the analytical framework developed here. We note that while the theorems here are presented using PRR×d as the protocol metric for evaluating links, they are easily customized to other forwarding protocols based on optimization of network metrics encapsulating PRR, SNR and d. 4.1

Analytic Model for 1-Dimensional and 2-Dimensional Uniform Graphs

The link usage statistics represent order statistics for the chosen protocol metric. It is straightforward to express the probability of attaining the maximum value among collection of random variables using the probability and cumulative density functions of the underlying random variables. However, for many transport protocols, the protocol metric for each link has a different probability distribution with non-identical support. This imposes a complex partition of the underlying multidimensional of link SNR’s with nonlinear boundaries prohibiting formulation of closed form expressions for link usage statistics. Instead of relying on computationally expensive numerical methods for calculating the relative volume of each partition, we derive analytical expressions for link usage statistics by approximating the partition boundaries with piecewise linear functions resulting in simple, accurate expressions for link usage. These provide computational savings over direct numerical integration method and, more importantly, allow generalization for the asymptotic case of the network size going to infinity. Theorem 1. For a protocol P that uses PRR × d as the metric for choosing forwarding links, the probability of choosing link li over link lj is expressed as follows: P (P RR(yi ) ∗ pi > P RR(yj ) ∗ pj )    (y −μ )2  a yi i 1 1 1 z3 √ = +√ z− + ... e− 2σ2 dyi 2 3 π −∞ σ 2π    (y −μ )2  ∞ yi i g(βj ) − μyj 1 1 1 √ √ + + erf e− 2σ2 dyi , σ 2π 2 2 2σ a

Reproducing Performance

535

yi −μy

where z = σ√2 j , and yi , yj are SNR (dB) values of link li , lj , and pi , pj are p progresses to destination for link li , lj , βj = pji , a = P RR−1 (min{βj , β1j }), g(βj ) = a if βj ≥ 1 and g(βj ) = ∞ if βj < 1 Proof. The probability of choosing li over lj with metric P RR × D, Prob(P RR(yi ) ∗ di > P RR(yj ) ∗ dj ) = Prob (P RR(yi ) > βj P RR(yj )) d

where we define the link length ratio as dji = βj . Without loss of generality, we can assume βj < 1. Then we can approximate the boundary between the protocol metrics calculated for two links as: P RR(yi ) > βj P RR(yj )  yi > h(yj ),

(8)

 yj , if yj < a h(yj ) = , a, if yj ≥ a

where

when a = P RR−1 (βj ) Prob (P RR(yi ) > βj P RR(yj )) = Prob(yi > yj , yj < a) + Prob(yi ≥ a, yj ≥ a) = Prob(yi > yj , yi < a) + Prob(yi ≥ a) Because yi , yj are Gaussian: Prob (P RR(yi ) > βj P RR(yj ))    (y −μ )2  a yi i yi − μyj 1 1 1 √ √ = + erf e− 2σ2 dyi 2 2 σ 2 −∞ σ 2π    (y −μ )2  ∞ yi i g(βj ) − μyj 1 1 1 √ √ + + erf e− 2σ2 dyi , σ 2π 2 2 2σ a where

 a, g(βj ) = ∞,

if βj ≥ 1 if βj < 1

Finally, we arrive at the expression in Theorem 1 retaining the first two terms (−1)n z 2n+1 of the Taylor series of erf (z), erf (z) = √2π ∞ n=0 n!(2n+1) . Theorem 2. For a protocol P which uses PRR × d as the metric for choosing forwarding links, the probability of choosing link l of length over all other links is expressed as follows:

536

T. Kwon, E. Ertin, and A. Arora





L(W, P, index(l)) = P P RR(y(l)) ∗ p(l) = max {P RR(yj ) ∗ pj } j=1...M

=

M −1  ak+1  ak

k=0

⎡ ⎣



M 

j=k+1

⎡

k   1

1 √ ⎣ σ 2π j=1

1 1 +√ 2 π



2

1 + erf 2

zj3 zj − + ... 3



g(βj ) − μyj √ 2σ



⎤ ⎦

⎤ (y−μy(l) )2 ⎦ e− 2σ2 dy,

where βj = pj /p(l), aj = P RR−1 (min{βj , β1j }), and the links are enumerated such that a1 ≤ a2 ≤ ... ≤ aM −1 , with a0 = −∞, aM = ∞ and yj is the SNR yj −μy experienced by link lj , zj = σ√2 j , and  aj , if βj ≥ 1 g(βj ) = ∞, if βj < 1 The proof is an immediate extension of Theorem 1 and is given in [14]. Both results rely on the approximation used in Equation 8 to convert the nonlinear boundary where one link is preferred over another to a piece-wise linear boundary suitable for close-form integration. 4.2

Validation of Analytical Approximations

We performed Monte-Carlo simulations to verify the accuracy of this analytical approximation of the link usage spectrum for various radio propagation environments and transmit power levels. The Monte-Carlo simulation results are nearly identical to the analytical results, supporting the approximations used in the derivation of the analytical expressions of link usage spectrum. We also compared the results of the analytical model with the observed link usage spectrum in the experiments described in the previous section. We observed that although the general behavior of link usage spectrum as a function 0.5

0.5 indoor (tx=-13dBm) outdoor (tx=-6dBm) outdoor (tx=-3dBm) outdoor (tx=0dBm)

indoor (tx=-13dBm) outdoor (tx=-6dBm) outdoor (tx=-3dBm) outdoor (tx=0dBm)

0.4 Usage Weights

Usage Weights

0.4 0.3 0.2

0.3 0.2 0.1

0.1

0

0 0

2

4

10 8 6 Distance (3 ft)

12

14

16

0

2

4

10 8 6 Distance (3 ft)

12

14

16

Fig. 4. Link Usage Spectrums. Left: Analytical (Theorem 2), Right: Monte-Carlo Simulation

Reproducing Performance

537

of transmit power is consistent between analytical and experimental results, the links chosen in the field experiments are in general shorter than the analytically derived link lengths. Careful analysis of the temporal variations in the experimentally observed tree structure suggested that temporal RSSI variations are one major reason for the observed gap. The analytical model derived in this section is a large scale fading model and does not consider the impact of temporal RSSI variations on link usage pattern. As shown in Figure 1, if we only consider spatial RSSI variations, PRR of all the chosen links are close to 1. However, if we consider temporal RSSI variations (typically, with a standard deviation of 3dB), long links will suffer in greater degree because they are likely to be closer to the threshold SNR and therefore subject to major fluctuations in their PRR values. As a result, the routing protocol will deselect long links encountering temporal fades. To account for this observed behavior, we consider only links whose unconditional (prior) expected PRR is larger than 1%. For example, using figure1, for an outdoor network with 0dB transmission power, this would correspond to all links whose length is less than or equal to 10. The result is shown in Figure 5.

1 Cumulative Usage Weights

Cumulative Usage Weights

1 0.8 0.6 0.4 indoor (tx=-13dBm) outdoor (tx=-6dBm) outdoor (tx=-3dBm) outdoor (tx=0dBm)

0.2 0

0.8 0.6 0.4 indoor (tx=-13dBm) outdoor (tx=-6dBm) outdoor (tx=-3dBm) outdoor (tx=0dBm)

0.2 0

0

5

10 Distance (3 ft)

15

20

0

5

10 Distance (3 ft)

15

20

Fig. 5. Cumulative Link Usage Spectrum (Left: Theorem 2 with Refinement, Right: Experiments)

2-Dimensional Simulation Study. We also compared analytical and simulation results of the link usage spectrum for the CTP protocol in a two dimensional 10×10 grid topology. The internode distance in this case is 6ft, and the sender is located at the leftmost bottom corner, and the destination is at the rightmost top corner. We use Theorem 2 to calculate the two dimensional link usage spectrum, and use TOSSIM 2 [15] simulator to obtain empirical data to validate analytical results. The source node is located at grid location (1,1) and it sends packets over the multihop network to the sink node located at (10,10). Figure 6 shows the analytical results calculated with Theorem 2. Figure 7 shows TOSSIM 2 [15] simulation results. We note that analytical expressions assume an infinite network model, calculate performance per unit distance, and scale the results to a finite size network. Analytical results show that a given indoor network of transmission power (-17dBm) is matched optimally with outdoor network of (-5dBm) transmission power, by minimizing the XPlantError.

538

T. Kwon, E. Ertin, and A. Arora 5

Indoor tx=−17dBm

4

Distance (1unit=6ft)

Distance (1unit=6ft)

5

3 2 1 0 0

1

2 3 4 Distance (1unit=6ft)

5

Outdoor tx=−5dBm

4 3 2 1 0 0

1

2 3 4 Distance (1unit=6ft)

5

Fig. 6. Link Usage Spectrum (Analytical). Left: Indoor -17dBm and Right: Outdoor -5dBm

Cumulative Usage Weights

1 0.8 0.6 0.4 Indoor:tx=-17dBm Outdoor:tx=-5dBm Outdoor:tx=0dBm

0.2 0 0

15 10 5 Progress to Destination (6ft)

20

Fig. 7. Cumulative 2-Dimensional Link Usage Spectrum (TOSSIM 2 Simulation)

Figure 6 shows contour plots for two dimensional link usage spectra for indoor (-17dBm) and outdoor (-5dBm). We observe analytically that the link usage is very similar at the optimal matching attenuation of −12dB. The simulation results of these two networks closely follow the analytical results. In Figure 7, the cumulative usage weights of indoor -17dBm matches well with outdoor -5dBm, and usage weights peak around progress 3 and saturate around progress 5, which conforms with Figure 6, where usage weights peak at coordinate (2,2) (progress 3) and saturate at coordinate (3.5,3.5) (progress 5). We also give cumulative usage weights for a second outdoor network of transmission level of 0 dB, as an example of mismatch of usage weights. Table 5 shows average end-to-end delays and link progresses for each environments. As predicted with analytical results, indoor -17dBm matches outdoor -5dBm better than it does 0dBm. Table 5. End-to-end Delay and Link Progress for 2-D Environments (TOSSIM 2 Simulation) Indoor tx-17dBm Outdoor tx=0dBm Outdoor tx=-5dBm e-2-e delay

5.23

3.85

4.68

link progress

2.34

3.9

2.81

Reproducing Performance

5

539

Conclusion and Future Work

In this paper, we studied a method for achieving comparable protocol performance across deployment and test environments. We have found the method to be valid across several protocols (which in turn were based on different forwarding metrics), various performance metrics, and diverse environments and presented some of those results here. We have also shown that our analytic methods for predicting link usage yielded good approximation in some environments, but need refinement in others (for instance, when temporal variation of links was significant). Error metrics other than XPlantError can be employed for matching performance. In particular, one can choose the transmit power attenuation to match the expected value of a single specified performance metric, such as at SNR, PRR, End-to-end Latency (1/P RR × D), ELD, P RR × D, as follows:   n n     ˜ (β), P, i) g(yi |li ↑)L(W, P, i) − g( y˜i |li ↑)L(W (9)    i=1

i=1

Outdoor Tx. Power (dBm)

where li ↑ means li is the conditional expectation given the link li was chosen. g is the performance metric chosen by the designer. It is easy to show that this error metric is bounded above by the l1 distance between the link usage spectrums proposed in this paper. Also, in this paper, we gave evidence that the use of the generic method of link usage spectrum matching suffices to obtain comparable performance over a wide variety of network metrics. Figure 8 compares the optimal attenuation level for link usage spectrum matching with the optimal attenuation required for matching end-to-end latency of test/deployment networks. We observe that the two approaches yield essentially the same attenuation factors.

0 -5 -10 -15

Analytical:LUS Monte-Carlo Sim:LUS Analytical:g=1/PRRxD

-20 -20 -19 -18 -17 -16 -15 -14 -13 -12 Indoor Tx. Power (dBm)

Fig. 8. Relation between Indoor Corridor and Outdoor transmit power for minimizing l1 distance between the link usage spectrums. Also, relation between Indoor Corridor and Outdoor transmit power to match end-to-end latency performance (cyan).

540

T. Kwon, E. Ertin, and A. Arora

References 1. Naik, V., Ertin, F., Zhang, H., Arora, A.: Wireless testbed bonsai. In: 2nd International Workshop on Wireless Network Measurement (2006) 2. Naik, V., Arora, A., Sinha, P., Zhang, H.: Sprinkler: A reliable and energy efficient data dissemination service for wireless embedded devices. In: 26th IEEE Real-Time Systems Symposium (2005) 3. Zhang, H., Arora, A., Sinha, P.: Learn on the fly: Data-driven link estimation and routing in sensor network backbones. In: 25th IEEE International Conference on Computer Communications (2006) 4. Arora, A., Ertin, E., Ramnath, R., Nesterenko, M., Leal, W.: Kansei: A high-fidelity sensing testbed. IEEE Internet Computing 10(2), 35–47 (2006) 5. Kansei Testbed, http://exscal.nullcode.org/kansei 6. Zuniga, M., Krishnamachari, B.: Analyzing the transitional region in low power wireless links. In: IEEE SECON (2004) 7. Seada, K., Zuniga, M., Helmy, A., Krishnamachri, B.: Energy-efficient forwarding strategies for geographic routing in lossy wireless sensor networks. In: Sensys (2004) 8. TEP 123, http://www.tinyos.net/tinyos-2.x/doc/html/tep123.html 9. Zhao, J., Govindan, R.: Understanding packet delivery performance in dense wireless sensor networks. In: Sensys (2003) 10. C. Technology, “Telosb datasheet”, http://www.xbow.com/Products/Product_pdf_files/Wireless_pdf/TelosB_ Datasheet.pdf 11. Chipcon, “cc2420 datasheet”, http://www.chipcon.com/files/CC2420_Data_Sheet_1_4.pdf 12. Lanzisera, S., Pister, K.: Theoretical and practical limits to sensitivity in ieee 802.15.4 receivers. In: IEEE International Conference on Electronics, Circuits and Systems (2007) 13. Couto, D., Aguayo, D., Bicket, J., Morris, R.: A high-throughput path metric for multi-hop wireless routing. In: Mobicom (2003) 14. Kwon, T., Ertin, E., Arora, A.: Transplanting protocols across different environments. Tech. Rep. OSU-CISRC-9/08-TR47, Ohio State University (2008) 15. TOSSIM 2, http://docs.tinyos.net/index.php/TOSSIM

Author Index

Hosseininezhad, Seyedali Huynh, D.T. 507

Alhawari, A.R. H 281 Alhemyari, Ali Z. 233 Arienzo, Loredana 249 Arora, Anish 524 Atanasovski, Vladimir 146 Awwad, Samer A.B. 281

Iannone, Luigi 82 Ismail, Abdallah 457 Jaekel, Arunita 400 Jiang, Yingjun 371 Juraschek, Felix 98

Bahi, Jacques M. 344 Bao, Miao 130 Bari, Ataul 400 Bezahaf, Mehdi 82 Biswas, Subir 314 Blywis, Bastian 98

Kakuda, Yoshiaki 114 Kang, Kyoung-Don 168 Kapitanova, Krasimira 168 Karnapke, Reinhardt 34 Kohno, Eitaro 114 Kunimune, Hisayoshi 130 Kwon, Taewoo 524

Cai, Wei 416 Ceesay, Omar.M. 233 Chehri, Abdellah 471, 495 Chen, Min 384, 416 Cui, Jun-Hong 185 Daigle, John N. 429 de Amorim, Marcelo Dias de Paz, Rodolfo 217

82

Egger, Christoph 48 Eltoweissy, Mohamed 1 Ertin, Emre 524

Lam, Nhat X. 507 Lambadaris, Ioannis 457 Lamont, Louise 447 Leung, Victor C.M. 64, 384, 416 Li, Jun 447 Longo, Maurizio 249 Lung, Chung-Horng 371, 457 Luo, Fangyun 400 Makhoul, Abdallah 344 Miˇsi´c, Jelena 314 Mizuta, Kazumasa 130 Mouftah, Hussein 471, 495

Fabini, Joachim 48 Fan, Xinxin 328 Fdida, Serge 82 Froese, Will 400 Fuwa, Yasushi 130 Gavrilovska, Liljana 146 Goel, Nishith 371, 457 Gong, Guang 328 Gonzalez, Sergio 416 Gonz´ alez-Valenzuela, Sergio G¨ une¸s, Mesut 98 Guyeux, Christophe 344 Happenhofer, Marco Hofmann, Sebastian

64

48 98

384

Nefzi, Bilel 265 Ng, Chee Kyun 233, 281 Nguyen, Trac N. 507 Niimura, Masaaki 130 Nilsson, Erik G. 17 Nolte, J¨ org 34 Noordin, Nor Kamariah 233, 281 Nose, Hiroaki 130 N¨ urnberger, Stefan 34 O’ Driscoll, Aisling 297 Ohta, Tomoyuki 114

542

Author Index

Okazaki, Tomoya 114 Olariu, Stephan 1

Song, Ye-Qiong 265 Son, Sang H. 168 Stølen, Ketil 17

Pesch, Dirk 217, 297, 359 Peytchev, Evtim 482 Pister, Kris 201 Prasad, Neeli Rashmi 160 Prasad, Ramjee 160 Qasem, Yaaqob Ali.A.

233

Rabbath, Camille-Alain 447 Radaideh, Ahmad 429 Rasid, Mohd. Fadlee A. 281 Rea, Susan 359 Risan, Øyvind 482 Rohokale, Vandana Milind 160

Tinka, Andrew

201

Villaverde, Berta Carballido Wallentin, Lukas 48 Watteyne, Thomas 201 Yoshikawa, Yasushi 130 Younis, Mohamed 1 Zheng, James Peng 185 Zhou, Robert Zhong 185 Zhou, Yifeng 447 Zhu, Yibo 185

359