Energy-Aware Communications - Eunice 2011

Lecture Notes in Computer Science Commenced Publication in 1973 Founding and Former Series Editors: Gerhard Goos, Juris...

Author: Ralf Lehnert

25 downloads 1136 Views 6MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Lecture Notes in Computer Science Commenced Publication in 1973 Founding and Former Series Editors: Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board David Hutchison Lancaster University, UK Takeo Kanade Carnegie Mellon University, Pittsburgh, PA, USA Josef Kittler University of Surrey, Guildford, UK Jon M. Kleinberg Cornell University, Ithaca, NY, USA Alfred Kobsa University of California, Irvine, CA, USA Friedemann Mattern ETH Zurich, Switzerland John C. Mitchell Stanford University, CA, USA Moni Naor Weizmann Institute of Science, Rehovot, Israel Oscar Nierstrasz University of Bern, Switzerland C. Pandu Rangan Indian Institute of Technology, Madras, India Bernhard Steffen TU Dortmund University, Germany Madhu Sudan Microsoft Research, Cambridge, MA, USA Demetri Terzopoulos University of California, Los Angeles, CA, USA Doug Tygar University of California, Berkeley, CA, USA Gerhard Weikum Max Planck Institute for Informatics, Saarbruecken, Germany

6955

Ralf Lehnert (Ed.)

Energy-Aware Communications 17th International Workshop, EUNICE 2011 Dresden, Germany, September 5-7, 2011 Proceedings

13

Volume Editor Ralf Lehnert Technische Universität Dresden Institut für Nachrichtentechnik 01062 Dresden, Germany E-mail: [email protected]

ISSN 0302-9743 e-ISSN 1611-3349 ISBN 978-3-642-23540-5 e-ISBN 978-3-642-23541-2 DOI 10.1007/978-3-642-23541-2 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: 2011935040 CR Subject Classification (1998): C.2, H.4, D.2, H.3, F.2, K.4 LNCS Sublibrary: SL 3 – Information Systems and Application, incl. Internet/Web and HCI

© Springer-Verlag Berlin Heidelberg 2011 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. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

Preface

It is my great pleasure to welcome you to the proceedings of the 17th EUNICE workshop held in Dresden. EUNICE has a long tradition in bringing together young researchers in communication network modeling and design from all over Europe. The single-track structure with suﬃcient time for presentations has always provided a platform for stimulating discussions. This year’s focus was on the actual topic of “energyaware communications.” Communication networks today account for 2 % of the worldwide emissions of CO2 with an exponentially rising trend. Currently, electrical energy is generated in a centralized fashion in a few power plants and distributed to the user. Eﬃcient use of energy requires distributed generation and distributed control of alternative energy generators. Therefore, the power grid has to become a so-called smart grid, which in turn requires a communication network for controlling the power grid. Smart grid communications can be realized by all communication technologies, wired or wireless. Applications range from meter reading to home automation and entertainment and the control of distributed power plants. EUNICE 2011 addressed research issues of energy-aware communication networks and communications for smart grids. EUNICE 2011 consisted of three keynotes on smart planet communications, network coding, and resource allocation. Sixteen full papers in seven sessions were accepted. Furthermore there was a session with seven poster presentations of ongoing research. My deep thanks go to our sponsors Comarch, Cracow, Poland, Detecon, Bonn, Germany, and Elcon Systemtechnik, Hartmannsdorf, Germany. Their generous support helped to reduce the registration fees signiﬁcantly. Many people worked hard to prepare this workshop: I would like to recognize the work of Stanislav Mudriievskyi, Stefan T¨ urk, Volker Richter, Roland Schingnitz, Rico Radeke and Jorge Robles. September 2011

Ralf Lehnert

Organization

EUNICE 2011 was organized by the Chair for Telecommunications, Technische Universit¨ at Dresden.

Executive Committee Conference Chair Ralf Lehnert

Chair for Telecommunications, TU Dresden, Germany

Local Organization Stanislav Mudriievskyi Martin Schuster Stefan T¨ urk Volker Richter Rico Radeke Jorge Robles

TU TU TU TU TU TU

Dresden, Dresden, Dresden, Dresden, Dresden, Dresden,

Germany Germany Germany Germany Germany Germany

Finance Chair Roland Schingnitz

TU Dresden, Germany

Technical Program Committee Finn Arve Aagesen Sebastian Abeck Marco Ajmone Marsan Laurie Cuthbert J¨ org Ebersp¨ acher Claudia Eckert Markus Fiedler Carmelita G¨ org Annie Gravey Jarmo Harju Yvon Kermarrec Paul K¨ uhn Oivind Kure Ralf Lehnert Maurizio Munafo

NTNU, Trondheim, Norway KIT, Karlsruhe, Germany University of Milan, Italy University of London, UK TU M¨ unchen, Germany TU M¨ unchen, Germany BIT, Blekinge, Sweden University of Bremen, Germany TELECOM Bretagne, France Tampere University, Finland TELECOM Bretagne, France University of Stuttgart, Germany NTNU, Trondheim, Norway TU Dresden, Germany PT Turin, Italy

VIII

Organization

Miquel Oliver Michal Pioro Aiko Pras Burkhard Stiller Robert Szabo Andreas Timm-Giel Samir Tohme

University of Pompeu Fabra, Spain University of Warsaw, Poland University of Twente, The Netherlands University of Z¨ urich, Switzerland Budapest University of Technology, Hungary TU Harburg, Germany University of Versailles, France

Organization

Sponsors Platinum sponsor: Comarch, Cracow, Poland

Gold sponsor: Detecon, Bonn, Germany

Silver sponsor: Elcon Systemtechnik, Hartmannsdorf, Germany

Other sponsors: Die Informationstechnische Gesellschaft im VDE (ITG), Germany

EUNICE

Technische Universit¨ at Dresden, Germany

Chair for Telecommunications, TU Dresden, Germany

IX

Table of Contents

Keynote Talks Physical Layer Network Coding for Improved Energy Eﬃciency . . . . . . . . Eduard Jorswieck

3

A Sense of a Smarter Planet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Matthias Kaiserswerth

4

Resource Management in a New Green-IT World . . . . . . . . . . . . . . . . . . . . . Mauro Biagi

5

Session Papers Network Architectures On the Beneﬁt of Forward Error Correction at IEEE 802.11 Link Layer Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Floris van Nee and Pieter-Tjerk de Boer

9

Simple Modiﬁcations in HWMP for Wireless Mesh Networks with Smart Antennas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Muhammad Irfan Raﬁque, Marco Porsch, and Thomas Bauschert

21

Ad-Hoc and Wireless Networks On the Evaluation of Self-addressing Strategies for Ad-Hoc Networks . . . Ricardo de O. Schmidt, Aiko Pras, and Reinaldo Gomes Considerations in the Design of Indoor Localization Systems for Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jorge Juan Robles

31

43

System Simulation Backoﬀ Algorithms Performance in Burst-Like Traﬃc . . . . . . . . . . . . . . . . . Ievgenii Tsokalo, Yamnenko Yulia, and Stanislav Mudriievskyi New IEEE 802.16-2009 Compliant Traﬃc Shaping Algorithms for WiMAX Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Volker Richter and Stefan T¨ urk

54

65

XII

Table of Contents

Network Planning, Optimization, and Migration Multiple-Layer Network Planning with Scenario-Based Traﬃc Forecast . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shu Zhang and Ulrich Killat

77

Optimization of Energy Eﬃcient Network Migration Using Harmony Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stefan T¨ urk and Rico Radeke

89

Self-management of Hybrid Networks: Introduction, Pros and Cons . . . . . Tiago Fioreze and Aiko Pras

100

Traﬃc Engineering Evaluation of Diﬀerent Decrease Schemes for LEDBAT Congestion Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mirja K¨ uhlewind and Stefan Fisches

112

Comparative Traﬃc Analysis Study of Popular Applications . . . . . . . . . . . Zolt´ an M´ ocz´ ar and S´ andor Moln´ ar

124

Flow Monitoring Experiences at the Ethernet-Layer . . . . . . . . . . . . . . . . . . Rick Hofstede, Idilio Drago, Anna Sperotto, and Aiko Pras

134

Quality of Experience A Survey of Quality of Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Qin Dai Investigation of Quality of Experience for 3D Streams in Gigabit Passive Optical Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ivett Kulik and Tuan Anh Trinh

146

157

Energy Eﬃcient Architectures A SystemC-Based Simulation Framework for Energy-Eﬃciency Evaluation of Embedded Networking Devices . . . . . . . . . . . . . . . . . . . . . . . . Daniel Horvath and Tuan Anh Trinh Energy Considerations for a Wireless Multi-homed Environment . . . . . . . German Castignani, Nicolas Montavont, and Alejandro Lampropulos

169 181

Poster Session Method for Linear Distortion Compensation in Metallic Cable Lines . . . . Albert Sultanov, Anvar Tlyavlin, and Vladimir Lyubopytov

195

Table of Contents

Multimedia Services Diﬀerentiation in 4G Mobile Networks under Use of Situational Priorities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Alexander Dyadenko, Olga Dyadenko, Larisa Globa, and Andriy Luntovskyy

XIII

199

Downlink Femtocell Interference in WCDMA Networks . . . . . . . . . . . . . . . Zolt´ an Jak´ o and G´ abor Jeney

203

Techno-economic Analysis of Inhouse Cabling for FTTH . . . . . . . . . . . . . . Navneet Nayan, Rong Zhao, and Kai Grunert

209

Impact of Incomplete CSI on Energy Eﬃciency for Multi-cell OFDMA Wireless Uplink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Alessio Zappone, Giuseppa Alfano, Stefano Buzzi, and Michela Meo An Eﬃcient Centralized Localization Method in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohamadreza Shahrokhzadeh, Abolfazl T. Haghighat, and Behrooz Shahrokhzadeh Mechanisms for Distributed Data Fusion and Reasoning in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ioannis Papaioannou, Periklis Stavrou, Anastasios Zafeiropoulos, Dimitrios-Emmanuel Spanos, Stamatios Arkoulis, and Nikolas Mitrou Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

213

217

221

225

Physical Layer Network Coding for Improved Energy Eﬃciency Eduard Jorswieck TU Dresden, Germany

In 2000, the concept of network coding was introduced by Ahlswede et al. and it lead to paradigm changes on several layers. On the physical layer, it changed the way how interference is handled. Interference is now rather exploited than avoided or canceled. On the network and transport layer, it changed the way how packets are forwarded. Instead of separate packets, combinations of packets ﬂow through the network. The concept of network coding was successfully applied to several research scenarios ranging from ad-hoc and sensor networks to peertopeer ﬁle sharing. τ1 NT = |XT |

τNT

R1

transmitter

RN

(information) sources

T

encoder network code encoder channel code

NN = |XN |

channel decoder channel code

(information) sinks

NR = |XR| ρ1

ρNR

decoder network code receiver

First, we will review the basic concept of network coding by two basic example scenarios: the butterﬂy and the two-way relaying network. Then, the two classes of network coding over ﬁnite ﬁelds (linear network coding, deterministic and random network coding) and physical layer network coding (also known as wireless network coding) are discussed. Finally, we apply the idea to a multiple antenna interference network scenario and show how the energy eﬃciency of the overall network can be signiﬁcantly improved. The presentation is concluded by pointing out open questions as well as an overview over relevant references. R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, p. 3, 2011. c Springer-Verlag Berlin Heidelberg 2011

A Sense of a Smarter Planet Matthias Kaiserswerth IBM Research, Zurich, Switzerland

In the next couple of years, there are expected to be 2 billion people connected to the Internet. At the same time, the instrumentation and interconnection of the worlds human-made and natural systems is explodingwhich could mean that there soon will be more things connected to the Internet than there are people who are connected. This Internet of Things promises to give people a much better understanding of how complex systems work, so they can be tinkered with to make them work better. But it also opens up a whole new sphere of opportunity to mine the data to create anything from smarter traﬃc to smarter healthcare to smarter food supply chains.

R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, p. 4, 2011. c Springer-Verlag Berlin Heidelberg 2011

Resource Management in a New Green-IT World Mauro Biagi Department of Information, Electronics and Telecommunications engineering (DIET) Sapienza, University of Rome

Wireless mobile, cabled xDSL, Fiber Optical Systems and Power Line Communications share a common problem: Bandwidth-Power eﬃciencies. While the eﬃciency required for the spectrum usage is diﬀerent from wireless to wireline since the former has to respect some standardization bodies recommendations and the latter is mainly tied to channel quality aspects, the power eﬃciency is mainly linked to electrical power that comes from diﬀerent sources like carbon, gas, oil, wind, solar, water. Unfortunately only few of these are renewable and require institutional investments for the medium-far future. In the meantime the only one way to be run appears to be the eﬃcient. From this the necessity, in a Green-IT vision, to propose schemes, architectures, protocols and algorithms able to optimally manage bandwidth and power so to guarantee the expected performance at a reasonable cost. This is the direction currently followed by scientiﬁc community for already existing systems and in-developing new technologies. Hence, each optimizing procedure with constraints on power/bandwidth usage or with the goal of power minimization or spectrum allocation, falls in the framework of a smart management of resources, so bit-loading for multi-carrier systems, coding and access techniques, scheduling and diversity schemes can be interpreted as an instance of Green-IT. In this context also the dualism between convex (closed-form) optimization and recursive algorithm takes place since also the complexity of the algorithms involves a power aspect, the power used for processing. This keynote, tries to show diﬀerent approaches to eﬃciency by highlighting strong aspects and drawbacks for the present and next future.

R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, p. 5, 2011. c Springer-Verlag Berlin Heidelberg 2011

On the Benefit of Forward Error Correction at IEEE 802.11 Link Layer Level Floris van Nee and Pieter-Tjerk de Boer Centre for Telematics and Information Technology (CTIT) Faculty of Electrical Engineering, Mathematics and Computer Science University of Twente, Enschede, The Netherlands [email protected], [email protected]

Abstract. This study examines the error distribution of aggregated MPDUs in 802.11n networks and whether or not forward error correction like raptor coding at the link layer would be useful in these networks. Several experiments with Qualcomm 4x4 802.11n hardware were performed. Two devices were used in a data link, while a third device sniﬀed all transmitted packets. The collected data was analyzed and used to calculate the packet error rate which would be obtained if FEC was used in order to determine whether FEC is useful at the link layer. It is shown that the error distribution of A-MPDUs does not follow the binomial distribution. Because of this, the performance of FEC in real networks is worse than for theoretical cases where a binomial distribution is assumed. Therefore, other ways to decrease the packet error rate have more impact than forward error correction.

1

Introduction

Wireless Local Area Networks (WLAN) play an important role in home video networking, because they are cost-eﬀective to install and easy to set-up. However, there are still areas where WLAN needs to improve in order to deliver the performance needed to stream multiple High Deﬁnition (HD) streams simultaneously. One of the reasons for this is that video streaming poses diﬀerent requirements on the transfer link than normal web browsing does. For example, a much lower Packet Error Rate (PER) is required, because video applications are more sensitive to errors in the communications link. Lost packets require retransmission which cause higher latency and lower throughput. To the end user this may be noticeable as artifacts on the video display. Over the past few years there have been many studies on improving the link quality to allow streaming of multiple video streams. As latency is the most important aspect of video streaming, most studies focus on keeping latency as low as possible by preventing packet loss. One study tried to achieve this by adjusting the algorithm for rate control to use the Signal-to-noise Ratio (SNR) in determining the ideal transmission rate [4]. Another study used forward error correction at the application layer in order to correct errors which occur during transmission [1]. Both studies show some positive results, however they do not R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 9–20, 2011. c Springer-Verlag Berlin Heidelberg 2011

10

F. van Nee and P.-T. de Boer

use devices which support the latest version of the IEEE 802.11 standard. This standard, IEEE 802.11n, introduces a couple of important improvements for transmitting large amounts of data which make it especially useful for video streaming. The most important improvements in the 802.11n standard for video streaming are higher data rates and the possibility to use frame aggregation [5]. MPDU aggregation makes it possible to send up to 64 data frames in one single Aggregated MPDU (A-MPDU) frame. Each of these MPDUs in the aggregated frame can be individually acknowledged using a method called block acknowledgement. This makes MPDU aggregation especially suited for real-time streaming, because an error in one MPDU does not result in the loss of one complete transmit opportunity (TXOP). The lost MPDU can simply be retransmitted together with new data frames as part of the next A-MPDU. One study on forward error correction did note these improvements. They investigated the use of forward error correction for multicast video data by doing simulations [9]. However, their study is completely theoretical and also makes the false assumption that an A-MPDU can contain as many as 1000 MPDUs. Moreover, 802.11a PHY rates are used in their simulation instead of the 802.11n rates. Because of the increasing data rate of HD video and the need to transmit more video streams simultaneously, it is important to achieve even higher throughput and lower latency. This research investigates whether the use of a forward error correction code at the link layer level would improve the performance of the link. In order to determine the usefulness of such an error correction code, one important aspect of the wireless link should be carefully examined. This aspect is whether individual errors in an A-MPDU are correlated. A forward error correction code is most valuable when the errors are not correlated, but instead follow a binomial distribution with an independent error probability p. If errors are correlated, there is a higher probability that there are too many errors to correct in an A-MPDU. In addition, the probability that the additional error correction MPDUs are received in error also increases. Therefore, there is little beneﬁt in using forward error correction when errors are strongly correlated, because only a small percentage of errors can be corrected in that case. In that case, the reduction in error probability may not be justiﬁed by the overhead of the extra error correction data. Two research questions are introduced: – What is the distribution of errors in an A-MPDU and which factors inﬂuence this distribution? – Would the performance of the link increase when a forward error correction code is introduced at the link layer? The answer to the ﬁrst question will give insight into the distribution of errors in A-MPDUs. This information can be used to answer the second question. In order to obtain the information necessary to answer the research questions, several experiments are conducted. Qualcomm 4x4 802.11n devices are used for

On the Beneﬁt of Forward Error Correction

11

these experiments, because these devices have full support for aggregation of MPDUs and are therefore well suited for the tests. They use up to four spatial streams to send and receive data enabling them to transmit at a physical layer data rate of up to 600 Mbps. The setup for the experiment is as follows. One device continuously sends data to another device, while a third device which is placed close to the sender, sniﬀs all packets. The data is sent without using FEC at the link layer. The experiment is performed for several types of data. The captured data is then analyzed for A-MPDU length and error rate to obtain an answer to the ﬁrst research question. This data is then used to calculate the improvement in PER if forward error correction would be used. This paper ﬁrst presents an introduction to the topic of A-MPDUs in the 802.11n standard and forward error correction codes. After that, the methods for measuring are explained as well as the methods for analyzing the data. Then, the results and a discussion of the results are presented and ﬁnally a conclusion is drawn and an answer to the research questions is formulated.

2 2.1

Background Frame Aggregation in the IEEE 802.11n Standard

Using the full potential of the 802.11n standard, physical layer data rates of 600Mbps can be obtained. Another important improvement is introduced at the link layer. This improvement is the aggregation of data link frames. By combining several link layer frames, the overhead at the link layer is decreased. At the highest data rates, aggregation of MPDUs can lead to a three to four times increase in channel utilization [3]. The MAC-PHY (medium access control-physical) interface combines a maximum of 64 MPDUs into one large aggregated MPDU. The receiver of the packet deaggregates the A-MPDU in its MAC-PHY interface and sends a special type of acknowledgement back, called a block acknowledgement. This block acknowledgement contains a bitmap of 64 bits, in which each bit corresponds to the success of the transmission of one individual MPDU. In the next A-MPDU, the sender retransmits each MPDU which was not received correctly, possibly together with new MPDUs. This is also illustrated in Figure 1. A B

A-MPDU Header

MPDU SN : 1022

MPDU SN: 1023

MPDU SN: 1024

A-MPDU Header

MPDU SN: 1025

MPDU SN: 1022

MPDU SN: 1024

MPDU SN: 1026

Block Ack , Start SN : 962 Bitmap (64bit): 1 1 ... 0 1 0 1

Fig. 1. Simpliﬁed ﬂow of transmission of aggregated MPDUs

The most important improvement when using aggregation is the following. In the old approach only one MPDU could be sent after which the sender had to wait before the acknowledgement was received. This process takes one TXOP, so sending ten MPDUs using the old approach takes at least ten TXOPs. In the new approach, these ten MPDUs can be sent in one aggregated MPDU, therefore decreasing the time it takes to transmit the ten MPDUs to only one TXOP.

12

2.2

F. van Nee and P.-T. de Boer

Forward Error Correction

Forward error correction has traditionally been used in wireless networks at the physical layer. Here, convolutional codes are used to encode each m-bit information symbol into an n-bit symbol where n is larger than m. Convolution codes are not block codes, but process a continuous bitstream instead. This makes them not suitable for use at the data link layer, because at this layer the individual MPDUs need to be corrected. Normally, the failure of receiving an MPDU leads to a retransmission of that MPDU. However, it is possible to add speciﬁc error correction MPDUs to an A-MPDU which are able to recover these lost MPDUs. These types of forward error correction code are called block erasure codes. In addition, when an error correcting algorithm is such that a message of k information symbols can be encoded into a potentially limitless number of encoding symbols, the code is called a fountain code. One well-known class of fountain codes are raptor codes. The latest generation of raptor codes, RaptorQ [8], are used in this paper when calculations are done on the performance of error correction codes in the measurements. RaptorQ is a systematic code, meaning that the symbols of the original message are included in the set of encoding symbols. Using RaptorQ coding on a source block of size k, the message can be decoded with 99% certainty from the reception of k encoded symbols. For k + 1 received symbols, the decoding probability is 99.99% and for k + 2 it is 99.999999%.

3 3.1

Measurements Equipment and Setup

All the measurements were done with Qualcomm 4x4 802.11n wireless devices. These devices use four spatial streams to send and receive data and have full support for aggregation of frames. The setup of the measurements is displayed in Figure 2. In essence, two wireless devices are sending data to each other, while a third one is sniﬃng all packets. As can be seen in Figure 2, A is connected to the access point with a gigabit ethernet connection. Consequently, when A sends data to B, the data is ﬁrst sent to the access point and then the data is wirelessly transmitted to B. B wirelessly sends an acknowledgement back which is received by the access point. Because it is a MAC level acknowledgement, it is not propagated all the way back to A. The sniﬀer passively participates in the link, capturing all wireless communications between the access point and B. The captured packets are collected on a computer using Wireshark [2]. Because the sniﬀer should capture as many packets as possible to get accurate results, care has to be taken where to place the sniﬀer. Data frames are sent at a much higher data rate than (block) acknowledgements. Therefore, acknowledgements have a much larger range than data frames. In the setup of the measurements, the sniﬀer should be placed such that both data and acknowledgement frames are captured. The best place for this is close to the sender of the data frames.

On the Beneﬁt of Forward Error Correction

13

Fig. 2. The setup of the measurements

3.2

Procedure

In order to obtain the data necessary for analyzing, several tests were performed. For all except one test, Iperf was used as the tool to generate the data to transmit [6]. For the remaining test, a high deﬁnition video was streamed from A to B. One requirement for the data set was to be suﬃciently large to draw conclusions. There have to be several hundred captured A-MPDUs for each possible A-MPDU size to get meaningful results. After some initial tests using diﬀerent sizes of data sets, it was concluded that a set of one million packets would be suﬃciently large. Therefore, the sniﬀer was set to capture approximately one million packets per test. Another important aspect to mention is the variety of the tests. It is important to perform tests with diﬀerent parameters to obtain a broad overview of the problem. Therefore, tests are performed at several application layer data rates and also using diﬀerent network layer protocols. The diﬀerence between using TCP and UDP as the network protocol is that UDP is one-way communication and TCP is two-way. When using TCP, acknowledgements are also sent at network layer level. These acknowledgements are normal data frames at the link layer. Because the communication is two-way, there is a probability that collisions occur. Therefore, three tests were performed: one using Iperf to transmit UDP data at 60Mbps, one using Iperf to transmit TCP data at the maximum available rate (which was 200Mbps) and one last test where an HD movie was streamed at 15Mbps. 3.3

Information Extraction

To obtain an answer to the research questions, the distribution of errors in an A-MPDU needs to be extracted from the acquired data. A plugin for Wireshark was developed for this purpose. The plugin collects statistics of the data set by having a function which is called once per captured packet.

14

F. van Nee and P.-T. de Boer

The easiest way to collect statistics about the error distribution of A-MPDUs is to simply split the data set by the size of A-MPDUs. Then, the probability can be calculated that exactly x MPDUs in an A-MPDU of size s are received in error. This leads to the error distribution per A-MPDU size. One diﬃculty in this approach is how to detect whether or not the transmission of a packet was successful. There are two methods to do this. First, it is possible to examine the block acknowledgements transmitted by the receiver of the data A-MPDU. This block acknowledgement contains information about which MPDUs were received. Second, the retransmission ﬂag in an MPDU can be used to detect whether the packet is retransmitted. Logically, when a packet is retransmitted, it must have been sent before, though not received. The second approach is better suited for the analysis, because it is more accurate. Sometimes block acknowledgements are not sent. For example, in the case that one complete A-MPDU fails, the receiver does not receive anything, thus it does not send a block acknowledgement back. Using the ﬁrst approach, this would be treated as if nothing happened. The second approach does detect such cases. The approach still relies on the assumption that the sniﬀer captures all packets. However, this is not the case. Sometimes the sniﬀer does not receive the packet, though the packet does arrive at the receiver. Also, sometimes the packet is received neither by the sniﬀer nor by the intended receiver. The ﬁrst case is not a problem, as the packet reaches the intended receiver so the sniﬀer does not have to count an error. The second case poses a larger problem. When the retransmit of the lost packets occurs, the sniﬀer does not know in which AMPDU the original packets were sent. Therefore, it does not know what the size of the original A-MPDU was and in which error distribution the retransmission should be placed. This problem is illustrated in ﬁgure 3. For the analysis, these retransmissions were placed in a special group with unknown A-MPDU size. original A -MPDU size

?

A B

A-MPDU Header

MPDU SN: 970

MPDU SN: 971

.. timeout ..

A-MPDU Header

MPDU (R) MPDU (R) SN: 970 SN: 971

MPDU SN: 972 Block Ack , Start SN : 909 Bitmap (64bit): 1 1 ... 1 1

Fig. 3. Analysis diﬃculty: the original A-MPDU is not received by the sniﬀer, making it impossible to obtain the size of that original A-MPDU

The data was also split per sender. For example, if TCP data is sent from A to B, the actual data goes from A to B, but TCP acknowledgements are also transmitted from B to A. It is possible that the distribution of errors diﬀers greatly between these two links, so it has to be split in the results.

4

Results

The results of the three experiments are presented in this section. Error distribution plots have been created for each experiment and A-MPDU size. However,

On the Beneﬁt of Forward Error Correction

15

because there is a lot of similarity between the ﬁgures of diﬀerent A-MPDU sizes of the same experiment, not all of them are shown. Instead, only a couple of ﬁgures which are representative are shown to provide a clear overview of the results. Each ﬁgure contains three lines, two of which show the results of the experiment. They diﬀer in the interpretation of cases where an entire A-MPDU was missed by the sniﬀer (typically due to a collision). For the ‘best-case’ line, such events are not counted as errors at all, thus estimating what the distribution would be in the absence of collisions. For the ‘worst-case’ line, it is assumed that the missed A-MPDU consisted of k MPDUs that were all received in error, with k being the number of retransmissions in the next A-MPDU (obviously, the real size of the missed A-MPDU is not known). For comparison, the third line shows the theoretical probability that exactly n errors occur, given that at least one error occurs, under the assumption that they follow a binomial distribution (i.e., are independent). The parameter p for this binomial distribution is set to the overall probability that a single MPDU fails in an A-MPDU of that size, as calculated from the measured data for that case. 4.1

UDP 60 Mbps

In Figure 6 the error distribution for A-MPDUs of size 5 can be found. 97% of the A-MPDUs in this measurement had a size between 1 and 11; as the results for these sizes are similar, they are not shown here. Too few A-MPDUs were sent at a size larger than 11 to make those results interesting. 4.2

TCP at Maximum Rate

There was a large spread in the size of A-MPDUs in these results. 93% of the TCP data A-MPDUs were sent at a size between 1 and 30. The TCP acknowledgement A-MPDUs were smaller, but showed similar results. Figure 7 shows the distribution for TCP data A-MPDUs of size 20 and Figure 4 of TCP acknowledgements A-MDPUs of size 10. The results for other A-MPDU sizes are similar. 4.3

TCP 15 Mbps Video

89% of the A-MPDUs of the TCP video data had a size between 1 and 20. Figure 5 shows the distribution for TCP video A-MPDUs of size 20. Figure 8 shows the individual error probability of a MPDU plotted against the A-MPDU size in this measurement. A similar plot was also obtained for the other experiments, but these are not shown here because of the similarity of the results. Also, Figure 9 shows a plot which illustrates the error probability at a speciﬁc location in the A-MPDU. On the horizontal axis the location in the A-MPDU is given. The vertical axis shows the number of errors on that location, relative to the number of errors of the ﬁrst MPDU in the A-MDPU.

16

F. van Nee and P.-T. de Boer

100

50

90

45

80

40 Binomial p=0.0671

Binomial p=0.073

35

Measured best 60

Probability (%)

Probability (%)

70

Measured worst

50 40

Measured best 30

Measured worst

25 20

30

15

20

10

10

5 0

0 1

2

3

4

1

5

6

11

16

Number of errors

Number of errors

Fig. 4. Error distribution for TCP ac- Fig. 5. Error distribution for TCP video knowledgements at maximum rate and A- data traﬃc and A-MPDU size 20 MPDU size 10 70

30

60

25 Binomial p=0.1338

50

Measured best Measured worst

30

Probability (%)

Probability (%)

20

Binomial p=0.0845 40

Measured best Measured worst

15

10

20 5

10 0

0

1

2

3

4

5

6

7

8

9

Number of errors

10

1

6

11

16

Number of errors

Fig. 6. Error distribution for 60 Mbps Fig. 7. Error distribution for TCP data UDP traﬃc and A-MPDU size 5 traﬃc at maximum rate and A-MPDU size 20

5 5.1

Discussion Distribution of the Errors

First, the results show a signiﬁcant diﬀerence compared to the theoretical distribution. In almost all cases, the theoretical probability for a single error is larger than the measured probability. Also, at a higher number of MPDU errors in the A-MPDU, the probability is in almost all cases smaller in the theoretical distribution than in the measurements. The theoretical distribution reaches zero quickly, while the measurements show a ﬂoor of a few percent for almost all number of errors. This ﬂoor could be caused by an error during the training phase of reception. When such an error occurs, the probability of successfully decoding an MPDU decreases signiﬁcantly, thereby increasing the probability that a large number of MPDUs in the A-MPDU fail. One clear result is that the error probability of individual MPDUs is strongly related to the size of the A-MPDU in which it was transmitted: MPDUs which

MPDU error percentage (%)

25 20 15 Best case

10

Worst case 5 0 1

6

11

16

A-MPDU size (MPDUs)

21

Number of errors relave to posion 1

On the Beneﬁt of Forward Error Correction

17

3 2,5 2 Size 20

1,5

Size 15 1 0,5 1

6

11

16

Posion in A-MPDU

Fig. 8. The error probability of a single Fig. 9. The number of errors relative to MPDU plotted against the A-MPDU size the number of errors on the ﬁrst place plotfor TCP video traﬃc at 15 Mbps ted against the location of an MPDU in an A-MPDU

are part of a large A-MPDU have a higher error probability than MPDUs in a small A-MPDU. The UDP traﬃc measurements follow the theoretical (binomial) distribution more closely than the TCP traﬃc results, though the curve is still more ﬂat than the binomial distribution. This indicates that even when using one-way traﬃc like UDP, errors are still slightly correlated. There is almost no diﬀerence between the best and worst case scenarios, which indicates that the sniﬀer captured almost all packets. The TCP traﬃc shows larger diﬀerences with the theoretical model. The best case, which ﬁlters out all completely failed A-MPDUs, is still similar to the best case in UDP traﬃc. However, the worst case series shows a large peak at the end, indicating that the probability that all MPDUs in an A-MPDU fail increases dramatically when using TCP traﬃc. A reasonable explanation for this is that the peak is caused by collisions. TCP is two-way communication, whereas UDP is one-way communication. This means that there is a probability that A starts sending TCP data exactly at the same time as B starts sending a TCP acknowledgement. This results in a collision where both complete A-MPDUs are lost. These collisions can be avoided by using RTS/CTS [7], which is why the best-case situation should be observed here when determining the usefulness of error correction. In the best case situation in the ﬁgures, the collisions are ﬁltered out of the results. Not only the distribution of the number of errors per A-MPDU diﬀers between reality and the binomial model, but Figure 9 shows that MPDUs transmitted in the latter part of the A-MPDU have a greater probability of failure. A possible explanation for this is that small variations in the channel since the channel measurement in the preamble of the packet lead to greater inaccuracies later in the reception of the A-MPDU. In general, using TCP at the highest data rate, more larger A-MPDUs are sent than at the lower data rate of TCP video. However, UDP traﬃc is sent in even smaller A-MPDUs, even though the data rate of the UDP traﬃc was higher. This is explained by the fact that UDP A-MPDUs can be sent with less

18

F. van Nee and P.-T. de Boer

Table 1. Estimated packet error rates (PER) with various amounts of FEC and data rate reduction 1 FEC symbol per A-MPDU equiv. FEC FEC rate binom. reduc.

Estimated PER 2 FEC symbols per A-MPDU equiv. FEC FEC rate binom. reduc.

3 FEC symbols Measured per A-MPDU PER equiv. A-MPDU without FEC FEC rate size FEC binom. reduc. UDP 5 6.71% 1.82% 0.84% <0.01% 0.47% 0.05% <0.01% 0.09% <0.01% <0.01% 10 12.35% 6.96% 5.02% 1.24% 3.92% 1.47% 0.01% 2.28% 0.31% <0.01% TCP data at max rate 10 5.13% 2.59% 1.03% 0.52% 1.49% 0.13% <0.01% 0.92% 0.01% <0.01% 20 7.30% 5.27% 3.41% – 3.99% 1.23% 0.73% 3.18% 0.35% 0.73% 30 13.56% 11.56% 10.27% – 10.05% 7.18% – 8.73% 4.54% 1.36% TCP video data 10 6.04% 3.47% 1.40% 0.6% 2.04% 0.21% <0.01% 1.25% 0.02% <0.01% 20 13.38% 11.19% 8.67% – 9.45% 4.83% 1.34% 7.99% 2.26% 1.34% 25 20.38% 17.79% 16.41% – 15.53% 12.51% – 13.52% 8.87% 2.04%

waiting time between the packets. When using TCP, the link also has to be utilized for transmitting the acknowledgements, which means the sender has less time to transmit the frames. 5.2

Estimating the Usefulness of FEC

In order to judge the usefulness of forward error correction, the main questions are how much the PER would decrease when FEC is introduced, and how this compares to the reduction in PER when switching to a lower physical layer data rate. We have used the measured error rates and distributions to estimate both of these PER reductions. The resulting PER when using FEC is computed by using the measured error distributions to calculate how many retransmissions are needed. With n FEC symbols it is possible to repair up to n errors in an A-MPDU. Therefore, when x errors occur, a retransmission is only needed when x > n. In that case, x − n MPDUs need to be retransmitted in order to obtain enough data to decode all MPDUs. (The 1% error rate for decoding RaptorQ encoding symbols when x = n is not taken into account here, but this makes only a small diﬀerence.) We can also do this computation under the assumption that the errors would be binomially distributed, by simply taking the measured PER as the parameter for the binomial distribution. The reduction in PER resulting from reducing the physical data rate can be obtained from Figure 5.12 in Perahia and Stacey (2008). E.g., this Figure indicates that for the high data rates in the range of 405 Mbps to 600 Mbps that were used in these measurements, switching to a 10% lower rate would lead to about 10 times smaller PER. The underlying assumption here is that the channel is AWGN. In our tests, this provided a good estimation, because the path did not change during the tests. However, in real situations this may not be true, a fact that needs to be taken into account when drawing conclusions. The results are shown in Table 1. For several of the A-MPDU sizes occurring in each of our experiments, both the measured PER without FEC is shown, and

On the Beneﬁt of Forward Error Correction

19

the calculated PER for 1, 2 and 3 FEC MPDUs per A-MPDU. For comparison, for each of the FEC cases also the calculated PER with FEC under the binomial assumption is shown, and the PER resulting from an equivalent physical data rate reduction. E.g., for 1 FEC symbol per A-MPDU and an A-MPDU size of 10, using FEC would eﬀectively reduce the data rate by 10%, so we compare it to a 10% physical data rate reduction without FEC. This is only possible if such a datarate is available in 802.11n; where this is not the case, the table shows a ‘–’. As can be seen, the PER would not beneﬁt much from adding forward error correction: in all cases, FEC performs worse than switching to an equivalent lower physical data rate. Also, it can be seen that at a given PER, the eﬀectiveness of FEC is reduced due to the correlation of the packet errors (i.e., the deviation from the binomial distribution).

6

Conclusion

Several conclusions can be drawn from the results of the measurements in the previous section. First the answer to the ﬁrst research question on the error distribution in A-MPDUs. It is shown that the distribution of errors in A-MPDUs does not follow the binomial distribution with independent probability p that a single MPDU fails. By comparing the binomial distribution to the distribution as found in the results, two important diﬀerences can be observed. In the ﬁrst place, the results of the measurements lead to the conclusion that there is a correlation between errors in A-MPDUs. Consequently, the presence of one error in an AMPDU leads to a higher probability of a second error in the same A-MPDU. Therefore, the slope of the distribution function of errors is ﬂatter than that of the binomial distribution function. Secondly, there is a probability that the entire A-MPDU fails due to a collision. This can be observed in the distribution function as a peak at the end. There are several factors which determine the distribution of errors as can be observed in the results of the measurements. The most obvious factor is the size of the A-MPDU itself. There is a higher probability that larger A-MPDUs contain errors. Also, the probability that a single MPDU fails increases with A-MPDU size. Another important factor is the existence of another device transmitting on the same frequency, thus also whether or not the link is bidirectional. The probability that a collision occurs is much higher in such links, which can be noticed in the distribution function by a higher peak for complete A-MPDU failure. Furthermore, the application layer data rate does not have much impact on the error rate of the MPDUs, because this does not change the physical layer data rate at which the packets are sent. For the second research question on the usefulness of forward error correction codes the following can be concluded. The performance gain when using a forward error correction code on the MAC layer depends heavily on the premise that errors on the link are binomially distributed and not correlated. When errors are correlated, there are often too many errors in the packet to correct.

20

F. van Nee and P.-T. de Boer

The results show that the distribution of errors can not be approximated by the binomial distribution function, because of the correlation of errors. In addition, A-MPDUs are sometimes received completely in error. While the latter can be avoided by using RTS/CTS, the ﬁrst remains a problem for FEC. Therefore, the performance of raptor coding is worse than suggested by previous studies that assume that MPDUs in an A-MPDU fail indepently [9]. It can be concluded that it is generally better to switch to a lower physical layer data rate than to use raptor coding. Acknowledgements. The authors would like to thank Qualcomm for making their latest 802.11n wireless devices available for this research.

References 1. Alay, O., Korakis, T.: An Experimental Study of Packet Loss and Forward Error Correction in Video Multicast over IEEE 802.11b Network. In: Proceedings of IEEE CCNC (2009) 2. Combs, G.: Wireshark (December 2010), http://www.wireshark.org 3. Ginzburg, B., Kesselman, A.: Performance analysis of A-MPDU and A-MSDU aggregation in IEEE 802.11n. In: 2007 IEEE Sarnoﬀ Symposium, pp. 1–5 (April 2007) 4. Haratcherev, I., Taal, J.: Automatic IEEE 802.11 rate control for streaming applications. Wireless Communications And Mobile Computing 5, 421–437 (2005) 5. IEEE Computer Society. Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Speciﬁcations. IEEE Draft P802.11-REVmb/D4.0 (June 2010) 6. NLANR/DAST. Iperf (December 2010), http://sourceforge.net/projects/iperf 7. Perahia, E., Stacey, R.: Next Generation Wireless LANs. Cambridge University Press, Cambridge (2008) 8. Qualcomm. Raptorq technical overview (December 2010), http://www.qualcomm.com/documents/files/raptorq-technical-overview.pdf 9. Samokhina, M., Moklyuk, K., Choi, S., Heo, J.: Raptor Code-Based Video Multicast over IEEE 802.11 WLAN. In: IEEE APWCS (2008)

Simple Modifications in HWMP for Wireless Mesh Networks with Smart Antennas Muhammad Irfan Rafique, Marco Porsch, and Thomas Bauschert Chemnitz University of Technology, 09126 Chemnitz, Germany {irfan.rafique,marco.porsch,thomas.bauschert}@ etit.tu-chemnitz.de

Abstract. In this paper we present simple modifications in the HWMP protocol for wireless mesh networks to incorporate different smart antenna transmission schemes. Our proposed amendments enable mesh nodes to select transmission techniques (spatial multiplexing, beamforming) adaptively according to the channel conditions. The advantages of diversity are exploited by transmitting control packets prior to beamforming communication. Simulation results illustrate that the modified HWMP (MHWMP) leads to a significant better throughput of the mesh network. Moreover, it enables a high degree of robustness wrt. wireless link failures for stationary mesh networks. Keywords: Wireless Mesh Networks, Routing, HWMP, Smart Antennas.

1 Introduction Wireless Mesh Networks (WMNs) have been recognized as low cost, reliable and scalable last mile access network solution. However the uncertain nature of the wireless channel and the multi-hop structure constrain the end to end performance of WMNs. In general, the efficiency of a wireless network can be improved up to a certain limit with suitable techniques at PHY layer. The IEEE 802.11n [2] amendment (Draft Part 11.0) is the latest addition under development of IEEE 802.11 providing an significant increase in the data rate up to 600 Mbps for traditional one hop WLANs (IEEE 802.11a/g). This increment is primarily achieved through PHY layer modifications: IEEE 802.11n relies on multiple input multiple output (MIMO) transmission and 40MHz channel bandwidth, while legacy IEEE 802.11 works with single input single output (SISO) transmission and 20MHz channel bandwidth. In recent years smart antennas are proposed as technology that besides MIMO (spatial multiplexing, MUX) transmission also allow for transmit/receive diversity, beamforming (BF) and null-steering. The current IEEE 802.11s draft standard for WMNs [3] contains some amendments at layer 2 - MAC modifications and a path selection protocol (layer 2 routing protocol) called Hybrid Wireless Mesh Protocol (HWMP) - but still relies on the legacy PHY with omnidirectional antennas. The incorporation of smart antenna features in IEEE 802.11s is quite challenging due to the mutual influence of the smart antenna R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 21–30, 2011. © Springer-Verlag Berlin Heidelberg 2011

22

M.I. Rafique, M. Porsch, and T. Bauschert

transmission scheme and the MAC mechanism. Furthermore the decision, which transmission technique to apply, strongly depends on the current mesh network topology and channel conditions, see [16]. For example, beamforming enables directional transmission with extended range thus being suitable for sparse network topologies whereas spatial multiplexing enables high bitrate omnidirectional transmission with lower range thus exploiting advantages in dense networks. Moreover, the diversity gain can be utilized to increase the reception range which might lead to improvements in non stationary networking scenarios. Thus, a transmitting node might influence a different number of neighboring nodes depending on whether beamforming, spatial multiplexing or diversity is applied. Hence it is advisable that MAC protocol, path selection mechanism and PHY layer transmission technique are operating in a well aligned manner. A number of mesh routing protocols exploiting some advantages of smart antennas in WMNs have already been proposed in the literature. But in our opinion they do not exploit the full potential of smart antennas transmission techniques wrt. dynamically changing networking environments and channel conditions. We present simple modifications in HWMP to incorporate the different capabilities of smart antennas in WMNs. These modifications enable mesh nodes equipped with smart antennas to choose different transmission schemes depending on the channel conditions. We consider spatial multiplexing to increase the transmission rate, diversity for range extension (of broadcast transmission) and beamforming for range extension (and also for link outage mitigation). We assume that VBLAST [18] is applied for spatial multiplexing and that diversity gain is achieved via simple space time block codes (STBC) [4]. Furthermore we determine the antenna gain for bidirectional beamforming according to the following formula, where M and N denote the number of antenna elements at the transmitter and receiver side [16]:

Garray =

(

M + N

)

2

(1)

As already stated above, the MAC protocol should be aligned to the PHY transmission scheme. However, most of the currently existing MAC layer amendments that have been proposed in different IEEE standards [3], [2], and [1] are agnostic to the PHY layer transmission scheme. For simplification, as an initial step we consider DCF at MAC layer and make some changes to cope with the different behaviors of the PHY transmission schemes of smart antennas. The paper is organized as follows: Section 2 describes the related work. Section 3 addresses the amendments of the MAC protocol whereas in Section 4 the HWMP modifications are explained. The results of our simulations are discussed in Section 5. Section 6 concludes the paper.

2 Related Work The MIR routing protocol [15] is the first one which incorporates some smart antenna features like spatial multiplexing and diversity. However MIR does not exploit benefits from beamforming. In [12] a routing protocol is presented that combines the advantages of multiplexing gain and interference cancellation. Lin et al. [11] proposes

Simple Modifications in HWMP for Wireless Mesh Networks with Smart Antennas

23

an algorithm that coordinates the scheduling, routing and power control mechanisms. Cai et al. [7] enhances an AODV-based routing protocol to consider spatial multiplexing at the PHY layer and focuses on the minimization of the route establishment time. Hu and Zhang [9] propose an extension of the IEEE 802.11 based MAC protocol that is aware of spatial diversity and analyzes its impact on the routing. To our knowledge, our work is the first one that considers the combination of multiple PHY layer transmission modes enabled by smart antenna technology and HWMP routing in an integrated approach for improving the performance of WLAN mesh networks.

3 MAC Amendments Before the data transmission with spatial multiplexing starts, we apply omnidirectional transmission for RTS/CTS exchange assuming that the transmission range of spatial multiplexing and single omnidirectional antenna transmission is the same. However, the RTS/CTS transmission with omnidirectional antenna is not suitable for data transmissions with beamforming. This is due to two reasons: The first reason is that due to the larger range of beamforming transmission a sender can cause interference to nodes that are beyond the default omnidirectional antenna transmission range. Secondly, due to the directional beamforming transmission a sender only interferes to those nodes that lie in the direction of the beam. To cope with the larger interference range of beamforming, MHWMP broadcasts RTS/CTS frames with STBC (S-RTS/S-CTS) prior to data transmission with beamforming - a similar approach is described in [14]. Here we assume that the range extension due to beamforming and diversity gain is approximately the same. The omnidirectional exchange of S-RTS/S-CTS might block more nodes from communication than necessary (thus decreasing the network throughput) as the directional beamforming transmission does not affect nodes that lie outside the beam. To avoid such situations we adopt the concept of a directional network allocation vector (DNAV) as suggested in [17]. The use of DNAV enables parallel communication of nearby located nodes with non-interfering beams.

4 Modified HWMP (MHWMP) The basic routing functions of MHWMP are the same as that of the original HWMP. The description of the original HWMP draft is beyond the scope of this paper, the interested readers are referred to [6]. Here we only outline our HWMP modifications. 4.1 Neighbor Discovery (ND) As the transmission range of beamforming and multiplexing is different and the former operates in directional mode, the neighboring nodes might be different for these two schemes. To exploit the advantage of beamforming and multiplexing, a node should have explicit informations about its respective neighbors. A number of neighbor discovery (ND) algorithms applicable for directional transmission can be found in the literature, see [13], [10], [8], [5] and [19], but so far there exists no suitable ND algorithm that can cope both with beamforming and

24

M.I. Rafique, M. Porsch, and T. Bauschert

multiplexing at the same time. In our new ND algorithm, each node has two neighboring tables - the Omnidirectional Neighbor Table (ONT) and the Directional Neighbor Table (DNT). In the DNT the neighbors are stored together with their respective directions. To find the directional and omnidirectional neighbors, beacons are broadcasted both with and without STBC respectively (similarly to S-RTS/S-CTS and normal RTS/CTS). Both beacons are transmitted alternately in 0.5 second intervals. We assume that the PHY layer is capable to differentiate both beacons and that a node can estimate the direction of arrival (DOA) from the received beacons. 4.2 Path Selection Metric Each node has two path tables i.e. an Omnidirectional Path Table (OPT) and a Directional Path Table (DPT). The former is used for communication with spatial multiplexing while the latter is considered for communication with beamforming. The path tables are calculated based on the minimum airtime link metric C:

B ⎤ 1 ⎡ C = ⎢Oca + O p + t ⎥ r ⎦ 1 − e fr ⎣

(2)

where Oca and Op are the channel access and MAC protocol overheads, Bt is the number of bits of a test frame and efr denotes the frame error probability. The DPT is calculated based on (2) by taking r as transmission rate without spatial multiplexing. r is determined by the specific rate adaptation algorithm RBAR [20] considering the channel conditions. The OPT is calculated based on (2) by using rmux instead of r. rmux denotes the transmission rate in case of spatial multiplexing and can be obtained from r as follows:

rmux = min(M , N )r

(3)

where M and N denote the number of antennas in the antenna array of the transmitter and receiver, respectively. 4.3 Path Discovery The path discovery procedure is depicted in Fig. 1. When a source S wants to communicate with destination D, first of all it looks in both path tables OPT and DPT. In case S does not find any entry for D, it broadcasts path request (PREQ) frames with STBC (S-PREQ) and without STBC (standard omnidirectional transmission) (OPREQ). The structure of both PREQ frames is the same as in HWMP. The destination only flag of the corresponding PREQ frames is not set (DO=0). Therefore an intermediate node responds with a Path Reply (PREP) frame to the PREQ frame if it has a valid path entry. When an intermediate node receives an S-PREQ it looks into its DPT and in case that an valid entry for the destination D exists, it replies back to the transmitter (with beamforming). Otherwise it updates its DPT and forwards the SPREQ provided that the sequence number of the S-PREQ is higher or, in case of equal sequence numbers if the airtime metric value of S-PREQ is lower than that of the corresponding path entry. If an intermediate node receives a O-PREQ it looks into it’s OPT and in case that an entry for the destination D exists it replies back to the

Simple Modifications in HWMP for Wireless Mesh Networks with Smart Antennas

25

transmitter (in omnidirectional mode). Otherwise it updates it’s OPT and forwards the O-PREQ if the conditions are met. If a valid entry for D exists only in OPT or DPT then the source starts the data transmission using either spatial multiplexing or beamforming and additionally initiates a path request to fill the missing entry in the respective path table. Thus, in case of the absence of a valid destination entry in one table, the source does not need to wait for the completion of the route discovery before beginning the transmission. When the destination node receives a new S-PREQ or O-PREQ it sends a Path Reply (PREP) on the return path towards the originator applying beamforming or omnidirectional transmission, respectively. While processing PREP frames, the role of the intermediate nodes is almost the same as explained in the original HWMP draft except that they update their respective path tables OPT or DPT and use the same transmission scheme for forwarding the PREP as that with which they received it. 4.4 Data Transmission Fig. 1 also explains the process of data transmission. Before transmitting a data packet to destination D, the source S compares the airtime metric value of the paths available in both path tables. It selects the next hop that belongs to the path with the lower airtime metric and starts its data transmission using the corresponding scheme (spatial multiplexing or beamforming). In case the destination entry in OPT has the lower path metric value, the node selects the next hop from the OPT and uses spatial multiplexing for transmission. If the destination entry in DPT owns the lower path metric value, the node selects the next hop from the DPT and sends the data packet with beamforming. Additionally it starts a O-PREQ in order to find a new path with possibly lower path metric - this is because applying spatial multiplexing for data transmission is always preferred due to the higher achievable data rate. The transmission with beamforming continues until a path with less airtime metric is found through this new request. The described method increases the overall throughput of the mesh network because the selected path always guarantees the highest transmission rate. If a valid destination entry exists only in OPT or DPT then the source initiates the data transmission using either spatial multiplexing or beamforming and additionally starts a path request to fill the missing entry in the respective path table. 4.5 Path Maintenance As we assume a stationary mesh network (with non-moving mesh nodes), mesh link outages might only occur due to bad radio channel conditions, i.e. poor SNR. In the original draft of HWMP, path interruptions due to link outages lead to the generation of path error (PERR) frames. In our algorithm, if the data frames were transmitted using spatial multiplexing, a node sends - after a link outage - a PERR frame towards the source applying omnidirectional transmission. It also looks in the DPT, whether a valid entry for the same destination exists and if so, the node continues the transmission of received packets on that path using beamforming. This approach avoids the loss of queued and on-air data frames. When the source receives the PERR frame, it erases the corresponding entry in the OPT and starts a path discovery procedure in

26

M.I. Rafique, M. Porsch, and T. Bauschert

Fig. 1. Path Discovery and Data Transmission of MHWMP

omnidirectional mode (O-PREQ). Meanwhile, the source uses the next hop entry out of the DPT and continues the transmission towards the destination with beamforming unless a better path is found through the O-PREQ. Intermediate nodes that receive a PERR frame delete the corresponding entries in the OPT and forward the PERR frame towards the source. If data frames are going to be transmitted with beamforming, a node sends - after a link outage - a PERR frame towards the source applying beamforming. Intermediate nodes that receive the PERR frame delete the corresponding entries in their DPT and forward the PERR towards the source accordingly.

5 Performance Evaluation The network simulator ns-3 is used for performance evaluation. MHWMP is compared with original HWMP that works with spatial multiplexing (O-MUX). For fairness reasons, O-MUX also uses standard RTS/CTS.

Simple Modifications in HWMP for Wireless Mesh Networks with Smart Antennas

27

We assume constant bit rate (CBR) traffic flows between source and destination nodes and a constant data packet size of 512 bytes. Each node is equipped with an uniform circular array comprising of four antenna elements. For the simulation of beamforming, we apply the keyhole model where the gain of the side lobes is set equal to zero. The maximum transmission rate for a single antenna element is 54 Mbps (according to IEEE 802.11a). For simplification, we assume that the nodes have perfect channel knowledge and packet losses occur only due to insufficient SNR. Moreover, it is assumed that the PHY layer is capable to differentiate between standard and STBC control frames. The maximum number of packets buffered in the output queue during route resolving is set to 255. After 3 successive PREQs reattempts, the destination is considered unreachable and the time duration between two successive PREQs is set to 102 ms. The same parameter values are mentioned in [3]. We consider a simple scenario where two nodes (source and destination) are placed at a distance of 50m. The life time of the active path and the path maintenance interval are set to 5s and 2s respectively. The total simulated time is 15s. The source generates traffic at 5Mbps and starts sending packets at 4s towards the destination. The SNR of the radio link is decreased by 10 dBm in the time interval between 8s and 11s. Fig. 2 shows the data throughput both for MHWMP and O-MUX. It illustrates that MHWMP shows robust behavior against link losses and maintains the throughput of the network even in situations where O-MUX fails.

Fig. 2. Data Throughput

In case of MHWMP, at the beginning, the destination entry in the sources' OPTs has a lower airtime metric compared to that in the DPTs. This is due to the higher PHY layer data rate of transmissions with multiplexing. Thus MHWMP chooses

28

M.I. Rafique, M. Porsch, and T. Bauschert

multiplexing for the transmission. When the link deteriorates at 8s, the high frame losses lead to a link outage. Then the source selects the available path in DPT and continues packet transmission with beamforming. At the same time it initiates a path discovery (for the path entry in OPT) by sending O-PREQ. However this path discovery remains unsuccessful in the time interval between 8s and 11s. At 11s, when the original SNR of the link is restored, the source successfully discovers a new path that is feasible for transmission with multiplexing (OPT). It compares the airtime metric with that of the DPT entry and finally selects multiplexing for the transmission because of the lower metric value. In the time interval from 8s to 11s O-MUX suffers with a complete link failure without any possibility to recover. However the source re-establishes the path when the link is recovered at 11s showing a temporary throughput peak. This can be explained as follows: during the link outage the source continuously generates traffic at the rate of 5 Mbps. When the path is re-established, most of the packets generated during the time of the link degradation already have been dropped from the output buffer due to its limited size. At 11s the buffered packets are sent out with maximum transmission rate of O-MUX which is much higher than 5 Mbps (the assumed source traffic rate) leading to the temporary throughput peak.

Fig. 3. Packet Time Delay

Fig. 3 explains the per packet delay behavior of O-MUX and MHWMP. As expected, both schemes show a small short delay peak at the beginning of the transmission due to the path discovery process. MHWMP and O-MUX exhibit similar time delay when the SNR of the link is not reduced. After the link deteriorates MHWMP successfully adapts the path from DPT and continues its transmission with

Simple Modifications in HWMP for Wireless Mesh Networks with Smart Antennas

29

beamforming. Since the transmission rate for beamforming is lower than for multiplexing and additionally a lower PHY layer transmission data rate is selected (due to the lower SNR value) by the rate adaptation algorithm, the packet delay is comparatively higher. At 11s, when the path is re-established, MHWMP switches to transmission with spatial multiplexing and resumes the low delay. O-MUX shows worse performance during the link deterioration period as no packets are sent but queued in the output buffer (or dropped if the buffer size is exceeded). As soon as the path is re-established at 11s, the source first transmits the buffered packets - this causes the delay time peak (for these packets). After the successful transmission of the queued packets, O-MUX again achieves the same packet delay time as MHWMP. Since it is assumed that the nodes have perfect channel knowledge, the degradation of the SNR can also be interpreted as caused by a larger distance between the two nodes. Thus by exploiting the benefits from beamforming transmission, MHWMP is expected to outperform O-MUX in sparse mesh network scenarios (with larger node distances) as well.

6 Conclusion In this paper we propose some modifications in HWMP for stationary wireless mesh networks with smart antennas. As a first step, we evaluated the performance of MHMWP by means of a simple two node scenario. It turns out that MHWMP improves the throughput and shows a quite robust behavior against link outages caused by SNR degradations. It exploits the advantages of smart antennas by adaptively using either spatial multiplexing or beamforming for data transmission and STBC for sending control packets prior to beamforming. Due to the different reach of these transmission techniques, MHWMP requires two separate path tables and also two separate sets of control frames. In our future work we will study the performance of MHWMP in medium and large scale wireless mesh network scenarios. Furthermore we intend to analyze the overhead caused by MHWMP in more detail - especially also for wireless mesh networks with non-stationary (moving) nodes. For stationary mesh networks this overhead is expected to be not a big hurdle as path failures and consequently path discovery actions normally do not occur frequently.

References 1. IEEE 802.11 Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, Amendment 8: Medium Access Control (MAC) Quality of Service Enhancements (2005) 2. IEEE P802.11n/D11.0, Part 11: Wireless Medium Access Control (MAC) and Physical layer (PHY) Specifications, Amendment 5: Enhancement for Higher Throughput (2009) 3. IEEE P802.11s/D3.0, Part 11: Wireless Medium Access Control (MAC) and Physical layer (PHY) Specifications, Amendment 10: Mesh Networking (2009) 4. Alamouti, S.M.: A Simple Transmit Diversity Technique for Wireless Communications. IEEE J. Selecl. Areas Commun. 16(8), 1451–1458 (1998) 5. An, X., Hekmat, R.: Self-adaptive Neighbor Discovery in Ad hoc Networks with Directional Antennas. Mobile and Wireless Communications Summit, 1–5 (2007)

30

M.I. Rafique, M. Porsch, and T. Bauschert

6. Bahr, M.: Proposed Routing for IEEE 802.11s WLAN Mesh Networks. In: 2nd Annual International Workshop on Wireless Internet, WICON 2006, New York, NY, USA (2006) 7. Cai, X., Li, J., Chen, D., Zhang, Y.: Cross-layer Design of Aodv-based Routing in Ad hoc Networks with MIMO Links. In: IEEE 20th International Symposium on Personal, Indoor and Mobile Radio Communications, pp. 2586–2590 (2009) 8. Cain, J.B., Billhartz, T., Foore, L., Althouse, E., Schlorff, J.: A Link Scheduling and Adhoc Networking Approach using Directional Antennas. In: IEEE MILCOM 2003, pp. 643–648 (2003) 9. Hu, M., Zhang, J.: MIMO Adhoc Networks: Medium Access Control, Saturation Throughput, and Optimal hop Distance. J. Communication and Networks, 317–330 (2004) 10. Haung, Z., Shen, C.C., Srisathapornphat, C., Jaikaeo, C.: Topology Control for Ad hoc Networks with Directional Antennas. In: Eleventh International Conference on Computer Communications and Networks, pp. 16–21 (2002) 11. Lin, Y., Javidi, T., Cruz, R.L., Milstein, L.B.: Distributed Link Scheduling, Power Control and Routing for Multi-hop Wireless MIMO Networks. In: IEEE Asilomar Conference on Signals, Systems, and Computers, pp. 122–126 (2006) 12. Mundarath, J.C., Ramanathan, P., Veen, B.D.V.: A Quality of Service Aware Cross Layer Approach for Wireless Adhoc Networks with Smart Antennas. J. Ad Hoc Networks 7(5), 891–903 (2009) 13. Ramanathan, R.: On the Performance of Adhoc Networks with Beamforming Antennas. In: ACM MobiHoc, pp. 95–105 (2001) 14. Rossetto, F., Zorzi, M.: On Gain Asymmetry and Broadcast Efficiency in MIMO Adhoc Networks. In: IEEE ICC 2006, pp. 3862–3867 (2006) 15. Sundaresan, K., Sivakumar, R.: Routing in Ad-hoc Networks with MIMO Links: Optimization Considerations and Protocols. J. Computer Networks 52(14), 2623–2644 (2008) 16. Sundersasan, K.: Network Protocols for Ad-Hoc Networks with Smart Antennas. PhD Thesis, Georgia Institute of Technology, USA (2006) 17. Takai, M., Martin, J., Bagrodia, R., Ren, A.: Directional Virtual Carrier Sensing for Directional Antennas in Mobile Adhoc Networks. In: 3rd ACM International Symposium on Mobile Adhoc Networking & Computing, MobiHoc 2002, New York, NY, USA, pp. 183–193 (2002) 18. Wolniansky, P.W., Foschini, G.J., Golden, G.D., Valenzuela, R.A.: V-BLAST: An Architecture for Realizing Very High Data Rates over the Rich-Scattering Wireless Channel. In: ISSSE 1998, pp. 295–300 (1998) 19. Zhang, Z., Ryu, B., Nallamothu, G., Huang, Z.: Performance of all-Directional Transmission and Reception Algorithms in Wireless Adhoc Networks with Directional Antennas. In: IEEE MILCOM 2005, pp. 225–230 (2005) 20. Holland, G., Vaidya, N., Bahl, P.: A Rate- Adaptive MAC Protocol for Multi-hop WirelessNetworks. In: MobiCom 2001, New York, NY, USA, pp. 236–251 (2001)

On the Evaluation of Self-addressing Strategies for Ad-Hoc Networks Ricardo de O. Schmidt1 , Aiko Pras1, and Reinaldo Gomes2 1

Design and Analysis of Communication Systems University of Twente, The Netherlands {r.schmidt,a.pras}@utwente.nl 2 Systems and Computing Department Federal University of Campina Grande - Brazil [email protected]

Abstract. Ad-hoc networks are supposed to operate autonomously and, therefore, self-* technologies are fundamental to their deployment. Several of these solutions have been proposed during the last few years, covering most layers and functionalities of networking systems. Addressing is one of the critical network operations on supporting others such as routing and even security of IP-based communcations. This paper has the goal of putting together diﬀerent strategies for self-addressing problem by evaluating ﬁve self-addressing protocols and pointing out their strenghs and drawbacks. At this time, we did not intend to come up with a novel approach for the problem, but to evaluate the existing ones in a non-isolated manner and also considering critical ad-hoc networking situations of partition and merging. Conclusions about the evaluated approaches and their applicability are drawn at the end of this paper.

1

Introduction

Autonomous networking systems represent a new paradigm mostly based on an innovative cooperation model of network devices to create and manage communication environments automatically. The idea of creating autonomous networking systems relies heavily on the concept of autoconﬁguration. The autoconﬁguration problem, also self-management, can be seen as the main reason for the emergence of self-* technologies. Due to its importance for the proper network operation, addressing can be seen as one of the most important (and challenging) issues in self-* solutions. According to documents, like [1], published within the IETF working group AUTOCONF [2], among the goals of autoconﬁguration in dynamic networks, like ad hoc networks, we must consider the conﬁguration of unique address for nodes. The well known DHCP (Dynamic Host Conﬁguration Protocol) has a very limited applicability when considering dynamic characteristics found in these networks, such as mobility, unpredictable number of nodes, and undeﬁned topology. Addressing is a critical functionality in IP networks given that it provides basic conﬁguration for other network functionalities such as routing and security. R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 31–42, 2011. c Springer-Verlag Berlin Heidelberg 2011

32

R. de O. Schmidt, A. Pras, and R. Gomes

Attempting to solve the addressing problem in dynamic networking systems, several self-addressing solutions were already proposed, like the ones surveyed in [3] and [4]. These solutions implement diﬀerent methodologies to allow selfconﬁguration of nodes interfaces with tentative of valid and unique addresses within the network. Such methodologies range from simple random selection of addresses from predeﬁned addressing spaces, to mathematical eﬀorts where equations are deﬁned to enable nodes to calculate their own interface(s) address(es). These solutions are roughly classiﬁed as stateless, stateful or hybrid approaches, discussed in more details in the next section. The main goal of this paper is to contribute to the speciﬁc area of addressing by comparing completely diﬀerent self-addressing approaches in critical ad hoc networking situations, like networks merging and partition. Past works have evaluated several strategies for self-addressing. However, most of them only submit the proposed solutions to simple network behaviors, what may not be the case in dynamic ad hoc networks. It is important to state that we do not intent to propose a novel approach for self-addressing, but to evaluate those we consider important methodologies for approaching the problem. With the results in this paper and also further experiments with speciﬁc functionalities (like merging/partition management), we plan to come up with a stable, scalable and reliable self-addressing solution, developing new ideas or using already existing ones that proved to be eﬃcient. The rest of this paper is organized as follow. The concepts of self-addressing are introduced in the next section. Section 3 presents the methodology used in the experiments as well as a brief description of the implemented and simulated self-addressing protocols. Results obtained from the simulations are presented in section 4. Finally, in section 5, conclusions about the evaluated strategies are drawn.

2

Self-addressing

Self-addressing is tightly related to the concepts of autoconﬁguration. It can be part of a set of technologies that enable a network to operate autonomously. Basically, a self-addressing protocol must provide a node the ability of or generating its own addressing conﬁguration or retriving such conﬁguration from another network entity. Self-addressing protocols can roughly be classiﬁed in one of the following categories: stateless, stateful or hybrid. In short, stateless approaches like Strong DAD [5] are those where nodes do not keep track of addresses in use in the network (i.e., the state of addresses is unknon). For instance, a stateless protocol can randomly select an address from a predeﬁned range and test it within the network to ensure that no other node is conﬁgured with such address. This testing procedure is generally named Duplicate Address Detection (DAD). Stateful approaches like SooA [6], on the other hand, allow nodes to be aware of addresses state. Usually, in a stateful approach, the nodes that keep track of addresses are responsible for performing addressing tasks on assignment and management of resources. For instance, information

On the Evaluation of Self-addressing Strategies for Ad-Hoc Networks

33

about addresses can be stored in tables. However, alternative strategies like Prophet Allocation [7] deﬁne mathematical approaches which allow nodes to track addresses that may be in use, thus characterizing a stateful solution. Finally, hybrid approaches mix properties of both stateless and stateful protocols. Hybrid solutions usually implement random address selection, DAD testing, and the address registration within one or more addressing authorities (like in HCQA [8]) or within tables spread through the network (like in MANETconf [9]).

3 3.1

Experiments Evaluated Protocols

Five self-addressing protocols were compared in our experiments. Three stateless protocols: Strong DAD, AIPAC and AROD; one stateful: Prophet Allocation; and one hybrid: MANETconf. The protocols AROD and AIPAC were fully evaluated. That is, we also considered their mechanisms for handling networks merging and partition. All protocols were implemented in C++ and added to the well known network simulator NS2 (Network Simulator 2) [10]. We opted for these protocols due to their diﬀerent proposals on addressing assignment, DAD procedures and partition/merging management. In our experiments we aimed a fair comparison among the ﬁve protocols, following AUTOCONF’s [2] goal of proposing a common and general self-addressing solution for ad-hoc networks. The protocols were submitted to the same network scenarios, disregarding their implementation (or not) of advanced operations for handling complex networking situations. Next, a brief description of each evaluated protocol is given. Strong DAD [5] is a stateless addressing protocol which implements random selection of addresses followed by DAD procedures. Each starting node is responsible for its own address conﬁguration. Upon starting a node selects randomly, from a predeﬁned range, two addresses: temporary and tentative addresses. The former is used as the node’s address during the DAD procedure on behalf of the latter. On succeding in the DAD procedure, the tentative address is used to conﬁgure the node’s interface. Otherwise, the tentative addresse selection and testing is done until no conﬂicts are identiﬁed. Strong DAD does not deal with network partition and merging. Consequently, its applicability to more critical networking scenarios is restricted. AIPAC [11] implements the concept of requester and initiator nodes. The former is a starting node and the latter an already conﬁgured node which negotiates an address within the network on behalf of the former. As well as in Strong DAD, a starting node randomly selects temporary address and uses it to communicate with its initiator node during the conﬁguration procedure. This temporary address is used until a negotiated one is received from the initiator. The negotiation procedure is executed by the initiator node. It randomly selects an address from a predeﬁned range and tests it with the network through a DAD procedure. AIPAC also deﬁnes functions for dealing with partition and

34

R. de O. Schmidt, A. Pras, and R. Gomes

merging of networks. A network identiﬁcation, deﬁned by the ﬁrst node (i.e., network’s founder), is added to the protocol’s messages, enabling nodes to identify diﬀerent networks that come into each others range. AIPAC’s main drawback is the dependancy on routing messages and tables. It requires modiﬁcations to the routing protocol operating in the network, making AIPAC unsuitable for dynamic networks that may even negotiate and change the routing protocol. AROD [12], although being also based on DAD procedure and requesterinitiator scheme, focuses on reducing conﬁguration time and control overhead. To do so, it implements address reservation and optimistic DAD. The main diﬀerence is that an initiator node may have reserved addresses, previously validated within the network, which can be allocated to a requester node without the DAD delay, reducing the conﬁguration time. A DAD procedure is usually executed for more than a single tentative address at the same time, which allow nodes to obtain spare addresses for future conﬁgurations of requester nodes. Prophet Allocation [7] is the only evaluated approach in this paper that does not implement DAD procedure (at least not during the procedures speciﬁed by the authors). Unlikely the other described solutions, Prophet Allocation does not select addresses randomly, but it obtain addresses inside a predeﬁned range from a formula. This formula is based on the fundamental theorem of arithmetic, which says that: every positive integer may be expressed uniquely as a product of primes. Each node has a 2-tuple identiﬁer, composed by their address and the formula’s state. Everytime an address is calculated through the formula, its state is incremented by one. However, as stated by the authors, this formula might generate duplicated addresses within a network. Therefore, when a node starts (requester) it receives from an already conﬁgured node (initiator) an address, generated by the formula, and a network identiﬁer. This identiﬁer is added to the protocol’s messages enabling nodes to identify messages from diﬀerent networks (merging situation). Unfortunately, no additional information on how the protocol deals with the merging problem was provided by the authors in [7]. Consequently, due to this the lack of information, we did not considered management of network partition and merging as Prophet Allocation function. Finally, MANETconf [9] is a completely distributed and hybrid addressing protocol, which implements allocation tables and DAD procedure. Its functionality is simple but it requires a lot of broadcasting for DAD and tables synchronization. MANETconf also implements requester and initiator nodes. The diﬀerence is that after successfully negotiating an address to a requester node, the initiator ﬂoods the network announcing that such address has been allocated. With this ﬂooding, all network nodes update their respective allocation tables. The allocation tables are used to increase the reliability of the DAD procedure. Despite the fact of ﬂooding the network with broadcasts, MANETconf has the advantage of not depending on any other technology. 3.2

Scenarios and Evaluation Metrics

The experiments were planned to evaluate the implemented protocols under critical situations of ad hoc networks. The protocols were set to operate with very

On the Evaluation of Self-addressing Strategies for Ad-Hoc Networks

35

limited addressing resources (256 available addresses) and submitted to conditions of random deployment and mobility of nodes, leading to situations of networks isolation, partition and merging. Two diﬀerent scenarios were simulated, as described below. In the ﬁrst scenario, named Scenario A, 100 nodes ramdonly positioned in a predeﬁned area, forcing situations of nodes isolation in the beginning of the simulation, due to geographical distance between nodes. In this scenario nodes are static and merging was resulted from the deployment of intermediate nodes between the isolated networks. In the second scenario, named Scenario B, two initially isolated 10-node networks were deployed and later merged, resulting in a 20-node network. The area of these networks was deﬁned so that no isolated nodes were deployed in the initial formation. Simulation time and nodes arrival rate did not play an important role in our scenarios. In addition, from previous experiences, like [6] and [12], we learned that variations on addressing space have a very low impact on protocol’s operation. We made use of a methodology described in [13] to determine the necessary number of simulation runs so that our results would fall suﬃciently close to the mean, in a conﬁdence interval of 95%. This guarantees that we perform a fair comparison between two or more solutions mainly due to the use of random variables (i.e., nodes positioning and random mobility). To evaluate the protocols, we considered three basic metrics: (a) control overhead, which quantiﬁes the data generated and transmitted during the protocols operation; (b) conﬁguration delay, which represents the time between the node deployment and its ﬁnal conﬁguration; and (c) address uniqueness, which measures the eﬃciency of protocols on avoiding problems with duplicated conﬁgurations (i.e., conﬂicting conﬁgurations).

4 4.1

Results Analysis Scenario A: Highly Populated Network

The graphics in Fig. 1 present the traﬃc generated and transmitted by the protocols while conﬁguring all 100 nodes. However, the lines do not represent all the traﬃc generated during the entire simulation time, which was 2000 seconds, but the traﬃc from the beginning of the simulation until the conﬁguration of the last node. Regarding the traﬃc in number of packets, as presented in Fig. 1(a) and 1(b), and the traﬃc i nnumber of bytes, as presented in Fig. 1(c) and 1(d), one can conclude in a ﬁrst moment that the best performances were reached by the protocols Strong DAD and AROD. However, we must consider that the former implements only allocation of addresses and no further maintenance operations. When the last node is conﬁgured with Strong DAD, the protocol concludes its operation. Therefore, all the traﬃc generated by the protocol is related to address assignment and DAD procedure. Thus, if after conﬁguring all nodes, n nodes happen to be conﬁgured with duplicated addresses, such problem will remain unsolved until the end of the simulation or, in another possible scenario,

36

R. de O. Schmidt, A. Pras, and R. Gomes

,$$$$ *$$$$ ($$$$ &$$$$ $

% %$ &$ '$ ($ )$ *$ +$ ,$ -$ %$$ !

% %$ &$ '$ ($ )$ *$ +$ ,$ -$ %$$ !

%&$$

%&$$

%$$$

%$$$

%$$$$$

($$$$ ')$$$ '$$$$ &)$$$ &$$$$ %)$$$ %$$$$ )$$$ $

!

!

,$$ *$$ ($$ &$$

,$$ *$$ ($$ &$$

$ % %$ &$ '$ ($ )$ *$ +$ ,$ -$ %$$ !

$ % %$ &$ '$ ($ )$ *$ +$ ,$ -$ %$$ !

Fig. 1. Protocols performance in Scenario A

only one of the conﬂicting nodes remains active in the network. Even being a ﬂood-based approach, Strong DAD generated signiﬁcantly less traﬃc than the other approaches. It is due to the fragmentation of the network resulted from the random positioning of nodes in the simulation area, where the less neighbors a node has the smaller is the impact caused by its ﬂooding procedure. In addition, the number of bytes transmitted by Strong DAD grows in diﬀerent rates than the number of packets. It occurs because in the DAD procedure a list of previous hops (reverse path) must be kept within a request message for allowing a node, which identiﬁes a conﬂict, to reply correctly to the request generator. It is also applicable to other protocols that implement DAD procedure with reverse path. As one can observe in Fig 1, regarding traﬃc overhead, AROD performed better that Strong DAD, even executing DAD procedures. The main reason for that is the address reservation implemented by AROD, where for testing several addresses with DAD, the protocol generates pratically the same overhead than Strong DAD would generate for testing a single address. With the address reservation, AROD avoids further DAD executions when conﬁgured nodes provide new ones with already tested addresses. It is important to inform that in our experiment nodes operating AROD were able to reserve only one address in addition to the address for their own conﬁguration. One can infer that the protocol performance may improve if, in cases that the addressing space is larger, the nodes are allowed to reserve more than one address for further allocations.

On the Evaluation of Self-addressing Strategies for Ad-Hoc Networks

37

Prophet Allocation, which does not implement DAD procedure on address allocation, did not generated high overhead to conﬁgure nodes. The simple handshake procedure for allocating an address allows the protocol to quickly conﬁgure new nodes. Given that in this experiments we did not considered cooperation between addressing and routing protocols, the strategy for address maintenance implemented by Prophet Allocation was the responsible for the most overhead. The periodically ack broadcasted by all nodes was the reason for the high values achieved by Prophet Allocation regarding the control overhead. The diﬀerence between the sent and received values achieved by the protocol is due to the number of neighbors each node has, where the higher the number of neighbors, the higher is the number of received packets/bytes. In addition, due to the small payload of ack messages, the numbers of packets is much higher than the number of bytes, as observed in Fig. 1. The traﬃc generated by AIPAC when conﬁguring nodes is similar to Strong DAD. However,the exceeding traﬃc generated by AIPAC, as observed in Fig. 1, is due to the execution of merging management procedure called "gradual merging". As one can observe in Fig. 1, in general MANETconf generated most traﬃc when conﬁguring all nodes. It is due to the excessive ﬂooding during the procedures of DAD and tables synchronization. The need of all nodes replying a DAD request received generates a very high control overhead, which is summed with the broadcasted message used to update allocation tables on every new address that is allocated. Due to post-allocation address maintenance, Prophet Allocation, AROD and AIPAC continued to generate traﬃc even after concluding the conﬁguration of all nodes. However, simulation parameters were set (e.g., nodes arrival rate) so that the conﬁguration of the last node happened close to the 2000 seconds. This way, the additional traﬃc did not impact signiﬁcantly in the ﬁnal results. Fig. 2 illustrates the average delay achieved by the the protocols for conﬁguring the nodes in Scenario A. Strong DAD usually has an average of 15 seconds for nodes conﬁguration, due to the three rounds of 5-second DAD procedure. However, given that several nodes were isolated in the beginning of the simulation, Strong DAD conﬁgured such nodes without executing the DAD procedure and, consequently, the average delay at the end of the simulation dropped to 9 seconds.

!"$%

+- +(

*-

*(

)- )(

-

( )

)(

*(

+(

,( -( .( /( !

0(

1( )((

Fig. 2. Conﬁguration delay

38

R. de O. Schmidt, A. Pras, and R. Gomes

The conﬁguration delay in MANETconf is strictly connected to the duration of the DAD procedure. Consequently, the average delay achieved by MANETconf was 5 seconds. In this case, the time for updating the allocation tables was not considered as part of the conﬁguration delay because such procedure is only triggered after the new node is already conﬁgured. The merging procedure implemented by MANETconf solves conﬂicts locally and, consequently, it did not impacted signiﬁcantly on the protocol’s performance. Unlike MANETconf, the highly fragmented network had a strong impact on AROD operation. The average conﬁguration delay achieved by AROD was 8 seconds, even operating with the address reservation scheme. The control overhead for addressing maintenance and the merging management in AROD compromised its performance. At the end, the reliable mering aimed by AROD spent much time and generated excessive traﬃc. Prophet Allocation achieved the best results reagarding conﬁguration delay due to its simple handshake for address allocation. However, we must consider that the protocol’s authors assume that the implemented mathematical equation may generate duplicate addresses in a certain moment. One can infer that by executing procedures for conﬂicts correction, which would be a ﬂooding-based DAD, a much higher maintenance load would be demanded from the nodes, what would result in performance degradation on overhead and conﬁguration delay. In this case, we believe that the averages presented as results in Fig. 2 for Prophet Allocation would be increased in 5 seconds. AIPAC achieved the highest conﬁguration delay. Its DADi procedure lasted 15 seconds in average. Two factors mainly contributed for such results. First, although merging problems were corrected later by the protocol, the procedures for gradual merging were unnecessarily triggered several times during nodes conﬁguration. The second reason is that AIPAC, unlike the other protocols, do not consider networks of one node. It means that an isolated node does not conﬁgure itself if it does not have at least one neighbor. Therefore, isolated nodes in the beginning of the simulation did not conﬁgured themselves until another node was deployed in their neighborhood. This waiting time increased the average conﬁguration delay for the ﬁrst nodes, as shown in Fig. 2. A serious problem identiﬁed on the results of this experiment was the number of nodes, at the end of the simulation, that were conﬁgured with conﬂicting addresses. Both netowrk fragmentation and reduced addressing space drastically impacted on the conﬁguration consistency. As presented in Table 1, all evaluated protocols concluded their operations with conﬂicted nodes conﬁguration. The weak performance of AIPAC, if compared to the other protocols, was rewarded by the success achieved on merging a highly fragmented network and ﬁnishing the simulation with the lowest number of conﬂicts. When compared to the other protocols, MANETconf also achieved good results on the number of conﬂicts. It is due to the allocation tables implemented in all nodes. As well as with AIPAC, the costly performance of the protocol was justiﬁed by the reliability of its conﬁgurations at the end of the simulations.

On the Evaluation of Self-addressing Strategies for Ad-Hoc Networks

39

Table 1. Address Conﬂicts in Scenario A Protocol

Conﬂicts

StrongDAD Prophet Allocation MANETconf AROD AIPAC

29 45 8 30 5

Strong DAD conﬁgured initially isolated nodes and later did not handle merging situations by locating and correcting possible conﬂicts. Consequently, the number of conﬂicts was very high with Strong DAD. Prophet Allocation achieved the highest number of conﬂicts at the end of the simulation. Given that it does not implement a procedure for handling merging, the fragmented network was also one of the reasons for its bad performance. In addition, Prophet Allocation allows an implementation of more states in its equation, which would generate more distinct sequences of addresses and, according to its authors, less conﬂicts. However, we decided for implementing its equation with only one state due to our goal of testing the protocols in very limited conditions. Although AROD ended the simulation with high number of conﬂicts, it is important to mention that the protocol’s procedure for conﬂict resolution was not implemented in this experiment. However, AROD successfully identiﬁed all the conﬂicts in the network. Such conﬂicts were resulted from the merging of all isolated nodes and small networks. Considering that AROD implements DAD procedure for resolving address conﬂicts, triggering such operation would degrade the protocol’s performance by increasing control overhead and conﬁguration delay. This leads us to conclude that it would bring AROD’s results close to AIPAC’s. 4.2

Scenario B: Handling Networks Merging

Past works have evaluated self-addressing protocols under more controlled networking scenarios, considering mainly the allocation procedure and not the resources maintenance and management after allocation. However, real ad-hoc network scenarios may be quite diﬀerent. In these, nodes arbitrarily come and go. In some situations not only the nodes move, but entire networks may move from one place to another, temporarily or permanently merging with each other. In this second scenario, we tested the merging procedures of AROD and AIPAC in a scenario of merging between two networks. Each network was composed by 10 nodes. In a predeﬁned moment, after the initial conﬁguration of all nodes, network N1 moved towards network N2, merging and creating a single network N3 with 20 nodes. Both protocols AROD and AIPAC successfully performed the network merging. The diﬀerence between the strategies implemented by the protocols for handling merging is clearly observed on the results presented in Fig. 3. The conﬂict

40

R. de O. Schmidt, A. Pras, and R. Gomes

Fig. 3. Merging traﬃc Vs. Allocation traﬃc

resolution procedures were not considered in this experiment, but seem it is seen as future work. As soon as AROD identiﬁes that two networks start to overlap, the mechanism for merging is triggered and a single network is quickly formed. The merging procedure in AROD aﬀects all the nodes in one of the overlapping networks. On identifying that it is necessary to join the other network, a node announce the merging to its network’s leader and the latter ﬂoods the network with a reconﬁguration message. It is a quick process and does not generate high control overhead. As one can observe in Fig. 3, in average only 5% of the total traﬃc generated by AROD was resulted from the merging procedure. On the other hand, AIPAC needs more time for executing the merging of networks. In addition, as illustrated in Fig. 3, if compared to AROD, a much higher percentage of its traﬃc is related to the merging procedure. After successfully merging the two networks, the traﬃc related to merging accounted for 80% of the total traﬃc generated by the protocol. It was due to the gradual merging strategy implemented in AIPAC, where some nodes migrated to the other network and later migrated back to their original network. Therefore, such situations doubled the traﬃc generated by these nodes during the merging procedure. Regarding the procedure delay, AROD detected and performed merging faster than AIPAC. AROD started the merging procedure as soon as the ﬁrst nodes identiﬁed the networks overlap, what happened in averages. AROD needed in average 40 seconds for completing the merging process. With the gradual merging strategy, AIPAC took longer for detecting the merging and concluded the procedure with and average delay of 94 seconds.

5

Conclusion

Several self-addressing solutions were already proposed for ad-hoc networks. From experiments, like the one presented in this paper, we can observe that

On the Evaluation of Self-addressing Strategies for Ad-Hoc Networks

41

such solutions still lack on providing alternatives for handling complex networking situations. Consistency of addresses is important due to its inﬂuence on other operations like routing. However, designed strategies for dealing with this issues result in problems with excessive control overhead and degradation on performance of basic addressing. In addition, dependency on other technologies is not a good approach for self-addressing protocols when considering Future Network scenarios. For instance, addressing solutions like AIPAC and Prophet Allocation, which depend on routing protocols, even requiring modiﬁcations on routing operations, have their applicability drastically limited to scenarios with the speciﬁc routing protocol. It is even more problematic if we consider scenarios where the routing is also dynamic and two or more routing protocols may coexist and cooperate. This paper presented the ﬁrst steps on evaluation of self-addressing approaches. As future work, the authors plan to carry more extensive and complete simulations with self-addressing solutions, also covering other network scenarios. Our main goals are: (a) to contribute with the decision on the best self-addressing solution (or combination of two or more solutions) for diﬀerent ad-hoc networks; (b) to develop a general evaluation framework for self-addressing solutions, based on AUTOCONF [2] documentation on requirements and guidelines; and (c) to propose an alternative self-addressing solution for SooA [6], which is going to be used on the absence of addressing servers, providing nodes with temporary conﬁguration.

References 1. Baccelli, E.: Address Autoconﬁguration for MANET: Terminology and Problem Statement. IETF Internet Draft (2008) 2. AUTOCONF: Ad-Hoc Network Autoconﬁguration. IETF Working Group, http://www.ietf.org (accessed in April 2011) 3. Bernardos, C., Calderon, M., Moustafa, H.: Ad-Hoc IP Autoconﬁguration Solution Space Analysis. IETF Internet Draft (2008) 4. Weniger, K., Zitterbart, M.: Address autoconﬁguration in mobile ad hoc networks: current approaches and future directions. IEEE Network Magazine, Special Issue on Ad Hoc Networking: Data Communications & Topology Control 18(4), 6–11 (2004) 5. Perkins, C.E., Malinen, J.T., Wakikawa, R., Belding-Royer, E.M., Sun, Y.: IP Address Autoconﬁguration for Ad Hoc Networks. IETF Internet Draft (2001) 6. de Schmidt, R.O., Gomes, R., Sadok, D., Kelner, J., Johnsson, M.: An Autonomous Addressing Mechanism as Support for Autoconﬁguration in Dynamic Networks. In: Proceedings of the Latin American Network Operations and Management Symposium, LANOMS (2009) 7. Zhou, H., Ni, L.M., Mutka, M.W.: Prophet Address Allocation for Large Scale MANETs. In: Proceedings of the 22nd Annual Joint Conference of IEEE Computer and Communication Societies (INFOCOM), vol. 2, pp. 1304–1311 (2003) 8. Sun, Y., Belding-Royer, M.E.: Dynamic Address Conﬁguration in Mobile Ad Hoc Networks. Technical Report 2003-11, University of California at Santa Barbara (2003)

42

R. de O. Schmidt, A. Pras, and R. Gomes

9. Nesargi, S., Prakash, R.: MANETconf: Conﬁguration of Hosts in a Mobile Ad Hoc Network. In: Proceedings of the 21st Annual Joint Conference of IEEE Computer and Communication Societies (INFOCOM), vol. 2, pp. 1059–1068 (2002) 10. Fall, K., Varadhan, K.: The ns Manual (formerly ns Notes and Documentation). The VINT Project (May 2010), http://www.isi.edu 11. Fazio, M., Villari, M., Puliaﬁto, A.: AIPAC: Automatic IP Address Conﬁguration in Mobile Ad Hoc Networks. Elsevier Computer Communications (COMCOM) 29(8), 1189–1200 (2006) 12. Kim, N., Ahn, S., Lee, Y.: AROD: An address autoconﬁguration with address reservation and optimistic duplicated address detection for mobile ad hoc networks. Elsevier Computer Communications (COMCOM) (30), 1913–1925 (2007) 13. Jain, R.: The Art of Computer Systems Performance Analysis: Techniques for Experimental Design, Measurement, Simulation and Modeling. Wiley-Interscience, New York (1991); ISBN 0471503361

Considerations in the Design of Indoor Localization Systems for Wireless Sensor Networks Jorge Juan Robles Chair for Telecommunications Dresden University of Technology, Dresden, 01069 Germany [email protected]

Abstract. The design of an indoor sensor network to support localization-based services is a challenging issue in that the protocols and localization algorithms have to be adapted to the capabilities of resources-constrained sensor nodes. Furthermore, the achieved position accuracy depends on many factors like the amount of signaling and quality of inter-node measurements, as well as the mathematical algorithm used for estimating the position. In this paper we present experiences and considerations in the design of such systems for IEEE 802.15.4 indoor sensor networks. Keywords: localization algorithm, position estimation, RSSI, wireless sensor networks, ranging protocols.

1 Introduction A Wireless Sensor Network (WSN) consists of many sensor nodes that gather information from the environment enabling the creation of interesting applications like animal monitoring, control of industrial processes and home automation. The position information plays an important role in many context-based applications, where the measurements have to be accompanied by the node position to make this information useful. Perhaps, the simplest way to know the position is to include a GPS receiver in the sensor node. Unfortunately, this is expensive in terms of energy consumption. Furthermore, the GPS signal cannot be detected correctly in indoor scenarios. For these reasons, alternative localization systems have to be designed for low-power sensor nodes. In general, we can consider a localization system as the integration of three main functional blocks: the data management, the localization algorithms and the network protocols. The data management refers to the acquisition and processing of the data necessary for the position estimation, like the connectivity, distance and motion information. The localization algorithms represent the mathematical methods used for the position estimation, e.g. the well-known Multilateration [21], Weighted Centroid Localization [22] or proposals based on particle filter. Finally, the networks´ protocols enable the communication required by the localization process, for example the routing and neighbor discovery protocols. Note that each functional block interacts R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 43–53, 2011. © Springer-Verlag Berlin Heidelberg 2011

44

J.J. Robles

with each other. If the goal is to keep the complexity as low as possible and additionally achieve acceptable position accuracy, optimizations in the three functional blocks are recommended. At the Chair for Telecommunications of the Dresden University of Technology, different research works related to localization in WSN have been conducted [1-14]. This paper presents an overview of the main conclusions obtained from these works and provides a guideline for the design of localization systems in 802.15.4 sensor networks. An indoor scenario is taken as reference in our work. Here, there are fixed sensor nodes with known positions called Anchors (ANs). They are external powered. The Mobile Nodes (MNs), whose positions are unknown, are battery powered. Our interest is to determine their positions and extend the lifetime of their batteries as much as possible.

2 Data Management Mostly, IEEE 802.15.4 transceivers have the capability to measure the received signal strength without the necessity of additional hardware. In this section we briefly discuss the suitability of this data to be used by localization algorithms. 2.1 Signal Strength Measurements for Localization The Received Signal Strength Indicator (RSSI) is an inter-node measurement and can be used as a proximity indicator in localization algorithms as well as to determine the link quality between nodes. The standard IEEE 802.15.4 does not clearly specify the format of RSSI. Thus, this indicator can change depending on the device manufacturers. In all our experiments we use the sensor node RCB230 from Dresden Elektronik [20]. It contains the low-power 802.15.4 transceiver AT86RF230 [18] operating at 2.45 GHz and the 8-bits microcontroller ATmega1281 [19]. As the RSSI indicator we use an internal register of the transceiver called Energy Detection (ED), which averages the signal strength measurements over 128 µs. This register can take 85 different values and has a resolution of 1dbm. The RSSI measurements are influenced by the antenna characteristics, temperature, obstacles, the human interaction and the multipath effect. Due to the fact that RSSI measurements present fluctuations, it is advisable to take several RSSI samples and apply statistical methods. Usually, the average value over a certain number of RSSI samples is used in localization algorithms. However, in cases with high dispersion, the median can be very useful because it can filter out unusual values or outliers in the measurements to obtain a stabile value. In order to analyze the spatial correlation of the RSSI, signal strength measurements were taken at defined distances in five different conditions. In Fig. 1 the first two tests were carried out in an outdoor scenario. The first one was conducted without the influence of persons (Outdoor 1), whereas in the second someone held a node while the measurements were taken (Outdoor 2). The next three tests in Fig. 2 were carried out in an indoor scenario (floor) without the influence of persons, where the height of the nodes took three different values (h=0m; h=10cm; h=1.2m). In each

Considerations in the Design of Indoor Localization Systems for WSNs

45

point we averaged more than 400 measurements. The error bars show the minimum and maximum value. In Fig. 1 the error bars in Outdoor 2 indicate a high dispersion in the measurements as compared to Outdoor 1. This is principally due to the contact of the person with the node. Furthermore, a big difference between the curves in all tests is clearly detected. In multiple cases, there is a difference of more than 20dBm for the same distance in different tests. This means that if only a general attenuation model is used for representing the signal strength in all the scenarios, then the error in the distance estimation is very high. Particularly in indoor scenarios, the multipath effect makes it difficult to estimate the distance because of the ambiguity of distance values for some RSSI regions. Our experiences show that RSSI is a poor indicator for the distance determination. In some cases, the RSSI could be used if the stationary characteristic of the scenario is taken into account and a previous calibration is carried out. Distance / m -28

1

5

13

20

30

40

RSSI / dBm

-38

50

60

Outdoor1 h=1.2m Outdoor2 h=1.2m

-48 -58 -68 -78 -88 -98

Fig. 1. Relationship between RSSI and distance in outdoor scenarios Distance / m

RSSI / dBm

-28

1

5

13

20

30

40

50

60

-38

Indoor h=0cm

-48

Indoor h=10cm Indoor h=1.2m

-58 -68 -78 -88 -98

Fig. 2. Relationship between RSSI and distance in indoor scenarios

46

J.J. Robles

3 The Localization Algorithms 3.1 Distance-Based Localization Algorithms Many localization algorithms make use of the distance to reference nodes to estimate the position. There are algorithms like Multilateration [21][15][12] that provide the exact coordinates of the blind node when the information of the distances and the position of the anchors do not contain errors. On the other hand, there are algorithms that approximate the position of the node although there is no error in the estimated distance and AN´s positions. Two well-known algorithms of this last group are MinMax [21] and WCL [22]. One of their advantages is the fact that they are more robust against distance errors compared to other exact methods. Given the low complexity of Min-Max and WCL, they can be easily implemented on sensor nodes. The position accuracy of Min-Max and WCL decreases considerably when the blind node are outside of the region bounded by the anchors of the periphery. In fact, these algorithms tend to locate their estimation inside the above mentioned region although the blind node is outside of the region. Therefore, the further the blind node is with respect to the internal region, the higher the position error is obtained. On the contrary, Multilateration is able to provide the coordinates of a blind node located outside of the region limited by the anchors. By using simulation, we analyzed the average position error in distance-based localization algorithms given different distance errors (Fig. 3). The 2D reference scenario is a square of dimension 40mx40m, where four anchors are placed at the corners. The estimated distance is modeled by adding a zero mean gaussian random variable to the real distance. The standard deviation defines the distance error and it is given as a percentage of the real distance. The MN takes different locations inside the scenario, where its position is estimated. More than 1500 estimations are averaged in each position. Trilateration uses only the information of three distances from ANs to estimate the position. Multilateration exploits the available distances for the position estimation. From Fig. 3 we can see that WCL and Min-Max have a better performance than lateration-based algorithms when the standard deviation is higher than 10%. Finally, it is important to note, that the relative position of the ANs impacts on the achieved position accuracy. For instance, consider a worst case where three ANs are collinear. If a MN requires calculating its position by using the distances to the ANs, then there is an ambiguous result because the MN cannot decide between two possible position estimations. Two important indicators that characterize the geometry of the ANs are the Cramer Rao Low Bound (CRLB) and Geometric Dilution of Precision (GDOP). They can be used to determine the optimal location of the ANs inside the scenario. For further information see [15]. 3.2 The Fingerprint Technique The “fingerprint technique” is a method often used in indoor scenarios given the fact that it takes the stationary characteristics of the scenario into account (e.g. the attenuation due to walls). This technique consists of two phases: the calibration phase and the online phase. During the calibration phase, RSSI measurements are taken at

Considerations in the Design of Indoor Localization Systems for WSNs

47

Average Position Error / m

25 WCL 20 15

Min-Max Trilateration Multilateration

10 5 0 5

10

20

40

Distance error %

Fig. 3. Robustness of localization algorithms against the distance error

known positions. These positions, together with their related RSSI measurements, represent a “fingerprint” inside the scenario. The fingerprints are saved in a database. In the online phase, which is the normal operation of the system, a blind node takes a new set of RSSI measurements. By comparing the online measurements with the stored fingerprints it is possible to approximate the position of the blind node. The position accuracy achieved by the fingerprint technique is in the range of 2-4 meters in an indoor scenario [23][17]. By using simulation we compare a low-complexity fingerprint-based algorithm with WCL in a 3D scenario [1]. Our results show that the fingerprint technique can improve the position accuracy although it is more sensitive to the RSSI dispersion than WCL.

4 Network Protocols The standard IEEE 802.15.4 [16] has been designed for low-rate sensor networks and defines the PHY and MAC layer of the communication model. In this section we analyze the more relevant features of this standard that impact the localization systems. 4.1 MAC Layer in 802.15.4 In 802.15.4, CSMA-CA is used for the communication between two nodes. The main drawback of the CSMA process is that the transceiver consumes extra energy when it waits a random time (backoff delay) and checks if the channel is idle or busy during the Clear Channel Assessment (CCA). If the channel is free the MN can transmit, if not, it has to repeat the aforementioned process. After a certain number of unsuccessful attempts, the MAC layer notifies the application layer by sending a Channel Access Failure (CAF). The application layer can discard the packet or try a retransmission. For further information about the operation of CSMA, please see [16].

48

J.J. Robles

By using an oscilloscope we measured the voltage in a resistance (2 Ohms) in series between the battery (2.8V) and the sensor node. Thus, it was possible to investigate the energy consumption during the transmission process [3]. The Fig. 4 shows a snapshot, where the microcontroller is always active and the transceiver wakes up to transmit a packet. Note, that it performs three CCA before the effective packet transmission. After this task, the transceiver goes into sleep mode again. Here, the duration of the transmission is about 1ms while the CSMA process takes more than 10ms. Thus, in these cases the whole transmission process can mean high energy consumption, especially if the channel is frequently busy. In a previous work of the authors [3] an analytical model for 802.15.4 sensor network is proposed, which considers the influence of the backoff delay on the energy consumption of the transmissions. For example, assume that each MN randomly broadcasts 8 packets of 1ms during a period de 300ms. Between the transmissions the MN sleeps, saving energy. The ANs take signal strength measurements, which are used for the MN´s position estimation. The Fig. 5 shows the expected MN´s energy consumption during the period. The values of the energy consumed in the different MN´s operation modes are taken from [3]. Additionally, Fig. 5 depicts the probability that the CSMA process is successful, leading to an effective transmission. The negative impact of a high number of MNs on both variables can be clearly identified from the figure.

CSMA CCA

CCA

Fig. 4. CSMA process during one transmission

Tx CCA

Considerations in the Design of Indoor Localization Systems for WSNs

1

6

0,95

5

49

4

0,85

3

0,8 0,75

mJ

Probabilty

0,9

2

0,7

Psuccess

0,65

1

Energy Consumption

0,6

0 1

3

5

7

9

11 13 15 17 19 21 23 25 27 N° of nodes

Fig. 5. MN´ energy consumption with increasing number of transmitting nodes

4.2 Listen or Transmit? As mentioned, RSSI measurements can be used for the position estimation in 802.15.4 networks. Basically, there are two different ways to obtain RSSI measurements: either the MN listens during a certain time to measure the RSSI of packets transmitted by ANs (scheme 1) or the MN broadcasts packets and the ANs take the corresponding RSSI measurements (scheme 2). Both schemes can be adequate depending on the scenario and application. In scheme 1, if many ANs broadcast a large number of packets, the MN could listen for a long time to receive the necessary RSSI samples from each AN. The number of required RSSI samples depends on the application; in general, if more RSSI samples are taken, a more stable RSSI value can be obtained. If the transmissions of ANs are not immediate, the MN can consume a high amount of energy during the idle listening periods. By using our 802.15.4 sensor nodes we investigated the listening period at the MNs. Fig. 6 is taken from a previous work [3]. In this test the MN sends a packet to indicate to the ANs that they should start transmitting. Thus, each AN tries to broadcast 10 short packets (p1, p2,…p10) and the MN registers the arrival of each packet and the source node. Although the ANs use CSMA in the MAC layer, a random time is additionally included in the application layer. Before the first transmission a random time of [0-6ms] is used to avoid collisions. Between the next transmissions another random time ([0-1ms]) is used by the ANs to increase the probability of having a successful transmission. We define Tm as the instant when the MN achieves registering the last pX sample. Tm is referenced at the first MN´s transmission. By way of example, according to Fig. 6 if there are three ANs (A=3), then it is expected that the MN has to listen for 20ms (average Tm= 20ms) to receive two samples (X=2) from each AN (assuming that all ANs successfully transmitted at least p1 and p2). Fig. 6 suggests a linear relationship between Tm and X when there are few ANs, although this relationship changes with an increasing number of ANs, principally due to the packets that cannot be successfully transmitted.

50

J.J. Robles 120 N° of ANs=1

100

N° of ANs=2 Tm /ms

80

N° of ANs=3 N° of ANs=4

60

N° of ANs=5

40

N° of ANs=6

20

N° of ANs=7 N° of ANs=8

0 1

2

3

4

5

6

7

8

9

X

Fig. 6. Listening period according to the required number of RSSI samples [3]. The confidence intervals (95%) are shorter than 2.5ms in all cases.

The information of the last arrival can be very useful to determine the expected listening period, and thus, the MN´s energy consumption. One advantage of scheme 1 is that the MN is able to estimate its position by using its measured RSSI values. As described, if there are many nodes in scheme 1 that are transmitting packets, the MN´s energy consumption can increase due to the time used in the backoff process (See Fig. 5). Furthermore, all the RSSI measurements taken by the ANs have to be sent to an operation unit, like a cluster head or a central computer, for the position calculation. This generates traffic between ANs reducing the scalability of the network. However, this scheme can be very attractive in terms of MN´s energy consumption compared to scheme 1. For instance, assume in scheme 1 that the node RCB230 listens for RSSI samples during 40ms and estimates its position in 1ms. Afterwards, it sleeps 959ms and repeats the cycle. In this configuration the life-time of its batteries (1500mh/A) is about two months. As an example of scheme 2, imagine the following configuration: the MN is sleeping, it wakes up to transmit a short packet (0.8ms) one time per second and then it goes to sleeping mode again. Here, the lifetime of the batteries is greater than two years (assuming that the channel is not frequently busy). The energy consumption in the transmission and reception modes depends on the used amplification and modulation in the transceivers. Actually both modes present similar energy consumption in many 2.4 Ghz 802.15.4 transceivers. For instance, in the transceiver ATRF230 of Atmel [18] 45mW are consumed in the reception mode and 50 mW in the transmission at 3dbm. The transceiver CC2430 of the company Texas Instrument [24] consumes 80mW during the reception and 81mW in the transmission at 0dbm. Very important is to consider the transitions between the different modes, because it cannot be negligible in terms of energy consumption and time. In the RCB230 we measured a transition time of 2ms between the sleeping mode, where transceiver and microcontroller are inactive consuming 72µW, and the transmission mode. The energy consumed only in this transition is about 81µJ, while the transmission of a typical beacon of 0.8ms consumes 54µJ. In few words, scheme 1 can be appropriate if there are many MNs in the network that require localization. In this scheme, the MN´s energy consumption rises with

Considerations in the Design of Indoor Localization Systems for WSNs

51

increasing numbers of ANs and required RSSI samples. On the other hand, scheme 2 allows the MNs to consume very low energy. However, the high traffic generated in the network can reduce the number of admissible MNs. The more MNs are transmitting, the higher the expected backoff delay, and thus, the MN´s energy consumption. In [5] we propose a new hybrid protocol, which combines the advantages of schemes 1 and 2. 4.3 Mobility in 802.15.4 Sensor Networks The 802.15.4 sensor nodes have two different addresses. The large address (64 bits) is unique and does not depend on the associated network. In contrast, the short address (16 bits) is dynamically assigned by a coordinator and can be used e.g. by a routing algorithm. By analyzing the header of the transmitted packet it is possible to obtain the address of the source. In 802.15.4 networks, an association process is executed when a new node joins the network. In the association procedure the new node has to scan the channel. After detecting a message from the coordinator, the node initiates an exchange of signaling to receive a short-address (16 bits) from the coordinator (parent node). If the node looses the contact with its parent node, and a new short address is necessary, the orphan node tries a re-association process with another coordinator of the network. The re-association process can mean high energy consumption at the MN. Therefore, efficient handover mechanisms have to be developed to support communication when the MN´s level of mobility is high. An alternative simple mechanism is implemented in [3], where the packets transmitted for localization purposes can also be exploited by the handover process. In this proposal, the ANs use the short-address and the MN uses the large address. In this way, it is possible to implement a low-complexity address-based routing algorithm between the fixed ANs, which decides the next hop according to the short address of the destination. The MN mainly makes use of the large-address to avoid the reassociation process defined in the standard. During the first phase of the localization protocol OLP described in [3], the ANs broadcast packets and the MNs take RSSI measurements for their position estimation (scheme 1). By comparing the RSSI measurements, it is also possible that the MN detects the nearest AN. Then, if the MN wants to transmit a packet to the sink, it directly sends its packet to the nearest AN without generating additional signaling. This “selected AN” sends a packet to the sink containing the MN´s large address and the received information. On the other hand, if the sink wants to transmit a packet to the MN, the sink has to send the packet to the selected AN for its posterior delivery. In order to minimize the impact of the MN´s mobility on the communication, the selection of the nearest AN has to be periodically updated. The disadvantage of the usage of the large address is that the header of the transmitted packets is larger as compared to those that use the short address.

5 Conclusions This paper deals with the design of localization system for resource-constrained 802.15.4 sensor nodes. In such networks the RSSI can be used for localization

52

J.J. Robles

purposes. Due to the dispersion of the RSSI measurements and other problems such as the multipath effect, only a coarse position can be estimated. Min-Max and WCL are two localization algorithms that are very robust against distance error. This characteristic make these algorithms appropriate when the RSSI is used for the distance estimation. The fingerprint technique is normally used in indoor scenarios due to the fact that it can take the stationary characteristics of the scenario into account. The transmission of packets and the taking of RSSI measurements are two different schemes that require optimization. In the transmission, the backoff delay in the MAC layer can be very large if the channel is often busy. On the other hand, if the arrival time between RSSI samples is large, the idle listening increases at the receiver. These two effects can means high energy consumption at the sensor node. The mobility can also impact the energy consumption due principally to the association process between nodes. In this paper we discuss the above mentioned problems and suggest some solutions.

References 1. Robles, J.J., Deicke, M., Lehnert, R.: 3D Fingerprint-based localization for Wireless Sensor Networks. In: International Workshop on Positioning, Navigation and Communication WPNC, Dresden, Germany (March 2010) 2. Robles, J.J., Tromer, S., Quiroga, M., Lehnert, R.: A low-power scheme for localization in wireless sensor networks. In: Aagesen, F.A., Knapskog, S.J. (eds.) EUNICE 2010. LNCS, vol. 6164, pp. 259–262. Springer, Heidelberg (2010) 3. Robles, J.J., Tromer, S., Quiroga, M., Lehnert, R.: Enabling Low-power Localization for Mobile Sensor Nodes. In: International Conference on Indoor Positioning and Indoor Navigation (IPIN), Zurich, Switzerland (September 2010) 4. Radeke, R., Robles, J.: Node Degree Improved Localization Algorithms for Ad-hoc Networks. In:10th International Conference on Ad-Hoc Networks and Wireless (ADHOC-NOW), Paderborn, Germany (July 2011) (to be presented) 5. Robles, J.J., Tromer, S., Perez Hidalgo, J., Lehnert, R.: A High Configurable Protocol for Indoor Localization Systems. In: International Conference on Indoor Positioning and Indoor Navigation (IPIN), Guimaraes, Portugal (September 2011) (to be presented) 6. Martinez. J: Leistungsbewertung eines Lokalisierungsalgorithmus in drahtlosen Sensoren - Diploma thesis, TU Dresden (February 2009) 7. Benavidez, M.: Investigation of the RSSI value for ranging - Student work, TU Dresden (December 2009) 8. Tromer, S.: Implementation of an energy efficient Indoor Localization Algorithm Student thesis, TU Dresden (February 2010) 9. Quiroga, M.: Graphical remote Application for an Indoor Localization System - Student thesis, TU Dresden (February 2010) 10. Mangili F.: Evaluation of energy consumption of sensor nodes in localization applications - Student thesis, TU Dresden (July 2010) 11. Oñate, R.: Performance analysis of an energy efficient localization algorithm - Student thesis, TU Dresden (July 2010) 12. Lorenz, M.: Optimierung von kollaborativen Lokalisierungsalgorithmen, Diploma thesis, TU Dresden (July 2010) 13. Miralles, P.: Adaptive Localization in a Localization System - Student thesis, TU Dresden (February 2011)

Considerations in the Design of Indoor Localization Systems for WSNs

53

14. Iribarren, M.: Central Localization for a Wireless Sensor Network - Diploma thesis, TU Dresden (November 2010) 15. Mao, G., Fidan, B.: Localization Algorithms and Strategies for Wireless Sensor Networks. Idea Group Publishing, USA (2009) 16. IEEE standard 802 (April 15, 2006), http://standards.ieee.org/getieee802/802.15.html 17. Lorincz, K., Welsh, M.: MoteTrack: a robust, decentralized approach to RF-based location tracking. In: Personal and Ubiquitous Computing, vol. 11(6), Springer, London (2007) 18. Datasheet IEEE 802.15.4 transceiver AT86RF230, http://www.atmel.com 19. Datasheet microcontroller ATmega1281, http://www.atmel.com 20. Datasheet RCB230 sensor node, http://www.dresden-elektronik.de 21. Langendoen, K., Reijers, N.: Distributed localization in wireless sensor networks: a quantitative comparison. The International Journal of Computer and Telecommunications Networking - Special issue: Wireless Sensor Networks Archive 43(4) (November 2003) 22. Blumenthal, J., Grossmann, R., Golatowski, F., Timmermann, D.: Weighted centroid localization in Zigbee-based sensor networks. In: Proceedings of the IEEE International Symposium on Intelligent Signal Processing (WISP 2007), pp. 1–6 (October 2007) 23. Liu, H., Darabi, H., Banerjee, P., Liu, J.: Survey of Wireless Indoor Positioning Techniques and Systems. IEEE Transactions on System, Man and Cybernetics 37(6) (November 2007) 24. Datasheet IEEE 802.15.4 transceiver CC2430, http://www.ti.com

Backoff Algorithms Performance in Burst-Like Traffic Ievgenii Tsokalo1, Yamnenko Yulia2, and Stanislav Mudriievskyi3 1

Chair for Industrial Electronics National Technical University of Ukraine ”Kyiv Polytechnic Institute” Prospect Peremogy 37, 03056 Kiev, Ukraine [email protected] 2 Chair for Industrial Electronics National Technical University of Ukraine ”Kyiv Polytechnic Institute” Prospect Peremogy 37, 03056 Kiev, Ukraine [email protected] 3 Chair for Telecommunications Technische Universitat Dresden 01062 Dresden, Germany [email protected]

Abstract. In the present paper several backoff algorithms are considered to analysis their behavior in burst-like traffic of smart metering network: YITRAN, Binary Exponential and Polynomial backoff algorithms. The comparison was performed by means of transmission successfulness probability estimation using Markov chain models for each algorithm. The results of the calculation showed that YITRAN algorithm outperforms the other ones for almost all kinds of burst-like traffic. The developed software can be used to implement other backoff algorithms and perform their comparison. A novel backoff algorithm was developed, which is basing on the results the comparison. Keywords: backoff algorithms, Markov chains, Smart Grid.

1 Introduction The Smart Grid topic is one of the very intensively investigated topics nowadays. This high attention to it was drawn by the prediction that introduction of the Smart Grid will increase the part of the “green” energy in the consumption and will reduce energy usage. In order to implement the Smart Grid a distributed energy generation is foreseen. In this context a Microgrid is defined as a “a localized grouping of electricity generation, energy storage, and loads” [10]. It is also defined that Smart Grid will consist of a number of Microgrids. One of the communications technologies considered to be used for control of the Smart Grid and especially Microgrid is Narrow-band Power Line Communications (NB-PLC) [3]. However most of the existing MAC (Multiple Access Control) protocols scheduling mechanisms for such communication network [1] are very sensitive to a kind of traffic. Thus, development of new MAC protocol with improved properties in the term of sensitivity and stable work under any kind of traffic is urgent and challenging task. R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 54–64, 2011. © Springer-Verlag Berlin Heidelberg 2011

Backoff Algorithms Performance in Burst-Like Traffic

55

Here was assumed that the traffic consists of traffic bursts. It can be described by two parameters: average value of active nodes variation and the number of packets in their queues. For each pair of these parameters there are certain scheduling mechanisms which are most preferable to be used. But there is no possibility to get accurate values of the parameters before each transmission frame. They can be tracked by collecting a statistical data. This does not allow fast reacting to a traffic change but most proper scheduling mechanism for some kind of network and day time could be chosen. Also switching between different mechanisms should be possible in the case if traffic kind changes and remains so for a long time. So, the target for this work is a development of scheduling mechanisms map that will permit choosing of the most proper mechanism for certain kind of traffic. The map actually shows a preference of one or another algorithm usage basing on two traffic parameters that are mentioned above. Usually value of active nodes variation in Microgrid [5] is high. CSMA/CA technology is chosen as channel access mechanism because it shows one of best accommodation possibility to the burst-like traffic [1, 6]. The main part of the scheduling mechanism in CSMA/CA is backoff algorithm. Thus, comparison of different backoff algorithm and choice of the most effective one is a necessary condition to develop proper scheduling mechanism. In second section of the paper different backoff algorithms models are proposed. In third and fourth sections static and dynamic characteristics of several algorithms are presented.

2 Backoff Algorithms Models Three algorithms were considered: binary exponential backoff (BEB), polynomial backoff (POL) and YITRAN backoff algorithm [2, 4, 7]. In the last one contention window width (CW) depends on the number of active nodes. YITRAN’s has a good performance [4] and can be effectively used to evaluate traffic parameters so it was considered as a basic algorithm. Performance of backoff algorithms can be estimated using method of Markov chains [8] or by simulation. The first one allows estimation of the backoff algorithm separately from the rest of scheduling mechanism and thus gives better results. That is why it will be considered in details further in the paper. One-dimensional Markov chain can be characterized with a transition matrix π where the rows represent previous states and the columns - new states:

CWnew CWold

π=

1

CWmax

2

1 2

p11

p12

...

p1CWmax

p21

p22

...

p2CWmax

...

...

CWmax

pCWmax 1

...

pCWmax 2 ...

... pCWmax CWmax

Therefore, matrix is square and each state corresponds to certain value of CW. The matrix dimension is equal to the difference between maximum ( CWmax ) and

56

I. Tsokalo, Y. Yulia, and S. Mudriievskyi

minimum ( CWmin ) width of CW (in π CWmin is shown to be 1 for visualization simplicity). Hereinafter accordingly to [4] CWmax = 500 and CWmin = 15 . Each

element of the matrix corresponds to the probability PCWold CWnew of transition from the old ( CWold ) to the new ( CWnew ) state. An algorithm for PCWold CWnew probability calculation is based on the method from [4]. At first probability density function (PDF) parameter should be estimated. This parameter is a probability that estimated number of active nodes r will be equal to real value R: r ⎧ S − 1 ⎛ CWold − S + 1 ⎞ ⎪ p( R = r ) = ⋅⎜ ⎟ ⎪ CWold ⎝ CWold ⎠ , ⎨ ⎪ CWold ⎪⎩ S = r + 1

(1)

where S – expected value of minimal backoff counter amongst all nodes in the network in one transmission frame. New value of CW is calculated using autoregression function for CWnew = f : f ( CWold , r ) = χ ⋅ CWold + (1 − χ ) ⋅ 12 ⋅ r .

(2)

Using (1) and (2) π matrix was filled with PCWold CWnew by following algorithm (value of r was varied in a range from 1 to 350):

⎧ for CWold ∈ (1, CWmax ) ⎪⎪ PCWold CWnew + = p (r ) . ⎨ ⎪ ⎪⎩CWnew = f ( CWold , r )

(3)

f ( CWold , r ) gives new state and p ( r ) - probability of transition to this state. A sum

of items from each row of π matrix should be equal to one. As far as p ( r ) is defined on infinite range of r each row of π should be normalized. Fixed row vector is found from π using Theorem 11.8 [8]:

w⋅π = w ,

(4)

where w = ( w1 , w2 , ..., wCWmax ) - matrix row of probabilities [1 x CWmax ]. Its maximum element should correspond to an expected average CW value for certain real number of active nodes R that can be found from following formula [4]: CWopt = 12 ⋅ R .

(5)

Backoff Algorithms Performance in Burst-Like Traffic

57

On Fig.1 fixed row vector versus CW state is plotted for several values of R.

3.2

CWopt = 72

Fixed vector values, %

2.8 2.4 2 1.6 CWopt = 228

1.2

CWopt = 384

0.8 0.4 0 0

50 R=6

100

150

200 250 State CW

R = 19

R = 32

300

350

400

R = 50

Fig. 1. Fixed probability values versus number of Markov chain state

Values calculated by (5) are almost coinciding with extreme points on Fig.1. Transition matrix for BEB algorithm was filled accordingly to the following sequence:

for CWold ∈[CWmin , CWmax ] CWnew = CWold >> 1

π CWold ,CWnew = PS

(6)

CWnew = CWold << 1

π CWold ,CWnew = PC _ CWold CWmin and CWmax for all algorithms have the same values as notified above. PC _ CW

old

- probability of collision in CWold state calculated as following [9]:

⎛ 1 ⎞ PC _ CWold = 1 − ⎜ 1 − ⎟ ⎝ CWold ⎠

R −1

`.

(7)

Transition matrix for polynomial algorithm was calculated by (6) but using following formula for calculation of new CW:

CWnew = (i + 1) a ⋅ CWmin , where minimal and maximal values of i are following [7]:

(8)

58

I. Tsokalo, Y. Yulia, and S. Mudriievskyi

imin = 0 ⎛ 1 ⎛ CWmax ⎞ ⎞ imax = (integer) ⎜⎜ ln ⎜ ⎟ − 1⎟⎟ ⎝ a ⎝ CWmin ⎠ ⎠ a = 1.5

In case of collision i increments, otherwise decrements.

3 Static Mode Evaluation To compare the performance of considered algorithms their models were investigated in the regimes of saturated throughput and dynamic traffic. In the first regime an average number of successful transmissions N succ was calculated as a variable for

performance measurement. It’s averaged for number of transitions_ n → ∞ and regarded to be equal to: N success

=

CWmax

∑

i =CWmin

(

)

wi ⋅ 1 − PC _ i ,

(9)

where wi - i-th value of the fixed vector; PC _ i is calculated by (7).

As it can be observed the proposed measurement is unitless. Thus in saturated throughput mode no preliminary condition because are required but the dimension of the transition matrix, which is defined above. On Fig.2 this number versus real number of active nodes is plotted.

Successful transmissions (maximal 1)

0.96 0.9 0.84 0.78 0.72 0.66 0.6 0.54 0.48 0.42 10

20

30

40 50 60 70 R, Number of active nodes

BEB Polynomial algorithm

80

90

100

YITRAN algorithm

Fig. 2. Average number of successful transmissions versus number of active nodes

Backoff Algorithms Performance in Burst-Like Traffic

59

It can be observed that YITRAN algorithm considerably outperforms BEB and POL.

As far as YITRAN algorithm uses the real probability distribution of active nodes number it adapts to this number change better and provides less average number of collisions.

4 Dynamic Mode Evaluation Saturated throughput performance does not give full information about the algorithms performance under real traffic. Therefore dynamic mode evaluation was performed as the algorithm reaction on a traffic burst. The burst can be completely defined with two parameters that define traffic kind: number of active nodes R that instantaneously appear (average value of active nodes variation) and the number of packets in their queues K In initial state there is no traffic at all. In terms of transition matrix π it corresponds to the state CW = CWmin . Then instantaneously R active nodes appear. In the beginning CW rises and then decreases due to the decreasing of R. The process goes on to the full extent of traffic. Time interval from initial state to the full extent of traffic is measured as a qualitative dynamic characteristic (hereinafter referred to as a traffic burst width). Markov chain model should be updated in order to take into account the decreasing of R. R can be changed in case of successful transmission. Usually MAC layer receives several packets to transmit at a time. So R decrements after K successful transmissions. This process can also be characterized by Markov chain. But each state now is defined with two variables: CW and R. Now transition matrix has order (CWmax − CWmin ) ⋅ Rmax . Let’s indicate it as π R . There is following connection between above defined π and π : R

π

R R =const

=π .

Matrix π for certain R=const will be indicated as π R . Transition matrix for one burst of traffic:

⎛ (1 − ps ) ⋅ π Rmax ⎜ 0 ⎜ R ⎜ π = 0 ⎜ ⎜ ... ⎜⎜ 0 ⎝

ps ⋅ π Rmax

(1 − ps ) ⋅ π Rmax −1

0

...

ps ⋅ π Rmax −1

...

0

(1 − ps ) ⋅ π Rmax −2

...

... 0

... 0

... ...

0⎞ ⎟ 0⎟ ⎟ 0⎟ ... ⎟ ⎟ π 0 ⎟⎠

where Rmax is the number of active nodes that appear instantly; p s - probability of successful transmission of all K packets after one attempt.

Using (7):

60

I. Tsokalo, Y. Yulia, and S. Mudriievskyi

R −1 1⎛ ⎛ 1 ⎞ ⎞. (10) ps = ⎜1 − ⎜1 − ⎟ ⎟ K ⎜ ⎝ CWold ⎠ ⎟ ⎝ ⎠ The observed time interval is calculated as an average number of transmission frames that are required to send all packets of all active nodes. Taking these states as end point of modeling we regard all states of π 0 to be absorbing (there are no active nodes remained) [8]. Software for these calculations should solve a system of linear equation with order (CWmax ⋅ ( Rmax − 1)) . For YITRAN algorithm this value is 4500. To simplify calculations instead of determining of the whole width of the traffic burst only time R interval that needed for each decrement of R will be defined. For this purpose π is transformed and new Rmax matrices are created from it. Each of these new matrices corresponds to one decrement of R. They are typical absorbing matrices:

⎛ (1 − ps ) ⋅ π Ri

π R = ⎜⎜ i

⎝

0

ps ⋅ π Ri ⎞ ⎟, I ⎟⎠

(11)

where i ∈ [ Rmax ,1] ; O is zero matrice; I is identity matrice. O and I have dimensions equal to the number of absorbing states. To find traffic burst width an average time for each R decrement is found. For this R purpose fundamental matrix N of the absorbing Markov chain π i is found as −1 N = ( I − Q ) . Item N ij denotes total number of transition times from state j to state i before being absorbed (usually float). Absorption time vector t (R decrement time) can be found as a sum of the rows from N. In order to find an average time of absorption from time vector a weighting coefficients vector should be found. Probabilities of being absorbed in each state play a role of the coefficients. Accordingly to Theorem 11.6 [8] absorbing probabilities can be calculated as following: B = N ⋅H ,

(12)

where H = p s ⋅ π Ri . Every item Bij denotes probability of being absorbed in a state j if i is an initial state. In the dynamic modeling preliminary parameters are denote by three variables: starting number active nodes, number of sent packets and the transition matrix dimension. On the beginning of the modeling we have R = Rmax and the number of sent packets pack = 0 . In this case we know that the initial state in transition matrix is

Backoff Algorithms Performance in Burst-Like Traffic

CW = CWmin . So time of absorption is t transitions in the matrix π

R

average time of absorption t

R = Rmax , pack = 0

61

= t0 . Hereinafter number of

will be marked by two indexes: R and pack, e.x an

= t 00 . For case R = Rmax and pack = 1 the

R = Rmax , pack = 0

initial state is determined by probability vector A, which is found from a matrix of 00

probability vectors B . First probability vector is found from B as a first row because in case if R = Rmax and pack = 0 process starts from the first state CW = CWmin . For next cases A will be calculated using all values of B because it’s non-zero probability that next transitions will start from any state. Probability vector is always calculate from matrix B calculated in previous step so A

01

=B

00 0

. Lower index

indicates first row; higher indexes indicate case R = Rmax , pack = 0 for B and R = Rmax , pack = 1 for A. Time of absorption is found as following: t=

CWmax −CWmin −1

∑ i =0

Ai ⋅ ti .

(13)

For the cases

( R = Ri , Rmin ≤ Ri < Rmax , 0 ≤ pack ≤ packmax ) & ( R = Rmax & pack > 1) initial state is determined by probability vector A CWmax −CWmin −1

A

nm k

=

∑

nm

. Each item of A

Bik ⋅ A

i =0 CWmax −CWmin −1

nm

is calculated as:

n[ m −1] i

,

∑

A

(14)

nm i

i =0

when pack variable is changing, R remains constant and CWmax − CWmin −1

A

n0 k

∑

=

[ n −1] packmax

Bik ⋅ A

i

i=0 CWmax − CWmin −1

∑

,

A

(15)

nm

i=0

i

when R changes too. R does not change till some node sends all K packets.

A

n[ m −1]

plays a role of weighting vector to find A CWmax − CWmin −1

performing (14) or (15)

∑ i =0

k max = CWmax − CWmin .

A

nm i

nm

. A

nm

is normalized, so after

= 1 . Length of this vector is equal to

62

I. Tsokalo, Y. Yulia, and S. Mudriievskyi

Time of absorption is calculated as follows:

t

nm

=

CWmax − CWmin −1

∑

A

n[ m −1] i

i=0

⋅ t nm i .

(16)

Then the width of traffic burst is calculated as:

BW = t

00

+t + 01

Rmax −1 packmax

∑ ∑ n =0

t

nm

.

(17)

m=2

BW, number of transmission frames

For BEB and polynomial backoff algorithms the same procedure for dynamic modelling was used. The traffic burst width versus K for several R for BEB, polynomial algorithm with above calculated values of YITRAN algorithm is plotted on Fig. 3. 120 110 100 90 80 70 60 50 40 30 20 10 0

Special point 2

Special point 1

2

3

4 5 6 7 8 K, starting number of packets in the queue

9

10

(YITRAN) R = 10,8,6,4,2 (BEB) R = 10,8,6,4,2 (POL) R = 10,8,6,4,2

Fig. 3. Ratio of traffic burst width for BEB and polynomial to improved YITRAN algorithm versus K for several R

Special point 1 on Fig. 3 is an intersection of the YITRAN (R = 2) and polynomial (R=2) characteristics and is a limit point. For K<4 and R=2 the last algorithm outperforms the former ones, i.e. for these combinations of K and R polynomial algorithm will show better performance. For all other results shown on Fig._3 YITRAN algorithm is the most effective. Special point 2 on Fig. 3 is the limit point between BEB and polynomial. But BEB shows worse performance than YITRAN algorithm for all K and R.

Backoff Algorithms Performance in Burst-Like Traffic

63

10

R, number of active nodes

9 8 7

YITRAN

6 5 4 3

POLYNOMIAL

2 1 2

3

4 5 6 7 8 K, starting number of packets in the queue

9

10

Fig. 4. Preferable backoff algorithms sections of K and R values

On Fig. 4 a line of the special points between polynomial and YITRAN algorithm performing modelling for all discrete values of R and K is plotted. An area to the left of the line is the area of R and K combination where polynomial algorithm is more preferable. BEB is not competitive with the others and has no area to be used. With the developed software other backoff algorithms can be investigated and their areas of effective application can be defined.

5 Conclusions Modified and improved backoff algorithm which can adapt to the network traffic kind is proposed. It was assumed that any traffic kind could be characterized by two parameters: the average variation of the number of active nodes and the number of packets that these nodes obtain to send in a beginning. Markov chain was created for all considered backoff algorithms (BEB, polynomial and YITRAN’s). The software was developed to compute the algorithms performance characteristic for static and dynamic traffic modes. In static mode YITRAN algorithm substantially outperformed BEB and polynomial one. In dynamic traffic modeling there was created the map that shows where it is preferable to use one or another algorithm. The map was defined with the traffic parameters that are discussed above. These parameters can be tracked during network operation and a backoff algorithm can be chosen in accordance to it. If the traffic kind changes for a considerable amount of time there is a possibility to switch between different algorithms to achieve better performance. A minimal switching time is defined by features of MAC protocol.

64

I. Tsokalo, Y. Yulia, and S. Mudriievskyi

BEB found no place on this map as far as the computation proved it to be not compatible with others. Polynomial algorithm outperforms YITRAN in a small area. Therefore among considered algorithms the YITRAN is a dominant one. The developed software can be used to add new algorithms to the map. In such a way all existing backoff algorithms can be compared. Until now only error-based algorithms were modeled with Markov chains. A reason for this is that traffic-based algorithms do not allow using of mathematical expressions for Markov chain representation. They can be interpreted by Markov transition matrix only.

References 1. Tannenbaum, A.S.: Computer Networks, 4th edn., p. 891. Pearson Education International, Inc., Upper Saddle River (2003) 2. Raphaeli, D., Grauer, A.: A New Power-Line-Communicatiovts Modem based on a Novel Modulation Technique. Tel Aviv University, Israel, Itran Communications 3. Lee, J.-h., Park, J.-h., Lee, H.-S., Leett, G.-W., Kim, S.-c.: Measurement, modeling and simulation of power line channel for indoor high-speed data communications. School of Electrical and Computer Engineering, Seoul National University 4. Gazit, R., Avidan, A.: An Innovative Adaptive Channel Access Mechanism. In: ISPLC (2004) 5. Carron, G., Ness, R., Deneire, L., Van der Perre, L., Engels, M.: Comparison of two modulation techniques using frequency domain processing for in-house nerworks. In: Interuniversity Micro Electronics Center (IMEC) - Kapeldreef 75, 3001, Heverlee, Belgium, 6. Lee1, K.-R., Lee1, J.-M., Kwon1, W.-H., Ko2, B.-S., Kim, Y.-M.: Performance Evaluation of CSMA/CA MAC Protocol in low-speed PLC Environments. School of Electrical Engr. and Computer Science, Seoul National University 7. Xu, D., Sakurai, T.: An Analysis of Different backoff Functions for an IEEE 802.11 WLAN. Department of Electrical & Electronic Eng, The University of Melbourne 8. Grinstead and Shell’s Introduction to Probability. The CHANCE Project. American Mathematical Society, Providence (July 4, 2006) 9. Vu, H.L., Sakurai, T.: Collision Probability in Saturated IEEE 802.11 Networks. Centre for Advanced Internet Architectures (CAIA). ARC Special Research Centre for UltraBroadband Information Networks (CUBIN) 10. Microgrid, http://en.wikipedia.org/wiki/ Distributed_generation#Microgrid

New IEEE 802.16-2009 Compliant Traﬃc Shaping Algorithms for WiMAX Networks Volker Richter and Stefan T¨ urk Dresden University of Technology Dresden, Mommsenstrasse 13 01062 {richter,tuerk}@ifn.et.tu-dresden.de www.ifn.et.tu-dresden.de/tk

Abstract. WiMAX based on the IEEE 802.16 standard allows the deployment of cellular mobile broadband access networks delivering all services through the Internet Protocol. An essential feature of such networks it the Quality of Service (QoS) support including the traﬃc shaping. Based on detailed analyses of the IEEE 802.16-2009 QoS concept, we propose two traﬃc shaping algorithms. The ﬁrst one is standard conform at the expense of high memory consumption. The second approximate algorithm has been developed for memory crucial system. We present parameter studies obtained by a detailed ns-2 WiMAX simulation environment. The simulation results show that both algorithms can be used as a base for QoS-aware WiMAX scheduling algorithms. Keywords: WiMAX; IEEE 802.16; QoS; Traﬃc Shaping; Traﬃc Policing.

1

Introduction

The IEEE 802.16 standard family guarantees the compatibility of WiMAX systems from diﬀerent vendors. WiMAX is one of the candidates for the 4th generation mobile networks, fulﬁlling the ITU-R IMT Advanced criteria [2]. The standard IEEE 802.16-2009 [1] speciﬁes three physical layers namely WirelessMANSC, WirelessMAN-OFDM and WirelessMAN-OFDMA supporting microwave radio relay connections through single carrier (SC) transmissions, cellular networks for ﬁxed subscribers and with WirelessMAN-OFDMA full mobile cellular networks. On the top of all these physical layers a mostly common Medium Access Control (MAC) layer has been speciﬁed. An integral feature of this MAC layer is the Quality of Service (QoS) support. Therefore, the standard deﬁnes ﬁve service classes adjusted to application requirements and speciﬁc traﬃc characteristics. The detailed requirements and constrains of a connection are signaled through QoS parameter sets during the connection establishment procedure. Assertion and observance of the accepted QoS parameters are tasks of scheduling and traﬃc shaping algorithms, which have been left unspeciﬁed by the standard for vendor diﬀerentiation. R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 65–76, 2011. c Springer-Verlag Berlin Heidelberg 2011

66

V. Richter and S. T¨ urk

In the last years many scheduling algorithms have been proposed [6, 3], but only few authors have taken the QoS parameters speciﬁed in the IEEE 802.162009 standard into consideration. Much less work has been done on traﬃc shaping algorithms. The authors of [5] propose a traﬃc shaping algorithm based on token buckets disregarding the standardized integration interval of the data rates. Token bucket approaches do not comply with the IEEE 802.16-2009 design. Therefore, the aim of this paper is to propose a traﬃc shaping algorithm accurately fulﬁlling the IEEE 802.16-2009 speciﬁcation. Furthermore, a second approximation algorithm is presented to overcome the drawback in terms of memory consumption of the accurate algorithm. In the ﬁrst section, the IEEE 802.16-2009 QoS concept is introduced based on our deep studies of the standard document. Subsequently, we describe our proposed algorithms in a detailed manner. These algorithms have been implemented in the network simulator ns2 based WiMAX model provided by the Application Working Group of the WiMAX Forum. Our simulation results are presented in section 4. Finally, we conclude our paper and give an outlook on future work.

2

The IEEE 802.16-2009 QoS Concept

The IEEE 802.16-2009 standard speciﬁes ﬁve service classes to fulﬁll various application requirements and to make use of their speciﬁc traﬃc characteristics. For the Downlink (DL) direction the service classes are denoted as data delivery services types (cp. [1], p.422) and for the Uplink (UL) direction they are speciﬁed as scheduling services (cp. [1], p.291). The service class Unsolicited Grant Service (UGS) has been designed for real-time applications generating traﬃc with ﬁxed packet sizes and constant inter-arrival times. Thus, it is especially suitable for Voice over Internet Protocol (VoIP) without silence suppression. To satisfy the needs of important real-time applications, such as VoIP with silence suppression, the service class Extended Real-Time Variable Rate (ERT-VR) has been created. It is denoted as Extended Real-Time Polling Service (ertPS) for UL transmissions. This kind of applications oﬀer packets with constant sizes and inter-arrival times within certain time periods. In contrast to UGS, these parameters change during connection times. Common video codex’s produce data steams with variable packet and interarrival times. The real-time service class Real-Time Variable Rate (RT-VR) for DL and respectively Real-Time Polling Service (rtPS) for UL transmissions is suitable for video streams. To support delay tolerant applications, which need minimum data rates for timeout avoidance the service class Non Real-Time Variable Rate (NRT-VR) or Non Real-Time Polling Service (nrtPS) has been designed. No QoS guarantees are given to applications which are using the Best Eﬀort (BE) service class. For example web traﬃc is delivered by this service class. Each service class is characterized by a set of mandatory and optional QoS parameters. To police incoming data streams, as required by the IEEE 802.16-2009

New IEEE 802.16-2009 Compliant Traﬃc Shaping Algorithms

67

standard, the parameters Maximum Sustained Traﬃc Rate (MSTR), Minimum Reserved Traﬃc Rate (MRTR), Time Base (TB) and Maximum Traﬃc Burst size (MTB) have to be taken into consideration. The MSTR deﬁnes the average maximum data rate which can be allocated for the corresponding connection. It is measured in bit per second, where MAC and physical overhead is excluded (cp.[1], p. 1285). This parameter allows to restrict the data rate of a connection according to the bandwidth contracted with the user. Therefore, this parameter is important for network operators to oﬀer contracts with diﬀerent maximum bandwidths and to limit their costs for backbone traﬃc. As shown in the left part of ﬁgure 1, it speciﬁes the upper data rate bound. Even in an idle WiMAX cell, a connection shall not receive more physical allocation than it is necessary to fulﬁll the MSTR. Because of the definition as an average over time, the MSTR can be temporarily overtaken. This is allowable when the previous data rate of the oﬀered traﬃc did not reach the MSTR. This case is depicted in ﬁgure 1 before the second traﬃc maximum in the idle example. In contrast to the MSTR, the MRTR describes the average

Idle

Overloaded MSTR MRTR Time

Fig. 1. System behaviour in idle and overload state

minimum data rate which must be allocated for connections, if data is available in their queues. It is also expressed in bit per second without MAC and physical overhead (cp. [1], p. 1285). This parameter deﬁnes a guaranteed rate for a connection, which is important for the functionality of the corresponding application. Also in case of a fully loaded system, all MRTRs have to be reached as illustrated in right part of ﬁgure 1. As stated before, both parameters are average values over time. Therefore, an integration interval has to be speciﬁed, which is done in the IEEE 802.16-2009 standard by deﬁning the QoS parameter Time Base (TB) given in ms (cp. [1], p. 1306). Furthermore, the optional QoS parameter Maximum Traﬃc Burst size (MTB) may have an inﬂuence on the traﬃc shaping. The MTB deﬁnes the maximum data volume which shall be allocated in one frame (cp.[1], p. 1285). With the help of this parameter a more regular data transmission can be achieved. Table 1 summarizes all service classes including their mandatory (X) and optional (O) QoS parameters. Because of the constant data rate the MSTR

68

V. Richter and S. T¨ urk

Table 1. Service classes and their mandatory (X) and optional (O) QoS parameters Service classes DL Service classes UL

UGS UGS

ERT-VR ertPS

RT-VR rtPS

NRT-VR nrtPS

BE BE

MSTR MRTR TB MTB

X X X O

X X X O

X X X O

X

X X O

X O

parameter is not intended for UGS. In contrast, BE does not support MRTR parameters as guaranteed rates are not provided. Our two traﬃc shaping algorithms are based on the described QoS parameters and will be introduced in the following section.

3

WiMAX Traﬃc Shaping Algorithm

In the following, we ﬁrstly introduce our exact traﬃc shaping approach. Afterwards, our approximate algorithm is described. Both algorithms are calculating the maximum admissible data rate based on the MSTR parameter as well as minimum guaranteed data rate based on the MRTR parameter for each connection. In the description of the algorithms we refer to the MSTR calculation. 3.1

Accurate Algorithm

In order to fulﬁll the average maximum data rate RM S deﬁned using the MSTR QoS parameter, the data volume S must be transmitted through the WiMAX system within the duration of one Time Base denoted as TB . S = RM S · TB

(1)

Therefore, transmission opportunities have to be distributed on a number of n frames. TB (2) n= TF Where TF denotes the frame duration in ms. To determine the maximum data volume Sk which might be allocated by the scheduler in the k th frame, all previous allocations Si within the current TB period have to be subtracted. Hence, all allocations made in frames from k − n to k − 1 have to be considered. This can be expressed in the following equation. Sk = RM S · TB −

k−1

Si

(3)

i=k−n

Sk is the maximum data volume allowed to be scheduled. To enable the calculation of the sum in equation 3 all previous allocation sizes of n − 1 frames have to be kept in memory.

New IEEE 802.16-2009 Compliant Traﬃc Shaping Algorithms

69

As mentioned in section 2, the IEEE 802.16-2009 standard requires, that only resources for available data shall be assigned. Therefore, a minimum function between Sk and the current queue size SQ of a connection is introduced. The resulting data volume allowed to be scheduled is denoted as Sk . Sk = min( Sk , SQ )

(4)

The optional QoS parameter MTB is denoted as SM T B and avoids large allocation sizes after periods where the oﬀered traﬃc of a connection dropped below the MSTR or respectively the MRTR. Figure 2 depicts an example, where the TB duration is four times longer than the frame duration. Immediately after

S ccc

TB

S MTB RMS TF

TB

4 TF

k

Fig. 2. Example of the Accurate Traﬃc Shaping Algorithm showing the inﬂuence of the MTB parameter

an idle period, where no packets have been transmitted the queue is ﬂooded. At this moment, the entire data volume of one TB interval is admitted. In the example of ﬁgure 2, the transmitted traﬃc equals four times the average volume. Therefore, the average data volume of four frames is transmitted in the ﬁrst frame k = 1. In the next call of the algorithm for the following frame k = 2, the possible data volume has already been used. Therefore, no further data can be sent until the ﬁrst transmission entry of the frame k = 1 gets older than one TB = 4 duration and is discarded. Due to that eﬀect no further data is transmitted in the frames k = 2, 3, 4. Afterwards, this procedure repeats beginning with the frame k = 5 and results in an uneven distribution of data transmissions. With the help of the QoS parameter MTB, the maximum transmitted amount of data in one frame is limited to SM T B assuring a more homogeneous transmission as illustrated in ﬁgure 2. The compliance to the MTB QoS parameter can also be realized with a minimum function. Where the maximum possible allocation size for scheduling is deﬁned as Sk . Sk = min( Sk , SM T B ) (5) The MTB value has to be greater or equal to the average allocation size in order to have the possibility to assign enough transmission opportunities to fulﬁll

70

V. Richter and S. T¨ urk

the MSTR QoS parameter. Otherwise, this parameter is violated and agreed QoS parameters for the corresponding connection can not be achieved. SM T B ≥ R M S · T F

(6)

The scheduling algorithm may allocate less data then Sk , according to the current system load and latency constrains of other connections. After segmentation and/or concatenation of incoming packets into MAC Protocol Data Units (PDUs) and their placement into the frame structure, the eﬀectively transmitted payload volume Sk is saved to calculate the data volume for the following frames. (7) 0 ≤ Sk ≤ Sk To avoid oscillations after establishment of a connection as shown in ﬁgure 2 all previous allocation sizes are set to the average allocation size. Sk = RM S · TF

for

k≤0

(8)

In conclusion, the accurate traﬃc shaping can be described with the following formula. k−1 Si (9) Sk ≤ Sk = min SM T B , SQ , RM S · TB − i=k−n

This proposal accurately fulﬁlls the deﬁnitions of the IEEE 802.16-2009 standard. It requires memory to save n − 1 allocation sizes. Therefore, the memory usage of the proposed exact traﬃc shaping algorithm scales linear with the Time Base duration Tb . In contrast, the computational complexity remains constant as the previous sum value is also saved and the next value used in equation 3 can be obtained as given in equation 10 . k

Si =

i=k−n−1

3.2

k−1

Si + Sk − Sk−n

(10)

i=k−n

Approximate Algorithm

For WiMAX systems,- where memory usage is crucial, we developed an approximate traﬃc shaping approach, which is based on [4]. To achieve a constant memory usage, the average eﬀective data rate of the previous frames Rk−1 is saved instead of all allocation sizes Si within one TB duration. Therefore, only memory for one value is needed. This algorithm is subdivided in two steps. Firstly, we have to determine the maximum allocation size Sk in for the current frame. Secondly, we use the amount of eﬀective transmitted data Sk to obtain the average rate for the next iteration Rk . The maximum allocation size Sk in for the current frame k results from the diﬀerence between the data volume fulﬁlling the MSTR RM S within the TB

New IEEE 802.16-2009 Compliant Traﬃc Shaping Algorithms

71

duration TB and the data volume allocated in the n−1 previous frames calculate with help of the previous average data rate Rk−1 . Sk = RM S · TB − Rk−1 (TB − TF )

(11)

In accordance to the exact algorithm, the constrains of the available data size in the queue SQ as well as the optional MTB size can be realized with a minimum function obtaining Sk . (12) Sk = min SM T B , SQ , RM S · TB − Rk−1 (TB − TF ) To avoid start oscillations as described for the accurate traﬃc shaping algorithm, R0 is set to the value of the MSTR parameter RM S for the ﬁrst frame k = 1. Rk = RM S

; for

k≤0

(13)

After the scheduling and frame building process the transmitted payload volume Sk is feed back to the algorithm. We adopted the time sliding window approach from [4] to calculate the future average rate Rk . Rk =

Rk−1 · TB + Sk TB + TF

(14)

Due to the time sliding window the memory duration depends also on scheduled data volumes and not only on the TB QoS parameter. Therefore, this approach cannot be considered as strictly standard compatible, but it is more eﬃcient in terms of memory consumption.

4

Simulation Results

Both algorithms have been implemented in a simulation environment consisting of the network simulator ns-2 and the corresponding Add-On for WiMAX modelling [8] developed by the Application Working Group of the WiMAX Forum. They are applied on the MSTR and MRTR of each connection calculating the maximum allowed allocation sizes and the minimum necessary data volumes to meet the application requirements. Our simulation environment includes a two phase scheduling algorithm. In the ﬁrst phase it allocates only data volumes to achieve guaranteed rates. Afterwards, MSTR based requests are handled. To study the proposed algorithms one subscriber station has been placed next to the base station, arising thereby the highest possible burst proﬁle 64 QAM 3/4 is used. One DL connection applying the RT-VR service class has been conﬁgured. The MSTR and MRTR were set to 14 M Bit/s and 10 M Bit/s in order to stay below the system capacity. In the ﬁrst presented simulation results the MTB size was set to inﬁnity. The frame duration of 5 ms and other none determining parameters have been chosen according to [7]. The oﬀered traﬃc was simulated by the ns-2 constant bit rate source model generating packets with a size of 576 Byte.

72

V. Richter and S. T¨ urk

Algorithm Comparison Figure 3 shows the achieved goodput over time without any activated traﬃc shaping algorithm. Therefore, the goodput corresponds to the oﬀered traﬃc. In

20 1.

Goodput [Mbit/s]

15

10

2.

3.

4.

RT-VR 6.

5.

MSTR

MRTR

5

0 10

20

30

40

50

60

70

80

90

100

Time [s]

Fig. 3. Goodput over Time without Traﬃc Shaping

the six shown test cases the oﬀered traﬃc alternates in all possible relations to the MSTR and MRTR boundaries. Thereby, it is possible to study the behaviour of algorithm in typical situations. In ﬁgure 4 the accurate traﬃc shaping algorithm was activated and the TB length was set to 20 ms equivalent to 4 frame durations. The comparison between the two ﬁgures shows, that the achieved goodput is limited to the MSTR boundary. Due to the short TB duration, unused data volumes in the idle periods are not taken to temporally exceed the MSTR. The MRTR has no inﬂuence of the graph as the system is not overloaded. The simulation result obtained with the approximate approach is shown in ﬁgure 5. In periods with constant oﬀered traﬃc the MSTR constraint is also kept by using the approximate algorithm. But especially after 40 s simulation time in test case three the disadvantage of this algorithm can be observed. If the oﬀered traﬃc changes heavily, the calculated average data rate by the sliding window method diﬀers heavily from the real one, leading to an overshoot of the achieved goodput.

New IEEE 802.16-2009 Compliant Traﬃc Shaping Algorithms

73

20 1.

Goodput [Mbit/s]

15

10

2.

3.

4.

RT-VR 6.

5.

MSTR

MRTR

5

0 10

20

30

40

50

60

70

80

90

100

Time [s]

Fig. 4. Accurate Traﬃc Shaping Algorithm

20 1.

Goodput [Mbit/s]

15

10

2.

3.

4.

RT-VR 6.

5.

MSTR

MRTR

5

0 10

20

30

40

50

60

70

80

Time [s]

Fig. 5. Approximate Traﬃc Shaping Algorithm

90

100

74

V. Richter and S. T¨ urk

Parameter Study After the comparison of the algorithms, we analyze the inﬂuence of the TB period and MTB size QoS parameters on the accurate traﬃc shaping approach. Therefore, the TB parameter was set to an extreme value of 1 s. Hence, the goodput exceeds the MSTR during one TB duration. Figure 6 shows the goodput over time in the above introduced scenario and a TB interval of 1 s. The simulation result shows, that the goodput is periodically

20 1.

Goodput [Mbit/s]

15

10

2.

3.

4.

RT-VR 6.

5.

MSTR

MRTR

5

0 10

20

30

40

50

60

70

80

90

100

Time [s]

Fig. 6. Accurate Traﬃc Shaping Algorithm - Time Base 1 s

higher than the MSTR in the test cases three, ﬁve and six, where the oﬀered traﬃc exceeds this boundary. Between 40 s and 50 s in case three the goodput reaches the system capacity at approximately 17 M Bit/s followed by a sharp decline to 0 M Bit/s. This behavior is repeated 10 times in this 10 s long test case. During the idle period no data has been allocated for this connection. When the packet queue is ﬁlled after 40 s the whole queue size may be allocated in the ﬁrst frame only limited by the system capacity, as described in section 3.1. Therefore, the following frames are also fully utilized. This leads to the allocation of the entire data volume of one TB within the ﬁrst frames. When this limit is reached no further data volumes are admitted until the ﬁrst saved allocation entry greater zero is older than one TB. This eﬀect results in repeated oscillations visible in ﬁgure 6. However, the IEEE 802.16-2009 speciﬁcation permits such a allocation scheme. But it should be mentioned, that these irregular transmissions

New IEEE 802.16-2009 Compliant Traﬃc Shaping Algorithms

75

can increase the average delay. Therefore, the TB value should not be greater as the QoS parameter maximum tolerable latency deﬁned in IEEE 802.16-2009. To achieve a more regular transmission the MTB QoS parameter can be applied. In ﬁgure 7 the MTB has been set to 8750 Byte corresponding to the average data volume to fulﬁll the MSTR. Due to this additional limitation of the

20 1.

Goodput [Mbit/s]

15

10

2.

3.

4.

RT-VR 6.

5.

MSTR

MRTR

5

0 10

20

30

40

50

60

70

80

90

100

Time [s]

Fig. 7. Accurate Traﬃc Shaping Algorithm - Maximum Traﬃc Burst 8750 Byte

allocated data volumes as described in equation 5, the goodput is always equal or below the MSTR. The allocations are equally distributed between all frames leading to a similar simulation result as shown in ﬁgure 4.

5

Conclusion and Future Work

In this paper, we have presented two approaches for the traﬃc shaping in WiMAX networks as speciﬁed by the IEEE 802.16-2009 standard. The ﬁrst proposed accurate traﬃc shaping algorithm fulﬁlls the speciﬁcation in an exact manner, but its memory requirements scales linear with the duration of the Time Base Quality of Service parameter. The second approximate proposal has a constant memory usage, but its memory duration additionally depends on the scheduled data. The simulation results show+, that both algorithms enforce the Maximum Sustained Traﬃc Rate boundary. The accurate algorithm shows a better behavior in cases of heavy changes of the oﬀered traﬃc. Based on both

76

V. Richter and S. T¨ urk

algorithms IEEE 802.16-2009 compatible schedulers can be developed, which satisﬁes the QoS requirements of all types of applications. Future research includes detailed investigations of the behavior of our algorithms in complex traﬃc scenarios and improvements of our scheduling algorithm to support all QoS parameters. Furthermore, an improvement of accurate algorithm is in development to reduce oscillations of the transmission after idle periods in an adaptable manner. We are going to extend our simulation environment with these algorithms to enable realistic WiMAX investigations. Acknowledgments.We would like to acknowledge our former student Mr. Tung Nguyen Khac for his contribution to the approximate traﬃc shaping algorithm. We would like to take this opportunity to thank our project partner Detecon Consulting for sponsoring this work. Special thanks are given to Dr. Petry and Dr. Knospe from Detecon Bonn for their great support.

References 1. IEEE standard for local and metropolitan area networks part 16: Air interface for broadband wireless access systems. IEEE Std 802.16-2009 (Revision of IEEE Std 802.16-2004) (2009) 2. Itu-r document 5d/558: Endorsement of candidate imt-advanced rit based on IEEE 802.16. WP 5D Meeting (October 2009) 3. Ali-Yahiya, T., Beylot, A.L., Pujolle, G.: Radio resource allocation in mobile wimax networks using service ﬂows. In: IEEE 18th International Symposium on Personal, Indoor and Mobile Radio Communications, PIMRC 2007 (Semptember 2007) 4. Fang, W., Seddigh, N., Nandy, B.: A time sliding window three colour marker (tswtcm). RFC 2859 (2000) 5. Ghazal, S., Ben-Othman, J.: Traﬃc policing based on token bucket mechanism for wimax networks. In: IEEE International Conference on Communications, ICC 2010 (May 2010) 6. Wang, W., Sharif, H., Hempel, M., Zhou, T., Wysocki, B., Wysocki, T.: Implementation and performance evaluation of qos scheduling algorithms in mobile wimax ns-2 simulator. In: 4th International Conference on Signal Processing and Communication Systems, ICSPCS 2010 (December 2010) 7. WiMAX Forum: Wimax system evaluation methodology Version 2.1 (July 2008) 8. WiMAX Forum - Application Working Group: Ns-2 mac + phy add-on for wimax (IEEE 802.16) (May 2009)

Multiple-Layer Network Planning with Scenario-Based Traﬃc Forecast Shu Zhang1 and Ulrich Killat2 1

The Quality Group it vision GmbH Hamburg University of Technology [email protected], [email protected] 2

One of the major tendencies in network planning is to take the interactions of multiple network layers into consideration, so that solutions much better than those restricted in single layers can be achieved. This also applies when planning for the extension of the network infrastructure to meet the increment of traﬃc in the future, in which minimizing the capital expenditure of new devices and the corresponding energy consumptions are among the greatest concerns of network operators. However, the planning relies on the forecast of future traﬃc demands, where some degree of uncertainty must be counted as a part. A special forecast mechanism currently used by some network operators is to make a series of forecasts, each being a scenario which describes a possible future situation. In this paper, we will introduce this mechanism, analyze its diﬀerences to other traﬃc uncertainty models, and suggest the corresponding planning methods based on the Integer Linear Programming (ILP) models. The performances are shown in the tests at a scale of real problems. Keywords: Network Planning, Multi-Layer, Traﬃc Uncertainty.

1

Introduction

As a basic principle, telecommunication networks belong to multi-layer architectures. Many researches have shown that the interrelations of multiple network layers can play an important role in network planning, and the planning restricted to single layers is generally sub-optimum, even may fail to meet some critical requests. In this paper, we will present our multi-layer network planning method developed to set a sensible trade-oﬀ between rigid multi-layer optimization and algorithmic complexity. Without loss of generality, we take the following abstract model as reference (see the layers in Fig.1): At the bottom of the network is the physical layer, which includes network elements (nodes) and the physical links (cables/ﬁbers) connecting them; a logical link layer is established above it, which has a quite diﬀerent topology; on the top there are a number of end-to-end traﬃc ﬂows according to the user demands, and they should be routed over the logical layer. Besides, special care must also be taken for the paths of traﬃc demands in the physical layer, so that resilience against single component (physical link or a node) failures in the network can be achieved when required. R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 77–88, 2011. c Springer-Verlag Berlin Heidelberg 2011

78

S. Zhang and U. Killat

In the short time period, the main task of the planner is to assign the paths and resource allocations for emerging new demands, as well as to create new logical links using spare resources in the physical layer when necessary. This is called Operative Planning. In the middle or long term, because the current network is exposed to rapidly increasing services and bandwidth demands, the extension of network infrastructure must also be taken into consideration. The corresponding task is referred to as Tactical Planning. The concepts of the planning task and our reference model of the tactical planning problem are shown in Fig.1.

Fig. 1. The reference model of multi-layer tactical planning

For network operators, the tactical planning is of great importance because it involves a large amount of investment (CAPEX) when the decision of installing new physical infrastructure is made, as well as the corresponding operational costs (OPEX, e.g. maintenance eﬀort, the energy consumption of the installed new devices, etc.). The planning decisions are made according to the forecasted traﬃc matrix. However, because making a precise forecast is practically impossible, some degree of uncertainty must be included in the planning model. Many researches have been made in this area, mainly focusing on two diﬀerent models: (1) The hose model [1][2][3], which only requires in/out bound of the traﬃc across the network, but yields quite loose result; (2) the polyhedral model where multiple traﬃc matrixes are considered [4][5]. However, a special traﬃc uncertainty mechanism, which is currently used by some network operators in Europe, is diﬀerent to any of the above ones. With this mechanism, a series of traﬃc matrix forecasts is made, each is referred to as a scenario. Unlike the polyhedral model, these matrixes follow a sequential order, and each scenario is a subset of the following one. Typically three scenarios, which represent the minimum, the average, and the maximum estimation of the future traﬃc matrix, are made. In the considered time range, the demands are expected to reach the minimum scenario at ﬁrst, and then continue moving to the following scenarios in order. However, it is not sure whether the average

Multiple-Layer Network Planning with Scenario-Based Traﬃc Forecast

79

and the maximum scenario may be reached in reality. A suitable plan must be optimized for all scenarios. Note that the solutions for the demands common to all scenarios must not diﬀer too much in order to minimize reconﬁgurations when moving from one scenario to the next. With these settings, the network planning task includes a suggestion for network infrastructure extension as well as the routing and resource allocation for each demand in all scenarios. The objective of the planning is to minimize the overall cost (CAPEX and OPEX) in each scenario, while the cost of reconﬁguration between neighbor scenarios must also be as low as possible. These features make our planning problem fundamentally diﬀerent from that of the standard polyhedral models. To the best of our knowledge, no research on this kind of problem settings has been made so far. In this paper, we will introduce our planning method for the above problem settings. The method is based on the Integer Linear Programming (ILP) model, while its compromised variant to cope with the scalability issues will also be discussed. The rest of the paper is organized as follows: In Section 2, the problem setting is deﬁned in detail. Then our ILP formulation and the two-phase optimization method are introduced in Section 3. Section 4 shows the numerical results from some tests cases. A conclusion is given in the last section.

2 2.1

Deﬁnition of the Planning Model Network and Traﬃc Modelling

According to the concept of Fig.1, we give the following deﬁnitions. A logical link llog is routed over one or multiple physical links, occupying resource Clog at each hop. This capacity can be used to support end-to-end traﬃc demands. Generally it is enough to set up one path for a demand. For some critical traﬃc with request on resilience, 1 + 1 protection must be established according to some policy of the network operator, so that any single failure in physical layer (e.g. device failure, cable cut) can be tolerated. In a telecom network, generally a number of logical links with diﬀerent capacities has already been established, and their resources are already being utilized to some extent. To route a new traﬃc demand, the spare capacity of a logical link Cllog can directly be used. Besides, since the current network conﬁguration is known, the routing of every logical link at the physical layer is also available. Denote the set of physical links that the path of a logical link llog traverses as Sphys (llog ), and the set of logical links which traverse the same physical link lphys as Slog (lphys ). It is quite common that the raw resources at physical links have not been fully occupied by the logical links at the moment. Therefore a physical link lphys may also have spare some resource C(lphys ). In our model, both physical and logical links are considered as abstract links l ∈ L in our planning model, each with spare resource C(l). Therefore, we are planning the routings for the forecasted traﬃc demands Di , i ∈ I, in a graph G(V, L) where V is the set of all nodes.

80

2.2

S. Zhang and U. Killat

The Optimization Model for Network Planning

The Plan in Each Scenario. For each individual scenario, the plan includes the following suggestions to the planner: 1. The routing for each future demand, in which eﬃcient resource utilization is preferred. In our model, the cost of resource utilization is expressed by a certain amount of OPEX. 2. The necessary hardware extensions, which claim a certain amount of CAPEX (purchasing and installation) as well as OPEX (maintenance, energy consumption, etc.). 3. Optionally, the modiﬁcation of the logical links which can help to reduce the cost of 1) and 2) but also requires some updating work and therefore the corresponding cost (considered as CAPEX in this paper). Since hardware extension involves the majority of cost, a suitable plan for that is the central part of our model. According to limitations in practice (devices type, free interfaces, etc.), we cannot arbitrarily install any type of physical link between a pair of nodes in the network. A speciﬁc hardware extension must be selected from a set of choices, which is referred to as an extension pattern in this paper. The pattern should be predeﬁned by the network operator (see for example Table.1). Table 1. Example of an extension pattern Option Type Amount Unit Cost

1 1 Gbps cable 1 1 million

2 1 Gbps ... 2 2 million

3 10 Gbps ... 1 4 million

... ... ... ...

In our model, the hardware extension is considered as an extra resource to be installed at a physical link lphys . Note that the link can be either an existing one, or a planned link which is a candidate of the possible future network infrastructure. The optimization algorithm is responsible for making a choice from the options in an extension pattern at each link, according to the principle of ensuring enough resources to support the forecasted traﬃc matrix with minimum cost. We deﬁne a binary variable t(s, l, k), which takes the value 1 when the kth option is chosen at link l in scenario s, and 0 for not making this choice. Note that the decision can be 0 for all k, which means no extension at this link is needed. The Reconﬁguration Between Scenarios. If each scenario is planned independently, waste of investment may possibly occur. Assume that the following plan for hardware extension was made for scenario s1 : t(s1 , l1 , 1) = 1 and t(s1 , l, k) = 0, ∀l = l1 , ∀k. Besides, an independent plan was suggested for scenario s2 : t(s2 , l2 , 1) = 1 and t(s2 , l, k) = 0, ∀l = l2 , ∀k. Although it seems that the cost in each scenario is minimized because there is only one hardware extension for each scenario, the whole plan is still problematic: The extension at

Multiple-Layer Network Planning with Scenario-Based Traﬃc Forecast

81

l1 has been installed in the ﬁrst scenario s1 , but is no longer utilized in the next scenario s2 , which means the waste of investment on l1 . This is deﬁned as a hard reconfiguration and should be avoided in the planning. Furthermore, in case each scenario is planned independently, the planning for the routing of demands and modiﬁcations of logical links can be quite diﬀerent. Although these diﬀerent conﬁgurations require only some parameter modiﬁcations, it is still a tedious task when the amount of necessary reconﬁguration is large, and the operation can be a signiﬁcant source of error. Therefore, the following diﬀerences between neighbor scenarios should also be minimized: 1) The routing of each demand; 2) The conﬁguration of each logical link. Because the two operations are very similar in nature (parameter updates to be made by technicians), they are both referred to as soft reconfigurations. 2.3

The Optimization Model

Based on the above analysis, we suggest a 2-phase optimization model. Phase 1. Each scenario is planned independently with the objective of minimizing the cost. The optimum solutions can show the minimum costs Costopt (s) of each scenario s ∈ S, which are used as reference values in the next phase. Phase 2. Now, the objective is to minimize the diﬀerences (i.e., in terms of hard and soft reconﬁgurations) between neighbour scenarios. We wish that the overall cost in each individual scenario is not signiﬁcantly larger than the optimum cost obtained in the ﬁrst phase. Practically, a relaxation factor F (e.g. F = 10%) is deﬁned so that the ﬁnal cost Cost(s) in each scenario is bounded by: Cost(s) ≤ Costopt (s) · (1 + F )

3 3.1

(1)

The ILP Formulation Model of Phase 1

The Flow Continuity Constraint. For each demand, an end-to-end path should be established. We deﬁne a binary decision variable x(s, i, l, p), which takes the value 1 if the path p of demand i traverses link l in scenario s, otherwise 0. By the node-link formulation of Eq.2, p end-to-end paths can be established (p ∈ P (i) where P (i) is the number of paths required by demand i for resilience). ⎧ ⎨ 1 if j = v (2) x(s, i, l, p) − x(s, i, l , p) = −1 if j = u ⎩ 0 otherwise l(m,j)∈L l (j,n)∈L ∀i(u, v) ∈ I(s), ∀j ∈ V, ∀p ∈ P (i) Here, I(s) is the set of all demands in scenario s, and i(u, v) is the demand i with source node u and destination node v. l(m, n) denotes a directional link from node m to n. The equation means that for a node j, the incoming traﬃc must be equal to the outgoing traﬃc, while traﬃc at the entrance and the exit are ﬁxed to 1.

82

S. Zhang and U. Killat

The Link-Disjointness Constraint. For a demand requiring multiple paths, each of the paths should be disjoint in the physical layer. However, we are making the plan in a multi-layer network, where both logical links and physical links are considered as an abstract link in the set L. Therefore, the end-to-end paths obtained by Eq.2 may contain a mix of both physical and logical links. In order to check the disjointness in the physical layer, we deﬁne a binary variable y(s, i, l, p) to track the exact path of demand i in the physical layer. It has the same meaning as x(s, i, l, p) but is deﬁned only at physical links. There are two cases to be checked. The ﬁrst case is when path p traverses a physical link. Eq.3 ensures that y(s, i, lphys, p) is equal to 1 if x(s, i, lphys , p) takes the value of 1 (in this case the path is routed over the physical link lphys ), otherwise unrestricted. y(s, i, lphys , p) ≥ x(s, i, lphys , p), ∀lphys ∈ Lphys , ∀i ∈ I(s), ∀p ∈ P (i)

(3)

The second case is when path p traverses a logical link. Now, as a tracking variable, y(s, i, lphys , p) must be 1 if x(s, i, llog , p) takes the value of 1 and lphys is a link on which the logical link llog is routed, otherwise unrestricted. The relationship of logical link routed on physical link is recorded in the set Sphys (llog ) (see Section 2.1), which can be obtained by analyzing the routing before establishing the ILP model. y(s, i, lphys, p) ≥ x(s, i, llog , p)

(4)

∀llog ∈ Llog , ∀lphys ∈ Sphys (llog ), ∀i ∈ I(s), ∀p ∈ P (i) Since y(s, i, l, p) records the paths in the physical layer, it is straightforward to establish the link-disjointness constraint. Eq.5 ensures that a physical link l cannot be traversed more than once by the paths p ∈ P (i) of demand i. y(s, i, l, p) ≤ 1, ∀l ∈ Lphys , ∀i ∈ I(s) (5) p∈P (i)

The Resource Utilization Constraint. This is the basic rule that the total resources claimed by new demands on a link should not exceed the spare resources at the link. A logical link l has spare resource C(l) which can be modiﬁed if necessary. We use the variable δ(s, l) to indicate the amount of modiﬁcation in scenario s at link l, so that the actual spare resource becomes C(l) − δ(s, l). Note that δ(s, l) can take positive as well as negative value, which implies reduction or increment of spare resources. Let R(s, i) be the resource requirement of ﬂow i in scenario s, the resource utilization constraint is therefore as follows: x(s, i, l, p) · R(s, i) ≤ C(l) − δ(s, l), ∀l ∈ Llog (6) i∈I(s),p∈P (i)

For a physical link, there are two ways that its spare resource may be modiﬁed: 1) From the modiﬁcation of logical links routed over it. 2) Due to the installation

Multiple-Layer Network Planning with Scenario-Based Traﬃc Forecast

83

of new hardware extension. Let the kth choice of the extension pattern at link l provides an extra resource Cext (l, k). Then, the constraint on physical links is: x(s, i, l, p) · R(s, i) ≤ i∈I(s),p∈P (i)

C(l) +

δ(s, llog ) + t(s, l, k) · Cext (l, k), ∀l ∈ Lphys

(7)

llog ∈Slog (l)

The discussion of the set Slog (l) is referred to Section 2.1. Eq.6 and Eq.7 imply that the modiﬁcation of logical links is restricted to the adjustment of capacity. A more general model should also allow the rerouting of logical links. Although it is theoretically feasible to establish an equation set of ﬂow continuity conditions for each logical link similar to Eq.2, such a multi-layer model can result in an explosive enlargement of the problem scale, which makes even much smaller problems than our test cases (see Section 4) unsolvable in acceptable time. In this paper, We will focus our discussion on the above analyzed model. The variable δ(s, l) determines the amount of capacity change at a logical link. Since a modiﬁcation to any value involves the same amount of reconﬁguration work, a binary variable d(s, l) is used to track whether any modiﬁcation has taken place. It should take the value 1 if δ(s, l) is nonzero, and 0 otherwise. Eq.8 describes such a relationship between d(s, l) and δ(s, l). d(s, l) · W ≥ δ(s, l), d(s, l) · W ≥ 0 − δ(s, l), ∀l ∈ Llog

(8)

Here, the constant W satisﬁes W |δ(s, l)|, ∀l. The Objective. The optimization objective of phase 1 is to minimize the realization cost in the given scenario s. There are 3 types of cost: 1) utilization of resources at each link (OPEX); 2) The modiﬁcation of logical links (CAPEX); 3) The implementation of hardware extensions (CAPEX and OPEX). Denote the cost for demand i to use a unit of resource at link l as Wx (i, l), the cost of modifying a logical connection as Ws , and the cost of implement the kth choice of hardware extension at link l as Wt (l, k), then the cost function is: x(s, i, l, p) · R(s, i) · Wx (i, l) Cost(s) = i∈I(s),p∈P (i),l∈L

+

l∈Llog

d(s, l) · Ws +

t(s, l, k) · Wt (l, k)

(9)

l∈Lphys ,k∈Kl

By the calculation in phase 1, the cost of the optimum solution Costopt (s) in each scenario s ∈ S can be obtained. These costs will serve as references for the next phase calculation. 3.2

Model of Phase 2

In phase 2, the network planning constraints of phase 1 remain unchanged, while there are the following new features:

84

S. Zhang and U. Killat

– All scenarios must be considered in an integrated planning problem, rather than only one scenario is planned independent of the others. – The soft and hard reconﬁgurations between neighbor scenarios should be minimized. – The actual cost of each scenario is upper-bounded. In the ideal case, it should also be minimized. Deﬁne the set SN of all pairs of neighbour scenarios: (s1 , s2 ) ∈ SN , and between these pairs comparisons should be made. We further require that the pairs in SN are ordered, where the ﬁrst scenario is a subset of the second one. The Soft Reconﬁguration. If the capacity modiﬁcations are diﬀerent for the logical links in the compared pair of scenarios, a reconﬁguration is necessary. Following the style of Eq.8, the corresponding inter-scenarios reconﬁgurations can be tracked by the binary variable ds (s1 , s2 , l) with the following equation: ds (s1 , s2 , l) · W ≥ δ(s1 , l) − δ(s2 , l) ds (s1 , s2 , l) · W ≥ δ(s2 , l) − δ(s1 , l)∀(s1 , s2 ) ∈ SN , ∀l ∈ Llog

(10)

The constant W should also be larger than any possible value of δ(s, l). For the routing of demands, a reconﬁguration dr (s1 , s2 , i) takes place when any segment of the paths in the neighbour scenarios is diﬀerent, as in Eq.11: dr (s1 , s2 , i) ≥ x(s1 , i, l, p) − x(s2 , i, l, p) dr (s1 , s2 , i) ≥ x(s2 , i, l, p) − x(s1 , i, l, p)

(11)

∀(s1 , s2 ) ∈ SN , ∀i ∈ I(s1 ) ∩ I(s2 ), ∀p ∈ P (i), ∀l ∈ L The Hard Reconﬁguration. A hard reconﬁguration happens when we make diﬀerent choices of hardware extension at a link. Deﬁne the binary variable dh (s1 , s2 , l) for this check: dh (s1 , s2 , l) ≥ t(s1 , l, k) − t(s2 , l, k), ∀(s1 , s2 ) ∈ SN , ∀l ∈ Lphys , ∀k

(12)

Note that the above equation only checks whether a hardware extension applied in s1 also remains in s2 , while not doing the same in the reversed direction. This is because s1 is generally a subset of s2 ; the extra hardware extension in s2 is natural. The Cost In Each Scenario. We wish that the ﬁnal solution for each scenario is not too far away from the optimum one, i.e. within a slightly relaxed bound as Eq.13. The excess cost recorded by the variable Costexc (s)(≥ 0) will later on be counted into the objective function of phase 2. Cost(s) ≤ Costopt (s) · (1 + F ) + Costexc (s)

(13)

Multiple-Layer Network Planning with Scenario-Based Traﬃc Forecast

85

The Optimization Objective. As the objective, we mainly wish to reduce the amount of reconﬁgurations. Let Ws and Wh be the cost of a soft and a hard reconﬁgurations. Then, the objective function can be deﬁned as: ds (s1 , s2 , l) · Ws Costcf g = (s1 ,s2 )∈SN ,l∈Llog

+

dr (s1 , s2 , i) · Ws

(s1 ,s2 )∈SN ,i∈I(s1 )∩I(s2 )

+

dh (s1 , s2 , l) · Wh

(14)

(s1 ,s2 )∈SN ,l∈Lphys

Although the actual cost in each scenario is bounded by Eq.13, in the ﬁnal solution it tends to locate at anywhere below the bound. If F is loose, e.g., 10%, there is still much room to improve. Therefore, it is meaningful to let the costs in each scenario also take place in the objective function so that they can be minimized as well. In this case, the value of Ws and Wh should be carefully chosen (generally a large value) so that to minimize the reconﬁguration remains our main objective in this phase. Besides, the excess cost from Eq.13 should also be counted with a punishment factor Wp . The overall objective function now becomes: Cost = Costcf g + Cost(s) + Costexc (s) · Wp (15) s∈S

4

s∈S

Numerical Results

Here we show some examples of numerical results obtained with the network topology derived from DFN XWiN [7], a scientiﬁc research network with nodes in major cities of Germany, and ﬁbres connecting them to form a meshed topology. Our test network consists of 54 nodes and 81 physical links from the core of XWiN. The capacities of links range from 1Gbit/s to 20Gbit/s as established in XWiN. Then, 100 logical links are randomly generated using the resource of physical links, which takes 70% to 80% of the physical resources into the logical layer. Finally, 20% to 80% (random even distribution) of the total capacity at each logical link is marked as occupied to emulate the current utilization. As a comparison, ﬁrstly we try to solve the same planning problem with a greedy method based on sequential shortest path calculation for each demand (for those requiring physical disjoint paths, the algorithm suggested in [6] is used). When a solution does not exist for the current demand, at ﬁrst a soft reconﬁguration is tried, then a hardware extension. This behavior is very close to that of a human planner. We denote the algorithm as Greedy Sequential Algorithm (GSA). The same procedure is repeated for ten thousand times and the best solution recorded. Then, we solve the same problem with our ILP model introduced in the previous section. There are two diﬀerent models: the exact model which strictly follows the formulation of Section 3; the compromised model in which paths for

86

S. Zhang and U. Killat

identical demands are same in each scenario, which is designed to alleviate the scalability problem. The results shown in the following are collected from 10 times repeating the same procedure: randomly generate a speciﬁc number of demands (to make the task heavy, each of them requires a pair of physical disjoint paths). 1/3 of the demands are assigned to the minimum scenario, 2/3 demands including those in the minimum scenario are assigned to the average scenario, and all demands are assigned to the maximum scenario. The capacities of the demands are randomly generated with an even distribution, where the value range is carefully selected so that the resource of the current network is only suﬃcient to accommodate roughly 80% to 90% of the new demands, i.e. hardware extension is necessary. In this test, we assume that all physical links follow hardware extension patterns with 3 diﬀerent choices as deﬁned in Table.1.

Cost in comparison to Greedy (%)

100

90

80

70 ILP ILP Compromised GSA

60

50 20

40

60

80

100

120

140

160

Number of demands

Fig. 2. Comparison of OPEX

To compare the performance of the planning methods, it is straightforward to compare the realization cost using Eq.14 and Eq.15. However, the optimization objective of phase 2 indicates an abstract cost strongly depending on the values of the weight parameters Ws Wh and Wp . Diﬀerent parameter settings may lead to signiﬁcantly diﬀerent results. Therefore in Fig.2, we separately compare the cost of resource utilization in the maximum scenario, taking the cost of the GSA solution as reference (100%). The situations in the average and minimum scenarios are very similar. This result shows the saving of OPEX by the optimized routing solution obtained through the ILP model. Note that the solution quality of the exact and the compromised models are very close. The plan of hardware extension involves the majority of costs in the network management. The results of Fig.3 shows fewer hardware extension in the solution obtained by our ILP model (same result from the exact and the compromised model), which indicates tremendous capital saving in reality.

Multiple-Layer Network Planning with Scenario-Based Traﬃc Forecast

87

Number of extension in comparison to Greedy (%)

100

80

60

40

20

GSA ILP

0 20

40

60

80

100

120

140

160

Number of demands

Fig. 3. Comparison of the number of hardware extensions

Finally, the calculation time of each method is compared in Fig.4. The scalability of the compromised ILP model, where routings of demands in all scenarios are the same, is much better than the exact model allowing the routing reconﬁguration. However, the qualities of both solutions are very close as shown in Fig.2 and Fig.3, and both solutions are much better than that of the greedy algorithm. Therefore, the compromised model is generally more suitable in solving practical problems, especially when the scale of the problem is large.

Time till gap below 0.01 (Sec)

20000

ILP ILP Compromised GSA (1E+4 rounds)

15000

10000

5000

0 20

40

60

80

100

120

Number of demands

Fig. 4. Comparison of scalability

140

160

88

5

S. Zhang and U. Killat

Conclusion

In this paper, we introduced the concept of tactical planning for a telecom network in the background of a multi-layer network optimization model. Here, the plan for the routing and resource allocation of future demands as well as the infrastructure extension in the middle to long term are to be made. Since the extension of network involves large amount of investment and directly inﬂuences the energy consumption, it plays a central role in the suggested planning method. Furthermore, the problem of the uncertain estimation of future traﬃc has been introduced, and a speciﬁc scenario-based forecast mechanism currently used by some network operators has been discussed. To solve the corresponding network planning tasks, we proposed a two-phase optimization method using ILP models. Through the tests with an XWiN network topology, it has been shown that our optimization model can ﬁnd satisfactory solutions in acceptable calculation time, while the quality of the planning solution is much better than that obtained by repeated calculation with greedy algorithms, which are close to the behavior of a human planner. Generally, the planning solution suggested by our ILP model requires much less hardware extensions than the greedy algorithms, which means a great reduction of one of the major cost factors for a network operator.

References 1. Duﬃeld, N.G., Goyal, P., Greenberg, A., Mishra, P., Ramakrishnan, K., van der Merive, J.: A ﬂexible model for resource management in virtual private networks. SIGCOMM Comput. Communication Rev. 29(4), 95–108 (1999) 2. Fingerhut, J.A., Suri, S., Turner, J.S.: Designing least-cost nonblocking broadband netowrks. Journal of Algorithms 24(2), 287–309 (1997) 3. Mulyana, E., Killat, U.: Optimizing IP netowrks for uncertain demands using outbound traﬃc constraints. In: Proceeding of INOC 2005, pp. 695–701 (2005) 4. Ben-Ameur, W., Kerivin, H.: Routing of uncertain demands. Optimization and Engineering 3, 283–313 (2005) 5. Applegate, D., Cohen, E.: Making Intra-Domain Routing Robust to Changing and Uncertain Traﬃc Demands: Understanding Fundamental Tradeoﬀs. In: Proceedings of ACM SIGCOMM, pp. 313–324 (2003) 6. Xu, D., Xiong, Y., Li, G.: Trap Avoidance and Protection Schemes in Networks with Shared Risk Link Groups. IEEE Network 18(13), 36–41 (2004) 7. German Research Network (Deutsches Forschungsnetz DFN), The Scientiﬁc Network X-WiN, http://www.dfn.de/xwin/

Optimization of Energy Eﬃcient Network Migration Using Harmony Search Stefan T¨ urk and Rico Radeke Dresden University of Technology Dresden, Mommsenstrasse 13 01062 {tuerk,radeke}@ifn.et.tu-dresden.de

Abstract. In this paper we describe the basic network migration problem in backbone networks to move from an existing to a new technology. Furthermore we use a generic harmony search algorithm to optimize the solution in terms of energy eﬃciency and costs. Harmony search is a probabilistic meta-heuristic which has been successfully adapted to many optimization problems. We analyze how harmony search can be used to calculate migration sequences with minimized energy consumption and ﬁnancial costs. Results of parameter studies for the heuristic will be shown to evaluate the method. The achieved resource utilization to cover increasing network demands and the point of introduction within a certain time interval will be presented also.

1

Introduction

Traﬃc demand in optical backbone networks is expected to increase rapidly in the next years [1]. Also the income of network providers is constantly decreasing due to shrinking tariﬀs for end users. Furthermore the hardware utilization and therefore the energy consumption at each Point of Presence will rise strongly. This perspective oﬀers opportunities for signiﬁcant cost and energy reduction. Our paper describes how network operators can cope with the increase of backbone traﬃc in the next years. Network migration describes a technical process of upgrading and exchanging existing hardware/software (infrastructure) to other network devices, which generate calculable cost savings for its owner. The shift of classical Synchronous Digital Hierarchy (SDH) services towards cost eﬃcient Ethernet services is currently under way, i.e. for high data rate Internet access or global interconnection of company locations [2, 3]. The introduction of Carrier Ethernet services is heavily investigated by most providers, since it promises signiﬁcant cost savings and simpliﬁcations in terms of administration and maintenance [4]. In this paper we will describe how a generic Harmony Search (HS) metaheuristic can be used to optimize the multi-layer and multi-period network migration problem for core networks. As a result a proposal for a possible resource utilization and scaling will be given. HS was chosen to be investigated since R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 89–99, 2011. c Springer-Verlag Berlin Heidelberg 2011

90

S. T¨ urk and R. Radeke

the heuristic is adaptable to the initial problem (see section 2) and was described as a fast converging algorithm by its authors [5]. Previously an Ant Colony Optimization (ACO) meta-heuristic [6] had been applied to the problem [7]. The system model of the migration approach was applied to a widely used IP/MPLS/SDH/DWDM1 network. A possible future scenario and migrated network is an IP/MPLS/ETH2/DWDM architecture. Both architectures are discussed in detail in [8]. Previous investigations on energy consumption in backbone networks have been shown in [9]. An energy optimization approach using the Nobel-2 [10] architecture has been discussed in [11]. Energy consumption of diﬀerent devices was analyzed in [12, 13] and used in our investigation. The remaining paper is structured as follows. Section 2 describes the harmony search agorithm and its adaptation to the migration process. In subsection 2.4 we shortly present the used simulation system and the deﬁned scenarios. In section 3 we discuss achieved results and ﬁnish the paper in section 4 with conclusions.

2

Algorithm Description

2.1

Network Migration

The process of network migration can be described as the stepwise insertion, replacement or removal of components in a network to change the technology. In our current approach (multi layer migration) Optical Cross Connect (OXC)s, Carrier Grade Ethernet switches (MPLS Transport Proﬁle (MPLS-TP) switches) and IP routers can be added to a network to satisfy yearly increasing demands. Within an incremental network migration only the next planning state is used to calculate the necessary hardware [14]. In our approach also potential future traﬃc is considered. Our network migration model implements for instance two important marketdriven factors: on the ﬁrst hand the network demand increase (40% annual trafﬁc increase [1]) and on the other hand a cost erosion of devices (20% yearly Capital Expenditures (CAPEX) decrease [15]). The detailed description of this cost model can be found in [8]. The energy model is device speciﬁc and was approximated for our investigations using the assumptions made by [13] (Fig. 1). Basically it has been stated that an IP Router (IPR) consumes about 1200 Watt/100GBit/s, an ETH switch about 800 Watt/100GBit/s and an OXC about 180 Watt/100GBit/s. 2.2

Basic Solution

The basic Migration Solution (MS) can be generated using diﬀerent approaches. The ﬁrst option is to directly calculate a minimal necessary device situation at 1 2

Internet Protocol(IP)/Multiprotocol Label Switching(MPLS)/Synchronous Digital Hierarchy(SDH)/Dense Wavelength Division Multiplex(DWDM). Ethernet(ETH).

Optimization of Energy Eﬃcient Network Migration Using Harmony Search

91

Fig. 1. Energy consumption for diﬀerent switch types [13]

each node for each time step. The second option is to generate a completely randomized but feasible solution within the solution space of the migration that is also able to cover all appearing demands for each step. Both approaches contain the Greenﬁeld planning result or a current device situation in the network. We denote this starting situation as migration step 0 (see Fig. 2).

Fig. 2. Exemplary network migration solution structure

The implemented device situation at each ﬁeld of the matrix in Fig. 2 contains information about each device at each node that has to be present at the node in a time step i to cover the demands of the backbone. These device granularities zk,i of node k mark the decision variables later used in section 2.3. Each variable has a minimum and maximum possible value (as deﬁned by the ﬂags, Tab. 1), which is deﬁnied by the used system model. Decision variables are stated in Tab. 1, a value of 0 indicates that a device of this type is not present at the node in this time step i. A ”v” deﬁnes a virtual meshed device, which would practically be a multi-chassis installation of IPR’s, ETH’s or OXC’s. The cost or

92

S. T¨ urk and R. Radeke

Table 1. Device granularities with cost model and power consumption [10, 12, 13] Type

Var

Flag

CAPEX [CU] Power[KW]

IPR IPR IPR ... IPR

z IP R k

f IP R M IN

0 16.67 111.67

0 2.92 14.94

f IP R 315.83 M AX

35.83

0 640 1280 5760

ETH ETH ETH ETH ... ETH

0 240 1200 2400(v)

OXC OXC OXC ... OXC ... OXC

0 3 Port 4 Port

9600(v)

9 Port(v) 18 Port(v)

z ET H f ET H 0 k M IN 12.5 50 100 f ET H 400 M AX

0 0.73 9.33 18.66 74.64

z OXC f OXC 0 k M IN 21.24 28.32

0 2.16 2.87

63.72

6.48

f OXC 127.44 M AX

12.96

energy model for virtual devices remains unclear as coupled devices loose capacity due to internal connections. These eﬀects are subject to further investigations. The generated basic MS can be optimized using meta-heuristics. One approach is to use harmony search as optimization algorithm, which will be described in the following. 2.3

Harmony Search

Harmony Search was ﬁrstly introduced 2001 by Geem et al. [5] and has been improved in several other works [16, 17]. The algorithm has been described very well in [5], but shall be shortly explained in the following to describe how it can be adapted to the network migration problem. HS is based on the idea, that musicians have several options to generate new songs. Basically they can modify an existing song or they can implement copied parts of multiple songs into a new song. They can also improve a totally new (random) harmony. HS was proposed as a very eﬀective optimization heuristic for Traveling Salesman Problem, routing and timetabling, containing ﬁve major steps [5]. I. Initialize solution variables and parameters II. Fill harmony memory with randomized solutions III. Generate a feasible random solution within search space and apply changes from a solution found in Harmony Memory (HM) IV. Vary the generated solution according to predeﬁned probabilities (called pitch adjusting) V. Repeat steps III and IV until a stop criterion is reached The deﬁned steps of the meta-heuristic can be applied to the initial network migration problem. In the ﬁrst step all parameters for the problem have to be set (e.g. the size of the HM or probabilities for algorithm decisions). The fret width

Optimization of Energy Eﬃcient Network Migration Using Harmony Search

93

fw , which is the range of possible granularity changes, has to be set in step I also. After the initialization of variables the HM is ﬁlled with randomized solutions until the HM’s maximum size LHM,max is reached. Here randomized solutions have to be provided since otherwise the best solution from the HM would be always the semi-optimal deterministic ”minimal necessary device”-solution. The ﬁrst part of step III is equal to the description provided in 2.2 as a new solution Sn is constructed here. The second part of the third step delivers the most important changes to a new and potential good migration solution Sn . According to the parameter Harmony Memory Considering Rate (HMCR) the algorithm d decides to fetch an existing decision variable zk,i (where d is the investigated device type) from the HM. The decision variable for current step and current node is randomly taken from the HM. With a probability of Pnew = 1 − HM CR a new randomized decision variable is generated. After the restructuring of the solution either via HM variables or with randomized variables a pitch adjustment can be applied. This means the current decision variable of the current step can be tuned up or down (in our case in terms of device granularities). After pitch adjusting the solution can be compared to the worst solution Sw of the HM. If the Sn has a smaller Total Cost of Ownership (TCO) (where TCO can also mean total power consumption) than Sw it is a potential good solution for the next iteration and replaces Sw in HM. When the stop criterion is reached the algorithm terminates and returns the best found migration solution. These major steps are stated in Alg. 1.

Algorithm 1. Pseudo code for HS implementation Specify parameters while LHM < LHM,max do Generate balanced random solution Add solution to HM end while Sort HM descending for all iterations do (Calculate PAR) Generate new solution Sn for all Solution Elements e ∈ E of Sn do for all Device types d ∈ D of e do Generate new random variable rnd1 if rnd1 < HM CR then Return random device z d from HM k,i Generate new random variable rnd2 if rnd2 < P AR then Do pitch adjusting for z d with fret width fw k,i end if Set device type z d for element e k,i Balance traffic at node else Generate new random device type z d k,i Balance traffic at node end if Store e in Sn end for end for Calculate cost cn of Sn Fetch worst solution Sw from HM Calculate cost cw of Sw if cn < cw then Replace Sw in HM with Sn end if end for print best S of HM

94

2.4

S. T¨ urk and R. Radeke

Scenarios

Two backbone networks were investigated, the German 17-node (Fig. 3) backbone network and the European 67-node backbone. Due to space limitations only the obtained results for GER-17 are discussed in this paper. To compare the results with a simple hardware insertion algorithm a Selective Random Search Heuristic (SRSH) has been implemented as well. This algorithm represents an optimization algorithm based only on feasible randomly generated solutions and their direct comparison either using TCO or power consumption only. The trafﬁc model of both scenarios is based on a population investigation that has been described more detailed in [7]. The total migration period is set to 5 years, where the ﬁrst year is a greenﬁeld planning result or existing network. Since HMCR and Pitch Adjusting Rate (PAR) strongly aﬀect the result of the migration solution both parameters have been investigated in detail (Tab. 2) to conﬁgure the algorithm optimal for the migration problem and to ﬁnd potential dependencies in the scenario. Table 2. Evaluated heuristic conﬁgurations Configuration HMS HMCR PAR version P ARmax P ARmin C-I C-IIa C-IIb C-IIIa C-IIIb C-IVa C-IVb

10 10 10 10 10 10 10

0.85 0.85 0.85 0.85 0.85 0.85 0.85

no PAR fix fix P ARv1 P ARv1 P ARv2 P ARv2

C-V C-VIa C-VIb C-VIIa C-VIIb C-VIIIa C-VIIIb

10 10 10 10 10 10 10

0.95 0.95 0.95 0.95 0.95 0.95 0.95

no PAR fix fix P ARv1 P ARv1 P ARv2 P ARv2

C-X

0

0 0.1 0.01

0.5 0.05 0.5 0.05

0.1 0.01 0.1 0.01

0

0 0.1 0.01

0.5 0.05 0.5 0.05

0.1 0.01 0.1 0.01

no migration

Improved dynamic PAR equations (Eq. 1 (proposed by [17]) and Eq. 2), had been applied to the problem to enhance the total optimization process.

with: it NI P AR(it) P ARmax P ARmin

... ... ... ... ...

P ARv1 (it) = P ARmin +

P ARmax − P ARmin ∗ it NI

(1)

P ARv2 (it) = P ARmax −

P ARmax − P ARmin ∗ it NI

(2)

current iteration step total number of iterations (predeﬁned) probability within iteration step it to do a pitch adjusting maximum predeﬁned probability to do a pitch adjusting minimum predeﬁned probability to do a pitch adjusting

Optimization of Energy Eﬃcient Network Migration Using Harmony Search

95

Load distribution German 17 node backbone

Ulm

Load Ulm

Munic

17

Munic

Nuremberg

Stuttgart

16

Karlsruhe

13

Nuremberg

Karlsruhe 14 15 Stuttgart

Leipzig

11

12

Frankfurt

Frankfurt Mannheim

Leipzig

Mannheim

10

Dortmund

Dortmund

8

Essen

4

9

Cologne

Hanover

Duesseldorf

7

Duesseldorf Cologne

Berlin 5

Berlin

6

Node From

Essen

Hanover

2 3

Bremen

Norden

1

Munic Ulm Stuttgart Karlsruhe Nuremberg Mannheim Frankfurt Leipzig Dortmund Cologne Duesseldorf Essen Hanover Berlin Bremen Hamburg Norden Bremen

Hamburg

Hamburg

Norden

Node To

(a)

(b)

Fig. 3. German 17 (GER-17) node backbone scenario (a) and demand situation (b) [8]

3

Results

The result of our research regarding migration algorithms is shown using different visualisations (algorithm performance and detailed view on migration result). All parameters have been investigated twice, once for cost and once for the power consumption. Stop criterion for the HS has been deﬁned with a maximum amount of 100k iterations. Optimization calculations have been repeated 80 times independenly to guarantee statistical conﬁdence. In Fig. 4 and 5 the performance of the algorithms and the HS variations (CI-CIIIb) are presented. Here CX is the best found cost solution if a network migration is not performed (just IP routers are upgraded), CX is used to compare our results to the providers usual incremental device upgrade. It can be seen that a usage of a too low HMCR can result in too many random changes. Therefore the ﬁnal power usage of the migration solution will be higher than necessary. Preliminary tests showed that on the other hand a usage of no random decisions results in a very bad algorithm performance. When using a HMCR of 0.95 (Fig. 5) the total algorithm performance is much better, as best algorithm conﬁguration CIIIb was found. A HMCR of 0.95 was chosen according to the default value of the HS algorithm, there might be HMCR values that perform better for the given network scenario. P ARv1 does not enhance the migration result in our case, due to many random decisions in the beginning of the iteration phase. P ARv2 improves the power savings of the migration solution up to 6.5% (compared to CVIb). The size of the HM has not been changed for the results of this paper. Fig. 6 and 7 present a detailed view on the respective best migration result. For cost eﬃcient solutions Ethernet switches (ETH Basic Node (ETHBN)) have to be purchased early (here year 2), for the energy optimal solution electric devices (also IP Port Card (IPPC)s and ETH Port Card (ETHPC)s) have to be bought as late as possible and optical equipment (Optical Basic Node (OPBN)) is preferred early. In both cases the algorithm sells IP Basic Node (IPBN) early if possible.

96

S. T¨ urk and R. Radeke

HS (HMCR=0.85), SRSH, 100000 Iterations, OPT=POW, 1−α=0.95 CI CIIa CIIb CIIIa CIIIb CIVa CIVb SRSH CX

Total power consumption (kW)

5000

4000

3000 10

100

1000

10000

100000

# of iterations

Fig. 4. Algorithm performance for HMCR=0.85 for diﬀerent conﬁgurations HS (HMCR=0.95), SRSH, 100000 Iterations, OPT=POW, 1−α=0.95 CV CVIa CVIb CVIIa CVIIb CVIIIa CVIIIb SRSH CX

Total power consumption (kW)

5000

4000

3000

2000 10

100

1000

10000

# of iterations

Fig. 5. Algorithm performance for HMCR=0.95 for diﬀerent conﬁgurations

100000

Optimization of Energy Eﬃcient Network Migration Using Harmony Search Resoure utilization for GER−17, OPT=POW, best of 100000 HMY iterations 500 POWERSUM OPBN IPBN IPPC ETHBN ETHPC

Partial power consumption (kW)

400

300

200

100

0 1

2 3 Migration step (years)

4

Fig. 6. Best found energy eﬃcient solution for GER-17 Resoure utilization for GER−17, OPT=TCO, best of 100000 HMY iterations CAPEX IMPEX OPEX OPBN IPBN IPPC ETHBN ETHPC

1600

1400

Partial expenses (CU)

1200

1000

800

600

400

200

0 1

2

3

4

Migration step (years)

Fig. 7. Best found cost eﬃcient solution for GER-17

97

98

4

S. T¨ urk and R. Radeke

Conclusion and Future Work

In this paper we have presented a methodology to solve the network migration problem using a harmony search algorithm. The evaluation of the results showed that harmony search is adoptable to the initial problem formulation and performs better than SRSH. The comparison of diﬀerent parameters for this approach showed promising working conditions for the algorithm, which will be further investigated. Generally it was shown, that simple improvements of the harmony search method can result in signiﬁcant energy savings for network providers. Future research in the ﬁeld of optimization of network migration is the evaluation and comparison of other types of algorithms (e.g. genetic) as well as the improvement of existing approaches. The investigation of device meshing and the inﬂuence on total migration cost and energy consumption will also be done. Acknowledgment. The work presented in this paper is a result of the CELTIC project 100GET-E3, which has been partially supported by Nokia Siemens Networks GmbH & Co. KG and the German Federal Ministry of Education and Research (BMBF) under grant 01BP0740.

References 1. Cisco: Cisco visual networking index: Forecast and methology, 2008-2013, Tech. Rep. (2009) 2. Kiy, N.: Carrier-ethernet: Transportnetz f¨ ur next generation networks. ntz 3-4, 28–29 (2009) 3. Michaelis, T., Duelli, M., Chamania, M., Lichtinger, B., Rambach, F., T¨ urk, S.: Network planning, control and management perspectives on dynamic networking. In: 35th European Conference on Optical Communication, Vienna, Austria, p. 7.7.2 (2009) 4. Ciena, The value of otn for network convergence and ip/ethernet migration (2009), http://www.ciena.com/files/ 5. Geem, Z.W., Kim, J.H., Loganathan, G.V.: A new heuristic optimization algorithm: Harmony search. Simulation 76(2), 60–68 (2001), http://sim.sagepub.com/content/76/2/60.abstract 6. Dorigo, M., Birattari, M., Stutzle, T.: Ant colony optimization. IEEE Computational Intelligence Magazine 1(4), 28–39 (2006) 7. T¨ urk, S., Radeke, R., Lehnert, R.: Network migration using ant colony optimization. In: 9th Conference of Telecommunication, Media and Internet TechnoEconomics (CTTE) (June 2010) 8. T¨ urk, S., Sulaiman, S., Haidine, A., Lehnert, R., Michaelis, T.: Approaches for the migration of optical backbone networks towards carrier ethernet. In: IEEE Workshop on Enabling the Future Service-Oriented Internet - Towards SociallyAware Networks, Honolulu, Hawaii, USA (2009) 9. Baliga, J., Ayre, R., Hinton, K., Sorin, W., Tucker, R.: Energy consumption in optical IP networks. Journal of Lightwave Technology 27(13), 2391–2403 (2009) 10. Ferreiro, A.: Nobel 2 project: Migration guidelines with economic assessment and new business opportunities generated by NOBEL phase 2. Tech. Rep. (2008)

Optimization of Energy Eﬃcient Network Migration Using Harmony Search

99

11. Palkopoulou, E., Schupke, D.A., Bauschert, T.: Energy eﬃciency and capex minimization for backbone network planning: is there a tradeoﬀ? In: ANTS 2009: Proceedings of the 3rd International Conference on Advanced Networks and Telecommunication Systems, pp. 34–36. IEEE Press, USA (2009) 12. Idzikowski, F.: Power consumption of network elements in IP over WDM networks. TU Berlin, TKN Group, Tech. Rep. TKN-09-006 (2009) 13. Tamm, O.: Scaling and energy eﬃciency in next generation core networks and switches. In: ECOC, Vienna (2009) 14. Meusburger, C., Schupke, D., Eberspacher, J.: Multiperiod planning for optical networks-approaches based on cost optimization and limited budget. In: IEEE International Conference on Communications, ICC 2008, pp. 5390–5395 (2008) 15. Verbrugge, S.: Strategic planning of optical telecommunication networks in a dynamic and uncertain environment. Ph.D. dissertation, University of Ghent (2007) 16. Lee, K.S., Geem, Z.W.: A new structural optimization method based on the harmony search algorithm. Computers & Structures 82(9-10), 781–798 (2004), http://www.sciencedirect.com/science/article/B6V28-4BWMP8Y-2/2/ cd0c48f22f516cc7df98796923cbe99a 17. Mahdavi, M., Fesanghary, M., Damangir, E.: An improved harmony search algorithm for solving optimization problems. Applied Mathematics and Computation 188(2), 1567–1579 (2007), http://www.sciencedirect.com/science/ article/B6TY8-4MM8B85-1/2/99776af451ff9cf34fec0d93a358aa6d

Self-management of Hybrid Networks: Introduction, Pros and Cons Tiago Fioreze and Aiko Pras University of Twente Enschede, The Netherlands [email protected], [email protected]

Abstract. In the last decade ‘self-management’ has become a popular research theme within the networking community. While reading papers, one could get the impression that self-management is the obvious solution to solve many of the current network management problems. There are hardly any publications, however, that discuss the drawbacks of selfmanagement. In this paper we will therefore introduce self-management for the speciﬁc case of hybrid networks, and discuss some pros and cons. In particular, this paper investigates the feasibility to employ selfmanagement functions within hybrid optical and packet switching networks. In such networks, large IP ﬂows can be moved from the IP level to the optical level, in an attempt to reduce the load at the IP layer and enhance the quality of service (QoS) of the ﬂow that is moved to the optical level. One of the typical management tasks within such networks, is the establishment and release of lightpaths. This paper identiﬁes the advantages and disadvantages of introducing self-management to control such lightpaths. Keywords: Future Internet, self-management, hybrid networks.

1

Introduction

In recent years there has been a considerable interest in what is called the Future Internet. Two fundamentally diﬀerent approaches are under discussion: evolution versus revolution. The evolutionary approach aims at moving the Internet from one state to another through incremental patches. The revolutionary approach, on the other hand, proposes a radical redesign of the current Internet architecture, and is therefore also called clean-slate approach [2]. Whatever approach will prevail, we can already foresee a future Internet in which optical communication will play a major role. At this moment we can already observe that the core Internet, which once solely relied on IP routing to deliver end-to-end communications, is moving towards a hybrid optical-IP network. Such network takes data forwarding decisions simultaneously at both IP and optical level [11]. It is composed of intermediate multi-service devices that are both switches at the optical level and traditional routers at the IP level. In such an environment, data ﬂows can traverse a hybrid network through either R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 100–111, 2011. c Springer-Verlag Berlin Heidelberg 2011

Self-management of Hybrid Networks: Introduction, Pros and Cons

101

an IP path or a lightpath. In this paper, we consider a lightpath as a direct connection over an optical ﬁber; the lightpath can consist of the whole ﬁber, a wavelength within the ﬁber (lambda), or a TDM-based channel within a lambda. Traditionally, the establishment and release of lightpaths is controlled by human managers. In this article we investigate the feasibility of removing the human manager from the loop and to introduce self-management capabilities. Such capabilities enable an autonomic decision process to conﬁgure lightpaths, based on measurement data received from the hybrid network. The human manager expresses what the self-managing system is expected to achieve, but not how this should be done. The human manager is therefore moved to a higher level in the management hierarchy, where he controls the autonomic decision process, rather than the whole hybrid network. Figure 1 shows this self-management approach. Network manager

Autonomic decision process Network traffic information

Configuration process

Hybrid network

Fig. 1. The self-management approach

The goal of this article is to introduce, in a tutorial style, the main advantages and disadvantages of self-management within hybrid networks. In fact, the material presented in this article summarizes four years of PhD research; readers interested in the technical details of our approach and the validation behind the conclusions, are encouraged to read the thesis and associated papers [3] [6] [7] [8]. The remainder of this article is structured as follows. Section 2 introduces the concept of self-management, and provides some deﬁnitions. Following that, Section 3 identiﬁes the related work in this area. Next, Section 4 introduces our approach to employ self-management in hybrid networks. Sections 5 and 6 then respectively present the pros and cons of employing our self-management proposal in hybrid networks. Finally, in Section 7 we draw our conclusions.

102

2

T. Fioreze and A. Pras

What Is Self-management?

Self-management encompasses the act of computer systems managing their own operation without (or with very little) human intervention. It was ﬁrst deﬁned by IBM in 2001 with the IBM Autonomic Computing Initiative (ACI) manifesto [9]. In such manifesto, IBM proposed an approach in which self-managed computing systems could work with a minimum of human interference. This approach is inspired from the human body’s autonomic nervous system. Many actions are performed by our nervous system without any conscious recognition, such as the act of adjusting our eye’s pupils depending on the amount of light or the act of sweating in order to regulate our body temperature. Below, we quote the main objective of IBM’s autonomic initiate that is: “to design and build computing systems capable of running themselves, adjusting to varying circumstances, and preparing their resources to handle most eﬃciently the workloads we put upon them. These autonomic systems must anticipate needs and allow users to concentrate on what they want to accomplish rather than ﬁguring how to rig the computing systems to get them there.” A system can be seen as a collection of computing resources bound together to achieve certain objectives. For example, a network router can constitute a system responsible for forwarding network traﬃc. When combined with other network routers, they can form a larger system, i.e., a Local Area Network (LAN). On its turn, a LAN network combined with other LANs can form a Metropolitan Area Network (MAN), and so on. Based on the IBM autonomic principle, each system must be able to manage its own actions (e.g., traﬃc forwarding), while collaborating with a larger, higher-level system. The same analogy can be found in the human body. From single cells to organs and organ systems (e.g., the circulatory system), each level maintains a measure of independence while contributing to a higher level of organization, culminating in the organism, i.e., the human body. In most parts of our daily life, we remain unaware of our vital organs (e.g., the heart) activities, since these organs (systems) take care of themselves and they only ascend to a higher level (e.g., the brain) when something is wrong and they need some assistance. 2.1

Self-management Aspects

IBM divided self-management into 4 aspects (yet other subdivisions exist), commonly referred as self-*, as follows: – Self-conﬁguration: consists of an automated conﬁguration process of components and systems based on high-levels policies. For example, when a new device is taken into a network, this device is expected to automatically conﬁgure itself and at the same time the rest of the network seamlessly adjust itself to incorporate this new device.

Self-management of Hybrid Networks: Introduction, Pros and Cons

103

– Self-optimization: means that components and systems are supposed to continuously improve their own performance. One example of this aspect is the automatic update process most operating systems provide to their users. Instead of requiring users to manually seek for updates, the operating system does that automatically. – Self-healing: consists of the capability of a system to automatically detect, diagnose, and repair problems found at certain components. As an example, a computer could self-heal every time a virus would strike the system, by automatically patching the damaged ﬁles. – Self-protection: is seen as a system automatically defending itself against malicious attacks or failures. A computer system could, for instance, prevent the infection by a certain email virus through analysis of email attachments. 2.2

Diﬀerent Deﬁnitions for Self-management

Although the term self-management has been widely used, there is no universal consensus among authors on what self-management actually means, which leads to diﬀerent deﬁnitions for the term self-management. Some of the most known deﬁnitions for self-management are as follows: – Autonomic management : is the most common synonym used to refer to the term self-management. That comes from the fact IBM considers selfmanagement as the essence for autonomic computing systems. As a result, the terms self-management and autonomic management are interchangeably used to mean the same. By analyzing the keywords attached to papers submitted to important network management conferences (e.g., IM, NOMS, CNSM), we found out that 80% of the papers were submitted with the keywords as self-*, whereas 20% were registered as autonomic. This leads us to a conclusion that even if they were constantly used as synonyms, the term self-management seems to be the most referred and used by the network management community. – Automatic management : is commonly confused as autonomic management (and thus with self-management). Automatic management refers as the act of managed devices automatically following explicit policies deﬁned by a network operator. In its turn, autonomic management refers as a specialized automatic process in the sense that the process is instructed to perform actions based on certain policies too, but with the capability of self-learning new actions. – Autonomous management : is another deﬁnition referring to self-management. Autonomous means that a process can operate independently from any human intervention. However, this lack of external control is, according to some, a contradiction. If an autonomous “management” system includes enough intelligence in order for the system to govern its own management, one can assume that there is no need whatsoever of managing such a system, which somehow invalidates the use of the term management to address this kind of management approach.

104

T. Fioreze and A. Pras

Degree of autonomy

It is worth saying that the foregoing diﬀerentiation among self-management deﬁnitions is not a common view in the community. On the contrary, this diﬀerentiation solely destined for being a reference to be used throughout this article. We see these deﬁnitions as following an evolution in the network management approaches as well as having diﬀerent degrees of autonomy (Figure 2).

Autonomous

Autonomic/ self-management

Automatic

Conventional

Evolution in the network management approaches

Fig. 2. Evolution in the network management approaches vs. their degree of autonomy

The simplest management approach is the conventional management approach. In the conventional management approach, the network management system is manually managed by network operators. There is no intelligence whatsoever and no (or very little) automation in the execution of management tasks. A next step in the evolution of management approaches is the automation of management tasks. In this case, the management system automatically performs explicit tasks deﬁned by network managers, but nothing beyond the scope of the deﬁned rules. Following to automatic management, autonomic management (or self-management) also performs these tasks, but it is capable of learning new rules by itself. The last step in the evolution process and the most complex one is the autonomous management. At this level, the management system is fully capable of deciding by itself the rules to follow. There is therefore no dependence on human intervention. The management system is intelligent enough to decide its own rules and following them according to its judgement.

3

Related Work on Self-management

Since the releasing of the ACI manifesto, several research works about the use of self-management have been reported. To name a few of these works, Lupu et al. [12] have been researching the use of self-management on healthcare practicing, in which a ubiquitous self-managed computing environment is used to monitor and report the health of patients under medical treatment. In another work, selfmanagement is investigated to be used in situations where there is a great risk

Self-management of Hybrid Networks: Introduction, Pros and Cons

105

for human beings, such as in military or disaster scenarios. Within this line of research, we point out the work by Asmare et al. [1] who has been investigating the use of self-management on Unmanned Autonomous Vehicles (UAVs). Not much diﬀerently, self-management has also being investigated in the area of communication networks [10]. Much of the focus of this investigation aims at developing highly distributed algorithms, with the objective to optimize several aspects of network operability (e.g., performance). This optimization is aimed through the provision of self-management capabilities to communication networks. Studies on self-management is also the focus of several research projects, such as Autonomic Internet (AUTOI), Self-Optimisation and self-ConﬁguRATion in wirelEss networkS (SOCRATES), and UniverSelf. A study that is closely related to ours is by Miyazawa et al. [13]. In their research, they propose a dynamic bandwidth control management mechanism based on the volume of IP ﬂows. In their work there is a centralized management system that observes the bandwidth of IP ﬂows, and decides about oﬄoading these ﬂows based on pre-deﬁned upper and lower threshold values. These threshold values are deﬁned in advance by a human operator and statically stored in the conﬁguration ﬁle of the management system. Once an IP ﬂow has a bandwidth utilization that exceeds the pre-determined upper threshold, the management system triggers an action to create a lightpath. In contrast, when the ﬂow decreases its bandwidth utilization below the lower threshold, the management system initiates a deletion process for deleting the established lightpath.

4

Self-management of Lightpaths

We focus now the use of self-management in the context of hybrid optical and packet switching networks, more speciﬁcally on the self-management of lightpaths in these networks. The use of self-management is aimed here at autonomically: 1) detect ﬂows at the IP level eligible to be moved to the optical level as well as 2) establish/release lightpaths for those ﬂows. In this paper we adopt the deﬁnition of ﬂows as described in the information model for the IP Flow Information eXport (IPFIX) protocol (RFC 5102). In this RFC, an IP ﬂow is deﬁned as a unidirectional sequence of packets that share the same properties (e.g., the same source and destination IP addresses, source and destination port numbers and higher level protocol). Network operators are only required to initially conﬁgure the self-management process with decision policies. After this initial setup, the self-management process autonomically runs by itself. Decision policies deﬁne a desired objective, which must be achieved by the self-management functions. In our research, the main objective is to oﬄoad as much traﬃc as possible from the IP level to the optical level. For that, our self-management approach aims at moving ﬂows to the optical level that are few in amount, but represent most of the traﬃc, namely the elephant ﬂows. Figure 3 depicts our approach for the self-management of lightpaths in hybrid networks.

106

T. Fioreze and A. Pras Selfmanagement

2

3 Monitoring station

1

IP domain B

Optical domain A

Caption Optical level Network level

IP router Optical switch

Elephant flow Lambda connections

Traffic information Configuration message

Fig. 3. Self-management of lightpaths in hybrid optical and packet networks

In Figure 3, IP routers located in the IP domain B are exporting network trafﬁc information to a monitoring station (step 1). Network information contains ﬂow information, such as source & destination IP addresses, protocol, ﬂow volume, amongst others. This information is then forwarded to our self-management module (step 2). Based on the information received, decisions are made by the module taking into account whether an elephant ﬂow is eligible or no longer eligible for a lightpath at the optical level. If the decision is in favor of creating a lightpath (i.e., the elephant ﬂow is eligible to be moved to the optical level), the self-management module conﬁgures the IP routers in the IP domain B and the optical switches in the optical domain A (step 3). The routers are informed that the elephant ﬂow is oﬄoaded to the optical level. On their turn, the optical switches are conﬁgured to establish a lightpath for the oﬄoaded elephant ﬂow. From that point on, the elephant ﬂow is switched at the optical level bypassing thus the network level in the IP domain B. For conﬁguring routers and switches, existing management technologies can be used, such as the Command Line Interface (CLI), the Generalized Multi-Protocol Label Switching (GMPLS) protocol, the Simple Network Management Protocol (SNMP) or the emerging Network Conﬁguration (NetConf) protocol. Note that this article can only summarize the operation of our self-management approach; details of this operation can be found in other papers [4] [5], as well as the thesis that resulted from this research [3].

5

Advantages of Self-management

Advantages and disadvantages depend on the context self-management is employed. We highlight in this section the main pros of employing self-management principles in the speciﬁc context of hybrid optical and packet switching networks.

Self-management of Hybrid Networks: Introduction, Pros and Cons

5.1

107

Network Performance

The use of self-management improves network performance by automatically reducing the burden of the IP level. When IP ﬂows are completely transported via lightpaths they bypass the per hop routing decisions of the IP level. As a result, the QoS oﬀered by hybrid networks is considerably better when compared to traditional IP networks. Big IP ﬂows that overload the regular IP level, for example, may be moved to the optical level where they experience better QoS (e.g., negligible jitter and larger bandwidth). At the same time, the IP level is oﬄoaded and can better serve smaller ﬂows. Last but not least, it is also cheaper to send traﬃc at the optical level than at the IP level. For the same traﬃc rate, the cost of an optical switch is 1/10th of an Ethernet switch or 1/100th of a conventional router. 5.2

Network Management

We believe that human factors have an impact on the management of lightpaths. For example, network operators of SURFnet reported, when informally interviewed, that it may take hours (intra-domain) or even days (inter-domain) before a lightpath is established by network operators when using a traditional network management paradigm. In such paradigm, a network manager regularly monitors a hybrid network. Based on his analysis of the collected data, he may decide to establish or release a lightpath. It is worth highlighting that this paradigm keeps the human in the management loop. That is, most of the management decisions have to go through the network manager. As a result, the management system does not go beyond any predetermined state or perform any unexpected action, unless explicitly triggered by the network manager. We argue that during the long periods lightpaths are established, several big IP ﬂows could have been transported via lightpaths, but due to the decision delay they remain being routed at the IP level. Moreover, in such long periods, many large IP ﬂows may be using resources at the IP level and, therefore, likely congesting the IP level. Moreover, by the time the lightpath is ﬁnally established, those large ﬂows may no longer exist. The human intervention required to select IP ﬂows and manage lightpaths may be considered therefore slow and ineﬃcient. We see our self-management proposal as an alternative to overcome this dependency on human intervention and therefore improve the network management. 5.3

Selection of Unknown Large Flows

Nowadays, IP traﬃc from several specialized applications, some of them requiring considerable amounts of bandwidth, already proﬁt from lambda-switched networks capabilities. Examples are: Grid applications, High-Deﬁnition Television (HDTV) broadcasting, and large-scale scientiﬁc experiments. The knowledge of the heavy-hitter behavior of ﬂows originated from these applications allows network managers to establish lambda-connections in advance for such ﬂows. However, there may be also other big IP ﬂows in current networks that

108

T. Fioreze and A. Pras

could also beneﬁt from being moved to lambda-connections, but since the network manager is not aware of their existence, they may not be selected. Selfmanagement comes handy here since ﬂows are monitored and selected by the self-managing system rather than by the human manager. 5.4

Dynamic Flow Selection

The selection of ﬂows to be moved to the optical level are traditionally made based on pre-deﬁned upper and lower threshold values. These threshold values are deﬁned in advance by a human operator and statically stored in the conﬁguration ﬁle of the management system (Section 3). The main shortcoming of using thresholds is that they are statically deﬁned and they are not adjusted depending on the current traﬃc. This can lead up to an unbalance between the IP and optical levels. If the upper threshold values are too restrictive, IP ﬂows may not be oﬄoaded over lightpaths, which may result in congestion in the IP level and underutilization of the optical level. Moreover, with a misadjusted lower threshold, a ﬂow can be inadequately removed from the optical level back to the IP level, where it can contribute to a congestion situation. Our alternative, on the other hand, aims at prioritizing ﬂows by merit (behavior) rather than by characteristics (i.e., port numbers, ToS, and so on). Their merit is measured based on the amount of traﬃc they are expected to generate. Flows that are expected to generate more traﬃc are chosen over ﬂows that are expected to generate less traﬃc.

6

Disadvantages of Self-management

Self-management of hybrid networks also introduces a number of problems, as will be explained in this section. 6.1

Complexity in the Network Management System

As shown in Figure 2, management approaches follow an evolution that is proportional to their degree of autonomy. In the simplest management approach, network managers are responsible for all management tasks. However, as more experience is obtained with these management tasks, some tasks can gradually be automated, which means that the need for human intervention can be reduced. In order to avoid problems related to centralization (single point of failure, possible performance bottleneck), subsequent cycles in the design of the management system may focus on distributing such management tasks. Therefore there is a shift from centralized and explicit management towards distributed and implicit management approaches, such as the aforementioned automatic and autonomic approaches [14]. However, the price to be paid for such evolution is the increased complexity of the management system. The chance that errors get introduced in the implementation of the management system therefore increases, and debugging possible failures becomes harder.

Self-management of Hybrid Networks: Introduction, Pros and Cons

6.2

109

Network Security

Network security consists of providing means to protect network resources from unauthorized access or malicious activities. Consistent and continuos monitoring of network activity is important to prevent or detect any misusage that may fall upon a managed network. Within the context of our self-management approach, network security is expected to prevent any inappropriate use of lightpaths. For instance, Denial of Service (DoS) attacks that may be transiting at the IP level should not be moved to lightpaths. Such move could make the attack more severe, since the increased bandwidth at the optical level allows the transfer of more packets to the attacked system. Other security concepts, such as authorization and intrusion detection should therefore be considered as well, in order to tighten security in hybrid networks. Flows could be veriﬁed prior to the oﬄoad to the optical level to detect malicious behavior. If ﬂows behave suspiciously, their oﬄoad over lightpaths could be blocked and their behavior could be logged for audit purposes and later analysis. 6.3

Temporarily Reduction In Throughput Performance

Another interesting question is whether there is any performance degradation when an active ﬂow is moved from the IP level towards a lightpath. We may expect that, at the moment ﬂows are moved, massive re-ordering takes place, since the ﬁrst packets transferred over the lightpath can arrive earlier than the last packets over the IP path. To analyze this eﬀect, we used ns-2 to simulate the behavior of TCP ﬂows during such movement, and identiﬁed which factors limited the throughput of such ﬂows [15]. For this analysis we used the network topology as shown in (Figure 4).

Fig. 4. Topology used in the simulations and limiting factors (Greek letters).

We observed diﬀerent kinds of impact on the throughput of TCP ﬂows. In all scenarios some throughput oscillation occurred during the transient phase, but TCP throughput recovered relatively fast after the transient phase was over. However, when the network link at the receiver side is the factor that limits the TCP throughput (thus the bandwidth of ξ is smaller than that of any other link), we found a huge impact on the performance of TCP. In this case, during the transient phase, router r3 tries to send the last data received over the IP path, together with the ﬁrst data received over the optical path, over the outgoing

110

T. Fioreze and A. Pras

link ξ. The outgoing router’s queue will be ﬁlled rapidly and packets must be dropped due to lack of queue space. It is interesting to note that the decrease in throughput was not caused by packet reordering, but by packet loss. This problem indicates that the transmission capacity of the link at the receiver side and the router’s buﬀer size should be considered before moving ﬂows on the ﬂy.

7

Conclusions

Based on the research presented in this article, our main conclusion is that selfmanagement is technically feasible to be deployed within hybrid optical and packet switching networks. From an implementation point of view, the decision process in a self-management approach can be built upon existing technologies, such as NetFlow/IPFIX to collect traﬃc information, and CLI, GMPLS, SNMP or NetConf to conﬁgure optical switches and routers. Compared to traditional management approaches, self-management of hybrid networks provide several advantages. These advantages include better network performance, faster lightpath establishment and release, the ability to move large ﬂows to the optical level, even in cases where such ﬂows have not been made known to the human manager in advance, and ﬁnally the possibility to avoid congestion or underutilization by dynamically changing the decision thresholds. Self-management of hybrid networks also has a number of disadvantages, however. An obvious disadvantage is that self-management increases the complexity of the management system, and therefore makes it harder to debug possible failures. Another problem is that large ﬂows generated by a DoS attack can be moved to the optical level, and in this manner strengthen the attack. Also the move of an IP ﬂow to the optical level may result in the temporary re-ordering and loss of packets, especially in cases where the bandwidth of the network links at the receiver side are the limiting factor that determine the TCP throughput. Acknowledgments: This work has been supported by the EC IST-Emanics NoE (26854), as well as the EC IST-UniverSelf (FP7-257513) and the GigaPort Next Generation projects.

References 1. Asmare, E., Gopalan, A., Sloman, M., Dulay, N., Lupu, E.: A Mission Management Framework for Unmanned Autonomous Vehicles. In: Mobile Wireless Middleware, Operating Systems, and Applications, Second International Conference (Mobilware). ICST Lecture Notes, vol. 7, pp. 222–235 (April 2009) 2. Feldmann, A.: The Internet Architecture - Is a Redesign Needed? pp. 147–164. Springer, Heidelberg (2009) 3. Fioreze, T.: Self-Management of Hybrid Optical and Packet Switching Networks. Ph.D. thesis. Universiteit Twente, Enschede (February 2010) 4. Fioreze, T., Granville, L., Pras, A., Sperotto, A., Sadre, R.: Self-management of hybrid networks: Can we trust netﬂow data? In: 11th IFIP/IEEE International Symposium on Integrated Network Management (IM 2009), Long Island, New York, USA, pp. 577–584 (June 2009)

Self-management of Hybrid Networks: Introduction, Pros and Cons

111

5. Fioreze, T., Granville, L., Sadre, R., Pras, A.: A statistical analysis of network parameters for the self-management of lambda-connections. In: Sadre, R., Pras, A. (eds.) AIMS 2009 Enschede. LNCS, vol. 5637, pp. 15–27. Springer, Heidelberg (2009) 6. Fioreze, T., van de Meent, R., Pras, A.: An Architecture for the Self-management of Lambda-Connections in Hybrid Networks. In: Pras, A., van Sinderen, M. (eds.) EUNICE 2007. LNCS, vol. 4606, pp. 141–148. Springer, Heidelberg (2007) 7. Fioreze, T., Pras, A.: Using Self-management for Establishing Light Paths in Optical Networks: an Overview. In: 12th EUNICE Open European Summer School, Stuttgart, Germany, pp. 17–20. Institut f¨ ur Kommunikationsnetze und Rechnersysteme, Universit¨ at Stuttgart, Stuttgart, Germany (2006) 8. Fioreze, T., Pras, A.: Self-management of Lambda-connections in Optical Networks. In: Bandara, A.K., Burgess, M. (eds.) AIMS 2007. LNCS, vol. 4543, pp. 212–215. Springer, Heidelberg (2007) 9. Horn, P.: Autonomic computing: IBM’s Perspective on the State of Information Technology (2001), http://www.research.ibm.com/autonomic/manifesto/ autonomic_computing.pdf 10. Jennings, B., van der Meer, S., Balasubramaniam, S., Botvich, D., Foghlu, M., Donnelly, W., Strassner, J.: Towards autonomic management of communications networks. IEEE Communications Magazine 45(10), 112–121 (2007) 11. Leon-Garcia, A., Widjaja, I.: Communication Networks: Fundamental Concepts and Key Architectures, 2nd edn. McGraw-Hill Companies, New York (2003) 12. Lupu, E., Dulay, N., Sloman, M., Sventek, J., Heeps, S., Strowes, S., Twidle, K., Keoh, S.L., Schaeﬀer-Filho, A.: AMUSE: autonomic management of ubiquitous e-Health systems. Concurrency and Computation: Practice and Experience 20(3), 277–295 (2008) 13. Miyazawa, M., Ogaki, K., Otani, T.: Multi-layer network management system with dynamic control of MPLS/GMPLS LSPs based on IP ﬂows. In: The 11th IEEE/IFIP Network Operations and Management Symposium (NOMS 2008), Salvador, Brazil, pp. 263–270 (April 2008) 14. Pras, A.: Network Management Architectures. Ph.D. thesis, Universiteit Twente, Enschede (February 1995) 15. Timmer, M., de Boer, P.T., Pras, A.: How to Identify the Speed Limiting Factor of a TCP Flow. In: 4th IEEE/IFIP Workshop on End-to-End Monitoring Techniques and Services (E2EMON 2006), Vancouver, Canada, pp. 17–24 (April 2006)

Evaluation of Diﬀerent Decrease Schemes for LEDBAT Congestion Control Mirja K¨ uhlewind1 and Stefan Fisches2 1

Institute of Communication Networks and Computer Engineering (IKR) University of Stuttgart, Germany [email protected] 2 [email protected]

Abstract. Low Extra Delay Background Transport (LEDBAT) is a new, delay-based congestion control algorithm that is currently under development in the IETF. LEDBAT has been proposed by BitTorrent for time-insensitive background traﬃc that otherwise would disturb foreground traﬃc like VoIP or video streaming. During previous evaluations the so called late-comer advantage has been discovered which makes a new starting LEDBAT ﬂow predominant against already running LEDBAT ﬂows. In this paper we evaluate diﬀerent decrease schemes which have been proposed to solve this problem. We found that the proposed solutions come with a lower utilization, sometimes increased completion times and are much more sensitive to noise, which is contra-productive for the considered traﬃc class. Furthermore, we propose extensions to both evaluated schemes. We show that our approach can help to yield more quickly to higher priority traﬃc. We argue that a fair and equal share is not required for the speciﬁc traﬃc class LEDBAT is designed for. But it is important to address diﬀerent application requirements in congestion control like LEDBAT as an approach for less-than-best eﬀort background traﬃc.

1

Introduction

A substantial portion of bandwidth in today’s Internet is used for background and time-insensitive traﬃc (e.g. P2P Traﬃc [1]). This traﬃc should not impede foreground and time-sensitive traﬃc. A novel congestion control algorithm designed for less-than-best-eﬀort traﬃc is the Low Extra Delay Background Transport (LEDBAT) [2]. It was proposed by BitTorrent in December 2008 and is now under development within an Internet Engineering Task Force (IETF) working group. It is a delay-based approach that can react earlier to congestion then loss-based schemes which are mostly used today in today’s Internet for TCP traﬃc. In this paper, we regard two sorts of traﬃc: 1. Less-than-best-eﬀort, low priority traﬃc using LEDBAT congestion control, e.g. automatic software Updates running in the background or Peer-to-Peer ﬁle sharing R. Lehnert (Ed.): EUNICE 2011, LNCS 6955, pp. 112–123, 2011. c Springer-Verlag Berlin Heidelberg 2011

Evaluation of Diﬀerent Decrease Schemes for LEDBAT Congestion Control

113

2. Higher priority best-eﬀort traﬃc using lost-based congestion control (or even sending with a constant bit rate), e.g. web-browsing or Voice over IP using UDP According to [2], the LEDBAT congestion control seeks to: 1. utilize end-to-end available bandwidth, and maintain low queueing delay when no other traﬃc is present, 2. add little to the queuing delay induced by concurrent TCP ﬂows, 3. quickly yield to ﬂows using standard TCP congestion control that share the same bottleneck link. With the current speciﬁcation of LEDBAT there is a so-called “late-comer’s advantage”, where a second, newly starting ﬂow can starve the ﬁrst, already running one. This only happens when two or more LEDBAT ﬂow compete on the same link and no other higher priority traﬃc is present. Thus it is an issue of intra-protocol fairness. Several mechanisms have been proposed to prevent this eﬀect e.g. a mandatory slow-start or multiplicative decrease. We argue however that a high link utilization is actually more desirable for a lower priority traﬃc class than fairness within that class. As LEDBAT is designed for background traﬃc that will yield for higher priority traﬃc, LEDBAT should be able to utilize the link as much as possible when no other traﬃc is present. Completing one ﬂow after another instead of transmitting all ﬂows in parallel will, whilst being unfair, minimize the mean completion time. When sending as much data as possible en-block, computational power and hence energy consumption will be minimized as well. To support our hypothesis we evaluate diﬀerent decrease schemes with regard to completion time and utilization. We evaluated the proposed linear decrease scheme, which is discussed in the ITEF, and a contra-proposal by Caroﬁglio et al. [3] for using a multiplicative decrease. Moreover, we introduce new extensions to either of the schemes. We implemented LEDBAT as TCP congestion control in Linux using the TCP Timestamp Option for one-way delay measurements and subsequently used this code within a simulation environment. Moreover, we show that the decrease behavior is not only important to mitigate the eﬀects of the late-comer’s advantage but is also important when LEDBAT needs to yield to higher priority traﬃc like standard TCP. The remainder of this paper is structured as follows: Section 2 summarizes related work. Section 3 is a general introduction into the LEDBAT algorithm and diﬀerent decrease schemes. In Section 4 we present our TCP implementation in Linux. Section 5 shows our results regarding fairness and utilization.Section 6 gives concluding remarks.

2

Related Work

A comparison of LEDBAT with standard TCP as well as other less-than-besteﬀort congestion control mechanism such as LP-TCP and TCP-NICE has already been performed by Rossi et al. [4] [5]. They also detected the late comer’s

114

M. K¨ uhlewind and S. Fisches

advantage [6]. Initially, the authors proposed TCP Slow-Start as a solution to this problem. Slow-Start will usually overshoot and induce losses that cause all competing ﬂows to basically restart their transmission. When two LEDBAT ﬂows start at the same time, they will equally share the available capacity. But SlowStart overshoot will actually aﬀect all competing ﬂows on the link, LEDBAT ﬂows as well as higher priority standard TCP ﬂows. This breaks one of the design goals of LEDBAT, as listed in 1. In fact, the current LEDBAT speciﬁcation leaves it to the implementor to use a speciﬁc start-up scheme, if necessary. [3] proposed a multiplicative decrease scheme to achieve fairness. We argue that equal sharing is a non-requirement for background traﬃc within its traﬃc class. [7] argues as well that it is not the right metric for fairness to share the available capacity equally between competing ﬂows with diﬀerent requirements. Another study about parametrization is provided by [8]. In this paper we did not look at any parametrization issues. This issues are widely discussed on the IETF LEDBAT mailing list and mostly addressed in the working group document.

3

LEDBAT

LEDBAT is a novel delay-based congestion control approach for low priority transmissions. It is under development within an IETF working group. It is based on one-way delay measurements. When the measured one-way delay increases, LEDBAT can react earlier to congestion than loss-based approaches, which are more predominant in today’s Internet. By slowing down its transmission rate earlier it will give the available bandwidth to presumed higher priority transmissions. Thus LEDBAT is friendly to most of today’s higher priority TCP traﬃc. Using timestamps, LEDBAT measures the queuing delay on a link. Assuming all queues on the path are empty at some point of the transmission, the sender will take the smallest delay measurement as the base delay. base delay = min(base delay, current delay) The base delay is the constant fraction of the time a packet needs from sender to receiver and thus the minimum transmission time that can be observed. Any additional, variable delays are presumed to be waiting times in network queues. Therefore the actual queuing delay is calculated based on the current delay measurement given by the receiver as following: queuing delay = current delay − base delay When the base delay changes during a transmission, e.g. because of re-routing, LEDBAT will automatically update to the new base delay if it gets smaller. To recognize a higher base delay a base delay history is kept which holds the measured base delay of the last n minutes and will discard old values after n + 1 minutes, thus adapting to the new base delay. LEDBAT aims to keep the queuing delay low but not to be zero since optimal resource utilization requires that there is always data to send. Therefore LEDBAT

Evaluation of Diﬀerent Decrease Schemes for LEDBAT Congestion Control

115

tries to achieve an extra delay of TARGET milliseconds. LEDBAT can determine how far oﬀ target it is and then increases or decreases its sending rate. LEDBAT is designed to not ramp up faster than TCP. Thus it will at maxium increase its sending rate by one packet per round-trip time (RTT). LEDBAT can use a ﬁlter to smooth or single out wrong delay measurements. But depending on the used ﬁlter scheme and length, e.g. minimum or average of the last CURRENT FILTER samples, the reaction to congestion might get delayed. 3.1

Late-Comer’s Advantage

Whenever a LEDBAT ﬂow uses a previously unused path, it will immediately measure the real base delay. Thus it will saturate the bottleneck link after a short time by maintaining an extra delay of TARGET milliseconds. If now a second ﬂow arrives, it can only measure the actual base delay plus the extra delay of the ﬁrst ﬂow. Wrongly, the second ﬂow will take this value as its base delay. While the latecomer will add its own target delay on top, the ﬁrst ﬂow measures an increased delay and begins to lower its sending rate. In the worst case, the second ﬂow will add an additional TARGET millisecond of extra delay. As long as the second ﬂow is not able to measure the actual base delay, it will push away the ﬁrst ﬂow completely. This eﬀect is called the “late-comer’s advantage“. A proposed way to mitigate this eﬀect is to change the decrease behavior to multiplicative decrease to empty the queue completely such that the second ﬂow can measure the real base delay [3]. 3.2

Managing the Congestion Window

Linear Controller. The current draft version of LEDBAT uses the same linear controller for an additive increase and additive decrease. of f target =

T ARGET − queuing delay T ARGET

of f target cwnd The congestion window (CWND), which gives the number of packets that can be transmitted in one RTT, is altered by the normalized oﬀ target parameter that can be positive or negative and thus determines how the CWND grows or shrinks. When GAIN is one, LEDBAT will at maximum speed up as quickly as standard TCP because oﬀ target will always be smaller than 1 and reach its maximum value when queuing delay is zero. Unfortunately, this approach allows LEDBAT to decrease very slowly if e.g. just one millisecond of extra delay above the TARGET is measured. To be friendly to standard TCP traﬃc, LEDBAT should at least decrease as quickly as standard TCP is increasing. The latest version of the LEDBAT draft in the IEFT allows a diﬀerent GAIN value for the decrease than for the increase. We propose to use GAIN = T ARGET ∗ N cwnd+ = GAIN ∗

M. K¨ uhlewind and S. Fisches

cwnd [packets]

116

80 70 60 50 40 30 20 10 0

Std. LEDBAT 1 Std. LEDBAT 2

0

20

40

60

100 80 Time [s]

120

140

160

180

Fig. 1. Two competing LEDBAT ﬂows with linear decrease

if oﬀ target is negative (decrease). N would need to be the number of parallel starting standard TCP ﬂows. As this number is usually unknown, we assume in most cases only one ﬂow starting at the same time and set N to 1. Thus LEDBAT would decrease the congestion window at least by one packet per RTT. Multiplicative Decrease. [3] proposes to decrease the congestion window for negative oﬀ target by multiplying it with a factor β such that 0 < β < 1: CW N D = β × CW N D The idea is that the multiplicative decrease allows the queues to drain and thus enable a correct measurement of the base delay. If β is chosen to low it is underutilizing the link. If β is to large it doesn’t drain the queue. While [3] is searching for a ﬁxed value, β is actually depending on the chosen value of TARGET in relation to the current RTT. This is because the CWND is depending on the RTT as it gives the number of packets that can be sent during one RTT. Whereas the TARGET gives the extra delay and thus determines the number of packets that will be stored in the network queue that needs to be emptied at once (in T ARGET one RTT). Thus β must be current RT T . For a negative oﬀ target value the congestion window is then calculated as CW N D =

T ARGET × CW N D current RT T

current RT T ≈ 2 × base delay + T ARGET In our implementation we subtracted additional 3 packets after this calculation to encounter measurement and computation inaccuracies. This value of 3 packets is selected through simulative studies. With every multiplicative decrease scheme we decrease only once per RTT as during the ﬁrst RTT after the decrease all delay measurements still reﬂect the situation before the decrease.

4

Linux Implementation

To provide a wide access to a less-than-best-eﬀort congestion control scheme, we decided to implement LEDBAT as a TCP congestion control module. The Linux

Evaluation of Diﬀerent Decrease Schemes for LEDBAT Congestion Control

117

kernel design provides an interface to extent the kernel functionality through additional modules. There is a speciﬁc module interface for congestion control. Thus our LEDBAT implementation follows the respective interface of the Linux kernel. The resulting c-ﬁle can be included in any current Linux kernel version. No further modiﬁcations were needed as TCP already provides an option to transfer and echo time-stamps. Based on this given functionality we do not calculate the delay at the receiver side, as speciﬁed in the LEDBAT draft, but at the sender. With the TCP Timestamp Option the receiver reﬂects the timestamp T Ssnd sent by the sender and adds an additional time-stamp T Srcv at sent-out of the ACK. By subtracting the echoed time-stamp (TSsnd) from the new time-stamp (TSrcv) we determine the one-way delay. OW D = T Srcv − T Ssnd This delay includes the processing delay at the receiver, but as we assume this processing delay to be constant it will not disturb our queuing delay estimation. The clocks of sender and receiver do usually not operate on the same time base. As we only use queuing delay, that means the variable part of the oneway delay in relation to a certain base delay, the absolute value of base delay in not important for us. Even if both timestamp have a diﬀerent resolution, this can be estimated by monitoring the ﬁrst samples in relation to the own clock. As we know the time-stamp resolution in our evaluation scenarios, we did not implement a speciﬁc logic. In order to archive an precise enough delay measurements to monitor the changes in on-way delay, we had to adjust the kernels timer frequency from its default 250 HZ to 1000 HZ, giving us a 1 ms resolution in the timestamps. Furthermore, TCP will acknowledge at least every second data packet and waits at most by default Linux conﬁguration 100 ms for an additional packet. This is the delayed ACK mechanism of TCP. But whenever two packets are acknowledged at once, only the timestamp of the ﬁrst packet will be echoed. Thus we have just half the number of measurement samples and some of them will have artiﬁcal delays. If packets arrive continuousely, there is a constant waiting time until the second packet arrives. This oﬀset is not a problem. But if the timer expires, there are high variations. With the linear decrease scheme the impact of one wrong sample is quite low but with multiplicative decrease we necessarily need noise ﬁltering to cope with this eﬀect. For our simulations the open source character of Linux also allowed us to patch the kernel and disable delayed acknowledgments at the receiver side.

5 5.1

Evaluation Results Scenario

To evaluate the presented decrease schemes in a real-world scenario we ﬁrst set up a small testbed with the just described LEDBAT kernel module. All presented results are extracted from simulations with the IKR SimLib [9] as

118

M. K¨ uhlewind and S. Fisches

cwnd [packets]

60

MD-NF2 1 MD-NF2 2

50 40 30 20 10 0 0

20

40

60

100 80 Time [s]

120

cwnd [packets]

60

140

160

180

160

180

160

180

MD-NF1 1 MD-NF1 2

50 40 30 20 10 0 0

20

40

60

100 80 Time [s]

120

cwnd [packets]

60

140

MDvar 1 MDvar 2

50 40 30 20 10 0 0

20

40

60

100 80 Time [s]

120

140

Fig. 2. Two competing LEDBAT ﬂows with diﬀerent multiplicative decrease, base RTT 40 ms

in the simulation delayed ACK could easily be deactivated. We used the IKR SimLib together with the TCP implementation of the real Linux kernel code provided by the Network Simulation Cradle [10] - a framework that makes kernel code usable within a simulation environment. In our simulation scenarios we used 2-5 parallel ﬂows with a bottleneck link capacity of 10 Mbit/s and one-way delays of 10, 19, 20 or 30 ms. In every scenario each ﬂow starts with an oﬀset of 15 s to the previous one. All ﬂows in a scenario want to transit the same data size which is either 30, 50, 100, 300, 500 or 1000 MBytes. The bottleneck node maintains a queue with a queue size that can hold 60 packets. For all simulations we used a TARGET value of 25 ms and a length CURRENT FILTER of 1 or 2, where 1 basically means no ﬁltering at all. 5.2

Two Competing Flows

To illustrate the behavior of the diﬀerent decrease schemes we present in detail the scenarios where two LEDBAT ﬂows compete with the same decrease mechanism. Figure 1 shows the linear decrease behavior. The ﬁrst ﬂow starts and increases its rate slowly until the bottleneck link is ﬁlled and 25 ms extra delay is introduced. The increase gets slower as it gets closer to the 25 ms target value. After 15 ms a second ﬂow starts. This ﬂow assumes the current delay as base

Evaluation of Diﬀerent Decrease Schemes for LEDBAT Congestion Control

cwnd [packets]

60

119

MD-NF2 1 MD-NF2 2

50 40 30 20 10 0 0

20

40

60

100 80 Time [s]

120

cwnd [packets]

60

140

160

180

160

180

160

180

MD-NF1 1 MD-NF1 2

50 40 30 20 10 0 0

20

40

60

100 80 Time [s]

120

cwnd [packets]

60

140

MDvar 1 MDvar 2

50 40 30 20 10 0 0

20

40

60

100 80 Time [s]

120

140

Fig. 3. Two competing LEDBAT ﬂows with diﬀerent multiplicative decrease, base RTT 20 ms

delay and starts increasing its rate as well. The ﬁrst ﬂows starts decreasing as it senses more extra delay than the already introduced 25 ms. It decreases only slowly as only little extra delay is introduced by the second ﬂow so far. Thus the second ﬂow, which never measures the correct base delay, adds an additional delay on top of the extra delay of the ﬁrst ﬂow. However, when two ﬂows start at the same time or restart after the currently dominating ﬂow ﬁnished, they will share the capacity equally as long as they have the same TARGET value. The upper and middle diagram in Figure 2 show the multiplicative decrease behavior as proposed by [3] and explained in 3.2 with a ﬁxed value for β of 0.6. We investigated two diﬀerent variants. The ﬁrst variant uses a simple noise ﬁlter of length 2. It takes always the minimum of the last two measurement samples of current delay. In the second variant the length is one, so there is no noise ﬁltering. But we deactivated delayed ACKs in our simulation. We labeled those variants with MD-NF2 and MD-NF1. All multiplicative decrease schemes aim to empty the queue when another LEDBAT is starting. Thus both ﬂows can measure the right base delay. Whenever some extra delay is introduced on top of TARGET these schemes will decrease the windows and the link becomes underutilized. This happens periodically even when no other ﬂow is starting as LEDBAT itself will exceed the TARGET for probing. When the second ﬂow starts it will quickly measure the correct base delay as the ﬁrst one decreases.

120

M. K¨ uhlewind and S. Fisches

cwnd [packets]

60

MD-NF2 1 MD-NF2 2 MD-NF2 3 MD-NF2 4 MD-NF2 5

50 40 30 20 10 0 0

20

40

60

80

100

120

140

Time [s]

cwnd [packets]

60

MD-NF1 1 MD-NF1 2 MD-NF1 3 MD-NF1 4 MD-NF1 5

50 40 30 20 10 0 0

20

40

60

80

100

120

140

Time [s]

cwnd [packets]

60

MDvar 1 MDvar 2 MDvar 3 MDvar 4 MDvar 5

50 40 30 20 10 0 0

20

40

60

80

100

120

140

Time [s]

Fig. 4. Five competing LEDBAT ﬂows with diﬀerent multiplicative decrease, base RTT 40 ms

Then both will increase in parallel and the decrease times get synchronized so that both get the same share of the bandwidth. MD-NF2 does not synchronize due to the noise ﬁltering and thus does not perfectly divide the bandwidth but approximately gives each ﬂow an equal share. The lower diagram in ﬁgure 2 shows our proposal