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!
Communications in Computer and Information Science
261
Tai-hoon Kim Hojjat Adeli Hyun-seob Cho Osvaldo Gervasi Stephen S. Yau Byeong-Ho Kang Javier García Villalba (Eds.)
Grid and Distributed Computing International Conference, GDC 2011 Held as Part of the Future Generation Information Technology Conference, FGIT 2011 Jeju Island, Korea, December 8-10, 2011 Proceedings
13
Volume Editors Tai-hoon Kim Hannam University, Daejeon, Korea E-mail: [email protected] Hojjat Adeli The Ohio State University, Columbus, OH, USA E-mail: [email protected] Hyun-seob Cho Chungwoon University, Chungnam, Korea E-mail: [email protected] Osvaldo Gervasi University of Perugia, Italy E-mail: [email protected] Stephen S. Yau Arizona State University, Mesa, AZ, USA E-mail: [email protected] Byeong-Ho Kang University of Tasmania, Hobart, TAS, Australia E-mail: [email protected] Javier García Villalba Universidad Complutense de Madrid, Spain E-mail: [email protected]
ISSN 1865-0929 e-ISSN 1865-0937 ISBN 978-3-642-27179-3 e-ISBN 978-3-642-27180-9 DOI 10.1007/978-3-642-27180-9 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: Applied for CR Subject Classification (1998): C.2, H.4, D.2, I.2, H.3, K.6.5
Grid and distributed computing is an area that attracts many professionals from academia and industry for research and development. The goal of the GDC conference is to bring together researchers from academia and industry as well as practitioners to share ideas, problems and solutions relating to the multifaceted aspects of grid and distributed computing. We would like to express our gratitude to all of the authors of submitted papers and to all attendees for their contributions and participation. We acknowledge the great effort of all the Chairs and the members of Advisory Boards and Program Committees of the above-listed event. Special thanks go to SERSC (Science and Engineering Research Support Society) for supporting this conference. We are grateful in particular to the speakers who kindly accepted our invitation and, in this way, helped to meet the objectives of the conference. December 2011
Chairs of GDC 2011
Preface
We would like to welcome you to the proceedings of the 2011 Grid and Distributed Computing (GDC 2011) Conference—one of the partnering events of the Third International Mega-Conference on Future-Generation Information Technology (FGIT 2011) held during December 8–10, 2011, at Jeju Grand Hotel, Jeju Island, Korea. GDC 2011 focused on various aspects of advances in grid and distributed computing. It provided a chance for academic and industry professionals to discuss recent progress in the related areas. We expect that the conference and its publications will be a trigger for further related research and technology improvements in this important subject. We would like to acknowledge the great efforts of the GDC 2011 International Advisory Board, Special Session Organizers, as well as all the organizations and individuals who supported the idea of publishing this volume of proceedings, including the SERSC and Springer. We are grateful to the following keynote, plenary and tutorial speakers who kindly accepted our invitation: Hsiao-Hwa Chen (National Cheng Kung University, Taiwan), Hamid R. Arabnia (University of Georgia, USA), Sabah Mohammed (Lakehead University, Canada), Ruay-Shiung Chang (National Dong Hwa University, Taiwan), Lei Li (Hosei University, Japan), Tadashi Dohi (Hiroshima University, Japan), Carlos Ramos (Polytechnic of Porto, Portugal), Marcin Szczuka (The University of Warsaw, Poland), Gerald Schaefer (Loughborough University, UK), Jinan Fiaidhi (Lakehead University, Canada) and Peter L. Stanchev (Kettering University, USA), Shusaku Tsumoto (Shimane University, Japan), Jemal H. Abawajy (Deakin University, Australia). We would like to express our gratitude to all of the authors and reviewers of submitted papers and to all attendees, for their contributions and participation, and for believing in the need to continue this undertaking in the future. Last but not the least, we give special thanks to Ronnie D. Caytiles and Yvette E. Gelogo of the graduate school of Hannam University, who contributed to the editing process of this volume with great passion. This work was supported by the Korean Federation of Science and Technology Societies Grant funded by the Korean Government. December 2011
Tai-hoon Kim Hojjat Adeli Hyun-seob Cho Osvaldo Gervasi Stephen S. Yau Byeong-Ho Kang Javier Garcia Villalba
Organization
General Co-chairs Hyun-seob Cho Osvaldo Gervasi Stephen S. Yau
Chungwoon University, Korea University of Perugia, Italy Arizona State University, USA
Program Co-chairs Byeong-Ho Kang Javier Garcia Villalba Myoung-kwan Oh Tai-hoon Kim
University of Tasmania, Australia Complutense University of Madrid, Spain Hyejeon University, Korea GVSA and University of Tasmania, Australia
Publication Chair Keun-Wang Lee
Chungwoon University, Korea
Publicity Chair Min-sun Kim
Hyupsung University, Korea
International Advisory Board Hak-min Kim Jong-sun Lee Ok-hwan Kim
Soonchunhyang University, Korea Daejin University, Korea Kongju National University, Korea
Publication Committee Dae-Kyu Kim Hyun-Tae Jang Jong-Chan Lee Jung-Hwan Park Kwang-Hwan Kim So-Yeon Min Sun-Yeob Kim Tae-Seon Cho Young-Chul Joo
Sangmyung University, Korea Hanseo University, Korea Kunsan National University, Korea Jeju National University, Korea Konyang University, Korea Seoil University, Korea Namseoul University, Korea Chungwoon University, Korea Soonchunhyang University, Korea
X
Organization
Publicity Committee Dae-seok Rho Hyoung-keun Park Hyun Hahm Sang-young Oh Shin-hyeong Choi Young-choon Kim
Korea University of Technology and Education, Korea Namseoul University, Korea Chungwoon University, Korea Youngdong University, Korea Kangwon National University, Korea Kongju National University, Korea
Program Committee Albert Zomaya Bilha Mendelson BongHee Hong Chao-Tung Yang Chia-Han Lee Ching-Nung Cho-Li Wang Chun-Hsi (Vincent) Huang Damon Shing-Min Liu Dan Grigoras Dan Meng Daniel S. Katz Danilo Gonzalez Dimitrios Serpanos Domenico Laforenza Domenico Talia Eung Nam Ko Farrukh Nadeem Gail-Joon Ahn Geoffrey Fox George Bosilca
Gi-sung Lee Mohamed Jemni Gyoo-soo Chae Mohand-Said Hacid Hai Jin Nabil Abdennadher Heui-seok Lim Omer F. Rana Hung-Chang Hsiao Ramin Yahyapour Hyeong-Ok Lee Ronald Perrott Jan-Jan Wu Ruay-Shiung Chang Javid Taheri Shiow-Yang Wu Jean-Louis Pazat Si-ho Cha Jiannong Cao Susumu Date Jung-won Cho Tom`as Margalef Keecheon Kim Wei-Jen Wang Kenichi Takahashi Won-ho Lee Kuan-Ching Li Yangwoo Kim Kwang-hyoung Lee Yeh-Ching Chung Liria Matsumoto Sato Yeong-Tae Song Marcin Paprzycki Yong Man Ro Marian Bubak A.G.H. Krakow Yongik Yoon Mark Baker Yong-Kee Jun Minglu Li Young Choon Lee Min-Xiou Chen
Table of Contents
Appropriate Education to Improve Computer-Aided Patient Record Ability for Apprentice Doctors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seong-Ran Lee and Kwang-Hwan Kim
1
An Image Processing for Omnidirectional Image Based on the Bresenham Raster Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jin Shin and Soo-Yeong Yi
Development of an Active Ranging System Based on Structured Light Image for Mobile Robot Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jin Shin and Soo-Yeong Yi
49
A Study on Fault Analysis of Wind Power (DFIG) in Distribution Systems Based on the PSCAD/EMTDC . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jintaek Jeon, Joonho Son, Munseok Lee, Munkbaht, and Dae-seok Rho Poor Periodontal Health in Type 1 Diabetic Youth . . . . . . . . . . . . . . . . . . . Ji-Hye Park, Youn-Hee Choi, Cheol-Woo Ko, Heung-Sik Kim, Sang-Gyu Lee, Hyung-Gyoo Kim, and Keun-Bae Song Performance Evaluation of MapReduce Applications on Cloud Computing Environment, FutureGrid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yunhee Kang and Geoffrey C. Fox Development of Evaluation Simulator of New Power Sources Using the LabVIEW Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kyung-Sang Ryu, Byung-ki Kim, Suk-hwan Song, Seob-han Kim, and Dae-Seok Rho
59
67
77
87
XII
Table of Contents
Evaluation Algorithm for Bi-directional Protection Coordination in Distribution System with Distributed Generations . . . . . . . . . . . . . . . . . . . . Sohee Kim, Seongho Gi, Bangwoon Song, Seungkyou Choi, and Dae-seok Rho Development of Remote Monitoring and Control Device for 50KW PV System Based on the Wireless Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jea-Bum Park, Byung-Mok Kim, Jian Shen, Sun-Young Kim, and Dae-Seok Rho
95
102
Operation Method for Protection Devices at the Primary Feeders with New Energy Sources Using EMTDC/PSCAD . . . . . . . . . . . . . . . . . . . . . . . . Hasan, Dae-seok Rho, Chanhyeok Kim, and Yongpeel Wang
110
The Design of AES-DN Embedded Cryptographic Processor for the RFID/USN Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seon-Keun Lee and Sun-Yeob Kim
An Enhanced Data Privacy Mechanism Suitable for Ubiquitous Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jeong-Hyo Park, Yong-Hoon Jung, Kwang-Hyung Lee, Keun-Wang Lee, and Moon-Seog Jun Grinding Ability and Characteristics in Normal QT Heat Treatment vs. Carbo-nitriding Heat Teatment Bearing Steel under CBN Wheel Grinding Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jeong Jin Kim, Sung Hoon Oh, and Byung Moon So Correlation between Low Polarization and Roughness in Low-k SiOC Thin Films by Chemical Vapor Deposition . . . . . . . . . . . . . . . . . . . . . . . . . . Teresa Oh A Semantic-Based Centralized Resource Discovery Model for Grid Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abdul Khalique Shaikh, Saadat M. Alhashmi, and Rajendran Parthiban Design of Synchronization Model Using Variance Buffer . . . . . . . . . . . . . . . Keun-Wang Lee and Si-Ho Cha Mobility Management Scheme for the Wireless Mesh Network Using Location Server and Mobility Anchor Point . . . . . . . . . . . . . . . . . . . . . . . . . Sung Hwan Jang and Gi Sung Lee
138
148
155
161
171
179
Table of Contents
XIII
The Comparative Study for ENHPP Software Reliability Growth Model Based on Mixture Coverage Function . . . . . . . . . . . . . . . . . . . . . . . . . Hee-Cheul Kim and Hyoung-Keun Park
187
A Study on Authentication and Access Control of Wireless Access Point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jong-Kyung Baek and Jae-Pyo Park
195
A Reliable Multicast MAC Protocol Based on Spread Spectrum Technique in Wireless Ad-Hoc Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jeonghun Kim, Junwoo Jung, and Jaesung Lim
202
Relationship between the Binding Energy and Boundary Condition in SiOC Film for ILD Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Teresa Oh, Chy Hyung Kim, and Chi Sup Jung
213
A Framework for Active Service Pattern Mining . . . . . . . . . . . . . . . . . . . . . Jeong Hee Hwang and Mi Sug Gu
219
Character Region Detection Using Structure of Hangul Vowel Graphemes from Mobile Image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jong-Cheon Park, Byoung-Min Jun, and Myoung-Kwan Oh
RFID Mutual Authentication Protocols Using Key Exchange Based on Hash Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seung-Jae Jang, Young-Hoon Jung, So-Yeon Min, Keun-wang Lee, and Moon-Seog Jun The Development of Passive Filter Control System Based on Power Measurement and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jeong-Chay Jeon, Jae-Hyun Kim, and Jae-Geun Yoo
246
256
Preparation and Characteristics of Poly(St-co-MAA) Microgels . . . . . . . . Young Sik Kim, Sang Hyun Moh, Suk Hyung Cho, Young Jun Kim, and Jea Chul Cho
263
Identifying Organizational Culture for Patient Safety . . . . . . . . . . . . . . . . . Yoo Kyung Boo, Hyeon Sil Lee, Ji Won Park, and Seung Woo Ham
271
The Effect of a Support Program for Family Caregivers of Dementia on Stress Hormones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hyoshin Kim
278
XIV
Table of Contents
The Effect of Educational Game on the Intrinsic Motivation by Learner’s Traits: Focus on Personality Types . . . . . . . . . . . . . . . . . . . . . . . . Hyung-sung Park, Jung-hwan Park, and Young-Tae Kim
286
Detection and Recovery for Kidnapped-Robot Problem Using Measurement Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chuho Yi and Byung-Uk Choi
293
Factors Affecting the Turnover Intentions of Small and Medium Construction IT Workers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dong Hwan Cho and Haeng Nam Sung
300
Multiple Case Studies on the Critical Success Factors of Internet Banks: Implications for Korean Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dong Hwan Cho and Jin Hyuk Hong
309
Development Plan of Pet Dog Safe Transportation System Using IP-RFID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hyung-Rim Choi, Byoung-Kwon Park, Young-Jae Park, Min-Je Cho, Chang-Hyun Park, Jin-Wook Lee, and Gwang-Rok Kim The Security Process Model of Embedded Systems . . . . . . . . . . . . . . . . . . . Shin-Hyeong Choi
Workload Prediction and Weighted Rule-Based Task Scheduling for Face Certification System on Distributed Parallel Computing . . . . . . . . . . Tae Young Kim and Jong Sik Lee
Appropriate Education to Improve Computer-Aided Patient Record Ability for Apprentice Doctors Seong-Ran Lee1 and Kwang-Hwan Kim2 1
Division of Health, Kongju National University, 182 Shinkwan-dong, Kongju, Chungnam, 314-701, South Korea [email protected] 2 Department of Hospital Management, Kon-yang University, 28 Gasuwon-dong, Seo-gu, Daejeon, 302-718, South Korea [email protected]
Abstract. This study was focused on appropriate education to improve computer-aided patient record ability for apprentice doctors. The records of 225 patients of a total of 4,519 patients, who had been discharged per a month from two tertiary care located in the area of metropolitan hospitals were reviewed to examine incomplete recording rate. In this work, the reduced value in incomplete recording rates after education were plotted as a function of time elapsed after education. The present research showed that although the incomplete recording rates in two hospitals were about 85.1-91.7%, it can be reduced to 43.952.4% by the education. In order to maintain the education effect well, it is very important to determine adequate education period and develop various education programs in consider hospital circumstances. Keywords: Computer-aided patient record, Apprentice doctors, Education, Ability.
information that is useful for the improvement of medical service quality[6],[7], Nevertheless, many apprentice doctors do not their best in performing patient record that is their obligatory duty, resulting in various problems such as delay of medical service or misdiagnosis. Thus, this study was done to examine appropriate education to improve computer-aided patient record ability for apprentice doctors.
2
Materials and Methods
The computer-aided patient records of 225 patients of a total of 4,519 patients, who had been discharged from two tertiary care located in the area of metropolitan hospitals during a period from January 10, 2011 to February 10, 2011, were reviewed to examine incomplete recording rate based on the JCAHO’s Comprehensive Accreditation Manual for Hospitals[5] and some other references[6],[8],[9]. Further, in order to investigate whether doctors realize the importance of computer-aided patient recording, 146 doctors among a total of 477 doctors in study hospitals were survey. Based on the examination result, six clinical departments, which revealed the highest incomplete patient recording rates, were selected from two hospitals. Then 62 doctors were selected from the six clinical departments in two hospitals for education. A patient recoding expert taught doctors from A hospital from March 21, 2011 to April 21, 2011 and a patient recording researcher performed education of doctors from B hospitals from May 9, 2011 to June 9, 2011. The education was performed based on Video, CD-ROM, teaching, case study, discussion, and others for four hours during one month period. Then, the education effect was estimated by the reduction in incomplete recording rate after education, compared with that before education. In this work, the reduced value in incomplete recording rates after education were plotted as a function of time elapsed after education; 7 days, 30 days, 60 days, and 90 days. Then the education effect was characterized as a function of time elapsed after education for six departments of two hospitals. The pairwise t-test was done to compare the incompleteness before and after education in two hospitals.
3
Results
3.1
Basic Informations for Study Hospitals
Table 1 was done to investigate on the basic informations for two study hospitals. The number of beds was 831 for A hospital and 918 for B hospital. Average of bed occupancy rate per a month was 88.7% for A hospital and 93.2% for B hospital. The number of discharge patients per a month was 1,804 for A hospital and 2,715 for B hospital. Average recording times in progress note for a survey period was 4.81 times for A hospital and 4.29 times for B hospital.
Appropriate Education to Improve Computer-Aided Patient Record Ability
3
Table 1. Basic Informations for Study Hospitals
Variables No. of beds Average bed occupancy rate per a month No. of discharge patients per a month No. of apprentice doctors Average recording times in progress note Total numbers of patient records reviewed
3.2
Hospitals A B 831 918 88.7 93.2 1,804 2,715 176 301 4.81 4.29 93 127
A Survey of Computer-Aided Patient Record Incompleteness
Table 2 was done to examine the computer-aided patient record incompleteness in study hospitals. Serious problems, which were most responsible for incomplete patient recording, were the missing of therapeutic plan and the use of abbreviation on the admission and discharge record. The therapeutic plan took 87.1% of the incomplete recording rate for A hospital and the use of abbreviation was 88.2% for B hospital. Table 2. A Survey of Computer-Aided Patient Record Incompleteness Unit : No.(%)
Patient Records Operation and procedure Therapeutic result Therapeutic plan Abbreviation Final diagnosis Complication Laboratory result Therapeutic plan Admission note Impression after physical exam Follow plan Recording rate after 4 days based on length of stay Recording of subjective, objective, assessment and plan Total
Comparison of the Incompleteness Before and After Education
Table 3 was done the pairwise t-test to compare the incompleteness before and after education in two hospitals. There was significantly higher in the lapse 30 days since the education than before education for A hospital(t=55.4, p<.05). The mean score of education effect durability after 30 days was 57.73. On the other hand, after the lapse of 60 days since the education for B hospital, the mean score of education effect durability was 55.38. It was significantly higher than before education(t=75.1, p<.05). Table 3. Comparison of the Incompleteness Before and After Education
Time Before education 1 week after education 30 days after education 60 days after education 90 days after education *p<.05
Fig. 1 was done to compare the durability of education effect as a function of time elapsed after education in two hospitals. It was investigated that the education effect was similarly high after the lapse of 7 days since the education in two hospitals. However, the education effect was lower after the lapse of 30 days, 60 days since the education.
Fig. 1. Durability of education effect after education * Slope = ΔY ΔX *Ratio = ΔYa ΔYb
Where ΔX : time interval. ΔY : variation of incomplete recording rate. Where ΔYb : incomplete recording rate before education. ΔYa : incomplete recording rate after education.
Appropriate Education to Improve Computer-Aided Patient Record Ability
3.5
5
Education Effect of Two Department Groups for A Hospital
Fig. 2. was done to compare the education effect for two different department groups for A hospital ; non-surgery and surgery. It was investigated that the education effect was higher at the non-surgery than the surgery after the lapse of 30 days since the education for A hospital. However, the education effect was lower at the non-surgery than at the surgery after the lapse of 60 days since the education.
Fig. 2. Education Effect of Two Department Groups for A Hospital
3.6
Education Effect of Two Department Groups for B Hospital
Fig. 3 was done to compare the education effect for two different department groups for B hospital; surgery and non-surgery. The education effect was investigated to be higher at the non-surgery than the surgery for B hospital, regardless of time elapsed of 30 days after the education. In particular, the education effect was investigated to decrease more rapidly with time elapsed of 60 days after the education at the surgery as compared to the non-surgery.
Fig. 3. Education Effect of Two Department Groups for B Hospital
6
4
S.-R. Lee and K.-H. Kim
Discussion
This study was attempted to identify the effect of education on the improvement computer-aided patient record completeness for two hospitals. Serious problems, which were most responsible for incomplete patient recording, were the missing of therapeutic plan and the use of abbreviation on the admission and discharge record. The incomplete recording rate of therapeutic plan indicated 87.1% for A hospital and the use of abbreviation was 88.2% for B hospital, respectively, before the education. But this result was slightly different on the items from other studies[3],[9]. The present works indicate that a system should be developed to reduce incomplete recording rate resulting from doctor’s mistake such as therapeutic plan missing throughout the reviewing of the patient recording input by a senior doctor. Moreover, periodic and persistent education should be performed for apprentice doctors in order to reduce incomplete recording rate fundamentally. Further, an evaluation on education is also necessary for the improvement of education quality throughout the reflection of trainee’s opinion as well as the strengthening of education effect. The present research was investigated that the education effect was higher at the non-surgery than the surgery after the lapse of 30 days since the education for A hospital. However, the education effect was lower at the non-surgery than at the surgery after the lapse of 60 days since the education. On the other hand, the education effect was investigated to be higher at the non-surgery than the surgery for B hospital, regardless of time elapsed of 30 days after the education. In particular, the education effect was investigated to decrease more rapidly with time elapsed of 60 days after the education at the surgery as compared to the non-surgery. The finding was consistent with the result of earlier research on diabetes mellitus patient education[10]. Thus, year-based education should be performed more often at the surgery than the nonsurgery for B hospital as well as A hospital. The present research showed that although the incomplete patient recording rates in two hospitals were about 85.1-91.7%, it can be reduced to 43.9-52.4% by the education, which is similar to data reported in the United States[8],[11]. However. it should be noted that the education effect does not maintain for so long. According, in order to maintain the education effect well, it is very important to determine adequate education period and develop various education programs in consideration of hospital circumstances. The present work elucidated throughout the statistical analysis how effectively the synthetic and systematic education contributes to the improvement of patient recording quality in two tertiary care metropolitan hospitals. However, there existed a limitation in the formulization of the present data due to the scarcity of raw data. For this example, I failed to analyze the education effect as a function of the clinical department because only few doctors participated in the study. The future work should focus on the study of the education effect as a function of hospital scale or classification thoughout more prolonged research based on a larger data base.
Appropriate Education to Improve Computer-Aided Patient Record Ability
5
7
Conclusion
In conclusion, in order to improve computer-aided patient recording quality throughout education, the hospital motivates apprentice doctors to attend education voluntarily, and so, it has to help apprentice doctors realize how much recording influences the improvement of medical service quality. Finally, the hospital should set up a budget for patient recording education and so, employ a patient recording expert who is entirely assigned for education. These efforts made by the hospital will reduce incomplete patient recording rate effectively, providing the hospital with benefits such as the improvement of medical service quality. Acknowledgments. We would like to thank the hospital members who supported in patient record use and apprentice doctors who participated in examination the education effect on the improvement of computer-aided patient recording completeness.
References 1. Counte, M.A., Kjerulff, K.H., Salloway, J.C., Campbell, B.C.: Implementation of a Medical Information System. Evaluation of Adaptation, pp. 317–320. Summver, HCM (2008) 2. Van De Velde, R.: Hospital Information System, pp. 205–209. Springer, New York (2009) 3. Jung, M.J.: Comparison of Documentation Completeness Between the Paper-Based Medical Record and the Electronic Medical Record, pp. 178–181. Graduate School of Yonsei University in Korea (2005) 4. Joint Commission Mission Oakbrook Terrace Illinois, pp. 231–235. Comprehensive Accreditation Manual for Hospitals JCAHO Department of Publication (1997) 5. Dean, F., Kuperman, G.J., Fisko, J.: Evaluating Physician Satisfaction Regarding User Interactions with an Electronic Medical Record System. AMIA, pp. 400–404 (2001) 6. Drazen, E.l., Metzger, J.B., Ritter, J.L., Schneider, M.K.: Patient Care Information Systems, pp. 151–154. Springer, Heidelberg (2003) 7. Shiffman, R.N., Cynthia, A.: Brandt. Transition to a Computer-Based Record Using Scannable. Structured Encounter Forums, pp. 37–39 (2002) 8. Trewwk: The Potential of Electronic Medical Record Systems to Support Quality Improvement Work and Research in Norwegian General Practice. J. Med. Information 15, 241–244 (2005) 9. Matsumua, Y., Takeda, H., Okada, T., Kuwata, S., Kakazawa, H., Hazumi, N.: Method of Structured Data Entry in Electronic Patient Record. J. Med. Information 17, 193–201 (2007) 10. Lee, H.Y.: Blood Glucose Variation Related to Exercise Physiology, vol. Korean Diabetes Association 11, 49–51 (2010) 11. Terry Hannan, M.J.: The Electronic Medical Record, a Missing Component of the Health Information Infrastructure in Australia. MD. Computing 14, 79–80 (2005)
An Image Processing for Omnidirectional Image Based on the Bresenham Raster Algorithm* Jin Shin and Soo-Yeong Yi** Seoul National University of Science and Technology [email protected], [email protected]
Abstract. In this paper, an image processing algorithm for omnidirectional image is proposed. Since the omnidirectional image contains many circular arc segments, a radial directional processing for certain features in the image is appropriate rather than the horizontal or vertical directional processing as in the usual rectangular shape image. Computation of the proposed radial processing algorithm is based on the well-known Bresenham raster algorithm for line drawing in discrete integer space, thereby it is very fast and efficient. Comparison between the proposed algorithm and the conventional horizontal processing in terms of feature detection and computational efficiency are presented in experiments. Keywords: omnidirectional structured light image, Bresenham raster algorithm, image feature point, search.
1
Introduction
Recently in the field of security and surveillance, mobile robots, and video conference, the omnidirectional imaging technology is popularly used. Compared with the o
conventional image having narrow FOV(Field-Of-View) less than 90 , the omnio
directional image has 360 all directional view containing much more information than the conventional one. The omnidirectional image can be acquired through a catadioptric (catoptric(reflective) + dioptric(refractive)) approach that uses an ordinary camera and a curved mirror. The omnidirectional imaging system combined with the structured light technique can be used as an omnidirectional ranging system [3,4]. Since it is simple and cost-effective, such omnidirectional ranging system can be an alternative for the ultrasonic ring sensors or motorized laser scanner [2]. An exemplar 360 all directional distance measurement system was proposed based on the catadioptric omnidirectional imaging approach with the structured light as illustrated in Fig. 1 [3,4]. From the o
omnidirectional structured light image shown in Fig. 1 (b), it is possible to remove the background image and extract only the structured light image shown in Fig. 1 (c) by e.g., differencing the pair of frames with structured light modulation or by using a narrow band-pass optical filter that passes only spectral frequency of the structured light [9]. To calculate the distance to an object from the extracted structured light image, an image processing such as differentiation between neighboring pixels and thresholding should be done in order to detect feature points of the structured light pixels. In general, the horizontal, vertical, or diagonal directional processing is appropriate for usual rectangular shape image from a conventional camera. However, for the omnidirectional image shown in Fig. 1 (b) or (c), a radial directional processing from the optical center is more suitable for detecting edge feature points of the structured light pixels as illustrated in Fig.2.
(a) Omnidiectional ranging system with structured light
(b) Omnidirectional structured light image
(c) Structured light pixel image extracted
Fig. 1. Omnidirectional structured light imaging system and structured light pixel image
10
J. Shin and S.-Y. Yi
In this paper, an efficient radial directional processing algorithm for an omnidirectional structured light image is proposed. In order to improve the computational efficiency, computational structure of the well-known Bresenham raster algorithm is adopted to the radial processing. Commonly used in the field of computer graphics, the Bresenham raster algorithm draws a line between given two points by using only the simple operations such as integer addition or subtraction. Comparison of the proposed radial processing with the vertical or horizontal processing shows a lot of advantages in feature point detection and computational efficiency. For demonstration purpose, a target image of the proposed image processing is an omnidirectional structured light image in this paper. However, the proposed radial processing algorithm is applicable for feature point detection of any omnidirectional image.
Fig. 2. Radial processing for omnidirectional structured light image
2
Radial Processing with the Bresenham Raster Algorithm
As shown in Fig. 1(b) and (c), the omnidirectional image contains many circular arc edges. Thus, the edge feature points in the image might be missed out by the vertical or horizontal directional processing in the quantized image space of discrete pixel units as demonstrated in Fig. 3. Fig. 3 (a) is an example of a circular arc image. Horizontal processing will give the result shown in Fig. 3 (b). For simplicity, it is assumed the horizontal directional differentiation of the image for the feature detection. On the contrary in the radial processing, the direction angle resolution can be set arbitrarily, as a consequence, it is possible to detect every feature pixels on arc edge as shown in Fig. 3 (c). To detect the feature points in an omnidirectional image through the radial directional algorithm, processing for the pixels on one directional angle from the optical center and repetition of the processing for all directional angles in 360 should be done. A simple brute-force radial processing for a certain directional angle θ is shown in Fig. 4. o
An Image Processing for Omnidirectional Image
(a) Circular arc segment of omnidirectional structured light image
(b) Result of horizontal processing
(c) Result of radial processing Fig. 3. Results of feature point detection for omnidirectional structured light image for ( r = 0; r < 400; r += Δ r ) {
x = xs + r cos θ ; y = ys + r sin θ ; processing ( round ( x ), round ( y )); }
Fig. 4. Simple radial processing algorithm for a directional angle θ
11
12
J. Shin and S.-Y. Yi
In Fig. 4, ( xs , y s ) is the starting point of the radial processing i.e., the optical center, ( xc , yc ) . Assuming that the optical center coincides with center of 640 × 480 sized image, the length of radius r would be 320 + 240 = 400 . The function round (⋅) casts the argument into an integer and processing ( m, n) is an image processing function such as differentiation between neighboring pixels and thresholding at location ( m, n ) . This simple approach includes floating point 2
2
operation in r sin θ and r cos θ , thus, much computations are required. Moreover depending on the size of Δ r , a certain pixel might be omitted or operated twice unnecessarily in the processing.
ε = 0,
y = ys ;
for ( x = xs ; x < xe ; x ++ ) {
putpixel ( x, y ) ; if (ε + m < 0.5)
ε = ε + m; else y ++ ,
ε = ε + m − 1;
}
Fig. 5. Pseudo code for Bresenham line drawing algorithm
Fig. 6. Bresenham line drawing algorithm
An Image Processing for Omnidirectional Image
13
To solve these problems, the computation of the well-known Bresenham raster algorithm is adopted to the radial processing in this paper. The Bresenham line drawing algorithm determines the points in discrete integer space in order to plot an approximate line between two given points ( xs , y s ) and ( xe , ye ) . The pseudo code representation for the Bresenham raster algorithm is given in y − ys Fig. 5 in case that the slope m = e between two points is 0 ≤ m ≤ 1 . In this xe − xs case, the reference coordinate is x . Starting from xs , the value of x is incremented
by 1 and the corresponding error ε in y is accumulated by m repeatedly. According to the fraction of the error, a pixel ( x, y ) is marked as an approximation point on the line at each step (Fig. 6). In the pseudo code, putpixel ( x, y ) is the marking function. The line of m > 1 is symmetrical to that of 0 ≤ m ≤ 1 with respect to the diagonal line y = x . Thus, the algorithm described in Fig. 5 is applicable simply by changing the variables x and y each other [6,7]. The radial processing proposed in this paper adopts computational method of the Bresenham raster algorithm as explained in Fig. 7. Without loss of generality, we assume 640 × 480 size of the image. It is noted that the optical center, ( xc , yc ) does not coincide with the image center at (320, 240) . For a certain directional angle θ , a pixel position ( x , y ) on the line through the image center should satisfy the following; tan θ = o
In case of 0 ≤ θ ≤ 45
o
y − yc
(1)
x − xc
or 0 ≤ m ≤ 1 , two points ( xs , y s ) and ( xe , ye ) for the radial
processing based on the Bresenham algorithm is obtained as follows; Let iys be round-off integer of y when x = 0 in (1), i.e., iy s = round ( yc − xc tan θ ) (Fig. 7). Then, if iy s ≥ 0 , the pair of points, ( xs , y s ) and ( xe , ye ) are
( xs , ys ) = (0, iys ), ( xe , ye ) = (639, iye ) where
iye
{
is
round-off
integer
of
(2) y
when
x = 639
i.e.,
}
iye = round yc − (639 − xc ) ⋅ tan θ . And if iy s < 0 , two points should be
( xs , ys ) = (ixs′ , ( xe , ye ) = (ixe′ ,
0)
(3) 479)
14
J. Shin and S.-Y. Yi
where ixs′ = round xc +
and tan θ
yc
ixe′ = round xc +
479 − yc tan θ
. It is noted that
processing region of x are 0 ≤ x ≤ 639 when iy s ≥ 0 and ixs′ ≤ x ≤ ixe′ when iy s < 0 respectively. o
On the other hand, in case of 45 < θ ≤ 90
o
or 1 < m < ∞ , it is possible to get
( xs , y s ) and ( xe , ye ) in similar manner. It is noted that the reference axis is y
instead of x in this case. It is also possible to get the two points in the same way as the case of m < 0 .
Fig. 7. Radial processing based on the Bresenham raster algorithm
To summarize, ( xs , y s ) and ( xe , ye ) for the Bresenham based radial processing are obtained by using (2) and (3) for a certain directional angle, θ and the radial o
o
processing should be done for all directional angles within the range, −90 < θ ≤ 90 .
3
Experiment Results and Analysis
This section presents performance of the proposed radial processing in terms of feature detection and computational efficiency compared with the conventional horizontal processing. As the image processing for detecting features, simple
An Image Processing for Omnidirectional Image
15
differentiation between neighboring pixels and thresholding are carried out in this paper. Fig. 8 shows results of the feature detection for the omnidirectional structured light image in Fig. 1 (c). The resultant feature image from the proposed radial processing in Fig. 8 (b) is much brighter than the image in Fig. 8 (a) because of the density of the structured light pixels detected in the image processing. Especially in the circle region in Fig. 8 (a), many image features of the structured light pixels are missed out. Result of the image processing depends on sampling interval of the directional angle of course. The sampling interval Δθ is 0.1o in this experiment. In order to compare the computational efficiency, repeat counts of for-loops in the image processing are analyzed. In the brute-force image processing presented in Fig. 4, a directional line processing requires 400 for-loop counts and overall processing needs o
o
400 × 360 × 10 = 1, 440, 000 counts in case of Δθ = 0.1 . In comparison, the repeat o
counts needed in the proposed radial processing with the same Δθ = 0.1 is as follows; When m ≤ 1 , the reference axis is x and the counts are o
640 × 90 × 10 = 576, 000 and when m
> 1 , the reference axis is
y and the counts
o
are 480 × 90 × 10 = 432, 000 . Thus, the total counts are 1, 008, 000 which is approximately 2 / 3 of the brute-force processing algorithm. It should be noted that the number of the for-loop repetitions in the proposed radial processing is less than the above analyzed since the explored interval in the reference axis is reduced according to the position of the optical center, ( xc , yc ) . Therefore, compared to the basic algorithm in Fig. 4, the proposed radial processing algorithm has an advantage not just because it uses only integers operation, but also because of its fewer for-loop repetitions which makes it more efficient in computation.
(a) Horizontal processing
(b) Radial processing
Fig. 8. Result of image processing
16
4
J. Shin and S.-Y. Yi
Conclusion
This paper proposes the radial directional image processing for feature detection in an omnidirectional image. To detect edge feature points of a circular omnidirectional image, the radial processing from an optical center is more suitable than the conventional horizontal or vertical directional processing. Computational structure of the radial processing algorithm suggested in this paper is based on the well-known Bresenham raster algorithm using only simple integer operations with fewer repeat counts that make the overall computation highly efficient. Through experiments on an omnidirectional structured light image, it is verified that the proposed radial processing algorithm gives a better performance of edge feature detection of structured light pixels.
References 1. Nayar, S.: Catadioptric omnidirectional camera. In: Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, pp. 482–488 (1997) 2. Jain, R., Kasturi, R., Schunck, B.G.: Machine Vision. McGraw-Hill (1995) 3. Joung, I., Cho, H.: An active omnidirectional range sensor for mobile robot navigation. Control Engineering Practice 6(3), 385–393 (1988) 4. Yi, S., Choi, B., Ahuja, N.: Real-time omni-directional distance measurement with active panoramic vision. Int’l Jour. of Control, Automation, and Systems 5(2), 184–191 (2007) 5. Shin, J., Yi, S., Hong, Y., Suh, J.: Omnidirectional Distance Measurement based on Active Structured Light Image. Jour. of Institute of Control, Robotics and Systems 16(8), 751–755 (2010) 6. Bresenham, J.: Algorithm for computer control of a digital plotter. IBM Systems Journal 4(1), 25–30 (1965) 7. http://en.wikipedia.org/wiki/Bresenham’s_line_algorithm 8. Kweon, G., Hwangbo, S., Kim, G., Yang, S., Lee, Y.: Wide-angle catadioptric lens with a rectilinear projection scheme. Applied Optics 45(34), 8659–8673 (2006) 9. Moigne, J., Waxman, A.: Structured light patterns for Robot Mobility. IEEE Journal of Robotics and Automation 4(5), 541–548 (1988)
A Novel Half-Bridge Resonant Inverter for CCFL-Driving Gang-Youl Jeong Department of Electronic Information Engineering, Soonchunhyang University 646, Eupnae-ri, Shinchang-myun, Asan-si, Chungnam, 336-745, South Korea [email protected]
Abstract. This paper presents a novel half-bridge resonant inverter for CCFLdriving. The proposed inverter uses half-bridge inverter topology and is controlled by an 8-bit one-chip microcontroller. Therefore, the inverter structure is very simple. The inverter reduces the ignition voltage and eliminates current spikes through burst dimming control with soft-starting. The design example and software implementation for a prototype of the inverter are shown, and the experimental results of the prototype are shown to verify the characteristics of the inverter. Keywords: Half-bridge, resonant inverter, soft-starting, CCFL-driving.
CCFL and the ground, which causes a large error in the sensed current of the CCFL and subsequently degrades the dimming performance. In contrast, burst dimming, also called digital dimming, dims a display by driving the CCFL between no current and a rated current, and thus it does not have the same problems as analog dimming. Three burst dimming control methods have recently been introduced in order to resolve the problems associated with the analog dimming control: duty-ratio control, frequency control, and voltage control [4]-[7]. The duty-ratio control regulates the lamp current, but the asymmetrical lamp current and poor lamp crest factor result in discoloration of the lamp [6]. The frequency control [8] is the most common technique in regulating the lamp current; however, the corresponding dimming range is restricted significantly by the switching frequency. Voltage control has acceptable dimming performance, but is too complex to implement, and hence not suitable for low power applications. Burst dimming involves the back-and-forth start and cut-off of the CCFL. This creates current spikes and reduces the lifespan of the CCFL. Therefore, resultant current spikes must be suppressed by appropriate control strategies [9]-[10] to extend the lamp life. In this paper, a novel half-bridge inverter based on burst dimming control with soft-starting for CCFL-driving is presented to overcome the disadvantages mentioned above. The design of and software implementation for the proposed inverter prototype are shown, and experimental results obtained using the prototype agreed closely with the theoretical analysis and explanation.
2
Power Stage of the Proposed Inverter Vdc 2
Lr1
Q1
1:n
a
CC
Vdc
Lm
Lr2
il CR
CB
Cs
b Vdc 2
Q2
T
(a) The half-bridge resonant inverter
Lr
vl CCFL (Rlamp)
Vs=nVrms
Cp
vo
R'lamp
(b) The simplified circuit
Fig. 1. The proposed half-bridge resonant inverter
The CCFL inverter helps supply power to the CCFL of the LCD backlight module. A circuit diagram of the proposed single-stage half-bridge resonant inverter is shown in Fig. 1(a), where the CCFL is represented as a variable resistor Rlamp and the parasitic capacitances in the transformer and between the lamp and ground are omitted. As shown in Fig. 1(a), the half-bridge resonant inverter is used to power Rlamp, which is composed of one main transformer T with a turn ratio n (=N2/N1), one DC blocking capacitor CC, one stabilization capacitor CB, one resonance capacitor CR, and two metal oxide semiconductor field effect transistor (MOSFET) switches Q1 and Q2. The transformer is represented by leakage inductances Lr1 and Lr2 at the primary and secondary, respectively, and by the magnetization inductance Lm. Based on the proposed control technique, Rlamp is driven by the sinusoidal voltage and current
A Novel Half-Bridge Resonant Inverter for CCFL-Driving
19
generated from the half-bridge resonant inverter. Since the CCFL characterizes negative dynamic impedance after ignition, CB is required to stabilize the lamp current [4]. Fig. 1(b) illustrates the most simplified circuit of Fig. 1(a), which is the series-resonant series-parallel-loaded circuit. The relationships between each circuit parameter are as follows: √
V
V
(1)
CC
C
(2) L L
L
L
L
L
CB
CR
C
(4)
CB R
R
R
(3)
CB R
(5)
where Vrms is the fundamental voltage value of the sinusoidal waveform and ω is the switching angular frequency. During the lamp startup time, the Rlamp value approaches infinity and a high voltage is applied to the lamp to ignite it. In this case, the inverter circuit operates in seriesresonant parallel-loaded mode with the following resonant frequency: f
,
.
L C ||C
(6)
At this time, the corresponding following quality factor Q is almost zero: L /C
Q
R
.
(7)
As the ionization phenomenon stabilizes, the equivalent resistor Rlamp is abruptly dropped and the effect of the parallel capacitance Cp on the circuit decreases. In this case, the circuit operates in series-resonant series-loaded mode with the following resonant frequency: f
,
L C
.
(8)
The startup should be first taken into account in the design of the backlight module. Since the CCFL must be driven by high voltage, there is a voltage-boosting transformer between the main power stage and load. The turn ratio n of the transformer is determined as follows: n
V √
V
,
(9)
,
where Vstart is the minimum rms voltage across the CCFL during startup, and Vdc,min is the minimum DC input voltage.
20
G.-Y. Jeong
3
Proposed Dimming Control Algorithm
The conventional burst dimming technique controls the on and off durations of the lamp through a low frequency dimming (LFD) gating signal. The average lamp brightness is proportional to the on-duty (HIGH-part) of the gating signal. However, this technique can cause a substantial reduction in the lifespan of the lamp since the CCFL is repeatedly struck by current spikes throughout the dimming operation. Furthermore, the boost transformer generates an annoying hum due to the repeated current spikes [12]. To reduce the high ignition voltage and eliminate the current spikes, a new burst dimming control is proposed. vGS1 di t
vl
ignition with current spikes
Vspike
il
df vGS2
Vnormal
df
normal operation point il
soft start
t vab
hard start di
df
t effective on-duty (d)
(a) The conceptual comparison of dimming controls
(b) The theoretical waveforms of the proposed control
Fig. 2. The burst dimming controls
Fig. 2(a) shows a conceptual comparison of two burst dimming controls with hardstarting and soft-starting. Soft-starting with a gradual on-duty increase can substantially reduce the ignition voltage. Thus, the magnitude of the lamp current spike can be reduced by the burst dimming control with soft-start. Fig. 2(b) shows the theoretical waveforms of the proposed burst dimming control with soft-starting, where di is the initial nominal on-duty, df is the final on-duty, and vab is the voltage between a and b in Fig. 1(a). The conventional burst dimming control that uses fullbridge inverter topology begins its effective on-duty (phase-shift) as a full duty df because the full duty is the resultant value of the current controller. Compared to this, the proposed control starts its on-duty as a minimum using the nominal starting onduty di and gradually increases its on-duty to df (di→df). It can then reduce the lamp ignition voltage and eliminate the lamp current spike. The proposed dimming control applied by soft-starting, shown in Fig. 3, is composed of an A/D converter for dimming reference, the decider of the soft-starting period, the open-loop on-duty controller of soft-starting, the current controller, and the gate driver of the power semiconductor switch MOSFET. The decider of the softstarting period is a control software algorithm that decides, by the timer of the
A Novel Half-Bridge Resonant Inverter for CCFL-Driving
21
decider of soft-starting period open-loop onduty controller of soft-starting
Y: soft-start dimming reference
period calculation & timer read
A/D conversion
PWM generator or PWM turn on/ off by LFD
soft-start period? normal current control
N: normal control
one-chip microcontroller
MOSFET
output
this part?
Fig. 3. The proposed dimming control applied by soft-starting
microcontroller, whether or not the operation time is the initial time of the LFD control period. The decider then chooses between the soft-starting open-loop on-duty controller and the current controller. Also, the open-loop on-duty controller of the soft-starting gradually increases the on-duty from the minimum to the current control output on-duty per unit time (per time of the timer interrupt). The open-loop on-duty controller then provides protection while the voltage is rapidly applied across the lamp. Thus, as the open-loop controller reduces the initial lamp ignition voltage, it gradually builds-up the lamp current and eliminates the initial lamp current spikes.
4
Design of the Inverter Power Circuit Vdc 2
Lr1
Q1
1:n
a
CC
Vdc
Lr2
Lm
il
CB vl
CR
b Vdc 2
T
Q2
Q1
CCFL (Rlamp)
Q2
MOSFET gate driver
burst dimming control algorithm
protection
one-chip microcontroller
Fig. 4. The total control block diagram of the proposed half-bridge resonant inverter
Fig. 4 shows the total control block diagram of the proposed half-bridge resonant inverter. The following is a real design example, based on the proposed concept, for a single-stage CCFL half-bridge inverter for a low power LCD monitor equipped with the proposed dimming control with soft-starting. The 18W-CCFL backlight inverter circuit with three CCFLs having a lamp voltage of 600Vrms and a lamp current of 10 mArms is designed and adopted to the driver of a 15-inch class LCD monitor. While
22
G.-Y. Jeong
the lamp operating frequency is set within the range of 50-90 kHz (fs=55 kHz), the starting voltage of the adopted CCFL is 1000 Vrms and the resonant frequency fr is 50 kHz (fr=50 kHz). The related parameters such as the turn ratio n, the magnetizing inductance Lm, stabilization capacitor CB, and the blocking capacitor CC are determined as follows: Using (9), the minimal turn ratio is obtained as: √ V
n
√
V
√ ,
√
131.
(10)
As a result, the actual turn ratio n is selected as n=131. It is good that the impedance of the stabilization capacitor CB is double that of the CCFL at the rating condition [11]. Thus, CB in steady state can be determined as: I
CB
V
24pF .
(11)
Using (2) and (8), the blocking capacitor CC is determined as follows: CC
L
1900 µF ,
(12)
where the total leakage inductance Lr= 92 H is measured from the boost transformer of the implemented prototype inverter as the designed turn ratio (10). CB and CC are set to 22 pF and 2,000 F, respectively. These values are similar to the designed and commercial values used in devices currently on the market.
5
Experimental Results
Based on the above explanation and Fig. 4, a prototype of the proposed inverter is implemented. The proposed dimming controller described in Section 3 is implemented at an 8-bit one-chip microcontroller, and the power circuit used is the same as that designed in Section 4. The prototype is applied to the target load described in Section 4, the 18 W-CCFL backlight module of a 15-inch LCD monitor. The LFD frequency of the proposed dimming controller is set to 150 Hz, which is the frequency typically used in the backlight of a general LCD monitor or TV. The voltage commands for the regulation of the soft-start duration and LFD frequency according to each backlight characteristic can be easily given to the microcontroller through the A/D converter input terminals connected to a variable resistor. Therefore, as the flexibility of the implemented system improves, the characteristic of the backlight applied using the proposed system can be optionally improved. Fig. 5 shows the operational waveforms of the dimming controller of the proposed inverter. Figs. 5(a)-(b) show that the turn-on ratios of the lamps are 50% and 70%. It is evident in Fig. 5 that each operational waveform for each dimming level coincides with the above theoretical explanation. Since the control law of the proposed controller operates effectively and soft-starts the lamp, the initial lamp voltage is reduced and the current spikes are eliminated.
A Novel Half-Bridge Resonant Inverter for CCFL-Driving
(a) 50% dimming (5 mA/div., 2 msec/div.)
23
(b) 70% dimming (5 mA/div., 2 msec/div.)
Fig. 5. The operational waveforms of the dimming controller of the proposed inverter
Fig. 6 shows the microscopic waveform of the lamp when the proposed inverter operates in a steady state of 100% dimming. Because the lamp current is sinusoidal, it can be inferred that the power circuit of the proposed inverter is well-designed and operates well.
Fig. 6. The microscopic waveform of the lamp when the proposed inverter operates at a steady state of 100% dimming (5 mA/div. 10 μsec/div.)
6
Concluding Remarks
A novel half-bridge resonant inverter for CCFL-driving is presented in this paper. The proposed inverter reduces the lamp ignition voltage by soft-starting of the proposed dimming control and eliminating the lamp current spikes. Thus, this inverter improves the ignition behavior of the CCFL and increases the lifespan of the CCFL. The inverter is easily implemented by an 8-bit single-chip microcontroller. The inverter power circuit utilizes the half-bridge resonant inverter topology. The design of the power circuit is shown, and the operational algorithm of the dimming control with soft-starting is explained in detail based on simple analysis. The experimental results for the prototype along with the designed parameters coincide with the analysis and explanation. The startup current is reduced by 50% compared to the conventional fullbridge phase-shift inverter, and the current spikes are eliminated.
24
G.-Y. Jeong
References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.
Lin, Y.L., Witsulski, A.F.: Proc. IEEE IAS Annu. Meeting, pp. 2149–2152 (1996) Jordan, M., O’Connor J.A.: Proc. IEEE APEC, pp. 424–431 (1993) Lin, C.H.: IEEE Trans. Power Elec., 1459–1466 (2005) Lin, M.S., Ho, W.J., Shih, F.Y., Chen D.Y., Wu, Y.P.: IEEE Trans. Ind. Electron, 249–255 (1998) Hsieh, G.C., Lin, C.H., Hsieh, H.I.: IEEE TENCON 2001, pp.717–723 (2001) Lee, S.W., Ko, D.Y., Huh, D.Y., Yoo, Y.I.: IEEE APEC 2001, pp. 447–453 (2001) Raiser, F.: IEEE IAS 2001, pp. 499–503 (2001) Lin, C.H., Pai, K.J.: IEEE PESC 2003, pp. 1000–1005 (2003) Lin, C.H., Pai, K.J.: IEEE APEC 2004, pp. 496–501 (2004) Hsieh, G.C., Lin, C.H.: J. Light Vis. Environ., 1–9 (1999) Hwu, K.I., Chen, Y.H.: IEEE PEDS 2007, pp. 1085–1090 (2007) Jeong, G.Y.: Jour. of KAIS, 461–468 (2010)
Quorum Based Mutual Exclusion in Asynchronous Systems with Unreliable Failure Detectors Sung-Hoon Park and Seon-Hyong Lee School of Electrical and Computer Engineering, Chungbuk National Univ. CheongjuChungBuk 361-763 {spark,lseon}@chungbuk.ac.kr
Abstract. This paper defines the quorum-based fault-tolerant mutual exclusion problem in a message-passing asynchronous system and determines a failure detector to solve the problem. This failure detector, which we call the modal failure detector star, and which we denote by M*, is strictly weaker than the perfect failure detector P but strictly stronger than the eventually perfect failure detector P. The paper shows that at any environment, the problem is solvable with M*.
In this sense, the problem stems from the famous impossibility result that consensus cannot be solved deterministically in an asynchronous system that is subject to even a single crash failure [7]. 1.2
Failure Detectors
In this paper, we introduced a modal failure detector M* and showed that the mutual exclusion problem is solvable with it in the environment with majority correct processes. The concept of (unreliable) failure detectors was introduced by Chandra and Toueg [3,4], and they characterized failure detectors by two properties: completeness and accuracy. Based on the properties, they defined several failure detector classes: perfect failure detectors P, weak failure detectors W, eventually weak failure detectors W and so on. In [3] and [4] they studied what is the "weakest" failure detector to solve Consensus. They showed that the weakest failure detector to solve Consensus with any number of faulty processes is Ω+Σ and the one with faulty processes bounded by n/2 (i.e., less than n/2 faulty processes) is W. After the work of [8], several studies followed. For example, the weakest failure detector for stable leader election is the perfect failure detector P [4], and the one for Terminating Reliable Broadcast is also P [1,3]. Recently, as the closest one from our work, Guerraoui and Kouznetsov showed a failure detector class for mutual exclusion problems that is different from the above weakest failure detectors. The failure detector, called the Trusting failure detector, satisfies the three properties, i.e., strong completeness, eventual strong accuracy and trusting accuracy so that it can solve the mutual exclusion problem in asynchronous distributed systems with crash failure. And they used the bakery algorithm to solve the mutual exclusion problem with the trusting failure detector. 1.3
Contributions
How about the quorum-based mutual exclusion problem? More precisely, what is the weakest failure detector to solve the quorum-based mutual exclusion problem? The bakery algorithm is completely different from the quorum-based ME in which the order of getting the critical section is decided based on a ticket order. In contrast to the bakery algorithm, the quorum-based ME algorithm should receive the permissions from all members of a quorum to exclusively use the critical section. In general, quorum-based mutual exclusion algorithms assume that the system is either a failure-free model [13,14,16,19], or a synchronous model in which (1) if a process crash, it is eventually detected by every correct process and (2) no correct process is suspected before crash [13,16]: with the conjunction of (1) and (2), the system is assumed to equipped with the capability of the perfect failure detector P [3]. In other words, the perfect failure detector P is sufficient to solve the fault-tolerant quorum-based mutual exclusion problem. But is P necessary? For the answer to the question, we present a modal failure detector star M*, that is a new failure detector we introduce here, which is strictly weaker than P (but strictly stronger than P, the
◇
Quorum Based Mutual Exclusion in Asynchronous Systems
27
eventually perfect failure detector of [3]). We show that the answer is “no” and we can solve the problem using the modal failure detector star M*. Roughly speaking, failure detector M* satisfies (1) eventual strong accuracy and (2) strong completeness together with (3) modal accuracy, i.e., initially, every process is suspected, after that, any process that is once confirmed to be correct is not suspected before crash. If M* suspects the confirmed process again, then the process has crashed. However, M* might suspect temporarily every correct process before confirming it’s alive as well as might not suspect temporarily a crashed process before confirming it’s crash. Intuitively, M* can thus make at least one mistake per every correct process and algorithms using M* are, in terms of a practical distributed system view, more useful than those using P. We here present the algorithm to show that M* is sufficient to solve fault tolerant quorum-based mutual exclusion and it is inspired by the well-known Grid-based algorithm of Maekawa [11,15,16,17]: a process that wishes to enter its CS first gets admissions from the one of quorums. M* guarantees that a crash of the process which has been confirmed at least once will be eventually detected by every correct process in the system. We show that, in addition to mutual exclusion and progress, our algorithm guarantees also a fairness property, ensuring that any process which wants to get in a CS is granted to access the CS as first come then first served (fairness property). We do not consider here probabilistic mutual exclusion algorithms [4, 7]. 1.4
Road Map
The rest of the paper is organized as follows. Section 2 addresses motivations and related works and Section 3 overviews the system model. Section 4 introduces the modal failure detector star M*. Section 5 shows that M* is sufficient to solve the problem, respectively. Section 6 concludes the paper with some practical remarks.
2
Motivations and Related Works
Actually, the main difficulty in solving the mutual exclusion problem in presence of process crashes lies in the detection of crashes. As a way of getting around the impossibility of Consensus, Chandra and Toug extended the asynchronous model of computation with unreliable failure detectors and showed in [4] that the FLP impossibility can be circumvented using failure detectors. More precisely, they have shown that consensus can be solved (deterministically) in an asynchronous system augmented with the failure detector S (eventually strong) and the assumption of a majority of correct processes. Failure detector S guarantees strong completeness, i.e., eventually, every process that crashes is permanently suspected by every process, and eventual weak accuracy, i.e., eventually, some correct process is never suspected. Failure detector S can however make an arbitrary number of mistakes, i.e., false suspicions. A quorum-base mutual exclusion problem, simply QME, is an agreement problem so that it is impossible to solve in asynchronous distributed systems with crash failures.
28
S.-H. Park and S.-H. Lee
This stems from the FLP result which mentioning the consensus problem can’t be solved in asynchronous systems. Can we also circumvent the impossibility of solving QME using some failure detector? The answer is of course “yes”. The grid-based algorithm of Maekawa [16] solves the QME problem with assuming that it has the capability of the failure detector P (perfect) in asynchronous distributed systems. This failure detector ensures Strong Completeness (recalled above) and Strong Accuracy, i.e., no process is suspected before it crashes [2]. Failure detector P does never make any mistake and obviously provides more knowledge about failures than S. But it is stated in [7] that failure detector S cannot solve the ME problem, even if only one process may crash. This means that ME is strictly harder than Consensus, i.e., ME requires more knowledge about failures than consensus. An interesting question is then “what is the weakest failure detector for solving the QME problem in asynchronous systems with unreliable failure detectors?” In this paper, as the answer to this question, we show that there is a failure detector that solves QME weaker than the perfect failure detector. This means that the weakest failure detector for QME is not a perfect failure detector P.
3
Model
We consider in this paper a crash-prone asynchronous message passing system model augmented with the failure detector abstraction [3]. 3.1
The Fault-Tolerant Quorum-Based Mutual Exclusion Problem
We define here the fault-tolerant quorum-based mutual exclusion problem (from now on -FTQME) using the terminology and notations given in [13,16]. We associate to every process i ∈ Π a user, ui that can require exclusive access to the critical section. The users can be thought of as application programs. As in [12,13], every process i ∈ Π and every user ui are modeled as state machines. A process i ∈ Π and the corresponding user ui interact using tryi, criti, exiti and remi actions. The input actions of process i (and outputs of ui) are the tryi action, indicating the wish of ui to enter its CS, and the exiti action, indicating the wish of ui to leave its critical section. The output actions of i (and inputs of ui) are the criti action, granting the access to its critical section, and the remi action, which tells ui that it can continue its work out of its critical section. A sequence of tryi, criti, exiti and remi actions for the composition (ui, i) is called a well-formed execution if it is a prefix of the cyclically ordered sequence {tryi, criti, exiti, remi}. A user ui is called a well-formed user if it does not violate the cyclic order of actions tryi, criti, exiti, remi, ... A mutual exclusion algorithm defines trying and exit protocols for every process i. We say that the algorithm solves the FTQME problem if, under the assumption that every user is well-formed, any run of the algorithm satisfies the following properties: Mutual Exclusion: No two different processes are in their CS at the same time. Starvation Freedom: Every request for its CS is eventually granted. Fairness: Different requests must be granted in the order they are made.
Quorum Based Mutual Exclusion in Asynchronous Systems
29
Note that Mutual exclusion is a safety property while Starvation freedom is liveness properties. Let Π denote a nonempty set of n processes as defined in the previous section. A coterie C is a set of sets, where each set Q in C is called a quorum. The following conditions hold for quorums in a coterie C under Π [6]: ∀Qi∈C : Qi ≠∅ ∧ Qi ⊆U Minimality Property : No quorum is a subset of another quorum. ∀Qi, Qj ∈C : Qi ≠ Qj :¬( Qi ⊆ Qj) Intersection Property : Every two quorums intersect. ∀Qi, Qj ∈C : Qi ∩ Qj ≠ ∅. For example, C = {{a, b}, {b, c}} is a coterie under Π = {a, b, c} and Qi = {a, b} is a quorum. The concept of intersecting quorum captures the essence of mutual exclusion in distributed systems. That is, process i executes its CS only after it has locked all the processes in a quorum Qj ∈C in exclusive mode. To do this, process i sends request messages to all the processes in Qj. On receipt of the request message, the process j of the quorum Qj immediately sends a reply message to i (indicating j has been locked by i) only if j is not locked by some other process at that time. The process i can access the CS only after receiving permission (i.e., reply messages) from all the processes in the quorum P. After having finished the CS execution, i sends release messages to all the processes in the quorum Qj to unlock them. Since any pair of quorums have at least one process in common (by the Intersection Property), mutual exclusion is guaranteed. The Minimality Property is not necessary for correctness, but it is useful for efficiency.
4
The Modal Failure Detector Star M*
Each module of failure detector M* outputs a subset of the range 2Π. Initially, every process is suspected. However, if any process is once confirmed to be correct by any correct process, then the confirmed process id is removed from the failure detector list of M*. If the confirmed process is suspected again, the suspected process id is inserted into the failure detector list of M*. The most important property of M*, denoted by Modal Accuracy, is that a process that was once confirmed to be correct is not suspected before crash. Let HM be any history of such a failure detector M*. Then HM(i,t) represents the set of processes that process i suspects at time t. For each failure pattern F, M(F) is defined by the set of all failure detector histories HM that satisfy the following properties: • Strong Completeness: There is a time after which every process that crashes is permanently suspected by every correct process: ─ ∀i,j∈Ω, ∀i∈correct(F), ∀j∈F(t), ∃ t’’:∀t’>t’’, j∈H(i, t’). •
Eventual Strong Accuracy: There is a time after which every correct process is never suspected by any correct process. More precisely:
Modal Accuracy: Initially, every process is suspected. After that, any process that is once confirmed to be correct is not suspected before crash. More precisely: ─ ∀i,j∈Ω: j∈H(i,t0), t0< t< t’ , j∉ H(i ,t) ∧ j∈ Ω-F(t’) => j∉ H(i, t’)
Note that Modal Accuracy does not require that failure detector M* keeps the Strong Accuracy property over every process all the time t. However, it only requires that failure detector M* never makes a mistake before crash about the process that was confirmed at least once to be correct. If process M* outputs some crashed processes, then M* accurately knows that they have crashed, since they had already been confirmed to be correct before crash. However, concerning those processes that had never been confirmed, M* does not necessarily know whether they crashed (or which processes crashed).
5
Solving FTQME Problem with M*
We give in Figure 1 an algorithm solving FTQME using M* in any environment where at least one quorum is available. The algorithm uses the fact that eventual strong accuracy property of M*. More precisely, with such a property of M* and the assumption of at least one quorum being available, we can implement our algorithm of Figure 1. Note here that we don’t consider the dead lock situation where two or more processes concurrently trying to obtain permissions from each number of quorums but only get in infinitely waiting. In this algorithm, we assume that there is a mechanism to resolve the dead lock. We give in Figure 1 an algorithm solving FTQME using M* in any environment E with any number of correct processes ( f < n ). Our algorithm of Figure 1 assumes:
Each process i has access to the output of its modal failure detector module Mi*; At least one quorum is available; Each process i is well-formed; A dead lock resolving mechanism is installed;
In our algorithm of Figure 1, each process i has the following variables: 1. A variable status, initially rem, represents one of the following states {rem, try, incs, wait }; 2. A boolean my_tokeni, initially true, indicating whether i has the its token; 3. A variable my_token_holderi, initially NULL, which denotes the token holder when i send its token to other node; 4. A list token_listi, initially empty, keeping the tokens that i has received from each member of a quorum. Description of [Line 1-6] in Figure 1; The idea of our algorithm is inspired by the well-known Quorum-based ME algorithm of Maekawa [11,12]. That is, the processes that wish to enter their CS first wait for a quorum whose members are all alive based on the information HM from its failure detector M*. Those processes eventually find
Quorum Based Mutual Exclusion in Asynchronous Systems
31
out the quorum by the eventual strong accuracy property of M* in line 3 of Figure 1 and then sets its status to “try”, meaning that it is try to get in CS. It sets the variable my_quorum with Qk and send the message “(ask_permit, i)” to all nodes in the quorum. Description of [Line 7-6] in Figure 1; The candidate asking for a permission to proceed from every process of one quorum does not take steps until the all permissions are received from the quorum. But it eventually received all permissions from a quorum and enter the CS due to the assumption of installed dead lock resolving mechanism in the system and it get in CS. Var status:{rem, try, incs, wait} initially rem Var my_token : initially true Var my_token_holder : initially NULL Var token : initially empty list Var my_quorumi : initially empty Periodically(W) do request M* for HM 1. Upon received (trying, upper_ layer) 2. if not (status = try) then 3. wait until Qk : j Qk : j HM 4. statusi := try 5. send (ask_permit, i) to j Qk 6. my_quorum:= Qk 7. Upon received (ok_pemit, j) 8. token := token { j } 9. If my_quorum = token then enter CS 10.On Exit CS 11. send (return_permit, i) to j my_quorum 12. status:= rem
13.Upon received (no_permit, j ) 14. send (return_permit, i ) to i token 15. token : = 16. goto tryi 17.Upon received (ask_permit, j ) 18. wait until j HM 19. if my_token = true then 20. send (ok_permit, j ) 21. my_token_holder:= j 22. my_token: = false 23. else 24. send (no_permit, j ) 25.Upon received (return_permit, j ) 26. my_token := true 27. my_token_holder := NULL 28.Upon received HM from Mi 29. if (my_token=false and 30. my_token_holder HM ) then 31. my_token := true
Fig. 1. FTQME algorithm using M* : process i
Description of [11-12] in Figure 1; On exit from the CS, the node sends “return _permit” to the every member of the quorum from which it received permissions. It set its status with “rem” meaning that it is in normal mode. Notice that no candidate i can be served if other process j is accessing the resource. That is because while other process j is serving but not yet releasing the resource, the candidate i can not obtain all permission from the quorum (line 13 in Figure 1). Description of [13-16] in Figure 1; If the candidate received the message “no_permit” from a node of quorum, it returns all received permissions from the quorum to every member of the quorum and after that it try again.
32
S.-H. Park and S.-H. Lee
Description of [17-24] in Figure 1; The node i, received “ask_permit” from node j, first checks that j is alive and if it is alive then the node i sends its “ok_permit” to the node i when it has its token. But if the node i has no token, it send the message “no_permit” to the node j. Description of [25-27] in Figure 1; When the node i received the failure detector history HM from M*, if it knows that that a node holding its token died, it regenerates its token again. Description of [25-27] in Figure 1; Upon received “return_permit” from node j, the node i sets its “my_token” with true meaning that it has its token. Now we prove the correctness of the algorithm of Figure 1 in terms of two properties : mutual exclusion and progress . Let R be an arbitrary run of the algorithm for some failure pattern F E (f < n). Therefore we prove Lemma 1 and 1 for R respectively.
∈
Lemma 1. (mutual exclusion property) No two different processes are in their CSs at the same time. Proof: By contradiction, assume that i and j (i≠ j) are in their CSs at time t’. According to the line 7-9 of the algorithm 1, no process enters its CS before receiving permissions from a quorum. Thus i must have received all permissions from each member of a quorum and j must have received all permissions from each member of a quorum before t’. Without loss of generality, assume the event that i received all permission from a quorum precedes the event that j received all permission from other quorum. That is, at some time t’’ < t’, j received all permissions from a quorum while i is entering CS but before exits from CS. That means that at some time t’’ < t’, j passed the (my_quorum = token) clause in line 9 while i is still in CS. Thus, one of the following events occurred before t’’ at every member of a quorum: (1) Every member of quorum j has a token and sends (ok_permit, j ): by the algorithm of Figure 1. But by intersection property of quorum, i is in the CS at t’ > t’’ and at least one member of the quorum does not have a token: a contradiction. (2) Every member of quorum j received HM from Mj and i ∈ HM by the algorithm of Figure 1, at some time t’’ < t’. Thus, we can assume that the following is true: i ∉ HM at time t’ and i HM at time t’’. By the Modal accuracy property of M, i is crashed at t’’. But it is in the CS at t’ > t’’: a contradiction. Hence, mutual exclusion is guaranteed. □
∈
Lemma 2. If a correct process request for the CS, then at some time later the process eventually enters in its CS. Proof: Assume that a correct process i volunteers at time t’, and no correct process is ever in its CS after t’. According to the algorithm, after t’, process i never reaches line 9 of the algorithm. In other words, i is blocked at some wait clause. The first wait clause (line 3 in Figure 1) is not able to block the process, due to the modal accuracy (1) property of M* and the fact that (n > f) processes are correct. Thus, eventually, statusi = try, and wait clause in line 5-6 in Figure 1 cannot block the process neither.
Quorum Based Mutual Exclusion in Asynchronous Systems
33
Thus, i issues send (ask_permit, i). The second received clause (more precisely, the statement in line 17 in Figure 1) is not blocking neither, because of the guarantee that any send message is eventually delivered by every correct process. Thus, i is blocked in the third clause (line 7-8 in Figure 1) while processing some token := token ∪ { j }. We show that if a correct process i is blocked while processing some token from j, then process j is blocked and it never sends (ok_permit, j) nor (no_permit, j). But j is never blocked since it is always in one of two states, i.e., my_token is true or not. So contradiction. □ Theorem 1. The algorithm of Figure 1 solves FTQME using M*, in any environment E with f < n/2, combining with two lemmas 1 and 2.
6
Concluding Remarks
Is it beneficial in practice to use a mutual exclusion algorithm based on M*, instead of a traditional algorithm assuming P? The answer is “yes”. Indeed, if we translate the very fact of not trusting a correct process into a mistake, then M* clearly tolerates mistakes whereas P does not. More precisely, M* is allowed to make up to n2 mistakes (up to n mistakes for each module Mi, i Π). As a result, M*’s implementation has certain advantages comparing to P’s (given synchrony assumptions). For example, in a possible implementation of M*, every process i can gradually increase the timeout corresponding to a heart-beat message sent to a process j until a response from j is received. Thus, every such timeout can be flexibly adapted to the current network conditions. In contrast, P does not allow this kind of “finetuning” of timeout: there exists a maximal possible timeout, such that i starts suspecting j as soon as timeout exceeds. In order to minimize the probability of mistakes, it is normally chosen sufficiently large, and the choice is based on some a priori assumptions about current network conditions. This might exclude some remote sites from the group and violate the properties of the failure detector. Thus, we can implement M* in a more effective manner, and an algorithm that solves FTQME using M* exhibits a smaller probability to violate the requirements of the problem, than one using P, i.e., the use of M* provides more resilience.
∈
References 1. Delporte-Gallet, C., Fauconnier, H.: The weakest Failure Detector to Solve certain Fundamental Problems in Distributed computing. In: Proceedings of the ACM Symposium on Principles of Distributed Computing. ACM, New York (2004) 2. Agrawal, D., Abbadi, A.E.: An efficient and fault-tolerant solution for distributed mutual exclusion. ACM Transactions on Computer Systems 9(1), 1–20 (1991) 3. Chandra, T.D., Hadzilacos, V., Toueg, S.: The weakest failure detector for solving consensus. Journal of the ACM 43(4), 685–722 (1996) 4. Chandra, T.D., Toueg, S.: Unreliable failure detectors for reliable distributed systems. Journal of the ACM 43(2), 225–267 (1996)
34
S.-H. Park and S.-H. Lee
5. Chockler, G., Malkhi, D., Reiter, M.K.: Backo. protocols for distributed mutual exclusion and ordering. In: Proceedings of the 21st International Conference on Distributed Computing Systems, ICDCS 2001 (April 2001) 6. Dijkstra, E.W.: Solution of a problem in concurrent programming control. Communications of the ACM 8(9), 569 (1965) 7. Fischer, M.J., Lynch, N.A., Paterson, M.S.: Impossibility of distributed consensus with one faulty process. Journal of the ACM 32(3), 374–382 (1985) 8. Gafni, E., Mitzenmacher, M.: Analysis of timing-based mutual exclusion with random times. SIAM Journal on Computing 31(3), 816–837 (2001) 9. Hadzilacos, V.: A note on group mutual exclusion. In: 20th ACM SIGACTSIGOPS Symposium on Principles of Distributed Computing (August 2001) 10. Joung, Y.-J.: Asynchronous group mutual exclusion. In: 17th ACM SIGACT-SIGOPS Symposium on Principles of Distributed Computing, pp. 51–60 (June 1998) 11. Keane, P., Moir, M.: A simple local-spin group mutual exclusion algorithm. IEEE Transactions on Parallel and Distributed Systems 12(7), 673–685 (2001) 12. Lamport, L.: A new solution of Dijkstra’s concurrent programming problem. Communications of the ACM 17(8), 453–455 (1974) 13. Lamport, L.: The mutual exclusion problem. Parts I&II. Journal of the ACM 33(2), 313–348 (1986) 14. Lodha, S., Kshemkalyan, A.D.: A fair distributed mutual exclusion algorithm. IEEE Transactions on Parallel and Distributed Systems 11(6), 537–549 (2000) 15. Lynch, N.A.: Distributed Algorithms. Morgan Kaufmann Publishers (1996) 16. Maekawa, M.: A√N algorithm for mutual exclusion in decentralized systems. ACM Transactions on Computer Systems 3(2), 145–159 (1985) 17. Manivannan, D., Singhal, M.: An efficient fault-tolerant mutual exclusion algorithm for distributed systems. In: Proceedings of the ISCA International Conference on Parallel and Distributed Computing Systems, pp. 525–530 (October 1994)
Location Prediction for Grid-Based Geographical Routing in Vehicular Ad-Hoc Networks Si-Ho Cha and Keun-Wang Lee* Dept. of Multimedia Science, Chungwoon University San 29, Namjang-ri, Hongseong, Chungnam, 350-701, South Korea {shcha,kwlee}@chungwoon.ac.kr
Abstract. Inter-vehicle communication (IVC) is wirelessly connected using multi-hop communication without access to some fixed infrastructure. Because of the rapid movement of vehicles and the frequent topology change, link breakages occur repeatedly and the packet loss rate increases. Geographical routing protocols are known to be very suitable and useful for vehicular ad-hoc networks (VANETs). However, they just select the node nearest to the destination node as a relay node within its transmission range. This increases the possibility of a local maximum and link loss in IVC. This paper presents the location prediction method for the grid-based predictive geographical routing (GPGR) protocol to overcome the problems. GPGR makes use of map data to generate the road gird and to predict the exact moving position of vehicles in the relay vehicle selection process. It regards every vehicle moves only along the road grid. Simulation results using ns-2 show performance improvement in terms of packet delivery rate than existing routing protocols. Keywords: Vehicular Ad-hoc Networks, Inter-Vehicle Communication, Gridbased Routing, Geographical Routing.
1
Introduction
VANET provides both vehicle-to-infrastructure (V2I) communication and vehicle-tovehicle (V2V) communication. V2I can provide real-time the road traffic conditions, weather, and basic Internet service through the communication with backbone networks. V2V can be used for providing information about traffic conditions and/or vehicle accidents based on wireless inter-vehicle communication (IVC). In V2V communication environments, vehicles are wirelessly connected using multi-hop communication without access to some fixed infrastructure [1]. Already, automobile manufacturers and research centers are investigating the development of IVC protocols for the establishment of VANETs which have been envisioned to be useful in road safety and many commercial applications [2]. VANET has unique characteristics such as high node mobility and a rapidly changing network topology compared to MANET. Current routing protocols in VANETs usually use relay nodes to forward data packets to the destination. Because of the rapid movement of vehicles *
and the frequent topology change of vehicles, link breakages occur repeatedly. And the frequent link disconnection is also cause by the characteristics of VANET such as vehicle movements are constrained in roads and traffic lights have great influence on the vehicle movement [3]. The frequent link disconnection may increase the possibility of local maximum. Due to these problems, geographical routing protocols such as GPSR [4] are known to be more suitable and useful to VANET than existing routing protocols designed for MANETs. Geographical forwarding is one of the most suitable solutions for routing in VANET because it maintains only local information of neighbors instead of perdestination routing entries. GPSR selects the node that is the closest to the destination among the neighbor nodes as the relay node. However, GPSR may generate the link loss problem in urban environments. Because GPSR does not take into account road structure and the speed and moving direction of vehicles, it is able to select stale nodes as relay nodes. As vehicle movements are constrained by roads, GPSR without taking urban environment characteristics is not applicable for VANET [3]. To solve this problem, greedy perimeter coordinator routing (GPCR) [5] and greedy perimeter urban routing (GPUR) [6] are proposed as possible solutions. However, GPCR may cause transmission delay and path selection error because it identifies nodes that are on a junction by detecting coordinator nodes to select relay nodes. GPUR selects nodes that have 2-hop neighbors for relay nodes. It will cause serious transmission delay. GPUR does not resolve local maximum problem due not to considering road specifications such as dead end roads. This paper proposes GPGR, a grid-based predictive geographical routing protocol for IVC. The protocol uses map data to generate road grids on the path of the moving vehicles, and predicts the exact moving position along the road grids. To do this, we assume that each vehicle knows its location through GPS like most of the related geographic routing protocols, and has a grid-based street map for road information. The rest of the paper is organized as follows. Section 2 discusses the related work, and Section 3 introduces the proposed routing protocol for IVC. The performance evaluation is discussed in Section 4. Finally, in Section 5, conclusions are made including the future research.
2
Related Work
GPSR [3] is one famous greedy forwarding protocol. GPSR makes greedy forwarding decisions using only information about the immediate neighbors in the network topology. GPSR may increase the possibility of getting the local maximum and link breakage because of the high mobility of vehicles and the road specifics in urban areas. This is because it just selects the nearest node to the destination as a relay node within its transmission region to make packet forwarding decisions. GPSR may also generate the link loss problem because it maintains stale nodes as neighbor nodes to select a relay node in greedy mode. The local maximum and link breakage problems can be recovered in perimeter mode forwarding, but packet loss and delay time may appear because the number of hops is increased in perimeter mode forwarding. This decreases the reliability of a VANET. GPCR [5] was proposed to improve the reliability of GPSR in VANET. The basic behavior of GPCR is similar to GPSR, but it selects a relay node by considering information about the road structure. GPCR makes routing decisions on the basis of
Location Prediction for Grid-Based Geographical Routing in VANETs
37
streets and junctions instead of individual nodes and their connectivity. However, GPCR forwards data packets based on the node density of adjacent roads and the connectivity to the destination. Thus, if the density of nodes is low or there is no connectivity to the destination, then the delay time increase and the local maximum problem are still not resolved. GPUR [6] selects a relay node based on information about the road characteristics, like as GPCR. However, unlike GPCR, GPUR selects a relay node among nodes that have 2-hop neighbors. It makes periodic beacon messages to estimate the presence of 2-hop neighbors about all relay candidates. The periodic beacon messages used to evaluate the presence of 2-hop neighbors will cause serious transmission delay. GPUR does not also resolve local maximum problem due not to considering road specifications such as dead end roads.
3
Proposed Algorithm
Our target is to select the optimal relay vehicle for IVC based on vehicles' movement information such as position, direction, and velocity as well as road topology. To do this, we assume that each vehicle knows its location through GPS like most of the related geographic routing protocols, and has a digital street map for road information. In our algorithm, the geographic area of VANET is partitioned into a two dimensional logical grid. Given any physical location, there should be a predefined mapping from the location to its grid coordinate. In case of each vehicle has a radio range of r, each grid size d is determined by
d=
r
to represent the maximum value of d such that a
2 2
vehicle located at position of a grid is capable of transmitting data to any vehicle of its 8 neighboring grids. Our beacon messages contain the grid coordinates rather than the position of vehicles. Since vehicles know their own position (xi, yi), each vehicle can calculate its current grid coordinates G ( xi , y i ) by the floor function as shown in formula (1).
x y G ( xi , y i ) = i , i d d In our GPGR, a sender
(1)
VS should first look at all relay candidates similar to GPUR.
Then, among these relay candidates, it selects the node nearest to the destination as a relay node within its transmission range based on the future vehicle’s position using the road grid. VS would find all neighbor vehicles V N = {V j , Vk ,..., Vl } in its r, and calculate the distance
D(VS , V N ) to find a relay vehicle VR whose distance is
maximal based on the road grid. For the distance between two grids, GPGR calculates the Euclidean distance between two centers of the grid cells as formula (2).
38
S.-H. Cha and K.-W. Lee
G ( xi , y i )G ( x j , y j ) = ( xi − y i ) 2 + ( x j − y j ) 2 × d
(2)
Let there be N vehicles present in a road segment space. The location of each vehicle at time t is given by G ( xi (t ), y i (t )) , where i = 1 to N. The next location of each object at time
(t + Δt ) is predicted to be G ( xi (t + Δt ), y i (t + Δt )) .
G ( xi (t − Δt ), y i (t − Δt )) and current position
Therefore, from previous position
G ( xi (t ), y i (t )) , GPGR can predict the next position of the relay candidate at t + Δt as G ( xi (t + Δt ), y i (t + Δt )) using the velocity and direction of the relay candidate. The velocity and the direction of the relay candidate are given by formula (3) and formula (4), respectively. Therefore, GPGR can calculate the next position of the relay candidate by formula (5).
V=
( x i (t ) − x i (t − Δt )) 2 + ( y i (t ) − y i (t − Δt )) 2 t − (t − Δt )
y i (t ) − y i (t − Δt ) xi (t ) − xi (t − Δt )
θ = tan −1
(3)
(4)
G ( xi (t + Δt ), y i (t + Δt )) ( x (t ) + V × cos θ × Δt ) ( yi (t ) + V × sin θ × Δt ) = i , d d The grid sequence of road grid generated
according
to
(5)
G = {G ( x1 , y1 ), G ( x 2 , y 2 ),..., G ( x k , y k } is
the available map information, where are grids in the road path from G ( x1 , y1 ) to
G ( x 2 , y 2 ),..., G ( x k −1 , y k −1 ) G ( x k , y k ) . For example, the grid sequence of road to move the vehicles on the road is {G (4,3), G ( 4,4), G ( 4,5), G (3,5)} as illustrated in Fig. 4. Therefore, all nodes can move only along the grid sequence: {G (4,3), G ( 4,4), G ( 4,5), G (3,5)} . As shown in Fig. 1, the predicted position of the node will be (3,5) instead of G ( 4,5) at t + Δt . Therefore, the position of relay candidate can be predicted correctly because the node might be in the grid sequence. The position prediction based on road girds is more realistic if two reads are superposed with or without different running directions. If G ( xi (t + Δt ), y i (t + Δt )) , the predicted next position of the relay candidates, is not on the grid sequence of road grid, GPGR should select the closest grid to the predicted grid instead of it.
Location Prediction for Grid-Based Geographical Routing in VANETs
X
t+∆t
(3,5)
39
(4,5)
(5,5) t
(3,4)
(5,4)
(4,4) t-∆t
(3,3)
(4,3)
(5,3)
Fig. 1. By restricting the position prediction in the sequence of road grid, GPGR can predict the next position of nodes and select the optimal relay node
4
Performance Evaluation
We analyze and compare the performance of the proposed GPGR and the existing GPSR, GPCR and GPUR using the ns-2 simulator. In this performance evaluation, we just considered packet delivery rate according to the number and velocity of nodes at this point. The simulated area is based on a real map from Seoul with a 1000 m * 1000 m size. Table 1 summarizes our simulation parameters. The simulations were performed for 180 secs, and the number of nodes was increased from 100 to 200. The moving velocity of the nodes was increased from 0 km/h to 80 km/h. The experiments were performed three times and average values were used. Maximum and minimum values were excluded. Table 1. Simulation parameters Parameter Topology size Transmission range MAC protocol Node number Node velocity Beacon time Bandwidth Packet size
Value 1000m*1000m 125m IEEE 802.11 100 to 200 0km/h to 80km/h 1sec 2Mbps 1000bytes
40
S.-H. Cha and K.-W. Lee
Fig. 2 and Fig. 3 show the packet delivery rate according to the number of nodes and the packet delivery rate according to the velocity variation of the nodes, respectively. The packet delivery rate of GPGR is higher than those of GPSR, GPCR, and GPUR. As shown in Fig. 2, the performance of GPGR appears similar to that of GPUR. However, GPGR shows better performance than GPUR according to the increase of the number of nodes. This is because that GPUR considers only the location of 2-hop nodes in selecting relay nodes without considering road topology. On the other hand, GPGR can provide a high delivery rate by selecting relay nodes based on the grid sequences of roads. 100
Paket Delivery Rate (%)
90 80 70 60 50 40 30 GPSR GPCR GPUR GPGR
20 10
0 100 110 120 130 140 150 160 170 180 190 200 Number of Nodes
Fig. 2. Packet delivery rate according to number of nodes
Packet Delivery Rate (%)
100 GPSR GPCR GPUR GPGR
90 80 70 60 50 40 30 20 10 0 0
10
20
30 40 50 Velocity (km/h)
60
70
80
Fig. 3. Packet delivery rate according to velocity of nodes
In Fig. 3, all protocols show the packet delivery rate drops with increasing node speed. However, the decreasing rate of GPGR is lower than the other protocols. This is because that GPGR knows the accurate position of nodes, it can be expected that the delay time to determine the position of nodes is lower than that of the other protocols.
Location Prediction for Grid-Based Geographical Routing in VANETs
5
41
Conclusion
This paper has presented GPGR algorithm to select the optimal relay vehicle for IVC based on vehicles' movement information such as position, direction, and velocity as well as road topology. GPGR generates road grids on the path of the moving vehicles, and predicts the exact moving position along the road grids to reduce the link breakage problem occurred in VANETs. GPGR can reduce the possibility of link breakage by selecting the relay node based on road grid sequence. Simulation results showed that our protocol has a high packet delivery rate compared with GPSR, GPCR and GPUR for VANETs. As the future work, we will consider the probability of local maximum and the packet breakage rate in our simulation.
References 1. 2.
3.
4.
5.
6.
Tse, Q.: Improving Message Reception in VANETs. In: The International Conference on Mobile Systems, Applications, and Services (MobiSys), Krakow, Poland (2009) Chen, M.-H., Dow, C.-R., Chen, S.-C., Lee, Y.-S., Hwang, S.-F.: HarpiaGrid: a geography-aware grid-based routing protocol for vehicular ad hoc networks. Journal of Information Science and Engineering 26, 817–832 (2010) Luo, J., Gu, X., Zhao, T., Yan, W.: A Mobile Infrastructure Based VANET Routing Protocol in the Urban Environment. In: The International Conference on Communications and Mobile Computing (CMC), pp. 432–437 (2010) Karp, B., Kung, H.T.: GPSR: Greedy Perimeter Stateless Routing for Wireless Networks. In: The International Conference on Mobile Computing and Networking (MobiCom), pp. 243–254 (2000) Nzouonta, J., Rajgure, N., Wang, G.: VANET Routing on City Roads Using Real-Time Vehicular Traffic Information. IEEE Transactions on Vehicular Technology 58(7), 3609– 3626 (2009) Ryu, M.-W., Cha, S.-H., Cho, K.-H.: A Vehicle Communication Routing Algorithm Considering Road Characteristics and 2-Hop Neighbors in Urban Areas. The Journal of Korea Information and Communications Society, KICS 36(5), 464–470 (2011)
Hybrid Routing Algorithm Considering Moving Direction of Vehicles in VANETs Min-Woo Ryu1, Si-Ho Cha2,*, Keun-Wang Lee2, and Kuk-Hyun Cho1 1
Dept. of Computer Science, Kwangwoon University 447-1, Wolgye-dong, Nowon-gu, Seoul, 139-701, South Korea {minu0921,chokh}@kw.ac.kr 2 Dept. of Multimedia Science, Chungwoon University San 29, Namjang-ri, Hongseong, Chungnam, 350-701, South Korea {shcha,kwlee}@chungwoon.ac.kr
Abstract. VANET (Vehicle Ad hoc Network) has attracted focus as nextgeneration networks to provide traffic information, flow control of vehicles, and safety messages. Unlike the existing networks, it is very important to predict the moving direction of vehicles because of the rapid movement of vehicles and the frequent change of topology in VANET. To do this, many researches have been performed to find new methods to predict the speed and direction of moving vehicles. However, it is very difficult to predict because the moving direction of vehicles is decided depending on the road types, traffic signals and vehicle's destination path. Therefore, this paper proposes the hybrid routing algorithm to solve this problem. The proposed algorithm ensures the reliability of data transmission throughout different routing methods in the straight roads and intersections. Simulation results using ns-2 revealed that the proposed algorithm performs much better than the existing algorithm for VANET. Keywords: VANET, Hybrid Routing, Prediction, Reliability, Inter-Vehicle Communication.
1
Introduction
Due to the development of ITS (Intelligent Transport System) [1], attention has been focused on vehicle and driver safety. VANET is a next-generation network technology to support communication between vehicles and between vehicles and infrastructure [2] [3]. It is a special kind of MANET (Mobile Ad-hoc Network). Unlike MANET, in VANET, it is very important to predict the moving direction and position of vehicles because the rapid movement of vehicles and the frequent change of topology [4]. To do this, many researches have been performed to find new methods to predict the speed and direction of moving vehicles. However, it is very difficult to predict because the moving direction of vehicles is decided depending on the road types, traffic signals and vehicle's destination path. If it is possible to move *
Hybrid Routing Algorithm Considering Moving Direction of Vehicles in VANETs
43
vehicle multiple direction such as intersections, the prediction of vehicle's location is more difficult. To resolve this problem, many algorithms [5] [6] are proposed that use GPS location data to select the next relay vehicle. However, these algorithms are still not solving the accurate prediction of the movement direction of vehicles. In this paper, we propose a hybrid routing algorithm to solve these problems. The proposed algorithm provides faster data transfer to the destination node through the different method from the straight roads and the intersections. The proposed algorithm also minimizes the prediction problem with the wrong direction of moving nodes in the intersections. The rest of the paper is organized as follows. Section 2 discusses the related work, and Section 3 introduces the proposed hybrid routing protocol for VANET. The performance evaluation is discussed in Section 4. Finally, in Section 5, conclusions are made including the future research.
2
Related Work
In this chapter, we examine and analyze the existing routing algorithms for VANET and describe the need for more advanced routing algorithms. GPSR (Greedy Perimeter Stateless Routing) [7] is a typical greedy forwarding protocol for VANET. GPSR uses greedy forwarding to forward packets to nodes that are always progressively closer to the destination. GPSR can generate the link loss problem because it is able to maintain stale nodes as neighbor nodes for next forwarding nodes in data transmission time. It can also face the local maximum problem due to changes in the direction of moving nodes. DGRP (Directional Greedy Routing Protocol) [8] detects the location of nodes by using GPS, and predicts the moved location of nodes by the moving speed. Therefore, DGRP can resolve the link breakage problem that resulted from maintaining stale nodes in its neighbor list. However, DGPR did not resolve the local maximum problem because its basic routing strategy is similar to GPSR. GPCR (Greedy Perimeter Coordinator Routing) [9] was proposed to improve the reliability of GPSR in VANET. The basic behavior of GPCR is similar to GPSR, but it selects a relay node by considering information about the road structure. GPCR makes routing decisions on the basis of streets and junctions instead of individual nodes and their connectivity. However, GPCR forwards data packets based on the node density of adjacent roads and the connectivity to the destination. Thus, if the density of nodes is low or there is no connectivity to the destination, then the delay time increase and the local maximum problem are still not resolved. GPRR (Greedy Perimeter Reliable Routing) [10] uses the 2-hop neighbor’s information to select the next relay node in greedy forwarding method so that it can solve the local maximum problem. However, GPRR increases time delay from the use of 2-hop neighbor's information. In this paper, we suggests a hybrid routing algorithm to resolve the link disconnection problem due to increased velocity of nodes, as well as the local maximum problem due to the direction of node movement.
44
3
M.-W. Ryu et al.
Hybrid Routing Algorithm (HRA)
Our hybrid routing algorithm (HRA) is to solve the problems occurred from the inaccurate prediction about the direction of moving nodes in existing VANET routing algorithms. HRA can minimize the data loss as well as time delay by using different forwarding methods depending on the road types. In HRA, if the node is on the straight road, it transfers data to the destination node by using the fast greedy mode. HRA also broadcasts data in case of difficult to predict the direction of moving nodes as intersections. Table 1 lists the symbols used in the proposed HRA. Table 1. Symbols used for defining HRA Symbol
Define
Si D N iS
Sender
S i
Set of neighbors of a sender
n
A neighbor of a sender
ΔV ( xiS )
Relative velocity on neighbors of a sender
niS pos RN MV (niS ) 3.1
Destination node
Position of
niS
Next relay node Node’s moving direction
Hybrid Routing Algorithm for Straight Roads (HRAS)
The proposed HRA in straight roads is similar to the existing greedy forwarding. However, if the sender can't transfer data using greedy forwarding, HRA selects the relay node through a comparison of the relative velocity of sender and neighbors. For example, as shown in Fig. 1, if there isn't any node that is near to the destination node than the sender S, S compares its velocity with neighbors and selects a neighbor with higher velocity than S as a relay node. This is because the packet loss may be occur through multi-hop transfer when it does not satisfy the greedy mode like the general greedy forwarding and if it do not find the node closer to destination, packet forwarding will be infinitely repeated. The proposed algorithm for straight roads is shown in Table 2.
Fig. 1. Routing in a straight road
Hybrid Routing Algorithm Considering Moving Direction of Vehicles in VANETs
45
Table 2. HRA for straight roads 1. if
( S i send Data to D ) then
2.
Si
request
3.
if
( Si , D)
4.
if
8.
<
ΔV (nis )
(niS , D) >
), where
N iS = {niS , n Sj , nkS ,...}
then
ΔV ( S i )
then
RN ← nis
5. 6. 7.
N iS ( pos , ΔV
end if else
RN ← nis
9. end if 10. end if
3.2
Hybrid Routing Algorithm for Intersections (HRAI)
The prediction of the moving speed and direction of vehicles in VANETs are the major challenges that must be addressed. The moving direction of vehicles can have a significant impact on the number of hops and on the routing path to the destination node. However, the moving direction of vehicles is very difficult to predict when the vehicles are on intersections because it changes according to the form of the road and the path to the destination node. Therefore, the proposed routing algorithm uses the broadcast method to resolve these problems on intersections. Where, if the node that is received data is moving in the same direction to the destination node, data is transferred using HRLS. However, if they do not have the same direction, the node discards the received data to avoid duplication of data.
Fig. 2. Routing in a intersection
46
M.-W. Ryu et al.
For example, as shown in Fig. 2, the sender S broadcasts data to its neighbors because it is on the intersection. In this case, node A, B, and C all discard all received data because their moving direction is different with the destination node. But node E should transfer the received data to the next node because its direction is same with the destination node. The proposed algorithm for intersections is shown in Table 3. Table 3. HRA for intersections 1. if
S i ( pos) == Intersection then
2.
Si
broadcasts to
3.
if
N iS
N iS , where N iS = {niS , n Sj , nkS ,...}
Receive Data then
4.
Compare
5.
if
MV (niS , D )
MV (niS ≠ D )
then
6. Drop Data 7. else if 8. do HRAS mode 9. end if 10. end if 11. end if
4
Performance Evaluation
We analyze and compare the performance of the proposed HRA and the existing GPBR and DGPR using the ns-2 simulator. In this performance evaluation, we just considered packet delivery rate and link breakage rate according to the velocity of nodes. The simulations were performed for 180 secs, and the number of nodes was increased from 10 to 100. The maximum packet size is 1000 bytes. The moving velocity of the nodes was increased from 20 km/h to 100 km/h. The experiments were performed three times. Maximum and minimum values were excluded and average values were used. Table 4 summarizes our simulation parameters. Table 4. Simulation parameters Parameter Topology size Transmission range MAC protocol Node number Node velocity Traffic type Bandwidth Packet size
Hybrid Routing Algorithm Considering Moving Direction of Vehicles in VANETs
47
Fig. 3 shows the packet delivery rate according to the velocity of nodes. In Fig. 3, DGRP and GPBR present the reduction in the packet delivery rate according to the increase of nodes velocity. However, HRA showed no significant decrease in the packet delivery rate, even if the velocity of the next node is high. It is because HRA selects the relay node using both the relative velocity of neighbors and the distance to the destination node. We are also guess that HRA minimized the data loss as well as time delay by using different forwarding methods depending on the road types.
Packet Delivery Rate (%)
100 GPBR DGPR HRA
90 80 70 60 50 40 30 20 10 20
40
60
80
100
Velocity (km/h)
Fig. 3. Packet delivery rate according to the velocity of nodes 100 GPBR DGPR HRA
Link Breakage (%)
90 80 70 60 50 40 30 20 10 20
40
60 Velocity (km/h)
80
100
Fig. 4. Link breakage rate according to the velocity of nodes
Fig. 4 shows the link breakage rate according to the velocity of nodes. In Fig. 3, HRA presents the lower link breakage rate than GPBR and DGRP. It is because HRA broadcasts data on intersections to ensure the link connectivity. We are also guess that
48
M.-W. Ryu et al.
if the node that is received data is moving in the same direction with the destination node, HRA transfers data through HRLS.
5
Conclusion
This paper proposes the hybrid routing algorithm, HRA, to ensure the reliability of data transmission throughout different routing methods in the straight roads and intersections. HRA can reduce the possibility of link breakage, and can increase the packet delivery rate because it transfers data to the destination node by using the fast greedy mode on straight road, and also broadcasts data on intersections. Simulation results showed that HRA has a high packet delivery rate and a low link breakage rate compared with GPBR and DGRP for VANETs. As the future work, we will consider the queuing problem in broadcasting on intersections and low node density environments.
References 1. ESTI: Intelligent Transport Systems, http://www.etsi.org/WebSite/technologies/ IntelligentTransportSystems.aspx 2. Wireless Access for Vehicular Environment, http://www.standards.its.dot.gov/fact_sheet.asp 3. Blum, J.J., Eskandarian, A., Hoffman, L.J.: Challenges of Inter-Vehicle Ad Hoc Networks. IEEE Transactions on Intelligent Transportation Systems 5(4), 347–351 (2004) 4. Ryu, M.-W., Cha, S.-H., Cho, K.-H.: A Routing Prediction Algorithm for Increasing Reliability in VANET. In: International Conference on Information Networking (2010) 5. Wang, Y.-B., Wu, T.-Y., Lee, W.-T., Ke, C.-H.: A Novel Geographic Routing Strategy over VANET. In: IEEE 24th International Conference (2010) 6. Menouar, H., Lenardi, M., Filali, F.: Movement Prediction-Based Routing (MOPR) Concept for Position-Based Routing in Vehicular Networks. In: IEEE Vehicular Technology Conference (2007) 7. GPSR, http://www.icir.org/bkarp/gpsr/gpsr.html 8. Kumar, R., Rao, S.V.: Directional Greedy Routing Protocol (DGRP) in Mobile Ad-Hoc Networks. In: IEEE Information Techology (2008) 9. Nzouonta, J., Rajgure, N., Wang, G.: VANET Routing on City Roads Using Real-Time Vehicular Traffic Information. IEEE Transactions on Vehicular Technology 58(7), 3609– 3626 (2009) 10. Ryu, M.-W., Cha, S.-H., Cho, K.-H.: An Enhanced Greedy Message Forwarding Protocol for Increasing Reliability of Mobile Inter-Vehicle Communication. The Journal of The Institute of Electronics Engineers of Korea 47-TC(4), 43–50 (2010)
Development of an Active Ranging System Based on Structured Light Image for Mobile Robot Application* Jin Shin and Soo-Yeong Yi** Seoul National University of Science and Technology, Korea [email protected], [email protected]
Abstract. This paper presents an active ranging system that is based on a laser structured-light image and is robust against environment illumination noise. Structured-light image processing is computationally more efficient than conventional stereo image processing since the burdensome correspondence problem is avoidable. In order to achieve robustness against illumination noise, we propose an efficient image processing algorithm, i.e., integration of sequential time-difference images with structured-light modulation. The proposed algorithm involves only elementary operations such as integer subtraction and addition and is appropriate for hardware implementation. For this algorithm, a specialized FPGA image processor is developed. A distance equation and sensitivity analysis are also presented. Keywords: ranging system, structured light, integration of time-difference images, distance equation, sensitivity analysis.
1
Introduction
A ranging system is essential for many engineering applications such as factory automation, intelligent buildings, and security. In recent years, ranging systems are being employed for mobile robots not only to detect obstacles but also to achieve selflocalization and autonomous navigation by constructing a spatial map of a robot’s environment. There are several methods of distance measurement [1]: stereo imaging, image blurring, time-of-flight (TOF), and structured-light (SL) imaging. A stereo vision system has an advantage in that it is able to obtain a rich set of distance information, but requires extensive computation to solve the correspondence problem and thus has a low refresh rate in the case of distance measurement [2]. Moreover, it is expensive to build and sensitive to illumination noise. The image blurring method uses only one camera, but requires a large computation time and has low accuracy [3]. The TOF methods have different performance characteristics depending on the signal type and signal *
processors. An exemplar TOF method using ultrasonic sensors is economical and has simple signal processing modules, but suffers from low accuracy [4][5]. TOF methods based on laser scanners have high accuracy and speed [6], but the precision electronics and mechanical systems used are costly and not very durable. On the other hand, SL imaging systems project light of a distinct frequency in a particularly structured pattern onto the environment and compute the distance on the basis of the distortion of the SL pattern, which is a function of the distance to objects. The SL imaging system avoids the computationally intensive correspondence problem of a conventional stereo vision system. Results are available for a number of studies conducted using this method [2][7]. As application examples of SL imaging systems, A. Escalera et al. applied an SL-based ranging system for mobile robot localization in [8], and similar application had been reported in [9]. The bulky laser and camera equipments and the image processing time have discouraged the use of SL imaging in the past, but recent advancements in semiconductor laser equipments and faster processors have made this system more viable and economically feasible. This paper presents a compact distance measurement system that is based on images acquired using laser SL. Recently, many researches have been reported on the SL imaging system for 3D shape reconstruction [10][11][12][13]. However, under the assumption that a mobile robot moves on a planar surface, 2D horizontal depth measurement is sufficient for navigational purposes, but not for the purposes of recognition. One of disadvantages of the SL imaging system is its vulnerability to background illumination. In order to overcome this shortcoming, we have developed an algorithm that can robustly extract SL pattern from the source image; the algorithm performs integration of time-difference images with the modulated SL projection. Since this image processing algorithm involves elementary operations such as integer subtraction and addition, it is easy to implement in hardware. Unlike ultrasonic or position sensitive device (PSD)-based systems that measure distance in one direction at a time, the ranging system based on SL imaging can simultaneously measure the distance to all points in wide direction corresponding to the FOV of the camera, and thus, it provides a fast and efficient solution to map building and autonomous navigation of mobile robots.
2 2.1
Structured-Light Imaging System System Overview
The distance measurement system based on SL imaging is shown in Fig. 1. This SL imaging system consists of a camera and an SL projection module; they have vertical displacement. In mobile robot applications where the horizontal distance is more important than the vertical distance, SL sheet should be projected horizontally at a constant height so as to obtain the horizontal distance to obstacles. A laser diode in the SL module generates a point light that is distinct from the ambient lighting. For the purpose of converting the point light into lines, a
Development of an Active Ranging System Based on SL Image
51
scanning mirror or cylindrical lens is commonly used [7]. A cylindrical lens is preferable to a scanning mirror for compact and efficient implementation since it eliminates the need to use motorized mechanical parts and can spread light stripes over a wider area, thereby speeding up the process of distance measurement. However, it lowers the light energy density by spreading the laser light; as a result, strong illumination noise can interfere with the extraction of the SL pattern from the image. Therefore, a special image processing method is required to improve the extraction of the SL pattern. In images acquired using SL, the light pattern is distorted according to the distance of the reflecting object surfaces, and the distance can be computed from the distortion pattern and the extent of distortion.
Fig. 1. Distance measurement system based on structured-light imaging
Fig. 2. Image difference for extraction of SL pattern
52
2.2
J. Shin and S.-Y. Yi
Integration of Time-Difference Images with SL Modulation
An easy method to remove the background image and extract only the SL pattern is to use an optical bandpass filter that only transmits the frequency of the SL; however, the projected SL is also attenuated through the filter. Moreover, when ambient light contains the same frequency component with the SL, the resultant image can be contaminated. Another method to extract an SL pattern is to obtain the difference between two images by using SL modulation: one image is acquired with the SL on and the other without the SL, as illustrated in Fig. 2. This is well-known technique to detect a motion in a static background. A laser modulator is needed to use the image difference method. As mentioned before, when a point laser is converted into a line by a cylindrical lens, energy density of SL is decreased. In order to enhance the weakened SL and make it resistant to the background illumination, sequential integration of the timedifference images is proposed in this paper. This image processing selectively emphasizes the SL pattern in the image. Its effectiveness is demonstrated through experiments in Section 5. The integration of time-difference image requires sequential image frames and it might undergo deterioration because of the afterimage effect in the case of nonstatic background. Thus, in order to mitigate this effect, we used a high-speed camera that can capture images at 200 FPS (frames per second) and developed a dedicated FPGA image processor for the camera. 200 FPS image acquisition rate is much faster than 30 FPS of standard camera.
3
Distance Computation and Error Analysis
Distance computation is described in Fig. 3, where the symbols denote the following quantities:
l : measured distance, b : displacement, p : pixel distance,
θ : camera view angle ρ : measured angle λ : camera focal length
Fig. 3. Camera model and distance measurement based on structured-light imaging
Fig. 3 shows the side view, i.e., a vertical cross-section of the system setup. The distance to the point that the SL is projected can be computed as
Development of an Active Ranging System Based on SL Image
p l = b cot θ − tan −1 λ
53
(1)
Note that the distance depends on the displacement between the camera and the SL source. The image sensor is a two-dimensional regular grid of discrete photo sensors, and therefore, the quantization of pixel distance can introduce an error in the measured distance. The error sensitivity of the measured distance to the pixel quantization error is given by b ∂l = ⋅ ∂p λ
1 1 ⋅ 2 p p sin 2 θ − tan − 1 1 + λ λ
(a) Distance
(2)
(b) Sensitivity according to pixel error
Fig. 4. Distance and sensitivity
Fig. 4 shows the plot of distance (1) and error sensitivity (2) as a function of the measured pixel distance on the image plane. The parameters in the two equations assume the values listed in Table 1 in Section 5. Pixel 0 denotes the center pixel on the image plane. Fig. 4 (a) shows that a pixel distance of about 230 pixels corresponds to a distance of about 2.9 m. Fig. 4 (b) shows that the sensitivity is 20 mm at the pixel distance of 230 pixels, which implies that 1 pixel error in that region of the image causes a 20-mm error in the computed distance. To sum up, our system has a less than 20-mm error when the distance is less than 3 m. If the displacement between the camera and the SL source is increased by a factor of two, the measurable distance and the error limit will be doubled as implied by (1).
4
Calibration of Distance Measurement
To compute the distance from the measured pixel data, the required parameter values are the displacement, b , the camera view angle, θ , and the camera focal length, λ .
54
J. Shin and S.-Y. Yi
These parameters can be estimated by minimizing the least-squared errors along with the calibration data of points of known distances. The least square of error terms is defined in (3): p 1 N E = ⋅ b cot θ − tan −1 n − ln 2 n =1 λ
2
(3)
Here, pn denotes the measured pixel data and ln represents the known actual distance. The desired parameter values are the ones that minimize the sum of the squared errors and can be obtained by the gradient optimization technique shown in (4). ∂E ∂E ∂E Ek +1 = Ek − Δ e ⋅ + + ∂b ∂θ ∂λ
(4)
∂E ak +1 = ak − Δ a ⋅ , k = 1, 2, ... ∂a
Here, a represents the parameters b , θ , and λ to be estimated, and k denotes the iteration index. The symbols Δ e and Δ a indicate the step sizes of the squared errors and the parameters during iteration. The derivatives of the squared errors with respect to the parameters are given as follows: p p ∂E N = b cot θ − tan −1 n − ln ⋅ cot θ − tan −1 n λ λ ∂b n=1 N p ∂E b = − b cot θ − tan −1 n − l n ⋅ ∂θ λ n =1 sin 2 θ − tan − 1 p n λ
N ∂E p = −b cot θ − tan−1 n − ln ⋅ ∂λ λ n=1
5
(5)
b p ⋅ 2 n 2 p p +λ sin2 θ − tan−1 n n λ
Experimental Results
The SL-based distance measurement system developed in this study is shown in Fig. 5. The dedicated FPGA image processor performs all the image processing and sends only the data on the measured pixel distance to the host processor through USB. In this manner, the computational burden on the host processor is reduced. As stated before, we used a high-speed camera with a 640×480 image resolution and a 200-FPS acquisition rate; time interval between two image frames is 5ms. Therefore, the resultant distance measurement frequency is 10 Hz as the number of integrations of time-difference images is 10 using 20 image frames with SL modulation. For generating the SL, we used a red semi-conductor laser with a wavelength of 660 nm and a power of 5 mw, a cylindrical lens that spreads the point laser into a line with a 90o width angle, and a laser on-off modulator that is synchronized with the camera image acquisition process.
Development of an Active Ranging System Based on SL Image
55
Fig. 5. SL-based distance measurement system
In order to verify the robustness of the proposed system against illumination noise, we carried out the following experiment; Fig. 6 (a) shows the setup where the lines in bold ellipsoids represent the SL, and the bright light in the dotted circle is the flashlight applied to the SL as an intentional illumination noise. At the center of flashlight, illumination intensity is about 1100 lux and that of ambient light is about 400 lux. The original image acquired by the proposed system with the SL ON is shown in Fig. 6 (b). It is difficult to extract the SL pattern from this image since the illumination noise is stronger than the SL. Fig. 6 (c) shows that the time-difference with the SL modulation removes the background flashlight in the image. However, a part of the SL pattern is still too weak to be extracted from this image. It is noted that the strength of the SL pattern depends on the object distance and the reflectivity of the object material. As shown in Fig. 6 (d), in accordance with the integration of the sequential time-difference images, the SL pattern becomes noticeably clear enough to be extracted from the image. This result demonstrates the robustness of the proposed system against the environmental illumination noise. In order to calibrate the distance measurement system, we set up the following experiment. An object is placed horizontally and its pixel distance data are obtained at every 2.5 cm intervals within a distance of 0.5 m ~ 3.0 m. Thus, a set of total 100 pixel distance data along with known actual distance data is prepared for parameter calibration. Since the resolution of the camera sensors is 640×480, the SL pixel distance range on the image plane is [–240, +240] with respect to the center pixel.
(a) Experiment setup
(b) Original image
Fig. 6. Results of experiment to verify the robustness of the proposed system against illumination noise
56
J. Shin and S.-Y. Yi
(d) 10th integration of difference images
(c) Difference image
Fig. 6. (Continued)
Table 1 lists the final parameter values for the distance equation after calibration as in (4)~(5). Table 1. Parameter values after calibration
Parameters
b ( mm)
θ (°)
λ ( mm)
d ( mm)
Convergent value
375.04
21.3
957.0
186.9
From the measured pixel distance and the calibrated system parameters, we can compute the distance data using (1). Fig. 7 (a) shows the comparison between the actual distances and the distances computed from the measurement; Fig. 7 (b) presents the corresponding distance errors. As shown in these figures, when the measured pixel distance is within 230 pixels, the approximate measurable distance range and the error bound are ~3 m and 10 mm, respectively. As explained in Fig. 4, when the measured pixel distance exceeds 230 pixels, the sensitivity of distance to pixel error is over 20 mm. We thus limit the range of high-confidence distance measurement to [0, 3] meters. It is possible to set the measurable distance by the displacement and the camera view angle with reduced accuracy, as stated before.
(a) Distance
(b) Error
Fig. 7. Result of distance measurement
Development of an Active Ranging System Based on SL Image
6
57
Conclusions
In the SL-based method of distance measurement, the distances to object surface points are determined from the distortion of structured laser patterns projected thereon. This method does not pose the classical correspondence problem and is thus computationally more efficient, and can obtain distances over large number of points in the camera’s FOV in a single image. Recently, low-cost semiconductor lasers and image sensors have enabled a cost-efficient implementation of this method. In this paper, a distance-measurement module is developed on the basis of the SL imaging. In order to compensate for the low energy density of the SL stripe spread by a cylindrical lens, an image processing algorithm based on the difference of modulated SL images and their sequential integration is developed. In addition, we implemented a compact FPGA image processor for a high-speed camera in order to accelerate the integration of time-difference image processing algorithm and thereby reduce the computational burden on the host computer. It has markedly improved the performance of the SL pattern extraction and made the extraction robust to background illumination noise, as verified through our experiments. It is expected that our system will be preferred in some applications such as factory automation or mobile robots over ultrasonic sensor or laser scanner that measure distance to a single point at a time. This is because our system simultaneously obtains the distances to a wide area of points in the camera FOV, and such efficiency is crucial for map building and autonomous navigation of mobile robots.
References 1. Oike, Y.: Smart Image Sensors and Associative Engines for Three Dimensional Image Capture. Ph.D. Thesis, Univ. of Tokyo (2004) 2. Jain, R., et al.: Machine vision. McGraw-Hill (1995) 3. Lewis, J., Maler, L.: Blurring of the senses: common cues for distance perception in diverse sensory systems. Neuroscience 114(1), 19–22 (2002) 4. Leonard, J., Durrant-Whyte, H.: Directed sonar sensing for mobile robot navigation. Kluwer Academic Publisher (1992) 5. Cameron, S., Probert, P.: Advanced guided vehicles-aspects of the oxford AGV project. World Scientific (1994) 6. http://www.sick.com 7. Besl, P.: Active, Optical Range Imaging Sensors. Machine Vision and Applications 1(2), 127–152 (1988) 8. Escalera, A., Moreno, L., Salichs, M., Armingol, J.: Continuous mobile robot localization by using structured light and a geometric map. Int’l J. of Systems Science. 27(8), 771–782 (1996) 9. Noh, D., Kim, G., Lee, B.: A study on the relative localization algorithm for mobile robots using a structured light technique. J. of Institute of Control, Robotics and Systems. 11(8), 678–687 (2005) (in Korean) 10. Kim, C., Park, J., Yi, J., Turk, M.: Structured light based depth edge detection for object shape recovery. In: Proc. of 2005 IEEE Conf. on Computer Cision and Pattern Recognition (CVPR 2005) (2005)
58
J. Shin and S.-Y. Yi
11. Koninckx, T., Gool, L.: Real-time range acquisition by adaptive structured light. IEEE Tr. on Pattern Analysis and Machine Intelligence 28(3), 432–445 (2006) 12. Scharstein, D., Szeliski, R.: High Accuracy Stereo Depth Maps Using Structured Light. In: Proc. of 2003 IEEE Conf. on Computer Cision and Pattern Recognition (CVPR 2003), vol. 1, pp. 195–202 (2003) 13. Narasimhan, S., Nayar, S.: Structured Light Methods for Underwater Imaging: Light Stripe Scanning and Photometric Stereo. In: Proc. of 2005 MTS/IEEE OCEANS, pp. 2610–2617 (2005)
A Study on Fault Analysis of Wind Power (DFIG) in Distribution Systems Based on the PSCAD/EMTDC Jintaek Jeon, Joonho Son, Munseok Lee, Munkbaht, and Dae-Seok Rho Korea University of Technology and Education {dsrho,jinfind}@kut.ac.kr
Abstract. Korean government (Ministry of Knowledge Economy) has estimated that the rate of wind power (WP) in the new energy sources will be occupied by 37% in 2020 and 42% in 2030, and also green energies such as photovoltaic (PV) and WP are expected to be interconnected with the distribution system according to the new national regulation of Renewable Portfolio Standard (RPS) starting from 2012. However, when a large scale wind power plant is connected to the traditional distribution system, the protective devices (mainly OCR and OCGR of re-closer) can cause the mal-function problems due to change of fault currents. Under these field conditions, this paper deals with analysis of faults current by using both the symmetrical components of fault analysis and PSCAD/EMTDC modeling and proposes the mechanism of fault current considering the WP. Keywords: Protective devices, Re-closer, Photovoltaic (PV), Protective Coordination, Symmetrical components, Distributed Generation (DG), Over Current Ground Relay (OCGR), Malfunction, Doubly-fed Induction Generator (DFIG), a single line ground fault (SLG), Over Current Relay (OCR).
Table 1 is the %impedance data from the data of substation, primary feeder and interconnection transformer locating at the interconnecting point between primary feeder and WP. Based on these data, this paper performs the modeling the distribution system with DFIG by using PSCAD/EMTDC[5]-[7] as shown in Fig 1. Table 1. % Impedance Data of "BUKP-YEONG" Substation % Impedance of Transformer Bank
Fig. 1. Modeling distribution system with DFIG by using PSCAD/EMTDC
A Study on Fault Analysis of Wind Power (DFIG) in Distribution Systems
3
Simulation and Analysis
3.1
Model System and Simulation Conditions
61
Fig. 2 shows the model system for distribution system with WP, which is expressed by 3 points. The first point of is assumed at 2[km] from substation, the second
① point of ② at 5[km] and ③ at 10[km]. As the simulation conditions of the
PSCAD/EMTDC, a single grounded fault occurs at 1.0s and the fault duration time is 0.2s, because fault current can be obtained as the effective values after a 3 cycle.
Fig. 2. Model distribution system
3.2
Analysis for Ground Fault Current Considering DFIG
③
At the point, the equivalent circuit of both zero phase sequence impedance and positive/negative sequence impedance are described as Fig. 3. The total of positive phase sequence % impedance (Z1) is calculated by 148.56[%]. And also zero phase sequence %impedance (Z0) is 370.67[%].
(a) Zero phase sequence
(b) Positive phase sequence
Fig. 3. % impedance of equivalent circuit
Thus, the fault current of three-phase short circuit(I3s), line to line short circuit(I2s) and the SLG current (Ig) at the primary feeder with DFIG can be obtained by the conventional symmetrical component method as follows.
62
J. Jeon et al.
(1)
(2)
(3)
To analyze the characteristics of fault current of distribution with DFIG of WP, this paper considers 3 cases as follows : (1) Distribution system without WP (2) Distribution system with only interconnection transformer (3) Distribution system with both interconnection transformer and WP As shown in Table 2, it is verified that the simulation result by EMTDC/PSCAD is almost equal to the value of the conventional method(CM) like Eq. (1), (2) and (3), and because % impedance of interconnection transformer and DFIG is parallel connection to the distribution system, so wind power system can cause increasing fault current. Namely, it is confirmed that if large scale wind power plant (over 3[MW]) is connected to the distribution system, fault current can be increased, and make OCR and OCGR of protection devices cause the mal-function problem. Table 2. Simulation cases
Kind of faults
Distribution line
Distribution line + Wyegrounded/Delta winding of interconnection transformer
Distribution line + Wyegrounded/Delta winding of interconnection transformer +1 [MW] wind power
CM[A]
EMTDC[A]
CM[A]
EMTDC[A]
CM[A]
EMTDC[A]
three-phase short
1449.07
1444.69 1444.68 1444.67
1449.07
1444.69 1444.68 1444.65
1697.08
1650.12 1651.34 1654.23
line to line short
1254.89
1286.54 1280.35 1280.35
1254.89
1290.1 1295.5 1292.2
1469.67
1441.82 1441.91 1442.24
single line ground fault
812.08
819.91 819.91 819.91
1052.52
1102.81 1102.54 1102.80
1132.62
1180.42 1181.12 1182.24
A Study on Fault Analysis of Wind Power (DFIG) in Distribution Systems
3.3
63
Analysis for a Single Line Fault Current
If SLG is occurred at the primary feeder with DFIG in Fig. 4, the simulation results for SLG current at the 4 points of I_CB4, I_DFIG, I_Fault, I_IDGN are obtained as shown in Fig. 5.
Fig. 4. Collection points of a single line ground
(a) A single line ground faults current(I_CB4) (b) A single line ground faults current (I_DFIG)
(c) A single line ground faults current(I_Fualt) (d) A single line ground faults current(I_IDGN) Fig. 5. Analysis for SLG current of 4 points
64
J. Jeon et al.
In order to do better understanding the above simulation results, this paper proposes the analysis diagram for fault current mechanism. As shown in Fig. 6, the SLG fault current flow is analyzed as follows : (a) Ia_CB4 - Ib_CB4 - Ic_CB4 = 973.09[A]-209.79[A]-162.82[A] = 604.48[A] (b) 604.48[A] ≒ I_IDGN(614.48[A]) (c) la_CB ≒ Ib_CB4 + Ic_CB4 + I_IDGN (d) Ia_CB4 + Ia_DFIG = 973.09[A] + 209.79[A] = 1182.88[A] (e) 1182.88[A] ≒ Ia_Fualt(1182.39[A]) (f) la_Fualt ≒ Ia_CB4 + Ia_DFIG
Fig. 6. Analysis diagram for SLG fault current mechanism
On the other hand, it is confirmed that if ground fault occurs in primary feeder interconnected with 1[MW] WP having Y-grounded/Delta winding type of interconnection transformer, the ground fault current is increased up to 320[A] before
A Study on Fault Analysis of Wind Power (DFIG) in Distribution Systems
65
without DFIG. From among this, 240 [A] is supplied by interconnection transformer side and 80 [A] is supplied by wind power side. 3.4
Analysis for the Capacity of DFIG(Interconnection Transformer)
This paper performs the simulation with parameter of the capacity of interconnection transformer and DFIG. Table 3 is the simulation results of SLG fault current. Namely, when decreasing the % impedance of interconnection transformer, SLG current can be increased, because the total % impedance can be decreased due to % impedance transformer in conjunction with parallel form to distribution system. Table 4 shows that SLG current is increase as capacity of wind power and interconnection transformer getting increased. Table 3. Fault current by the capacity of interconnection transformer
DFIG [MW]
1
% impedance of transformer
Ia_CB4[A]
Ia_DFIG[A]
Ia_Fault[A]
I_IDGN[A]
0.04[pu](j320)
1027.13
367.69
1354.95
821.40
0.06[pu] (j480)
973.10
255.92
1213.15
614.65
0.08[pu] (j640)
949.04
237.37
1127.09
510.09
0.1[pu] (j800)
933.24
204.97
1071.32
435.81
Table 4. Fault current by the capacity of interconnection transformer and WP
DFIG [MW]
Load capacity [MW]
Ia_CB4[A]
Ia_DFIG[A]
Ia_Fault[A]
I_IDGN[A]
0.5
0.5
906.49
146.57
1041.09
356.68
1
1
972.99
254.53
1212.30
614.21
2
2
1068.80
426.23
1462.33
985.70
3
3
1132.64
560.57
1641.24
1244.79
66
6
J. Jeon et al.
Conclusions
This paper has been dealt with the fault analysis of the distribution system with DFIG(WP), and proposed the mechanism of mal-functional problem of protective devices by using both the conventional symmetrical component method and PSCAD/EMTDC. The results are summarized as follows (1) It is confirmed that if ground fault occurs in primary feeder interconnected with WP having the Y-grounded/Delta winding type of interconnection transformer, the ground fault current is divided by 2 ways. One is supplied by interconnection transformer side, and the other is supplied by wind power side. (2) It is found that because the size of fault current depends on the capacity of interconnection transformer and WP, the change of it can affect the possibility of the mal-function of the protection devices. Acknowledgment. This Work was support by the Power Generation & Electricity Delivery of the Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Korea government Ministry of Knowledge Economy. (No. 20101020300430)
References [1] Son, J., Kim, B., Rho, D.: A Study on the Optimal Operation Method for Protective Devices at the Primary Feeders Inter connected with New Energy Sources. In: IASTED Conferences, pp. 696–701 (2010) [2] New Energy and Renewable Energy Development, Use, and Spread Promotion Law [3] Barker, P.P., de Mello, R.W.: Determining the impact of distributed generation on power systems: Part 1 - Radial distribution systems. In: Proc. IEEE Power Eng. Soc. Summer Meeting, vol. 3, pp. 1645–1656 (2000) [4] Arritt, R.F., Dugan, R.C.: Distributed generation interconnection transformer and grounding selection. In: Power and Energy Society General Meeting - Conversion and Delivery of Electrical Energy in the 21st Century, pp. 1–7. IEEE (2008) [5] Shin, D.-y.: An Analysis and Countermeasure for the Reverse-Power-Flow Phenomena of Distribution System interconnected with Distributed Generations. Chungnam National University (February 2009) [6] Jeong, B.-C.: Wind Generator Control Algorithm for Inrush Current Restraint and Torque Ripple Compensation of Doubly-Fed Induction Generator. Chonbuk National University (2006) [7] Jang, M.-S., Bang, H.-J.: The Current Status and the Prospects of Wind Energy. Journal of the Environmental Sciences 18(8), 933–940 (2009)
Poor Periodontal Health in Type 1 Diabetic Youth Ji-Hye Park1,2, Youn-Hee Choi1, Cheol-Woo Ko3, Heung-Sik Kim4, Sang-Gyu Lee5, Hyung-Gyoo Kim6, and Keun-Bae Song1 1
Department of Preventive Dentistry, School of Dentistry, Kyungpook National University, Daegu, Korea {mahal23,cyh1001,kbsong}@knu.ac.kr 2 Department of Dental Hygiene, Ulsan College, Ulsan, Korea 3 Department of Pediatrics, Kyungpook National University Hospital, Daegu, Korea [email protected] 4 Department of Pediatrics, Keimyung University School of Medicine, Daegu, Korea [email protected] 5 Department of Preventive Medicine and Public Health, School of Medicine, Dankook University, Chenan, Korea [email protected] 6 Graduated School of Business and Public Administration, Anyang University, Anyang, Korea [email protected]
Abstract. The purpose of this study is to provide evidence of the relation between the two diseases while evaluating the periodontal health of the type 1 diabetic children and adolescents. The total number of participants were 148; 41 diabetics, and 107 controls. Logistic regression revealed that the possibility of contracting diabetes is 2.71 times higher (p=0.042) as the number of brushing increases, 26.08 times higher (p<0.001) as the gingival index gets higher, 23.37 times higher (p=0.010) for those who have any attachment loss than who have not, and 5.57 times higher (p=0.009) for those involved in their family history than who are not. The gingival index and the attachment loss of the type 1 diabetic youth were likely to be higher than those in healthy controls. Keywords: Periodontal health, Type 1 diabetes, Youth.
precise blood glucose level through this method, it has high possibility of developing more than 1 complications such as neurosis, retinosis, coronary disease, chronic renal failure, and so on [2, 3]. Diabetes is one of the risk factors for periodontal disease [4, 5]. A recent metaanalysis of periodontal status in subjects with and without diabetes concluded that the diabetics had a significantly higher severity but the same extent of periodontal disease than non-diabetics [6]. This surely attributes to the fact that the increase of the blood glucose level must have led to debility in the immunity system, so have reduced the synthesis, maturation, and homeostasis of collagen in one's periodontium [7]. Related to this, periodontal disease was identified as the 6th complication of diabetes since the 1990s, and the American diabetic association [1] once also stated it as one of the complications of diabetes. Recently, researchers have begun to investigate the relation between the diabetes and the periodontal diseases not only with adults but also with the children divided by the diabetics and the non-diabetics [8]. After Pinson et al. [9] had experimented 26 type 1 diabetes from 7 to 18, average of 13.42 and 20 controls from their family or friends in the pediatric and adolescent diabetic clinic at medical college of Georgia, they could find out that most of the cases represented higher average gingival index at almost every teeth than the controls. Moreover, they could notice that the type 1 diabetic youth increase the severity of inflammatory gingival disease much more than the controls who were at the same age as them. Likewise, Lal et al. [10], with selecting 164 diabetics age from 6 to 13, 99% of type 1 diabetes who were in the Naomi Berrie diabetes center at Columbia university medical center, New York and 191 controls from 6 to 13 who were in the Colombia university college of dental medicine, stated that the diabetes were at a significantly higher risk for gingival bleeding. On the review paper about juveniles and adults who had diabetes, the type 1 diabetics initiate the periodontal diseases thereabout their puberty, and progress by age [11]. Finally, on the cross-sectional study with the diabetic children, Lalla et al. [8] suggested that periodontal destruction can start very early in life in diabetes and become more prominent as children become adolescents. Nonetheless, a few of documents so far are not comprehensive in that they do not cover all the races except some people observed by investigators in some developed countries. Furthermore, the evidences do not support the consequences sufficiently. Accordingly, the purpose of this study is to provide evidence of the relation between the two diseases while evaluating the periodontal health of the type 1 diabetic children and adolescents.
2
Materials and Methods
2.1
Study Population
The total number of participants were 148; 41 diabetics, and 107 controls. The study design and procedure of the research was reviewed and approved by Kyungpook National University Hospital Institutional review board (74005-1434) and Keimyung
Poor Periodontal Health in Type 1 Diabetic Youth
69
University Dongsan Hospital Institutional review board (09-103), respectively. Moreover, all of their parents were acquainted with the purpose and the process with filling in their agreement of the experimental involvement. The diabetics were selected among the patients in the pediatrics of Kyungpook National University Hospital and of the Keimyung University Dongsan Hospital, locate in Daegu, South Korea. Participants were reffered by their pediatrician to the department of preventive dentistry at Kyungpook National University to receive oral examination. The extent of qualified age of the subjects was from 7 to 23, as well as they had to diagnosed with type 1 diabetes prior to participation. Also, those who had the pathological condition or any other diabetic complications were excluded. The controls were selected in three local community based schools placed in the same metropolitan city, and they were healthy students who do not embrace any systemic disease including diabetes. Then, the 1:2 frequency matching was implemented within the two groups based upon the sex and age. 2.2
Oral Examination
One of the trained dentists had oral examination all the teeth except the wisdom teeth to notify the decayed, missing, and filled surface/teeth (DMFS/DMFT)index and pocket depth, attachment loss, gingival index, plague index, bleeding on probing, and the presence of dental calculus in upper and lower (upper right jaw, lower left jaw), respectively. In addition, measured the stimulated salivary rate for 5 minutes while the participants were chewing the paraffin wax. 2.3
Diabetes-Related Factors and Other Information
The fasting glucose level, the Hemoglobin A1c (HbA1c), and the concentration of albumin and creatinine were brought by the medical records of the pediatrics, as well as the measurement of another HbA1c (NycoCard READER II, Axis-Shield Poc AS, Norway) and the irregular glucose levels (CareSens II, I-Sens Korea) in which time the participants were visited in the preventive dentistry. Another diabetes-related factors such as the duration of the diabetes, the age when the moment diagnosed, and the family history were collected by two dental hygienists' interview. Socioeconomic factors were revealed by self-administered questionnaires; the sex, age, frequency of in-taking snacks in daily, numbers of taking exercises on weekly, times of brushing teeth in daily, whether using auxiliary oral hygiene materials or not, whether checking dental conditions occasionally or not, the educational level of the parents, the average amount of household's incomes per month, whether their parents are smoking or drinking liquors. Also, had found the conditions of the general health of children like weights, heights, waists to use for calculating the Body Mass Index (BMI). 2.4
Statistical Analysis
The bleeding points on probing had converted into the bleeding points proportion; total number of bleeding points were divided by the all sites that have investigated
70
J.-H. Park et al.
than multiplied by 100. We had carried out the t-test and χ2-test to compare the socioeconomic aspects and dental status between the diabetics and the controls. In addition, the logistic regression analysis to be acquaint with the relation between the diabetes and periodontal status. All of the statistical analysis were performed by using SPSS (17.0 for window, Korea) and considered the level of significance as 0.05.
3
Results
Sociodemographic and health characteristics of study population are given in Table 1. The average age was 14.5 (range 7-23); the male was 54.1%, and the female 45.9%. Their average number of involving in sports per week was 2.5 times, and taking snacks per day 1.1 times. They brush their teeth 2.2 times per day, and visited dental clinics about 11.1 times in the latest 6 months. Moreover, merely 9.3% of the participants replied that they use dental auxiliaries. Among the health behaviors, only the number of taking snacks showed significance between the diabetics and the controls. Table 1. Sociodemographic and health characteristics of study population Cases (N=41)
Controls (N=107)
p-value*
Socio-demographic characteristics Age (years) 13 14-17
11(26.8) 24(58.5)
40(37.4) 59(55.1)
18 Gender
6(14.6)
8(7.5)
Boy Girl
22(53.7) 19(46.3)
58(54.2) 49(45.8)
0.952
Education (father) High School
0.671
0.269
23(57.5)
62(61.4)
College Education (mother)
17(42.5)
39(38.6)
High School College
27(67.5) 13(32.5)
69(69.7) 30(30.3)
0.800
Household income (10,000 won per month) 199 10(28.6)
47(49.0)
0.110
200-299 300
18(18.8) 31(32.3)
10(28.6) 15(42.9)
Poor Periodontal Health in Type 1 Diabetic Youth
71
Table 1. (Continued) Health behaviors Smoking (parents) Yes No
23(56.1) 9(22.0)
45(43.3) 45(43.3)
9(22.0)
14(13.5)
Yes No
30(73.2) 11(26.8)
70(68.0) 33(32.0)
0.540
Sports (hours per week) 1
18(43.9)
66(64.1)
0.085
2 3
9(22.0) 14(34.1)
14(13.6) 23(22.3)
2( 5.3)
12(12.0)
20(52.6) 16(42.1)
62(62.0) 26(26.0)
Former Drinking (parents)
Brushing (times per day) 1 2 3 Dental auxiliary use Yes
3(7.7)
9(9.0)
No Dental visits (6 months)
36(92.3)
91(91.0)
None 1
19(46.3) 8(19.5)
55(51.4) 13(12.1)
2 Snacks (times per day)
14(34.1)
39(36.4)
1 2
8(19.5) 12(29.3)
53(52.0) 31(30.4)
3
21(51.2)
18(17.6)
0.051
0.136
0.805
0.515
<0.001
N(%). * Denote the significance by Ȥ2-test.
Oral health and diabetes-related characteristics of study population are given in , and the DMFT index 4.07. Table 2. The average stimulated flow rate was 6.50 Among the oral health factors, gingival index manifested the relatable contrasts (p<0.001) between the diabetes (1.17) and the controls (0.83). Also, the attachment loss of the diabetes (0.13 ) was significant(p=0.038) with that of the controls (0.02 ). Furthermore, among the diabetes-related characteristics, the HbA1c of the cases (8.68%) was significant (p<0.001) with that of the controls(5.56%); as well as the 41.0% of the cases had family history of DM comparable to 20.4% of the controls.
㎖
㎜
㎜
72
J.-H. Park et al. Table 2. Oral health and diabetes-related characteristics of study population Cases (N=41)
Oral health Stimulated flow rate ( min)
㎖/5
6.10±2.75
Controls (N=107)
6.65± 2.75
p-value*
0.289
DMFS index
7.49±8.70
6.42± 6.67
0.431
DMFT index
4.54±4.64
3.88± 3.64
0.372
Gingival index
1.17±0.36
0.83± 0.38
<0.001
Plaque index
1.09±0.55
1.11± 0.59
0.918
12.78±8.12
11.13±11.14
0.389
1.22±0.25
1.23± 0.22
0.874
0.13±0.32
0.02± 0.14
0.038
None
17(41.5)
52(48.6)
0.548
Mandible
21(51.2)
51(47.7)
3( 7.3)
4( 3.7)
8.68±2.71
5.56± 0.47
<0.001
Yes
16(41.0)
23(20.4)
0.013
No
20(59.0)
78(79.6)
BMI (weight/height^2)
20.71±4.39
24.00±25.42
0.418
Waist circumference (inches)
28.59±4.80
28.10± 3.72
0.511
Bleeding points proportion (%)
㎜) Attachment loss (㎜)
Pocket depth (
Dental calculus
Both Diabetes-related characteristics HbA1c (%) Family history of DM
Mean±SD or N(%). * Denote the significance by t-test or χ2-test.
The results of logistic regression analysis for diabetes are given in Table 3. The possibility of contracting diabetes is 2.71 times higher (p=0.042) as the number of brushing increases, 26.08 times higher (p<0.001) as the gingival index gets higher, 23.37 times higher (p=0.010) for those who have any attachment loss than who have not, and 5.57 times higher (p=0.009) for those involved in their family history than who are not.
Poor Periodontal Health in Type 1 Diabetic Youth Table 3. Estimated ORs and 95% CIs from logistic regression model for diabetes OR(95% CI)
p-value
Age
1.01(0.78- 1.31)
0.924
Brushing
2.71(1.04- 7.08)
0.042
Stimulated flow rate
0.92(0.74- 1.15)
0.483
DMFT index
1.05(0.87- 1.26)
0.639
BMI
0.94(0.80- 1.12)
0.502
Gingival index lower higher
1.00 26.08(4.72-144.03)
<0.001
Plaque index 1st tertile
1.00
2nd tertile
0.50(0.11- 2.37)
0.386
3rd tertile
0.23(0.05- 1.20)
0.082
Bleeding points proportion 1st tertile
1.00
2nd tertile
0.47(0.09- 2.41)
0.365
3rd tertile
0.51(0.08- 3.24)
0.473
1.00
0.010
Presence of attachment loss No Yes
23.27(2.09-259.11)
Family history of DM No
1.00
Yes
5.67(1.54- 20.80)
0.009
73
74
4
J.-H. Park et al.
Discussion
According to the study, since the gingival index was 1.17 for the diabetes and 0.83 for the controls, they can be considered as significant (p<0.001). Also, attachment loss was in relatable contrasts; 0.13 for the diabetes and 0.02 for the controls (p=0.038). Furthermore, the results of calculating the relevance between the periodontal index and the diabetics by the logistic regression model were that the percentage of contracting diabetes for those who had high gingival index embraced about 26.08 times (p<0.001) higher than the others who were low. Similarly, the group who had attachment loss demonstrated higher possibility for about 23.37 (p=0.010) times higher than those who had not (p=0.009). The periodontitis especially begins at puberty for those who aborted to control their blood glucose level from the time of infancy, after suffering from slight gingivitis and gingival recession with bleeding, finally progress to severe periodontitis. Microangiopathy, impaired immune response, different bacterial microflora, and collagen metabolism play a role of outbreaking the diabetic periodontitis. The gingival flora is mostly composed of Gram-negative, anaerobic bacteria, while collagen has a lower solubility and is atrophic and inadequate to support the occlusion forces [12]. Sbordone et al. [13] had followed up 16 diabetic juveniles for 3 years, and come with the fact that the Prevotella inermedia increases relatively in after 3 years comparable to the base line significantly. Several researches [14, 15] had revealed that a diabetic child has lower salivary flow rate than a normal, and this was subject to the alteration of salivary gland vascularization. Even though the flow rate of diabetics (6.10 ) were lower than that of controls (6.65 ) in this study too, their numerical values were not relatable enough (p=0.289). Most of the researchers were consonant with the idea that the relevancy between the periodontal health and the plaque index were lack of [15-18]. Sandholem et al. [18] had also reported that, after comparing the oral health of the diabetic children and that of normal children, the diabetic children have absolutely high gingival index (p<0.0025), even if there was not relatable contrast in plaque index. Glickman [19] had noticed that the diabetic patients react abnormally to common irritating factors responsible for periodontal disease. Similarly, while the plaque index of diabetics (1.09) and that of controls (1.11) had not relatable contrasts (p=0.918) in this research, too, there surely occurred meaningful conclusions in terms of the gingival index; the value of 1.17 for the diabetics was relatively higher than that of 0.83 for the normal (p<0.001). As the duration of contracting the diabetes gets longer, the periodontal disease affects not only in gingiva but also in periodontium; that is, after developing to deeper structures such as alveolar bone, it bring about severe periodontitis with abscess, enlargement of periodontal ligament, the alveolar bone loss, the formation of pocket, the mobility of the teeth, and even the loss of the tooth [13]. Firatli et al. [20] briefed that there is correlation between the duration of diabetes and the attachment loss. Though it does not appear in the table above, the average duration of diabetes of the participants of the study was 4.85 years, and the average attachment loss of the cases (0.13 ) were statistically relatable (p=0.038) compared to that of the controls (0.02 ). Aren [21] also said that after comparing the pocket depth to the controls, while the numeral value was similar with the diabetics who
㎜
㎜
㎖
㎖
㎜
㎜
Poor Periodontal Health in Type 1 Diabetic Youth
75
were told to have a diabetes, the other diabetics who had the long duration had higher significant values. After analyzing the variables which influence the diabetes through the logistic regression model in this study, the possibility of contracting diabetes is 2.71 times higher (p=0.042) as the number of brushing increases, 26.08 times higher (p<0.001) as the gingival index gets higher, 23.37 times higher (p=0.010) for those who have any attachment loss than who have not, and 5.57 times higher (p=0.009) for those involved in their family history than who are not. The probable reason for high numbers of brushing per day in the diabetic group is that such compensational mentality that concerns the complication of diabetes might have affected. Besides, after Lalla et al [22] had implemented the logistic regression with the 350 diabetics (range 6-18) in Naomi Berrie Diabetic Center of Columbia University in the USA and 350 controls (range 618) in the pediatric dental clinic of East University, they reported that the contrasts of the destruction of the periodontium (gingival bleeding and the attachment loss, OR: 1.84-3.72) were absolutely significant between the diabetics and the controls. And this result is surely similar with this study, too. Additionally, it is no use to say that the family history is one of the main risk factors for diabetes. One of the limitations of the study is that there might occur some selection biases owing to the fact that the diabetes were those who had volunteered from a few hospitals placed in limited areas in South Korea. Also, the controls cannot be considered as the representation of the whole Korean population since they were recruited from a few schools in some confined areas. In addition, it is not convincible enough owing to insufficiency of the subjects. Likewise, as the study was progressed by cross-sectional study, the chronical sequence of the diabetes and the periodontal status is remained doubtful. However, this study is surely worthwhile in that this was the first research which had tried to identify the relevancy between the diabetes and the dental health among the type 1 diabetic children and juveniles in South Korea.
5
Conclusion
The gingival index and the attachment loss of the type 1 diabetic youth were likely to be higher than those in healthy controls. Consequently, the prevention of periodontitis for the type 1 diabetic adolescents is crucial, and they have to visit dental clinics to take a regular check-up for gingival health. Acknowledgements. This study was supported by National Research Foundation of Korea (2009- 0069115).
References 1. American Diabetes Association, http://www.diabetes.org 2. Korean Diabetes Association, http://www.diabetes.or.kr 3. Merchant, A.T., Jethwani, M., Choi, Y.H., Morrato, E.H., Liese, A.D., Mayer-Davis, E.: Associations between periodontal disease and selected risk factors of early complications among youth with type 1 and type 2 diabetes: a pilot study. Pediatr. Diabetes (2011)
76
J.-H. Park et al.
4. Garcia, R.I., Henshaw, M.M., Krall, E.A.: Relationship between periodontal disease and systemic health. Periodontol 25, 21–36 (2001) 5. Genco, R.J.: Current view of risk factors for periodontal diseases. J. Periodontol 67, 1041–1049 (1996) 6. Khader, Y.S., Dauod, A.S., El-Qaderi, S.S., Alkafajei, A., Batayha, W.Q.: Periodontal status of diabetics compared with nondiabetics: a meta-analysis. J. Diabetes. Complications 20, 59–68 (2006) 7. Tsai, C., Hayes, C., Taylor, G.W.: Glycemic control of type 2 diabetes and severe periodontal disease in the US adult population. Community. Dent. Oral. Epidemiol. 30, 182–192 (2002) 8. Lalla, E., Cheng, B., Lal, S., Tucker, S., Greenberg, E., Goland, R., Lamster, I.B.: Periodontal changes in children and adolescents with diabetes: a case-control study. Diabetes. Care. 29, 295–299 (2006) 9. Pinson, M., Hoffman, W.H., Garnick, J.J., Litaker, M.S.: Periodontal disease and type I diabetes mellitus in children and adolescents. J. Clin. Periodontol. 22, 118–123 (1995) 10. Lal, S., Cheng, B., Kaplan, S., Softness, B., Greenberg, E., Goland, R.S., Lalla, E., Lamster, I.B.: Gingival bleeding in 6- to 13-year-old children with diabetes mellitus. Pediatr. Dent. 29, 426–430 (2007) 11. Manouchehr-Pour, M., Bissada, N.F.: Periodontal disease in juvenile and adult diabetic patients: a review of the literature. J. Am. Dent. Assoc. 107, 766–770 (1983) 12. Iughetti, L., Marino, R., Bertolani, M.F., Bernasconi, S.: Oral health in children and adolescents with IDDM–a review. J. Pediatr. Endocrinol. Metab. 12, 603–610 (1999) 13. Sbordone, L., Ramaglia, L., Barone, A., Ciaglia, R.N., Iacono, V.J.: Periodontal status and subgingival microbiota of insulin-dependent juvenile diabetics: a 3-year longitudinal study. J. Periodontol. 69, 120–128 (1998) 14. Ben-Aryeh, H., Serouya, R., Kanter, Y., Szargel, R., Laufer, D.: Oral health and salivary composition in diabetic patients. J. Diabetes. Complications 7, 57–62 (1993) 15. Markopoulos, A.K., Belazi, M.: Histopathological and immunohistochemical features of the labial salivary glands in children with type I diabetes. J. Diabetes. Complications 12, 39–42 (1998) 16. de Pommereau, V., Dargent-Pare, C., Robert, J.J., Brion, M.: Periodontal status in insulindependent diabetic adolescents. J. Clin. Periodontol. 19, 628–632 (1992) 17. Firatli, E.: The relationship between clinical periodontal status and insulin-dependent diabetes mellitus. Results after 5 years. J. Periodontol. 68, 136–140 (1997) 18. Sandholm, L., Swanljung, O., Rytomaa, I., Kaprio, E.A., Maenpaa, J.: Periodontal status of Finnish adolescents with insulin-dependent diabetes mellitus. J. Clin. Periodontol. 16, 617–620 (1989) 19. Glickman, I.: The relation of experimental diabetes to periodontal disease. Am. J. Orthod. 33, 703–722 (1947) 20. Firatli, E., Yilmaz, O., Onan, U.: The relationship between clinical attachment loss and the duration of insulin-dependent diabetes mellitus (IDDM) in children and adolescents. J. Clin. Periodontol. 23, 362–366 (1996) 21. Aren, G., Sepet, E., Ozdemir, D., Dinccag, N., Guvener, B., Firatli, E.: Periodontal health, salivary status, and metabolic control in children with type 1 diabetes mellitus. J. Periodontol. 74, 1789–1795 (2003) 22. Lalla, E., Cheng, B., Lal, S., Kaplan, S., Softness, B., Greenberg, E., Goland, R.S., Lamster, I.B.: Diabetes mellitus promotes periodontal destruction in children. J. Clin. Periodontol. 34, 294–298 (2007)
Performance Evaluation of MapReduce Applications on Cloud Computing Environment, FutureGrid Yunhee Kang1 and Geoffrey C. Fox2 1
Division of Information and Communication, Baekseok University, 115 Anseo Dong, Cheonan, Korea 330-704 [email protected] 2 Pervasive Technology Institute, Indiana University, 2719 E 10th St. Bloomington, Indiana, IN 47408, USA [email protected]
Abstract. This paper describes the result of performance evaluation of two kinds of MapReduce applications running in the FutureGrid: a data intensive application and a computation intensive application. For this work, we construct a virtualized cluster system made of a set of VM instances. We observe that the overall performance of a data intensive application is strongly affected by the configuration of the VMs. It can be used to identify the bottleneck of the MapReduce application running on the virtualized cluster system with various VM instances. Keywords: MapReduce application, Cloud computing, FutureGrid.
computational intensive application. For this work, we construct a virtualized cluster system made of a set of nodes of which are VM instances in the FutureGrid [2]. To monitor and measure the performance of nodes in the virtualized cluster system while the MapReduce applications are running on nodes, we use a tool, top command. To take a system snapshot of the cluster system, we write a shell script that extracts information including load average and memory/swap usage from the results of top command. The paper is organized as follows: Section 2 describes the related works including brief overview of MapReduce, Twister and the FutureGrid. We report the results from our experimental study and the observations from the result of our experiments in Section 3. Conclusions are presented in Section 4.
2
Related Works
2.1
MapReduce
MapReduce programs are designed to compute large volumes of data in a parallel fashion. It is a kind of data parallel languages aimed at loosely coupled computations that execute over given data sets [3]. This requires dividing the workload across a large number of machines. The degree of parallelism depends on the input data size. MapReduce, introduced by Dean and Ghemawat at Google, is the most dominant programming model for developing applications in cloud computing environment [3-5]. In a MapReduce application supported by a MapReduce library, all map operations can be executed independently. Each reduce operation may depend on the outputs generated by any number of map operations. All reduce operations can also be executed independently. The following describes the MapReduce programming model: • The computation takes a set of input (key, value) pairs, and produces a set of output (key, value) pairs. The computation is expressed as two functions: Map and Reduce. • The Map takes an input pair and produces a set of intermediate (key, value) pairs. The MapReduce library groups together all intermediate values associated with the same intermediate key key and passes them to the Reduce function. • The Reduce accepts an intermediate key key and a set of values of that key. It merges together these values to form a possible smaller set of values. Traditional parallel applications are based on a runtime library for message passing such as MPI [10] and PVM [11] that have some programming features of communication and synchronization. The feature provided by a runtime library is a low-level primitive. In MapReduce, a programmer is able to focus on the problem that needs to be solved since only the map and reduce functions need to be implemented, and the framework takes care of computing the programmer has to deal with lowerlevel mechanisms to control the data flow [2,4].
Performance Evaluation of MapReduce Applications
2.2
79
Twister
There are some existing implementations of MapReduce such as Hadoop [6] and Sphere [7]. Twister is one of MapReduce implementations, which is an enhanced MapReduce runtime with an extended programming model that supports an iterative MapReduce computing efficiently [8]. In addition it provides programming extensions to MapReduce with broadcast and scatter type for transferring data. These improvements allow Twister to support iterative MapReduce computations highly efficiently compared to other MapReduce runtimes. It reads data from local disks of the worker nodes and handles the intermediate data in distributed memory of the worker nodes. All communication and data transfers are performed via a pub/sub messaging system NaradaBrokering that is an open-source, distributed messaging infrastructure [9]. Twister uses a publish/subscribe messaging infrastructure to handle four types of communication needs; (i) sending/receiving control events, (ii) send data from the client side driver to the Twister daemons, (iii) intermediate data transfer between map and reduce tasks, and (iv) send the outputs of the reduce tasks back to the client side driver to invoke the combine operation. 2.3
FutureGrid
FutureGrid is a distributed testbed for developing research applications and middleware, which employs virtualization technology to allow the testbed to support a wide range of operating systems. This project provides a capability that makes it possible for researchers to tackle complex research challenges aimed at minimizing overhead and maximizing performance in computer science related to the use and security of grids and clouds. It has been offering a flexible reconfigurable testbed based on dynamically provisioning software to support deploying a specific image to variety of environments composed of virtual machines [2]. A machine image is used as a template that is an abstraction of software stack including such as operating system, middleware and end-user access solutions. Hence a machine instance is an actual instantiation of the template. One of the goals of this project is to understand the behavior and utility of cloud computing approaches. FutureGrid dynamically provides diverse configurations that have different operating systems and middleware configurations. The goal of dynamic provisioning is to partition a set of resources in an intelligent way to provide a user-defined environment to any user that makes such a request.
3
Experimental Results
3.1
Experiment Environment
In this experiment, a virtualized cluster system consists of a set of nodes that are allocated from a cluster named India, which is one of FuturGrid environments. The image of an instance basically contains Linux 2.6.27 and Java VM 1.6. Each instance provides a predictable amount of dedicated computing capacity that is
80
Y. Kang and G.C. Fox
defined in FutureGrid. Table 1 shows an overview of the types of VM instances to be used in the experiments. Hereafter CPU is used to represent core. Table 1. Main specification of VM instance type
Type of VM instance
Main HW Features CPU
Memory (Mbyte)
Disk
c1-medium
1
1,024
7
m1-large
2
6,000
10
m1-xlarge
2
12,000
10
We make a configuration, which is based on the type of VM instance described in Table 1, for a virtualized cluster system as testbed and use various configurations that are used to evaluate performance of the aforementioned MapReduce applications. A configuration has various middleware setups. It is used to represent a specific workload. For example, Type-2 represents an unbalanced load allocation and Type4 represents a balanced load allocation. Table 2 shows the list of configurations to be used in our experiments except gf14-gf15 and India. The gf14-gf15 is composed of two Linux machines. India is a multi-core machine having 1,024 cores in 128 nodes in the FutureGrid. Table 2. Configuration of virtual cluster systems Main Configuration
To set up the virtualized cluster systems, we deploy images and run the instances in India. A MapReduce application is implemented on a system using: • Twister 0.8 • NaradaBroker 4.2.2 • Linux 2.6.x running on Xen
Performance Evaluation of MapReduce Applications
81
We gather and analyze OS-level performance metrics, without requiring any modifications to Twister, its applications or the OS to collect these metrics. For data collection, we choose a Linux command to be used as a system-monitoring tool, top that provides a dynamic real-time view of a running system, including information about system resource usage and a constantly updated list of the processes that are consuming the most resources. 3.2
Experiment: Data Intensive Application
In this experiment, two different computing environments are evaluated, which are running a data intensive application with various configurations: one is a cluster system composed of physical machines and the other is a virtualized cluster computing system. For this work, we construct a MapReduce application that is used to transform a data set collected from a music radio site, Last.fm(http://www.last.fm/) that provides the metadata for artists include biography by API, on the Internet. The goal program is to histogram the counts referred by musicians and to construct a bi-directed graph based on similarity value between musicians in the data set. We compare both environments with application’s performance metrics in terms of elapse time and standard variation. The graph in Figure 1 plots the results using the MapReduce application. In Type-1, there is starvation problem as the application run that is caused by no enough computing resources including CPU and memory capability. In the part of the graph, Type-2 to Type-5, we see that as the resources of VMs including CPU and memory increase, the elapse time of the application and the value of its standard variation decrease.
Fig. 1. Elapse time of similarity: 5 configurations – Virtualized cluster systems (4 configurations) and a Physical cluster systems (1 configuration)
What we observed is that the number of CPUs has less impact on the elapse time in comparison with the results of Type-3 and Type-4. Though performance degrades as the application runs in the virtualization environment, the performance of Type-3 still provides 80.9% of the average performance of gf14-gf15 and India when running the physical computing environment. However, the elapse time of type Type-4 is 98.6 % of the elapse time of gf14-gf15.
82
Y. Kang and G.C. Fox
As shown in Figure 2, the average memory usage is almost 100% of real memory and swap area in the node having a resource shortage problem during the most of its running time. Hereafter NB stands for NaradaBroker that plays a role as message. Hence the application is delayed due to the heavy disk I/O caused by high swap value.
Fig. 2. Memory/swap area usage of Type-2 (NB running on the node typed with c1-medium)
As shown in Figure 3, the average memory usage is around 80% in the node during its running time. But the average swap area usage is less than 1% in running time. As a result, it can properly handle the I/O requests from the application.
Fig. 3. Memory/swap area usage of Type-3 (NB running on the node typed with m1-medium)
Based on the performance evaluation we may choose the configuration of a virtualized cluster system to provide 80% of performance of a real cluster system. This observation induces that load balancing is helpful in spreading the load equally across the free nodes when a node is loaded above its threshold level.
Performance Evaluation of MapReduce Applications
83
• The performance of the application running on the Twister strongly depends on the throughput of a message broker, Naradabroker. • The pending of the application is caused by broken pipe between a Twister daemon and a Naradabroker server when Naradabroker has a threshold of the limitation to accept a connection from Twister due to its QoS requirement. • The capability of Naradabroker in the middleware configuration affects the performance of an application as the application runs in the same configuration computing resource. 3.3
Experiment: Computation Intensive Application
To do performance evaluation of a MapReduce application typed computation intensive, one configuration, type-5, is added to the configurations of this experiment. In this experiment, we use a parametric k-means algorithm with 100,000 data points, which is to organize these points into k clusters. We compare environments, a virtual cluster system and a physical cluster system, with application’s performance metrics in terms of elapse time and standard variation.
Fig. 4. Elapse time of k-means: 5 configurations - virtualized cluster systems (4 configurations) and a physical cluster system (1 configuration)
As shown in Figure 4, our experiments indicate that the average of elapse time increases over 375.5% in the virtualized cluster computing system, in comparison with a physical cluster system, represented by India. In Type-1, there is also the same starvation problem that is occurred as the data intensive application runs. Besides, the elapse time decreases proportionally as VM’s CPU capability is added to the virtualized cluster computing system. Furthermore, the standard deviation is less
84
Y. Kang and G.C. Fox
affected by configuration change and the size of input data. In the physical cluster system, the value remains vey low at about 1-2% of the variation of elapse time due to the capability of system mainly related with CPU power. In addition, the standard variation in the three configurations of the virtualized cluster computing system remains low at about 2.0-3.78%. A similar trend is observed in the values of standard deviation of all configurations. Hence we can expect that as the number of available VMs increases, there is a proportional improvement of elapse time. 3.4
Summary of the Experiments
In summary, performance evaluation based on the metrics, load average and memory/swap area usage, according to the type of specific application is essential to choose properly a configuration that consists of a set of instances in the FutureGrid. This observation induces that load balancing is helpful in spreading the load equally across the free nodes when a node is loaded above its threshold level. Though load balancing is not so significant in execution of a MapReduce algorithm, it becomes essential to handle large files in the case of having a limited computing resources in the FutureGrid. As a highlight, it enhances hardware utilization in resource-critical situations with a slight improvement in performance. It is important to determine a configuration of a virtual cluster system in order to run a MapReduce application efficiently in the FutureGrid. Eventually we conclude that the appropriate selection of a set of VM instance types increases the overall utilization of resources in the FutureGrid. This approach is the way to identify the relationship between the type of applications and resources allocated for running them. The result obtained from the experiments leads that it is important to determine a configuration of a virtual cluster system in order to run a MapReduce application efficiently in cloud computing. This approach is the way to identify the relationship between the type of applications and resources allocated for running them. 3.5
Observation of System Behavior with Anomaly
This section is mainly focused on the resource starvation. Figures 5 and 6 show the snapshot of a virtualized system having resource shortage problem. A buffer value is used to measure the throughput of network I/O between a MapReduce application and a middleware to be used as message broker. Figure 5 shows the buffer value is so low. The buffer value indicates how much of the memory in use. The buffer is currently being used for I/O buffering. Basically an I/O request happens when a Map task is finished and a Reduce task is started. Simultaneously there is no available cache memory in 7 seconds, while the memory usage is sharply high by up to 100 % as shown in the Figure 5. Hence system is still pending caused by little or no memory available. It can exacerbate failures.
Performance Evaluation of MapReduce Applications
85
Fig. 5. Buffer variations in VM
We also observed that the variation of swap area is happened. Figure 6 shows the usage of swap area is sharply increased. It triggers a write burst to a disk and affects the response time of a MapReduce application. Because of the swapping, the system has slowed down, and heavy disk drive activity can be happened. There is still a small amount of free memory.
Fig. 6. Memory and swap variations in VM
We have observed an anomaly of system behavior as a data intensive MapReduce application runs in a virtualized cluster system that consists of VMs in the FutureGrid. It is caused by having no enough computing resources including CPU and memory capability to run the application and by having an inappropriate configuration of a set of VMs associated with a middleware setting.
4
Conclusion
Cloud computing is designed to provide on demand resources or services over the Internet. What we have observed in the experiments is that the overall performance of data intensive application is strongly affected by the throughput of the messaging
86
Y. Kang and G.C. Fox
middleware since it requires to transfer data when a map task sends the intermediate result to a reduce task. When it is close to limit of available memory as a data intensive MapReduce application runs on the specific configuration of nodes, the elapse time sharply increases and its standard deviation is high. The performance of the MapReduce application is so strongly affected by the configuration of VM. However the performance of computational intensive application is associated with CPU throughput. It is less affected by the configuration of VMs having the same CPU power. This work represents a first step towards understanding the relationship between the configuration of VMs and performance effects associated with the type of applications. The result of the experiments can be used for selecting the proper configuration based on the proposed guideline in cloud computing. It can be used to identify the bottleneck of a MapReduce application running on the resource given VM configuration. It will be used to extend the information service system associated with the middleware for cloud computing.
References 1. Armbrust, M., Fox, A., Griffith, R., Joseph, A.D., Katz, R.H., Konwinski, A., Lee, G., Patterson, D.A., Rabkin, A., Stoica, I., Zaharia, M.: Above the Clouds: A Berkeley View of Cloud Computing. EECS Department. University of California, Berkeley (2009) 2. https://portal.futuregrid.org/ 3. Dean, J., Ghemawat, S.: MapReduce: A Flexible Data Processing Tool. Commun. ACM 53, 72–77 (2010) 4. Morton, K., Friesen, A., Balazinska, M., Grossman, D.: Estimating the Progress of MapReduce Pipelines. In: IEEE 26th International Conference on Data Engineering (ICDE), Long Beach, CA, pp. 681–684 (2010) 5. Dean, J., Ghemawat, S.: Mapreduce: Simplified data processing on large clusters. Commun. ACM 51, 107–113 (2008) 6. Wang, F., Qiu, J., Yang, J., Dong, B., Li, X., Li, Y.: Hadoop high availability through metadata replication. In: Proceeding of the First International Workshop on Cloud Data Management, Hong Kong, China, pp. 37–44 (2009) 7. Grossman, R., Gu, Y.: Data mining using high performance data clouds: experimental studies using sector and sphere. In: Proceeding of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 920–927. ACM, Las Vegas (2008) 8. Ekanayake, J., Li, H., Zhang, B., Gunarathne, T., Bae, S.-H., Qiu, J., Fox, G.: Twister: A Runtime for Iterative MapReduce. In: The First International Workshop on MapReduce and its Applications (MAPREDUCE 2010) - HPDC 2010 (2010) 9. Fox, G., Pallickara, S.: Deploying the NaradaBrokering Substrate in Aiding Efficient Web & Grid Service Interactions. Invited paper for Special Issue of the Proceedings of the IEEE on Grid Computing 93(3), 564–577 (2005) 10. MPI. MPI(Message Passing Interface), http://www-unix.mcs.anl.gov/mpi/ 11. PVM. PVM(Parallel Virtual Machine), http://www.csm.ornl.gov/pvm/
Development of Evaluation Simulator of New Power Sources Using the LabVIEW Software Kyung-Sang Ryu, Byung-ki Kim, Suk-hwan Song, Seob-han Kim, and Dae-Seok Rho* Korea University of Technology and Education {ryu3212,dsrho}@kut.ac.kr
Abstract. This paper deals with a evaluation simulator which is composed of distribution system simulator, PV systems simulator, and control and monitoring systems with the LabVIEW S/W. The simulator can analyze the customer voltage characteristics of normal conditions, considering the 3 parameters of the introduction capacity for PV systems, system configuration and load factor. Also, the simulator can provide the analysis of the power quality problems in distribution systems, when new power sources like photovoltaic (PV) systems are interconnected. By performing the simulation for power quality operation characteristic, this paper presents the optimal countermeasures for power quality for the new distributed power sources(NDS).
1
Introduction
Recently, according to the expansion of green growth energy policy of government, many independent power producers(IPP) have a plan to introduce the new distributed power sources(NDS) such as PV, WP, FC, etc. However, many technical problems between IPP’s operators and KEPCO of power company has been occurred, because of insufficient skills and knowledges on the operation and management of distribution systems interconnected with NDS. In order to overcome these problems, we propose a reasonable and comfortable evaluation simulator for NDS using the LabVIEW S/W, to analyze technical problems of power quality and steady state conditions. The simulation results through characteristic test based on the several parameters such as power factor, line length, capacity and location of NDS show that the simulator can be a practical tool.
2
Design of Evaluation Simulator
Being different with existing distribution systems, the distribution system interconnected with NDS has to be operated under the condition of bi-directional power flow including the loads and sources. So there are many problems such as increase of fault current, protection coordination, voltage control and power quality. Especially the *
power quality problems are related with sag, swell, interruption and harmonic. Accordingly, it is getting an urgent issue to review and prepare a countermeasure about the problems mentioned above. So, this paper proposes evaluation simulator designed by real test device of NDS. 2.1
Voltage Drop Calculation Algorithm Considering NDS
The conventional algorithms for the voltage drop calculation can consider the one way power flow from distribution substations(sources) to customers(loads). However, the voltage rise by the reverse power flow of NDS output has to be considered if the NDS is interconnected to distribution systems. For doing this, this paper proposes an idea that the load current (I) is divided into 2 separate parts of real part (Ip) and imaginary part (Iq), in order to properly consider the bi-directional power flow and power factor of NDS. As shown in Fig. 1, the conventional method can consider the forward power flow from sources to loads and the lag power factor only using the 1rd quadrant, but the proposed method can consider the reverse power flow and lag power factor of NDS using all of the quadrants, 1st - 4th quadrants.
Fig. 1. Concept for voltage drop calculation algorithm using 4 quadratics
The typical equation for voltage drop is expressed by V = Z (impedance) × I (current). The new equation can be expanded by considering the concepts of 4 quadratics and real and reactive currents as : I Sq(n ) + I Rq( n ) I Sp(n ) + I Rp(n ) ΔV(n ) = k ⋅ ⋅ r(n ) + ⋅ x( n ) 2 2
(1)
where, Isp(n) and Isq(n) are real and reactive inflow current in nth section in Fig.2, IRp(n) and IRq(n) are real and reactive outflow current in nth section, r(n) and x(n) are resister and reactance in nth section, V(n) is voltage profile in nth section.
VSS 䚷, I SS
I R n I S n 1
I S n
SS Ạᴸn
Section n
I R n I S n 1 I L n
I L n
Fig. 2. Inflow and outflow current in n section
Development of Evaluation Simulator of New Power Sources Using the LabVIEW S/W
2.2
89
Configuration of the Simulator
The evaluation simulator proposed in this paper is composed by 4 parts including the sources of substation, distribution feeders of line impedance, loads of end user and new distributed power source of NDS as shown in Fig.3. And also, Fig.4 shows a flowchart of NDS evaluation simulator, which is performed on the basis of proposed voltage drop calculation algorithm considering NDS. Thus, the operation characteristics for distribution system between before NDS and after NDS can be measured and analyzed.
Fig. 3. Configuration of NDS Evaluation Simulator System
Fig. 4. Flowchart of NDS Evaluation Simulator System
3
Implementation of NDS Evaluation Simulator
Fig. 5 shows total concept of NDS Evaluation Simulator System. This is designed by front panel and block diagram on LabVIEW S/W. Front panel is able to analyze characteristics to feed each initial data based on substation, end user and NDS points.
90
K.-S. Ryu et al.
Fig. 5. NDS Evaluation Simulator System
Evaluation simulator is mainly composed of 4 components including distribution system, customers(loads), NDS and harmonic analysis as following. 3.1
Distribution System
Fig 6 shows the component of the distribution system which has power sources and distribution lines. At the power sources, system parameters such as voltage, current, and frequency can be controlled and displayed in an easy manner. And the distribution line is modeled by the impedance values of R and X. Therefore, the line voltage can be calculated by the change of the system parameters and impedance values. Simulated Substation
System Display
Line Control
Fig. 6. Evaluation of the simulator system and the line configuration
3.2
Customer Side
As shown in Fig. 7, customer side is designed by resistance(R), capacitor(C) and inductor(L) load. The current, voltage and power factor in the customer side can be
Development of Evaluation Simulator of New Power Sources Using the LabVIEW S/W
91
Consumer Data Input
System Analysis
Fig. 7. Evaluation of the simulator consumer configuration
calculated and displayed by changing the R, L, C parameters. In addition, the data saving function is designed to analyze the system parameters. 3.3
New Energy Sources(NDS)
The capacity of NDS is designed by controlling the NDS’s current according to the inverter system voltage. The current, voltage and power factor in the NDS side can be displayed by changing the parameters as shown in Fig. 8. New energy Power
Display
Fig. 8. New energy system and waveform diagram of simulator Evaluation
92
3.4
K.-S. Ryu et al.
Harmonic Analysis
In the simulator, this paper designs the harmonic analysis related to the PV system based on the real data as shown in Fig. 9. The principle of the harmonic analysis is that the harmonic is determined by the changing rate which was measured in a real system.
Fig. 9. Harmonic analysis of program configuration
4
Simulation and Analysis
By using the proposed simulator, this paper performed the several simulations for the operation characteristics of distribution system both with NDS and without NDS. 4.1
Voltage Sag
Table 1 shows the comparison results between NDS evaluation simulator and real test device(theoritical value), when a source(substation) voltage drops from 222V to 207V in a instantaneous manner and also the load is assumed as 400W and line impedance is 5Ω and the capacity of NDS is 1,200W. It is confirmed that the load-side(customer) voltage is getting bigger than source-side voltage because the reverse power flow of DNS can raise the load-side voltage, like Ferranti phenomenon. Table 1. Results of Instantaneous Voltage Drop(sag)
Test device Theory value Simulator
Substation Voltage 207 207 207
Customer Ampere 1.48(Reverse A) 1.93(Reverse A) 1.94(Reverse A)
Customer Voltage 217.37 220.51 220.58
Development of Evaluation Simulator of New Power Sources Using the LabVIEW S/W
4.2
93
Voltage Swell
Table 2 shows the comparison results between NDS evaluation simulator and real test device(theoritical value), when a source(substation) voltage raises from 222V to 233V in a instantaneous manner and also the load is assumed as 400W and line impedance is 5Ω and the capacity of NDS is 1,200W. It is confirmed that the load-side(customer) voltage is getting bigger than source-side voltage. Table 2. Results of Instantaneous Voltage Rise(Swell)
Test device Theory value Simulator
4.3
Substation Voltage 233 233 233
Customer Ampere 1.19(Reverse A) 1.72(Reverse A) 1.72(Reverse A)
Customer Voltage 241.36 245.04 245.04
Customer Voltage Variations
Table 3 shows the comparison results between NDS evaluation simulator and real test device(theoritical value), when the P + jQ load is assumed as 800W + j270Var and line impedance is 5+j2Ω and the capacity of NDS is 1,200W. It is confirmed that the load side(customer) voltage is affected by the power factor. Table 3. Voltage variations of lag power factor
Table 4 shows the comparison results between harmonic analyzer and NDS evaluation simulator, when the P load is assumed as 700W and line impedance is 3Ω. The simulation results by changing the capacity of PV systems from 600W to 1,200W, shows that the TDD(Total Demand Distortion) of harmonic depends on the output power of PV inverter. Table 4. Harmonic fluctuations characteristic test data 3th Harmonic(%) Measurement Device PV 600W PV1200W
5th Harmonic(%)
7th Harmonic(%)
Analyzer
Simulator
Analyzer
Simulator
Analyzer
Simulator
7.59 3.27
7.48 3.51
21.24 10.85
21.77 11.3
20.14 1.95
20.71 2.12
94
K.-S. Ryu et al.
5
Conclusions
By performing characteristic tests with NDS simulator using LabVIEW, their effectiveness can be illustrated and demonstrated as follow. (1) It is confirmed that the theoretical value(test devices) and simulation results of NDS evaluation simulator are equivalent through the characteristic tests on Sag and Swell. (2) It is founded that the load-side(customer) voltage is getting bigger than source-side voltage because the reverse power flow of NDS can raise the load-side voltage, called Ferranti phenomenon.
Acknowledgment. This Work was support by the Power Generation & Electricity Delivery of the Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Korea government Ministry of Knowledge Economy. (No. 20101020300430)
References [1] Rho, D., Kim, M.: A study on the Optimal Operation of Line Voltage Regulator (SVR) in Distribution Feeders. In: 2004 IFAC Symposium, Seoul, South Korea (September 2004) [2] Kim, M., Rho, D., Kim, J., Kim, K.: Optimal Operation Method of Multiple Voltage Regulators in Distribution Systems Interconnected with Dispersed Storage and Generation Systems. Trans. KIEE 54A(2) (February 2005) [3] Hartmann, W.: Winding Arrangements for Distributed Generation. On-Peak Performance (December 2002) [4] Burke, J.J.: Power Distribution Engineering, pp. 320–348. Dekker, Inc. (1994) [5] Kirkham, H., Das, R.: Effects of Voltage Control in Utility Interactive Dispersed Storage and Generation Systems. IEEE Trans. Power Apparatus and Systems PAS-103(8) (1984) [6] Chen, C.: The Effect of Voltage Control to Efficiency and Operation of Electric Distribution Systems, Ph.D. Thesis, University of Texas at Arllington (1982) [7] Kim, C., Park, O., Lee, B., Rho, D.: IEEE Optimal Algorithms for Voltage Management in Distribution Systems Interconnected with New Dispersed Sources. IEEE T&D Asia (2009)
Evaluation Algorithm for Bi-directional Protection Coordination in Distribution System with Distributed Generations Sohee Kim, Seongho Gi, Bangwoon Song1, Seungkyou Choi1, and Dae-Seok Rho1 1
Korea University of Technology and Education {sh-kim2579,dsrho}@kut.ac.kr
Abstract. Recently, with the increasing of the interest about global warming, pollutions and so on, a number of distributed generations (DGs) such as photovoltaic (PV) and wind power (WP) are interconnected with distribution systems. However, installing of DGs makes power flow changes such as directions, one-direction to bi-direction, and increasing/decreasing of fault current. Therefore, it may cause the critical problems. This paper proposes an evaluation algorithm for bi-directional protection coordination and presents an evaluation system for protection coordination based on the algorithm. By performing simulations, the result shows that the existing method may cause critical problems, and also the effectiveness of proposed method is verified. Keywords: Distributed Generation (DG), bi-directional protection coordination, Protective devices, Over Current Relay (OCR), Over Current Ground Relay (OCGR).
1
Introduction
Recently, with the increasement of the interest about global warming, pollutions and so on, a number of distributed generations (DGs) such as photovoltaic (PV) and wind power (WP) can be interconnected with distribution systems. However, the power producers with DG and power company like KEPCO (Korea Electric Power Corporation) have sticken to the existing one way protection coordination method because of the lack of technical skills and experiences. Therefore, it may cause the critical problems due to reverse power flow and change of fault current (for example, dividing effect) by the DGs. In order to solve these problems, this paper proposes an evaluation algorithm for bi-directional protection coordination and presents an evaluation system based on the algorithm. Additionally, coordination policies between the protection devices (such as re-closers and relays of substation) of power company and relays of DG producers are reviewed using the proposed system.
<Step 1> Collecting the data of distribution system and DGs. <Step 2> Selecting the suitable type of protection devices. <Step 3> Obtaining the % impedance map of positive and zero sequence and performing load flow and fault current analysis of distribution system with DGs. <Step 4> Reviewing the setting values between forward and back-up protection devices. <Step 5> Checking the TCC (Time-Current Characteristic) curve in order to evaluate the performance of protection coordination whether the operating time difference between forward and back-up protection devices is longer than the guideline value or not. <Step 6> If the protection coordination is not suitable, adjusting the TCC curve with the change of the time difference. <Step 7> After the TCC curve is modified, resellecting the proper protection devices if the protection coordination is still not suitable. <Step 8> Performing the iteration of the above process until the protection coordination is satisfied.
The flowchart of the above algorithm is as shown in Fig.1.
Fig. 1. Evaluation algorithm for protection coordination considering DGs
2.2
Composition of Evaluation System
Fig. 2 shows the main menu of the evaluation system for protection coordination based on the presented algorithm in this paper. It has several icons such as S/S (Substation), R/C (Re-Closer), DG (including customer relay) and so on. This system is made by window-based HMI (Human Machine Interface) program in order to make users comfortable by creating virtual distribution systemsion a easy manner. The evaluation system can automatically calculate the fault current and voltage profile considering DG and also evaluate the protection coordination. This program is supposed to be designed for the customer users like DG producers, not power company.
Evaluation Algorithm for Bi-directional Protection Coordination
97
Fig. 2. Evaluation system for protection coordination
3
Simulation and Analysis
3.1
Model Distribution System with DG
In order to analyze the protection coordination between the power company protection device(re-closer) and customer relay of DG producers, a model distribution system with a wind power system of 3MVA capacity is assumed as shown in Fig. 3. In Korea, re-closer of primary feeder is generally located(installed) at the recommended section which is half of rated load current of a feeder. Accoding to the below quation, the peak load current is calculated by 252A because one primary feeder has 10MVA capacity in normal case. Therefore, re-closer is installed at the point(section) of 126A of load current.
Fig. 3. Model distribution system with DG
98
3.2
S. Kim et al.
Analysis of Simulation Result
First of all, the fault current should be calculated to obtain the setting values of each protection devices (OCR, OCGR, etc.). Fig. 4 is the % impedance map of both positive and zero sequence, in case where a fault is occured at the customer side of DG. By using this % impedance map, theoretical value of three-phase short circuit current is below equation. Fig.5 shows a simulation result of three-phase short circuit current in proposed evaluation system. The theoretical value obtained from % impedance map is equal to the simulation result. Therefore, it is confirmed that the evaluation system can automatically calculate the fault current and voltage profile considering DG and also make users comfortable by creating virtual distribution systems in a easy manner.
Fig. 4. % impedance map of model distribution system with DG
Fig. 5. Simulation result of three-phase short current
The peak load current at re-closer location is 126A, before WP is interconnected, but after WP is connected, it is reduced to around 50A, due to the output of WP, 76A, which is the reverse power flow. Table 1 is re-closer setting values (minimum operation current) recommended by KEPCO guidelines. According to below equation, the minimum operation current difference between the maximum operation current without WP and minimum operation current with WP is up to 360A. Fig.6 is a simulation result of operation time gap(difference) of protection devices based on the TCC curve. In this TCC curve, PU(Pick-Up) which means fault current divided by minimum operation current may be changed from 16.3 to 20.1 due to WP (DG can affect the Operation time of protection devices).
Evaluation Algorithm for Bi-directional Protection Coordination
99
Table 1. Setting Parameter for Re-closer (guidelines of KEPCO) Type Minimum Operation Current
Phase (OCR) Gound (OCGR)
Font size and style Over Maximum Load Current x 2.8 Under Maximum Load Current x 4.0 Maximum Load Current x 0.3
Without DG: 126A×2.8 < setting value < 126A×4.0 With DG: 50A×2.8 < setting value < 50A×4.0 (126A×4.0) – (50A×2.8) = 364A 360A
≒
Fig. 6. Operation time characteristics and PU change between S/S relay and re-closer without (left) and with (right) WP
Fig. 7 shows the operation time characteristics between the re-closer and DG customer’s relay, in case of before and after WP(3MVA). In this simulation, the OCR operation time difference between the re-closer and DG relay is more than 6 cycles (which is guidline value of KEPCO) at 0.5 levers for DG relay (7.3 cycles and 6.8 cycles). Therefore, the protection coordination problem may not be occurred, in the case where DG is interconnected with distribution system.
Fig. 7. Operation time characteristics between re-closer and DG relay without (left) and with (right) WP
100
S. Kim et al.
Table. 2 shows the operation time difference each protection devices in case of before and after WP(3MVA). In this table, the operation time differences satisfy coordination operation time(guidline value of KEPCO). Therefore, the protection coordination problem may not be occurred, in the case where DG is interconnected with distribution system. However, it is found that there are some possibilities to malfunction of protection devices because DG can affect the change of fault current and load current. Table 2. Setting Parameter for Re-closer Capacity of Protection DG Devices S/S Phase (OCR) R/C Ground (OCGR)
R/C DG producer S/S DG producer
4
Operation time difference (Cycles) Without DG
With DG (WP, 3MVA)
17.8
17.2
Coordination operation time (Cycles) 10
40
36
10
Phase (OCR)
6.2
6.8
6
Ground (OCGR)
6.2
7.3
6
Phase (OCR)
20
21.6
17
Ground (OCGR)
27
30
17
Conclusions
There are some possibilities to mal-function of protection devices since DG can effect the change of fault current and load current. But, until now there was no evaluation system for evaluating protection coordination, at customer side, considering DGs. Therefore, this paper proposes an evaluation algorithm for bi-directional protection coordination and presents an evaluation system based on the algorithm. From the simulation result, it is verified that the proposed method is one of useful tools for protection coordination anslysis when DGs are connected. Acknowledgment. This Work was support by the Power Generation & Electricity Delivery of the Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Korea government Ministry of Knowledge Economy. (No. 20101020300430)
References [1] [2] [3]
Rho, D.-S.: An Operation Method for Bi-directional Protection Coordination in Smart Grid with New Energy Sources. KIEE Magazine 5(9), 39–43 (2010) New Energy and Renewable Energy Development, Use, and Spread Promotion Law Transmission & Distribution Laboratory Korea Electric Power Research Institute, Distribution protection descriptions, pp. 70–91 (October 2008)
Evaluation Algorithm for Bi-directional Protection Coordination [4]
[5] [6] [7]
[8]
[9]
[10]
101
Shin, D.-y.: An Analysis and Countermeasure for the Reverse-Power-Flow Phenomena of Distribution System interconnected with Distributed Generations. Chungnam National University (February 2009) Jang, M.-S., Bang, H.-J.: The Current Status and the Prospects of Wind Energy. Journal of the Environmental Sciences 18(8), 933–940 (2009) IEEE 1547.1, IEEE Standard conformance Test Procedures for Equipment Interconnecting Distributed Resources with Electric Power System (June 2005) Rho., D.S., et al.: A study on development of interconnected algorithm of distribution system in photovoltaic. In: Proceedings of the KAIS (The Korea Academia-Industrial Cooperation Society) spring Conference (May 2008) Rho., D.S., et al.: A study on the fault analysis algorithm for dispersed sources of distribution system. In: Proceedings of the KAIS (The Korea Academia-Industrial Cooperation Society) spring Conference (May 2009) Rho., D.S., et al.: A study on protective coordination method in distribution system with new energy sources. In: Proceedings of the KAIS (The Korea Academia-Industrial Cooperation Society) spring Conference (May 2009) (Paper Award) IEEE 1547, IEEE Standard for Interconnecting Distributed Resources with Electric Power System (July 2003)
Development of Remote Monitoring and Control Device for 50KW PV System Based on the Wireless Network Jea-Bum Park1, Byung-Mok Kim, Jian Shen, Sun-Young Kim, and Dae-Seok Rho1 1
Korea University of Technology and Education {pjbwwkd,dsrho}@kut.ac.kr
Abstract. This paper deals with the efficient management for the intelligent distribution system related with the renewable energy sources, using the wirewireless monitoring and control device. This device is mainly composed of 2 sections. One is monitoring device using the Autobase S/W and the other is control device using PLC. This paper proposes a wire and wireless monitoring and controlling device which can monitor and control the 50Kw PV system installing remotely (about 1Km) in campus of the Korea University of Technology and Education. By the analysis of PV output characteristic using the device proposed in this paper, it is confirmed that the device can contribute the establishment of Smart Grid. Keywords: Wired and wireless communications, 50KW PV monitoring and control.
1
Introduction
Globally, Green-Energy industry has already been expected to trigger the Cascade Effect which leads the change of society and culture as other industries go green. Accordingly, developed nations have arranged the active plan of promotion, and have continuously supported and invested for a long time. Also, this situation is what Clean Edge, Inc. which is the specialized research institution in USA expects the Green energy to be highly developed that the growth of Green energy market would be up to 15.1% as much as the IT Industrial Revolution (growth in early 2000, 17%). Under these backgrounds, this paper proposes one of several solutions that collect and analyze the real time output data from solar power system by using the wire-wireless monitoring and controlling device. It is expected to overcome the operating system problems caused by the installation of remote locations like mountain and seaside.
Development of Remote Monitoring and Control Device for 50KW PV System
103
with distribution systems, may cause a terrible change of power output under the conditions of weather (windy, rainy), and lead to some problems with voltage management of system operation of the distribution system and quality of power. Therefore, this paper proposes a wire and wireless monitoring and controlling device which can monitor and control the 50Kw PV system that installed remotely (about 1Km) in campus of the Korea University of Technology and Education as shown in Fig.1.
Fig. 1. Concept of 50Kw PV system
3
Configuration of Wire-Wireless Monitoring and Control System
The wire-wireless monitoring and control system is mainly consist of 2 sections as in Fig.2. One is the monitoring and controlling section for receiving data and the other is transmitting section for collecting and sending data. The former is made up of HMI software and router, and the latter is made of 50KW inverter, PLC(Program Logic Controller), Relay and MC elements.
Fig. 2. Configuration of wire-wireless monitoring and controlling device
104
3.1
J.-B. Park et al.
Monitoring Device Using HMI Program
As shown in Fig. 3, the monitoring device using HMI program is designed to communicate the protocol of components which are wire-wireless converter and router and Autobase S/W.
Fig. 3. Concept of monitoring device
3.2
Control Device Using PLC
Fig.4 shows the control device is made of PLC(Program Logic Controller), Relay and MC elements for 50KW PV system. PLC plays a main role at the control device, which turns on the output contact of LADDER by operating internal contact of PLC from the control signal of the Autobase(S/W).
Fig. 4. Concept of control device
3.3
50Kw PV System with 3 Phase Inverter
As shown in Table 1, the input and output data which of PV inverter used in this test for grid connection is transferred by using the RS232 serial communication and itself protocol.
Development of Remote Monitoring and Control Device for 50KW PV System
105
Table 1. Contents of 50KW inverter Protocol
4
Implementation of Monitoring and Controlling Device Using Wireless-Network
4.1
Real Time Remote Monitoring Device
The concept for data flow of monitoring device is shown in Fig. 5. First, the data of the 50Kw solar power system is transferred to the wireless converter of CSW-H80 by transforming into Serial communication. The transferred data is transferred to the monitoring place wirelessly by transforming into TCP/IP. And the data is received by the router at the monitoring place, then is transferred by transforming from Ethernet into Serial by using wire converter of CSW-H20-2.
Fig. 5. Algorithms of monitoring device using wireless-network
4.2
Real Time Remote Control Device
The concept for data flow of control device is shown in Fig. 6. At first, the signal to control ON/Off switch of the 50Kw solar power system is produced by HMI S/W. The control signal is transformed from Serial communication into Ethernet communication with CSW-H80, and is sent the transformed signal to the PV system in a wireless manner. And the signal is transferred to PLC to control the ON/OFF switch like MC. Then the 24V of output voltage by using LADDER program of PLC is sent into the power of relay.
106
J.-B. Park et al.
Fig. 6. Algorithms of controlling device using wireless network
4.3
Confirmation of Communication
Fig 7 shows that the control device proposed in this paper is operated well by checking whether the MC works or not from the voltages of MC using MyDAQ.
Fig. 7. Test results of control device
5
Analysis of Simulation
Fig. 8 shows the main menu of remote monitoring and controlling system using HMI S/W. The panel can monitor the weather conditions and the 3 phase voltage and current from 50Kw solar power system, and also, monitor the frequency and power factor of each phase. The panel can also control the ON/OFF switch of PV system to solve the maintenance and management problems.
Development of Remote Monitoring and Control Device for 50KW PV System
107
Fig. 8. Monitoring panel using HMI S/W
Fig 9 shows that the measured data from 50KW inverter of PV systems is equals to the data from monitoring system proposed in this paper. Therefore it is verified that the monitoring and control system is prctical system.
Fig. 9. Comparison between data of inverter and Autobase(S/W) through wireless network
Fig. 10 shows the daily output pattern of PV systems by using HMI S/W. he output power of PV system is analyzed as the highest value from 12 PM to 3 PM, by using the saving data function of HMI S/W. And it is confirmed that predicting the output pattern of PV system through the proposed system can be effectively performed.
108
J.-B. Park et al.
Fig. 10. Daily output pattern of solar power system
6
Conclusions
This paper proposes a wire-wireless monitoring and controlling device which can monitor and control the 50KW solar power in a remote manner. In purpose of remote communication, used directional and omnidirectional antenna to monitor and control the PV source 1KM away. And it is confirmed that predicting the output pattern of PV system through the proposed system can be effectively performed and this system can contribute to the infra construction of SmartGrid. Acknowledgment. This Work was support by the Power Generation & Electricity Delivery of the Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Korea government Ministry of Knowledge Economy. (No. 20101020300430)
References [1] Smith, T.F., Waterman, M.S.: Identification of Common Molecular Subsequences. J. Mol. Biol. 147, 195–197 (1981) [2] May, P., Ehrlich, H.-C., Steinke, T.: ZIB Structure Prediction Pipeline: Composing a Complex Biological Workflow Through Web Services. In: Nagel, W.E., Walter, W.V., Lehner, W. (eds.) Euro-Par 2006. LNCS, vol. 4128, pp. 1148–1158. Springer, Heidelberg (2006) [3] Foster, I., Kesselman, C.: The Grid: Blueprint for a New Computing Infrastructure. Morgan Kaufmann, San Francisco (1999) [4] Czajkowski, K., Fitzgerald, S., Foster, I., Kesselman, C.: Grid Information Services for Distributed Resource Sharing. In: 10th IEEE International Symposium on High Performance Distributed Computing, pp. 181–184. IEEE Press, New York (2001) [5] Foster, I., Kesselman, C., Nick, J., Tuecke, S.: The Physiology of the Grid: an Open Grid Services Architecture for Distributed Systems Integration. Technical report, Global Grid Forum (2002)
Development of Remote Monitoring and Control Device for 50KW PV System
109
[6] National Center for Biotechnology Information, http://www.ncbi.nlm.nih.gov [7] IEEE 1547, IEEE Standard for Interconnecting Distributed Resources with Electric Power System (July 2003) [8] IEEE 1547.1, IEEE Standard conformance Test Procedures for Equipment Interconnecting Distributed Resources with Electric Power System (June 2005) [9] Saadat, H.: Power System Analysis. McGraw-Hill (1999) Printed in Singapore
Operation Method for Protection Devices at the Primary Feeders with New Energy Sources Using EMTDC/PSCAD Hasan1, Dae-Seok Rho1,*, Chanhyeok Kim1, and Yongpeel Wang2 1
Abstract. This paper deals with the optimal operation method for the malfunction of protective devices at primary feeder in distribution systems when new power sources like photovoltaic (PV) systems are interconnected, based on the symmetrical components of short circuit studies and the EMTDC/PSCAD S/W. When new power sources are considered to be interconnected to distribution systems, bi-directional power flow and interconnection transformer connection of new power sources may cause the operation problems of protective devices (mainly re-closer), since new power sources can change typical characteristics of distribution systems. Therefore, this paper shows an analysis skill of the mal-functional mechanism of protective relay and proposes the optimal solution for the mal-function problem using the symmetrical components of fault analysis and the EMTDC/PSCAD simulation. And, this paper also shows the effectiveness of proposed method by the simulation at the field distribution systems. Keywords: Protective devices, New energy sources, Protective coordination, Symmetrical components, Re-closer, EMTDC/PSCAD.
1
Introduction
As power systems have been deregulated and decentralized thanks to the technology development making small-sized generators more competitive and reliable, new energy sources (NES) such as photovoltaic cell, solar energy, fuel cell, wind power, energy storage, and etc. have been actively interconnected and operated in distribution systems. Moreover, since unstable oil price and environmental consciousness are strongly promoting higher penetration level of such renewable energy in these days, much more NES are expected to be interconnected into distribution systems. In several countries, the protective coordination guidelines on the interconnection of NES have been established and even special laws also have been enacted to promote the interconnection of NES. However, the bi-directional power flow and *
Operation Method for Protection Devices at the Primary Feeders
111
interconnection transformer connection of NES may cause the various operation problems of protective devices which are mainly re-closer, because NES can change typical operation characteristics of protective coordination in distribution systems. The reverse power flow and connection type of NES have severe impacts on typical distribution systems, for example power quality problems, protection coordination problems, and so on. Specially, it is frequently reported in the field systems that the protective devices (Re-closer) located at the un-faulted primary feeder are malfunctioned when a single line ground fault (SLG) is occurred at a primary feeder of same bank (Main Trans-former: M.Tr). The malfunction of protective device could result the severe damage of many customers like a several interruptions. Therefore, this paper proposes an analysis skill for the mal-functional mechanism of protective relay considering the abnormal operation characteristic of NES and presents the countermeasure for the mal-function problem through the modeling and numeric calculation by both using symmetrical components of fault analysis the EMTDC/ PSCAD simulation.
2
Mal-Functional Cases of Protection Device in Distribution Systems
Fig. 1 shows the configuration of "BUK-PYEONG" substation (154/22.9KV) of "DONG-HAE" branch, "KANG- REUNG" office, in Korea Electric Power Corporation (KEPCO). As shown in Fig.1, 1MW Photovoltaic (PV) system is located at the "BUKPYEONG" primary feeder, and the connection type of inter-connection transformer is Yground-Delta connection. Many cases of malfunction of Re-closer located at the unfaulted primary feeder are reported in the time when a single line grounded fault is occurred at the other primary feeders at the same bank. In other words, when a single line grounded fault by external contact, lighting arrestor (LA) damage and burn-out of transformer at "SAM-WHA", "DONG-SAM" and "EO-DAL" primary feeder is happened, Re-closer located at the "BUK-PYEONG "primary feeder, un-faulted primary feeder, is operated, so called, malfunctioned. So, many customers of the primary feeder suffer severe interruption problems. Actually, it is reported that the OCGR (Over Current Ground Relay) of Re-closer at "BUK-PYEONG" is operated and the neutral current (N phase) was measured as 130 ~ 400A.
Fig. 1. Distribution system in “BUK-PYEONG” substation
112
Hasan et al.
3
Analysis for Mal-Function of Protective Devices
3.1
Analysis Based on the PSCAD/EMTDC
Main Transformer (154/22.9KV) of "BUK-PYEONG" substation and the modeling based on the PSCAD/EMTDC is represented in Fig. 2. Fig. 3 shows the % impedance data of several primary feeders and the modeling is represented in Fig. 3. And also, Fig. 4 describes the modeling of PV system through inverter using the PSCAD / EMTDC. Fig. 5 shows modeling for distribution system using PSCAD/EMTDC.
Fig. 2. Modeling for Main Tr. in "BUK-PYEONG" substation using PSCAD/EMTDC
Fig. 3. % Impedance data of distribution systems
Fig. 4. Modeling for PV systems using PSCAD/EMTDC
Operation Method for Protection Devices at the Primary Feeders
113
Fig. 5. Modeling for distribution system using PSCAD/ EMTDC
3.2
Analysis Based on the Symmetrical Component Method
Modeling for PV System and Distribution Systems To make modeling for PV systems and distribution systems, the impedance configuration of field system shown as Fig. 1 and specification of main transformer in Table 1 are collected and arranged from the field sites. By using symmetrical components, the equivalent circuits of zero phase sequence impedance and positive /negative phase sequence impedance are presented in Fig. 6 and Fig. 7. Table 1. % Impedance Data of "BUKP-YEONG" Substation % Impedance of Transformer Bank
NGR Z12
Z23
Z31
(45MVA)
(15MVA)
(15MVA)
15.970
2.230
8.460
35.499
14.867
56.400
100MVA
15.970
6.690
25.380
45MVA
Base
(2nd)
Selfcapacitance
0.6Ω
#3 11.14 Line Connection
Y – Y - Delta
Fig. 6. Zero phase sequence component of equivalent circuit
Analysis Formal-Function of Re-closer Based on the above equivalent sequence circuit by using symmetrical methods, the single line fault (SLG) currents of each primary feeder at the same bank are acquired as following steps. Firstly, the zero phase sequence impedance at the faulted point of "SAM-HWA" primary feeder is obtained as :
And, the positive and negative phase sequence impedances are calculated as follows.
Thus, SLG current at the faulted point is
And, also the SLG current (neutral line current) at the un-faulted primary feeder of "BUK-PYEONG" which is interconnected 1MW PV systems, can be obtained by using current divider theory as follows. Namely, the SLG current on neutral line of "BUK-PYEONG" primary feeder can be calculated as about 123A to 493A depending on the position of grounding fault at the faulted feeder of "SAM-HWA". Thus, the possibility of malfunction of re-closer located at "BUK-PYEONG" primary feeder can be checked, because N phase OCGR of re-closer is normally set to 70A less than
Operation Method for Protection Devices at the Primary Feeders
115
the SLG current on neutral line. It was confirmed that this was not due to the influence of fault current supplied by photovoltaic systems, which is about 30A which is about 1.5 times of maximum rated current, but due to connection type of interconnection transformer of NPS. This paper shows that protective device (OCGR) is malfunctioned by grounding current on N phase flowing through the zero phase sequence path supplied by Yground-Delta connection of the transformer of NPS.
4
Countermeasure for the Malfunction of Protection Devices
4.1
Analysis by the Symmetrical Method
As eliminating the zero phase sequence current path of interconnection transformer is the best method for malfunction of protective device, this paper presents the method to reduce fault current of neutral line at the un-faulted primary feeder under 70A by inserting proper neutral grounding resistance (NGR) to connection line of the transformer. Namely, it is supposed that NGR of 100Ω(300×19.1%) is inserted to neutral line by changing connection type of the transformer from Y-ground-Delta to Y-high resistance ground-Delta connection. By considering this concept, the equivalent circuit of zero phase sequence impedance and positive/negative sequence impedance are described as Fig. 8 and Fig. 9. Thus, the SLG current (neutral line current) at the un-faulted primary feeder of "BUK-PYEONG" can be obtained by the equivalent circuit.
Namely, it is found that if Y-high resistance grounding - delta connection type is applied; the malfunction of protective device (re-closer) can be prevented by remarkably reducing the ground fault current of neutral line as only 13A. This paper shows that the inserting NGR to interconnection transformer is one of the effective counter -measures. 4.2
Analysis by the PSCAD/EMTDC
Simulation Results without PV System Fig. 10 shows the simulation results for a single grounded fault, in the case where 1MW PV system is not interconnected with "BUK-PYEONG" primary feeder. This paper assumes that a single grounded fault occurs at 0.2s and the duration time is 0.2s, by the modeling based on the PSCAD/EMTDC. The results show that the fault current (zero sequence current) at the "BUK-PYEONG" primary feeder is only around 12A, so the re-closer located at the feeder is not operated (malfunctioned) because the OCGR setting value of the re-closer is set as 70A.
116
Hasan et al.
Fig. 8. Zero phase sequence component of equivalent circuit
Fig. 10. Single grounded fault currents OF "BUK-PYEONG" feeder for primary feeders without PV system
Simulation Analysis with PV System This paper considers two simulation cases for the fault locations. One is a sending point near the substation; the other is a ending point of last customer. In the Fig. 6, the sending point and ending point on "SAM-WHA" primary feeder is considered as the locations of single grounded fault. Fig. 11 and Fig 12 show the simulation results for a single grounded fault of "SAM-WHA" primary feeder, in the case where 1MW PV system is interconnected with "BUK-PYEONG" primary feeder at the same bank. This paper assumes that a single grounded fault occurs at 0.2s and the duration time is 0.2s by modeling based on the PACAD/EMTDC. The results show that the fault current (zero sequence current) at the "BUK-PYEONG" primary feeder is around 160A to 400A, depending on the fault location, sending point and ending point. The simulation shows that the re-closer located at "BUK-PYEONG" primary feeder can be operated (malfunctioned) because the fault current exceeds the OCGR setting value of the re-closer (70A). Therefore, we confirm that interconnection with PV system can cause the problems for the malfunction of re-closer.
Operation Method for Protection Devices at the Primary Feeders
117
(a) Single grounded fault current of "BUK-PYEONG" substation (main transformer)
(b) Single grounded fault current of "SAM-WHA" feeder
(c) Single grounded fault current of "BUK-PYEONG" feeder. Fig. 11. Single grounded fault currents for primary feeders with PV system
Fig. 12. Single grounded fault neutral currents for primary feeders with PV system (malfunction of re-closer)
118
4.3
Hasan et al.
Countermeasure for Malfunction of Re-closer
The simulation results show that the fault current (zero sequence current) at the "BUK-PYEONG" primary feeder is around 160A to 400A, in the case where 1MW PV system is interconnected with "BUK-PYEONG" primary feeder. The fault current of PV system actually can be provided as small as 1.5 times of a rated current (about 32A). So, the simulation results show that the Y-grounded-delta connection type of interconnecting transformer offers a passage of zero sequence current. To solve this problem, this paper assumes that the path of zero sequence current is reduced by inserting NGR at the Y-grounded-delta connection of the transformer. As shown in Fig. 13, The results for the countermeasure show that the fault current (zero sequence current) at the "BUK-PYEONG" primary feeder is only around 10 - 15A, so the recloser located at the feeder is not operated (malfunctioned) because the fault current is not reached to OCGR setting value of the re-closer(70A).
Fig. 13. Single grounded fault current of "BUK-PYEONG" feeder with NGR
5
Conclusions
This paper proposes the mechanism of mal-functional problem of protective devices and the countermeasure for the problem by using the symmetrical component method and PSCAD/EMTDC. The results are summarized as follows. (1) It is confirmed that if ground fault occurs in other primary feeder of same bank in the distribution substation, very high ground fault current flows in the neutral line(N phase) of un-faulted primary feeder which is interconnected with NPS, thereby causing OCGR of the feeder to result in malfunction. (2) It is confirmed that the fault current flowing on the neutral line is not affected by fault current by NPS, but by zero phase sequence current path of interconnection transformer of NPS. (3) This paper shows that protective device (OCGR) is malfunctioned by grounding current on N phase flowing through the zero phase sequence path supplied by Yground-Delta connection of the transformer of NPS. (4) This paper shows that the inserting NGR at the Y-grounded-delta connection of the transformer can reduce the ground fault current of neutral line. It is confirmed that
Operation Method for Protection Devices at the Primary Feeders
119
the inserting NGR to interconnection transformer is one of the effective countermeasures for the malfunction of protective device (re-closer). Acknowledgment. This Work was support by the Power Generation & Electricity Delivery of the Korea Institute of Energy Technology Evaluation and Planning (KETEP) grant funded by the Korea government Ministry of Knowledge Economy. (No. 20101020300430)
References [1] Nagpal, M., et al.: Dispersed Generation Interconnection – Utility Prospective. IEEE Trans. on Industry Application 42(3) (2006) [2] Interconnected the technique basis in distribution system of distribution generations, Korea Electric Power Corporation (KEPCO) (April 2004) [3] Tech Spec. of interconnection of Japanese distributed power system. Japan Electrics Organization (2001) [4] Guide of technique requisites of interconnection of distribution of Japanese system, Japan (1998) [5] Work procedure and technique guide of interconnection of distributed power of Hokkaido, Japan (2002) [6] Technical system of distribution system of distributed power of Hokkaido, Japan (2004) [7] Rho., D.S., et al.: A study on development of interconnected algorithm of distribution system in photovoltaic. In: Proceedings of the KAIS (The Korea Academia-Industrial Cooperation Society) spring Conference (May 2008) [8] Rho., D.S., et al.: A study on the fault analysis algorithm for dispersed sources of distribution system. In: Proceedings of the KAIS (The Korea Academia-Industrial Cooperation Society) spring Conference (May 2009) [9] Rho., D.S., et al.: A study on protective coordination method in distribution system with new energy sources. In: Proceedings of the KAIS (The Korea Academia-Industrial Cooperation Society) spring Conference (May 2009) (Paper Award) [10] IEEE 1547, IEEE Standard for Interconnecting Distributed Resources with Electric Power System (July 2003) [11] IEEE 1547.1, IEEE Standard conformance Test Procedures for Equipment Interconnecting Distributed Resources with Electric Power System (June 2005) [12] Saadat, H.: Power System Analysis. McGraw-Hill (1999); Printed in Singapore
The Design of AES-DN Embedded Cryptographic Processor for the RFID/USN Environment Seon-Keun Lee1 and Sun-Yeob Kim2,* 1
Faculty of Materials & Chemical Engineering, Chonbuk National University Jeonju, Korea [email protected] 2 Department of Information and Communication Engineering, Namseoul University Cheonan, Korea [email protected]
Abstract. With the widespread use of embedded system based ubiquitous system, there is a rapid increase of the RFID/USN environment.. However, the security for the RFID/USN environment has been far behind the speed of rising security threats. In this regard, this study proposed a AES-DN embedded encryption system for RFID/USN. Compared to the existing AES, the proposed AES-DN processes 120% faster and is equipped with the DN-R method reducing node loads, which is most important for the RFID/USN environment. Therefore AES-DN is considered to be an ideal algorithm to address limited resources of the RFID/USN environment. Keywords: RFID/USN, AES, High Speed, network management, Condition State.
1
Introduction
Rapid development of embedded processors has led to the widening use of RFID/USN, raising the importance of security of node networks. Although diverse security methods have been applied to ubiquitous based RFID/USN, ways to protect from various attacks are still insufficient[1]. In this respect, this study aimed to provide enhanced processing capacity and better security for the AES algorithm standardized in 2000 to improve the performance of encryption algorithm for RFID/USN. The proposed method is a AES-DN(AES with Distributed Network) technique to improve processing capacity and ensure easy management of the network for AES, and we carried out an analysis of its performance by comparing to AES and other symmetric encryption systems[2,3]. As the RFID/USN environment has many sensor nodes, encryption algorithm used for the RFID/USN environment should have a distributed processing feature. For this, the proposed AES-DN encryption algorithm used PRN (Pseudo-Random Number) *
The Design of AES-DN Embedded Cryptographic Processor
121
that performs bit and byte operations together and divides a cycle by N in order to manage networks more easily unlike the existing symmetric encryption algorithm and improve processing rate. Therefore, we believe that AES-DN will become one of the indispensible encryption systems for the RFID/USN environment with complex node networks and in the applications of embedded processor[4,5].
2
The AES-DN Cryptographic Algorithm
Even though there are a variety of methods to strengthen the security for RFID/USN, general encryption algorithms cannot be applied to the RFID/USN environment due to limited environmental characteristics. In this paper, we proposed a AES-DN method to solve realization problems in consideration of processing time, consumption power, and size and ensure the efficiency of security resources amid growing nodes and applied the method to AES Rijndael encryption algorithm. The proposed AES-DN method uses exclusive OR as a fundamental operator like existing AES and byte as a processing unit. Byte operations have a highly faster processing speed and are excellent in raising security as it is hard for back tracking. Since the method also uses existing Feistel and SPN structure together, encryption and decryption occur at the same time, the efficiency does not decrease, unlike Rijndael[2,3] or Serpent[6,7] encryption algorithm. These features fully meet the prerequisites for AES and presents ideas for the next version of AES, and therefore, encryption algorithm covering DN can solve real time processing and security, realization and management problems caused by growing nodes. Figure 1 show how PRN is formed by means of a register with an arrangement using GF 2 .
Fig. 1. 2-DSRS
For 2-DSRS, in the state of during every event, mn–dimensional column vector is expressed in Equation (1): (1)
122
S.-K. Lee and S.-Y. Kim
Here, column sub-vector
is expressed in Equation (2): (2)
In addition, matrix
,
state transformation matrix consists of return . If event time is 0, the next state of s function is expressed in
Equation (3). 1
(3)
From these characteristics, it is possible to build a safe network including all nodes in a single PN structure regardless of the number of nodes by setting s for sensor node ID after applying sub-vector s whenever state changes prior to the iteration of AES. Primitive polynomial used for this type of PRN distributed processing is as in Equation (4) [8]. 1
(4)
1 The same as PRN used for CDMA mobile communication networks, Equation (4) generates output, regarding two bytes as a single equation. Existing AES performs rounds n times following exclusive OR with initial round key. Upon completion of all rounds, encryption/decryption ends. Here, DN-PRNG gives sub-codes to each round for distributed processing of PN code. The entire flow of this AES-DN encryption algorithm is shown in Figure 2. Most importantly, AES-DN encryption algorithm performs encryption and decryption at the same time as shown in Figure 2. Control signals determine encryption and decryption mode, and the operations occur reversely and sequentially. Data of Inv/SubByte, Inv/ShiftDiagonal, Inv/MixColumn, Inv/AddRoundKey for every round works separately by DN-PRN, and the number of rounds determines security. Therefore, every round occurs independently. These four transformations (DN-R : DN PRNG-Round) occur as in Equation (5): Existing AES changes data and key length into 128, 192, and 256 bits, the optimal number of rounds is determined depending on the length changed. Accordingly, in case of existing AES, the size of data block determines the number of rounds. Therefore, if data symbol size is known, the number of rounds is revealed, and DC and LC of the number of rounds, key, and data makes cracking possible. Nevertheless, as data content of AES-DN changes depending on the number of rounds even if block and key size is fixed, the number of rounds cannot be measured by data symbol size alone[5,9]. For the first time in the DN-R transformation, Inv/SubByte transformation changes nonlinearly independent bytes and thereby forms a set of changed nonlinear bytes.
The Design of AES-DN Embedded Cryptographic Processor Plain/Ciphertext
123
Key
DN-PRNG
AddRoundKey
Key Scheduler
ByteSub ShiftRow MixColumn
Round key Round 1 key 2 Round key 9
AddRoundKey
ByteSub ShiftRow
Round key 10
AddRoundKey
Cipher/Plaintext
Fig. 2. AES-DN Cryptographic Algorithm
/
/
/ /
/
(5)
/ S-box used for Inv/SubByte transformation is a set of functions that allow for inverse transformation and nonlinear transformation which applies affine transformation to GF(2 ) which has the inverse transforms of the multiplication in Finite Field GF(2 ) and is defined as Equation (6), where refers to the bit block of which transforms internal information out of output information of DN-PRNG, in case of 0 7, refers to the i th bit of independently working byte, and refers to the i th bit of a certain byte block. (6)
where refers to the bit block of which transforms internal information out of output information of DN-PRNG, in case of 0 7, refers to the i th bit of independently working byte, and refers to the i th bit of a certain byte block. Although 8 works by byte for operation, actual operation is based on bit.1 though 8 works by byte for operation, actual operation is based on bit. Therefore, to represent bits and bytes separately, moduler operations are expressed on Finite Field GF(2 ), which indicates that AES-DN encryption algorithm is based on 1-byte operations.
124
S.-K. Lee and S.-Y. Kim
Secondly in the DN-R transformation, Inv/ShiftDiagonal transformation occurs at the same time diagonally from columns and rows of data state arrangements. This transformation is expressed in Equation (7). 03,12,21,30 13,22,31,00 23,32,01,10 33,02,11,20
00,11,22,33 01,12,23,30 02,13,20,31 03,10,21,32
(7)
Equation (7) shows that columns and rows of each CS are transformed diagonally. This diagonal transformation makes it easy to perform inverse substitution and calculation and guarantees randomization. Consequently, compared to Rijndael or Serpent encryption algorithm, Inv/ShiftDiagonal transformation is twice faster and doubles security. Since Inv/ShiftDiagonal transformation uses diagonal transformation, inverse transformation occurs in the same way. Thirdly in DN-R transformation, Inv/MixColumn transformation forms a newly transformed arrangement by multiplying a fixed polynomial expression . Here, the entire transformation equation is basically a multiplication operation. Namely, transformed function is formed by multiplying the function before transformation by . The is as in Equation (8). multiplication 02 01 03 03
01 03 03 02
03 03 02 01
03 02 01 03
(8)
In Equation (8), like Inv/ShiftDiagonal transformation, diagonal transformation occurs but, Inv/MixColumn transformation carries out a diagonal transformation for only. Namely, in case of performing diagonal transformation of AES-DN encryption algorithm, the diagonal transformation is differentiated as in Equation (9). / /
(9)
As in Equation (9), if diagonal transformation is performed in the directions of both and for a single CS, all the diagonal angles are transformed. This diagonal transformation can reduce the inconvenience of performing multiplication operation like existing encryption algorithm. Lastly, AddRoundKey transformation of the DN-R transformation performs the binary addition of round key and DN-PRNG. Each Round Key generates different value whenever succession happens from key schedule, and operations of the generated values are performed as in Equation (10), ,
,
,
,
,
,
(10)
refers to the range of performing Round for DN-PRNG. Even where though frequency of operations vary by the range of Round in AddRoundKey
The Design of AES-DN Embedded Cryptographic Processor
125
transformation, simple multiplication operations are performed in the same volume. General encryption algorithm performs key generation algorithm and key scheduling, which is absolutely necessary for the safety of algorithm to create complicated keys. The AES-DN encryption algorithm does not use the method that changes key scheduling by round, allowing for independent operations for every round by having DN-PRNG set the range of pseudo noises. In this way, different keys for individual rounds can be created efficiently, and if the created keys are used for encryption, there is no need to set a separate reference point to differentiate rounds. Changing values of every PRNG event are used to create keys, and internal encrypt text changes in DN-R block, or at each of the four stages of transformation. PRN values control transformation. Therefore, in other encryption algorithm, the length of keys is extended or shorten to adjust the size of keys and encrypt text, but AES-DN encryption algorithm does not require any adjustment of key length.
3
Design of AES-DN Cryptographic System and Simulation
AES-DN encryption algorithm was realized in top-down manner using VHDL, Synopsys Design Analyser Ver. 1999.10, QUARTUS 7.0 was used for circuit composition, and Synopsys VHDL Debegger, ModelSim 5.8C was hired for experiments. For a test-bed for the realization, ALTERA Cyclone EP1C6Q240C8N device was used. The DN-PRNG processor as a key part of the proposed AES-DN encryption algorithm serves to carry out actual AES-DN encryption algorithms. DN-R block creates condition state from data input and CS of AES-DN encryption algorithm, and s(t) as arrangements of the created condition state goes through a byte substitution. In regards to input data 128 bit, the key value is also 128 bit, and some data of the key value 128 bit is used as DN-PRNG SEED following the binary multiplication with input data. DN-R block inside the AES-DN encryption system performs diagonal transformation in both directions of , at the same time. The DN-R block works for Inv/SubByte, Inv/ShiftDiagonal, Inv/MixColumn, AddRoundKey blocks sequentially. Following the operation of AddRoundKey in conditional state, actual data transformation is carried out; after a byte transformation in the Inv/SubByte block, Inv/ShiftDiagonal block performs diagonal transformation, followed by diagonal transformation carried out once again by Inv/MixColumn block. By encryption/decryption mode, RN-R receives key information in the form of SEED value and generates random key information, which undergoes preliminary SEED operations with part of encrypt/decrypt text. The key information following SEED formating process is used as an input of AddRoundKey block in the DN-R block. 7 Here, rounds of AES-DN encryption system are set to 10 times. Following the ten operations, round screening and final substitution entails the creation of decrypt text or encrypt text. In this way, keyscheduler enters key information of DN-R block and creates random key information. AES-DN encryption system produces 128-bit output for 128-bit input and plain text or encrypt text for encrypt text or plain text. The key
126
S.-K. Lee and S.-Y. Kim
information necessary for encryption/decryption is not needed for actual encryption/ decryption, but serves as a basic data to form new key information. Figure 3 shows the entire circuit composition chart of the AES-DN encryption system AES-DN encryption system including DN-R block performs encryption by sending independent pseudo noises depending on the range of 2-DSRS.
Fig. 3. Overall block diagram of cryptographic processor
Table 1 compares existing encryption system and the proposed AES-DN encryption system [10-16]. As outlined in Table 1, AES-DN encryption system processed 1.2 times faster than the existing symmetric block encryption system, AESDN encryption system is advantageous among because it can perform independently encryption of as many sensor nodes as by s , of DN-PRNG. Table 1. The evaluation of performance for AES-DN @50MHz
The Design of AES-DN Embedded Cryptographic Processor
127
These findings reveal that AES-DN encryption algorithm is more efficient in an environment with limited resources like RFID/USN than the existing block encryption algorithms.
4
Conclusion
With the widespread use of embedded system based node networks such as RFID/USN, there is a increasing need for the security accordingly. However it is hard to select encryption algorithm to maintain security in the RFID/USN environment due to limited resources. In this study, we proposed AES-DN block encryption algorithm to address the problem of limited resources of RFID/USN. The proposed AES-DN encryption system has internal information only as a resource for encryption. In addition, processing time and security of the system increased 120% thanks to its all-at-once operations, compared to existing symmetric encryption system. By including DN method that enables distributed processing, which is most important, we realized asymmetric features using PRNG capacity, which is ideal for the RFID/USN environment, a distributed network with high node load, not using complicated asymmetric encryption system. Compared to existing encryption algorithms, the proposed AES-DN encryption algorithm has a higher transmission and system efficiency. Since it is a structural base algorithm requiring no key determination for some time, system complicatedness is extremely low, and processing is faster. To conclude, the proposed new AES-DN encryption algorithm is considered to be fit to overcome limited resources of the RFID/USN environment and serve as an encryption algorithm replacing asymmetric encryption system.
References 1. Kim, C.-W.: Ubiquitous environments in the information protection technology (November 2008), http://www.eic.re.kr 2. Ferguson, N., Kelsey, J., Lucks, S., Schneier, B., Stay, M., Wagner, D., Whiting, D.L.: Improved Cryptanalysis of Rijndael. In: Schneier, B. (ed.) FSE 2000. LNCS, vol. 1978, pp. 213–230. Springer, Heidelberg (2001) 3. NIST, AES Algorithm (Rijndael) Information, http://csrc.nist.gov/archive/aes/rijndael 4. NIST, pseudo-random number generator, http://www.itl.nist.gov/div897/sqg/dads/HTML/ pseudorandomNumberGen.html 5. Gottfert, R., Neiderreiter, H.: On the Linear Complexity of Products of Shift Register Sequences. In: CRYPTO 1993, pp. 151–158. Springer, Heidelberg (1994) 6. Damaj, I., Itani, M., Diab, H.: Serpent Cryptography on Static and Dynamic Reconfigurable Hardware. In: IEEE International Conference on Computer Systems and Applications, Aiccsa, pp. 680–684 (2006) 7. A candidate block cipher for the advanced encryption standard, http://www.cl.cam.ac.uk/~rja14/serpent.html
128
S.-K. Lee and S.-Y. Kim
8. Lee, W.C.Y.: Overview of Cellular CDMA. IEEE Trans. Vehicular Tech. 40, 291–302 (1991) 9. Matsui, M.: Linear Cryptanalysis Method for DES Cipher. In: Helleseth, T. (ed.) EUROCRYPT 1993. LNCS, vol. 765, pp. 386–397. Springer, Heidelberg (1994) 10. Microelectronic Systems Laboratory, Implementation of DES Algorithm Using FPGA Technology (2002), http://www.alagger.com/des-vhdl/report.pdf 11. DES and 3DES cores, http://www.heliontech.com/des.htm 12. SEED block cryptographic algorithm, http://service2.nis.go.kr/pw_certified/seed.jsp 13. Nyberg, K., Knudsen, L.R.: Provable security against a differential attack. Journal of Cryptolog 8(1), 27–37 (1995) 14. Silverman, R.D.: A cost-based Security analysis of symmetric asymmetric Key lengths. Bulletin 13, RSA Lab (2000) 15. NIST, Draft FIPS for the AES (February 2001), http://csrc.nist.gov/publications/drafts.html 16. Silverman, R.D.: A cost-based Security analysis of symmetric asymmetric Key lengths. Bulletin 13, RSA Lab (2000)
Validity of Developmental Delay Screening Test; Using Latent Class Model Seong-A. Lee1, Chung-Yill Park2, and Kang-Suk Lee2 1
Occupational Therapy, SoonChunHyang University, Shinchang-myun, Asan-si, Choogchungnam-do, Korea [email protected] 2 Prevent medicines Catholic University, Seoul, Korea {cypark,leekang}@catholic.ac.kr
Abstract. The purpose of this study was measured and compared validity including sensitivity, specificity and positive predictive values using latent class analysis. Data were collected from 229 children from 3 to 5 years old in local nurseries whose parents agreed with the purpose of study. Especially this study, validity of each test was measured for three developmental screening tests after establishing the last one developmental screening test as a standard and utility of the four screening tests was calculated four two latent class model. When a child was included in the developmental delay class through a latent class analysis. Positive predictive values were 86%-54% in K-DIAL3, K-DDSTII, K-ASQ and K-CDI respectively and those of combination of two of them were 95%-78%. It was suggested that tests of children of two or more rather than one and shall be used to of developmental screening tests for detecting developmental delay. Keywords: Latent Class Analysis, Positive Predictive Value, Effectiveness of developmental screening tests.
predictive value of test tools regarding use of screening test tools are being required. A latent class analysis is a statistical method which classifies potential classes with values of several observation variables and which finds out relations between each observation variable and potential classes and in this study, a latent class analysis was used for confirming utility regarding screening test tools [2]. A latent class analysis was used in calculating positive predictive values of developmental screening tests. A latent class analysis is generally used in classifying respondent groups as a subcategory in multivariate scope data and is a method completed by Goodman in 1972 after it was first introduced by Lazasfeld in 1950 [2, 3]. A latent class analysis can be used additionally together with feasibility verification when we want to estimate diagnoses or positive rate accurately. In addition, it is appropriate when there is no gold standard test or an accurate diagnostic test and it also has advantage to suggest probabilistic information about which potential class shall related objects belong based on observed categorical data [2]. Regarding development delay and normal children class established through a latent class analysis, a positive predictive value was obtained in finding out how effectively the results of each screening test can predict [3, 4]. This study carries out validity of developmental screening tests of 3-5 years old children in K-DIAL3, K-DDST , K-CDI, and K-ASQ [5-14] that using a latent class analysis method, screening effects according to combination of screening tests and positive predictive values of screening tests were compared. These research results can be suggested as basic data for the prevention policy regarding infant-child health screening test projects.
Ⅱ
2
Methods
Setting and Population Object organizations of the study were nurseries located in Cheongju city of Chungcheongbuk-Do and out of total 230 nurseries, we explained about research purpose, research goal, screening test, questionnaires and test methods and requested cooperation to 23 nurseries which were about 10% of the total number of nurseries and 16 nurseries decided to participate. Data Source and Extraction This study had 229 children of 3-5 years old (36 months-60 months), who were born from February 29, 2003 to February 29, 2005, as the object people. Finally recovered questionnaires out of guardians Q&A type questionnaires distributed to total 290 people were 235 and total 229 people of 120 male and 109 female were selected as final object people of the analysis by excluding six people due to inappropriateness of their answers. Data was collected papers were distributed before the survey under the consents of managers of nurseries which agreed to the purpose of the research and the papers were recovered later. With children, who received consent papers of their guardians,
Validity of Developmental Delay Screening Test; Using Latent Class Model
131
as the objects, surveyors visited children’s homes and had interviews of K-DIAL3 and K-DDST for the survey. K-CDI and K-ASQ, which are guardians Q&A type questionnaires, were also sent to homes and were collected later.
Ⅱ
Measures
Ⅱ
Evaluation tools used in this study are types of surveyors K-DIAL3 and K-DDST were used. As Q&A type questionnaires for guardians, K-CDI and K-ASQ were used. K-DIAL3 was translated into Korean by Byeong-Un Chun and et al and the object age was full 3-6 years old. Test-retest reliability was 0.92~0.96 and internal consistency was 0.82. Total time required for the test was about 20-30 minutes [5-7]. K-DDST was developed by re-standardization research in 2002 and is composed of a total 110 questions. Test-retest reliability of K-DDST was 0.90 and internal consistency was reported as 0.99 [8-10]. K-CDI is a questionnaire developed to evaluate children between 15 months and six years old. Implementation time of CDI questionnaires was about 20~30 minutes according to age and internal consistency regarding questions was 0.70 and test-retest reliability was from 0.81 to 0.96 [11, 12]. K-ASQ is published In Korea, Seoul Welfare Hall for Individuals with Disabilities and Test-retest reliability was 0.94 and standard error of measurement was 0.10 [9-14].
Ⅱ
Ⅱ
Data Analysis Scores calculated from experiences depended on diagnostic tests because development delay was suspicious in behaviors of the object people in the results of cutoff points of each test. Scope of cutoff was set corresponding to 5% in K-DIAL3, 10% in KDDST , 2% in K-CDI, and 10% in K-ASQ. Judgment of the ‘suspicion of delay of development’ in this study was based on the standard that there are cases where under one or more detailed development fields exist in cutoff point as the result of screening tests and this is the application standard of K-ASQ and K-DDST [15]. Collected data used SAS statistical program and MLLSA. Verification of validity in sensitivity, specificity, and positive predictive values regarding the remaining three tests was obtained based on each screening test. For utility of developmental screening test tools, positive predictive values were calculated and the two potential classes and developmental screening test tools were compared after dividing into the two potential classes of development delay and normal children by using the results of the four tests.
Ⅱ
Ⅱ
3
Results
Sample Characteristics The age of the children who participated in the study was 36-41 months 42(18.3%), 42-47 months 68(29.7%), 48-53 months 53(23.2%), 54-62 months 66(28.8%) and subject’s gender was mans 120(52.4%), females 109(47.6) (Table. 1). Family member of this study was frequently 4 ones 142(62%), more 5 ones 50(24.9%), 3 ones less
132
S.-A. Lee, C.-Y. Park, and K.-S. Lee
than 30(13.1%). Father ages was below 34 years old 97(42.3%), above 35 years old 132(57.7%), mother’s ages were below 34 years old 159(69.5%), above 35 years old 70(30.5%) (Table. 1).
Table 1. Characteristics of subjects characteristics Age(months)
Gender
Birth order
Family number
Age of father
Age of mother
Total
No(%) 36-41
42(18.3)
42-47
68(29.7)
48-53
53(23.2)
54-60
66(28.8)
Male
120(52.4)
Female
109(47.6)
First
119(52.0)
Second
98(42.8)
Third ≤3
12( 5.2) 30(13.1)
4
142(62.0)
≥5
50(24.9)
≤34
97(42.3)
≥35
132(57.7)
≤34
159(69.5)
≥35
70(30.5) 229(100)
Developmental Delay Positive Rate of Individual Screening Tests Developmental delay positive rate of Individual screening test were K-DIAL3 69 (30.1%), K-DDST 70(30.6%), K-CDI 90(39.3%), and K-ASQ 86(37.6%) (Table 2).
Ⅱ
Validity of Developmental Delay Screening Test; Using Latent Class Model Table 2. Developmental delay cases of subjects by each tests Frequency N(%) Developmental area Test Delay
Normal
Motor
28(12.2)
201(87.8)
Cognition
16( 7.0)
213(93.0)
Language
25(10.9)
204(89.1)
Motor+Cognition+Language
18( 7.9)
211(92.1)
Self-care
19( 8.3)
210(91.7)
Social
21( 9.2)
208(90.8)
Total(delayed children)
69(30.1)
160(69.9)
Social-personal
26(11.4)
203(88.6)
Fine-motor
25(10.9)
204(89.1)
Language
18( 7.9)
211(92.1)
Gross-motor
33(14.4)
196(85.6)
Total(delayed children)
70(30.6)
159(69.4)
Social
55(24.0)
174(76.0)
Self-help
20( 8.7)
209(91.3)
Gross-motor
20( 8.7)
209(91.3)
Fine-motor
10( 4.4)
219(95.6)
Expressive
41(17.9)
188(82.1)
Comprehension
19( 8.3)
210(91.7)
Letters
19( 8.3)
210(91.7)
Numbers
13( 5.7)
216(94.3)
Total(delayed children)
90(39.3)
139(60.7)
KDIAL3(5%)
K-
ซ
DDST (10%)
K-CDI(2%)
133
134
S.-A. Lee, C.-Y. Park, and K.-S. Lee Table 2. (Continued) KCommunication
29(12.7)
200(87.3)
Gross motor
21( 9.2)
208(90.8)
Fine motor
13( 5.7)
216(94.3)
Problem-solving
54(23.6)
175(76.4)
Personal/social
10( 4.4)
219(95.6)
Total(delayed children)
86(37.6)
143(62.4)
ASQ(10%)
Positive Predictive Rate of Developmental Screening Tests Using Latent Class Analysis As the result of implementing a latent class analysis using the results of four screening tests, it was found out that a model designating the potential classes into two was in a good conformity with the analyses. Probability to be classified as a delay of development class among the two potential classes was 0.32% and probability to be classified as a normal development potential class was 0.68%. When children belong to a delay class, the probability to be screened as a delay in each test was 82% in K-DIAL3, 70% in K-ASQ, 67% in K-CDI and 56% in K-DDST . On the contrary, when children belong to a normal class, the probability to be screened as normal in the results of the test was in a sequent order of 94% in K-DIAL3, 81% in K-DDST , 77% in K-ASQ and 73% in K-CDI (Table 3).
Ⅱ
Ⅱ
Table 3. Class probabilities rate as the 2 class model Result of tests∖Latent class Latent class probabilities K-DIAL3
Ⅱ
K-DDST K-CDI K-ASQ
a
Fail
b
Delayc
Normald
0.32
0.68
0.82
0.05
Pass
0.17
0.94
Fail
0.56
0.18
Pass
0.43
0.81
Fail
0.67
0.26
Pass
0.32
0.73
Fail
0.70
0.22
Pass 0.29 0.77 2 x =10.625, df=6, p>.05. a Fail: screening cases of tests results as delay, b Pass: screening cases of tests results as normal, c Delay: latent class of developmental delay, d Normal: latent class of normal developmental.
Validity of Developmental Delay Screening Test; Using Latent Class Model
135
Probability to be a development delay when it was screened as a delay in the results of the test was 86% in K-DIAL3, 59% in K-ASQ, 58% in K-DDST and 54% in K-CDI. As the result, probability for a single result of K-DIAL3 to predict as a development delay was higher than K-DDST , K-CDI and K-ASQ (Table.4).
Ⅱ
Ⅱ
Table 4. Probability of being included in developmental delay class in a single test Delayc predict rate
Test Faila
Passb
K-DIAL3
0.86
0.07
K-DDST
0.58
0.19
K-CDI
0.54
0.16
Ⅱ
K-ASQ 0.59 0.14 a Fail: screening cases of tests results as delay, b Pass: screening cases of tests results as normal, c Delay: latent class of developmental delay.
As the result of screening as a delay in the results of the two tests of K-DIAL3 & K-DDST and K-DIAL3 & K-ASQ, probability to be classified as a development delay was 95%, in K-DIAL3 and K-CDI it was 94%, in K-DDST & K-ASQ, it was 81%, in K-DDST & K-CDI and K-CDI& K-ASQ, it was 78% (Table. 5).
Ⅱ
Table 5. Probability of being included in developmental delay class in a two type tests Delayc predict rate Fail& Faila
Pass & Passb
K-DIAL3 & K-DDST
0.95
0.04
K-DIAL3 & K-CDI
0.94
0.03
K-DIAL3 & K-ASQ
0.95
0.03
K-DDST & K-CDI
0.78
0.09
Test 1 & Test 2
Ⅱ K-DDSTⅡ& K-ASQ
Ⅱ
0.81
0.08
K-CDI & K-ASQ 0.78 a Fail: screening cases of tests results as delay, b Pass: screening cases of tests results as normal, c Delay: latent class of developmental delay.
4
0.07
Discussion
This study has compared validity and utility according to the use of tests based on results of development screening of K-DIAL3, K-DDST , K-CDI, and K-ASQ with the objects of 3-5 years old children. As the result, a sequent order of tools, which can
Ⅱ
136
S.-A. Lee, C.-Y. Park, and K.-S. Lee
predict classification of delay of development children and normal children the most well, was K-DIAL3, K-ASQ, K-CDI, and K-DDST . In addition, in case of using over two tools together, prediction of delay of development was higher than using only one tool and a careful screening test tool is required because positive predictive values are also different according to screening tools. Positive predictive values of delay of development used a latent class analysis based on the results of each test. By dividing children of delay of development and normal children into dependent variables, values screened as normal and delay in developmental screening test with the object people of each test were used as independent variables. As the result, probability to screen as a delay in case children are classified as a development delay in screening test was in a sequent order of K-DIAL3 (82%), K-ASQ (70%), K-CDI (67%), and K-DDST (56%). K- K-DDST had a lower prediction rate(56%) regarding children of delay of development rather than prediction rate regarding normal children(81%) was higher. This is the result that is not complied with the purpose that the developmental screening tests have to distinguish children of delay of development from normal children [16, 17]. In case of using a single developmental screening test, probability to predict development delay, when it was screened as a delay, was 86% only in K-DIAL3 and other tests predicted 54-59%. [2, 4]. In addition, in case of using a two test tools, all predicted development delays were over 90% except K-DDST and K-ASQ were 81%, K-DDST & K-CDI and K-CDI& K-ASQ were 78%. Therefore, in case of conducting development screening evaluations, it is predicted to screen development delay better when using more than two tests together than one test but efficient screening tool must be established in order to reduce cost according to the use of tests. As limits of this study, first, there was a limit in recruiting the object people of the study because only children of 3-5 years old (36-60 months) visiting nurseries were selected as the object. Therefore, it is difficult to generalize by representing only 3-5 years old children. Second, since positive predictive values were obtained from the results of developmental screening tests of children of delay of development and normal children, it will be difficult to generalize the result. In addition, because the secondary diagnoses evaluation of children classified as a development delay was not carried out in this study, it was not investigated what diseases include development delay. Therefore, in the future researches, researches on prediction of obstacles diagnoses of children with development delay and a positive rate for general delay of development must be carried out continuously.
Ⅱ
Ⅱ
Ⅱ
Ⅱ
Ⅱ
References 1. Glascoe, F.P., Foster, M., Wolraich, M.L.: An economic analysis of developmental detection methods. Pediatrics 99, 830–837 (1997) 2. Sohn, W.K.: Exploration for the development of developmental screening test. Open Infant. Edu. Research 9, 81–100 (2004) 3. Pepe, M.S., Jenes, H.: Insights into latent class analysis good diagnostic test performance. Biostatistics 8, 474–484 (2007)
Validity of Developmental Delay Screening Test; Using Latent Class Model
137
4. Garrett, E.S., Eaton, W.W., Zeger, S.: Methods for evaluating the performance of diagnostic tests in the absence of a gold standard: a latent class model approach. Stat. Med. 21, 1289–1307 (2002) 5. Cho, G.S., Chun, B.Y., Lee, J.S.: Korean edition sensitivity and clearness, consent index of DIAL-3. Emotion Behaviors Obstacle Research 21, 25–46 (2005) 6. Chun, B.Y., Cho, G.S., Lee, K.H., Lee, E.S., Im, J.T.: Feasibility research for standardization of (DIAL-3) of an infant development selection tool. Infant Edu. Research 24, 119–140 (2004) 7. Chun, B.Y., Cho, G.S., Lee, K.H., Lee, E.S., Im, J.T.: K-DIAL3 guideline book. Special Education, Seoul (2003) 8. Shin, H.S., Kwon, B.S., Im, S.O.: Research on feasibility of Korean type Denver II test for danger children selection of development problem. Children Nursing J. 11, 316–321 (2005) 9. Frenkenburg, W.K., Dodds, J., Archer, P., Shapiro, H., Bresnick, B.: The Denver II: A major revision and re-standardization of the Denver Developmental Screening Test. Pediatrics 89, 91–97 (1992) 10. Glascoe, F.P., Byrne, K.E., Ashford, L.G., Johnson, K.L., Chang, B., Strickland, B.: Accuracy of the Denver-II in developmental screening. Pediatrics 89, 1221–1225 (1992) 11. Kim, J.M., Shin, H.S.: K-CDI children development test guideline book. Special Education, Seoul (2006) 12. Ieton, H.: Child Development Inventory Manual. Behavior Science system, Minneapolis (1992) 13. Huh, G.H., Moon, J.S., Chung, Y.R.: Preliminary research for standardization of Parents preparation infant & children development selection tool of ASQ. Special Educational Science Research 36, 275–296 (2003) 14. Huh, G.H., Squires, J., Lee, S.Y., Lee, J.S.: Guideline book to use K-ASQ. Seoul Welfare Hall for Persons with Disabilities, Seoul (2006) 15. Glascoe, F.P., Squires, J.: Issues with the new developmental screening and surveillance policy statement. Pediatrics 119, 861–863 (2007) 16. McPhillips, M., Jordan-Black, J.A.: The effect of social disadvantage on motor development in young children: a comparative study. J. Child Psychol. Psychiatry 48, 1214–1222 (2007) 17. Palfrey, J.S., Singer, J.D., Walker, D.K., Butler, J.A.: Early identification of children’s special needs: a study in five metropolitan communities. J. Pediatr. 111, 651–659 (1987)
An Enhanced Data Privacy Mechanism Suitable for Ubiquitous Sensor Networks Jeong-Hyo Park1, Yong-Hoon Jung2 , Kwang-Hyung Lee2 , Keun-Wang Lee3 , and Moon-Seog Jun1 1
3
Department of Computer Science, Soongsil University, Korea {helios914,mjun}@ssu.ac.kr 2 Department of Internet Information, Seoil University, Korea {jyh0178,dreamace}@seoil.ac.kr Department of Multimedia Science, Chungwoon University, Korea [email protected]
Abstract. Sensor network maintains high limitation of resource because it performs many communications in order to create encryption keys. To solve this problem, this research suggests LA2 EP. LA2 EP Protocol can minimize resource and provides a new scheme for authentication and encryption that can provide anonymity of node for safe communication. To analyze the performance of the suggested scheme, a degree of anonymity that is provided by the scheme suggested by using an Entropy-based modeling was measured. As a result, when the suggested scheme was used, the degree of anonymity of sensor node was high. It showed that an important element to increase the degree of anonymity was to let the sensor’s ID not known correctly. Also, as a result of calculating spaces for operation, communication, and storage while considering the characteristic of sensor network, which is limited in resource, it showed suitability for sensor network environment. Keywords: Sensor Network, LA2 EP Protocol, Anonymity, Authentication, Encryption.
1
Introduction
A method of security that is specialized for sensor network, not the one that is used in the perspective of the existing network, is studied [1]. However, the research on the existing sensor network security has been focused on the security services that provide authentication, confidentiality, integrity, and availability [2]. Recently, the interest in the problem of guaranteeing the anonymity of node in wireless sensor network is rising [3] [4] [5]. In case when user uses real ID, instead of a false name for message in the middle of communication, attacker can snatch the network traffic and easily analyze the traffic or easily learn not only the identifier of the sending sensor that communicates with the base station, but also the move of the sensor’s location [6] [7] [8]. Therefore, it is very important to make the moving sensor node not to let the transmitter that exchanges T.-h. Kim et al. (Eds.): GDC 2011, CCIS 261, pp. 138–147, 2011. c Springer-Verlag Berlin Heidelberg 2011
An Enhanced Data Privacy Mechanism
139
data and the third party besides the receiver to easily distinguish the identifier of the sender-receiver in the middle of communication in the environment of doing the daily surveillance or tracking certain object. Also, in case when monitoring some important asset, since the information acquired through sensor’s monitoring should not be exposed to the third party, a security device such as encryption will be used to protect the information [9]. However, one point to pay attention here is that though information about the environment or asset itself that senor monitors is important, the source that transmitted the information can be also important. Therefore, an additional device for preventing exposure of source that is right for the characteristics of sensor network is required besides protecting information through encryption [10]. This thesis is organized like the following. In Chapter Two, it explains the problems of authentication and key distribution in the existing sensor network, and in Chapter Three, it proposes LA2 EP Protocol that provides authentication and encryption to provide safe communication in the next generation sensor network environment. In Chapter Four, it analyzes LA2 EP Protocol qualitatively and quantitatively, and at last, in Chapter Five, it gives conclusion of this thesis.
2
Assumptions and Design Goals
LA2 EP Protocol is the one that provides authentication and encryption to provide safe communication in the next generation sensor network environment. LA2 EP Protocol has assumption details and purpose like the following. 2.1
Network and Assumptions
In order to prevent things like loss and theft of sensor network device and unauthenticated access point, authentication is necessary. Therefore, a research on authentication scheme to prevent this is necessary. Also, the very first thing to consider for composing key in sensor network is the problem of setting encrypted key. This key is used to exchange information between sensor nodes or to protect the exchanged information. For authentication and key management, the following assumption is made. It is not limited only to LA2 EP in this thesis, but it is used in many general encryption protocols. – Attacker can wiretap all the messages that are exchanged through key. – Attacker can interrupt the process of key. Especially, he can change, insert, and block messages, and he can also convey them to other destination. – Attacker can be a user who can normally participate in key, or the third party. – Old session key can be exposed to attackers. 2.2
Design Goals
Sensor must have encryption algorithm in order to perform security functions of sensor network. Here, since sensor nodes have highly limited resource such
140
J.-H. Park et al.
as memory, communication, operation, and power, memory should be used as small as possible so that it fits the limitation of resource of sensor network, and encryption algorithm with small amount of calculation should be applied. The most important resource in sensor node is energy; therefore, power consumption should be decreased as much as possible to increase the life of sensor node. Power consumption is determined by the amount of operation and that of communication, so it is one of the details to be considered primarily when selecting encryption algorithm. Therefore, encryption algorithm is necessary that can let it stay in the sleep mode for the most of the time and that can decrease the amount of operation or that of communication as much as possible even in the execution mode. Also, anonymity authentication of sensor node should not let the moving sensor node to allow the sending node that exchanges data and the third party besides the receiving node to easily distinguish the identifier of the sender-receiver in the middle of communication in the environment of doing the daily surveillance or tracking certain object.
3
LA2 EP : Lightweight Anonymous Authentication and Encryption Protocols
Sensor network is a vulnerable side in security due to the reasons such as the environment exposed to dangers, dynamic network topology, weakness in wireless communication, risk of node’s capture, and limited resource. Therefore, in the USN environment, there is difficulty in applying the security policies that were applied to the existing network as they were. Thus, in this passage, it proposes LA2 EP Protocol (key management and authentication technology) that is designed based on the required details in information protection of sensor network. 3.1
Establishing Anonymous Authentication
This research suggests LA2 EP Protocol that makes the moving sensor node not to allow the sending node that exchanges data and the third party besides the receiving node to easily distinguish the identifier of the sender-receiver in the middle of communication in the environment of doing the daily surveillance or tracking certain object. Here, even if a malicious attacker snatches the content of the message in the middle of communication and attempts traffic analysis, he cannot conjecture the actual identifier of sensor node A. To create a phantom ID, sensor node A performs the process like the following. Step 1. Sensor Node A(A) Node A utilizes a part of information among what it possesses and attempts authentication with BS. Here the information that is used prevents replay attack, H(A), its identifier through the hash operation, and creates random number NA , which is to help create ID, and the value of MACH(A) to guarantee the integrity of message that is in the middle of message.
An Enhanced Data Privacy Mechanism
141
Fig. 1. PID Agreement between the base station and sensor A
Step 2. From sensor node A to the Base Station(A→BS) To guarantee the confidentiality within the same sensor network, message is encrypted by the master key and sent to BS. During this process, there is a risk of wiretapping, but since message is a part of the information that is used to make identifier, the actual identifier of node cannot be induced, even if it is exposed. Step 3. The Base Station(BS) The contents of the encrypted message are decrypted by the master key that is imposed in the same sensor network. The base station that has received the message from node A examines MACH(A) to first check integrity. If there is a problem in the integrity, it ignores the message, and if there is none, it continues proceeding. First, among the contents of the message, it examine the table, utilizing H(A) as index, and save the NA value in the searched column. And it creates arbitrary random number NBS that participates in the creation of node A and H(BS), which is its identifying information that is matched and hashed, and identifier. And it creates MACH(BS) to guarantee the integrity of the message that is in the middle of communication. Step 4. From the Base Station to sensor node A(BS→A) In order to prevent a malicious attacker’s wiretapping or message fabrication in the same sensor network, it is encrypted with the secret key of node A, which has been defined previously at node distribution, and sent. Step 5. Sensor Node A(A) Sensor node A decrypts the received encrypted contents of the message by using the secret key KA , which is already saved. To verify the integrity of the message that is received from BS, sensor node A checks MACH(BS) . If there is a problem in the integrity, it ignores the message, and if there is none, it continues proceeding. It saves H(BS), which is the necessary information in the received message, and the arbitrary random number NBS , which is created at the base station.
142
J.-H. Park et al.
Step 6. Sensor Node A and the Base Station(A/BS) This process is the process of making anonymous identifier that has different location by sensor node A and the base station together. Anonymous identifier is created based on the information that has been sent and received previously in communication. The information that A and BS each possesses is X(BS, A), Y(NA , NBS ), Z(H(BS), and H(A)). Each X, Y, Z show on the plane coordination, and they can be expressed in the matrix of 2 columns and 3 rows like the following. BS NA H(BS) A NBS H(A) To create anonymous identifier, operation like the following is performed. TA ,TB = (BS ⊕ NA ⊕ H(BS) / 3, A ⊕ NBS ⊕ H(A) / 3) TA and TB are the temporary values to create sensor node A’s phantom ID. Later, PID is created by the following operation. PID= TA ⊕ TB 3.2
Establishing Optimized Encryption
In LA2 EP Protocol, light-weighted key management scheme is proposed that has been designed to suit the limitation of resources of sensor network, which has highly limited resource such as sensor nodes memory, communication, operation and power. The following picture shows the process of key distribution of LA2 EP Protocol.
Fig. 2. The process of distributing key with BS and sensor node A
Step 1. The Base Station(BS) During the initial process of key distribution, the base station creates the arbitary random number value, NBS , which will be used as the
An Enhanced Data Privacy Mechanism
Step 2.
Step 3.
Step 4.
Step 5.
Step 6.
143
material for key, and creates MACH(BS) to guarantee the integrity upon the communication. When the period of using sensor node key expires, it re-distributes automatically. From the Base Station to Sensor node A(BS→A) In order to prevent the exposure of the encryption key in the same sensor network, it is encrypted as the secret key KA of sensor node A and sent. Sensor Node A(A) The sensor node that received the message decrypts the encrypted message with its secret key. Also, it confirms MACH(BS) in order to verify the integrity of the message. When there is a problem in the messages integrity, it terminates the message, and if there is no problem, it performs the next procedure. It creates random value NA as the material for the key that will be used in creation of the encryption key, and after going through the operation of BS’=BS+NA and A’=A+NBS , it creates MACH(A ) to guarantee the integrity of the message. From sensor node A to the Base Station(A→BS) In order to prevent the exposure of the key materials that will create the encryption key in the same sensor network, it is encrypted using the master key KM and then sent. The Base Sation(BS) The encrypted message is decrypted using the master key. If there is a problem in the messages integrity, it terminates the message, and if there is no problem, it follows the next procedure. The calculations BS’=BS+NA and A’=A+NBS are performed. And it saves the acquired resulting values of BS’ and A’ and also saves NA and NBS . Sensor Node A and the Base Sation(A/BS) It creates symmetric key with X(BS’, A’), Y(NA , NBS ), Z(H(BS), and H(A)), which are the values that were acquired through communicating between the sensor node and the base station. The operation is conducted after composing the matrix of two columns and three rows.
BS ⊕ NA NA H(BS) A NBS H(A)
TA and TB , which were acquired through the formula, TA ,TB =(BS’ ⊕ NA ⊕ H(BS) / 3, A’ ⊕ NBS ⊕ H(A) / 3), create the symmetric key KA−BS through TA ⊕ TB operation.
4
Performance Evaluation
In this passage, it was distinguished between the security aspect and efficiency aspect and was analyzed qualitatively and quantitatively in order to measure the performance of LA2 EP.
144
4.1
J.-H. Park et al.
Measuring Degree of Anonymity
In this passage, LA2 EPs degree of anonymity is analyzed by using the entropy formula that was defined above. It can be known that the degree of the defined anonymity has the range of 0 ≤ d ≤ 1 through the following equation.
d=
H(X) HM
(1)
Here, when the value of entropy is the largest (the value that d is proximate to 1), the degree of anonymity could be said to be high; when it is the smallest (the value that d is proximate to 0), the degree of anonymity could be said to be low. In other words, the fact that the value of d is close to the maximum value means that it satisfies HM ≈ H(X).
HM = −
N 1 1 [ × log2 ] = log2 N N N
(2)
i=1
HM is the value of entropy in the case when the probability that all the nodes will be distinguished by attacker is identical. H(X) is the value of entropy after the event happened and the value that can be measured after the signal was occurred by the node.
H(X) = −
N
[Pi × log2 (Pi )]
(3)
i=1
To compare H(X) and HM when determining d, it can be known that the most important factor is Pi. It is the probability that attacker will catch the source of sensor when the event of sending the data happens. If Pi = 1/N, then the same value comes out for both H(X) and HM , so the degree of anonymity, d, can be said as the maximum value. Pi = Pr (X = i)
(4)
For the probability to catch the source where the event happened correctly in the network with N numbers of sensors, attacker must be able to correctly identify the sensor that works in certain time and presume the location. In other words, since the attacker knows all the value of ID and information for all the sensors in the network that the anonymous scheme is not applied, the equation is Pi =1 like the following. However, the range of key-pool that N numbers of nodes can have in LA2 EP is like the following picture. The range of key-pool can be determined by the random number, which was exchanged between sensor node A and BS, the value acquired from the hash operation, and the individual key of each. We used matrix to draw the acquired value on the plane coordination. When the matrix of two columns and three
An Enhanced Data Privacy Mechanism
145
Fig. 3. The range of key-pool in LA2 EP protocol
rows was expressed in three points, the area of circle that wraps around the surroundings is the area of total creation key-pool. In other words, the probability pi of malicious attacker finding random ID with brute-force is the same as πr1 2 . Like so, when the value of Pi of LA2 EP scheme is applied to H(X) is like the following equation.
H(X) = −
N 1 1 [ 2 × log2 2 ] = log2 πr2 πr πr i=1
(5)
As a result, since it is defined the same as H(X) and HM and above, the degree of anonymity, d, is like the following. d=
H(X) log2 (πr2 ) = HM log2 (N )
(6)
Just like what has been verified in d, the degree of anonymity, the larger the number of N of sensor node becomes, the wider the radius of , the range of keypool, expands. It can be correctly known through the entropy formula that it is hard for the malicious attacker to conjecture the identifier regarding the random sensor node when the range of key-pool increases. 4.2
Computational Cost
When renewing the key in the same sensor network, the encryption operation is performed to send the new key. The key that is used at this time is pairwise key of sender-receiver node. Therefore, how many encryption operations will be performed is decided by the number of sender-receiver nodes. When the sender sensor node is do, each legitimate receiver nodes can be defined as di, i=1,2,,d0. In other words, Nthe number of performance of encryption operation using pairwise keys is i=1 di. The number of performing decryption operation by the node that received the encrypted message, using pairwise key, is also the same
146
J.-H. Park et al.
as the number of performance of encryption operation. Therefore, when the total amount of consumption of operation is calculated, the total sum of encryption and decryption operations can be defined as 2S, and when the network size is N, it could be analyzed as 2S/N in average. 4.3
Communication Cost
It could be said that when creating, renewing, and consuming secret key of sensor node, the average of communication consumption is similar to the operation consumption that was calculated above. What the number of encryption/decryption operation means is exchanging the communication with base station and neighboring nodes. For example, in the sensor network environment where the network size is N and the degree is d, the average amount of sender-receiver communication can be analyzed as (d − 1)2 /(N-1). 4.4
Storage Cost
When it is supposed that there are d numbers of neighboring nodes in the random sensor node to calculate the storage consumption, each sensor node stores d number of pairwise keys, d number of cluster keys, and one group key. Also, authentication and encryption key can be renewed with random numbers at update stage, not at the bootstrap stage, where the key is issued. Therefore, the storage is separately needed to create the key, besides the key storage. When the storage is defined as T, the storage that each sensor node needs can be analyzed as 2d+2+T.
5
Conclusions
In the anonymity authentication scheme, the moving sensor node does not allow the sending node that exchanges data and the third party besides the receiving node to easily distinguish the identifier of the sender-receiver in the middle of communication in the environment of doing the daily surveillance or tracking certain object. In the key management scheme, it suggested encryption algorithm that uses memory as minimal as possible and has less amount of calculation so that it would suit the sensor networks limitation of resources. By analyzing the performance of LA2 EP protocol qualitatively and quantitatively, the security and efficiency have been verified. The entropy formula was applied to the aspect of security and was analyzed quantitatively. For the degree of anonymity, d, the larger the number of sensor node N becomes, the wider the radius of πr2 , which is the range of key-pool, expands. Through the entropy formula, it can be correctly known that when the range of key-pool increases, it is hard for the malicious attacker to conjecture the identifier of the random sensor node. In the aspect of efficiency, it was quantitatively analyzed in spaces for operation, communication, and storage. In the aspects of operation consumption and communication consumption, the differences in result value were large, based on the network
An Enhanced Data Privacy Mechanism
147
composition, but viewing it in average, it showed suitability for the sensor network, and largely increased efficiency was confirmed in storage consumption. As a result, it showed that it was efficient to the network and hypothesis that this thesis suggested.
References 1. Akkaya, K., Younis, M.: A survey on routing protocols for wireless sensor networks. Ad Hoc Networks 3, 325–349 (2005) 2. Li, N., Zhang, N., Das, S.K., Thuraisingham, B.: Privacy preservation in wireless sensor networks: A state-of-the-art survey. Ad Hoc Networks 7(8), 1501–1514 (2009) 3. Gruteser, M., Grunwald, D.: Anonymous usage of location-based services through spatial and temporal cloaking. In: Procs. of the 1st Int. Conf. on Mobile Systems, Applications and Services, pp. 31–42. ACM (2003) 4. Sweeney, L.: Achieving k-anonymity privacy protection using generalization and suppression. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 571–588 (2002) 5. Wadaa, A., Olariu, S., Wilson, L., Eltoweissy, M., Jones, K.: On providing anonymity in wireless sensor networks. In: ICPADS 2004, pp. 411–418 (2004) 6. Mehta, K., Liu, D., Wright, M.: Location privacy in sensor networks against a global eavesdropper. In: ICNP 2007, pp. 314–323 (2007) 7. Xi, Y., Schwiebert, L., Shi, W.: Preserving source location privacy in monitoringbased wireless sensor networks. In: Procs. of the 20th Int. Parallel and Distributed Processing Symposium. IEEE Computer Society (2006) 8. Ozturk, C., Zhang, Y., Trappe, W.: Source-location privacy in energy-constrained sensor network routing. In: Procs. of the 2nd ACM Workshop on Security of Ad Hoc and Sensor Networks, pp. 88–93. ACM (2004) 9. Misra, S., Xue, G.: Efficient anonymity schemes for clustered wireless sensor networks. IJSNet, 50–63 (2006) 10. Pietro, R.D., Mancini, L.V., Mei, A.: Energy efficient node-to-node authentication and communication confidentiality in wireless sensor networks. Wirel. Netw. 12(6), 709–721 (2006)
Grinding Ability and Characteristics in Normal QT Heat Treatment vs. Carbo-nitriding Heat Treatment Bearing Steel under CBN Wheel Grinding Process Jeong Jin Kim1, Sung Hoon Oh2, and Byung Moon So3 1
Graduate School, Mechanical Engineering, Chon-buk National University 664-14, Dukjin-dong, Jeon-ju city, Jeollabuk-do, 561-756 [email protected] 2 Division of Mechanical system engineering, Chon-buk National University, Dukjin-dong, Jeon-ju city, Jeollabuk-do, 561-756 [email protected] 3 Department of IT applied system engineering, Chon-buk National University, Dukjin-dong, Jeon-ju city, Jeollabuk-do, 561-756 [email protected]
Abstract. For enhance fatigue resist ability in hardened bearing steel especially contaminated lubrication condition, Nitriding treatment applied. This diffuse nitrogen and carbon into the surface of steel during austenitizing. This enhanced surface condition gave physical strength with tribological function but this cause grinding difficulties in conventional grinding process. Its main reason is intensified toughness in the steel surface. Due to increase in Nitrogen and carbon in the steel surface, it contain more retained austenite than normal QT heat Treated condition. With grinding power trend it shows increasing in Nitriding condition. Also Same with grinding load behavior. Surface roughness slightly impaired in nitriding condition In roundness, QT treated condition shows conventional high speed adaptability but Nitriding Condition shows different behavior when accumulated wheel speed and depth of cut amount. Keywords: Carbo-nitriding, QT heat treatment CBN wheel grinding, Fatigue resist, Surface roughness.
Normal QT Heat Treatment vs. Carbo-nitriding Heat Treatment Bearing Steel
149
rolling element life especially in fatigue resist life, Carbo-nitriding heat treatment widely used in the material processing area. This diffuse nitrogen and carbon into the material surface during austenitizing. In consideration of bearing Circumstances its rolling element sometimes exposure to the contaminated lubrication conditions, under this severe operation condition whole rolling element easily destructed and anticipated life cycle reduces with definitely low level. With carbo-nitriding heat treatment conditions, enhance fatigue resist life with increased retained austenite at the surface and temper resistance of surface region. On the other hand, these enhanced surface properties will damage grinding ability and cause basic adversities in achieving geometrical tolerances and proper quality characteristics. In this study we analyze and check the internal grinding ability between normal termperd hardened bearing steel and Carbo-nitriding heat-treated bearing steel with using CBN grinding wheel. During the test check the grinding power, grinding load, surface roughness and roundness checked under various kinds of grinding conditions. With these results suggest optimized grinding conditions in the actual industrial filed and developing optimal grinding wheels for high speed and sustainable form shape processing area.
2
Material Composition and Heat Treatment Conditions Table 1. Chemical composition of STB2 bearing steel (wt. %)
Fig. 1. Normal QT heat treatment cycle
Fig. 2. Carbo-nitiriding heat treatment cycle
150
J.J. Kim, S.H. Oh, and B.M. So
Material is high chromium bearing steel, so called STB2. Standard Heat treatment condition between Normal QT and Carbo-nitriding is in Fig. 1 and Fig. 2. For diffuse nitrogen and carbon in the steel surface, during the hardening stage carbon potential with 0.9~1.05%, addition of NH3 with typically 8 vol % of base gas. Cooling is quenching in oil with 60~100 . After heat treatment hardness kept HRC62 both QT & CarboNitiriding heat treated conditions. Nitrogen diffusion depth kept minimum 0.5mm and the core microstructure is fine acicular. After carbo-nitriding treatment, material property changed with increased wear resistance and corrosion resistance ability by nitride network formed in the steel surface. Also Carbo-nitriding allow much lower quenching temperature compared with Normal QT heat treatment cycle, it showed much less distortion after heat treatment. Nitrogen is generally sourced by the decomposition of ammonia in the following reaction sequence during the gas nitriding procedure using heat as the method of the decomposition and the steel as the catalyst. But most of all, retained austenite amount in microstructure gave with toughness and durable features.
℃
2NH3 ↔ N2 + 3H2 Table 2. Inspection result after heat treatment with hardness and microstructure
Fig. 3. Microscopic image: QT treated (X1000)
(1)
Normal QT Heat Treatment vs. Carbo-nitriding Heat Treatment Bearing Steel
Machine and Measuring Device Specification Table 3. Experimental condition for grinding operation
Fig. 5. System configuration of Grinding machine & Power
151
152
J.J. Kim, S.H. Oh, and B.M. So
Fig.5 is the experimental machine set-up for grinding power and grinding load check. Wheel spindle main power input is 350V-18kW. Grinding power output monitored in the Warner KT4060-10 converter, grinding load output is checked in the 16bit AD converter results through PRI-3400 indicator. Surface roughness check after grinding operation is by Taylor-Hobson Form taly-surf and roundness is checked by Mahr MMQ-400 measuring device. Work-piece rpm is 1,042rpm, wheel speed is variable to 30, 45, 60m/s. Depth of cut is designated to 100, 200 in each wheel speed conditions. Oscillation during the grinding operation is Set-up by 1.5mm over at each material edge end. Speed is 2500m/min which travel back and forth along the bore axis. Spark-out time is 0.5sec and grinding material guided shoe Support during the grinding operation.
4
Results and Discussions
4.1
Grinding Power and Load
Fig. 6. Grinding power results in various grinding condition
Fig. 7. Grinding load results in various grinding condition
Fig.6 is grinding power results under various grinding conditions. In QT heat treated condition grinding power slightly increased as wheel speed increased. With cutting depth influences in low wheel speed grinding power increased while wheel speed
Normal QT Heat Treatment vs. Carbo-nitriding Heat Treatment Bearing Steel
153
increased with 60m/s, grinding power slightly decreased. This can be explained by “Size effect” which cutting depth is increased cutting pressure is decreased. But Carbo-nitriding heat treated condition it showed totally different behavior. Cutting depth increasing directly affect to grinding power trend. This mainly depends on nitrogen network in steel surface. Nitriding layer placed over 0.5mm depth in steel surface and this increase toughness property in grinding operation. Also it can be clearly found than in high speed grinding probably effect on grinding operation but with enhanced surface layer hinder optimal grinding and with cutting depth increased grinding power increased in proportion. This tendency also can be found in grinding load results. In QT heat treated conditions load variation is slightly low but in Carbonitriding heat treated conditions load behavior showed much higher results these consequences. It can be inferred that nitriding layer caused grinding hinder and with cutting depth increased grinding load increased in proportion. 4.2
Surface Roughness and Roundness
Fig. 8. Surface roughness results in various grinding condition
Fig. 9. Roundness results in various grinding condition
Fig.8 is surface roughness results in various grinding conditions. Both Quenching and nitriding conditions show roughness improving trend with wheel speed increase. In Carbo-nitriding condition, show similar trend with roughness improved in high speed
154
J.J. Kim, S.H. Oh, and B.M. So
area and roughness impaired with cutting depth increased. Fig.9 is roundness results in various grinding conditions. In Quenching conditions roundness trend became stable with wheel speed increased. On the other hand Nitriding conditions show quite contrary to this.
5
Conclusions
1. In QT treated conditions grinding power increased with wheel speed increased. But when cutting depth increased grinding power slightly decreased. In nitriding conditions cutting depth increasing affect to grinding power in proportion. 2. In nitriding conditions grinding load show much higher tendency in comparison with QT conditions. 3. Surface roughness improved in QT conditions when wheel speed increased. In nitriding conditions surface roughness improved when wheel speed increase. 4. Roundness results show quite contrary to QT and Nitriding conditions. In QT conditions roundness improved with wheel speed increased but it impaired in Nitriding conditions.
References 1. Marinescu, I.D., Hitchiner, M., Uhlmann, E., Rowe, W.B., In-asaki, I.: Handbook of Machining with Grinding Wheels. CRC Press (2007) 2. Totten, G.E.: Steel heat treatment, Ph.D. FASM. CRC Press (2007) 3. Choi, B.Y., Kim, D.K., Kim, C.S., Jin, J.K.: Effect of retained Austenite on rolling contact fatigue of Nitrocaburized high-carbon Chromium bearing steel (1995)
Correlation between Low Polarization and Roughness in Low-k SiOC Thin Films by Chemical Vapor Deposition Teresa Oh Semiconductor Engineering, Cheongju University 36 Naeduckdong Sangdangku, Cheongju 360-764, Korea [email protected]
Abstract. SiOC films prepared by the chemical vapor deposition were analyzed by photo luminance spectra and atomic force microscopy to define the correlation between the chemical shift by PL spectra and roughness of SiOC films as low dieclectric constant materials with amorphous structures. The amorphous structure was owing to decrease the polarity in SiOC film. The effect of weak boundary conditions due to the low polarization induced the chemical shift in PL spectra, and the shifting to high wavelength produced to decrease the roughness. These effects such as the red shift, lowering of roughness and polarity were required for an amorphous structure. The lowest dielectric constant could be obtained at SiOC film with the lowest polarity as ideal amorphous structure. Keywords: SiOC film, Polarity, Roughness.
mechanical properties of CVD-SiOC film analyzed by photo luminance and roughness. In order to optimize the deposition condition with lowering the polarization of the SiOC films deposited on Si wafer, SiOC film was deposited with various flow rate ratios by the ICP-CVD system and the origin of lowering the dielectric constant was explained.
2
Experiments
SiOC films were obtained using the mixed gases of oxygen and methyltrimethoxy silance (MTMS) by inductive coupled plasma chemical vapor deposition (ICP-CVD). The films were processed at annealing temperature 500 for 30 minute at various MTMS flow rate ratios. The MTMS was vaporized and carried by argon gas at 35 with a thermostatic bubbler. The base pressure of the mixture was kept at 3 Torr and the rf power was 800 W in each experiment. Total flow rate of the precursors was maintained as 15 sccm and the sample number was the same as the flow rate of MTMS (carried by Ar) precursor. The dielectric constant was estimated from the C-V (capacitance-voltage) measurement at 1MHz using the HP4284A on the MIS (Al/SiOC film/Si substrate) structure, and the PL spectra were obtained from the Photo Luminescence (SPEX, SPEX) and the roughness was measured by the atomic force microscopy (S.I.S, GmbH).
℃
3
℃
Results and Discussion
Figure 1 is the PL spectra of the as-deposited SiOC film as a function of various flow rate ratios. Sample number is the MTMS (Ar) flow rate. Figure 1(a) is PL spectra of the samples of 3~9, and Fig. 1(b) is about the samples 9~13 results. The spectrum consists of the band of 350~650 nm, including peaks with 370 nm, 405 nm and 450 nm. There are a lot of changes between samples at 7~10.
As deposited film (b) 9~13
As deposited film (a) 3~9
13
PL intensity (a.u.)
PL intensity (a.u.)
9 8 7 6
5
12 11
10 9
4 3 300
400
500
600
Wavelength (nm)
700
800
300
400
500
600
700
800
Wavelength (nm)
Fig. 1. PL spectra of as deposited SiOC film with various flow rate ratios, (a) PL spectra of sample 3~9, (b) PL spectra of samples 9~13
The wavelength of 370 nm, 405 nm and 450 nm in PL spectra correspond to the energy band gap of 3.35 nm 3.06 nm and 2.76 eV, respectively.
Correlation between Low Polarization and Roughness in Low-k SiOC Thin Films
157
Figure 2 is the PL intensity of SiOC film with increasing the MTMSM (Ar) gas flow rate ratios. The PL intensity increased until sample 8 with increasing the MTMSM gas flow rates, but decreased at sample 9 notwithstanding the increase of MTMSM gas flow rate. 220000
Fig. 2. PL intensity of SiOC film with various flow rate ratios
Figure 3 is the roughness of as deposited SiOC film with various flow rate ratios obtained from AFM. Sample 8 with high intensity of PL were higher roughness than any others. But sample 12 is high intensity and low roughness as shown in Fig. 2 and 3. There is existed the difference in polarity due to the chemical reaction between alkyl and hydroxyl group as opposite sites according to the flow rate ratios during the deposition. Sample 9 has low intensity in PL and low roughness. These figures show an extreme results in spite of very few amount in the difference of flow rates.
2.5
As deposited film
Roughtness (nm)
2.0
1.5
1.0
0.5 4
6
8
10
12
Sample number
Fig. 3. Roughness of as deposited SiOC film with various flow rate ratios
Figure 4 displays the comparative figures of PL spectra of as deposited SiOC films and annealed SiOC films to research the effect of after annealing processes. Solid line is PL intensity of annealed SiOC films and dot line is that of as deposited SiOC films. Annealed SiOC film mostly decreased the intensity of PL spectra. Especially, the decreasing range of samples 6~11 is large. SiOC film was formed by the chemical reaction due to the decreasing the polarization between the alkyl and hydroxyl group as opposite sites. Therefore the dielectric constant was also decreased by the lowering
158
T. Oh
the polarization. Low intensity of annealed samples at 6~11 is related to the reduction of polarization. To obtain the dielectric constant, it was researched the thickness and capacitance from the MIS structure and measured the dielectric constant. 3
4
6
8
7
Intensity (a.u.)
5
300
400
500
600
700
800
900
300
Wave number (nm)
400
500
600
700
800
300
900
400
500
600
700
800
900
300
400
Wave number (nm)
Wave number (nm)
9
600
700
800
900
300
400
500
600
700
800
900
Wave number (nm)
300
400
500
600
700
800
900
Wave number (nm)
13
12
Intensity (a.u.)
10
500
Wave number (nm)
11
300
400
500
600
700
Wave number (nm)
800
900
300
400
500
600
700
800
900
300
400
500
600
700
800
900
300
Wave number (nm)
Wave number (nm)
400
500
600
700
800
900
300
400
Wave number (nm)
500
600
700
800
900
Wave number (nm)
Fig. 4. Comparison of PL spectra of as deposited SiOC films and annealed SiOC films, dot line of as deposited filma and solid line of annealed films
Figure 5 is the thickness and dielectric constant of as deposited films obtained from MIS structure. 350
40
As deposited film 35 30
250
25 200 20 150 15 100
Dielectric constant
Thinckness (nm)
300
10 50
5
0 4
6
8
10
12
14
Sample number
Fig. 5. Thickness and dielectric constant of SiOC film with various flow rate ratios
The thickness was generally in proportional to the dielectric constant, and the lowest dielectric constant was obtained at sample 8. The steep reduction of polarization from samples 6~11 causes the lowest polarity at sample 8, which has the lowest dielectric constant. Sample 8 also has high intensity of PL spectra as shown in Fig. 2. Annealed film
PL intensity (a.u.)
13 12 11 10 9 8 7 6 5 4 300
400
500
600
700
3
800
Wavelength (nm)
Fig. 6. PL intensity of annealed SiOC film with various flow rate ratios
Correlation between Low Polarization and Roughness in Low-k SiOC Thin Films
159
Figure 6 is the PL intensity of annealed SiOC film with various flow rate ratios. There are a lot of difference between samples 7~11 because of the weak boundary condition due to the low polarization, so then this effect induced to decrease the roughness and flat surface. To define the correlation between the chemical shift and roughness, it was researched the roughness about annealed SiOC films as next sentence. 3.0
Annealed film
Roughtness (nm)
2.5
2.0
1.5
1.0
0.5 4
6
8
10
12
Sample number
Fig. 7. Roughness of annealed SiOC film with various flow rate ratios
From the result of Fig. 7, the roughness decreased at samples 4, 9and 13, which were shown the changing to low wavelength of peak in PL spectra. These results are similar to the previous result of the figures 1 and 2. 25
Deposited film
Dielectric constant
20
15
Annealed film
10
5
0 4
6
8
10
12
Sample number
Fig. 8. Dielectric constant of annealed SiOC film with various flow rate ratios
Figure 8 is the dielectric constant of annealed SiOC film with various flow rate ratios. The lowest dielectric constant was obtained at the annealed SiOC film 9, which has low roughness and low wavelength in PL spectra.
4
Conclusion
The SiOC films were prepared at various flow rate ratios by the inductive coupled plasma chemical vapor deposition and annealing at 500 . The correlation between the chemical properties and roughness in SiOC film was observed by analysis of PL spectra and AFM. The weak boundary condition due to the attractive force between
℃
160
T. Oh
alkyl and hydroxyl group as opposite sites decreased the polarization in SiOC film, so induced the amorphous structure, and then decreased the dielectric constant. The intensity of PL spectra was decreased after annealing process, and the roughness decreased at the annealed SiOC film with the lowest dielectric constant and shifting to higher wave length in PL analysis. The low dielectric constant, low roughness and low polarization were good for low-k materials to require the amorphous structure and insulators.
References 1. Oh, T.: Jpn. J. Appl. Phys. 44, 1409–1413 (2005) 2. Yu, L.D., Lei, S., Dong, Z.S., Yi, W., Yan, L.X., Qi, H.R.: Chin. Phys. Soc. 16, 240–245 (2007) 3. Oh, T., Choi, C.K.: Journal of the Korean Physical Society 56, 1150–1155 (2010) 4. Oh, T.: IEEE Transactions on Nanotechnology 5, 23–29 (2006) 5. de Rouffignac, P., Li, Z., Gordon, R.G.: Electrochemical and Solid-State Letters 7, G306– G308 (2004) 6. Kim, C.Y., Navamathavan, R., Lee, H.J., Choi, C.K.: Journal of the Korean Physical Society 202, 5688–5692 (2008) 7. Tsai, K.C., Wu, W.F., Chao, C.G.: Journal of Electronic Materials 35, 1523–1529 (2006) 8. Cheng, Y.L., Wang, Y.L., Lan, J.K., Chen, H.C., Lin, J.H., Wu, Y.L., Liu, P.T., Wu, Y.C., Feng, M.S.: Thin Solid Films 469-470, 1783–1787 (2004)
A Semantic-Based Centralized Resource Discovery Model for Grid Computing Abdul Khalique Shaikh1, Saadat M. Alhashmi1, and Rajendran Parthiban2 1
School of IT, Monash University, Sunway Campus, Bandar Sunway, Malaysia {shaikh,alhashmi}@monash.edu 2 School of Engineering, Monash University, Sunway Campus, Bandar Sunway, Malaysia [email protected]
Abstract. Grid computing aggregates computing resources to execute computationally complex jobs. Identifying an appropriate resource selection mechanism for Grid jobs is a major concern because an appropriate mechanism could helps to schedule and allocate resources better. In existing real grid systems, centralized and hierarchical resource discovery models are very common, where providers publish resources with a fixed schema and users follow the same schema in order to find appropriate resources for jobs. Lack of coordination between users and providers in a highly heterogeneous Grid environment often results in user jobs failing in finding relevant resources. One of the reasons for rejection of jobs is the usage of fixed schema between user requests and provider resources. We present a semantic-based centralized resource discovery model to improve job success probability and utilization of resources. Preliminary simulations carried out using GridSim simulator show improved success probability and enhanced utilization of resources. Keywords: Grid Computing, Resource Discovery, Semantic, GridSim, Centralized.
resource selection in a Grid system not only involves finding appropriate resources based on users’ requirements, but also includes tracking, matching, selecting, requesting and accessing appropriate resources to execute computationally complex jobs [3]. The overall effectiveness of the system depends on the level of coordination and cooperation among users, providers, resources and services [4]. In order to maximize resource utilization and enhance job success probability, there is a need for better coordination between users and providers. Job success probability could be defined as the ratio of the number of jobs accepted to the total jobs requested by users, whereas utilization of resources is the ratio of utilized resources to the number of requested resources. It is noted that relatively high percentage of jobs submitted by users to a Grid system are rejected since current schedulers are unable to find relevant resources. For example, the technical report in [5] states that the WISDOM project submitted 6500 jobs to EGEE grid, of which 35% jobs were rejected. One of factors for rejection is the usage of a fixed schema between user requests and provider resources. A fixed schema usually increases the possibility of missing relevant resources. To overcome this issue, we use a semantic-based approach where metadata and ontology are used to describe resources, the approach provides more opportunity to discover and monitor grid computing resources. In this paper, we use the semantic-approach in a centralized resource discovery model which can improve job success probability and maximum utilization of resources in a Grid system. Our approach can be used with any existing Grid middleware which provides centralized or hierarchical grid scheduling and resource discovery services such as Globus Toolkit, Condor or gLite grid middleware. Centralized and hierarchical resource discovery models are very common in real grid environments and are used for small and medium size Grid systems. In this paper, we assume medium size grid where the number of resource nodes is not beyond 1000. The remaining sections of this paper are organized as follows: Section 2 reviews existing centralized and hierarchical-based resource discovery systems. Section 3 explains semantic deployment with explanation of resource ontologies. Section 4 presents experiments and Section 5 shows results. Finally, Section 6 provides concludes the paper with possible future work.
2
Related Work
Significant research has been carried out to enhance grid resource discovery mechanism in both academia and industry. Centralized and hierarchical resource discovery models are recommended for small and medium size grid systems, since they are easy to design and manage. On the other hand, decentralized resource discovery models suite for a large-scale grid system in which load balancing, scalability, heterogeneity and fault resilience could be maintain very effectively [6]. Since, we assume a medium size Grid system, we only focus on centralized and hierarchical resource discovery models in this paper. Condor [7] uses centralized queries for resource discovery process and works on the principle of matchmaking [8]. In the matchmaking approach, resource providers
A Semantic-Based Centralized Resource Discovery Model for Grid Computing
163
and clients place their service characteristics and requirements using classified advertisements (Class Ads). The matchmaking framework includes several components of a resource discovery process and provides stateless matching services [9]. Condor resource discovery process is based on a syntactical approach that works on exact keyword matching and provides limited results. Moreover, users are unable to find relevant resources for their jobs, if they do not use the same schema as providers’. Thus, users’ jobs could be rejected and utilization of resources could be low. Monitoring and Discovering Services (MDS) [10] is a suite of components offered by Globus, for monitoring and discovering resources and services. It uses a hierarchical approach and utilizes standard interfaces defined within the WS-Resource Framework (WSRF) and WS-Notification (WS-N) specifications. MDS allows users to stipulate basic configurations, such as the file path, maximum CPU power and required memory. However, it does not support job matching/scheduling at the global level and does not provide semantic support [11] [12]. Thus, these could be affects the results in searching process to find out some relevant resources. Pastore [13] proposed a Universal Description Discovery and Integration (UDDI) mechanism using a registry deployed in Globus-based Grid system. UDDI is a standard directory service mechanism for web services. Pastore states that the registry engine’s function limits automation of the search mechanism. Semantic web-based web services could enhance the search mechanism. In the gLite grid middleware, a Workload Management System (WMS) service is responsible for the management of users’ jobs. It matches the jobs’ requirements to the available resources and schedules the jobs for execution using appropriate computing resources. Also, an IS (Information System) component provides information about the existing resources and their status. The information is published by the individual resource and copied into central databases [14]. The gLite uses syntactical approach in matching process, so users are bound to follow the providers naming scheme in a Job Description Language (JDL). An exact keyword matching is used in above centralized and hierarchical resource discovery mechanisms. In this approach, resource providers and users have to agree on the same schema which can affect job success probability negatively. Dikaiakos [15] utilized Globus toolkit and UNICOR (Uniform Interface to Computing Resources) Grid middleware. The resource discovery in this model is based on hierarchical structure. Authors of [15] uses semantic approach to described resource attributes. The basic mapping between UNICORE resource ontology and Globus resource ontology has been done. However, the proposed technique was not experimentally evaluated. In our work, we present a semantic centralized resource discovery model in which we add semantic features and compute semantic similarity threshold values using existing ontology concepts. We also simulate our work on GridSim and compare our model’s performance with that of non-semantic model.
3
Semantic Deployment
This section explains a semantic deployment along with resource ontologies. It also presents the features of our semantic framework used for resource ontology discovery in a Grid system.
164
A.K. Shaikh, S.M. Alhashmi, and R. Parthiban
Grid resources belong to different virtual organizations with their own rules and policies, so it is possible for same resources to be published with different terms. A semantic approach can be useful to identify the relationship between those resources [16]. Moreover, ontology defines resource descriptions and builds a semantic bridge between the resource descriptions and users’ demands [17]. Semantic similarity is defined as the relationship between ontology concepts. The similarity of concepts represents the degree of commonality between these concepts. Andreasen et al. [18] derive conceptual similarity using the notion of “similarity graph”. In this, ontology is represented as a graph with concepts as nodes and relationships connecting these concepts as edges. Andreasen et al. introduced following equation and use sim (x,y) function to measure the degree of similarity i.e. the degree of common concepts x and y share. By using following equation, we compute semantic similarity values between concepts of resources. The semantic similarity distance between concepts of resources is also known as a semantic threshold value. ,
|
| |
|
1
|
| |
|
(1)
In (1), is a factor that determines the degree of influence of generalization of ontology concepts. The value of is between 0 and 1. If value of is 1, that means perfect generalization, with each and every concept defined properly and 0 means very poor generalization. We set the value of = 2/3 (0.67) as in [18] with the same level of generalization. α x is the set of nodes reachable from x and α x α y is the set of reachable nodes shared by x and y. , 0 means x and y are completely dissimilar and , 1 means full similarity. To compute semantic similarity, we used two existing ontologies for resources Processor and Operating System from [19]. The Processor ontology is structured as in Fig.1.
Fig. 1. Ontology for Processor Resource
A Semantic-Based Centralized Resource Discovery Model for Grid Computing
165
With the help of sub-domain specific ontologies and (1), we compute semantic similarities between concepts of resources. For example, the following method can be used to compute the semantic similarity between Intel concept and AMD concept for Processor resource. By using (1) ,
|
|
| |
|
|
,
|
. ,
|
. .
Table 1. Semantic threshold values for Processor ontology
The computed semantic threshold values for Processor ontology concepts are shown in Table 1.
4
Experimental Setup
To simulate semantic features on a centralized resource discovery model, we use GridSim simulator, which supports discrete event and time based simulations in a Grid environment. GridSim also provides facilities for the modeling and simulation of resources and network connectivity with different capabilities, configurations and domains [20]. As semantic API classes are not provided with this simulator, we have extended some existing classes to add semantic features. We evaluate the performance of the resource discovery model in terms of successful jobs, utilization of resources and response time. We ran two experiments to evaluate and compare the performance of semantic and syntactical-based (nonsemantic) resource discovery models. We use the following parameters for both semantic and non-semantic experiment. Semantic threshold values are only applicable for semantic experiment.
166
A.K. Shaikh, S.M. Alhashmi, and R. Parthiban
Number of Resources = 512 Jobs query rate =50- 500 per/second in steps of 50 Total Processor Architecture Concepts = 16 Total Operating System Concepts = 17 Semantic Threshold= 0.7 The experiment is aimed at enhancing job success probability and resource utilization by adding semantic features in a centralized resource discovery model. We used subdomain specific ontology for each type of resource separately. In this implementation, we computed semantic similarity threshold values using equation (1). We observe the successful jobs, utilization of resources and response time under both semantic and non-semantic centralized resource discovery models. In our experiment, we utilize resource attributes including processor architecture and operating system from existing ontology concepts. Given below is an example of one generated resource and one gridlet/job entity in GridSim simulator. •
All resources are registered in Grid Information Service (GIS) that provides grid resource registration, indexing and discovery services. Subsequent to registering the resources, users are able to send job requirements to available resources in a grid system. In order to find appropriate resources for jobs, each job deals with existing resources to find the match and then pick those resources to submit the job. We consider space shared allocation policy for grid resources that behaves exactly like First Come First Serve (FCFS).
A Semantic-Based Centralized Resource Discovery Model for Grid Computing
167
Each job finds its requirement such as number of CPUs, processor architecture and operating system by comparing with each resource. If the requirements of jobs match with any existing resources, then the jobs are submitted successfully. Otherwise the jobs are rejected. In this way, we measure the relationship between successful jobs and query rates. Since we are not considering semantic matching in the first experiment, users’ requests match with resource availability based on exact keywords matching. In the second experiment , we add semantic features and use two subdomain specific ontologies for Processor and Operating System resources. We run this experiment under 512 nodes network with various job query rates i.e. 50 to 500 in steps of 50. In order to find appropriate resources for jobs, each job deals with existing resources with semantic data to find a match and then picks those resources. Once this is done, the job is considered submitted. The semantic search mechanism helps to find relevant resources for a job based on semantic similarity threshold 0.7 in case the exact match is not found. We use an average semantic threshold value of 0.7 in this model. However, the values can be varied and selected based on users’ requirement. The results of both semantic and non-semantic experiments are shown in the next sub-section.
5
Results
Fig. 2 shows the relationship between percentage of successful jobs and job query rates for semantic and non-semantic cases. The result shows that some jobs are rejected under non-semantic case because of tight coupling between users’ requirements and available resources, since this is based on exact keyword matching. Thus, the resources are not utilized well. For high query rates, the job success probability becomes less. For the semantic case, percentage of successful jobs is higher than non-semantic case and high query rates do not affect its performance. Fig. 3 shows the difference in utilization of resources between semantic and non-semantic cases.
110 100
Successfu ll Jo bs %
90 80 70 60 Semantic Non-Semantic
50 40 30
Nodes = 512 Semantic Thr=0.7
20 10 0 50
100
150
200
250
300
350
400
450
500
JOb Query Rates (per sec)
Fig. 2. Percentage of Successful Jobs vs. Job query rates (per second)
168
A.K. Shaikh, S.M. Alhashmi, and R. Parthiban 100 90
Utilization of Resources %
80 70 60 50
Non-Semantic Semantic
40 30
Nodes = 512 Semantic Thr=0.7
20 10 0 50
100
150
200
250
300
350
400
450
500
Job Query Rates (per sec)
Fig. 3. Utilization of Resources vs. Job Query rates (per second)
For non-semantic case, resource utilization increases at first and then decreases when jobs query rate increases. In semantic case, the utilization of resources is directly proportional to the jobs query rate. Hence, the utilization of resources in semantic case is higher than that of non-semantic case. The reason is that semantic matching gives high success probability of jobs. In Fig. 4, the response time is plotted against the query rate. The results show that the non-semantic model takes more time to find the requirement for the jobs than the semantic model. With the semantic approach, there is high probability to find resources and it helps to make the finding process faster. The overall simulation results show that by using semantic features in a centralized resource discovery model, percentage of successful jobs and utilization of resources can be enhanced and the response time can be reduced. 2200 2000
Responce Time (Second)
1800 1600 1400 Semantic Non-Semantic
1200 1000 800 600
Nodes = 512 Semantic Thr=0.7
400 200 0 50
100
150
200
250
300
350
400
450
500
Job Query Rates (per second)
Fig. 4. Response Time (seconds) vs. Jobs Query rate (per second)
6
Conclusion and Future Directions
Performance of a Grid system depends on an effective resource discovery process. Hence, there is a need for comprehensive resource discovery model to reduce job rejection rate and to enhance utilization of resources. In this paper, we added semantic
A Semantic-Based Centralized Resource Discovery Model for Grid Computing
169
features in a centralized resource discovery model and compared the performance of semantic and non-semantic cases. Two existing ontologies for Processor and Operating system were used to remove the tight coupling between user requirements and resource availability. The overall results show that job success probability and utilization of resources improve with semantic approach. Our proposed model can be used with Globus Toolkit, CONDOR and gLite grid middleware. As future work, we plan to add semantic features in decentralized or P2P overlay network which is recommended for a high scale grid system, and compare the performance of semantic and non-semantic cases.
References 1. Lacks, D., Kocak, T.: Developing reusable simulation core code for networking: The grid resource discovery example. Journal of Systems and Software 82(1), 89–100 (2009) 2. Caminero, A., Sulistio, A., Caminero, B., Carrión, C., Buyya, R.: Extending GridSim with an Architecture for Failure Detection. In: 13th International Conference on Parallel and Distributed Systems 2007, pp. 1–8 (2007) 3. Naseer, A., Stergioulas, L.K.: Resource discovery in Grids and other distributed environments: States of the art. Multiagent and Grid Systems: An International Journal 2, 163–182 (2006) 4. Ranjan, R., Buyya, R.: Decentralized overlay for federation of Enterprise Clouds. Handbook of Research on Scalable Computing Technologies, 191 (2009) 5. Fragopoulou, V., Markatos, E.: Failure Management in Grids:The Case of the EGEE Infrastructure (2006) 6. Buyya, R., Abramson, D., Giddy, J., Campus, C., Melbourne, A.: An Economy Driven Resource Management Architecture for Global Computational Power Grids. In: The 7th International Conference on Parallel and Distributed Processing Techniques and Applications (PDPTA 2000), Las Vegas, USA (2000) 7. Krauter, K., Buyya, R.: A Taxonomy and Survey of Grid Resource Management Systems for Distributed Computing. Software: Practice and Experience 32(2), 135–164 (2002) 8. Kaur, D., Sengupta, J.: Resource Discovery in Web-services based Grids. In: World Academy of Science, Engineering and Technology, pp. 284–288 (November 2007) 9. Kim, J., Chandra, A., Weissman, J.B.: Using Data Accessibility for Resource Selection in Large-Scale Distributed Systems. IEEE Transactions on Parallel and Distribution System 20(6), 788–801 (2009) 10. Schopf, J., D’arcy, M., Miller, N., Pearlman, L., Foster, I., Kesselman, C.: Monitoring and discovery in a web services framework: Functionality and performance of the globus toolkit’s mds4. Argonne National Laboratory, Preprint ANL/MCS-P1248-0405 (2005) 11. Wang, C.-M., Chen, H.-M., Hsu, C.-C., Lee, J.: Dynamic resource selection heuristics for a non-reserved bidding-based Grid environment. Future Generation Computer Systems 26(2), 183–197 (2009) 12. Liangxiu, H., Berry, D.: Semantic-Supported and Agent-Based Decentralized Grid Resource Discovery. Future Generation Computer Systems 24(8), 806–812 (2008) 13. Pastore, S.: The service discovery methods issue: A web services UDDI specification framework integrated in a grid environment. Journal of Network and Computer Applications 31(2), 93–107 (2008)
170
A.K. Shaikh, S.M. Alhashmi, and R. Parthiban
14. Campana, S., Litmaath, M., Sciaba, A.: LHC Computing Grid, http://www.grid.org.tr/servisler/dokumanlar/LCG-mw.pdf (accessed on March 2011) 15. Dikaiakos, M.D., Brooke, J., Fellows, D., Garwood, K., Goble, C.: Semantic Matching of Grid Resource Descriptions. In: Dikaiakos, M.D. (ed.) AxGrids 2004. LNCS, vol. 3165, pp. 240–249. Springer, Heidelberg (2004) 16. Chen, L., Tao, F.: An Intelligent Recommender System for Web Resource Discovery and Selection. In: Intelligent Decision and Policy Making Support Systems, pp. 113–140 (2008) 17. Ranjan, R., Harwood, A., Buyya, R.: Peer-to-Peer Based Resource Discovery in Global Grids: A Tutorial. IEEE Communications Surveys & Tutorials 10(2) (2008) 18. Andreasen, T., Bulskov, H., Knappe, R.: From ontology over similarity to query evaluation. In: 2nd CologNET-ElsNET Symposium-Questions and Answer: Theoretical and Applied Perspective 2003, pp. 39–50 (2003) 19. Vidal, A.C.T., Silva, F.J.S., Kofuji, S.T., Kon, F.: Applying semantics to grid middleware. Concurrency and Computation: Practice and Experience 21(13), 1725–1741 (2009) 20. Buyya, R., Murshed, M.: GridSim: a toolkit for the modeling and simulation of distributed resource management and scheduling for Grid computing. Concurrency and Computation: Practice and Experience 14(13-15), 1175–1220 (2002)
Design of Synchronization Model Using Variance Buffer Keun-Wang Lee and Si-Ho Cha* Dept. of Multimedia Science, Chungwoon University San 29, Namjang-ri, Hongseong, Chungnam, 350-701, South Korea {kwlee,shcha}@chungwoon.ac.kr
Abstract. In this paper, we propose a new PBMSM (Petri-net and Buffer for Multimedia Synchronization Model) for the synchronization specification. It can address QoS requirements for jitter that represents the time difference in the media by using Petri-net and variance buffer to express multiple media with respect to multimedia synchronization. Simulation results show the proposed PBMSM improved QoS such as a high playout rate and a low loss rate. Therefore, PBMSM is the appropriate model for multimedia systems that require QoS guarantees. Keywords: Multimedia Synchronization, QoS, Jitter, Petir-net, Buffer.
1
Introduction
Synchronization is a critical prerequisite to ensuring an adequate quality of service in providing multimedia services. The reason behind that lies in the disruption of original timing relations resulting from the occurrence of random delays in mobile networks or mobile host (MH) system clock inconsistencies among inter-media, all of which is attributable to difference in the time of arrival of data being transferred from the server side to the MH over wireless communications networks [1][2][3]. Mobile multimedia service is based on the concept of a wireless extension of wirebased multimedia services, which combines multimedia data connectivity with user mobility over wired networks [4][5]. Contrary to wired networks where the user location is fixed, in wireless networks, a mobile host (MH) has a smaller memory footprint, and a base station (BS) has tighter resource constraints. Due to the constraints of limited memory and resource availability, in particular, MHs are susceptible to buffer underflow or overflow. Therefore, an adaptive synchronization scheme is needed in mobile multimedia networks.
2
Related Work
The problems of the existing model for standard multimedia synchronization model are as follows. *
In OCPN model, the media are severely damaged under the influence of the delayed media. This is because the media are fired only if all media must be reached. Therefore, OPCN model is the synchronization model that is not considered QoS aspects at all [6][7]. RTSM model does not represent a variety of media. It has no way to handle events especially. In addition, if the audio objects arrived late as key media, it is fired by the time media. If the audio object has severe damage, other objects will also inflict considerable damage. Therefore, RTSM model is not regarded as a complete model because it does not consider QoS [6][7]. MMSM model is not sufficiently deal the delayed packets when considering the real-time problem of multimedia as well as the random delay of packet/cell networks. Furthermore, QoS parameters of inter-media and intra-media are also not included in the MMSM modeling [8]. Many existing studies pose significant problems in that delay in wireless networks causes a reduction in the maximum media playout latency. For the I-strategy, frames with end-to-end delays longer than logical time are discarded. For the E-strategy, belatedly arrived frames are presented at the subsequent playout time, so the playout time for all subsequent frames is delayed. However, the I-strategy has a disadvantage; all frames are discarded if frames with end-to-end delays become longer than logical time. Likewise, the E-strategy has some weak points; if frames with end-to-end delays become longer than logical time, all frames withheld from playout are presented at the subsequent playout time. This paper proposes a dynamic synchronization scheme that adjusts intervals between synchronization activities using the maximum available delay variation or jitter, reduces data loss caused by variations in latency, and synchronizes inter-media and intra-media.
3
Proposed Synchronization Model
The proposed model is a new synchronization model to remedy the problems of OCPN, RTSM, and MMSM. It covers analysis and verification of its model conformity using the four properties and two methods of Petri-net for the safety of the model. 3.1
Definition of Synchronization Model
The proposed synchronization model for specifying Petri-net in any BS is defined as follows: The model is specified by the tuple [P, T, K, A, B, D, J, Re, M], where
P = p1 , p 2 ,..., p n ; Regular places (single circles). T = t1 , t 2 ,..., t m ; Transitions. K = k1 , k 2 ,..., k i ; Key places. X = P ∪ K ; All places. A : ( X × T ) ∪ (T × X ) → I , I = 1,2,3,...; Directed arcs.
Design of Synchronization Model Using Variance Buffer
173
B : X → R ' , R ' (Real Number); Variance duration time. ' D : X → R ; Duration time.
J : X → R ' ; (Maximum delay jitter). Re : X → r1 , r2 ,..., rk ; Type of media. M : X → I , I = 0,1, 2; State of places.
Places are used to represent a medium unit and their corresponding actions. The place can have tokens. A pace doesn’t token means that it is currently inactive. A place with a token is active and could be in one of the two states; blocked or unblocked. When a transition is fired and then a token is added into a place, the action of the place is executed, and the token is blocked before finishing the action. The token is unblocked after finishing. Each place has some parameter that determines its relative importance compared with other medium. 3.2
Firing Rule of PBMSM
The PBMSM model can easily represent the synchronization procedure between streams and events, and determine the playing conditions by its precedence relations by correcting the problems of OCPN, RTSM and MMSM model. The simulation procedure for PBMSM playing algorithm is summarized as follows. It starts from the early marking ( M ( p j ) = 1, ∀j : A( p j , t i ) > 0) . That is, all active transitions are fired and create a new marking. The proposed algorithm is shown in Table 1. Table 1. Firing rule of PBMSM 1. begin 2.
M ( p j ) = 1, ∀j : A( p j , t i ) > 0
3. 4. 5.
key medium decision begin if !(Bover or (key medium playout)) then wait
6.
else
7.
if
8.
M ( p j ) < 1, ∀j : A( p j , t i ) > 0 fire
then
t j , t j : A(t j , p j ) > 0
9. 10.
end if end if
11.
M ( p j ) = M ( p j ) − 1, ∀j : A( p j , t i ) > 0
12.
M ( p k ) = M ( p k ) + 1, ∀k : A(t i , p k ) > 0
13. end 14. end
174
K.-W. Lee and S.-H. Cha
As described in firing rule, if transition
ti is fired before removing a token from
its input place, the backtracking is performed for removing forcibly the remained tokens. 3.3
Presentation of Synchronization Model '
Fig. 1 shows an overall view of Petri-nets. 1 (125,3,1,1,1,0,4,1,1,1,1) indicates the information of a control medium. The first parameter 125 refers to the absolute time, 3 the number of key media, and 6 the number of input places. As Control indicates type, Control_in is applicable to all conditions. HS indicates a hierarchical sub module that goes to a lower level module for job processing before going to the next phase. HS goes to a relative duration routine - one of its lower level modules - to find the relative duration before executing After_duration. Therefore, HS has synchronization intervals three times, which results in effective synchronizations in a way that computes the relative duration before executing sync1, computes the compensation time for jitters before executing sync2. Control control_in
(125,3,1,1,1,0,4,1,1,1,1) Control
Check_control
Medium_on
variance_durat ion_time
variance_com pensate_time
HS
HS
Medium
Medium After_ duration
Au1
Au2
After_ jitter
Au2
Control Medium_on
variance_durat ion_time
variance_com pensate_time Medium
Medium I1
After_ duration
Medium_on
variance_durat ion_time
Control
I2 sync 1
After_ jitter
variance_com pensate_time
I3 sync 2
Medium
Medium Ani1
After_ duration
Medium_on
variance_durat ion_time
Ani2
After_ jitter
Ani3
Control variance_com pensate_time Medium
Medium V1
After_ duration
V2
After_ jitter
V3
Fig. 1. Presentation of the synchronization model
Verifications of the proposed multimedia synchronization model have been performed in order to make sure that it corresponds to the analysis methods of Petri-nets including a reachability graph and a matrix-equation.
Design of Synchronization Model Using Variance Buffer
175
3.3.1 Verification of Reachability Graph The reachability tree represents the reachability set of Petri-nets. In the proposed model, the initial marking is (1,1,1,1,0,0,0,0,0,0,0,0) . The transition t1 is generated
t1 is fired, t1 [giving (0,0,0,0,1,1,1,1,0,0,0,0)] is obtainable, and a transition to t 2 is possible. If t 2 is fired, t 2 [giving (0,0,0,0,0,0,0,0,1,1,1,1)] is obtainable. The result of this tree is
from this initial marking. If
represented as follows:
(1,1,1,1,0,0,0,0,0,0,0,0) ↓ t1 (0,0,0,0,1,1,1,1,0,0,0,0) ↓ t2 (0,0,0,0,0,0,0,0,1,1,1,1) 3.3.2 Verification of Matrix-Equation Verifications of the proposed model were performed using matrix-equations as the
D − and D + indicating input function and output function, respectively. Matrices D − and D + are expressed as follows:
second analysis method of the petri-nets. The following are two matrices
111100000000 000011110000 + , D = D− = . 000011110000 000000001111 And the matrix D is expressed as follows: − 1 − 1 − 1 − 11111 D = D+ − D− = . 0000 − 1 − 1 − 1 − 1 The result of applying the equation the following:
x is ( x1 , x 2 ) and − 1 = −1 ⋅ x1 , 0 = x1 − x 2 , 1 = x 2 . For this, x1 = 1 , x2 = 1 . Therefore, x1 − x2 = 0 .
Where
Through the verification described above, the proposed synchronization model proved to be consistent with the reachability graph and matrix-equation.
176
4
K.-W. Lee and S.-H. Cha
Synchronization Scheme
This paper proposes a jitter strategy complementing the shortcomings of both the I-strategy and E-strategy in which belatedly arrived frames wait for as long as delay jitter times until they are presented, instead of being deleted unconditionally or waiting indefinitely until they are presented at the next playout time. Fig. 2 shows the proposed jitter strategy that complements the shortcomings of both the I-strategy and E-strategy. As depicted in Fig. 2, the frames b and c are played out in the units 4 and 6 by compensating for the maximum latency or jitter. The frame f indicates that it has not arrived within the variable delay jitter time. In this case, the frame f cannot be played out even if it waits for as long as the variable latency or jitter times. Therefore, the unit 8 in the frame f indicates that it cannot be compensated for the maximum delay jitter time due to excessive delay.
Fig. 2. Proposed strategy applied under reduced network traffic
Fig. 3 shows that the unit 4 in the frame c was skipped due to excessive delay and the unit 5 in the frame d was compensated by applying the maximum delay jitter time.
Fig. 3. Proposed strategy applied under heavy network congestion
Design of Synchronization Model Using Variance Buffer
5
177
Performance Evaluation
With focus on the playout time and loss time, we performed a comparative analysis between the existing scheme and the delay jitter & playout scheme using the maximum delay jitter time. Comparisons among the I-strategy, E-strategy, and the proposed strategy were made in consideration of the following two cases; Circumstances of reduced network traffic and circumstances of heavy network congestion. As a result, the proposed strategy proved to be vastly superior to other two strategies. For the purpose of this paper, we have assumed that the average delay is 100ms and the variance is 20ms in the event of delay in the audio stream. Fig. 4 shows the result of comparisons of playout rates among the I-strategy, E-strategy, and the proposed strategy that were applied under reduced network traffic. The playout rates of the proposed strategy were obtained by conducting experiments ten times. The proposed strategy showed more improved playout rates than the I-strategy and E-strategy by 18.36% and 8.83%, respectively. As shown in Figure 5, under heavy network congestion, the proposed strategy showed more improved playout rates than the I-strategy and E-strategy by 16.57% and 7.38%, respectively.
Fig. 4. Comparison of playout rates under reduced network traffic
6
Fig. 5. Comparison of playout rates under heavy network congestion
Conclusion
This paper has proposed a scheme for implementing intra-media and inter-media synchronizations by means of smooth buffering at the BS in mobile environments. The proposed scheme delivered optimized synchronizations without causing any degradation of quality of service. The superiority of the proposed scheme was demonstrated by extending intra-media synchronization intervals using the maximum delay jitter time of the audio media as a key medium, as well as by synchronizing irregularly arriving packets within the extended intervals through applications of the
178
K.-W. Lee and S.-H. Cha
said maximum delay jitter time to inter-media synchronizations. Furthermore, it allows us to take 3D animation training for electrical safety while on the road. Future work needs to focus not only on standard schemes for mobile multimedia synchronizations which take interactions with users into account, but also on optimized synchronization mechanisms which employ minimum buffering.
References 1. Oh, M.-K., Lee, K.-W.: Design and Verification of Multimedia Synchronization Model using PetriNet. Journal of The Korea Academia-Industrial Cooperation Society 11(2) (2010) 2. Boukerche, A., Hong, S., Jacob, T.: MoSync: A Synchronization Scheme for Cellular Wireless and Mobile Multimedia Systems. In: Proc. of the Ninth International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems. IEEE (2001) 3. Lee, K.-W., Lee, J.-H., Lee, K.-H.: Multimedia Synchronization Model for Two Level Buffer Policy in Mobile Environment. In: Yazıcı, A., Şener, C. (eds.) ISCIS 2003. LNCS, vol. 2869, pp. 1060–1068. Springer, Heidelberg (2003) 4. Blakowski, G., Steinmetz, R.: A Media Synchronization Survey: Reference Model, Specification, and Case Studies. IEEE J. Selected Areas in Communications 14(1) (1996) 5. Nam, D.-H., Park, S.-K.: A Smooth Playback Mechanism of Media Streams in Mobile Computing Environment. In: ITC-CSCC 1998 (1998) 6. Yang, C.-C., Huang, J.-H.: A Multimedia Synchronization Model and Its Implementation in Transport Protocols. IEEE J. Selected Areas in Communications 14(1) (1996) 7. Steinmetz, R.: Human Perception of Jitter and Media Synchronization. IEEE J. Selected Areas in Communications 14(1) (1996) 8. Lee, K.-W., Cho, H.-S., Lee, K.-H.: Mobile-Based Synchronization Model for Presentation of Multimedia Objects. In: Bubak, M., van Albada, G.D., Sloot, P.M.A., Dongarra, J. (eds.) ICCS 2004. LNCS, vol. 3036, pp. 381–388. Springer, Heidelberg (2004)
Mobility Management Scheme for the Wireless Mesh Network Using Location Server and Mobility Anchor Point Sung Hwan Jang and Gi Sung Lee School of Information & Communication Engineering, Sungkyunkwan University, 300 Cheoncheon-Dong, Jangan-Gu, Suwon, Gyeonggi-Do 440-746, Republic of Korea School of Computer Game, Howon University, 64 Wolha-ri, Impi-mveon, Gunsan-si, Jeollabuk-do 573-718, Republic of Korea [email protected], [email protected]
Abstract. Some WMN researches support mobility using router protocol such as AODV and DSR because WMN has the similar features with Ad-hoc network in terms of operation. However, WMN has got different operations with additional functions of multi-interface and internet connect, it can't be a good method to manage efficient mobility any more. This study suggests method to apply Location Server and MAP function of HMIPv6 to WMN as a way to manage WMN mobility. Through simulation, operation was verified and function was appreciated. Through the results of evaluation, the suggest mobility management had shorter handover delay and packet loss reduced, which improved whole network throughput rate, compared with mobility management using AODV. Keywords: Wireless Mesh Network, Hierarchical Mobile IPv6, Mobility Anchor Point, Mobility Management Scheme, Location Server.
1
Introduction
Compared with the existing wired infrastructure, Wireless Mesh Network (WMN)[1] can be quickly established at inexpensive cost by making use of wireless communications technologies. As shown in Fig. 1, WMN comprises Mesh Routers (MR) and Mesh Clients (MC). A MR, which has usually minimum mobility, forms a wireless backbone and serves as a bridge and a gateway, which help to connect different networks. A MC, which can be stationary or mobile, connects to a MR to receive services. WMN is similar to ad hoc network in the aspect that it features multi hop communications. But, ad hoc network, which communicates between mobile nodes without infrastructure, is connected to WMN and thereby uses the existing single-hop wireless network’s ad hoc routing protocols such as Ad-hoc On-demand Distance Vector (AODV)[2] and Dynamic Source Routing (DSR)[3], so it has difficulties in using handover as it was. Accordingly, it is necessary to find a mobility management scheme suitable to WMN.
In order to provide handover for WMN, this thesis proposes a mobility management scheme for WMN on the basis of Location Server (LS) and Hierarchical Mobile IPv6[4]’s Mobility Anchor Points (MAP). Since the existing client WMN is classified as subnet and MR manages its subnet information, handover delay and packet losses, which are caused by MC’s mobility, reduces. Following this introduction, chapter 2 of this thesis consists of related works, which are conducted on HMIPv6, and the current WMN mobility management schemes. Chapter 3 will explain the algorithm proposed in this thesis, and chapter 4 will show the analysis of the performance results about the proposed method. Lastly, chapter 5 sums up and concludes this thesis.
2
Related Works
HMIPv6 provides regional mobility by defining MAP with new functions and making a hierarchical network structure, and it also defined Regional CoA(RCoA), which a mobile node creates based on MAP domain’s prefix, and On-Link CoA (LCoA), which a mobile node creates based on Access Router(AR)’s prefix. RCoA us a temporary address presenting MAP’s subnet position and LCoA is a temporary address on the basis of a physical location of Access Router (AR). Studies on schemes for WMN other than general mobility management schemes have been conducted, and they are largely divided into Tunnel-based, Routing-based, and Multicast-based schemes. Tunnel-based scheme has Ant[5] and Mesh Mobility Management(M3)[6]; Routing-based scheme has iMesh[7] and Mobile Party[8]; Multicast-based scheme has SMesh[9]. Details about each scheme are presented in Table 1.
Mobility Management Scheme for the Wireless Mesh Network
181
Table 1. 1 Existing Mobility Management Schemes
Type Layer LocationServer Routing Macro mobility Overhead reason Overhead
3
Ant Tunnelbased Layer-3 Yes
M3 Tunnelbased Layer-3 Yes
iMesh Routingbased Layer-2+3 No
Mobile Party Routingbased Layer-3 No
SMesh Multicasttbased Layer-3 No
OLSR No
No No
OLSR No
Mobile Party No
No Yes
Location update Normal
Hierarchical delay Normal
Routing
Routing
Normal
High
Group managemeent High
Mobility Manag gement Scheme in the WMN
Fig. 2 presents a logical stru ucture of WMN proposed in this thesis. Internet compriises Gateway Mesh Routers refflecting MAP function (MAP-GMR) and Location Servver, and Infrastructure/Backbon ne WMN consists of MAP-GMR and MR. Client WM MN consists of MCs and is linkeed to MR to receive communications support.
Fig. 2.. Logical Structure of the proposed WMN
Table 2 indicates a messsage newly proposed and changed for the mobility m management scheme of this thesis. As MC enters the networrk, MR receives Router Advertisement messages which are periodically broadcast. If MR M fails to receive the messages, MC broadcast a Rouuter Solicitation message. Wh hen MR receives the message, it transmits MC C a RouterAdvertisement messaage including MAP-GMR’s prefix and its own prefix. M MC uses MR and MAP-GMR’s prefix to create LCoA and RCoA.
182
S.H. Jang and G.S. Lee Table 2. Message information
Message
Field/Flag
Signification
I
Intermediate MCs LCoA update
C
Intermediate MCs Hop Count update
Path Solicitation
S
Destination MC RCoA request
Location Update
R
RCoA update
S
Destination MC RCoA request
A
Acknowledgement (Destination MC RCoA)
Local Binding Update
Location Solicitation Tunnel Setup
T
Tunnel Setup
R
RCoA update
After that, the Local Binding process in which MC is registered in MR and MAPGMR is displayed. MC which entered the network transmits a Local Binding Update message to MR. If it is linked to MR through multi hops, it updates both LCoA and hop count of relay MCs into the message’s I field and C flag and then transmit the updated message. MR updates MC’s LCoA and the LCoA and hop count of relay nodes into Visitor List, and transmits a Local Binding Update Message to MAP-GMR. MAP-GMR performs Duplicate Address Detection (DAD) process regarding RCoA, and updates MC’s RCoA and LCoA into Binding Cache. After that, MAP-GMR sends MC a Local Binding Acknowledgement, and if MR is linked with MC through multi hops, relay MCs’ LCoA is updated into relay MCs’ Binding Cache. In this process, a tunnel between MAP-GMR and MR is established. After that, MAP-GMR sends LS a Location Update message containing MC’s RCoA. LS saves MC’s RCoA into its Location Table, and sends MAP-GMR a Location Update Acknowledgement. The process to start communicating between MCs is presented as follows. To communicate with target MC, source MC sets the S flag of a Path Solicitation message as 1 and then sends it to MAP-GMR. MAP-GMR identifies if the target MC’s RCoA exists in its own Binding Cache, and unless the target MC belongs to its own network, then the MAP-GMR transmits a Location Solicitation message to LS. The LS sends MAP-GMR a Location Solicitation Acknowledgement message containing the information on the target MC. After that, MAP-GMR sends the source MC a Path Solicitation Acknowledgement message containing the target MC’s RCoA. The source MC which received the message updates the target MC’s RCoA into its own Binding Cache, and then transmits the packet. When MC performs micro handover to other client WMN, it receives a Router Advertisement Message which nMR periodically broadcasts. MC creates a new LCoA with nMR’s prefix, but does not create RCoA since the received prefix is equal to the MAP-GMR’s prefix of the MC. And then, the Local Binding process is performed just as the process on the network entrance was. When Binding Life Time expires, pMR prior to MC’s movement and the paths saved into relay MCs’ Visitor List and Binding Cache are removed.
Mobility Management Scheme for the Wireless Mesh Network
183
In the case of Macro handover, MC uses nMR’s prefix and nMAP-GMR’s prefix of a Router Advertisement message received from nMR to create a new LCoA and RCoA. And then, Local Binding process is to be performed. The nMAP-GMR sends LS a Location Solicitation. The LS checks out Location Table, and sends nMAP-GMR a Location Solicitation Acknowledgement message containing pMAP-GMR’s information. The nMAP-GMR sets T flag as 1, saves the targe MC’s new RCoA into R field, and then sends pMAP-GMR a Tunnel Setup message. The pMAP-GMR updates a previous RCoA and a new RCoA into its Binding Cache, and sends a Tunnel Setup Acknowledgement message to the nMAP-GMR. In this process, a tunnel between the pMAP-GMR and the nMAP-GMR is established. When the target MC performs handover, it does not let the source MC register the target MC’s new RCoA, and packets are sent to the target MC through the tunnel between MAP-GMRs.
4
Performance Analysis
The ns-2[10] was used to evaluate the performance of the mobility management scheme proposed in this study. Multi-interfaces and multi-channel allocation function was applied to the existing 802.11 based network to implement WMN, and AODV was used as a routing protocol. In addition, MAP function was applied to mobilwan [11] based WMN to implement the proposed WMN. Fig.3 presents the simulation composition.
Fig. 3. Simulation Structure
Constant Bit Rate (CBR) is used as a packet type, and a packet with 200byte is transmitted in a 20ms cycle. As for Wireless MAC, IEEE 802.11b was applied. The target MC is linked to MR1 one second after the simulation starts; 16 seconds after the simulation, the target MC performs micro handover from MR1 to MR2; at the 31 seconds after the simulation, a macro handover is performed from MR2 to MR3; and finally, at the 46 seconds after the simulation, a micro handover is performed from MR3 to MR4.
184
S.H. Jang and G.S. Lee
4.1
Performance Analysis
Fig. 4 displays the comparative graph about a delay as the existing WMN and the WMN reflecting the proposed mobility management scheme perform handover, respectively.
Fig. 4. Handover Delay
In the case of the WMN using AODV, when the target MC performs its mobility in the same network, a delay records about 165.28ms and 227.32ms, and when it moves to a different network, a delay records about 1173.79ms. The reason for the high record of delay in the latter case is that the AODV searches for the paths of the entire network again to explore a new path, so the time loss increases. In the case of the proposed mobility management scheme, when the target MC performs micro handover, a delay records about 82.17ms and 82.16ms, when it performs macro handover, a delay records about 102.34ms. The reason for the high record of delay on the performance of macro handover is that the target MC’s network changed, and thus the time of registering mobility information to LS and of establishing a tunnel between MAP-GMRs was added. Fig. 5 presents the graph about packet losses caused by handover.
Fig. 5. Packet losses caused by handover
In the case of the existing WMN using AODV, packets are lost a lot as a node moves to a different network. When the distance between MRs is close, searching
Mobility Management Scheme for the Wireless Mesh Network
185
paths goes fast, but when it is far, the path searching time increases because the entire network should be sheared for a path. In the case of the proposed mobility management scheme, there was no difference in packet losses caused both by micro handover and by macro handover. That is because nMAP-GMR does not let the source MC register the target MC’s mobility, but let only pMAP-GMR register it, and so the time of reconfiguring a path reduces.
5
Conclusions
This thesis proposed MAP function to manage mobility in the existing WMN and a method of applying LS to the existing WMN, and analyzed the performance of the mobility management in the existing WMN and of the proposed mobility management in the WMN. The proposed mobility management scheme reflecting MAP function showed no big difference of handover delay regarding micro handover and macro handover, as MC performs handover. Regarding packet loss rate, the proposed scheme also had better rate than the WMN using the existing ad hoc routing protocol. In addition, in the case of the proposed scheme, the delay of micro handover and macro handover recorded about 82 ~ 102ms, a figure which satisfies the maximum delay of 150ms [12] necessary to provide real-time service. Therefore, the proposed mobility management scheme turned out to be suitable for real-time service. Acknowledgements. This work was supported by the Howon University Fund. 2011.
References 1. Akyildiz, I.F., Wang, X., Wang, W.: Wireless mesh networks: a survey. Computer Networks 47, 445–487 (2005) 2. Perkins, C., Royer, E.B., Das, S.: Ad hoc On-Demand Distance Vector (AODV) Routing. IETF RFC 3561( 2003) 3. Johnson, D., Hu, Y., Maltz, D.: The Dynamic Source Routing Protocol (DSR) for Mobile Ad Hoc Networks for IPv4. IETF RFC 4728 (2007) 4. Soliman, H.: Hierarchical Mobile IPv6 mobility management. Internet Draft (2004) 5. Wang, H., Huang, Q., Xia, Y., Wu, Y., Yuan, Y.: A network-based local mobility management scheme for wireless mesh networks. In: IEEE Wireless Communications and Networking Conference, pp. 3792–3797 (2007) 6. Huang, R., Zhang, C., Fang, Y.: A mobility management scheme for wireless mesh networks. In: IEEE Global Telecommunications Conference, pp. 5092–5096 (2007) 7. Navda, V., Kashyap, A., Das, S.R.: Design and evaluation of iMesh: an infrastructuremode wireless mesh network. In: Sixth IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks, pp. 164–170 (2005) 8. Sabeur, M., Sukkar, G.A., Jouaber, B., Zeghlache, D., Afifi, H.: Mobile party: A mobility management solution for wireless mesh network. In Third IEEE International Conference on Wireless and Mobile Computing, Networking and Communications, pp. 45–53 (2007)
186
S.H. Jang and G.S. Lee
9. Amir, Y., Danilov, C., Hilsdale, M., Musaloiu-Elefteri, R., Rivera, N.: Fast handoff for seamless wireless mesh networks. In: Proceedings of the 4th International Conference on Mobile Systems, Applications and Services, pp. 83–95 (2006) 10. Network Simulator version 2, http://www.isi.edu/nsnam 11. MobiWan: NS-2 extensions to study mobility in Wide-Area IPv6 Networks, http://www.inrialpes.fr/planete/pub/mobiwan 12. The VOIP quality standard applying the wideband network, Telecommunications Technology Associations, TTAK.KO-01.0136 (2008)
The Comparative Study for ENHPP Software Reliability Growth Model Based on Mixture Coverage Function Hee-Cheul Kim and Hyoung-Keun Park Department of Industrial & Management and Electronics Engineering, Namseoul University, Korea {kim1458,phk315}@nsu.ac.kr
Abstract. Finite failure NHPP models presented in the literature exhibit either constant, monotonic increasing or monotonic decreasing failure occurrence rates per fault. Accurate predictions of software release times and estimation of the reliability and availability of a software product require quality of a critical element of the software testing process: test coverage. This model called Enhanced non-homogeneous Poisson process (ENHPP). In this paper, exponential coverage and S-shaped (Yamada-Ohba-Osaki) model was reviewed, proposes modified mixture model, which make out efficiency application for software reliability. Algorithm to estimate the parameters used to maximum likelihood estimator and bisection method, model selection based on SSE statistics for the sake of efficient model, was employed.1 Keywords: Test Coverage, ENHPP, Mixture model, SSE.
1
Introduction
In the software testing phase, if the access repair software failure phenomenon by number of software failures and time between failures, the valuation of the software can be more easily. Reliability assessment measures (the number of software failures, software fault occurs every time, software reliability and failure rate) is estimated that in the future failure times may predict. To modeling the number of software failures in the observation time, non-homogeneous Poisson process (NHPP) is widely have been using qualitative. These NHPP, intensity function and the mean function are dependent on [1]. In order to release times of software forecast or predict the effectiveness and reliability, using important factor, test coverage, in the process of software testing, effective testing can work. These models from the existing NHPP models included test coverage. In this model, ENHPP (Enhanced non-homogeneous Poisson process) is called [2]. In this study, the lifetime distribution for finite failure ENHPP model with the exponential model and the Goel-Okumoto S- shaped pattern and mixture models used, present a measure to estimate the reliability based on this to evaluate the characteris1
Funding for this paper was provided by Namseoul University.
tics of reliability. Section 2 of this paper as a test-related research coverage, simple coverage and proposed mixture model was described. Parameter estimation methods of ENHPP model in section 3 was listed. In section 4, the actual failure data for each model was analyzed. Finally, conclusion was listed in section 5.
2
Related Researches
2.1
Test Coverage on Software Reliability
Test coverage on software reliability can be applied to finite failure NHPP. This applies to the extended model is called ENHPP. ENHPP model developed states that the rate at which the faults are detected is proportional to the product of the rate at which potential fault sites are covered and the expected number of remained faults, the ENHPP model is based on the following basic assumptions [2]: (A1) Faults are uniformly distributed over all potential faults sites. (A2) When a potential fault- site is covered, any fault present at that site is detected with probability K (t ) . (A3) Repairs are affected instantly and without the introduction of new faults. This assumption is the same as in case of finite NHPP models. (A4) Coverage is a continuous monotonic non-decreasing function of testing time. d m(t ) d c(t ) = a K (t ) dt dt
(1)
or t
K (τ ) c ' (τ ) dτ
m(t ) = a
,
(2)
0
Where a is defined as the expected number of faults that would be detected given infinite testing time, perfect faults detection coverage implying K (t ) = 1 , and complete test coverage leading to lim t →∞ c(t ) = 1 . If K (t ) is assumed to be a constant value K , denoted by m(t ) is given by m(t ) = a K c (t )
(3)
Equation (3) is intuitively simple – the expected number of faults detected by time is equal to the expected total number of faults in the product times the probability of detecting a fault given that it is covered times the fraction of potential faults sites covered by time . Substituting a = a K , Equation (3) can be written as m(t ) = a c(t )
This results in a failure intensity function λ , given by
(4)
The Comparative Study for ENHPP Software Reliability Growth Model λ (t ) =
d m(t ) = ac '(t ) dt
189
(5)
The failure intensity function can also be rewritten as [2, 3]: λ (t ) = [a − m(t )]
c '(t ) 1 − c(t )
(6)
Equation (6) clearly shoes that failure intensity depends not only on the number of remaining potential fault sites are covered divided by the current population of uncovered potential fault sites. From Equation (6), the failure occurrence rate per fault or hazard function h(t ) is given by h(t ) =
c '(t ) 1 − c(t )
(7)
Therefore, the testing process models as ENHPP is each rate of failure that can occur because of the time-varying time-dependent as the rate of failures can occur. In addition, this hazard function at time t that corresponds to the distribution function can be evaluated in the coverage. These ENHPP model for (incomplete) coverage for the function can be applied to practical situations c(∞) < 1 . As a result, the conditional reliability R(t | s) is known as follows [2, 3]: R(t | s ) = e
Where
2.2
t denote
−
s +t
s
mission time and
λ (t ) dt
s is
=e
−
s +t
s
a c '(t )
= e − a[c( s + t ) − c( s )]
(8)
the last failure time.
NHPP Model with Coverage Function (ENHPP)
2.2.1 Exponential Coverage Function Well-known Goel-Okumoto model [2, 3] is the basic model in this field. The failure occurrence rate per fault is constant. If access to these NHPP model, the mean value function and intensity function is known as follows: m(t | a, g ) = a (1 − e − g t ) (a > 0, g > 0)
λ (t | a , g ) = a g e − g t
(9) (10)
Where a can be denoted the expected number of faults that would be detected given infinite testing time in case of finite failure NHPP model and g is search ratio rate per fault. Therefore, the coverage function c (t ) and the hazard function h (t ) are derived as follows: c (t ) = 1 − e − g t
(11)
190
H.-C. Kim and H.-K. Park
h(t ) =
c '(t ) =g 1 − c(t )
(12)
2.2.2 S- shaped Coverage Function The S-shaped coverage function was known as Yamada-Ohba-Osaki the model [2, 4]. If access to these NHPP model, the mean value function and intensity function is known as follows: m(t | a, g ) = a[1 − (1 + gt )e− g t ] (a > 0, g > 0)
(13) (14)
2 −g t
λ (t | a , g ) = a g t e
Therefore, the coverage function function is derived as follows:
c (t )
and the hazard function
h (t )
for the S-shaped
c(t ) = 1 − (1 + gt )e − g t
h(t ) =
2.3
(15)
2
c '(t ) g t = 1 − c(t ) 1 + g t
(16)
Proposed Mixture (Transformation) Coverage Function
Software systems grow in complexity, it may cause a malfunction of one does not happen only by the intensity function is several possible causes could be detected. In this case, if you know the intensity function and the mean value function can be modeled. In other words, case of the distribution of the lifetime of components or systems with different groups of k subpopulations, mixture model can be applied to in field. For example, a combination of the entire population with parts or manufacturing defects early failure group and the normal group, consisting of two groups of cases are those in the mixed model. In this mixed-model, probability density function and the distribution function consist of the form of a weighted sum. In other words, the distribution function and probability density function is expressed as follows [6]: F (t | Θ, p ) = Pr(T < t ) =
Where
k i =1
k i =1
pi Fi (t | θi )
(17)
pi = 1 .
Therefore, the mean value function and intensity function of the coverage mixed pattern for exponential and S- shaped can be expressed as follows: −g −g λ (t ) = pλ 1(t ) + (1 − p)λ 2 (t ) = a[ pg1 e 1t + (1 − p) g22 t e 2 t ]
m(t ) = pm 1(t ) + (1 − p) m 2 (t ) = ap (1 − e
Where a > 0 , p ≥ 0
g1 > 0
and g 2 > 0 .
−g t 1
) + a(1 − p)[(1 − g2 t ) e
−g t 2
(18) )]
(19)
The Comparative Study for ENHPP Software Reliability Growth Model
Therefore, the coverage function c(t ) and the hazard function sition function is derived as follows: c(t ) = ( p − 1)(1 + g 2 t ) e
h (t ) =
pg1 e pe
3
−g t 1
−g t 1
−g t 2
− pe
−g t 1
2
−g t 2
+ (1 − p ) g 2 t e
+ (1 − p )(1 + g 2 t ) e
191
for the superpo-
h(t )
+1
(20) (21)
−g t 2
Parameter Estimation for ENHPP
Given the time truncated model testing until time t , the order epochs of the observed n failure times are denoted by x1 , x2,, xn . The likelihood function is given by
∏
LNHPP ( a, g | Dt ) = [
n k =1
λ ( xk )exp(− m(t )]
(22)
Where D t = {x1 , x2,, xn ; t} is the data set? In this paper, using the maximum likelihood estimation (MLE) parameter estimation is carried out. Log likelihood function on exponential coverage model, using equation (22) is expressed as follows:
∏
LE xp (a, g | Dxn ) = (
n k =1
age
−g x
k )exp[− a (1 − e − g xn )]
(23)
To use the MLE this log likelihood function is derived as follows. ln LE xp (a, g | Dxn ) = n ln a + n ln g − g
Using equation (24), the likelihood estimators following equation:
aˆ MLE
n − gˆ xn =1− e aˆ n = gˆ
n k =1
xk + aˆ xn e
n k =1
xk − a(1 − e
and
−g x
gˆ MLE
n)
(24)
must satisfy the
(25) − gˆ xn
(26)
Using (25) and (26) equation, numerically solve the two parameters can be obtained Similarly, the likelihood estimators aˆ MLE and gˆ MLE for S- shaped (Yamada-OhbaOsaki) must satisfy the following equation: n − g xn = 1 − e− g − gˆ xn e a n = g
n k =1
xk + a gxn 2 e
− gxn
(27) (28)
192
H.-C. Kim and H.-K. Park
When using equation (18) and (19), the likelihood function for mixture model is expressed as follows:
∏
LMix (a, g | Dxn ) = [
n k =1
pλ 1( xk ) + (1 − p )λ 2 ( xk )] exp[ pm 1( xn ) + (1 − p ) m 2( xn )]
(29)
−g −g Where λ 1(t ) = a g1 e 1t , m 1(t ) = a (1 − e 1t ) , a > 0 , g1 > 0 and g 2 > 0 .
λ 2 (t ) = a g 2 2 t e
−g t 2
, m 2(t ) = a[1 − (1 + g2 t ) e− g 2 t ] ,
p is
constant.
Using equation (29), the likelihood estimators aˆ MLE , pˆ MLE satisfy the following equation:
, gˆ1MLE and
gˆ 2MLE
−g x −g x n = ( p − 1)(1 + g 2 xn ) e 2 n − pe 1 n + 1 a
4
−g x −g x 1 k − pg 1 k −g x 1xk e = pa xn e 1 n −g x −g x −g x k =1 2 2 g1 e 1 k p − p g 2 xk e 2 k + g 2 xk e 2 k
( p − 1)[ g 2 xk e
n k =1
(30)
pe
n
g1 e
−g x 1 k
2
−g x 2 k
p − p g 2 xk e
(2 − g 2 x k )]
−g x 2 k
2
+ g 2 xk e
−g x 2 k
= a (1 − p )(1 − g 2 xn 2 ) e
must
−g x 2 n
(31)
(32)
Software Failure Data Analysis
In this chapter, using actual fault data is to analyze the model. Failure data NTDS [10] (Naval Tactical Data System) is caused by a software failure data Goel and Okumoto [3], Mazzuchi and Soyer [7, 8 and 9] and using the failure data has proposed a software model. This study will also want to use this data. In this paper, for the parameter estimation was performed using the MLE estimation. Nonlinear equations, numerical analytical method of calculating was used to bisection method. These calculations and the initial value 1 and 10−5 , tolerance for width of interval 10−5 , ensuring the availability of sufficient convergence and repeated 100 times, using C-language, parameter estimation was carried out. In table 1, result of parameter estimation for Goel-Okumoto model coverage, S- shaped (Yamada-Ohba- Osaki) coverage, depending upon the mixed is summarized. Table 1. Parameter Estimators of each model
Table 2. SSE of each model
The Comparative Study for ENHPP Software Reliability Growth Model
193
In general, model selections in this area as a way of model bias trend (mean squared error; SSE [9]) are available. In Table 2, this result of SSE is summarized. In this table, mixture model are relatively efficient compared to Goel-Okumoto model and S- shaped coverage as p-value decreases.
Fig. 1. Hazard function of each model
Fig. 2. Hazard function of mixture model
Fig. 3. Coverage function of each mode
Fig. 4. Reliability of each mode
In Figure 1, the hazard function for each model is a drawing. In this figure, the exponential coverage model is constant, S-coverage model and mixed model (p = 0.5) have the form of increases shows. In Figure 2, the hazard function for the mixed model as p-value decreases have a high probability pattern. In Figure 3, the coverage function for each model is a drawing. In this figure, ENHPP assumed in the model as a continuous forging is shown in the form of non-decreasing function. In Figure 4, patterns for the reliability are discussed. This figure, too, has shown highly mixed model.
194
H.-C. Kim and H.-K. Park
Acknowledgment. Funding for this paper was provided by Namseoul University, Korea. Dep. Industrial & Management and Electronics Engineering is also gratefully acknowledged.
5
Conclusions
In this paper, ENHPP of mixed model, existing NHPP models to include test coverage, were studied. In other words, the superposition and mixed model of the failure occurrence rate per fault with reduction form were suggested. Using time between failure data, maximum likelihood estimation for existing models, mixed model were carried out. Numerical example, the relatively large number of faults measured data, NTDS, was analyzed. Using mean squared error, the comparing results of efficient model, mixed model than existing models have shown that better. Thus, mixed ENHPP model that can be used effectively in the field of the model was able to find that.
References 1. Lawless, J.F.: Statistical Models and Methods for Lifetime Data. John Wiley & Sons, New York (1981) 2. Gokhale, S.S., Trivedi, K.S.: A time/ structure based software reliability model. Annals of Software Engineering 8, 85–121 (1999) 3. Goel, A.L., Okumoto, K.: Time-Dependent Error-Detection Rate Models for Software Reliability and Other Performance Measures. IEEE Trans. on Reliability R-28(3), 206–211 (1979) 4. Yamada, S., Ohba, M., Osaki, S.: S-Shaped Reliability Growth Modeling for Software Error Detection. IEEE Trans. on Reliability R-32(5), 475–485 (1983) 5. Cinlar, E.: Introduction to Stochastic Process. Prentice-Hall, New Jersey (1975) 6. Berry, D.A., Stangl, D.K.: Bayesian Bio-statistics, pp. 601–616. Marcel Dekker Inc. (1996) 7. Mazzuchi, T.A., Soyer, R.: Bays Empirical Model for Software Reliability. IEEE Transactions on Reliability 37, 248–254 (1988) 8. Gokhale, S.S., Philip, T., Marinos, P.N., Trivedi, K.S.: Unification of Finite Failure NonHomogeneous Poisson Process Models through Test Coverage. In: Proc. of Int. Symposium on software Reliability Engineering, White Plains, NY (1996) 9. Horgan, J.R., London, S., Lyu, M.R.: Achieving Software Quality with Testing Coverage Measure. IEEE Computer, 60–69 (September 1994) 10. Kuo, L., Yang, T.Y.: Bayesian Computation of Software Reliability. Journal of the American Statistical Association 91, 763–773 (1996)
A Study on Authentication and Access Control of Wireless Access Point Jong-Kyung Baek1 and Jae-Pyo Park2 1
Course of Computer Science, Soongsil University, 1-1 Sangdo-Dong, Donggak-Ku, Seoul, Korea 2 Graduate School of Information Science, Soongsil University, 1-1 Sangdo-Dong, Donggak-Ku, Seoul, Korea {jkbaek,pjerry}@ssu.ac.kr
Abstract. When a client is accessing a wireless AP, many methodologies which prevent from external attack or internal information leakage using only authorized wireless AP after deciding authorized/unauthorized have proposed. But there are securities and cost problem in server based authentication, and availability problem in client based authentication. In this paper, we use MAC Address to get an authentication from an AP. It searches the NDIS Intermediate Driver from the wireless network card, and then controls the packets after operating the scope of IP and PORT. We have implemented the proposed model, and it showed no more security and cost problem. It also showed that client based authentication availability is very flexible and has high scalability. Keywords: Wireless Authentication, Access Control, Network Driver, MAC Address, Access Point.
1
Introduction
These days, Smartphone, Wibro(Wireless Broadband Internet), HSDPA(High Speed Downlink Packet Access), and other wireless technologies are developed and widely used. Because of that, the number of attacking wireless device vulnerability continues to grow. Leaking internal information to external case has also found using wireless AP mobility and scalability. There are proposals such as server based control against wireless AP attack, wireless sensor used, client side control, etc. But these have security vulnerabilities and less controlling availability in external network. In this paper, we propose that the client decides authorized/unauthorized AP by authenticating wireless AP and we increase strength of the security by controlling the unauthorized AP packets and also increase availability by using part of IP and PORT.
goes out externally. Only internal wireless AP is controllable but external wireless AP. It is hard to control if an internal authorized person takes a notebook out to external and send information. We need additional device to monitor packets, therefore it is expensive than client based. The controlling method using device controls by placing a sensor which can control wireless AP in controlling area. 500 AP's can be controlled per a sensor and it is possible to defend around 20 wireless threats at the same time. Because of the sensor are placed in each areas, it is easy to track problems. But it is very vulnerable when the sensor goes out of the range or internal authenticated person sends information after taking the notebook to the external area. It costs a lot by adding sensors to broadening the range. Client based control is needed to supplement the limitation and the cost. Client decides if the AP is authorized or not with MAC Address whenever client tries to get access from the AP. But it has less availability by using only permit/block against wireless AP. An event though unauthorized wireless AP, it needs availability in secure way. Also, because of the client based, it has vulnerable security by attacking malicious code, hacking tool and reversing tool.
3
Wireless AP Authentication and Control Method
3.1
Security Model Configuration and Definition
In this paper, we have configured security model to authenticate and control wireless AP. The model is shown as follows in Fig.1
Fig. 1. Wireless AP Authentication Security Model
There are three different modules such as Server(AW-Server), Client User Lever(AW-Client), Client Kernel Level(NDIS Intermediate). Network transmitting packet goes out like User-level Client, TDI Layer, NDIS Layer, Network device in this order. AW-Server receives requested wireless authenticating AP information from AWClient and send only authorized AP information to AW-Client. AW-Client receives authorized wireless AP list from AW-Server, after that client decide if the AP is
A Study on Authentication and Access Control of Wireless Access Point
197
authorized when it access to wireless AP. It controls network packets by the policy after receiving unauthorized wireless AP diver name. 3.2
Wireless Authenticating Process in Server
After checking authorized list from AW-Client, and it authorized internal network wireless AP, and then it updates and send it to AW-Client. Client update the list as soon as it receives, it deletes unauthorized wireless AP information. 3.3
Wireless AP Authentication Process in Client
When SW-Client is initiating, it receives authorized wireless AP information and stores it in the memory. If it monitors wireless AP access and detect access request, it checks the network card and finds connected network card. It requests MAC Address to related wireless network card and decide if it is authorized wireless AP. In case of unauthorized wireless AP, if it finds device name by searching network card, it sends network card's MAC Address to NDIS Filter Driver using the device name. When NDIS Filter Driver sending a data, if it is unauthorized network card MAC Address, it controls the network packet by the unauthorized policy. 3.4
NDIS Intermediate Driver Packet Control Policy
Unauthorized wireless AP controlling scope is calculating the value of the OR and AND operations with each policy IP and PORT scope. If it blocks unauthorized wireless AP unconditionally, the availability lowers so that it increase the availability by permitting part and decide the policy to maintain the security. This is an example of the control policy operation as follows. We made unauthorized wireless AP to use internal network FTP Server (192.168.1.224, 21 PORT) and internet(80 PORT) Possible IP scope goes in the first operation and PORT scope in second. Next condition will be decided in third, and if there is anymore next decision, then decide the control scope after AND operation. First IP scope and second PORT scope set the control scope after OR operation. When sending packets, it detects packet from NDIS Intermediate Driver, and it blocks if it is violated the policy by scanning IP and PORT, in other case, it permits and makes the availability high.
4
Implementation and Performance Analysis
4.1
Implementation Environment and Method
The test environment configured as table 1, except wireless AP device, we have added the smart phone which switches in AP device using WiBro tethering technology.
198
J.-K. Baek and J.-P. Park Table 1. Test Enviroment Detail
O/S
Server
Windows 2008 R2
Client Wireless AP Smart Phone Wibro
Windows 2000NT ~ Windows 7 (x86/x64) I-OS/ Android -
Client side shows wireless AP authentication requested list, and we configured the WEB(JSP) to authenticate for the requested wireless AP. Authenticated wireless AP is stored in file and DB(MS-SQL 2008), and stored information sends to AW-Client. We have used C/C++ language to configure the AP authentication process in AWClient, and ASEM/Kernel C language was used to control unauthorized wireless AP. Smart phone and WiBro device was used as connecting device which can turn wireless AP into AP device. After implemented DB and WEB server in the server, we have set unauthorized wireless AP control policy. The policy reflects if these Usage, explanation, IP scope, PORT scope, used time, used date are set. We need to set the test client PC area, but because we need to assort authorized and unauthorized, we need more than 2 of them. It the policy is established, we need to install client SW-Client. Wireless authentication registration and unauthorized wireless AP scenario are as follows.
① Install wireless LAN card more than 1 in client. ② Request wireless AP to authenticate to SW-Server. ③ Bring the possibly connecting list and signal through wireless LAN card if the
wireless AP search button in clicked. And after selecting the wireless LAN on the list, it request to the server by clicking. The requested wireless AP to the server is able to check on the WEB page as Manager click delete button if it is inappropriate wireless AP, if it is not, manager checks in check box and reflects the policy.
④
Fig. 2. Wireless AP Registration in Server
A Study on Authentication and Access Control of Wireless Access Point
199
⑤ Client tries to contact wireless AP, and unauthorized wireless AP is controlled by policy Blocked wireless AP requests to use if it is wireless AP, it leads to use different IP and PORT by changing controlling policy. 4.2
The Implementation Result and Performance Evaluation
The test environment configured as table 2, except wireless AP device, we have added the smart phone which switches in AP device using WiBro tethering technology. We have checked if it controls by connecting to authorized wireless AP and unauthorized wireless AP. As soon as the AW-Client module start, it bring the OS version and list of wireless AP. If the wireless AP is connected, it performs the authentication process after wireless network card. If the wireless AP is connected, it compares with authenticated list. And it blocks unauthorized wireless AP and inform to user by using alert window if it is not on the list. Control methods for known wireless AP; one is blocking relevant frequency using electronic jamming device in server, and the other one is placing authentication server to control in NAC level. In client side, there is control method for wireless AP by installing program to user PC. We have compared securing method in the server and controlling method by using proposed model in table 2. Table 2. Control Method and Suggestion Model in Server Detail control method scope limitatino external network normal scalability cost
frequency based frequency high
NAC based NAC high
proposed model Client control none
normal
low
high
low high
low low
high low
Frequency based range is until the disturbing frequency, and NAC based scope is in internal network range, if a user goes out with a notebook it can't be controlled. But, because of the proposed model controls in client level, it is possible to control external packets. Therefore, if it goes out externally, frequency and NAC based external networks have security vulnerability. Because it is limiting frequency and NAC based scope, scalability is very low, cost will incurred by installing the devices, but the proposed model doesn't need the cost because it controls the program itself. In table 2, Server based controlling method is very vulnerable if it is located outside, client side controlling method was chosen. We have compared client controlling method and proposed model in table 3.
200
J.-K. Baek and J.-P. Park Table 3. Control Method and Suggestion Model in Client Detail confidentiality availability scalability program security
client based high low low low
proposed model high high high high
Between client based and proposed model show high confidentiality, but it also showed many differences in availability, scalability and program security part. The client based show high security strength on unauthorized AP by controlling permit/block, it can't control every each packet. But security model can use internet, FTP while maintaining the security strength by manager setup and can monitor after saving used logs. Because of client based only blocks, it cannot expand, but proposed model can expand additional functions such as mail, messenger, WEB hard, etc Because of the proposed model only authenticates User-Level and controls in Kernel-Level, it doesn't get attacked by debugs. Proposed model doesn't have limitation compared with frequency, NAC based, it also can secure in external network, and it is cost effective. Also, the availability is higher than client based model, it is also high in scalability and it has strong security in program.
5
Conclusion
In wireless authentication, existing methods have security weaknesses and low availability in limited scope. Therefore, in this paper, we proposed client based model which can authenticate vulnerable server based external network and beyond the scope. And also, we have installed network driver to increase availability by controlling packets instead of permit/block method. Proposing models can have various polices depends on the packet algorithm, and it can have high availability by users having different policies. Wireless based services have continuously developed, whereas various attacks have increased. We should control the wireless packets and prevent various attacks and leakage from the internet information by mixing server based authentication and client authentication method.
References 1. Industrial Network Team : Windows Device Driver Programming. SamYangBook, pp. 571–578 (2000) 2. ANSI/IEEE Std 802.11 : Wireless LAN Medium Access Control(MAC) and Physical Layer(PHY) Specification (1999)
A Study on Authentication and Access Control of Wireless Access Point
201
3. Lee, J.H., Lee, M.S., Ryou, J.H.: Implementation of a Secure Wireless LAN System using AP Authentication and Dynamic Key Exchange. Korea Information Processing Society 11-C(4), 497–508 (2004) 4. Kwo, J.H., Park, J.T.: User Pre-Authentication Method for Support of Fast Mobility in IEEE 802.11 Wireless LAN. The Institute of Electronics of Korea 44(11), 191–200 (2007) 5. Kim, M.H., Lee, J.W., Choe, Y.G., Kim, S.J.: DoS-Resistance Authentication Protocol for Wireless LAN. Korea Institute of Information Security 14(5), 3–10 (2004) 6. Go, Y.M., Kwon, K.H.: Detecting and Isolating a Cloned Access Point IEEE 802.11. The Korea Contents Association 10(5), 45–51 (2010) 7. Song, C.L., Jung, B.H.: Wireless LAN Security Mechanism. Korean institute of Information Scientists and Engineers 20(4), 5–13 (2002) 8. IEEE Standard for Port Based Network Access control, IEEE Draft P802.1X/Dll (1998) 9. IEEE Standard 802.11i: Medium Access Control(MAC) Security Enhancements, Amendment 6 to IEEE Standard for Information technology - Wireless LAN Medium Access Control (MAC) and Physical Layer(PHY) specifications. IEEE (2004) 10. Im, S.C.: A Study of Hand-off Scheme Using Mutual Authentication between APs in Wireless LAN Environments. Korean Institute of Information Technology 8(9), 95–101 (2010)
A Reliable Multicast MAC Protocol Based on Spread Spectrum Technique in Wireless Ad-Hoc Networks Jeonghun Kim, Junwoo Jung, and Jaesung Lim Ajou University, Woncheondong san 5, Suwon, Korea
Abstract. We propose a reliable multicast MAC protocol based on the directsequence code-division multiple access (DS-CDMA) scheme to enhance channel efficiency in wireless ad-hoc area networks. In conventional multicast MAC protocols based on IEEE 802.11 standards using RTS-CTS handshake exchange CTS and ACK packets consecutively for reliable transmission. Although this protocol guarantees transmission reliability by avoiding the hidden node (HN) problem, multiple CTSs and ACKs degrade network efficiency. The multicast sender must receive multiple CTSs and ACKs from all multicast receivers in its own multicast group for reliable multicast. In the proposed protocol, CTS/ACK messages of multiple receivers can be transmitted concurrently through the DS-CDMA mechanism. The proposed protocol reduces the total transmission delay due to the reduction of the overhead caused by consecutive CTSs and ACKs. The proposed protocol is thoroughly examined by simulation and theoretical methods. Results show that the proposed protocol significantly reduces the aforementioned overhead, and thus improves the performance of wireless ad-hoc networks. Keywords: DS-CDMA, Multicast MAC, Reliability, Wireless ad-hoc networks.
1
Introduction
Mobile Ad hoc networks (MANETs) have recently been the topic of extensive research. The lack of infrastructure and the peer-to-peer nature of ad hoc networking make random access protocols the natural choice for medium access control in ad hoc networks. Indeed, most proposals of MAC protocols for ad hoc networks are based on the random access paradigm; in addition, the CSMA/CA scheme was selected by the IEEE802.11 committee as the basis for its standards due to the inherent flexibility of this scheme [1]. Multicasting is an efficient communication service for supporting multi-point applications in the Internet. In mobile Ad hoc network (MANET), the role of multicast services is potentially even more important due the bandwidth and energy savings that can be achieved through multicast packet delivery [2].
A Reliable Multicast MAC Protocol Based on Spread Spectrum Technique
203
Various multicast schemes have recently been introduced into wireless communication systems to improve channel efficiency. The multicast technique achieves better channel efficiency in transmitting the same packets to receivers than the unicast method. Unlike broadcast, whose transmission targets all devices within radio range, multicast targets devices belonging to a particular receiver group. Several issues, including reliable multicast MAC protocol design, challenge the design of a multicast MAC protocol. IEEE 802.11 standards specify multicast services, as well as unicast services in layer 2. However, most parts of the protocol specifications are limited to the unicast environment, and multicast has mainly been studied to focus on the relatively higher layers, such as the network and transport layers, as seen in [3-4]. Several MAC layer multicast protocols [5-8] have been introduced. However, they have some problems including MAC overhead and reliability. Until now, there was a tradeoff between guaranteeing reliability and reducing overhead in multicast MAC protocols. In this paper, we propose a new reliable multicast MAC protocol assisted by DSCDMA, to reduce the multicast MAC overhead with guaranteeing reliability. When multiple receivers receive the RTS/DATA message from a sender, each receiver sends a CTS/ACK message spread with a given DS-CDMA code. This MAC protocol can reduce the MAC overhead of multiple CTSs/ACKs via concurrent transmissions. The proposed protocol is thoroughly examined by simulation and analytical methods. The results show that it significantly reduces the aforementioned overhead, and thus improves the performance of the system. The remainder of the paper is structured as follows. Section 2 reviews previous wireless multicast research. In section 3, the proposed protocol is detailed. In section 4 and 5, we evaluate the proposed protocol with mathematical and simulation methods, respectively. Finally the conclusion is given in section 6.
2
Related Works
2.1
Multicast MAC Protocol Based on IEEE 802.11
Multicast aware MAC protocol (MMP) [6] and a multiple access collision avoidance protocol for multicast services [7] were proposed for reliable multicast. These kinds of protocol use multiple CTSs and ACKs sent by each receiver for reliable transmission, as shown in Fig. 1(a). This can be reliable. However, it increases overhead caused by the reception time of multiple CTSs and ACKs. Leader-based protocol (LBP) was proposed for multicast to reduce the overhead [8]. In this work, when a sender transmits a packet, it selects a leader station among the multicast group member stations. Then, only the leader responds with the CTS and ACK corresponding to the RTS/DATA message. If any member station receives the packet with an error, it transmits a NACK message to the sender. This NACK causes a collision with the ACK. LBP, as shown in Fig. l(b), can reduce overhead.
204
J. Kim, J. Jung, and J. Lim
However, when one of the member stations fails to demodulate the multicast RTS, the leader CTS, or the MAC header, it cannot send a NACK, since it cannot recognize if the received packet is multicast and if it is the destination of the packet. Since the station does not send a NACK, no collision is experienced at the sender. Therefore the LBP may not be reliable in terms of the detection of failed transmission.
(a)
(b) Fig. 1. Example scenarios of (a) Reliable multicast with multiple CTSs and ACKs and (b) LBP (Leader Based Protocol)
Orthogonal frequency division multiplex (OFDM) based multicast ACK (OMACK) was proposed to reduce MAC overhead and to guarantee reliability [9]. This protocol uses one OFDM symbol for ACKs from all member stations. Each member station indicates its packet reception status by utilizing a sub-carrier within the OFDM symbol. This can reduce MAC overhead. However, this can only be adopted for small-sized packet transmissions. Until now, the research on the multicast MAC protocol based on IEEE 802.11 standards has been conducted to reduce overhead or to guarantee reliability. There is a tradeoff between the overhead reduction and guaranteeing reliability. 2.2
Research on Wireless Ad-Hoc Networks Based on Spread Spectrum Technique
In a CDMA system, each user signal is spread over a large bandwidth using PN (Pseudo-random Noise) sequence, or codes with low cross-correlation, before transmitting over the channel. It can help multiple stations can transmit their own packet concurrently. However, the use of spread spectrum technique is restricted by the constraints imposed due to power control, synchronization, hardware complexity of the medium access protocol. When CDMA is used in ad-hoc networks, where multiple contending links try to transmit simultaneously, the near-far effect causes multiple access interference (MAI) and leads to collisions and packet loss.
A Reliable Multicast MAC Protocol Based on Spread Spectrum Technique
205
Most of the earlier schemes suggested for medium access in CDMA networks did not consider the effect of multiple access interference, and they assumed that two overlapped transmissions spread using two unique spreading codes do not collide with each other [10, 11, 12, 13]. Orfanos et al. [14] propose a broadcast scheme using Multi-Carrier CDMA (MC-CDMA) where available frequency channel is divided into multiple narrow-band sub-channels, and a node which has to send packets to multiple nodes in the neighborhood, can use all of the sub-channels to simultaneously transmit packets to all of such nodes. In contrast to the model assumed in the above schemes, in practice, two or more simultaneous links can concurrently communicate if and only if the SIR at the receiver of each link is above the required threshold. Due to the near-far effect and the asynchronous channel, spreading of the signal only minimizes the effect of interference but cannot completely eliminate it. Minimum separation and power control of the contending links is thus necessary for them to communicate concurrently. ElBatt and Ephremides [15] propose a centralized link scheduling policy for power control in CDMA/TDMA ad-hoc networks. The effect of multiple access interference has been considered by Muqattash and Krunz [16] in the design of CDMA-based medium access protocol for ad-hoc networks. Muqattash et al. make use of the multiple access interference suppression provided by the binary phase code modulated CDMA channel employing random spreading sequences [17], for increasing the link-packing capability of the network. [18 ] assume the use of multiuser receiver, specially minimum mean square equalizer (MMSE) receivers, by each node in the network. Multiuser receivers are known for their near-far effect resistance and asymptotic efficiency.
3
Proposed MAC Protocol
The previous section mentioned the multicast MAC problem that we want to solve. In this section, we propose a new multicast MAC protocol, to reduce the multicast MAC overhead, while guaranteeing reliability, assisted by the DS-CDMA scheme. We adapt the assumption of [ 18]. To lessen MAI effect, each station use MMSE for multi user detection. And let us assume the chip synchronization of the allocated codes coincide with each other for the successful reception of CTS/ACK messages, and each station is able to control the power of the transmission, as the processing gain varies. The mechanism of the proposed multicast MAC protocol is as follows. When multiple receivers receive a RTS/DATA message from a sender, each receiver spreads a CTS/ACK message with the allocated code based on DS-CDMA. Each code allocated to multiple receivers is given by the multicast RTS (mRTS) message from the sender. After waiting for the SIFS time, each receiver sends the CTS/ACK message concurrently. The specific feature of the proposed protocol compared to conventional protocol using multiple CTSs/ACKs is as follows. First, the proposed MAC protocol can reduce the overhead of multiple CTSs and ACKs through concurrent transmissions, as shown in Fig 2(a). However, a single CTS/ACK time spread with the given code is
206
J. Kim, J. Jung, and J. Lim
longer than the previous ones. Additionally, the number of multicast group members is restricted by the codelength. If the number of multicast group members exceeds the supportable number, the CTS/ACK is sent at least twice. If the codelength is enlarged, the CTS/ACK time is longer. We will compare this influence in later section in this paper. Also, this protocol can reduce both transmission and carrier sensing range of multiple receivers through the processing gain of the codes as shown in Fig. 2(b). This can elevate the spatial reuse to avoid packet collision between stations.
(a)
(b) Fig. 2. (a) Proposed MAC Protocol (b) Mitigating carrier sensing range assisted by DS-CDMA
In the following section, we will show the margin between the multiple consecutive CTS time and the proposed spread CTS time and find the appropriate codelength, as the number of multicast group members is varied. The contribution of the proposed multicast MAC protocol is this protocol can reduce the MAC overhead while guaranteeing reliability.
4
Performance Evaluation
4.1
Throughput Analysis
An analytical model is suggested for throughput calculation under saturation. That is, each station immediately has a packet immediately available for transmission [19].
A Reliable Multicast MAC Protocol Based on Spread Spectrum Technique
207
This work models the MAC events of successful transmission, collision and idle times. The backoff process is then modeled by a two-dimensional Markov chain. τ is probability that the station transmits a packet in a randomly chosen slot time. p is probability of a collision seen by a packet being transmitted on the channel. The following two equations are found for τ and p.
τ=
2(1 − 2 p) (1 − 2 p)(W + 1) + pW (1 − (2 p) m ) p = 1 − (1 − τ ) n −1
(1) (2)
where n is the number of contending stations, W = CWmin, Let m, “maximum backoff stage”, be the value, such that CWmax = 2mW, and let us adopt the notation Wi = 2iW, where i ∈ (0,m) is termed the “backoff stage”. Solving these equations using numerical methods, one can find τ and p based on the known values of W, m, and n. PT = 1 − (1 − τ ) n
PS =
nτ (1 − τ ) n −1 nτ (1 − τ ) n −1 = PT 1 − (1 − τ ) n
(3)
(4)
Let PT be the probability that there is at least one transmission in the considered slot time. The probability Ps that transmission occurring on the channel is successful, We now express throughput, S as the ratio S=
E[payload informatio n transmit ted in a slot time] E[ slot ]
(5)
where, E[slot] denotes the length of a slot time. As E[P] is the average packet payload size, the average amount of payload information successfully transmitted in a slot time is PTPSE[P]. E[slot] is obtained considering that, with probability 1 − PT, the slot time is empty; with probability PTPS it contains a successful transmission, and with probability PT(1 − PS) it contains a collision. Hence, (5) becomes
Here, TS is the average time the channel is sensed busy (i.e., the slot time duration) due to a successful transmission and TC is the average time the channel is sensed busy by each station during a collision. σ is the duration of an empty slot time. In conventional multiple CTS/ACK systems, like [6]-[7], we can express TS and TC as follow. TSbas = RTS + n( SIFS + δ + CTS ) + SIFS + δ + H + E[ P ] + n( SIFS + δ + ACK ) + DIFS + δ TCbas
= RTS + DIFS + δ
(7)
208
J. Kim, J. Jung, and J. Lim
Let H = PHYhdr + MAChdr be the packet header, and δ be the propagation delay. In the proposed protocol, TS is modified. TSmc = RTS + SIFS + δ + CTS mc + SIFS + δ + H + E[ P] + SIFS + δ + ACKmc + DIFS + δ
(8)
If the number of receivers (R) exceeds the codelength (LC), the CTS/ACK has to transmit more than twice. If LC < R, equation (8) has to be modified. R TSmc = RTS + ( SIFS + δ + CTS mc ) + SIFS + δ LC R + H + E[ P] + ( SIFS + δ + ACK mc ) + DIFS + δ LC 4.2
(9)
Packet Delay Analysis
We use an approach similar to [20] to derive the average delay of the successful packet transmissions. The average delay for a successfully transmitted packet is defined to be the time interval from the time the packet is at the head of its MAC queue ready to be transmitted, until an ACK for this packet is received. If a packet is dropped, as it has reached the specified retry limit, the delay time for this packet will not be included in the calculation of the average delay. The average packet delay E[D] for packets is given by
E[ D ] = E[ X ]E[ slot ]
(10)
Where, E[slot] is given by (6) and (7), E[X] is the average number of slot times required for successfully transmitting a packet and is given by W −1
Wi + 1 i bi k = ⋅ 2 k =0 i =o m
E[ X ] =
Wi + 1 Wi + 1 ⋅bi.0 = 2 2 i =o m
=
W −1
i Wi − k Wi + 1 ⋅ bi.o W 2 i i =0 k =0
m
2
Wi + 1 ⋅ bi.0 2 i =o m
2 W + 1 2 b W ⋅ 2 m + 1 m i 0.0 p p b ⋅ ⋅ + ⋅ ⋅ i 0.0 2 2 1 − p i =o m
=
=
m −1
i Wi + 1 τ ⋅ (1 − p ) p ⋅ 2 i =0
2
m Wm + 1 +τ ⋅ p 2
2
2 2 m−1 W i + 1 W +1 + p m ⋅ m = τ (1 − p ) p i ⋅ 2 2 i =0
Where, bi.k value can be referred to [19].
(11)
A Reliable Multicast MAC Protocol Based on Spread Spectrum Technique
5
209
Numerical Results
The analytical model and simulation results are implemented by a simulator to verify the performance of the proposed protocol. We use the MATLAB simulator. Table 1 summarizes the values of the parameters used to obtain the numerical results for the analytical model and the simulation runs. The values of these parameters are based on the IEEE 802.11a standard [21]. The proposed protocol is compared to the conventional protocol using multiple CTSs and ACKs, such as [5] and [6]. We only compare this basic multiple CTSs/ACKs protocol, since we are only interested in a reliable transmission, as explained in section 1. Among the protocols explained in section 2, the only protocol supporting a reliable service with RTS-CTS handshake is the MAC protocol using multiple CTSs/ACKs, such as [5] and [6].
Table 1. Simulation Parameter Values
Parameter
Value
Parameter
Value
CWmin
15μsec
MAC header
272bits
CWmax
1023
PHY header
46bits
RTS time
52μsec
Preamble
16μsec
SIFS time
16μsec
CTS/ACK time
44μsec
DIFS time
34μsec
Data rate
6Mbps
Slot time
9μsec
Packet payload
8192bits
δ
1μsec
Fig. 3 shows the variation of the throughput of the proposed MAC and the multiple CTS/ACKs MAC with the number of multicast receivers (R), where n = 50, the codelength is 8,16 and 32, the R is varied from 5 to 50. The performance of the proposed MAC is better than for the basic one. Furthermore, as the R value is higher, the gain difference between the basic protocol and the proposed one gets bigger and bigger. The throughput of both decreases as the number of multicast receivers increases. The curve of the proposed scheme is bent at a point (the point that the number of multicast receivers exceeds the codelength) of multiple numbers of codelength, because the CTS/ACK is sent more than twice.
210
J. Kim, J. Jung, and J. Lim 0.8
bas Lc=8 Lc=16 Lc=32
Throughput
0.6
0.4
0.2
10
20
30
40
50
Number of Multicast Receivers(R) Fig. 3. Number of multicast receivers and throughput 0.8
bas Lc=8 Lc=16 Lc=32
Delay(sec)
0.6
0.4
0.2
10
20
30
40
50
Number of Multicast Receivers(R)
Fig. 4. Number of multicast receivers and packet delay
Fig. 4 shows the packet delay time, calculated by equations (10) and (11), as R increases from 5 to 50. Fig. 4 shows the delay time increases with increasing R. The curves of the proposed scheme are also bent at the point of the multiple number of the codelength.
A Reliable Multicast MAC Protocol Based on Spread Spectrum Technique
6
211
Conclusion
In this paper, we proposed a new reliable multicast MAC protocol over the IEEE 802.11-based ad-hoc networks. The major innovation over previous reliable multicast protocols using consecutive CTSs/ACKs is that this can send multiple CTS/ACK messages concurrently via the spread spectrum technique. When multiple receivers receive a RTS/DATA message from a sender, they respond with a CTS/ACK message spread with the pre-assigned code in the same time interval. This can reduce the overhead of multiple CTSs/ACKs. Our proposed protocol and the multiple CTSs/ACKs based multicast protocol were compared and evaluated using an analytical method and simulation. The proposed protocol outperforms the conventional protocol based on multiple CTSs/ACKs in terms of throughput and delay. That is, our proposed protocol provides an efficient reliable multicast MAC mechanism with minimum overhead. Acknowledgments. "This research was supported by MKE (The Ministry of Knowledge Economy), Korea, under the ITRC (Information Technology Research Center) support program supervised by the NIPA (National IT Industry Promotion Agency" (NIPA-2011-C1090-0902-0003).
References 1. Chlamtac, I., Conti, M., Liu, J.: Mobile ad hoc networking:imperatives and challenges. Ad Hoc Networks J. 1(1) (January-March 2003) 2. De, M.C.C., Gossain, H., Agrawal, D.P.: Multicast over wireless mobile ad-hoc networks: present and future directions. IEEE Network (2003) 3. Obraczka.: Multicast transport protocols: A survey and taxonomy. IEEE Commun. Mag. 36(1), 94–102 (1998) 4. El-Sayed, A., Roca, V., Mathy, L.: A survey of proposals for an alternative group communication service. IEEE Netw. 17(1), 46–51 (2003) 5. Sum, M.T., Huang, L., Arora, A., Lai, T.H.: Reliable MAC layer multicast in IEEE 802.11 wireless networks. Wireless Commun. Mob. Comput. 3(4), 439–453 (2003) 6. Gossain, H., Nandiraju, N., Anand, K., Agrawal, D.P.: Supporting MAC layer multicast in IEEE 802.11 based MANET’s: Issues and solutions. In: IEEE LCN, pp. 172–179 (2004) 7. Lee, K.-H., Cho, D.-H.: A Multiple access collision avoidance protocol for multicast service in mobile ad hoc networks. In: Vehicular Technology Conference (2003) 8. Kuri, J., Kasera, S.K.: Reliable multicast in multi-access wireless LANs. Wireless Netw. 7(4), 359–369 (2001) 9. Kim, B., Kim, S.W., Ekl, R.L.: OFDMA-based reliable multicasting MAC protocol for WLANs. IEEE Transactions on Vehicular Technology 57(5), 3136–3145 (2008) 10. Bertossi, A.A., Bonuccelli, M.A.: Code assignment for hidden terminal interference avoidance in multihop packet radio networks. IEEE/ACM Transactions on Networking 3(4), 441–449 (1995) 11. Jiang, S., Hsiao, M.-T.T.: Performance evaluation of a receiver-based handshake protocol for CDMA networks. IEEE Transaction on Communications 43(6), 2127–2138 (1995)
212
J. Kim, J. Jung, and J. Lim
12. Jao-Ng, M., Lu, I.-T.: Spread spectrum medium access protocol with collision avoidance in mobile ad-hoc networks. In: IEEE INFOCOM, vol. 2, pp. 776–783 (1999) 13. Rodoplu, V., Meng, T.H.: Position based CDMA with multiuser detection (PCDMA/MUD) for wireless ad hoc networks. In: Proceedings of IEEE International Symposium on Spread Spectrum Techniques and Applications, vol. 1, pp. 336–340 (2000) 14. Orfanos, G., Habetha, J., Liu, L.: MC-CDMA based IEEE 802.11 wireless LAN. In: Proceedings of 12th IEEE International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunications Systems (MASCOTS 2004), pp. 400– 405 (2004) 15. ElBatt, T., Ephremides, A.: Joint scheduling and power control for wireless ad-hoc networks. In: Proceedings of IEEE INFOCOM, vol. 2, pp. 976–984 (2002) 16. Muqattash, A., Krunz, M.: CDMA-based MAC protocol for wireless ad hoc networks. In: Proceedings of the 4th ACM International Symposium on Mobile Ad Hoc Networking and Computing, MobiHoc 2003, June 1-3, pp. 153–164. ACM, New York (2003) 17. Sousa, E., Silvester, J.A.: Optimum transmission ranges in a direct-sequence spreadspectrum multihop packet radio network. IEEE Journol on Selected Areas in Communications 8(5), 762–771 (1990) 18. Hasan Raza Naqvi, S., Patnaik, L.M.: A medium access protocol exploiting multiuserdetection in CDMA ad-hoc networks. Wireless Netw. 16, 1723–1737 (2010) 19. Bianchi, G.: Performance analysis of the IEEE 802.11 distributed coordination function. IEEE J. Select. Areas Commun. 18, 535–547 (2000) 20. Chatzimisios, P., Boucouvalas, A.C., Vitsas, V.: IEEE 802.11 packet delay – a finite retry limit analysis. In: IEEE Globecom, San Francisco, USA (2003) 21. Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications: Higher-Speed Physical Layer in the 5 GHz Band, IEEE Std. 802.11a (1999)
Relationship between the Binding Energy and Boundary Condition in SiOC Film for ILD Application Teresa Oh1,*, Chy Hyung Kim2, and Chi Sup Jung3 1
Semiconductor Engineering Department of Applied Chemistry 3 Department of Laser and Optical Information Engineering, Cheongju University 36 Naeduckdong Sangdangku, Cheongju 360-764, Korea [email protected] 2
Abstract. Low dielectric constant SiOC film was deposited by the chemical vapor deposition. To research the change of polarization by the chemical reaction between opposite polar sites, the synthesized SiOC films at various flow rates of BTMSM were studied by Fourier Transform Infrared Spectroscopy, X-ray diffraction patterns, X-ray photoelectron spectroscopy, and Scanning Electron Microscope. By using these analytical methods, the correlation between the chemical shift and boundary condition on surfaces of SiOC film was observed. The binding energies of Si 2p, C 1s, and O 1s in SiOC film with the low polarization were higher than other films, and the 2θ value in X-ray diffraction pattern was also moved to the higher degree in the film of the lowest polarity. The chemical shift was caused by the polarization due to the lowering of the ionic polarization by the chemical reaction between alkyl and hydroxyl group as polar sites. The Scanning Electron Microscope of SiOC film showed an image of flat surface without pin-holes due to the low surface energy, which improves the properties of SiOC film to become an ideal low-k material. On the other hand, SiOC film with low binding energies of Si 2p, C 1s, and O 1s displayed the pin-holes in SEM images owing to the existence of polar sites. The grain boundary such as pin-holes indicates that an unbalance exists in the uniformity of electron distribution on a surface and a low degree of amorphism. Keywords: SiOC film, FTIR spectra, SEM, XRD, XPS spectra.
1
Introduction
To get high-speed integrated circuits, insulating materials with low-k (low dielectric constant) is required, which can be replaced instead of SiO2 film because of a lot of problems such as signal noise, cross talk, and power consumption in small size devices. SiOC film as a low-k material is made by the SOD (spin on coating) or CVD (chemical vapor deposition). SiOC film as a promising low-k material possesses a lower dielectric constant due to the presence of lighter C and H atoms, as opposed to Si and O atoms [1-4]. The reduction of dielectric constant in the SiOC film by the *
CVD originates from the decrease of the polarization in the film. Low polarization is essential to low-k materials, which become the amorphous structure to a great degree because of the reduction of the electron density that resulted from the C-H bond elongation effect and the carbon insertion in Si-O network. Therefore, SiOC film must satisfy a number of requirements including a low dielectric constant of the film to be successfully integrated. The dominant fraction of the C-H bond length behavior comes from the straightforward electrostatic polarization of the proton donor. Regarding to the chemical shift of hydrogen bonds, electrostatic interactions illustrates a major portion of the shift in FTIR spectrum, with polarizabilities increasing the shift further. The C-H bond should also be weakened upon the hydrogen bond formation and elongated. The extent of the chemical shift has been correlated with the strength of the hydrogen bond, the proton donor-acceptor distance, and the ionization potential of the proton acceptor [5-7]. SiOC film is usually prepared by using the CVD such as ICP (inductive coupled plasma)-CVD, CCP (capacitive coupled plasma)-CVD, and TCP (transformer coupled plasma)-CVD. Each pattern of FTIR spectrum has some differences in accordance with the plasma energy during the deposition. However, in order to understand the C-H bonding effect, the results obtained from the FTIR spectra of SiOC films in previous studies are not enough to understand the bonding structure. Moreover, for the analysis of SiOC film more than two kinds of analyzers are needed [8-10]. The CVD-SiOC film was analyzed by the Fourier transform infrared spectroscopy, X-Ray Photoelectron Spectroscopy, Scanning Electron Microscope, and X-ray diffraction. The results were utilized focusing on the relationship between boundary conditions and physical-chemical properties based on the incorporation of carbon compounds at the SiOC film formation. To research the surface properties of SiOC film, the pentacene was deposited on the SiOC film, because the trend of pentacene growth provided us the information such as electron distribution in SiOC films depending on the roughness of surface or weak boundary condition.
2
Experiments
The carbon doped silicon oxide films were obtained by the capacitive coupled plasma chemical vapor deposition (CCP-CVD) using the mixed gases of oxygen and bistrimethylsilylmethane (BTMSM). The precursor of bistrimethylsilylmethane was purchased from the Applied Materials Corporation. The deposition on substrate was carried out at 100℃ for 10 sec. The BTMSM was vaporized and transferred by argon gas at 35 ℃ through a thermostatic bubbler. Although the carbon doped silicon oxide films were prepared at various flow rate ratios of BTMSM precursors, the oxygen gas flow rate was kept at 60 sccm. The base pressure was 3 Torr and the rf power was 450 W in each experiment. To study on the surface properties of SiOC film, the pentacene was deposited on SiOC film by thermal evaporation at the base pressure of 1 Torr approximately, due to the dependence of the substrate properties on the growth patterns of pentacene molecule. The substrates were held at 80 °C and the pentacene was deposited at 0.1~0.3 nm/s, with a final average thickness of 40 nm. The chemical shift
Relationship between the Binding Energy and Boundary Condition in SiOC Film
215
was researched by Fourier Transform Infrared spectrometer (FTIR, Galaxy 7020A) and X-Ray photoelectron spectroscopy (XPS, ESCALAB 210). The surfaces of films were measured by using the Scanning Electron Microscope (HITACHI S-4200) and the crystalline structure was studied by utilizing X-ray diffractometer (D/MAX2200H, Rigaku).
3
Results and Discussion
Figure 1(a) shows the FTIR spectra of SiOC film at various flow rate ratios. The BTMSM flow rates are 24~28 sccm respectively, and the oxygen gas flow rate is 60 sccm. Sample number indicates the BTMSM gas flow rate of 24~28. In the range of 1400~1800 cm-1, the figure 1(b) is the normalized FTIR spectra. The intensity of Si-CH3 bond near 1250 cm-1 decreased at the sample 25. The reduction of intensity of Si-CH3 bond is due to the dissociation of the bond by plasma energy and then the nucleophilic reaction between alkyl group and hydroxyl group because dissociated sites induce to decrease the polarization and the chemical shift in FTIR spectra. The main peak near 1110 cm-1 observed in FTIR spectra originates from the dissociation of Si-CH3 groups broken by the oxygen of neighboring high electronegative atom. Then the Si-O bond or O-CHn(n=1, 2, 3) in SiOC film is formed. The O-CH bond changes to the Si-O-CH bond because the Si atom attacks O-CH bond in accordance with the increase of BTMSM flow rate. Thus, the weak Si-O-C bond displays the chemical shift in FTIR spectra. The main bond of 970~1190 cm-1 of samples 26~28 is the Si-O-C broad bond due to the recombination of the dissociated radicals. The chemical shift observed between 25 and 26 samples, which showed the red shift owing to the increase of the intensity of left shoulder in the main bond of Si-O-C. The chemical shift of SiOC film provides the information about the polarity of the final materials. Non-polarization is one of the conditions to make an amorphous structure as a low-k material. In terms of the wavelength, the red shift results from the decrease of the polarization and low electron density due to the C-H bond elongation, which results in the weak boundary condition and changes the surface energy as described in Figure 2. (b)
Absorbance (arb.units)
Absorbance (arb.units)
(a) as deposited films O2 = 60 sccm
BTMSM(sccm)= 24 25 26 27
24 25
26 27
28
28 750
1000
1250
1500 -1
Wave number (cm )
1750
700
800
900
1000
1100
1200
1300
-1
Wave number (cm )
Fig. 1. FTIR spectra of SiOC films at various flow rate ratios, (a) samples of 24~28 in the range of 650~2000 cm-1, (b) normalized FTIR spectra
216
T. Oh, C.H. Kim, and C.S. Jung
After the pentacene deposition, the SEM shows the images of many grains on SiOC film at various flow rate ratios. The pentacene is deposited with gradient growth depending on the surface’s properties of the substrate. The gradient growth strongly influences the grain formation on the surface of SiOC film in accordance with the polarity [8]. The samples 24 and 25, which have the characteristic of polarization produced by hydroxyl, show the SEM images of a number of grains involving the pinholes after the pentacene deposition on SiOC film. The pin-holes induce the electronic localization inside the film causing a higher surface energy and providing less flat surface. The samples 27 and 28 display images of the small grains. These grain boundaries and pin-holes prevent the carrier transfer for the devices and reduce the mobility of the carrier [10]. However, sample 26 observed first in the chemical shift displays a relatively weak boundary and large grain due to the coalescence of small grains, so the surface becomes flatter than other samples. This weak boundary condition is strongly related to the low polarization. So, the non-polar property of the SiOC film results from a nucleophilic reaction and C-H bond elongation due to the electron deficient group, improving the surface property by lowering the density of the grain boundary.
Fig. 2. SEM images after pentacene deposition on SiOC film in accordance with the increase of the precursor, BTMSM
Figure 3 indicates the Si 2p, C 1s, and O 1s spectra obtained by using X-Ray photoelectron spectroscopy (XPS). Of the samples 24~28, sample 26 in which the chemical shift occurs first, increases the binding energies of the Si 2p, C 1s and O 1s the most, owing to the reduction of polarization that causes the fine cross link structure and flat surface. So, the bonding structure of sample 26 is stronger than that of other samples. Therefore, the sample 26 demonstrating the chemical shift has a low surface energy due to the decreased polarity and weak boundary condition, which improve the hardness of the film surface. Figure 4 is the X-ray diffraction of SiOC film at various flow rate ratios of BTMSM. X-ray diffraction pattern near 54.4 degree of 2 theta (2θ) changed with increasing the flow rate of precursor, BTMSM.
Relationship between the Binding Energy and Boundary Condition in SiOC Film
217
104.8 286.8
(a) Si 2p
535.0
(b) C 1s
(c) O 1s 534.8
104.4
104.2
104.0
286.6
Binding Energy (eV)
Binding energy (eV)
Binding Energy (eV)
104.6
286.4
286.2
534.6 534.4 534.2 534.0 533.8
103.8 286.0 22
22
24
26
28
24
26
533.6
28
22
BTMSM flow rate (sccm)
BTMSM flow rate (sccm)
24
26
28
BTMSM flow rate (sccm)
Absorbance (arb.units)
Fig. 3. X-Ray photoelectron spectroscopy of SiOC film at various flow rate ratios, (a) Si 2p electron orbital spectra, (b) C 1s electron orbital spectra, (c) O 1s electron orbital spectra
23 26 27 29 34 51
52
53
54
55
56
57
58
o
2 Theta ( )
Fig. 4. X-ray diffraction of SiOC films at various flow rate ratios of BTMSM
The XRD pattern of sample 26 shows the highest peak intensity near 54.4 degree, and most of the samples display the XRD peaks below 54.4 degree. The bonding structure between alkyl and hydroxyl group fabricated by the chemical reaction becomes finer cross link structure like amorphous structure of non-polarization and induces the reduction of the dielectric constant of SiOC film. Therefore, there is a close relationship between the low polarization and dielectric constant in low-k materials.
4
Conclusion
The red chemical shift in FTIR and the increase of the binding energy of C 1s orbital in XPS occurred in accordance with changing the flow rate ratios of precursor in the CVD-SiOC film. The growth of pentacene on surface of SiOC film depends on the properties of polarity of SiOC film. The terminal bond of Si-CH3 in the precursor (BTMSM) was dissociated by the plasma energy. When the terminal bond was not dissociated and deposited on the substrate surface, the polarity of SiOC film increased. Thus, the SEM of pentacene/SiOC film with polar sites showed the images of many pin-holes caused by the electron rich group. As lowering the polarity from sample 25 to sample 26, the grain size decreased and then the coalescence of the grains were induced, so the final film became the flat surface because of the weak boundary
218
T. Oh, C.H. Kim, and C.S. Jung
condition and improvement of the uniformity, which is one of the important factors to fabricate the small sized-semiconductor device. The binding energy of C 1s orbital in XPS also increased in SiOC film as the polarization was lowered. Low polarity is an important key in synthesizing SiOC film, which induces the flat surface owing to the low surface energy and weak boundary condition, thus forming a strong bonding structure in the film.
References 1. Navamathavan, R., Choi, C.K.: Journal of the Korean Physical Society 48, 1675–1679 (2006) 2. Cho, S.J., Bae, S., Boo, J.H.: Thin Solid Films 518, 6417–6421 (2010) 3. Oh, T., Choi, C.K.: Journal of the Korean Physical Society 56, 1150–1155 (2010) 4. Grill: Diamond and Related Materials 10, 234–239 (2001) 5. Oh, T., Kim, C.H.: IEEE Trans. Plasma Science 38, 1598–1602 (2010) 6. Damayanti, M., Widodo, J., Sritharan, T., Mhaisalkar, S.G., Lu, W., Gan, Z.H., Zeng, K.Y., Hsia, L.C.: Materials Science and Engineering B 121, 193–198 (2005) 7. Yu, L.D., Lei, S., Dong, Z.S., Yi, W., Yan, L.X., Qi, H.R.: Chin. Phys. Soc. 16, 240–245 (2007) 8. Huang, W.J., Li, B.Q., Zuo, J.M.: Surface Science 12, 157–160 (2005) 9. Li, H.M., Zhang, G., Yoo, W.J.: Thin Solid Films 518, 6382–6384 (2010) 10. Oh, T.: IEEE transactions on Nanotechnology 5, 23–29 (2006)
A Framework for Active Service Pattern Mining* Jeong Hee Hwang1 and Mi Sug Gu2 1
Department of Computer Science, Namseoul University Database/BioInformatics Laboratory, Chungbuk National University [email protected], [email protected] 2
Abstract. It is important to consider both location and time information which is related to all object and user activity to supply suitable services to users in ubiquitous computing environments. In this paper, we design a spatial-temporal ontology considering user context and propose system architecture for active mining user activity and service pattern. The proposed system is a framework for active mining user activity and service pattern by considering the relation between user context and object based on trigger system. Keywords: Ontology, Context-awareness, Spatio-temporal Ontology, Active Mining.
1
Introduction
In the ubiquitous environment, each computer which has the functions such as information collection, process, and communication, is connected functionally and spatially each other, so it needs the method to provide the necessary information or services to the users immediately[2], [3]. In the previous works, there have been context ontology generation which considers the user context[1], [4], and also researches about the data mining technologies using the ontology[7], [8]. In this paper, to provide the behavior patterns and the optimum services according to the user context, we construct the ontology about the spatio-temporal information which is related with the user behavior. And also we propose active mining framework based on spatiotemporal ontology which uses trigger function of the event based intelligent movement performance. This framework is the architecture which provides the most suitable services according to the user context and can automatically find out the user behavior and the service pattern which is provided to the users, so it can be a basic research to provide the intelligent services. The proposed framework has a characteristic that includes ECA(Event-Condition-Action) rule based trigger system responding automatically to the events according to the user context which is collected from sensor devices. Ubiquitous computing is the basic technology which connects with all the objects including people functionally and spatially and then provides the necessary information or services immediately [1], [2]. *
Funding for this paper was provided by Namseoul university.
In Ubiquitous computing there are software architectures for the context recognition service such as ECA pattern[9], multi agent architecture, blackboard pattern and so on. ECA pattern of them has some advantages for the event based process. Therefore, our mining engine combined with trigger system is based on the context information and performs the active mining of the user behavior and service pattern by the trigger event. Accordingly, this paper is not only basic research for providing the intelligent information service needed in Ubiquitous environment but also can be useful for basic information which can provide the optimum service according to the user behavior pattern. This paper comprises as follows. In chapter 2, we explain the context ontology which considers the user activities and context proposed in this paper. And in chapter 3, we propose the active service pattern mining framework and show the example of active rule. And in chapter 4, we explain the mining algorithm and the experimental result. Finally, in chapter 5 conclude this paper.
2
Context Ontology Design
Recognizing the context information is modeling the environment and the objects that are around the users and providing the service based on this. Time exists with space together. The action of the people is related to the time and space[10], [11]. Spatio-temporal movement pattern is the location pattern of the moving objects, and also it is the spatio-temporal rules that are personalized according to the user location and can provide the suitable contents and service[6], [10]. Therefore, through the ontology which considers the time and space together, it is needed to provide the service to the user behavior. The temporal topology relation which analyzes the temporal relation includes the similar definition to the topology relation of the space, Allen[11] defines thirteen temporal relation operators such as before, after, starts, started_by, finishes, finished_by, during, contains, meets, met_by, equals, overlaps, overlapped. In this paper, we used eight major basic temporal relation operators. Spatio-temporal ontology considers temporal phase relation, spatial phase relation [10] and also the relationships between the user behavior and the objects about the spatio-temporal phase relation. Fig. 1 shows the whole domain of the context ontology to provide the service recognizing the context information of the user in home. Sptio-temporal ontology is classified into SPACE ontology and TIME ontology. and the spatio-temporal information of the user is generalized through relation ontology which connects the relationships between time and space. To express the spatiotemporal based user information we define the spatio-temporal information, the user domain and the bounds as follows. Definition 1. The sptio-temporal based context information of the user, U includes three phases such as object, spatio-temporal information, behavior, and also contains the identifier for recognizing of the only object. It is represented as U={OID, (S, T), A}. Here, OID is the object identifier used by the user, S is the space, T is time, and A is the user activity and context.
A Framework for Active Service Pattern Mining
221
The spatio-temporal information which identifies the user space and time comprises the object identifier, OID for the user behavior and the location information according to the object movement. Location information is composed of a generalized value about spatial coordinate and the valid time which the real event occurs. It is represented as {(S1, T1), (S2, T2), ...,and (Sn, Tn)}. Here, (Sn, Tn) means nth location and time information about the user and the object movement. Sn is a generalized location information about the spatial coordinate (xi, yi).
Fig. 1. Context Ontology
The generalization of the location identifies the location of the user and the objects, generalizing the coordinate of the space as a certain zone. For example, if the user U is watching the television in the living room, the coordinate of the user is generalized into the zone value of the living room area. The general domain of the time is divided into two such as the effective time which means that the object exists in the real world, and the transaction time which is recorded in the system. That is, within the domain of each phase there is a scope. The time domain has timestamp, interval, and period of the temporal object on the line of time. For example, there are the timestamp which the user comes into home, the time interval which it takes for the user to have meal, and the periodic occurrence of the event. And the user context information contains the user activity. The detail information of user activity is the behavior attribute that represents speed, direction, and status. For example, if there are two switches in a row, the habitual action that turns the light on the left is stored as detail information through the attribute information and then
222
J.H. Hwang and M.S. Gu
the user can be provided the suitable services. To represent the context ontology, we used OWL(Web Ontology Language)[12] suggested by W3C. And Fig. 2 shows the generation of the ontology using protege[13].
Fig. 2. Ontology generation using protege
3
Service Pattern Mining Framework
Fig. 3 shows the framework of active service pattern mining, which intelligently provides the service according to the user environment and context. The structure of the system comprises the application layer, the service provider layer, and the physical layer. The application layer provides the users with the optimum services according to the query from the users. And the service provider layer comprises the context ontology composed of the basic context information, the data mining engine which finds out the user behavior patterns and the service rules, and the database which generates new service and stores the existing service. The physical layer is the lowest layer that collects the context information through the sensor network. And ECA trigger engine of the proposed system is the system that combines the ontology engine including the spatio-temporal ontology with the data mining engine to perform data mining and provide the service intelligently. Through the events such as the user behavior, time, location recognition and so on, the condition of the previously given rules is evaluated, and the exact service to the user is searched and
A Framework for Active Service Pattern Mining
223
provided automatically. Also the trigger system searches the service and performs the data mining according to the user behavior event. Data mining engine finds not only the user behavior but also the related service based on the spatio-temporal ontology. Data mining generates the repetitive pattern rules about the user behavior and the provided service in given interval period and the special context. Here, the given interval period means the time interval period over the collected data amount or the window size. Data mining is defined by the window size based on the number of transaction and then is performed. And performing the data mining using the ontology can find the more concrete rules.
Fig. 3. Service pattern mining architecture
The ontology engine stores the previous service history provided to the user and also a new rule from the data mining results is added and stored consistently. The ontology provides the spatio-temporal ontology information for performing the data mining and is the base for extracting the suitable service to the user. That is, the ontology database stores not only the previously provided service rules but also a new rule generated by the data mining. And if a new is searched, the priority is given to the new rule. so the newly searched rules is performed over others. The trigger rule performs the data mining if the time is fixed or a given number of transaction is collected to perform the data mining periodically by the trigger event. An example of the trigger rule(INVOKE_MINING_TRIGGER) is defined in Fig. 4. To provide the service automatically if the user behavior event happens, the related service from the service set is searched and stored into the service providing table. The other in Fig. 4 is the trigger rule(SEARCH_SERV_TRIGGER) for storing into the service table. Trigger monitors the operation(insert, delete, update) about the target table and defines the suitable action. If the previously defined event(data operation)
224
J.H. Hwang and M.S. Gu
on target table is searched, and then performs the action, checking the condition defined in the rule. In this paper, using INSERT trigger the condition and the action of the trigger rule is performed. And the type of a newly inserted service is same to the existing service, the priority is given to the new service rule and then it is inserted. If the inserted new rule recognizes the same context event of the user related to the service, the service that has the top priority is provided to the user. // INVOKE_MINING_TRIGGER CREATE OR REPLACE TRIGGER INVOKE_MINING_TRIGGER AFTER INSERT ON TRANSACTION_TABLE FOR EACH ROW BEGIN IF DATE_TIME = RESERVED_TIME OR TRAN_SEQ_NUM ≥ SET_NUM THEN ALERT MINING_MANAGER END IF; END; // SEARCH_SERV_TRIGGER CREATE OR REPLACE TRIGGER SEARCH_SERV_TRIGGER AFTER INSERT ON EVENT_TABLE DECLARE w_service_id INT; w_user_id INT; BEGIN IF ACT_CONTEXT(OID, Space, Time, Activity) EXIST IN Rsrv_Table THEN w_service_id:= Rsrv_Service_id; w_user_id:= Rsrv_User_id; INSERT INTO SERV_TABLE VALUES (w_user_id, w_service_id); ENDIF; END; Fig. 4. Trigger rule example
4
Algorithm and Experiment
Data mining algorithm which searches the user activity pattern is as follows. We use the sequential pattern algorithm[14] which searches the relationship over time. And we can predict the following action by the searched rules Algorithm ActPatterns Input : Transaction Database(TD)
A Framework for Active Service Pattern Mining
225
Output : Activity Sequential Patterns Initialize Activity Patterns For each transaction t TD Activity Patterns = PrefixSpan(α,β,S|α) //extract activity patterns using PrefixSpan algorithms If Activity Patterns not exists in RuleBase then add Activity Patterns to ActivityRuleBase End if End for
∈
Data mining in designing the active mining framework proposed in this paper is important for extracting the basic data to provide the user the optimum service. Accordingly, in the experiment, we will show the efficiency of the proposed system through the number of the service searched by the data mining. For the experiment we used Java and Oracle9i in Window XP with Pentium 4 CPU 2GHz and 1GB RAM. And we assumed that the service rule (Si) provided according to the user behavior (Ai) is stored in the ontology rule base. Table. 1 shows a part of the rules. Table 1. Service rule Activity
Service Rule
A1
S111, S103
A2
S110, S125, S101
A3
S115, S106, S118
A4
S107, S102
A5
S104, S126, S109
To find out the service relation rules, the service history information is regarded as a transaction, and we let each transaction contain 2 to 3 of basic service information randomly to 10000 transactions. and then using Apriori algorithm the related service rules are searched. Fig. 5 is the result compared the number of the service rules (ServRule) which
can be stored and provided in the ontology with the number of the service rules (ServMineRule) which can be provided by data mining. The experimental result shows that the number of the service rules by data mining can provide on average 2.5 times more than the number of the service rules which are stored in the ontology. It means that in addition to the service rules previously stored in the ontology a new relation service rules can be found based on the service history. And it is possible to provide and recommend better service to the user, adding the service combination or the mined service. In the following experiment, to find out a variation on the number of the service, we inserted the transaction of 1000 units sequentially and then measured the change of the number of the searched service. In Fig. 6 it is proven that the more the number of the transaction is over time, the more the number of the searched service is. That is, if the number of the transaction is increasing, the number of the related service is also so increasing that the number of the service relation rules with a given threshold is increasing.
226
J.H. Hwang and M.S. Gu
Fig. 5. The number of services
5
Fig. 6. Variation of the number of services
Conclusion
In this paper, to provide the automatic service we proposed a framework of the active service pattern mining which includes ECA trigger system. It is based on the ontology which includes the spatio-temporal information, and mines the service pattern rules relating to user activity. And we proved the efficiency of finding out the service pattern rules through the experiment. In the future, we will evaluate the quantity about the efficiency of the service system using the trigger.
References 1. Harry, C., Tim, F.: An Ontology for Context-aware Pervasive Computing Environments. In: Workshop Ontologies and Distributed Systems. IJCAI Press (2003) 2. Khedr, M., Karmouch, A.: Negotiating Context Information in Context-aware Systems. IEEE Intelligent Systems (2004) 3. Strimpakou, M., et al.: Context Modeling and Management in Ambient-Aware Pervasive Environments. In: Workshop on Location and Context-aware (2005) 4. Strimpakou, M.A., Roussaki, L.G., Anagnostou, M.E.: A Context Ontology for Pervasive Prevision. National Technical University of Athens (2004) 5. Lee, C.H., Helal, S.: Context Attributes:An Approach to Enable Context-Awareness for Service Discovery. In: Symposium on Applications and the Internet, pp. 22–30 (2003) 6. Maffioletti, S., Mostefaoui, S.K., Hirsbrunner, B.: Automatic Resource and Service Management for Ubiquitous Computing Environments. In: The Second IEEE Annual Conference on Pervasive Computing and Communications Workshops (2004) 7. Brisson, L., Collard, M.: An Ontology Driven Data Mining Process. In: The Tenth International Conference on Enterprise Information Systems (2008) 8. Bellandi, A., Furletti, B., Grossi, V.,, R.: Ontology-driven Association Rules Extraction: a Case of Study. In: The International Workshop on Contexts and Osntologies: Representation and Reasoning (2007)
A Framework for Active Service Pattern Mining
227
9. Beer, W., et al.: Modeling Context-Aware Behavior by Interpreted ECA Rules. In: Kosch, H., Böszörményi, L., Hellwagner, H., et al. (eds.) Euro-Par 2003. LNCS, vol. 2790, pp. 1064–1073. Springer, Heidelberg (2003) 10. Abraham, T.: Knowledge Discovery in Spatio-Temporal Databases. School of Computer and Information Science, University of South of Australia, Ph. D dissertation (1999) 11. Allen, J. F., Kautz, H. A.:A Model of Native Temporal Reasoning. In: Formal Theories of The Commonsense World (1985) 12. http://www.w3.org/2004/OWL 13. http://protege.stanford.edu 14. Pei, J., Han, J., et al.: PrefixSpan: Mining Sequential Patterns by Prefix-Projected Growth. In: The International Conference on Data Engineering (2001) 15. Agrawal, R., Srikant, R.: Fast Algorithms for Mining Association Rules. In: The 20th International Conference on Very Large Data Bases (1994)
Character Region Detection Using Structure of Hangul Vowel Graphemes from Mobile Image Jong-Cheon Park1,*, Byoung-Min Jun1, and Myoung-Kwan Oh2 1
Abstract. Recently, many researchers have been proposed utilizing mobile images, whereupon the various smartphone applications have been developed on the basis of images. These applications have analyzed the images and mined the information, so they can let people search without typing keywords on the web. However, most of conventional methods for character region detection are based on clustering technique without using structural feature of character. Therefore, if a character in mobile image has complex backgrounds, these methods are difficult to localize the character region. We proposes the method to detect the Hangul character region from mobile image using structure of Hangul vowel graphemes and Hangul character type decision algorithm. First, we transform a mobile image to a gray-scale image. Second, feature extraction performed with edge and connected component based method, Edge-based method use a Canny-edge detector and connected component based method applied the local range filtering. Next, if features are not corresponding to the heuristic rule of Hangul character, extracted features filtered out and select candidates of character region. Next, candidates of Hangul character region are merged into one Hangul character using Hangul character merging algorithm. Finally, we detect the final character region by Hangul character type decision algorithm. Experimental result, proposed method could detect a character region effectively in images that contains a complex background and various environments. As a result of the performance evaluation, the recall rate is 82.33% and the proposed method showed the advanced results about detection of Hangul character region in mobile image. Keywords: Character Region, Detection, Hangul Vowel Grapheme Structure.
1
Introduction
Mobile devices have been spreading like a digital camera, a cell phone, a smartphone, PDA, and so on and users have made good use of the images recorded with a camera built in these devices. The image is combined with network technology and computer vision technique and used as major data of diverse application. A letter of the natural *
Character Region Detection Using Structure of Hangul Vowel Graphemes
229
image needs to be recognized and detected the information of the region. So, it is an important matter to detect the region in it precisely. These research fields are helping system for visually impaired person [1], indoor mobile robot navigation [2], a Robust License-Plate Extraction [3], video or image coding/ index on contents [4], etc.. A letter of the natural image can be damaged, overlapped each other, made up with any other color, distinguished from background color because it is mixed with structure, trade mark, white noise. So, it sometimes has a low contrast with these differences. And as it has the feature changed the color of letter due to various font, size and slope of a letter and lighting and shadow, detection method should be considered of various feature of character and character region [5]. The old research detecting the character region from natural images couldn’t make full use of structural feature of Hangul but had usually used the group feature of general letters. As a result, the existing method had a difficulty in detecting Hangul in the intricate background effectively. So, this article suggests the method to detect the character region by using the structural feature in order to solve this problem. New method detects the character region by checking whether that is a Hangul letter or not with algorithm to merge character region by using the location of vowel, the structural feature and the Hangul geometric relation consisting of initial consonant, vowel and final consonant and to detect the type of the letter. Overall procedure on this article is shown blow Fig. 1.
Input image G Image transform & preprocessing
Detection based on the edge
Detection based on connecting elements
Detection of the edge & labeling Filtering feature value of edge Distinction and detection of candidates region
Detection of connecting elements & labeling Filtering feature value of connecting elements Distinction and detection of candidates
Merging candidate regions
Detection of the character region
Fig. 1. Character Region Detection Procedure
230
2
J.-C. Park, B.-M. Jun, and M.-K. Oh
Detection of Hangul Character Region
The first step for the detection of Hangul region is that we convert RGB 24 bit color image into gray-scale image with brightness value 256, do with median-filter to remove the noise in the pre-process and use the end-in searching method to improve the image contrast. In the second step, the procedure extracting character was conducted with edge and connecting elements. Canny-edge detector was used in extracting edge character [6], it applies the threshold value to the outcome after conducting the range filter finding the remainder between minimum and maximum of the 8 directions neighbor pixel value adjacent to pixel in order to extract the character of connecting element. Extracted value is detected by labeling edge and connected component as objects. In the third step, filtering is conducted to remove anything not to correspond with the character region of labeling the detected edge and connected component. In [7], we comment on the method to filter feature value by using an edge and the edge and connected component based method. This is to make the character value selected as a filtering procedure using the algorithm that merges each element of Hangul into the one character region. Merged character region detects the final character region by testing whether the character is with algorithm that tells Hangul character type.
Fig. 2. Hangul Character Structure. C1: First Consonant, C2: Final Consonant, V1: Vertical Vowel, V2: Horizontal Vowel
2.1
Fig. 3. Hangul Character Font Face
Hangul Grapheme Merging
Old researches related with merging each element of Hangul have dealt with the method to isolate Hangul, written English, numbers, and special characters on documentary image mainly [8]. We suggest two algorithms merging character to detect character region with color and leaning of the character from natural images after detecting connecting elements with these features [9]. Hangul consists of first consonant, vowel and last consonant and can have sense to merge every element into one character. Accordingly, extracted value is merged into one character region using an algorithm of merging Hangul graphemes. Hangul grapheme merging algorithm is made by merging Hangul graphemes repeatedly from top-down and left to right fashion. Table 1 shows the location of vertical vowel, the
Character Region Detection Using Structure of Hangul Vowel Graphemes
231
number of Runs and the position of stroke as a threshold standard. Hangul letter is classified into 6-types as depicted in Fig. 2. We can correctly merge one Hangul letter among neighboring letters with the feature of vertical vowels considering that a letter always starts with a consonant. Hangul grapheme merging algorithm uses the geometric structural feature, distance value between features and slope of the detected candidate feature values. But every Hangul character is not detected as Hangul character region merged with the element merging algorithm. Fig. 3 shows the various Hangul font, ‘ ’, ’ ’, ’ ’ in the second line have already combined in one letter with first consonant, vowel, last consonant. Characters with like this structure cannot detect with this algorithm. Instead, they should be detected using Hangul character type decision algorithm, the next grade.
당 구 장
Table 1. Features of Hangul Vowel Vertical Vowel
ㅣ ㅏ ㅑ ㅓ ㅕ ㅐ ,ㅒ 2.2
The Number of Run Horizontal Vertical 1 1
Vowel Position Vertical left
Stroke Position Horizontal null
left
center up center /below
1
1
1
2
left
1
1
right
2
1
right
[1, 2]
2
left, right
center up center /below center, up center /below
Hangul Character Type Decision
The detected candidate character region with Hangul grapheme merging algorithm can be detected using the 6-type-character decision algorithm. In table 2, the features are presented according to the Hangul character 6-types. It is showing the feature of position and length of Run in the horizontal and vertical direction. The number of Runs means the number of black pixel equated to character in binary image. That is, it means the number of pixel that consists of the character region. The number of Runs marked with [minimum value, maximum value]. As Hangul character has structurally horizontal or vertical vowel surely, we can distinguish whether the character is in Hangul character region by extracting the feature value of horizontal or vertical vowel according to each Hangul character type and confirming that position. Too small region should be excluded from the character region by setting minimal size of the character region in order to remove but this noise not the character region detected generally. On this article, a minimal size of width and height is set up 16 pixels. The character region is removed to consider of noise region unless it meets the condition that the rate of width versus height is bigger than 0.4 and smaller than 2.5.
232
J.-C. Park, B.-M. Jun, and M.-K. Oh
If the character region is detected with the edge-based method, more than 2 edges in the character region should be consisted of less than 50% of width and height. We could divide the noise region from the character region by setting threshold value for the standard deviation calculating the standard deviation of the character region as the character region has big differences in the brightness value and the background. As an experimental result, the standard deviation is set up ’25’ as a threshold value. Firstly, the noise should be removed as a feature condition of the character region and then it is detected using Hangul character type decision algorithm. Table 2. Features of Hangul Character 6-Types 6-Types
3
Number of Run Horizontal Vertical
Vowel Position Vertical Horizontal
I
[2,3]
[2,4]
right
null
II
[2, 4]
[2, 4]
null
middle below
III
[2, 3]
[2, 5]
up right
null
IV
[3, 3]
[2, 6]
null
middle
V
[2, 3]
[2, 4]
right
middle
VI
[2, 4]
[4, 5]
up right
middle
Experimental Result
We presented the method using the feature of Hangul vowel grapheme structure and Hangul character type decision algorithm in natural image captured by mobile phone camera. Experimental image data base for detection of proposed character region was used from KAIST artificial intelligence laboratory site [10]. It was taken with mobile camera and made of images taken in various backgrounds like indoor, outdoor, lighting and shaded image. We evaluated detection performance with precision rate in formula 1 and recall rate formula 2 [11]. They are revised with pixel to apply formula 1 and 2 to image- based. Precision rate is refined to formula 3 recall rates to formula 4. That is, it means the evaluation is conducted for the square detected as character region. p r p
∑
E
|E|
E T ,T
(1) (2) (3)
Character Region Detection Using Structure of Hangul Vowel Graphemes
r
∑
T
|T|
,T
233
(4)
The c is the number of the character region(c) in formula 1and 2 and it is refined into formula 3 and 4 and marked with number in the detected square. |E| in the formula 3 means the number of detected character region and |T| in the formula 4 means the number of all character regions existing in the image. So, precision rate stands for the correctness of character region detected from the experiment and recall rate stands for how many character regions are detected from the image. Fig. 4 is the detected result in natural image taken in various environments as remarked above. We could find out that character region which couldn’t be detected with the edge-based method is able to detected with connecting element-based method. Like this, they can be done complementarily. Recall rate as experimental result was improved. But precision rate had detected a noise as the character region and the more detected noise region, the lower precision rate by combining with results of the edge-based and connected component-based method. In Fig. 4(a) indicates that indoor image has a high recall rate with a relatively simple background. Because the most characters in indoor image have a small size and connecting element-based method cannot get a feature value of the character region during the pre-procedure, it makes precision rate and recall rate low. We found out the skill to detect the feature of the small character region effectively with connecting element-based method. In Fig. 4(b) is as an image taken outdoors and we detected the character region when the background is simple. It is easy to detect the character region having a small character, too. But precision rate is lower than indoor image outdoor image has various environments and a complex background. In Fig. 4(c) has feature that the contrast value of the character region is high as the character region is shaded. So, recall rate is high when the bigger differences of the character region are, the more highlighted it is; however, there is an error in detecting the character region and recall rate becomes lower because the noise region is highlighted in accordance with the change of contrast. In Fig. 4(d) shows the character region is in the lighting and surroundings around the character are blurred. But because there is difference of the brightness between the character region and surroundings, rather the character region is detected easily. Therefore, it is the highest recall rate of the experimental results; on the contrary, precision rate is low with other effect of lighting. In Fig. 4(e) indicates recall rate and precision rate of proposed character region detecting methods are low entirely as a vertical character image is that the character we usually see on the store sign is vertical aligned. The reason is that the character aligned toward the right side has the unique feature not to start with vertical vowel as structural feature during the Hangul element merging procedure but the character aligned vertically doesn’t have the feature like that. Because the character region aligned vertically is affected according to the distance with other letters, this kind of trouble is arose when elements are not merged correctly if the character is relatively close to aligned letters in the procedure merging elements. This trouble needs to be solved by applying additional Hangul structural feature to vertical character merging procedure.
234
J.-C. Park, B.-M. Jun, and M.-K. Oh
Fig. 4. Result of Performance Evaluation about Character Region Detection
Performance evaluation for the result detecting the character region of the image taken from various environments is shown with each image type using the bar graph in Fig. 4. The result of indoor image is like that recall rate which is 84.6% and precision rate is 74.8% respectively. Indoor image has a higher detection rate relatively than other images due to a simple background. The image with the lowest performance is the vertical character image and it is resulted because the feature of merging Hangul element is not distinguished in the character region aligned vertically. As an overall experimental result, it turns out that the edge-based method has a high recall rate. It is because the edge feature of character region can be detected relatively easily, on the other hands, the connecting elements are affected from many environments and they cannot get the enough feature value. Instead, the edge-based method has a weakness that precision rate lowers by detecting not the edge of the character but the noise when the background is complex or thickness is distinct from each other. The edge-based method is affected less from the lighting and shadow and because a small sized character region can be detected with an edge feature, it turns out it is a major fact. The weaknesses of this method are that the character loses the structural feature as the edge is not connected and the edge is connected with noise around the character. These problems can be supplemented by detecting connecting elements using connecting element-based method. Proposed methods in this article have weaknesses that while recall rate goes up, precision rate goes down by merging two results of the edge-based method and connecting element-based method. The letter is low when the character region is not detected correctly with two methods by taking merging or if two regions are not overlapped with each other, the matching rate of the detected region goes down to half. Accordingly, we can reduce these problems if we conduct with merging the
Character Region Detection Using Structure of Hangul Vowel Graphemes
235
(a)
(b)
(c)
(d)
(e)
Fig. 5. Result of character region in various natural images. Left images are test images, Right images are results of character region detection. (a)Indoor (b)Outdoor (c)Shadow (d)Lighting (e)Vertical Text.
edge-based and connected component-based method before the detection of the character candidates. Therefore, the research should be conducted like above. Because the character region has local grouping generally, another method should study on the local grouping of the character region for detection result with the edge-based and connected component-based method or if there is any grouping like that, it can make precision rate high by detecting not a character region. Also, it can be improved by distinguishing the character region from background set up the threshold value after detecting distribution feature because pixel has a high contrast rate with the background.
236
4
J.-C. Park, B.-M. Jun, and M.-K. Oh
Conclusion
On this article, the method to detect the character region with Hangul structural feature was considered as the method to detect the Hangul in the natural image. Hangul elements merging algorithm was suggested to detect the character region among extracted values after detecting feature value of the character region using the method based on an edge and connected component. It could merge precisely Hangul elements in the character region as a special feature due to the feature which is distinguished from surroundings in merging procedure. The character region detected with Hangul elements merging algorithm was detected by procedure detecting Hangul character type. To detect Hangul character region, the 6 - type - detection algorithm was suggested. We got the result similar to the character detection with that of Hangul through Hangul character region detection algorithm by testing whether the character is in Hangul character region with using the location information and the number of Run from Hangul vertical vowels. After testing image taken with the mobile phone in diverse environments, whereupon the character region detection performance is 82.3%, the relatively high reproduction rate, but we got the relatively lower performance as the correctness rate is 57.9%. The next research project needs to be conducted on the method that merges the outcomes into one using the edge and revision connecting element-based method after extracting the character value. We should do the research on the slop detection and revision to detect the when the natural scene image is tilted slightly. Finally, we merged Hangul elements effectively in the character region based on the structural features of Hangul elements for the detection of character region. And then, we proposed the algorithm of vertical vowel detection for merging Hangul elements. The vertical vowel is needed to find out the unique feature value of the additional Hangul character.
References 1. Ezaki, N., Bulacu, M., Schomaker, L.: Character detection from natural scene images: towards a system for visually impaired persons. In: Proceedings of the 17th International Conference on, vol. 2, pp. 683–686 (2004) 2. Liu, X., Samarabandu, J.: An Edge-Based Character Region Extraction Algorithm for Indoor Mobile Robot Navigation. International Journal of Signal Processing 3(4), 273–280 (2006) 3. Kim, S., Kim, D., Ryu, Y., Kim, G.: A Robust License-Plate Extraction Method under Complex Image Conditions. In: Proceedings of International Conference on Pattern Recognition, vol. 3, pp. 216–219 (2002) 4. Smith, M.A., Kanade, T.: Video Skimming for Quick Browsing Based on Audio and Image Characterization. Carnegie Mellon University, Technical Report CMU-CS-95-186 (1995) 5. Lee, C.-M., Kankanhalli, A.: Automatic Extraction of Characters in Complex Scene Images. Pattern Recognition 9(1), 67–82
Character Region Detection Using Structure of Hangul Vowel Graphemes
237
6. Canny, J.: A Computational Approach to Edge Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-8(6), 679–698 (1986) 7. Cheon, P.J.: Detection Method of Character Region Using Hangul Structure from Natural Image, Department of Computer Engineering, Doctoral Thesis, University of Chungbuk National (1993) 8. Oh, I.G.: Study on the Extraction of Character and Special Character from Hangeul Documents with English, Master’s Thesis, Department of Computer Science, University of Kwangwoon (1993) 9. Yi, C., Tian, Y.: Character String Detection from Natural Scenes by Structure-based Partition and Grouping. IEEE Transactions on Image Processing, PMID:21411405 (2011) 10. KAIST Scene Character Database, http://ai.kaist.ac.kr/home/DB/SceneCharacter 11. Lucas, S.M., Panaretos, A., Sosa, L., Tang, A., Wong, S., Young, R.: CDAR 2003 Robust Reading Competition. In: 7th International Conference on Document Analysis and Recognition (ICDAR), pp. 682–687 (2003) 12. Lucas, S.M., Panaretos, A., Sosa, L., Tang, A., Wong, S., Young, R., Ashida, K., Nagai, H., Okamoto, M., Yamamoto, H., Miyao, H., Zhu, J., Ou, W., Wolf, C., Jolion, J.-M., Todoran, L., Worring, M., Lin, X.: ICDAR 2003 Robust Reading Competitions: Entries, Results and Future Directions. International Journal on Document Analysis and Recognition (IJDAR) 7(2-3), 105–122 (2005)
Task Scheduling Algorithm with Minimal Redundant Duplications in Homogeneous Multiprocessor System Inseong Song, Wanoh Yoon, Eunmee Jang, and Sangbang Choi Department of Electronics Engineering, Inha University, Incheon, Korea [email protected], [email protected]
Abstract. The performance of a homogeneous multiprocessor system depends on an algorithm which schedules input DAG. Among various task scheduling algorithms, ones that are using full duplication provide superior performance by duplicating tasks recursively. However, redundant duplications lead to increase of total execution time and decrease of processor idle time. In this paper, we propose GDRA which reduces total execution time and increases processor idle time by eliminating redundant duplications. Comparisons with other full duplication algorithms, such as CPFD, SD, and MDJ, show that GDRA provides the better performance as the CCR increases. Keywords: DAG, full duplication, homogeneous, multiprocessor, scheduling.
Task Scheduling Algorithm with Minimal Redundant Duplications
239
And the method that pre-confirms redundant duplications on join nodes and fork nodes before their actual allocation is adapted to increase processor idle time. The more idle time stands for the more opportunities for task insertion and as a result, total execution time can be reduced. Input graphs created with various parameters are applied to CPFD, SD, MDJ, and proposed GDRA for a performance evaluation. The results show that GDRA outperforms CPFD, SD, and MDJ, in terms of total execution time.
2
Problem Definition
An input task graph, which models parallel application that is to be executed in homogeneous multiprocessor system, is generally modeled as a DAG (Directed Acyclic Graph) G = (V, E). A DAG consists of nodes and edges. V is a set of n nodes which stand for tasks, and E is a set of e communication edges which stand for relations between two tasks. If there exists a directed edge from ni to nj, denoted by ei,j, then ni is called as a parent node of nj, and nj is called as a child node of ni. The set of all parent nodes of ni is called as pred(ni), and the set of all child nodes of ni is called as succ(ni). In Fig. 1, each node is represented by a circle, and each edge is represented by an arrow. The numbers in the upper and lower half denote the node number ni and the computation cost of the node, respectively. A number on an edge denotes a communication cost from ni to nj. A node without a parent node, such as n1, is called as a start node, and a node without a child node, such as n13, is called as an exit node. If there is more than one start or exit node, they have to be connected to a pseudo start or exit node with a pseudo edge which has zero computation cost and zero communication cost, so as not to affect the schedule. The cp(n) is a parent node from which the data arrives at the latest time. By duplicating the cp(n), the node n’s finish time can be reduced[4].
Fig. 1. An example of directed acyclic graph
240
I. Song et al.
Homogeneous multiprocessor system consists of a set P of p homogeneous processors, fully connected with each other. Assume that all the inter-processor communications are done without contention, and task executions are nonpreemptive. Also, assume that computation and communication can be overlapped. Task ordering is carried out according to a node’s ranku(ni) and rankd(ni) value. The ranku(ni) is the sum of computation costs and communication costs along the longest path from an exit node to the ni, and is calculated as follows[5]: ranku(ni) = c(ni) + maxn
j
succ(ni){d(ni,
nj) + ranku(nj)} .
(1)
The rankd(ni) is the sum of computation costs and communication costs along the longest path from a start node to the ni, and is calculated as follows[5]: rankd(ni) = maxn
3
j
pred(ni){rankd(nj)
+ c(nj) + d(ni, nj)} .
(2)
Proposed Algorithm
In this paper, we propose a novel task scheduling algorithm which removes redundant duplications in homogeneous multiprocessor system, called GDRA. GDRA consists of three major phases; task prioritizing phase, RN (Redundant Node) selection phase, and processor allocation phase. 3.1
Task Prioritizing Phase
In task prioritizing phase, the given DAG is traversed in a top-down fashion in order to group the nodes that are independent of each other. This is called the DAG leveling. As a result, nodes in the same group can be executed in parallel. Given a DAG, level 0 contains start node. Level i consists of all nodes nk such that, for all ej,ks, node nj is in a level less than i and there exists at least one ej,k such that nj is in level i-1. The last level comprises of exit node. The DAG leveling is carried out with the breadth first search. Then, calculate each node’s ranku(ni) and rankd(ni) with equation (1) and (2) to determine a critical path. Nodes with the equal ranku(ni) and rankd(ni) constructs critical path. In each level, a node on the critical path gains the highest priority and the other nodes gain priorities according to their ranku(ni). The higher ranku(ni), the higher priority. 3.2
RN Selection Phase
In RN selection phase, the candidate nodes that may cause redundant duplications are selected. RN selection phase consists of two minor stages: RJN (Redundant Join Node) selection stage, and RFN (Redundant Fork Node) selection stage. In the RJN selection stage, the join nodes ns which may cause the redundant duplication of cp(n) are selected and grouped into set RJNS (Redundant Join Nodes Set). Followings are the conditions of selecting the RJN.
Task Scheduling Algorithm with Minimal Redundant Duplications
241
Condition 1: A join node that is the only child of at least one of its parent. Condition 2: The communication cost between the node, the node which satisfies the condition 1, and its parent with the highest priority is larger than the computation cost of the parent. In the RFN selection stage, the fork nodes which may cause redundant duplication of themselves are selected and grouped into set RFNS (Redundant Fork Nodes Set). Following are the conditions of selecting the RFN. Condition 1: A fork node that is a RJN is not a RFN. Condition 2: A node which has at least one child node whose level is larger than the fork node as much as 1. Condition 3: The communication cost between the node, the node which satisfies the condition 1 and 2, and its child with the highest priority is larger than the computation cost of itself.
Group the nodes that are independent to each other Calculate ranku(ni) and rankd(ni) of all nodes Determine critical path for each level do Select and queue the node that are on the critical path Queue nodes in decreasing order of ranku(ni) endfor Make a final ordered list for input DAG Find RJN and group into RJNS Find RFN and group into RFNS for each node n in final ordered list do if duplication of n’s parent is possible then if the node n is not the parent of RJN or not the RFN then Assign the node with duplication technique else Assign the node with end technique endif elseif insertion is possible then Assign the node with insertion technique else Assign the node with end technique endif endfor Fig. 2. The pseudo code for the proposed GDRA
242
3.3
I. Song et al.
Processor Allocation Phase
In processor allocation phase, the nodes are assigned to the processor which provides the fastest execution finish time with task insertion technique or duplication technique[6], considering RN. First, the GDRA selects a first node n in the task priority list created in task prioritizing phase. Then GDRA tries to assign the node n with duplication technique considering RN, if it satisfies duplication conditions. At here, if the current node n is a parent of RJN or the current node n is a RFN, duplication does not take place. And if the node n does not satisfy duplication conditions, then GDRA tries to assign the node with insertion technique. If the node n cannot be assigned with both duplication technique and insertion technique, then the node is assigned with end technique[6]. Fig. 2 shows the pseudo code for the proposed GDRA, and Fig. 3 shows the flowchart for the proposed GDRA. DAG leveling
Calculate priority of each node Determine critical path
Order tasks in the level
NO
Last level?
YES Find RJN
Find RFN
Can use duplication?
NO Assign the node with insertion
YES
YES
RJN or RFN?
NO
YES
Can use insertion?
NO Assign the node with end technique
NO
Last node in ordered list?
YES Finish
Fig. 3. The flowchart for the proposed GDRA
Assign the node with duplication
Task Scheduling Algorithm with Minimal Redundant Duplications
4
243
Performance Evaluation
The performance of GDRA is compared with previous algorithms, CPFD, SD, and MDJ. Input DAGs are from STDGP (STanDard task Graph Project)[7-8]. The number of nodes, the number of processors, and CCR (Communication to Computation Ratio) are used as the input parameters for the simulations. Input parameters are restricted to the following values in the Table 1. Table 1. Input parameters Parameters The number of nodes The number of processors CCR
The number of nodes refers to the number of nodes in a DAG. STDGP provides standard DAGs with 50, 100, 300, 500, 750 nodes, and so on, 180 DAGs for each graph size. The number of nodes in a DAG can be regarded as a size of a parallel application. The number of processors refers to the number of processors in the homogeneous multiprocessor system, which greatly affects the performance. CCR is the ratio of the average communication cost to the average computation cost. Lower CCR value indicates that it is computation-intensive, and has lower chance of duplication. Total execution time, the Makespan, is used as the comparison metric. Makespan is the time that execution of given input DAG is finished. That is, the finish time of the exit node. The task scheduling algorithm that gives the lowest Makespan is the best algorithm in terms of the performance. For each graph size, 180 DAGs are used. By combining the number of processors and CCR, the total number of 22,500 graphs is used in this simulation. CPFD SD MDJ GDRA
Average Makespan
800
600
400
200
50
100
300
500
750
Number of Nodes
Fig. 4. Average Makespan with respect to the number of nodes
244
I. Song et al.
Fig. 4 shows the average Makespan with respect to the number of nodes. As shown in the figure, GDRA outperforms CPFD, SD, and MDJ. Furthermore, GDRA shows better performance as the number of nodes increases. Fig. 5 shows the average Makespan with respect to the CCR. As above, GDRA shows more improved performance as CCR increases. Since higher CCR stands for larger communication costs, this leads to more chances of duplication. GDRA is an algorithm which enhances performance by reducing redundant duplications, and as a result, it shows better performance as CCR increases. CPFD SD MDJ GDRA
460
Average Makespan
440 420 400 380 360 340
1
2
3
4
5
CCR
Fig. 5. Average Makespan with respect to the CCR
800
CPFD SD MDJ GDRA
Average Makespan
700
600
500
400
300
200 4
8
16
32
64
Number of Processors
Fig. 6. Average Makespan with respect to the number of processors
Task Scheduling Algorithm with Minimal Redundant Duplications
245
Fig. 6 shows the average Makespan with respect to the number of processors. The result shows that GDRA outperforms CPFD, SD, and MDJ.
5
Conclusion
In this paper, we presented the GDRA (Generalized Duplication Reduction Algorithm) for homogeneous multiprocessor system. GDRA reduces the total execution time and increases processor idle time by removing redundant duplications on join nodes and fork nodes. Decreased number of redundant duplications and increased processor idle time led to increase of insertion possibilities. Since insertion can enhance the performance of the algorithm significantly, GDRA was able to provide improved performance. As shown in the comparison results, the proposed algorithm, GDRA, provides better performance when compared to CPFD, SD, and MDJ with varying the number of nodes, CCR, and the number of processors in terms of total execution time, the Makespan. Acknowledgment. This work was supported by Key Research Institute Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology (2011-0018394).
References 1. 2.
3.
4. 5.
6. 7. 8.
Ahmad, I., Kwok, Y.K.: On Exploiting Task Duplication in Parallel Program Scheduling. IEEE Transactions on Parallel and Distributed Systems 9, 872–892 (1998) Bansal, S., Kumar, P., Singh, K.: An Improved Duplication Strategy for Scheduling Precedence Constrained Graphs in Multiprocessor Systems. IEEE Transactions on Parallel and Distributed Systems 15 (2004) Shin, K.S., Cha, M.J., Jang, M.S., Jung, J.H., Yoon, W.O., Choi, S.B.: Task Scheduling Algorithm Using Minimized Duplications in Homogeneous Systems. J. Parallel Distrib. Comput. 68, 1146–1156 (2008) Sinnen, O., To, A., Kaur, M.: Contention-aware scheduling with task duplication. J. Parallel Distrib. Comput. 71, 77–86 (2011) Topcuoglu, H., Hariri, S., Wu, M.Y.: Performance-Effective and Low-Complexity Task Scheduling for Heterogeneous Computing. IEEE Transactions on Parallel and Distributed Systems 13, 260–274 (2002) Sinnen, O.: Task Scheduling For Parallel Systems. Wiley (2007) Tobita, T., Kasahara, H.: A Standard Task Graph Set for Fair Evaluation of Multiprocessor Scheduling Algorithms. Journal of Scheduling 5, 379–394 (2002) http://www.kasahara.elec.waseda.ac.jp
RFID Mutual Authentication Protocols Using Key Exchange Based on Hash Functions Seung-Jae Jang1, Young-Hoon Jung1, So-Yeon Min2, Keun-wang Lee3, and Moon-Seog Jun1 1
Department of Computer Science, Soongsil University, Sangdo-Dong, Dongjak-Gu, Seoul 156-743, Korea {hezc81,jyh,mjun}@ssu.ac.kr 2 Department of Internet Information, Seoil University/Seoildaehak-gil 22 Jungnang-Gu Seoul 131-702, South Korea [email protected] 3 Department of Multimedia Science, Chungwoon University, Namjang-ri 29, Hongseong-eup, Hongseong-gun, Chungcheongnam-do 350-701, South Korea [email protected]
Abstract. A RFID system is a key technology used to create a ubiquitous society through non-contact radio identification technology, expected to provide us with information far above time and space. However, several security issues including a matter of privacy have been raised on account of restricted features of RFID communication environment. A variety of studies have been underway to bring solutions to them, but existing studies require excessive processing capabilities for security problems, tags and back-end database. To solve out this problem, this paper suggests mutual authentication RFID protocols that can guarantee safety from spoofing and re-transfer attack and also reduce the number of comparative operations in the back-end database, with the use of hash functions, XOR operation and key exchange. Keywords: RFID, Authentication, Key Exchange, Hash, XOR.
RFID Mutual Authentication Protocols Using Key Exchange Based on Hash Functions
247
restricted resources and features of low-cost RFID tags, creating a necessity of protocols which are simple and enable privacy protection. In particular, though a RFID tag enables various attacks, it is essential to have a mutually authenticable mechanism between the tag and the reader, and between the reader and the tag, as it can carry out an attack to acquire tag information illegally and deceive tag information by using an unauthorized and unjust tag or reader. Of them, existing main authentication mechanisms using hash functions include hash-lock technique, extended hash-lock technique, hash chain technique and hash ID modification technique. But these techniques are very vulnerable to spoofing or re-transfer attack. This paper explains in Chapter 3 it gives an insight into an introduction to and problems with the existing authentication techniques. In Chapter 4, the paper suggests light weight and highly stable protocols strong against spoofing and re-transfer attack while exchanging keys between tags and readers, not requiring much of the operation on back-end database, with the use of hash functions and XOR operations to be possibly embodied in tags. Chapter 5 of this paper suggests an analysis of the stability in authentication protocols and performance assessments, followed by Chapter 6 dealing with a conclusion.
2
Related Work
2.1
Hash-Lock Technique
Fig. 1. Hash-Lock Technique
The operating process of a hash-lock technique [2,3] is shown in Figure 1. Tags in the hash-lock technique store the metaID = H(key) with a random key hashed. When the reader requests an approach to the tag, it sends a metaID to the reader. The reader delivers metaID through the safe channel to the back-end database system, and receives the key to the tag and an ID value. The reader forwards the tag key again to the tag which then calculates hash values for the key from the reader compare the values to its metaID for identification and then delivers the identical ID to the reader. This technique uses a hash just one time, but it can trace down the tag location as metaID values are always fixed. When the legitimate metaID is received and metaID is sent to the reader afterward, it may bring a problem of acquiring Key and ID.
248
2.2
S.-J. Jang et al.
Extended Hash-Lock Technique
Fig. 2. Extended Hash-Lock Technique
The extended hash-lock technique [2,3], as shown in Figure 2, allows the tag to generate a random number R which is connected to its ID, calculate a hash value and then transmit R, H(ID||R) to the reader. Back-end database receives the value from the reader, make comparison to see if there is an identical value to what the tag operates, using all ID values and a random number R in its DB. Then it forwards the identical ID value to the reader, if any. This technique is strong against the spoofing attack as it uses a random number which changes information transferred each time from the tag to the reader. On the other hand, the IDk value can be exposed and when re-transferring the R, H(IDk || R) it can be posed as an legitimate tag. This technique also causes server overload as the back-end database needs to work on hash operation and comparison operation for all identification information to see if any of the information is identical to the value from the tag. Besides, the number of gate is too insufficient to embody the random number generator in the tag. 2.3
Hash-Based ID Modification Technique
Fig. 3. Hash-based ID Modification Technique
RFID Mutual Authentication Protocols Using Key Exchange Based on Hash Functions
249
Figure 3 shows Hash-based ID Modification technique[4]. A hash-based ID modification technique keeps changing ID in each session using hash functions. Tags include the following information : current ID, current session number TID, and last success session number LST. The back-end database system contains the same information as tags and H(ID) values at the same time. Once the system begins to operate, ID and TID are initialized to a random value, while △TID and TID hold the same value. This technique does not guarantee the integrity of ID and contains a disadvantage of necessarily maintaining the modified ID continually if applied to other RFID systems. In addition to this, it enables getting an approval as a legitimate tag when carrying out a re-transfer attack on the reader, pretending to be a tag prior to the ID value being changed, by posing as a legitimate reader to acquire H(ID), H(TID ⊕ ID), and △TID values, followed by a legitimate tag making the next session. H(ID) values remain the same anytime until the legitimate session is complete, so the technique holds a weakness of allowing the location tracking. 2.4
Improved Hash-Based ID Modification Technique
An improved hash-based ID modification technique [4] is a technique with the addition of partial security to the re-transfer and spoofing attack in a hash-based ID modification technique.
Fig. 4. Improved Hash-based ID Modification Technique
Figure 4 shows improved hash-based ID modification technique [5]. Like the hashbased ID modification technique, this technique still contains problems, such as allowing the location tracking as H(ID) values remain the same until the legitimate session is complete, as well as ID integrity and compatibility with other RFID systems.
3
Proposed RFID Mutual Authentication Protocol
The structure of a proposed RFID mutual authentication protocol, as shown in Figure 5, enables safe communications by renewing periodically the key value for the
250
S.-J. Jang et al.
tag and back-end database through a mutual authentication process between the reader and the tag, which makes the key useless in the following session though a hacker takes the key.
Fig. 5. Proposed Protocol Structure
3.1
Initial Information and Assumptions in Each Component
A tag stores ID, a unique identification number, previously hashed H(ID) and KeyT, the key value. DB has IC, the key value KeyDB and H(KeyDB) with KeyDB hashed, putting the KeyDB value in DB and the KeyT of a tag in the same initial configuration. Besides, a tag is able to do hash and XOR operations, while the reader enables hash operation, generating a random number, timer generation, XOR and a connection, supposing DB contains hash functions. 3.2 • • • • • • • • • •
Term Definitions
H() : Unidirectional hash functions ID : Unique bit identifying object information KeyDB : Key transferred from DB KeyT : Key transferred from a tag ⊕ : eXclusive-OR operation || : Connection C : H(ID||Timestamp) value used to create the following session key R : Random number H(ID)’ : H(ID) calculated from a tag Timestamp : Time generated by clock
RFID Mutual Authentication Protocols Using Key Exchange Based on Hash Functions
3.3
251
Operating Process of Proposed Protocol
Fig. 6. Operating Process of Proposed Protocols
As shown in Figure 6, the operating process of proposed protocols is as follows. 1. The reader transfers a request message to the tag. 2. The tag transfers H(KeyT) to the reader. H(KeyT) is used as a primary key to find out individual identification information and keys in DB. 3. The reader transfers H(KeyT) to DB. DB finds out, using its H(KeyDB) previously calculated, the requested key and ID in comparison with the H(KeyT) received from the reader. 4. DB transfers KeyDB and ID to the reader which then uses the KeyDB and ID received from DB and generate a random number to create a random number R and H(KeyDB||R), and create C = H(ID||Timestamp) using a timer. C is used as a new key for an authentication and hashes ID and timestamp as it should be a unique value. The reader hashes the C value to keep the KeyDB=H(C) value renewed. 5. The reader transfers a random number Rand H(KeyDB||R) and C⊕KeyDB to a tag. The tag hashes and compares its KeyT and R before authenticating the reader. Once the reader has been authenticated, the value C is extracted via C=(C⊕KeyDB)⊕KeyT) using the KeyT, then the KeyT=H(C) value of the tag is renewed. 6. The tag transfers H(ID)’⊕KeyT to the reader. The reader, which shares the same key with the tag, uses its KeyDB to extract H(ID)’. 7. The reader delivers H(ID)’, KeyDB to DB. DB compares H(ID)’ from the reader to H(ID) of its own before authenticating if it is a proper tag. If the tag has been authenticated, renew KeyDB and H(KeyDB).
252
S.-J. Jang et al.
4
Stability Analysis and Performance Assessments
4.1
Stability Analysis
(1) Re-transfer Attack A reader may forward the information in phase 5 to the tag without any suspicion, if a malicious attacker is posed as a legitimate reader, as shown in Figure 5, receives information H(KeyT) in phase 2, send the H(KeyT) received by pretending to be a legitimate tag as a response to a reader request. Besides, information in phase 6 will forward the information also to the reader if the reader stays near the tag. This way, all the information in phases 2, 5, 6, communications between the reader and the tag, this information cannot be used for a re-transfer attack whatsoever. Information (H(ID)'⊕KeyT) in phase 6 is the combination of the past key and data, so the reader extracts the wrong H(ID)' when the reader needs to extract it following the transfer made to the reader. Then, DB stops the authentication process when it fails in receiving this data and comparing it to H(ID)'=H(ID). (2) Spoofing Attack As discovering all information in phases 2, 5, 6 does not stop the data including key values different from random numbers in each session from being transferring, the tag posed as legitimate is not authorized to keep using this data until the following session. (3) Search Speed of Back-end Database DB prepares previously hashed. H(KeyDB) on a table. Like the extended hash-lock technique, if there are over 10,000 DB documents to search, all in hash formatting each time, it may increase overhead and bring the system shutdown. Previously hashed H(KeyDB) compares only data values which is therefore advantageous in the aspect of speed. This advantage is applied to the hash-based ID modification technique and improved hash-based ID modification technique, except for the hashlock technique and extended hash-lock technique. It features a fast –pacing search, but has a limitation of allowing a location tracking. (4) ID Integrity Maintenance The hash-based ID modification technique and improved hash-based ID modification technique modify ID values. Therefore, when applied to the RFID systems different from each other, one weakness is that they need to keep the modified ID values continually. ID integrity was maintained by modifying the key value, metadata, rather than changing the unique ID value. As the key value should be a unique value, H(ID||Timestamp) was used to maintain the unique value. (5) Growing Complexity in Attacks Using XOR Operation XOR operation is a technique used for stream coding, showing a weakness of allowing one operand known while the other becomes known automatically.
RFID Mutual Authentication Protocols Using Key Exchange Based on Hash Functions
253
All information between the tag and the reader contains all different operands, so it is difficult to analogize ID or key values. (6) Increase in Security via Mutual Authentication Regarding the hash-lock technique, it has a structure where a tag simply has provided meta information one-sidedly without checking if it is legitimate. But, proposed protocols improved security through mutual authentication in which the tag clearly certifies the reader and the reader certifies the tag, using key values. (7) Damage by Key Leak If a key is lost while used continually without getting renewed, numerous tags may have to be wasted, so the key renewal may cause an overhead but the damage by a key loss may be minimized. 4.2
Performance Assessment
In the RFID system, tags have restrictions on space and cost with respect to hardware design. The reader and back-end database, on the other hand, have no space limitations in hardware design, and it is possible to embody, without demanding great expense, such simple functions as hash functions, random number generation and XOR. Therefore, operation processing capability of a tag should be primarily considered in designing protocols. Table 1. Compare the Function of Each Protocol
Function
Classification
HL
Tag Reader DB Tag Reader DB Tag Reader DB Tag Reader DB
-
-
Authentication Step
6
5
Hash Function Random Number Generator ⊕
||
EHL
Proposed Protocols
HIdM
IHIdM
-
-
-
-
-
-
-
-
-
-
-
5
5
7
254
• • • •
S.-J. Jang et al.
HL : Hash-Lock EHL : Extended Hash-Lock HIdM : Hash-based ID Modification IHIdM : Improved Hash-based ID Modification
Table 1 shows a comparison of functions needed for each authentication protocol. Proposed protocols, compared to other protocols in the tag, do not need a special operation device, but need a timer for generating the timestamp in the reader and a random number generator for creating a random number. The reader, however, has better operation ability than the tag, so it is not too much to have a timer and a random number generator loaded extra. With respect to phases of authentication process, an element called a key was added to them for security reinforcement, which shows a couple of more phases than the conventional techniques. As the authentication between the reader and the tag is processed within a very fast period of time and the recognition speed of a microchip and the reader is growing with the development of hardware, seven phases of operation process in authentication do not work as a major obstacle.
5
Conclusions
RFID technology, a core technology to establish the ubiquitous computing environment, is applied in overall industry yielding a lot of profits. Regarding its adverse, however, greater damages may occur due to threats to the RFID system. To troubleshoot this, many conventional authentication techniques have been studied, each of which contains several problems. The mutual authentication protocol based on a key exchange as suggested in this paper is designed to be free from re-transfer attack, spoofing attack and location tracking attack which are the biggest issues in previously studied RFID authentication techniques. This was confirmed through the stability analysis and performance assessment and strengths of other proposed protocols were described beforehand. When applying the proposed protocols, it is expected to provide various utilities under the environment in demand of authentication and security which uses RFID including the RFID radio authentication system and the RFID-based home network system.
References 1. Sarma, S.E., Weis, S.A., Engels, D.W.: RFID Systems and Security and Privacy Implications. In: Kaliski Jr., B.S., Koç, Ç.K., Paar, C. (eds.) CHES 2002. LNCS, vol. 2523, pp. 1–19. Springer, Heidelberg (2003) 2. Weis, S.A., Sarma, S.E., Rivest, R.L., Engels, D.W.: Security and Privacy Aspects of Low-Cost Radio Frequency Identification Systems. In: Hutter, D., Müller, G., Stephan, W., Ullmann, M. (eds.) Security in Pervasive Computing. LNCS, vol. 2802, pp. 201–212. Springer, Heidelberg (2004)
RFID Mutual Authentication Protocols Using Key Exchange Based on Hash Functions
255
3. Weis, S.A.: Security and Privacy in Radio-Frequency Identification Devices: MS Thesis. MIT (May 2003) 4. Henrici., D., Muller, P.: Hash-based enhancement of location privacy for radio-frequency identification devices using varying identifications. In: PerSec 2004, pp. 149–153 (March 2004) 5. Lee, S., Hwang, Y., Lee, D., Lim, J.: Efficient Authentication for Low-Cost RFID Systems. In: Gervasi, O., Gavrilova, M.L., Kumar, V., Laganá, A., Lee, H.P., Mun, Y., Taniar, D., Tan, C.J.K. (eds.) ICCSA 2005. LNCS, vol. 3480, pp. 619–627. Springer, Heidelberg (2005) 6. Ohkubo, M., Suzuki, K., Kinoshita, S.: Hash-chain based forward-secure privacy protection scheme for low-cost RFID. In: Proceedings of the SCIS 2004, pp. 719–724 (2004) 7. Yang, J., Park, J., Kim, K.: Security and Privacy on Authentication Protocol for Low-Cost Radio. In: The 2005 Symposium on Cryptography and Information Security (2005) 8. Juels, A., Rivest, R., Szydlo, M.: The blocker tag: Selective blocking of RFID tags for consumer privacy (2002) 9. Sarma, S.E., Weis, S.A., Engels, D.W.: Radio Frequency Identification: Secure Risks and Challenges. RSA Laboratories Cryptobytes 6(1), 2–9 (2003) 10. Rhee, K., Kwak, J., Kim, S., Won, D.H.: Challenge-Response Based RFID Authentication Protocol for Distributed Database Environment. In: Hutter, D., Ullmann, M. (eds.) SPC 2005. LNCS, vol. 3450, pp. 70–84. Springer, Heidelberg (2005)
The Development of Passive Filter Control System Based on Power Measurement and Analysis Jeong-Chay Jeon, Jae-Hyun Kim, and Jae-Geun Yoo #27 Sangcheon-ri, Cheongpyeoung-myeon, Gapyeong-gun, Gyeonggi-do, 477-814, Rep. of Korea [email protected]
Abstract. The passive filter is economic and efficient in suppressing harmonics. But the effect of the passive filter is constantly dependent on power system impedance or working conditions of loads. This paper presented the control system can automatically control the passive filter according to working conditions of loads and measured harmonics, reactive power and power factor. Experimental results in the power system with passive filter are presented in order to verify the performance of the control system. Keywords: Harmonics, Passive filter, Control, Reactive power, Power Factor.
the reactive power of the passive filter when some harmonic generation loads are stopped, there are concerns regarding the increase of power losses of the power system and transformer, and rise of power system voltage. In particular, in the event that power system impedance is varied and the harmonic component in the power source is included, there is a need to protect the filter by monitoring it because of series and parallel resonance problems. The passive filter must be controlled to avoid overload of the other branches when troubles with some branches of the passive filter are generated. In order to solve the problems in dealing with passive filters, this paper developed a control system to automatically control the passive filter according to operating conditions of loads, along with a program to monitor operating states of the passive filter as well as the control system, harmonics, power factor, voltage, and current in the power system.
2
Passive Filter
Passive filters are largely classified into a tuned filter to eliminate low order harmonics and a high pass filter to eliminate high order harmonics (e.g. 17th up) . The parameter to be considered prior to the selection of R, L and C during passive filter design is the quality factor (Q), and the quality factor determines the sharpness of tuning. In this respect filters may either be of high Q type or low Q type. The high Q type filter is sharply tuned to one of the lower harmonic frequencies and a typical value is between 30 and 60. The low Q type filter, typically in the region of 0.5 5, has a low impedance over a wide range of frequencies [1]. A single tuned filter as shown in Figure 1 consists of a series R-L-C circuit and is generally tuned to a lower characteristic harmonic. Its impedance ( Z f ) and quality
～
factor are given by equations (1) and (2), which at the resonant frequency (
X0 )
reduces to R
1 Z f = R + j wL − wC X Q= 0 R where
(1) (2)
X 0 is reactance of inductor or capacitor in ohms at the tuned frequency, and it
is given by equation (3).
X 0 = wn L =
L 1 = wnC C
where wn is tuned angular frequency in radians per second.
(3)
258
J.-C. Jeon, J.-H. Kim, and J.-G. Yoo
In the high pass filter, impedance and quality factor (Q) are equations (4) and (5) respectively. −1
1 1 1 − Z f = + jwC R jwL
(4)
R R = wn L X 0
(5)
Q=
3
The Passive Filter Control System
The passive filter control system, which is proposed in this paper, consists of a signal input voltage component, current measuring sensor and filter, A/D conversion part converting analog signal into digital signal, digital signal processing element controlling circumference installations and performing operation functions, and an output part sending signals to control the passive filter as shown in Figure 1.
Fig. 1. Block-diagram of passive filter control system
The passive filter control system in Figure 1 adjusts the analog signal into a signal suitable to the A/D converter after inputting voltage and current signal by voltage and current sensor and converting it into digital signal by the A/D converter. In order to simultaneously convert the analog signal of voltage 3 and current 4 channels into a digital signal, the A/D conversion component was designed using Analog Devices 14bit A/D converter AD 7865 having an analog input 4 channel of AC level and conversion start signal input ability. A/D converted data is transmitted to Microchip’s microprocessor PIC24 HJ256GP210A of the digital signal processing part, and the microprocessor performs frequency analysis by Fourier conversion on each cycle of voltage and current, and calculates power (active, reactive and apparent power), power factor and so on. Picture 1 shows the developed passive filter control system.
The Development of Passive Filter Control System
259
Pic. 1. The developed passive filter control system
4
Test and Results
4.1
Measurement Accuracy Test of the System
First, we tested accuracy of voltage and current measurement by the proposed system. Test system, which is consisted of standard AC power source (Fluke 6100A) and digital oscilloscope (Tektronix TDS7104), was constructed as shown Figure 2. Table 1 and 2 show the test results that the proposed system yielded less than 0.3 % error in all case.
G Fig. 2. Test Circuit Table 1. Voltage measurement results Measurement Value and Accuracy
Vref [V]
Vr[V]
Accuracy[%]
Vs [V]
Accuracy [%]
Vt [V]
Accuracy [%]
12
12.0
100.0
12.0
100.0
12.0
100.0
25
25.0
100.0
25.0
100.0
25.0
100.0
50
49.9
99.8
49.9
99.8
49.9
99.8
110
109.8
99.8
109.8
99.8
109.8
99.8
220
219.8
99.9
219.8
99.9
219.8
99.9
330
330.0
100.0
330.0
100.0
330.0
100.0
380
379.6
99.9
379.6
99.9
379.4
99.8
260
J.-C. Jeon, J.-H. Kim, and J.-G. Yoo Table 2. Current measurement results
Iref[A]
Measurement Value and Accuracy Ir [A] Accuracy[%] Is [A] Accuracy[%] It[A] Accuracy[%] In [A] Accuracy[%]
5.0
5.1
102.0
5.1
102.0
5.1
102.0
5.1
102.0
10.0
10.0
100.0
10.0
100.0
10.0
100.0
10.0
100.0
20.0
20.0
100.0
20.0
100.0
20.0
100.0
20.0
100.0
40.0
40.1
100.3
40.1
100.3
40.0
100.0
40.0
100.0
60.0
60.0
100.0
60.0
100.0
60.0
100.0
60.0
100.0
4.2
Field Test of the Developed System
Table 3 lists the design and production specifications of the passive filter for test of the developed system. The passive filter consists of a tuned filter for third, fifth, seventh order and a high-pass filter for the 11th and high orders. SSR (Solid State Relay) was used to switch each branch of the passive filter by the signal outputted in the passive filter control system. Table 3. Passive filter design and production specification
Section Capacitor (kVAR) Inductor (mH)
Third order Fifth order 6.1 2.6
5.7 0.95
Seventh order
Hi-pass
5.5 0.4
10 0.09
As shown in Figure 3, the control system developed in this paper was connected with a passive filter in the power system using a 100HP DC motor, an extruding machine, an air blower and other loads. The power system of Figure 3 used a DC motor to produce automobile soundproofing material, and it was difficult to optimize effects of the passive filter due to repeated start and stop of the DC motor.
Fig. 3. Wiring diagram of the passive filter control system
The Development of Passive Filter Control System
261
Figure 4 shows reactive power magnitude variation according to load operation condition variation. In Figure 4, because the passive filter is automatically closed by the control system when the DC motor is being operated, reactive power is reduced to 5kVAR by compensation effectiveness of the passive filter. Also, there is no reactive power after a set time because the passive filter is automatically opened by the control system after a set time (five minutes according to the operating pattern of the DC motor in this paper) when the DC motor is stopped.
s
s
s m vu
m vmm
Fig. 4. Variation of reactive power magnitude
s s
s
m vu m vu
m vmm m vmm
Fig. 5. Variation of power factor
5
Conclusion
This work developed a passive filter control system that can automatically control the passive filter according to operating conditions of loads, harmonics, reactive power, power factor and so on. The passive filter control system was tested with the passive filter. Test results illustrated that the control system suppressed the occurrence of reactive power by automatically opening the passive filter when the DC motor is
262
J.-C. Jeon, J.-H. Kim, and J.-G. Yoo
stopped and reduced reactive power by automatically closing the passive filter when the DC motor is operated. It is judged that the passive control system could solve both the over compensation problem of reactive power and the resonance problem with the power system by automatically controlling the passive filter.
References [1] Arrillaga, J., Bradley, D.A., Bodger, P.S.: Power System Harmonics, pp. 41–109. John Willy & Sons (1985) [2] IEEE Std. 519-1992, IEEE Recommended Practices and Requirements for Harmonic Control in Electric Power Systems (1993) [3] McGRanaghan, M.F., Mueller, D.R.: Designing harmonics filters for adjustable-speed drives to comply with IEEE-519 harmonic limit. IEEE Trans. on Industry Application 35(2), 312–318 (1999) [4] Sato, Y., Kawase, T., Akiyama, M., Kataka, T.: A Control Strategy for General-Purpose Active Filters Based on Voltage Detection. IEEE Trans. on Industry Application 36(5), 1405–1412 (2000) [5] Chang, T.-T., Cang, H.-C.: An Efficient for Reducing Harmonic Voltage Distortion in Distribution Systems with Active Power Line Conditioners. IEEE Transaction on Power Delivery 15(3), 990–995 (2000) [6] Peng, F.Z.: Harmonic sources and Filtering Approaches. IEEE Industry Application Magazine, 18–24 (August 2001) [7] Wu, C.J., Chiang, J.C., Yen, S.S., Liao, C.J., Yang, J.S., Guo, T.Y.: Investigation and Mitigation of Harmonic amplification problems caused by single-tuned filters. IEEE Trans. on Power Delivery 13(3), 800–806 (1998)
Preparation and Characteristics of Poly(St-co-MAA) Microgels Young Sik Kim1,*, Sang Hyun Moh2, Suk Hyung Cho3, Young Jun Kim4, and Jea Chul Cho5 1
Department of Chemical Engineering, Kangwon National University, 1 Joongang-ro, Samcheok, Gangwon, Korea 245-711 2 BIO-FD&C Co., Ltd, Phytomecca Co., Ltd, Pilot plant 7-50, Songdo-Dong, Yeonsu-Gu, Incheon, Korea 406-840 3 Department of Fire Safety and Management, Hyejeon University, Hongsung-up Chungnam, Korea 350-800 4 Department of Cosmetic Science, Chungwoon University, Hongsung-up Chungnam, Korea 350-800 5 Department of Hotel Culinary Arts & Food Service, Hyejeon University, Hongsung-up Chungnam, Korea 350-800 [email protected]
Abstract. By using suspension polymerization, molecularly imprinted porous poly(St-co-MAA) gels were prepared from styrene (St), methacrylic acid (MAA) monomer and ethylene glycol dimethacrylate (EGDMA) as a crosslinking agent. The polymerization condition was chosen that the formation of solution complexes between methacrylic acid and the template(caffeine) prior to polymerization gave better results. Polymerization parameters such as mole ratio of monomers, concentration of monomers, concentration of a crosslinking agent and concentration of a blowing agent on the average particle size and pore size distribution were investigated. Keywords: suspension polymerization, imprinted porous polymer, microgels microgels.
1
Introduction
As every manufactured products are diversified with industrial development, analysis of product ingredients and separation and refinement of compounds demand a high level of accuracy. Thus, analysis methods such as HPLC and GPC are widely used. Porous poly gels are usefully used as fillings of column for HPLC and GPC [1-5]. Monodisperse polymer microgels have a variety of size, pattern and structure of particles upon preparation method. Thus, polymer microgels with appropriate characteristics can be manufactured with monodisperse polymer microgels according to the purpose of use. These polymer microgels are rather easy to functionalized by employing manufacturing method and functional groups and a lot of studies on its *
application is being tried due to its large surface area. Especially, the studies on porous polymer microgels with uniformed particle size are popular: such as column fillings of HPLC and GPC, carriers for enzyme fixations and protein adsorption [6-11], chelate resin that selectively adsorbs and separates ionic materials in heavy metal and seawater[12-13], adsorption and separation of protein and enzyme, adsorption and separation of metal ions, and development of application as catalyst[14-16]. Recently, poly gels are used for cosmetic fillings, toner for digital images, gap controller for liquid crystal display panel and standard particle of coulter counter. Studies have been focused on application of poly gels to biomedical area, e.g. as a diagnostic reagent and drug carrier. To produce polymer microgels, emulsion polymerization, dispersion polymerization and suspension polymerization are mostly used. According to polymerization methods, poly gels have different characteristics in particle size, dispersion of particles and molecular weight of polymers. Generally, monodisperse particles of polymers with are produced Emulsion polymerization, while polydisperse particles under 1 are made by dispersion polymerization, which particles in size within 1-10 molecular weight cannot be controlled [17-18]. Seed polymerization is the way to synthesize poly gels with uniformed dispersion of particles which monodisperse polymer seed, linear or cross-linked synthesized by emulsion polymerization and dispersion polymerization are increased in particle size by swelling by monomer or organic solvent [19-21]. Also, particles over 10 can be obtained by suspension polymerization. However, its polydisperse particles may cause reduced column efficiency, decreased pressure and shortened polymer life due to fine particles if it is used as column fillings. On the other hand, its manufacturing method is simple and various vinyl monomers can be used rather than other polymerization methods. Also, it is widely used for various area by employing functional groups to polymer particles. In this study, porous foamed polymer microgels with uniformed size of particles were synthesized by suspension polymerization. Also, the experiments were performed on mol ratio of monomer as factors that affect the characteristics of poly gels, particle size and particle size distribution by the amount of cross-linking agent.
㎛
㎛
㎛
2
Experiment
2.1
Reagent
Monomer styrene (St) used in this study was purchased from Junsei Chemical, Japan. After washed with 10% NaOH, it was vacuum distilled under nitrogen stream and stored at -5°C. and Methacrylic acid (MAA) was purchased from Aldrich and vacuum distilled under nitrogen stream. Ethlene glycol dimethacrylate (EGDMA), a grosslinking agent was purchased from Aldrich. After washed with 10% NaOH, it was vaccum distilled under nitrogen steam. Benzoyl peroxide (BPO), a polymerization initiator was purchased from Wako Pure Chemical and recrystallized with methanol. Also, polyvinyl pyrrolidinone (PVPd) as suspension stabilizer, n-hexane as blowing agent, toluene as swelling agent, and acetonitrile as porogen were purchased from Junsei Chemical.
Preparation and Characteristics of Poly(St-co-MAA) Microgels
2.2
265
Synthesis of Copolymer Gel
Synthesis of Poly(St-co-MAA) gel. dissolve PVPd the suspension stabilizer in 200ml of distilled water. Then, add cross-linking agent, EGDMA and stirred in the water. And then, dissolve benzoyl peroxide in the water and add monomer styrene and methacrylic acid. Put this liquid into 500ml erlenmeyer flask. Add nitrogen stream to remove dissolved oxygen and stir and polymerize it at 300rpm at 70°C for six hours. At this time, fix monomer concentration in 10wt.% and change mol ratio of St/MAA to 95/5, 90/10, 85/15 and 80/20 each and copolymerize it. Change EGDMA concentration to 0.25 5 wt% and synthesize poly(St-co-MAA) gel. To remove unreacted reactants, wash them more than three times using methanol and distilled water and vacuum dry it at 50°C.
～
2.3
Characteristics of Microgels
Analysis of chemical composition. Furrier Transfer Infrared Spectroscope(IFS66/FRA 106S, Bruker) was used to examine IR spectrum of microgel in order to find the composition of prepared poly gels. Measurement of particle pattern and particle size distribution. Scanning electron microscope(LEO-1530FE) was used to observe particle size and the surface pattern at 10 kV. The prepared poly gel was spreaded on the sample plate of the microscope and observed. Also, ultrasonic generator and agitator were used to uniformly distribute specific amount of poly gel. Particle analyzer (Malvern Mastersizer 2000) was used to measure particle size of suspension. The particle size distribution was calculated by Span value using the following equation: Span=represents volume per cent that microcapsule accounts for and Dn% indicates average particle size in the area. Accordingly, small span value indicates monodisperse microcapsule.
3
Results and Discussion
Poly(St-co-MAA) gel was prepared by suspension polymerization changing mol ratio of St/MAA and using BPO as initiator and EGDMA as cross-linking agent. while monomer concentration was changed to 10, 15 and 20 wt%; St/MAA mol ratio was to 95/5, 90/10, 85/15, 80/20; cross-linking agent, EGDMA concentration was changed to 0.5, 1.0, 1.5, 2.0, and 2.5 wt%, particle size of prepared poly gel was measured by SEM image and PCS method. Sphere type of fine particles were obtained with different particle sizes upon mol ratio of monomer and concentration of cross-linking agent. Prepared poly(St-co-MAA) gel particles are distributed within the range of 93-643 . SEM images of the gel particles prepared upon each reacting condition were shown in Figure 1, Figure 3 and Figure 5. As shown in Figure 1 and 2, as St/MAA mol ratio changed to 95/5, 90/10, 85/15 and 80/20, particle size of microgel tended to increase.
㎛
266
Y.S. Kim et al.
Fig. 1. SEM images of the microgels prepared with various molar ratio of monomer: (a) St/MAA(95/5); (b) St/MAA(90/10); (c) St/MAA (85/15); (d) St/MAA(80/20)
Exp. No
Particle size ( )
㎛
Span
a
134
1.66
b
164
1.39
c
206
1.99
d
271
1.95
Fig. 2. Particle size and size distribution of the microgels prepared with various molar ratio of monomer: (a) St/MAA(95/5); (b) St/MAA(90/10); (c) St/MAA(85/15); (d) St/MAA(80/20)
Preparation and Characteristics of Poly(St-co-MAA) Microgels
267
Also, as shown in Figure 3, if St/MAA mol ratio was 90/10, as EGDMA increased, particle size decreased. As shown in Figure 4, as cross-linking agent increased, the range of particle size distribution got narrow. When the amount of hydrophilic MAA increased than that of hydrophobic St which forms nucleus, gel particles are polydispersed and the average size of particle increased. This is because swelling
Fig. 3. SEM images of the microgels prepared with various concentration of crosslinker: (a) 0.5 wt%; (b) 1.0 wt%; (c) 1.5 wt%; (d) 2.5wt%
Exp. No
Particle size ( )
㎛
Span
a
634
2.29
b
367
1.99
c
180
1.56
d
93
0.84
Fig. 4. Particle size and size distribution of the microgels prepared with various concentration of crosslinker: (a) 0.5 wt%; (b) 1.0wt%; (c) 1.5 wt%; (d) 2.5 wt%
268
Y.S. Kim et al.
occurs when polymerized when more hydrophobic MAA were contained and cohesion event occurred while swollen gel was drying. Also, as shown in Figure 3 and 4, when cross-linking agent increased to 0.5, 1, 1.5, 2.0 and 2.5 wt% while preparing poly (St-co-MAA) gel, the particle type got uniformed and the size got smaller as crosslinking agent increased. It is because cross-linking density increased and particle internal structure got dense as cross-linking agent increased.
Fig. 5. SEM images of the microgels prepared with various concentrations of monomer: (a) 10 wt%; (b) 15 wt%; (c) 20wt%
Exp. No
Particle size ( )
㎛
Span
a
164
1.39
b
334
8.84
c
473
7.86
Fig. 6. Particle size and size distribution of the microgels prepare with various concentration of monomer: (a) 10 wt%; (b)15wt%; (c) 20 wt%
Preparation and Characteristics of Poly(St-co-MAA) Microgels
269
When St/MAA mol ratio was 90/10 and cross-linking agent was 2 wt%, monomer to solvent was changed to 10, 15 and 20 wt%. As shown in Figure 5 and 6, as the level of monomer increased, particle size increased and particle distribution also increased. As dispersed medium increased, monomer bubbles relatively decreased. Thus, the opportunity for them to cohere each other increased and more stable and uniformed bubbles are created. Figure 7 indicates FT-IR results that indicates the structure of poly(St-co-MAA) prepared. Absorption peak by C-H bond was shown in 3100-3000 cm-1; the absorption peak by C=C and =CH by double bond of aromatic series was shown in 1600-1400 cm-1 and 800-700 cm-1; the characteristic peak of vinyl group (CH2=CH-) was shown in 1600 cm-1. Crossed-linked poly(St-co-MAA) was proved in that peak of C=O bond of MAA was shown in 1700 cm-1 and absorption peak of C-O of EGDMA was shown in 100-1100cm-1.
Fig. 7. FT-IR spectrum of poly(St-co-MAA) gels
4
Conclusion
In this study, St and MAA were used as monomers and EGDMA was used as crosslinking agent to produce poly(St-co-MAA) gel upon mol ratio of monomers and the level of cross-linking agent. 1. When monomer level was 10 wt%, mol ratio of St/MAA, 90/10 and cross-linking agent, 25 wt%, poly(St-co-MAA) with the most uniform particle size was obtained. Here, relatively monodisperse poly gel was produced with 93 in particle size and 0.84 in span value.
㎛
㎛
㎛
to 271 2. As mol ratio of St/MAA increased, particle size increased from 134 while as level of crossing agent increased, particle size decreased from 643 to 93 .
㎛
㎛
3. When mol ratio of St/MAA was 90/10, level of cross-linking agent, 2.0 wt%, and blowing agent level, 20 wt% and swelling agent level, 20 wt%, foamed poly(Stco-MAA) gel with uniformed particle size was obtained. Here particle size was 943 and foaming magnitude was approx. 5.8.
Ogino, K., Sato, H.: , 46, 667 (1989) Kulin, L.I., Ellingsen, T., Ugelstad, J.: J. Chromatogr 514, 1 (1990) Ellingsen, T., Aune, O., Ugelstad, J., Hagen, S.: J. Chromatogr 535, 147 (1990) Cheng, C.M., Vanderhoff, J.W., El-Asser, M.S.: J. Polym. Sci. 30, 245 (1992) Ugelstad, J., Kaggerud, K.H., Hansen, F.K., Berge, A.: Markrom. Chem. 180, 737 (1979) Gehrke, S.H., Andrews, G.P., Cussler, E.L.: Chem. Eng. Sci. 41, 2153 (1986) Kawaguchi, H., Hoshinon, F., Ohtsuka, Y.: Makromol. Chem. Rapid Commun. 7, 109 (1986) Freitas, R.F.S., Cussler, E.L.: Chem. Eng. Sci. 42, 97 (1987) Okubo, M.: Polymer International 30, 469 (1993) Park, T.G., Hoffman, A.S.: J. Biomed. Mater. Res., 24, 21 (1990) Park, T.G., Hoffman, A.S.: Biotech. Bioeng. 35, 152 (1990) Egawa, H., Nakayama, M., Nonaka, T., Uemura, K.: J. Appl. Polym. Sci. 34, 1557 (1987) Kim, K.S., Kang, S.H., Park, M.H., Lee, S.H.: J. Ind. and Eng. Chem (Korea) 2, 45 (1996) Kawaguchi, H.: Gendai Kagaku 170, 34 (1985) Kawaguchi, H.: Kobunshi Kakou 30, 477 (1984) Scott, C.D.: Clin. Chem. 14, 521 (1986) Shiroya, T., Yasui, M., Fujimoto, K., Kawaguchi, H.: Colloid and surfaces B: Biointerfaces 4, 267 (1995) Achiha, K., Ojima, R., Kasuya, Y., Fujimoto, K., Kawaguchi, H.: Polym. for Adv. Tech. 6, 534 (1995) Kawaguchi, H.: Prog. Polym. Sci. 25, 1171 (2000) Pelton, R.H., Chibante, P.: Colloids Surf. 120, 247 (1986) Clarke, J., Vincent, B.: J. Chem. Soc., Faraday Trans. 77, 1831 (1981) Liu, Z.S., Xu, Y.L., Yan, C., Gao, R.Y.: Anal. Chem. Acta 523, 243 (2004) Chen, X., Yia, C., Yan, X.Q., Wang, X.R.: J. Chromatohr, B 812, 149 (2004) Kugimiya, A., Takeuchi, T.: Anal. Chim. Acta 395, 251 (1999) Yano, K., Karube, I.: Trends in Anal. Chem. 13(3), 199 (1999) Kempe, M., Mosbach, K.: J. Chromatogr. A 664, 276 (1997) Cheng, S.H., Oh, C.Y., Seo, J.I., Park, J.K.: Korean J. Bioeng. 16, 115 (2001)
Identifying Organizational Culture for Patient Safety Yoo Kyung Boo1, Hyeon Sil Lee2, Ji Won Park3, and Seung Woo Ham4 1
Department of Healthcare Management, Eulji University Department of Health Administration, Korea University 3 Department of Medical Management Administration, Gimcheon University 4 Department of Medical Record, Korea Institute of Radiological & Medical Sciences, 215-4, Gongneung-Dong, Nowon-Gu, Seoul 139-706, Korea [email protected] 2
Abstract. A survey on organizational culture for patient safety was conducted among a total of 391 physicians, nurses, and health professionals working at four general hospitals in Seoul, using the Hospital Survey on Safety Culture instrument developed by AHRQ of the U.S. The results of the study showed that the patient safety awareness significantly differed by gender, age, career, average work hours, work units, contact with patients, and frequency of reported events. A good work climate and the immediate supervisor’s high level of patient safety awareness had positive influences on patient safety, and the hospital management’s low patient safety awareness had a negative influence thereon. To ensure patient safety in a hospital, the management’s leadership is required. In addition, the education of the medical professionals who come in direct contact with the patients should be reinforced, and a patient safety management system that is linked to the hospital information system should be developed. Keywords: patient safety, quality of care, organizational safety culture, medical error reporting, health information technology.
care [5]. Data on medical errors, however, are very scarce. As such, there is a need to lay the foundation for the fast establishment of a patient safety culture in local medical institutions by investigating the medical institutions’ climate, organizational culture, and medical errors regarding patient safety. This study aimed to provide the basic data required for the early establishment of a patient safety culture through the investigation of the hospital climate, organizational culture, and current status of medical errors in various medical professions’ reports with regard to patient safety.
2
Materials and Methods
2.1
Subjects
This study was conducted on physicians, nurses, and health professionals (pharmacists and medical technicians) of four randomly sampled general hospitals with over 200 beds in Seoul City. 2.2
Instruments
The survey questionnaire that was used in this study for the investigation of the medical professionals’ patient safety awareness was based on the Hospital Survey of Patient Safety Culture instrument developed by the Agency for Healthcare Research and Quality(AHRQ) of the U.S. as well as on other related studies and literature [6]. The survey questionnaire consisted of three areas (hospital climate, organizational culture, and medical errors) and 53 items. The survey questionnaire items were graded based on a 5-point Likert scale, where “1” corresponds to strongly disagree, “2” to disagree, “3” to neither agree nor disagree, “4” to agree, and “5” to strongly agree. 2.3
Data Collection and Analysis
The survey questionnaire was self-administered by the subjects from May 10 to June 9, 2011, in cooperation with their respective hospitals. A total of 500 questionnaire forms were distributed, and 400 (80%) were retrieved. Of the 400 retrieved forms, 391 (78.2%) were used for the final analysis, excluding nine that contained what were suspected as being insincere answers, or where many items were not answered. For the reliability of the survey, Cronbach’s alpha was 0.945. Frequency test, chi-square test, t test, one-way ANOVA, multiple-regression analysis, and logistic-regression analysis were conducted on the results of the survey, using SPSS 18.0.
3
Results
3.1
General Characteristics of the Survey Respondents
The general characteristics of the survey respondents are presented in Table 1. The test of the association of patient safety awareness with medical profession showed significant
Identifying Organizational Culture for Patient Safety
273
differences in gender (p=0.000), age (p=0.000), entire career (p=0.000), career in the current work unit (p=0.000), work hours per week (p=0.000), work units (p=0.000), contact with patients (p=0.000), and number of reported events (p=0.014) (Table 1). Table 1. General characteristics of respondents
Career - years Current specialty - years Working hours per week Work area
*p<.05, **<.01
3.2
N
%
female male
309 72
81.1 18.9
~30 31~40 40~ ~1 1~5 6~10 11 ~
189 112 59 69 158 75 86
~1 1~5 6~10 11 ~
1)
Physician
X2
p
186.24
*
32.858
*
106.57
*
112.52
*
251.29
*
226.44
*
13(.033)2 5(.064)
39.937
*
12(.031) 18(.046) 7(.018) 0(.000)
16.01
*
Intensive care unit.
Medical Professionals’ Patient Safety Awareness
The medical professionals’ awareness of the hospital climate, organizational culture, medical-error reporting, and patient safety are shown in Table 2. The medical professionals’ overall patient safety awareness was equivalent to 3 points or below (moderate) out of the perfect score of 5 points. The physicians showed a slightly higher level of awareness than the nurses and health professionals in all the items, except for patient safety awareness. The post-hoc Dunnett test for ANOVA showed a significant difference in hospital climate between the physician group and the nurse/health professional group, in the immediate supervisor’s patient safety awareness between the physician group and the nurse/health professional group; and in the hospital management’s patient safety awareness between the physician group and the nurse/health professional group.
274
Y.K. Boo et al. Table 2. Patient safety awareness among the medical professionals Domain
Specialty
Hospital climate Supervisor/ Manager Communications Frequency of events reported Patient safety grade Hospital management’s patient safety awareness
3.3
N
Mean
Sd
F
p
Dunnett
Physician(a) Nurse(b) Health Prof.(c)
79 269 38
2.77 2.37 2.36
.596 .327 .236
32.916
.000
a>b,c
Physician(a) Nurse(b) Health Prof.(c) Physician(a) Nurse(b) Health Prof.(c) Physician(a) Nurse(b) Health Prof.(c) Physician(a) Nurse(b) Health Prof.(c)
Factors That Had an Influence on the Assessment of Patient Safety
The stepwise selection of multiple-regression analysis showed that the hospital climate (p=0.000) and the immediate supervisor’s patient safety awareness (p=0.006) were independent variables with a positive (+) influence on the assessment of patient safety, and that the hospital management’s patient safety awareness (p=0.000) was an independent variable with a negative (-) influence on the assessment of patient safety (Table 3). Table 3. Influence factors on the assessment of patient safety dependent variable Patient safety grade
independent variable (constant) x1 x2 x3
β
t
-.105 .210 -.238
20.892 -1.814 3.582 -.3.116
P
Tolerance
.000 .000 .000 .002
.706 .681 .404
VIF
1.417 1.468 2.473
R=.340, R2-square=.116, Adj-R2=.106, F=12.305, p=.000 x1) Frequency of Events Reported, x2) hospital management’s patient safety awareness, x3) Hospital climate.
Identifying Organizational Culture for Patient Safety
3.4
275
Influence Factors on the Patient Safety in Relation with Patient Contacts
The logistic-regression analysis that was performed to analyze the influence of the medical professionals’ contact with patients on patient safety predicted that the possibility of a medical professional coming in contact with a patient will increase by e1.341=3.823-fold in “one team helps another team if the work unit is really busy”, by e-.979=.376-fold in “a work unit’s procedure or system is helpful for preventing errors”, and by e-.859=.424-fold in “things fall between the cracks when transferring patients from one unit to another” compared with when a medical professional does not come in contact with a patient if one unit increases in each variable, when it is assumed that the scores of the other independent variables were equal (Table 4). Table 4. Prediction of patient safety according to contact with patients B x1 x2 x3 constant
x1: one team helps another team if the work unit is really busy, x2 : a work unit’s procedure or system is helpful for preventing errors, x3: things fall between the cracks when transferring patients from one unit to another.
4
Discussion
Patient safety can be defined as the avoidance and prevention of patient injuries or adverse events resulting from the process of healthcare delivery [7]. Nowadays, patient safety has become the minimum principle that the entire area of medical services should comply with. In South Korea, studies on patient safety have been extensively conducted due to the influence of the accreditation and assessment of medical institutions and the increased awareness of medical errors in developed countries [8, 9]. This study was conducted to identify hospital climate and organizational culture for patient safety as well as the current status of medical-error reporting, to improve the quality of medical care by establishing a patient safety culture. There was nearly no difference in patient safety awareness between the physicians, nurses, and health professionals, and the longer the career and work hours were, and the higher the number of reported events was, the higher the patient safety awareness, although the overall patient safety awareness was low. In this study, the patient safety awareness was low possibly because the hospitals that were included in this study were not fully prepared for medical-institution accreditation as they were general hospitals, even though patient safety awareness is
276
Y.K. Boo et al.
increasing in the society due to the accreditation of medical institutions. Thus, hospital climate improvement is needed to ensure patient safety. The results of this study showed that an increase in hospital climate improvement and in the immediate supervisor’s patient safety awareness had positive influences on patient safety, whereas a decrease in the hospital management’s patient safety awareness had a negative influence on patient safety. Thus, it was revealed that the leadership of the hospital management is needed to establish a patient safety culture, as has been shown in other studies, both local and abroad [10, 11, 12, 13]. To prevent patient-safety-related accidents, a patient safety management system that can remove and prevent risk factors in advance should be established. The patient safety management system should be linked to the hospital information system to ensure its expeditious and effective operation [14]. Enhanced prescription delivery and clinical-decision support systems and clinicalpractice guidelines should be interlinked within the hospital information system. In addition, a voluntary medical-error reporting and feedback system where the confidentiality is ensured is required for the effective prevention of near-miss and medical errors. To create a patient safety culture at the soonest possible time, a climate that can ensure patient safety, safety education of the staff members, a change in the level of patient safety awareness, establishment of patient safety and medical-error reporting systems, cooperation between work units, and open communication are required. By doing all these, the quality of medical services could be improved.
References 1. Blendon, R.J., DesRoches, C.M., Brodie, M., Benson, J., Rosen, A., Schneider, E., Altman, D., Zapert, K., Herman, M., Steffenson, A.: Views of Practicing Physicians and the Public on Medical Errors. N Engl. J. Med. 347, 1933–1940 (2002) 2. Institute of Medicine: To Err Is Human: Building a safer Health System. National Academy Press, Washington, DC (2000) 3. Joint Commission on Accreditation of Health Organizations, http://www.jointcommission.org/ 4. Radecki, R.P., Sittig, D.F.: Application of electronic health records to the Joint Commission’s 2011 National Patient Safety Goals. JAMA 306(1), 92–93 (2011) 5. Park, J.H.: Understanding of the new Korea Healthcare Accreditation System. J. Korean Med. Assoc. 54(2), 142–145 (2011) 6. Agency for Healthcare Research and Quality (AHRQ) Hospital Survey on Patient Safety, http://www.ahrq.gov/qual/patientsafetyculture/ hospscanform.pdf 7. Kim, M.R.: Concept Analysis of Patient Safety. J. Korean Acad. Nurs. 41(1), 1–8 (2011) 8. Nam, M.H., Kang, S.H.: The Cognition Level on the Patient Safety and Safe Nursing Activities among Hospital Nurses in Busan. J. Health and Social Sciences 28, 197–221 (2010) 9. Kim, J.E., Kang, M.A., An, K.E., Sung, Y.H.: A Survey of Nurses’ Perception of Patient Safety Related to Hospital Culture and Reports of Medical Errors. Clinical Nursing Research 13(3), 169–179 (2007)
Identifying Organizational Culture for Patient Safety
277
10. Sammer, C.E.: What is patient safety culture? A review of the literature. J. Nurs. Scholarsh 42(2), 156–165 (2010) 11. Ruchlin, H.S., Dubbs, N.L., Callahan, M.A.: The Role of Leadership in Instilling a Culture of Safety: Lessons from the Literature. J. Healthc Manag. 49(1), 47–58 (2004) 12. Ginsburg, L.R.: The Relationship between Organizational Leadership for Safety and Learning from Patient Safety Events. Health Serv. Res. 45(3), 607–632 (2010) 13. Firth-Cozen, J., Mowbray, D.: Leadership and the Quality of Care. Quality in Health Care 10(suppl. II), ii3–ii7 (2001) 14. Bates, D.W., Gawande, A.A.: Improving safety with information technology. N. Engl. J. Med. 348(25), 2526–2534 (2003)
The Effect of a Support Program for Family Caregivers of Dementia on Stress Hormones* Hyoshin Kim Dept. of Nursing, Chungwoon University, San29, Namjang-ri, Hongseong-eup, Hongseong-gun, Chungnam, 350-701, Republic of Korea [email protected]
Abstract. This study aims to evaluate the effect of a support program for family caregivers of seniors with dementia on stress hormones (Epinephrine and Norepinephrine). A total of 44 family caregivers participated in the experiment: 25 were in the experimental group, and 19 were in the control group. The experimental group took part in a support program once per week for 8 weeks, with 2-3 hours per session. Blood Epinephrine and Norepinephrine of both groups were measured before and after the program. As a result, Norepinephrine increased significantly less in the experimental group than in the control group after the support program. This results shows that the support program had the positive effect of reducing the increase of Norepinephrine. In conclusion, it was found that the support program can be used as a safe and effective strategy for managing stress control for the family caregivers of seniors with dementia. Keywords: Dementia, Family Caregiver, Stress Hormone, Epinephrine, Norepinephrine.
1
Introduction
1.1
Necessity
The elderly population is increasing rapidly due to a number of factors, including the extension of the average life span, economic and medical development, and the declining birthrate. By 2026, Korea is expected to be a super-aged society, with seniors representing 20.8% of the population [1]. Corresponding to this, the prevalence of dementia, one of the typical diseases experienced by the elderly, has increased sharply. The prevalence rate is reported to be between 6.3 ~ 13% in many epidemiologic surveys in Korea [2]. Family caregivers for a person with dementia suffer from chronic psychological problems, including depression, anxiety and tension. In certain cases, they have to be treated for depression, and visit hospitals or are administered with medicines due to *
This study was supported by a research grant of the Life Insurance Philanthropy Foundation.
The Effect of a SP for Family Caregivers of Dementia on Stress Hormones
279
their physical sacrifices, such as skipping meals or working too hard to care for their patients.[3]. Economic burdens are also frequently increased, as a family member may have to quit a job or withdraw from social activity to support their seniors with dementia, or to pay the related care service cost [4]. As dementia, unlike other diseases, is a cognitive and behavioral disorder that requires 24-hour protection and care, family caregivers experience huge physical and mental stress [5]. In particular, the frequent problematic behavior and low social support for elderly sufferers of dementia increase the burden and depression of caregivers [6][7]. The support program is designed to reduce the burden to family caregivers, based on the knowledge and demand of dementia by caregivers. Previous studies on family caregivers developed programs mainly based on a literature survey [8][9][10]. But this study attempted to develop a customized support program through interviews, a literature survey and the participation of professionals. In previous studies, the effect of these programs were measured mainly as social and psychological variables, but rarely was any attempt made to measure stress hormones. Thus, in this study Epinephrine and Norepinephrine were measured. 1.2
Purpose of the Study
The purpose of this study is to develop a support program and to verify its effects on Epinephrine and Norepinephrine of family caregivers of senior citizens with dementia.
2
Methodology
2.1
Design
This study is quasi-experimental research employing the nonequivalent control group pretest-posttest design. The support program (SP) is an independent variable, and the Epinephrine and Norepinephrine of family caregivers are dependent variables. It was developed using interview, literature survey, experience of caregivers, and participation of professional groups. The experimental group received the support program once per week for 8 weeks, with one 2-3 hour session per week. The control group did not receive any intervention. 2.2
Subjects
The experimental group consisted of family caregivers who consented to join the support program and to take the 8-week support program and blood test before and after the program. The control group was made up of family caregivers who also agreed to join the experiment and took a blood test before and after the program. Both group had blood tests at the same time and period before and after the test.
280
H.S. Kim
The homogeneity test of both groups showed no difference in general characteristics (Table 1). Before the test, there was no difference between the two groups in the Epinephrine and Norepinephrine levels (Table 2). 2.3
Instruments
2.3.1 The Blood Test of Epinephrine and Norepinephrine The blood test was carried out at 9:30~10:30 a.m. in a comfortable state after a 12hour fast from 7:00 p.m. of the previous day, as scheduled in advance. A trained examiner and a blood-gathering nurse controlled the process and the subjects. For the blood Epinephrine and Norepinephrine content, the High Performance Liquid Chromatography (HPLC) method was applied, and the sample was analyzed using an HPLC-Acclaim device. 2.4
Resource Collection Method and Research Procedure
Family caregivers were recruited with the assistance of workers in dementia support centers, day centers for the elderly, and domiciliary care facilities in Seoul, as well as through internet and leaflet advertisement. The program content was constructed by interviewing 44 family caregivers of senior citizens with dementia. This was followed by the advice of a professional group. The program included information on the concept and symptoms of dementia, diseases related to dementia, home care methods, health management of family caregivers of senior citizens with dementia, exercise therapy, art therapy, resolution of family conflict caused by senior citizens with dementia, skills for communication with senior citizens with dementia, handling problematic behavior by senior citizens with dementia, stress control for family caregivers, information about local community resources, and an information exchange among family caregivers of senior citizens with dementia. Subsequently, the program was given to 26 Experimental group members for 8 times with one session per week, and 2-3 hours per session. The questionnaire survey and blood test were performed in the preliminary meeting, and the program was carried out 8 times. The total number in the experimental group became 25 as one dropped out during the test. The control group was 19 after excluding 3 who did not join the post test. The test was carried out from July to November of 2009. 2.5
Resource Analysis
A total of 44 cases were analyzed. The collected resources were analyzed using SPSS 18.0k for Windows. (1) The general characteristics of family caregivers were analyzed by calculating frequency and percentage. (2) The general characteristics of family caregivers and the homogeneity test of Epinephrine and Norepinephrine were analyzed using the χ2-test.
The Effect of a SP for Family Caregivers of Dementia on Stress Hormones
281
(3) In the preliminary homogeneity test of family caregivers, all general characteristics such as gender, age, residence, marital status, number of family members, number of children, religion, education level, occupation, monthly income, relation with the senior with dementia, years caring for the senior with dementia, and symptoms showed no difference between the experimental group and the control group. Therefore, the paired t-test was employed within the group before and after the test, and the unpaired t-test was used to analyze the two groups before and after the test.
3
Research Result
3.1
Homogeneity Test
3.1.1 General Characteristics of Family Caregivers of Seniors with Dementia Table 1 shows the preliminary general characteristics and homogeneity of subjects in the experimental group and control group. In all general characteristics, such as gender, age, residence, marital status, number of family members, number of children, religion, education level, occupation, monthly income, relation with the demented elderly, caring years of the demented elderly, and symptoms, there was no difference shown between the experimental group and the control group (Table 1). Table 1. General characteristics and homogeneity of subjects in the experimental and control groups
(N=44) General characteristics
Category
Gender
Age(yr)
Place of Residence Marital status Number of family members (living together) Number children Member Religion
of of
a
Female Male Below 40 40-49 50-59 60-69 70-79 80 and above Seoul Gyeonggi Unmarried Married 1 2-4 5-6 0 1-3 4 or more Yes No
Years caring for the demented senior Does the caregiver have symptoms
Elementary school or below Middle or High school College or above Yes No Below 100 100~Below 200 200~Below 300 300 or higher Daughter-in-law Son Daughter Spouse Sister 5 or less 6-15 16 or more Yes No
3.1.2 Preliminary Measurement of Epinephrine and Norepinephrine of the Experimental Group and the Control Group In the preliminary measurement of Epinephrine and Norepinephrine, there was no significant difference between the experimental group and the control group [Table 2]. Table 2. Homogeneity of Epinephrine and Norepinephrine of subjects in the experimental and control group
(N=44) Exp.
Cont.
t
p
Epinephrine (pg/mL)
27.84±15.48
29.60±9.00
0.47
0.661
Norepinephrine (pg/mL)
303.01±108.92
349.24±151.07
1.18
0.245
Exp. : Experimental group (n=25) Cont. : Control group (n=19) Data represents mean±S.D.
3.2
Hypothesis Testing
3.2.1 Epinephrine Epinephrine in the experimental group was 27.84±15.48 pg/mL before the test and became 28.66±15.01 pg/mL after the test, which was not a statistically significant difference. Epinephrine in the control group was reduced from 29.60±9.00 pg/mL to 28.87±13.90 pg/mL, which was also not significant. After the test, Epinephrine was
The Effect of a SP for Family Caregivers of Dementia on Stress Hormones
283
increased by 0.83±17.32 pg/mL in the experimental group and decreased by 0.73±15.96 pg/mL in the control group. Nonetheless, both tests showed no significant difference between the two groups (Table 3). Table 3. Epinephrine and Norepinephrine before and after support program in the experimental and control groups
(N=44) Before
After
t
p
Difference (After-Before)
t
p
0.06
0.953
2.63
0.012
Epinephrine (pg/mL) Exp.
27.84±15.48
28.66±15.01
0.24
0.816
0.83±17.32
Cont.
29.60±9.00
28.87±13.90
0.20
0.844
-0.73±15.96
Norepinephrine (pg/mL) Exp.
303.01±108.92
337.76±98.46
1.77
0.090
34.76±98.21
Cont.
349.24±151.07
514.72±337.81
2.24
0.038
165.50±321.32
Exp. : Experimental group (n=25) Cont. : Control group (n=19) Data represents mean±S.D.
3.2.2 Norepinephrine Norepinephrine in the experimental group was 303.01±108.92 pg/mL before the test and became 337.76±98.46 pg/mL after the test, which was not a statistically significant difference. Norepinephrine in the control group was increased from 349.24±151.07 pg/mL to 514.72±337.81 pg/mL, which was statistically significant (p=0.038). After the test, Norepinephrine was increased by 34.76±98.21 pg/mL in the experimental group, and by 165.50±321.32 pg/mL in the control group, which showed significant differences between the two groups (p=0.012) (Table 3) (Fig. 1).
Fig. 1. Stress hormone (Norepinephrine) changes before and after support program in the experimental and control groups
284
4
H.S. Kim
Discussion
Stress is a generic term for an individual's physiological, psychological and behavioral reactions to control and to adapt to internal and external pressures. Basically, a stress reaction is an adjustment reaction for survival. If this reaction is over-activated, it can cause anxiety and depression. If it continues chronically, it can damage the body [11]. According to Wellis-Federeman et al. (1995) [12], stress is a non-specific response of the body against a stressor, and may cause a negative psychological reaction such as anxiety when environmental stimulation (the stressor) is recognized as threatening. Stress can cause stress reaction through the central nervous system, the autonomic nervous system, and the hypothalamus - pituitary gland - adrenal that increases stress hormones. There are three catecholamines - Epinephrine, Norepinephrine, and Dopamine - secreted from the adrenal medulla, of which Epinephrine and Norepinephrine are present in the greatest amount. As dementia is an irreversible disease that chronically progresses after its occurrence, it imposes a very serious psychological, physical and economical burden to the family. Kim and Jeon (1996) [13] defined dementia as a disease that devastates whole families, and as the disease of the century. Park (1997) [14] enumerates five common burdens to the families of seniors with dementia: first, the restriction of social activity; second, an increased psychological burden; third, an increased economic burden; fourth, an increased burden on the health of caregivers; and fifth, the aggravation of conflict between family members. Among the factors that reduce the burden of family caregivers, the importance of support has been proven in previous studies. Support helps to improve individual psychological adaptation, to promote problem-solving ability and to achieve stability [15]. The support program for family caregivers described in this study is a program designed to reduce stress to family caregivers and to improve adaptation. In this study, Norepinephrine levels in the experimental group showed a significantly smaller increase than in the control group, when measurements taken before and after the test were compared. This demonstrates that the support program contributed to reducing the increase of Norepinephrine, a factor of stress hormone, in the experimental group, as has been shown in similar studies [16][17]. In conclusion, the support program for family caregivers had a positive effect of reducing the increase of Norepinephrine, and was proven to be safe and suitable to contribute to stress control. Therefore, if the support program is applied to family caregivers of the demented elderly, it seems to have a positive effect through stress hormone control.
References 1. Statistics Korea Estimating future population (2006), http://sgis.kostat.go.kr 2. Cho, M., et al.: The Prevalence and Risk Factors of Dementia in the Korean Elderly. Ministry of Health and Welfare (2008)
The Effect of a SP for Family Caregivers of Dementia on Stress Hormones
285
3. Parks, M., Novielli, K.D.: A practical guide to caring for caregivers. American Family Physician 62, 2613–2620 (2000) 4. Song, M.I., Choe, G.G.: The Decision Factor on Primary Care-giver’s Burden for Senile Dementia Elderly. Journal of Welfare for the Aged. 37, 131–160 (2007) 5. Lee, S.J., Park, H.A., Cho, N.O., Oh, J.J.: Study on the Effectiveness of Caregiver Education Program on the Home Care of Senile Dementia Patients. Journal of Academy of Nursing 25, 45–60 (1995) 6. Donalson, C., Tarrier, N., Burns, A.: The Impact of the Symptoms of Demetia on Caregivers. British Journal of Psychiatry 170, 62–68 (1997) 7. Lee, Y.M., Yoo, I.Y.: Care Burden, Depression and Social Welfare Service Utilization among Family Caregiver for Demented Elderly. Journal of Korean Gerontological Society 25(3), 111–121 (2005) 8. Kim, S.Y., Chang, J.K.: A Study on Education Programs for Senile Dementia Family. Living Science Research Institute 16, 13–39 (2001) 9. Han, H.S.: The Development and Its Effectiveness of the Group Program on Empowerment for the Senile Dementia Caregivers, Master’s thesis of Honam university, Korea (2004) 10. Lee, E.H.: Effects and Development of Cognitive-behavioral Group Intervention in Reducing Psychological Distress in Primary Caregivers of Persons with Dementia: Caregiving Burden, Depression, and Anxiety, Master’s thesis of Chonnam national university, Korea (2005) 11. Yu, B.H., Jong, M.W.: Neural Circuits Mediating Stress 9(1), 81–92 (2001) 12. Wellis-Federman, C.V., Stuart, E.M., Deckro, J.P., Mandle, C.L., Baim, M., Medich, C.: The Mind-body Connections: The Psychophysiology of Many Traditional Nursing Intervention. Clinical Nurse Specialist 9(1), 59–66 (1995) 13. Kim, T.H., Jun, G.Y.: A Study on Caregiving Situations and Resources for Adjustment among Family Caregivers of the Elderly with Dementia. Journal of the Korean Home Economics Association 34(4), 145–160 (1996) 14. Park, T.R.: Senile Dementia and Support Countermeasure for the Family. Journal of Social Science Research 4(2), 251–269 (1997) 15. Suh, M.H., Oh, K.S.: A Study of well-being in Caregivers Caring for Chronically Ill Family Members. Journal of Academy of Nursing 23, 467–486 (1993) 16. Yun, S.Y., Choi, B.J.: Effect of Horticultural Therapy on the Stress and Serum Cortisol of Demented Elders. Korean Journal of Horticultural Science and Technology 28(5), 891–894 (2010) 17. Yoo, K.H.: The Effects of Massage on Stress Hormone in Prematuer Infants. Journal of Korean Academy of Child Health Nursing 11(1), 125–131 (2005)
The Effect of Educational Game on the Intrinsic Motivation by Learner’s Traits: Focus on Personality Types Hyung-sung Park, Jung-hwan Park, and Young-Tae Kim Korea National University of Education, Jeju National University, Wonkwang University [email protected], [email protected], [email protected]
Abstract. The purpose of this study was to identify the effect of the action levels of educational game on the intrinsic motivation of the learners by personality types. In this study the action levels of the educational game and the learners’ personality types are main variables to be explored. The result of this research, there are meaningful difference in intrinsic motivation between learners who use games on the higher level of action and learners on the lower level of action in game-based learning. And also, the intrinsic motivation of extrovert learners' improved more than that of introvert learners' in game-based learning. It is expected that the results of this study would suggest how to utilize educational games in teaching and learning. It is considered that the result can be enough information for learners, parents of students and teachers who try to use and develop educational games. For example, it can be an idea to construct menus considering characteristics of learners in various steps of designing and developing educational games. And it can help to improve learners' intrinsic motivation especially if games are developed, distinguishing games for extrovert learners from those for introvert learners. Keywords: Educational Game, Personality Type, Intrinsic Motivation.
The Effect of Educational Game on the Intrinsic Motivation by Learner’s Traits
287
learning contents and gaming situations occur simultaneously while balancing fun and learning. This balancing act between enjoyment and educational value continues to be a hot topic in discussions about developing educational games (Kim, Park, & Baek, 2007). With the attention of the games deeply, educators begin to focus on the educational potentials of games. First game as a media, there are sound music, vivid animations, intricate plots the characters with distinct personality, so the involvement of game can be called a kind of enjoyment of art. Second, the player through various designed circumstances, and to complete the task, in the process of playing, their ability of resolving problems and corporation promote increasingly. Specially, the MMORPGs which can incorporate some education content, players can obtained in the course of the geography, history, culture, daily life, animals and plants, tourism, a few interesting physical and chemical knowledge, because such knowledge is relatively easy to express by the scenes, maps, tasks, skills and other elements of the game. Regardless of how the game was demonized, if we can objectively understand the positive effects of games, and make the best use of it to promote learning, it should have great advantages and potential. Many educators and scholars at home and abroad have initiated research and development of the educational game based on the pros and cons, they try to explore application model, and already have achieved some positive results (Dickey, 2007; Kafai, 2006; Killi, 2005; Prensky, 2001; Squire & Jenkins, 2003; Squire, 2003). Like this, much effort to use educational games for students’ learning is being made actively at educational filed. Roles of learners and instructors are undergoing considerable changes in current view of the constructivism, and the changes in the teaching-learning method are dynamic, correlating with changes in the whole contemporary socialites. For this reason, this research was started. In this study the action levels of the educational game and the learners’ personality types are main variables to be explored. It is expected that the results of this study would suggest how to utilize educational games in teaching and learning.
2
Educational Game, Motivation and Learning by Doing
Games provide an engaging and immersive learning environment which is different from traditional learning. Many digital video games create a sense of sensory and physical immersion by using technologies resulting in one’s feeling “inside” an environment. These environments are psychologically immersive in ways that the world-to-the-desktop interface is not (Oblinger, 2006). In the environment game create which interactive fantasy, motivated, the player can play, practice actively. game-based learning environment proved to increase participants' interest and motivate participants in exploring and engaging in the learning activities(Kuo, 2007). Digital game as a learning environment is a powerful pedagogy that allows for practice, operate, rehearse immersive in safe environment. Squire (2003) examines the history of games in educational research, and argues that the cognitive potential of games have been largely ignored by educators. Contemporary developments
288
H.-s. Park, J.-h. Park, and Y.-T. Kim
in gaming, particularly interactive stories, digital authoring tools, and collaborative worlds, suggest powerful new opportunities for educational media. These elements we argued will contribute to a good learning environment. Games provide a rich learning environment in which students involved in and can operate by themselves with mentoring and apprenticeships in communities of practice. Begoña Gros (2003) argued game with tremendous potential for reaching, motivating, and fully involving learners. Through the gaming, students can learn by doing and work with others through games, game can support children’s learning inside and out of school. Learning by doing, this is great for Digital game-based learning, because “it turns out that doing is something that computer games are especially good at- they allow us to interact with them” (Prensky, 2001, p.158). Learn-by-doing is permeated into digital games, drill and practices, exploring, discovery, and problem solving are forms of doing required in games, in Game-based learning, students pursue a goal by practicing target skills and using relevant content knowledge to help them achieve their goal. The essential components of games: the goal, mission, storytelling scenario, operation, resources, feedback and ubiquitous coach are good at supporting learning by doing in video games. Most often, the more fun a game is the less effective it is as a learning tool and vice versa. In fact, the quintessential combination of fun and learning seems almost impossible to achieve. This ideal scenario is an equal mixture of both learning and game content. It’s not easy to determine the intrinsic balance where both learning and gaming happen simultaneously. This has continued to be a hot issue in the development of educational games. In the constructivist point of view on cognition, knowledge is generated from experience with tasks in authentic contexts rather than from isolated activities like memorizing facts and practicing skills separately (killi, 2005). Skills and knowledge are thus best acquired within a complex problem-rich context such as those presented by educational game, where a variety of problems can vary greatly in complexity and difficulty. As far as constructivists are concerned, using games for learning can play an important role in knowledge construction and composition and as an environment for problem solving skills.
3
Characteristics of Personality Type
The Myers-Briggs Type Indicator (MBTI), developed by Isabel Briggs Myers and her mother, Katharine Cook Briggs, is a well known and widely used personality inventory based on the psychological theories of Carl Gustav Jung. Jung's theory of psychological types assumes that much apparently random human behavior is really quite orderly and consistent and reflects the different ways persons take in information and make decisions. The Myers-Briggs Type Indicator is a self-report instrument based on Jung's theory of psychological types. The instrument returns the respondent's preferences on each of the four dimensions extraversion/introversion, sensing/intuition, thinking/feeling and judging/ perceiving (Rosati, 1998). The classification scheme is based on four dimensions or scales of preferences: (1) extraversion-introversion
The Effect of Educational Game on the Intrinsic Motivation by Learner’s Traits
289
(EI)-whether to direct perception outward toward the environment or toward the inner world of ideas; (2) sensing or intuition (SN)-whether the person perceives things directly through one or a combination of the five senses (sensory), j or whether he/she tends to add ideas or associations to the incoming information (intuitive); (3) thinking or feeling (TF)-whether the person prefers to make judgments (or decisions) based on an impersonal or true-false basis (thinking), or based on personal values (feeling); and (4) judging-perceiving (JP), which “reflects whether the person relies primarily upon a judging process [thinking or feeling] or upon a perceptive process [sensing or intuition] in her/his dealings with the outer world” (MYERS, 1967).
4
Research Methods
4.1
Main Research Hypothesis
Hypothesis I: There will be a meaningful difference in intrinsic motivation between learners who use games on the higher level of action and learners on the lower level of action in game-based learning. Hypothesis II: The intrinsic motivation of extrovert learners' will be improved more than that of introvert learners' in game-based learning. Hypothesis III: There will be interaction effects between the level of action and the learner's personality type on learners' intrinsic motivation in game-based learning. 4.2
Participants, Experimental Design and Questionnaire
The learners were divided into four groups, considering the educational games' levels of action and the learners' personality types and were treated with the experiment process in the same condition for six weeks. The educational games were selected through operational definitions and expert examination. Those had same subject matters but different levels of action. The learners' personalities were evaluated by extraversion (E) or introversion (I) test of MBTI. Intrinsic Motivation Assessment Scale which was made by Harter’s(1981) assessment scale to have examined middle/high school students' intrinsic motivation, was used to check the learners' changes in their intrinsic motivation. That scale was reconstructed for questions suitable for educational games through expert examination, the reliability of the questions was satisfying as Cronbach's α = .84.
5
Results and Conclusion
The followings are results of the experiment. First, as shown Table 1, according to the pre-test result, high action level group and low action level group could make sure that there is no significant between experimental group and control group as statistical by action levels of educational game (t=1.545, p>.05). The result of the post-test of intrinsic motivation indicates that there is a 0.38 points difference between learners
290
H.-s. Park, J.-h. Park, and Y.-T. Kim
who use games on the higher level of acting(M=3.68, SD=.385, t=5.649, p<.001) and learners on the lower level of action(M=3.11, SD=.649, t=5.649, p<.001). Therefore There will be a meaningful difference in intrinsic motivation between learners who use games on the higher level of action and learners on the lower level of action in game-based learning has been proved affirmative. To identify the experimental treatments effect, we calculated the effect size. As the result, effect size was Cohen's d= 1.07. It means that was effective rather high action level games than low action level games. Table 1. Intrinsic motivation by action levels Test Pre Post
Action Level
N
M
SD
Low High Low High
59 56 59 56
2.85 3.04 3.11 3.68
.78 .56 .65 .38
t
Sig.
1.545
0.125
5.649
0.000
Second, as shown Table 2, according to the pre-test result introvert learner group and extrovert learner group could make sure that there is no significant between experimental group and control group as statistical by learner's personality type (t=1.041, p>.05). Extrovert learners (M=3.21, SD=.47, t=2.924 p<.001) according to MBTI have got 0.26 points more than Introvert learners (M=3.47, SD=.46, t=2.924, p<.001). Therefore The intrinsic motivation of extrovert learners' will be improved more than that of introvert learners' in game-based learning has been proved affirmative. To identify the experimental treatments effect, we calculated the effect size. As the result, effect size was Cohen's d= 0.56. It means that was effective rather extrovert learners’ than introvert learner’s in the intrinsic motivation. Table 2. Intrinsic motivation by personality types
Test
Pre Post
Persona lity types
N
M
SD
E I E I
67 48 67 48
3.20 3.08 3.21 3.47
.69 .52 .47 .46
t
Sig.
1.041
0.300
2.924
0.004
Third, as shown Table 3, the result of analyzing the interaction effects indicates that there are meaningful figures on each of the main effects (F=16.22, p<.001), the level of action (F=35.78, p<.001) and the type of the personality (F=10.30, p<.001) but there isn't on the interaction effect (F=0.76) between the level of action and the type of the personality. Therefore There will be interaction effects between the level of action and the learner's personality type on learners' intrinsic motivation in game-based learning has been proved negative.
The Effect of Educational Game on the Intrinsic Motivation by Learner’s Traits
291
Table 3. Result of interaction effect (*** p<.001) Source Main Effect Action Levels Personality Types Interaction Error Total
df 3 1
SS 7.99 5.88
MS 2.66 5.88
F 16.22*** 35.78***
1
1.69
1.69
10.30***
1 111 115
0.12 18.22 1298.46
0.12 0.16
0.76
The above is the result of the research about the level of action of educational games, the type of the personality, and the intrinsic motivation. It is considered that the result can be enough information for learners, parents of students and teachers who try to use and develop educational games. For example, it can be an idea to construct menus considering characteristics of learners in various steps of designing and developing educational games. And it can help to improve learners' intrinsic motivation especially if games are developed, distinguishing games for extrovert learners from those for introvert learners. The main result of the research is that it is possible to stimulate learners' motivation enough by only games.
References Amory, A.: Building an Educational Adventure Game: Theory, Design, and Lessons. Journal of Interactive Learning Research 12(2), 249–263 (2001) Barab, S., Thomas, M., Dodge, T., Carteaux, R., Tuzun, H.: Making learning fun: Quest Atlantis, a game without guns. Educational Technology Research and Development 53(1), 86–107 (2005) Bos, N.: What do game designers know about scaffolding? Borrowing SimCity design principles for education (2001), http://playspace.concord.org/Documents/ Learning%20from%20SIMCITY.pdf (retrieved April 5, 2010) Bottino, R.M., Ferlino, L., Ott, M., Tavella, M.: Developing strategic and reasoning abilities with computer games at primary school level. Computers & Education 49, 1272–1286 (2007) Dickey, M.D.: Game design and learning: a conjectural analysis of how massively multiple online role-playing games (MMORPGs) foster intrinsic motivation. Education Tech. Research Dev. 55, 253–273 (2007) Eck, R.V.: Digital Game-Based Learning: It’s Not Just the Digital Natives Who Are Restless. Educause Review 41(2), 17–30 (2006) Freitas, S. D.: Learning in Immersive worlds: A review of game-based learning JISC report (2007), http://jisc.ad.uk/media/documents/programmes/elearning innovation/gamingreport_v3.pdf (retrieved April 23, 2010) Garris, R., Ahlers, R., Driskell, J.E.: Games, motivation, and learning: A research and practice model. Simulation and Gaming 33(4), 441–467 (2002) Gee, J.P.: Learning by Design: Good Video Games as Learning Machines. Interactive Educational Multimedia 8, 15–23 (2004) Gros, B.: Digital games in education: the design of game-based learning environments. Journal of Research on Technology in Education 40(1), 23–38 (2007)
292
H.-s. Park, J.-h. Park, and Y.-T. Kim
Gros, B.: The impact of digital games in education (2003), http://www.firstmonday. org/issues/issue8_7/xyzgros/#author (retrieved April 7, 2011) Harter, S.: A new self-report scale of intrinsic versus extrinsic orientation in the classroom: Motivational and informational components. Developmental Psychology 17, 300–312 (1981) Kafai, Y.B.: Playing and making games for learning: Instructionist and constructionist perspectives for game studies. Games and Culture 1(1), 36–40 (2006) Killi, K.: Digital game-based learning: Towards an experiential gaming mode. Internet and Higher Education 8, 13–24 (2005) Kim, B.K., Park, H.S., Baek, Y.K.: Metacognitive Strategies for Successful Game-Based Learning. In: 5th International Conference for Media and Education, Korea, pp. 125–131 (2007) Kirriemuir, J., McFarlane, A.: Literature Review in Games and Learning (2005), http://www.futurelab.org.uk/resources/documents/lit_reviews/ Games_Review.pdf (retrieved April 1, 2008) Kuo, M.J.: How does an online game based learning environment promote students’ intrinsic motivation for learning natural science and how does it affect their. Paper presented at The First IEEE International Workshop on Digital Game and Intelligent Toy Enhanced Learning (2007) McFarlane, A., Sparrowhawk, A., Heald, Y.: Report on The educational use of games: an Exploration by TEEM of the Contribution which Games Can Make to the Education Process. Teachers Evaluating Educational Multimedia report (2002), http://www.teem.org.uk/publications/teem_gamesined_full.pdf (retrieved April 11, 2011) Moshirnia, A.: The Educational Potential of Modified Video Games. Issues in Informing Science and Information Technology 4, 511–521 (2007) Myers, I.B.: The Myers-Briggs Type Indicator Manual. Princeton, NJ (1967); Educational Testing Service Oblinger, D.G.: Games and Learning: Digital games have the potential to bring play back to the learning experience. Educause Quarterly 29(3), 5–7 (2006) Prensky, M.: Digital Game-Based Learning. McGraw-Hill (2001) Robertson, J., Good, J.: Children’s narrative development through computer game authoring. Tech. Trends 49, 43–59 (2007) Rosati, P.: Academic Progress of Canadian Engineering Students in terms of MBTI Personality Type. Academic Progress of Canadian Engineering Students 14(5), 322–327 (1998) Sandford, R., Ulicsak, M., Facer, K., Rudd, T.: Teaching with Games: Using commercial off the-shelf computer games in formal education (2006), http://www.futurelab.org.uk/resources/documents/project_ reports/teaching_with_games/TWG_report.pdf (retrieved April 11, 2010) Sandford, R.: Virtual Worlds: more than a game (2007), http://futurelab.org. uk/resources/publications_reports_articles/web_articles/ Web_Article477 (retrieved April 12, 2010) Shaffer, D.W.: How computer games help children learn. Palgrave Macmillan (2006) Squire, K.: Video games in education. International Journal of Intelligent Simulations and Gaming 2(1), 49–62 (2003) Squire, K., Jenkins, H.: Harnessing the power of games in education. Insight 3, 5–33 (2003) Squire, K., Giovanetto, L., Devane, B., Durga, S.: Building a self-organizing game-based learning environment. Tech.Trends 49, 34–42 (2007)
Detection and Recovery for Kidnapped-Robot Problem Using Measurement Entropy Chuho Yi and Byung-Uk Choi Department of Electronics and Computer Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul, 133-791 Korea {d1uck,buchoi}@hanyang.ac.kr
Abstract. In robotics, the kidnapped robot problem commonly refers to a situation where an autonomous robot in operation is carried to an arbitrary location[1]. This is a very serious issue, and because the robot only computes its position in accordance with mathematical models, it is difficult to determine whether or not its position is being accurately reported. In this paper, to solve the kidnapped-robot problem, we suggest a method of automatic detection and recovery, using the entropy extracted from measurement likelihoods in our own technique of semantic localization. We verify the usefulness of the proposed procedure via repeated indoor localization experiments.
1
Introduction
In the kidnapped-robot problem, a robot in unable to estimate its own position via the localization process [1][2]. To solve this problem, it must first be detected, and a recovery process must then be initiated. Many robot localization methods have been studied over the past decade, using grid-based maps [3], feature-based maps [4][5], topological maps [6], semantic maps [7], or adaptively selecting various types of maps [8]. In their research on the probabilistic Bayesian model, Thrun et al. applied Bayes’ rule to robot localization [9]. A topological map is represented in semantics using ontology, and relies on semantics for navigational inferences [10] [11]. Lee et al. performed localization and solved the kidnapped-robot problem using sonar sensors and a camera to create a feature map [13]. In this paper, we apply the particle distribution of the robot to calculate the entropy used to detect the problem, and the other particle set is added to recover from it. In [12], we used object recognition and metric measurements with a single camera for semantic localization. In the present paper, the amount of measurement information is less than that of traditional techniques. Our localization method uses a particle filter, and the probability distribution of the robot locations is computed for each particle. If there is not enough measurement information, divergence could occur. To solve this problem, we calculate the entropy of measurement between observed metric data and existing map data. This value is used to estimate the robot’s current position from the similarity of the observed data and the map data. In other words, the entropy is low (due to the high T.-h. Kim et al. (Eds.): GDC 2011, CCIS 261, pp. 293–299, 2011. c Springer-Verlag Berlin Heidelberg 2011
Fig. 1. Example of a topological-semantic distance map. (a) Semantic representation comprising a set of spatial object contexts and spatial robot contexts. (b) Specific example of a topological-semantic distance map comprising a set of spatial object contexts and spatial robot contexts. (c) Relationship between the topological-semantic distance map and the real environment.
measurement likelihoods) when the observations are similar to the map data; otherwise, the entropy is high. When the robot is subjected to the kidnappedrobot problem, this value increases, and continuous measurement detects the increase. In this paper, we provide a brief description of semantic localization, and evaluate the validity of the proposed method via indoor experiments.
2
Semantic Localization
In [12], we proposed a topological-semantic distance map consisting of spatial object and spatial robot contexts. In a topological-semantic distance map, a node is one of the components of a general topological map, playing the role of a standard and containing information on spatial object contexts. In addition, the spatial robot contexts in our semantic representation can be used to determine an approximate distance and bearing from one assigned node to another. The node-to-node (n-n) distance context is an approximate qualitative distance and the n-n bearing context is an approximate qualitative bearing. Therefore, global localization is made possible by the objects in a topological-semantic distance map, as well as the spatial object contexts. In addition, localization is processed more specifically and locally, based on the information gathered from global localization, by deriving the probability distribution of possible robot locations around a node. Figure 1 shows a semantic representation consisting of observed objects and their respective spatial symbols. Here, a spatial context includes distance, bearing, and relationship contexts. The robot-to-object (r-o) distance context, denoted by sr , is the distance between the object and the robot. Each distance context is represented by a set of distance symbols; that is, sr = {nearby, near, far}. The robot-to-object (r-o) bearing context, denoted by sω = { front, left front, left,
Detection and Recovery for Kidnapped-Robot Problem
295
left rear, rear, right rear, right, right front}, is the bearing of the object relative to the robot. The object-to-object (o-o) relationship context, denoted by sζ = { left far, left near, left nearby, right nearby, right near, right far}, is the relationship between objects. Probabilistic robot location represents a set of beliefs as a conditional probability distribution. We denote belief in a state xt by Bel(xt ), which satisfies the following:
[y]
Bel(xt ) = η · wt
Y y
[y]
[y]
p(xt |xt−1 , ct , ut, m)Bel(xt−1 )
(1)
[y]
where each particle xt (1 ≤ y ≤ Y ) is a concrete instantiation of the state at [y] time t. The importance wt is the probability of a measurement at the given particle, and is obtained from [y] target distribution = p(st , ot , ot |xt , m) (2) proposal distribution We represent the features extracted from the image by z, and the map is denoted by m. We have already discussed the type of map in question, namely the topological-semantic distance map. Robot location is denoted by x = [v(i) e(i,j) ρ(i,j) θ(i,j) ]T , where v(i) and e(i,j) are the topological node and edge, respectively. In addition, ρ(i,j) and θ(i,j) are the metric distance and bearing, respectively, on edge e(i,j) . The object is denoted by o. [y]
wt =
3
Proposed Method
The measurement employed in this paper is expressed by Equation (2), and its value is used to detect the kidnapped-robot problem. This value is converted to entropy as follows: [y] [y] Hm (p) = − p(st , ot , zt|xt , m) log p(st , ot , zt |xt , m) (3) [y]
xt
This value is managed depending on the type of each object. The detection process for the problem is as follows: 1) The variables are initialized when nothing is observed. HT = 0, lobj = {0}
(4)
where HT is the current measurement entropy and lobj is the object list. 2) The entropy is calculated from the measurement value, using Equation (3). 3) If Hm (p) > β is true, go to step 1), otherwise go to step 4). If the object is not in lobj , add it to the list and set HT = HT +Hm (p). Here, β is a constant that determines whether or not there is a problem with the current value of the measurement entropy.
296
C. Yi and B.-U. Choi
4) If HT ≥ π is true, the robot is determined to have a problem; if false, go to step 2). If a problem is detected, go to step 5). Here, π is an entropy constant for deciding whether or not there is a problem. 5) The particles are initialized and computed during the localization progress. Detecting and recovering from the kidnapped-robot problem can be accomplished via the steps listed above. In general, the computational cost of such a procedure is high because the localization, detection, and recovery processes should operate concurrently. It has been verified that by simplifying the measurement likelihood to an entropy value, the proposed method lowers the computational cost. In addition, the effectiveness of the procedure for solving the problem has been experimentally confirmed.
4
Experimental Results
4.1
Experimental Environment
To validate the technique proposed in this paper, the following experimental environment was constructed, and appropriate experiments were carried out. • Robot platform: Pioneer 3DX • Camera: Logitech Quickcam 4000. • Vision system for object recognition: ERSP(Evolutionary Robotics. Software Package) on Linux. • Control module: Aria + Microsoft Visual Studio 2008 on windows. • Semantic localization: Matlab 2007 on windows. • Environment: Corridor (14 x 26.5 m) • Sensor assumption: a visual sensor that can measure up to a maximum object distance of 2.5 m. The distance from the robot to the observed object is estimated with a single camera to derive a fragment of the spatial context. After an object is recognized, its height in the image space is measured using a set of corresponding features, and the metric distance is then estimated by ERSP [14]. 4.2
The Topological-Semantic Distance Map
Figure 2 illustrates the topological-semantic distance map, which consists of 15 nodes (yellow rectangles) and 42 objects (green circles). The solid lines between nodes are the edges that represent the n-n contexts of distance and bearing. The dotted blue lines denote the r-o contexts and the dotted red lines represent the o-o contexts. To evaluate the proposed method for solving the kidnappedrobot problem, an experiment was repeated on six floors of the Information and Communications building at Hanyang University.
Detection and Recovery for Kidnapped-Robot Problem Bo ardA
297
Sdoor B
n ear, front nea r, left rear ele v
Win dowB N ode 8
5 step, front
Node 9
Nod e 1 0
6 step, right rea r
fa r, left rear
Nod e 1 1
6 step, left fron t
left near
far , r igh t
far, left re ar 3 step, r ig ht
e le v
Node 7 ne ar, right r ear
ne ar, right near, re ar
far , left
far , fro nt
near, left
Corr idorA left ne ar
Sdoor A Doo rD
D oorA
Window B
left nearby
elev
elev
le ft near by
left ne ar
4 s te p, fron t
Sdoo rB
8 step, front
Board A
near , r ear near , left rear
le ft near by ne ar, left front
Boar dB nea r, front
N ode 12 D oorC
near, right front
N ode 6 near, front Boar dC ne ar, right fr ont
near , r ig ht fron t DoorA
D oorA 4 step, front
3 step, front
Do orB PanelB n ear, right
ne ar, rear Nod e 5
Node 13 nea r, left
left nearb y
Corrid orC
near , right ne ar, right rear right nearby
Door C 3 step, fr ont
BoardD
Doo rA
7 step, front D oorB
WindowA
Co rridor A Cor ridor A
left nea r
far , left Node 4
far , left fr ont
nearby, front 5 step, front
Do orB
near, right rea r
nea rby, right n ear, fr ont
No de 3
PanelA Doo rA
near , left rear
nea r, left fr ont
PanelB
Node 2 4 step , right
righ t nea rby
le ft near
Door A
D oorB near, right rea r
left n earby
Window A
righ t near
near , right 2 step, left fron t
Nod e 1
BoardB
Window A
PanelB
far , left
near by, r ig ht r ear
near , re ar
No de 14 r ight ne arby
nearb y, re ar far , r igh t fron t
far, left
2 step, right far, right
2 step, fr ont
Node 15 nea r, r ight fro nt
Windo wA Pane lB
Fig. 2. Result of topological-semantic distance-map building
Fig. 3. The result if the robot fails in the kidnapped-robot problem
PanelB
298
C. Yi and B.-U. Choi
Fig. 4. The results of detecting the kidnapped-robot problem and recovering from errors
4.3
The Results of Detecting and Recovering from the Kidnapped-Robot Problem
Figure 3 shows the results of repeated measures experiments in which a robot was subjected to the kidnapped-robot problem. The Y-axis represents the error in the estimated robot location, and the X-axis represents the node where the robot was located. In cases 2, 4, 5, 6 and 7, the robot succumbed to fatal errors after passing node 3. The rest of the cases also exhibited some errors, and semantic localization with a single camera had relatively low accuracy because object recognition and metric measurement was somewhat affected by noise. However, semantic localization performance is not a big problem in navigational tasks such as fetching and delivering objects. Figure 4 shows the results of experiments repeated more than 10 times, using the proposed method. Cases 2, 3, 7, and 8 indicate detection of and recovery from the kidnapped-robot problem. Unlike the cases shown in Figure 3, the problem was detected, and the recovery process initiated immediately after the problem occurred. These results verify that our technique is efficient for detecting and recovering from the problem.
5
Conclusions
We introduced a procedure for detecting and recovering from the kidnapped-robot problem, using our semantic localization technique. Measurement likelihoods are
Detection and Recovery for Kidnapped-Robot Problem
299
converted to entropy through observation, and this value is checked continuously. We assumed that the value will increase beyond a certain threshold when the robot is subjected to the problem. The validity of the proposed method was demonstrated via experiments, and we plan to validate and verify the localization technique with other sensors in the near future.
References 1. Engelson, S., McDermott, D.: Error correction in mobile robot map learning. In: Proc. of IEEE Int. Conf. on Robotics and Automation, vol. 3, pp. 2555–2560 (1992) 2. Choset, H., Lynch, K., Hutchinson, S., Kantor, G., Burgard, W., Kavraki, L., Thrun, S.: Principles of robot motion: theory, algorithms, and implementation, p. 302 (2005) 3. Elfes, A.: Using occupancy grids for mobile robot perception and navigation. IEEE Computer 22, 44–57 (1989) 4. Miro, V., Zhou, J., Dissanayake, W.: Towards vision based navigation in large indoor environments. In: Proceeding of IEEE the IROS, pp. 2096–2102 (2006) 5. Karlsson, N., Bernardo, E.D., Ostrowski, J., Goncalves, L., Pirjanian, P., Munich, M.E.: The vSLAM Algorithm for Robust Localization and Mapping. In: Proceeding of the ICRA (2006) 6. Beeson, P., Kuipers, B.: Towards Autonomous Topological Place Detection Using the Extended Voronoi Graph. In: Proceeding of IEEE ICRA, pp. 4384–4390 (2005) 7. Kuipers, B.: The Spatial Semantic Hierarchy. Artificial Intelligence, 191–233 (2000) 8. Tully, S., Moon, H., Morales, D., Kantor, G., Choset, H.: Hybrid Localization using the Hierarchical Atlas. In: Proceeding of IEEE the IROS, pp. 2857–2864 (2007) 9. Thrun, S., Burgard, W., Fox, D.: Probabilistic Robotics. MIT Press, Cambridge (2005) 10. Krieg-Br¨ uckner, B., Frese, U., L¨ uttich, K., Mandel, C., Mossakowski, T., Ross, R.J.: Specification of an Ontology for Route Graphs. In: Freksa, C., Knauff, M., Krieg-Br¨ uckner, B., Nebel, B., Barkowsky, T. (eds.) Spatial Cognition IV. LNCS (LNAI), vol. 3343, pp. 390–412. Springer, Heidelberg (2005) 11. Bateman, J.A., Farrar, S.: Modelling Models of Robot Navigation Using Formal Spatial Ontology. In: Freksa, C., Knauff, M., Krieg-Br¨ uckner, B., Nebel, B., Barkowsky, T. (eds.) Spatial Cognition IV. LNCS (LNAI), vol. 3343, pp. 366–389. Springer, Heidelberg (2005) 12. Yi, C., Suh, I.H., Lim, G.H., Choi, B.-U.: Bayesian Robot Localization using Spatial Object Contexts. In: Proceedings of the 2009 IEEE/RSJ International Conference on Intelligent RObots and Systems, pp. 3467–3473 (2009) 13. Lee, J., Lee, K., Ahn, S., Choi, J., Chung, W.K.: mplementation of Global Localization and Kidnap Recovery for Mobile Robot on Feature Map. The Journal of Korea Robotics Society, 29–39 (2007) 14. Munich, M., Pirjanian, P., Bernardo, E., Goncalves, L., Karlsson, N., Lowe, D.: SIFT-ing Through Features with ViPR. IEEE Robotics and Automation Magazine, 72–77 (2006)
Factors Affecting the Turnover Intentions of Small and Medium Construction IT Workers Dong Hwan Cho1 and Haeng Nam Sung2 1
Venture & Business, Gyeongnam National Universtiy of Science and Technology, 150, Chilam-dong, Jinju, Gyeongsangnam-do, 660-758, Republic of Korea 2 Management Information System, Gyeongsang National University, 900, Gazwa-dong, Jinju, Gyeongsangnam-do, 660-701, Republic of Korea
Abstract. The aim of this study is to measure job embeddedness(fit, links, sacrifice) and work satisfaction(career satisfaction, job satisfaction) and figure out their relations to turnover intention, targeting small and medium construction IT workers through the related research. Through analytical research, three hypotheses are adopted; the more sacrifice, the higher career satisfaction, and the higher job satisfaction the employees have, the less turnover intention they have. However, these two hypotheses are rejected: the higher small and medium construction IT workers’ embeddedness(fit and links), the lower turnover intention. The results of this research can be used to identify the factors affecting turnover intention, especially in the small and medium construction IT industries, and to actually lower turnover intention. Keywords: turnover intention, job embeddedness, fit, links, sacrifice, work satisfaction, career satisfaction, job satisfaction.
Factors Affecting the Turnover Intentions of Small and Medium Construction
301
affecting the construction IT businesses, as evidenced by the development of BIS (Building Information System), ITS(Intelligent Transport System), u-city, and so on.
2
Theoretical Background and Related Studies
2.1
Research on Turnover Intention
There are various definitions of turnover suggested by previous research. In this study, turnover is defined as the action of leaving an organization by his/her will. This type of turnover is categorized as voluntary/involuntary turnover, avoidable /unavoidable turnover (Kim, 2005), and functional/disfunctional turnover. There are also planned and unplanned turnovers. 2.2
Researches on the Job Embeddedness
Job embeddedness means that the life of an individual is entangled by various people, or organizations, and an individual is in the center of the complicated web. The number of the connections an individual has is varied, and the degree of the attachment and the strength of the connections also vary according to the individual. Turnover confuses the individual web, which Mitchell et al.(2001) named job embeddedness. Fit. Fit is how compatible the employees feel with the company, organization or the local community he/she belongs to. Mitchell at al.(2001) suggested that the more the individual feels fitted to the job, colleagues and the business culture, the lower turnover is. Links. Links means that each individual is linked to other people, teams, and organizations officially or unofficially. A lot of organizations put a lot of importance on these links. For example, accounting firms try to increase the degree of the attachment by using a system of mentors and role models. The links outside of work have a great influence on the employees. Some firms provide their employees with information on their support and activities to the local community. These strategies of the organizations are helpful to reinforce the employees’ links between organizations and local community. Sacrifice. Sacrifice means the opportunity cost of turnover, which is the perceived cost of physical or psychological convenience sacrificed when leaving a current job. Recent research suggests that job attitude has relatively little influence on the stay or turnover of the employees. Other factors than job satisfaction, organizational commitment, and job alternative are stand out to understand turnover (Park, Lee, 2004). To measure the effect of job embeddedness itself, it is reasonable to measure factors both inside and outside of work. However, this study is aimed at identifying methods that contribute to the organization by improving the sense of unity through
302
D.H. Cho and H.N. Sung
developing and retaining employees. Therefore, it is appropriate that this study only measure factors inside work. Job Embeddedness and Turnover Intention. According to the results of Mitchell at al (2001), fit, in relation to turnover, has a negative interrelation in both hospitals and grocery stores, and links have a meaningful but lower negative interrelation i, but does not have a meaningful interrelation in grocery stores. Sacrifice has a high negative interrelation in both hospitals and grocery stores. In Kim(2002)’s research on the turnover decisions of the research workforce, job embeddedness(fit, sacrifice) contributes to reducing the turnover while job embeddedness (links) has no meaningful effect on the turnover. 2.3
Research on Work Satisfaction
Career Satisfaction. Career is the path which an individual walks on through his/her work, Career satisfaction is the overall emotional directivity about one’s career (Gattiker & Larwood, 1988) Therefore, career is behavioral process people have been pursuing and choosing to achieve career satisfaction. Job Satisfaction. Locke (1976) defined job satisfaction as pleasure caused when an individual carried out an important job. Smith (1975), also looked upon job satisfaction as a balanced feeling of liking and disliking or an array of attitudes resulting from harmony between good feelings and bad feelings. Hoppock (1935) regarded it as integration of psychological, physiological and environmental situations when a worker felt satisfied at his/her job. McCormick & Trffin (1974) defined it as the degree of the need satisfaction experienced or obtained from a job. Cranny et al.(1992) defined it as an emotional response caused by the comparison of expected results with actual ones when performing a job. To sum up the various definitions about the job satisfaction, it is an emotional attitude about the job given about an individual’s values and need satisfaction. Employees with higher job satisfaction feel stable and become loyal to the job. Work Satisfaction and Turnover Intention. The research on career satisfaction has been carried out centering around factors affecting career satisfaction. Blau(1989) suggested that career satisfaction has a negative interrelation to turnover. Aryee & Tan (1992), Aryee et al. (1994) also suggested career satisfaction has a negative interrelation to career change intention. Research on career satisfaction and turnover intention has been carried out targeting various organizations, and this research has shown that there is a relationship between job satisfaction and turnover intention. However, while this research shows that there is negative interrelation between the two, there is a little difference in the degree. Newman (1974) and Munchisky(1977) showed that there is a negative interrelation between them, which is trivial in its degree. However, Cotton & Tuttle(1986) suggested that there was a meaningful interrelation between most of the job related factors and turnover when the precedent studies were examined comprehensively through meta analysis methods.
Factors Affecting the Turnover Intentions of Small and Medium Construction
3
Research Model and Hypotheses
3.1
Definition of a Dependent Variable
303
In this study, turnover is defined as the action of leaving an organization by his/her will. To measure this, the questions used in this survey are those which Mobley(1977) & Becker(1992) developed. 3.2
Definition of an Independent Variable
In this study, job embeddedness is defined as the phenomenon where the inner /outer performance of the job and its result are affected by the composition of the overall network to which the individual belongs (Dicken et al.(1994)). Definition of Job Embeddedness. Mitchell at al.(2001) defined fit, links, and sacrifice as follows. Fit to an organization was defined as the degree perceived in how compatible an employee feels to the job he/she is performing and the company he/she belongs to. Links to the organization was defined as how connected he/she is to other people, team or organizations. Sacrifice to the organization was defined as the perceived cost of physical or psychological convenience sacrificed when leaving a current job. All these definitions are adopted here. Definition of Work Satisfaction. In this study, based on the definition of career satisfaction by Cheong(1999), it is defined as a psychological attachment to his/her job, the hope that he/she have related to the job, and the satisfaction about individual values coming from the career role expected by the members of the organization and other fields and the results of involvement of individual knowledge and ability. Regarding the individual values and need satisfaction used by Locke(1976), they are defined as the emotional attitude about the given job. Employees with higher job satisfaction will feel stable and become loyal to their jobs. Table 1. Operational Definition Concept
Factors, Source fit [18]
job embeddedness
links [18] sacrifice [18]
Work Satisfaction
career Satisfaction [6] job Satisfaction [3]
turnover intention [4], [19]
Operant Definition as the degree perceived as to how compatible the employees feel to the job he/she is performing and the company he/she belongs to how connected he/she is to other people, teams or organizations as the perceived cost of physical or psychological convenience sacrificed when leaving the current job the satisfaction about individual value coming from the career role expected by the members of the organization and other fields the emotional attitude about the job given about an individual value and needs satisfaction the action of leaving the organization by in his/her will.
304
3.3
D.H. Cho and H.N. Sung
Research Hypotheses
Based on Mitchell et al.(2001) and Kim(2002), the following hypotheses are set about the interrelation between turnover and organization fit, links, and sacrifice, each which are the variables of job embeddedness. Hypothesis I. The higher work internal job embeddedness (fit, links, sacrifice) the small and medium construction IT workers have, the lower turnover will appear. Hypothesis I-1. The higher work internal fit the small and medium construction IT workers have, the lower turnover will appear. Hypothesis I-2. The higher work internal links the small and medium construction IT workers have, the lower turnover will appear. Hypothesis I-3. The higher work internal sacrifice the small and medium construction IT workers have, the lower turnover will appear. According to Blau(1989), Aryee & Tan(1922), Aryee et al.(1994) and Mobley(1977, 1979), when an individual experiences job dissatisfaction, he/she considers turnover, or does deviant acts such as absence and passive job performance. Therefore, the following hypotheses are set. Hypothesis II. The higher satisfaction the small and medium construction IT workers have, the lower turnover will appear. Hypothesis II-1. The higher career satisfaction the small and medium construction IT workers have, the lower turnover will appear. Hypothesis II-2. The higher job satisfaction the small and medium construction IT workers have, the lower turnover will appear.
4
Empirical Analysis
During the period of the 20 days of survey, of the 200 questionnaires that were distributed to the workers working in information technology department in construction companies, 177(collect rate 88.5%)were collected and analyzed. They were analyzed using SPSS 15.0E. 4.1
Descriptive Statistic of Sample
To figure out the characteristics of the demography of the sample, a frequency analysis was conducted. There were 90 respondents of in their 30’s(50.9%), 100 college graduate respondents (56.6%), 71 respondents serving for 1~4 years(40.1%), and 141 respondents who had experience with other companies(79.7%).
Factors Affecting the Turnover Intentions of Small and Medium Construction
305
Table 2. Profile of the Respondents
Frequency Distribution
Category
152 (85.9%) male female 25 (14.1%) less than 30 28 (15.8%) 31 to 35 47 (26.6%) 36 to 40 43 (24.3%) 41 to 45 46 (26.0%) 46 to 50 10 ( 5.6%) over 51 3 ( 1.7%) yes 141 (79.7%) no 36 (20.3%)
gender
age
other companies experience
4.2
Frequency Distribution
Category educ ation level
yea rs of wor k
High school graduated
College graduated university graduated finished Grad school
less than 1 year 1 to 4 years 5 to 9 years 10 to 14 years over 15 years subtotal
Validity and Reliability Analysis of the Measurement Items
To verify the validity, an empirical factorial analysis was conducted, and to extract the composition factors, a principle component analysis was used for measurement variables. For reliability analysis, Cronbach’s α a value was used. Table 3. Validity Analysis & Reliability Analysis Conception Construct
All the correlation coefficient were meaningful under 0.01 of significance level. In detail, the turnover intention had a high correlation of -.657 to sacrifice of job embeddedness and had a higher correlation to career satisfaction (-.620) than to job satisfaction (-.506). There was, in addition, a correlation of .676 between career satisfaction and job satisfaction. Table 4. Correlation Analysis
Construct
Mean
SD
1 fit 2 links 3 sacrifice 4 career satisfaction 5 job satisfaction 6 turnover intention
In the process of verifying the hypotheses, each effect was analyzed using a hierarchical regression analysis to evaluate the relative influence. Next, the hierarchical regression model affecting turnover intention was considered. Evaluating the relative influence which can affect turnover intention, sacrifice (β=-.452) is the variable which had the greatest influence, career satisfaction(β=-.326) had the second greatest influence and the job satisfaction (β=.208) hadh the third greatest influence on the turnover intention. Based on the analysis above, hypothesis I suggest the following results. Firstly, the hypothesis that the higher work internal sacrifice the small and medium construction IT workers have, the lower turnover will show(hypothesis I-3) had a negative Table 5. Hierarchal Regression Analysis
Independent Variable fit links sacrifice career satisfaction job satisfaction
Factors Affecting the Turnover Intentions of Small and Medium Construction
307
influence. Secondly, the organization related links of the small and medium construction IT workers had no meaningful influence on the turnover intention. It is acceptable to adopt hypothesis II. It was concluded that the turnover intention, sacrifice and career satisfaction and job satisfaction play important roles.
5
Conclusions
There is a gradual rise in the importance of competent employees in the competitive society and in the fast changing environment. While companies put more efforts into recruiting and managing human resources, there is more turnover taking place. Therefore, strategic approaches to the concern and management of turnover are needed. Here, the new viewpoint about the turnover was suggested about job embeddedness and work satisfaction. To sum up this research, the higher the work internal sacrifices of the small and medium construction IT workers is, the lower turnover is. The hypothesis that the higher fit and links there are, the lower turnover will show was rejected, which is a difference from other precedent research. However, there was also a study which showed that the significance of the fit, and links was at the low level(Mitchell et al.,2001) and there have been several studies which have shown that links had no meaningful influence(Kim, 2002; Park & Lee, 2004; Chun at al., 2005; Lee, 2007). The implication of this study is that there needs to be comparative studies with other types of businesses since this study is focused on the construction IT workers. Also, there needs to be more research on whether the variable of demography has an influence on turnover intention, embeddedness(fit, links, sacrifice) and work satisfaction(career satisfaction, job satisfaction).
References 1. Aryee, S., Chay, Y.W., Chew, J.: An investigation of the predictors and outcomes of career commitment in three stages. Journal of Vocational Behavior 44(1), 1–16 (1994) 2. Aryee, S., Tan, K.: Antecedents and outcomes of career commitment. Journal of Vocational Behavior 40(3), 288–305 (1992) 3. Bayfield, A., Rothe, H.: An index of job satisfaction. Journal of Applied Psychology 35, 307–311 (1951) 4. Becker, T.E.: Foci and bases of commitment: Are they distinctions worth making? Academy of Management Journal 35(1), 232–244 (1992) 5. Blau, G.J.: Testing the generalizability of a career commitment measure and its impact on employee turnover. Journal of Vocational Behavior 35(1), 88–103 (1989) 6. Cheong, E.I.: A Study about Organizational Member’s Career Plateauing and Job Attitude & Moderating Effect of Individuals’ Characteristics. Master. diss. Dept. of Business Administration. Sogang Univ. (1999) 7. Chun, W.B., Park, I.K.: Effect of Job Embeddedness on Turnover Intention Organizational commitment for Hotel Employees. Tourism Research 21(10), 291–308 (2005)
308
D.H. Cho and H.N. Sung
8. Cotton, J.L., Tuttle, J.M.: Employee Turnover: A Meta-analysis and review with Implications for Research. Academy of Management Review 11, 55–70 (1986) 9. Cranny, S.J., Smith, P.C., Stone, E.F.: Job Satisfaction. Lexington Books, New York (1992) 10. Dicken, P., Forsgren, M., Malmberg, A.: The local embeddedness of Transnational Corporations. In: Amit, A., Thrift, N. (eds.) Globalization, Institutions and Regional Development in Europe, pp. 23–24. Oxford University Press, New York (1994) 11. Gattiker, U.E., Laurie, L.: Predictors for Managers’ Career Mobility, Success, and Satisfaction. Human Relations 41(8), 569–591 (1988) 12. Hoppock, R.: Job Satisfaction. Harper & Row, New York (1935) 13. Kim, E. A.: The Study on the Relationship between Job Embeddedness and Turnover Intention. Master. diss. Dept. of Information and Science. Ewha Womans Univ. (2005) 14. Kim, J. J.: The Empirical Study on the turnover decision process of research scientist – Replication of Unfolding Model. Ph.D. diss. Dept. of Management. Korea Univ. (2002) 15. Lee, J. I.: A Study on the Effects of Job Embeddedness on Turnover Intention, Staying in Organization and Organizational Citizenship behavior for Employees in Travel Agency. Master. diss. Dept. of Management. Kyoungwoon Univ. (2007) 16. Locke, E.A.: The nature and cause of job satisfaction, handbook of industrial and organizational psychology. Rand Mcnally College Publishing Company, Chicago (1976) 17. McCormick, E.J., Tiffin, J.: Industrial Psychology. Prentice-Hall (1974) 18. Mitchell, T.R., Holtom, B.C., Lee, T.W., Sablynski, C.J., Erez, M.: Why People Stay: Using job embededness to Predict Voluntary Turnover. Academy of Management Journal 44(6), 1102–1121 (2001) 19. Mobley, W.H.: Intermediate linkages in the relationship between job satisfaction and employee turnover. Journal of Applied Psychology 62(2), 237–240 (1977) 20. Munchinsky, D.M.: Organizational relationship to organizational climate and job satisfaction. Human Communication research 12(3), 395–419 (1977) 21. Newman, J.: Predicting absenteeism and turnover: A field comparison of Fishbein’s model and traditional job attitude measure. Journal of Applied Psychology 59, 610–615 (1974) 22. Park, K.K., Lee, K.E.: A Study on the Relationship between Job Embeddedness and Turnover Intention in Korea. Korean Management Review 33(5), 1423–1440 (2004) 23. Smith, H.C.: Psychology of Industrial Behavior. Mcgraw-Hill, New York (1975)
Multiple Case Studies on the Critical Success Factors of Internet Banks: Implications for Korean Market Dong Hwan Cho* and Jin Hyuk Hong** Department of Venture and Business, Gyongnam National University of Science and Technology, 150, Chilamdong, Jinjushi, Gyongsangnamdo, Republic of Korea
Abstract. Special banks primarily using Internet have different properties in transactions with customers, risk management, capital strength, and branch networks compared with general banks. The success and failure factors of Internet Banks have been explained only by the economics of scale, and the learning and experience effects of Internet Banks based on the perspective of ecologies. In this study, the success factors of Internet Banks are investigated based on the strategic choice by organizations and the resource-based view of the firms instead of an organizational ecology perspective. To this end, 31 major Internet Banks have been classified by the size and profitability of banks and three strategic groups have been derived. The three representative company cases were analyzed, and the critical success factors of Internet banks were derived. Implications for Korean market were discussed. Keywords: Internet Primary Bank, Internet Bank, Business Model, Critical Success Factors.
1
Introduction
Recently, Internet Primary Banks have been increasingly concerned about the various aspects⎯the extension of Internet Banking Service, the expensive efficiency, the service differentiation. The possibility of establishment of Internet Primary Banks will be raised by the mitigation of financial regulation(bring down the fund scale by the Bank’s Law, take some preventive measures by the real-name financial transaction system) in Korea. Although there has been much concern about the establishment of Internet Primary Banks in Korea, the research related to this issue has been insufficient. The Internet Primary Banks have a similar feature by general banks⎯e.g., trade types with customers, risk management, capital power, and local networks. Considering the emerging variety of Internet Primary Banks due to the continual development of the Internet and information technology, the fact that little research has been conducted about this topic is remarkable in the near future. One study on the survival, extinction, * **
success, and failure of Internet Primary Banks(eg. [3][4]) have been based on the perspective of organizational ecologies. Due to the fact that the history of and research about Internet Primary Banks is very short, the perspective of organizational ecologies is based on the success and failure factors of Internet Primary Banks. Research about Internet Primary Banks has focused on the strategic choice by organization and resource-based view of the firms. In this study, the success factors of Internet Primary Banks are explored based on the strategic choice by organization and resource-based view of the firms instead of the organizational ecology perspective. Because Internet Primary Banks don’t have an appearance cases, this study will analyze case studies by the foreign Internet Primary Banks in the U.S., E.U., and Japan. This research recognized the fact that foreign Internet Primary Banks would be discriminated against by the strategic types, and this would be indicated to show the strategic groups by the representative three companies.
2
Literature Review
Internet Primary Banks have been defined as banks using the Internet as the main channel of sales instead of face trade through a branch office. The Internet Banking System, by offering the dimension of channel diversification in the existing banks, has been discriminated against[1]. The theoretical Internet Primary Banks have preserved attractive features by distinguishing themselves from traditional banks[4]. By eliminating a physical location and hiring fewer employees, Internet Primary Banks have been able to decrease operation and labor costs.1 The primary stage of Internet Primary Banks was mainly established as a type of complete non-storage, for the Internet overcomes all boundaries was reinforced the scale power to use for the part of off-line channel by the ATM and the Internet Cafe in this recently. In the first stage, since only the Internet was unitized as the sales channel, this was called an Internet Only Bank, or an Online Bank, recently the Internet has been used as the main channel, and off-line channels have been used as complementary channels; this let to them being called Internet Primary Banks, or Online Primary Banks. The success and failure factors of Internet Primary Banks from a theoretical base have mainly been based on the perspective of organizational ecologies by the existence and the extinction of neo-companies. From the perspective of organizational ecologies, the cause by the existence and the extinction of firms has been explained as well as the disadvantages of neo-companies, the disadvantages of scale, and the disadvantages of business capacity[5][6]. DeYoung(2005)[3] explored the success and failure factors of Internet Primary Banks as a disadvantage of neo-company based on both the learning and experience
—
1
According to the investigation made by the Booz, Allen & Hamiltion, the account transfer by Internet was spent the one cent, the service by ATM(automated teller machine) was consumed the twenty-seven cents, the service by teller was expended the one dollars.
Multiple Case Studies on the Critical Success Factors of Internet Banks
311
effects and the bank scale economics. This research analyzed the primary stage of Internet Primary Banks in 1997-2001; these results similarly have an effect on the traditional and new banks. According to the earning rate of Internet Primary Banks, the interest scale is also similar to be derived in the same research result. Yom(2005)[4] similarly deliberated on the success and failure factors of Internet Primary Banks based on the perspective of organizational ecologies. He particularly points out the lower customer demand as a handicap of new companies. Because of the fact that Internet Primary Banks were recently established, the lower business volume compared with the traditional banks can roughly be explained. In addition, he indicated three points: Fundamentally, the depositor was managed and sales for the loan. The timing composure was necessary to use the profit chance. The customer demand has restricted the Internet Banking Service. No sure about this the success and failure factors of firms actually has diverse the opinions, the present state of research about the Internet Primary Banks is more rarely based on the perspective of organizational ecologies. According to this paradigm, the success of companies has determined the firm of external factors as an organizational strategy, and structure, competency rather than the firm of internal factors as the firm of internal industry, and the country of structural property. Therefore, the success factors of Internet Primary Banks are explained based on the new perspectives, and this will be discussed from the perspective of strategic choice and the resource-based view. The perspective of strategic choice that the subject of mind and action for the strategic plan and the strategic practice have been important factors that have decided a firm’s success[7]. The perspective of the resource-based view has been determined by internal resources and competency[8]. The success factors of firm are carried out by established strategies. Due to the fact that the perspectives of strategic choice and resource-based view have shared the mutually exclusively rather than the common property, this research adopted or integrated two perspectives.
3
Research Methodology
This research selected the methodology of one case study that mainly identified the success factors of Internet Primary Banks. The case study is a variable tool for the restricted cases due to the fact that it relies on a survey design so as to generate the statistical data rather than relies on the experiment so as to the analytical generation[9]. This case analysis sorts out the methodology of a multiple-case study. The mythology of multiple-case study has merits to present a high result of validity and reliability through between replication and pattern match by the multiple-case as compared with a single case study[9]. Since the selection of case study distinguished the types of Internet Primary Banks, the representative company cases were selected to represent different strategic groups. 31 major Internet Primary Banks operating in overseas markets in the U.S., E.U., and Japan were analyzed and three strategic groups were derived: 1-Group(Moderate): size bank – large, profitability bank – medium. 2-Group(High): size bank – small,
312
D.H. Cho and J.H. Hong
profitability bank – large. 3-Group(Low): size bank – small, profitability bank – small. In the kinds of 3-strategic groups, the 2-Group(High) recorded the business performance compared with the 1-Group(Moderate) and 3-Group(Low). As the classification of strategic group that matches and classifies in the generic competitive strategy by the Michael E. Porter[10], The following is classified: 1-Group (Moderate): cost leadership. 2-Group(High): Differentiation. 3-Group (Low): Reactor(refer to Figure 1). This research selected three companies that represented the three strategic groups, and followed the process of historical-evolution by looking at the success and failure factors of the three firms, and the causes and the nature of the success or failure factors of Internet Primary Banks. In addition, the researcher examined the cause and the nature in the success and failure of private companies regarding the multiple-case mythology, totally what was clarified as the main success factors of Internet Primary Banks. Each representative company case was chosen from a strategic group: 1Group(cost leadership): ING Direct in Canada. 2-Group(Differentiation): Seven Bank in Japan. 3-Group(Reactor): Net Bank in U.S. Since each strategic group was investigated, the representative cases were examined by the main statues and core factors.
FC5 FC5 HchU`5ggYhg V]``]cb#
Fig. 1. Scale and Profitability of Internet Primary Banks
4
Research Findings
4.1
ING Direct
ING Direct was established as a direct-banking franchise of ING Group in 1997 in Canada, and it began to make profits as the ING Group in 2001. ING Direct has17.5 million customers from 9-countries(U.S., Canada, UK, Germany, French, Spain, Italy, Netherland, Australia etc.), and recorded the pretax profit of 7.17 billion Euro in 2006. The core success factors of ING Direct are summarized as glocalization. The scale of business is expanded to entry into the countries based on the successful business in
Multiple Case Studies on the Critical Success Factors of Internet Banks
313
one country(globalization); at the same time the emerging countries of customer and market compatibly gains complete success(localization). ING Direct is effective for adopting a clear or simple strategy for the type of investment commodity and utilizing higher interest rates commodity through the cost reduction by offering or dividing in the stage of market entry. Another main success factor is that it is able to operate the “think global and act local” for other countries by the banks through an effective controlling structure based on the principle of a “fleet of companies”. The consist of nine the countries for ING Direct where the most of bank as like “the fleet of companies” is the part of form to be permit in the fairly autonomy, it has the part of form to be make in the autonomy for the business decision making. According to the fact that the knowledge share is cleared over national boundaries through the various societies or the conferences, the organizational learning effect can try to advance to the maximization. The organizational learning effect for the knowledge share also could be maximized by various societies⎯e.g., directors’ conferences, marketing conferences, sale and operations conferences, IT conferences, and global conferences attended annually by 200 managers. The approach of global dimension and local dimension has synthetically utilized customer management, and customer management has suitable analyzed the market situation. The reliability and flexibility of IT through the standard IT-Infra and the partial alteration system have been secured, and this reliability, flexibility, and userconvenience are some of the main success factors of ING Direct. 4.2
Seven Bank
Since the appeasement policy of financial industry in Japan, Seven Bank first established the name of IY Bank in April 2001; after that the Seven Bank was later changed in October 2005 and has attained higher growth and a sustainable profit in 2003. Throughout Japan, there are over 15,000 ATMs(automated-teller machine), and over 500 finance agency constitute the network. Recently, the financial demand of Japanese tourists has extended the business, and so ATMs offer services that allow customers to withdraw money from a bank card or credit card. The stockholders are widely organized(September 2006): SevenEleven: 24.9%, Itoyogato: 16.1%, Mitbishi UFJ BANK: 4.9%, Cooperate Value Up Fund: 4.5% etc. The success factors of Seven Bank are clearly defined as well as what core resources and capability of the firm are. Seven Bank is utilized in tactics by the effective focused-strategy. More costs are spent on a computation system instead of a wholesale construction, and by not costing much, the ATM Network effectively uses resources. The sales strategy using ATM Network by the Seven Bank was clearly defined as a core competency, as has been indicated by the results, and has been so successfully evaluated and received that is has come to be called ‘Convenience Financing’. Establishing ATMs is not necessary an additional resource in Seven Bank; it is clearly
314
D.H. Cho and J.H. Hong
defined that what is a mean the core resource that ATM is able to use in the practical strategy by the various finance service. The utility of ATM network is steadily increased so that it has formed an alliance with various finance agencies. This is an additional confirmation of the effective strategy of using core resource through ATMs. SevenEleven and Mitbishi UFJ BANK in Tokyo have each demonstrated a professional capability or created the synergy, and this also points to the main success factors for the retrenchment of marketing cost by changed the name in the Seven Bank. 4.3
Net Bank
The Internet Primary Bank of Net Bank, which was established in 1966, initially preformed in the black and, have been evaluated as a successful model of Internet Primary Bank in 3 years hence by making profits in 1999. Thereafter, Net Bank took over both the Market Street Mortgage Corporation in Florida in 2001 and the Resource Bancshares Mortgage Group in 2002, and it pursued to increase its profit base by focusing on the house secured loan. However, in the process of expanding by the Retail Banking, the Financial Intermediary, and Transaction Processing, which was caused by the irrational investment, it recorded a deficit in 2005, became bankrupt in September 2007, and was taken over by ING Direct. The diversification of unreasonable business by large scale construction of ATM networks in the U.S. was determined to be the main cause of bankruptcy. In the primary object of individual and small and medium sized enterprises, the retail financial services as part of the Retail Banking, the Financial Intermediary, and the Transaction Processing in business is extended to attempt an unreasonable enterprise. To ensure Critical Mass of customers in the early stages of business, this target as based on extending the profit basis did not effectively realize the stage of business in the conversion. In the same of Internet Primary Banks, the early stage of customer security which offered a high interest rate of commodity by reducing the local operating costs was successful; therefore, the later stages could not be prepared the success factors of convert strategy. After this customer security of Critical Mass, this profit extended strategy was demanded, and the strategy of Net Bank organized a new customer security by aligning the various firms and extended customer encounters, such as ATMs. In an attempt to offer the various commodities through the variety of needs satisfaction in finance agencies by the customers, and it is incurred from the uncertain of target customer to the external outsourcing of core competency. All sorts of the deposits and loan commodities were basically classified as commodity strategies, and this alienated professional companies. To improve the weak point of brand power of Internet Primary Banks, the absence of marketing strategy attained an exaggerative situation both in the marketing costs and indirect costs of other Internet Primary Banks.
Multiple Case Studies on the Critical Success Factors of Internet Banks
5
315
Summary of Results and Implication for Korean Market
The success factors of Internet Primary Banks for this study are explored based on the strategic choice by organizations and resource-based view of the firms instead of an organizational ecology perspective. This research discovered the internal factors by the organizational strategy, the organizational structure, the process, and the technology. The Internet Primary Banks operating in overseas markets have been classified by the size and profitability of banks, and three strategic groups have been derived. The first types of groups by the cost leadership strategy pursued the cost efficiency, the larger-size, and the standardization; the second types of group by the differentiation strategy utilized the core competency in the niche-market(the representative company by Seven Bank); the last types of groups by the Reactor strategy was consistent with the definite strategy(the representative company by Net Bank). In the three strategic groups, the differentiation groups were firstly ranked as having the highest performance, the cost leader groups were secondly ranked in the middle of performance, and the lower performance groups exposed the Reactor. The main success factors of Internet Primary Banks understood the multiple-case study. Firstly, the perspectives of strategy are critical to understanding the main success factors for the acquisition and preservation of competitive advantages. Secondly, the perspectives of resource-view emphasized the acquisition of completive position in the firm by the firm-specific organizational mechanism. In the case study of research results, ING Direct organized the core competitive of assets operation, and Seven Bank adopted the organizational strategy on the channel by the ATM. On the other side, Net Bank confused the core competitive and imitated the offline system of traditional banks. The organizational perspectives were successful in reducing the marketing cost and public relation cost by the brand power in the mother of Internet Primary Banks. The process perspectives were successful in offering loans with high or lower interest commodity so as to have not held by the lower operating cost in the branch offices and employees. The technology perspectives were successful in retaining the internal of the internet and information systems by the operating business of Internet Primary Banks. Internet Primary Banks are necessary to arrange the conditions for the stable growth from both a short-term extended perspective, As shown in the overseas cases, Internet Primary Banks were need to overcome the increase of marketing costs, the expansion of service costs, and the limitation of customer encounters, etc. The differentiation strategy can be important for competitive achievement and competition mitigation. Emerged the Internet Primary Banks, the existing general banks devoted the diversification of profit structure and entry of overseas market, as well as the enhancement of cost efficiency against the new competitive condition[2]. The policies authorities will continuously endeavor for political consideration through the development of financial markets and stability of the financial system by each complementary equipment and regulation system in the system of Internet Primary Banks.
316
6
D.H. Cho and J.H. Hong
Conclusions
In this study, the success factors of Internet Primary Banks have been investigated based on the strategic choice by organizations and a resource-based view of the firms instead of an organizational ecology perspective. In the three representative company cases that were analyzed, the critical success factors of Internet Primary Banks were derived and considerable points when operating Internet Primary Banks were discussed. The theoretical contribution of this research is that it offers a new based theory about the success factors of Internet Primary Banks and derives the success factors of Internet Primary Banks. The practical contributions of this research are that it offers guidelines for the successful operation of Internet Primary Banks in Korean Market and also has practical implications. In spite of its contributions, the research has certain limitation. First, the failure factors of Internet Primary Banks show significant differences in the growth stages. Secondly, this research investigated the success and failure factors of Internet Primary Banks through attached importance of common properties, and the property of inherent business is not reflected deeply. Despite these limitations, this study offers a variety of data that implications for both qualitative and quantitative studies in the future about the success factors of Internet Primary Banks.
References 1. Jung, Y.S.: The prospect and Current State of Internet Primary Bank. Settlement Payable and Information Technology (Korean), 1–28 (2006) 2. Goo, B.S.: The Methodology of Internet Primary Bank. The Korea financial Research Institute (Korean), 1–29 (2008) 3. DeYoung, R.: The Performance of Internet-Based Business Models: Evidence from the Banking Industry. Journal of Business 78(3), 893–947 (2005) 4. Yom, C.: Limited-Purpose Banks: Their Specialties, Performance, and Prospect. FDIC Banking Review 17(1), 19–36 (2005) 5. Bradley, B., Rubach, M.J.: Small Business Bankruptcies: A Comparison of Cause from 1981 and 1995. Journal of Business and Entrepreneurship 11(1), 135–151 (1999) 6. Shepherd, A., Douglas, E., Shanley, M.: New Venture Survival: Ignorance, External Shocks and Risk Reduction Strategies. Journal of Business Venturing 5(2), 393–410 (2000) 7. Cho, D.S.: Strategic Management in the 21 Century. Revised edn. Seoul Economics and Business, Seoul (1998) 8. Barney, J.: Firm Resource and Sustained Competitive Advantage. Journal of Management 17(1), 99–120 (1991) 9. Yin, K.: Case Study Research Design and Method, Revised edn. Sage Publications, NewYork (1989) 10. Porter, M.: Competitive Strategy. Free Press, NewYork (1980) 11. ING Direct, http://www.ingdirect.com 12. NetBank, http://www.netbank.com 13. Seven Bank, http://www.sevenbank.co.jp
Development Plan of Pet Dog Safe Transportation System Using IP-RFID Hyung-Rim Choi1, Byoung-Kwon Park2, Young-Jae Park3, Min-Je Cho2, Chang-Hyun Park1, Jin-Wook Lee1, and Gwang-Rok Kim1 1
Abstract. Since the numbers of single-person family and nuclear family increase, the pet market is getting bigger to meet the increasing demand of pet purchase. Different from other cargos, the pet transportation requires real-time monitoring of pet status. However, at present, there is no monitoring system on pet during transportation. In fact, there are many incidents of pets dying after the transportation because of the stress they had suffered caused by the problems that had occurred during transportation which remain unknown and unidentified. This study suggests a system that can monitor the location and environment information of pet dog during its transportation using IP-RFID. This system enables the real-time remote-monitoring of pet dog situation while being transported using Internet Protocol communication. In addition, the system also enables to take immediate actions on the problematic issues by bidirectional communication. It is expected that pet dogs would be transported in safe and comfortable manner without stress if this system would be realized and be used. Keywords: IP-RFID, Pet, Transportation.
status of pets during transportation, and many problems occur during the transportation because of lack of the monitoring system during transportation process. Accordingly, a system that can monitor the status of pets and enables to take proper and quick actions when problems occur during transportation is urgently required for the pet transportation service market. This study suggests a system that can enable the safe transportation of dogs, the most common pet. The suggested system uses IP-RFID to resolve the current issues. IP-RFID with IP communication enhances its mobility and expandability. The existing RFID can collect the information through the reader; however, the IP-RFID can collect the information at real-time using the Internet infrastructure such as Access Point(AP). Application of the IP-RFID technology to the transportation of pet dog enables the real-time tracking while being transported. The system suggested by this study enables the monitoring of pet dog status by bidirectional communication to take proper measures when there is a problematic issue.
2
General Information of Pet Transportation
2.1
Pet Market Status in Overseas and Korea
The pet market of the U.S. is growing every year. In 2001, the market was $28.5 billion with annual growth of 6% and reached total revenue of $48.3 billion today which is $19.8 billion increase from year 2001. The biggest market is the pet food market and the second biggest market is the pet care market. The pet care market accounts for 27% of the entire pet market and it has the revenue of $13.1 billion. In the pet care market, there are pet medical service, pet lodging service and pet transportation service, of which the detail revenue breakdown is not known. Nevertheless it is clear that the pet transportation service market is continuously growing together with other pet related markets.
Development Plan of Pet Dog Safe Transportation System Using IP-RFID
319
Table 1. U.S pet market size breakdown for 2010
Unit: billion
Sector
Market Size
Share
Care
13.1
27%
Food Supplies Purchases Pet Grooming Total
18.7 10.9 2.1 3.5 48.3
39% 23% 4% 7% 100%
American Pet Products Association(2011) The number of people who keeps pets at home is also rapidly increasing in Korea as the society enters into the aging society. The increase in single person families and nuclear families together with low birth rate also make the number of people with pets increase. According to the Korea Kennel Club, approximately one million people in Korea keep pets at home. The market size is around 2 trillion KRW at present and it shows approximately 15% growth every year. Different from the past, the current pet market is expanding to various fields such as pet hotel and pet grooming in addition to existing pet foods and pet medical service market. Even the pet market has such high growth potential; the pet transportation service market is not growing at all. The market is still in early stage because there is no way to properly manage the pets during transportation since there is no valid method to monitor the problems. 2.2
The Present Status of Pet Transportation
The process of pet transportation is as following. First, the destination and the departing date are confirmed when the reservation is made in advance. Then the breed and the status of the pet would be checked. At this time, if there is any potential possibility that the pet would make problem during transportation, it would be identified. Then the transportation company takes the pet on the designated departating date. After the trip, the pet arrives at the destination and the transportation company is paid.
Reservation
Confirm status of pet
Take pet
Transportation
Fig. 2. Pet transportation process
Hand over pet and be paid of transportation fee
320
H.-R. Choi et al.
Most of the problems of pet transportation happen during the transportation process. The pet cannot adapt to the unfamiliar environment of a motor vehicle, and there is no way to monitor the current location and the status of the pet. Different from other general cargos, pets are sensitive to the change in the environment. Pets get severe stress because they cannot adapt to the unfamiliar transportation environment or the noise and vibration during transportation. Many pets become physically weaker after the transportation and even some of them die after transportation because of the stress. The major reason is that it is not possible to monitor the transportation process and there is no way to know the problem and take proper actions during the transportation process. In some cases, the pet gets suffocated to death due to the rise in temperature in the cargo section during transportation. Therefore, pet transportation companies sometimes decline to transport certain pets that have the high possibility of problem during transportation in order to avoid the accidental death. 2.3
Case Analysis
Till today, most of the available systems in the pet market were too efficiently manage the history of the pets. There is no existing system that takes care of pets under transportation except a system called “uPetCare”. The “uPetCare” system attaches a sensor to the pet and the sensor information is transmitted to the sink node which enables the monitoring of pet’s physical status. However, the “uPetCare” system monitoring works only in the limited space and the system cannot monitor the pet that is being transported and take required action. Table 2. Characteristic of uPetCare System Items
uPetCare System
Major technology
Sensor, web application
Communication method
Multi-hop Communication
Function
Identification of current pet status · ·
Features
Can monitor the status Collect the information through sink node and send it to a web server · Works only in the place with required infrastructure
Like the above, the current issue in the pet transportation is that there is no system that can perform the monitoring of pet status while being transported to a far destination and take necessary actions when problems occur. Accordingly, this study took the pet dogs as the most common pet and suggests a system that can monitor the location and status information during the entire transportation process and prevent the problem in advance by taking action on the potential problem by of bidirectional communication.
Development Plan of Pet Dog Safe Transportation System Using IP-RFID
3
321
IP-RFID General
The IP-RFID technology is the core technology in this study. The IP-RFID technology took the advantages of USN and RFID technology and combined it with minimum IPv6 technology. Its tag has an IP so that the synergy effect between existing IP infrastructure, USN and RFID infrastructure would be maximized. It guarantees the wide range of expandability and mobility by directly managing and controlling the tags.
Fig. 3. Function of IP-RFID
The existing RFID system should communicate through a middleware in order for the user to request specific data to the RFID tag or to receive data from the RFID tag. The middleware should have all information about the tag and there should be a separate mapping table to enable communication between the user and the tag. However, the IP-RFID system does not require separate infrastructure for communication because the tag can provide with information through Internet. The IP-RFID can transmit information of moving objects under transportation because it can transmit information in any place as long as the Internet is available. In addition, the existing RFID system does not allow the user to directly approach the tag and get the information. Also, it is impossible for multiple users to approach a single tag and register their information. And, it is not possible to set up a critical value for certain information and let the system notify the user if the value exceeds or falls short of this critical value in the existing RFID system. However, the tag in the IP-RFID system has its IP address different from the existing RFID system. It can provide desired parties with information and it enables immediate actions based on the information transmitted from the IP-tag using the bidirectional communication function. In this study, such a system was developed using the characteristic of bidirectional communication for taking required action when a pet dog has a problem.
322
H.-R. Choi et al.
Fig. 4. Existing RFID environment > IP-RFID environment
4
The Pet Dog Safe Transportation System
4.1
System Configuration and Operation Procedure
This system monitors the status of a pet dog being transported at real-time and enables to take necessary actions. Sensors are first attached to the neck and hind leg of the pet dog to monitor its body temperature and pulse. The transporting kennel or the container will have IP-tag and a fan for temperature control. The IP-tag has sensors to measure the temperature and humidity, an IP-camera for video transmission and a speaker for voice transmission. Figure 5 is the system configuration of Pet Dog Safe Transportation System.
Fig. 5. The structure of the pet dog safe transportation system
Development Plan of Pet Dog Safe Transportation System Using IP-RFID
323
The operation procedure of this system consists of Pet Dog Safe Transportation System transaction functions and the bidirectional communication functions. There are; (1) procedure for IP-tag itself resolving the problem if the cause is the simple temperature change of the surrounding, (2) procedure for the pet owner taking actions at real-time using bidirectional communication if the pet dog is seriously disturbed by the environment change, and, (3) procedure for the driver taking direct actions after consulting the vet if the pet dog is sick or gets injured.
Fig. 6. Sequence Diagrams(IP-Tag utilization)
Figure 6 is the scenario in which the IP-tag itself resolves the problem if the cause is the simple temperature change of the surrounding and the problem would be resolved by simple adjustment of the temperature. If the temperature around the transportation container rise, the temperature sensor attached to the IP-tag will detect the temperature rise. If temperature exceeds the average temperature that had been stored in the IP-tag before, the IP-tag itself instructs the fan attached to the transportation container to start . When the temperature goes down to the average temperature level, the IP-tag instructs the fan to stop. Figure 7 is the scenario in which the pet owner takes actions at real-time using bidirectional communication if the pet dog is seriously disturbed by the environment change. A living organism can be calmed by psychological measure. If the pet dog would hear the familiar voice of the owner at real-time, the disturbance of the pet dog can be calmed and its stress can be reduced. The IP-tag detects the body temperature rise of the pet dog and transmits the information on current status to the knowledge server. The knowledge server will compare the information from the IPtag with the pet dog status information previously saved in the server and send the required action based on the comparison result. The pet dog owner will learn the situation from the application server and the pet dog owner will send his/her voice to the IP-tag. The IP-tag will let the pet dog hear the voice of the pet dog owner by the attached speaker.
324
H.-R. Choi et al.
Fig. 7. Sequence Diagrams(Pet owner action)
Fig. 8. Sequence Diagrams(Vet action)
Figure 8 is the scenario when a person should resolve the problem, instead of the system, if the pet dog is sick or gets injured. For pet dogs, an action is required in accordance with the judgment of a specialist depending on the situation. The IP-tag detects the body temperature rise and pulse rise of the pet dog and transmits the current status information to the knowledge server. The knowledge server will
Development Plan of Pet Dog Safe Transportation System Using IP-RFID
325
compare the information from the IP-tag with the pet dog status information previously saved in the server and send the required action based on the comparison result. The application server will transmit the information to the vet. The vet will learn the situation from the application server and make diagnosis on the pet dog situation. The vet will transmit his/her diagnosis to the application server. The transportation company will notify the driver of the action required in accordance with the vet’s diagnosis and recommendation. The driver will take required action and report the result to the application server. 4.2
Major Service
In the Pet Dog Safe Transportation System using IP-RFID, there are (1) location information provision service, (2) status information provision service, and, (3) realtime status management service. Table 3. Main service available Service
Location information provision service
Status information provision service
Real-time status management service
Detail Learn the location of the pet dog
Contents The location information of the pet dog can be provided at real-time using the unique information technology of IP-tag and the GPS technology
Transmit information to the IP-tag managing parties
The location information is provided to the parties related to the transportation of the pet dog using the IP of the IP-tag
Detect the status information through sensors
Attach sensors to the pet dog and check the body temperature and pulse of the pet dog Check the surrounding environment by attaching sensors on IP-tag
Transmit information to the IP-tag managing parties
The location information is provided to the parties related to the transportation of the pet dog using the IP of the IP-tag
Real-time management using IPcamera
The photo of the pet dog and the voice information of pet dog owner can be received through Internet. The pet dog can hear the voice of the pet dog owner through the speaker.
Bidirectional information exchange using IP communication
Information exchange service using bidirectional communication with pet dog at real-time using IP communication function
The purpose of this system is the safe transportation of pet dog. The system collects the location and status information of pet dog and transmits the information to the related parties. By enabling the immediate action on the problematic issue by way of bidirectional communication, the system achieves its purpose of safe and pleasant transportation of pet dog.
326
4.3
H.-R. Choi et al.
User Interface
This system has user interfaces so that the users can get desired information easily and utilize those. The user interfaces are designed separately for the pet dog owner, the transportation company, the vet and the driver.
Fig. 9. U.I (Pet owner)
Figure 9 is the interface for the pet dog owner. It provides the pet dog owner with location and status information of his/her pet dog. It is possible to know the status of pet dog at real-time. The interface shows the history of status change so that the status change trend of the pet dog can be easily learned. Pet dog owner can confirm the photograph information of pet dog by way of IP-camera and the pet dog can hear the voice of the pet dog owner through the speaker.
Fig. 10. U.I (Company)
Figure 10 is the interface for the transportation company. It shows the interface of current location and the status of the pet dog. It is possible to know the status of pet dog at real-time. The interface shows the history of status change so that the status change trend of the pet dog can be learned. The tags are managed by customer record so that the pet dog data can be managed by customer. There is a button that can transmit the action instruction to the driver if there would be a problematic situation.
Development Plan of Pet Dog Safe Transportation System Using IP-RFID
327
Fig. 11. U.I (Vet)
Figure 11 is the interface for vet. It consists of mainly the status information of the pet dog. The vet can see the photograph of the pet dog taken by the IP-camera. The status of pet dog such as body temperature and pulse are managed by both values and graph so that the history and trend can be more accurately managed and presented. If there is a problematic situation, the interface enables a more accurate status diagnosis.
Fig. 12. U.I(Driver)
Figure 12 is the interface for the driver. It shows the present status of the pet dog. If there is a problematic situation, it is possible to transmit an alarm message to the smart-phone of the driver. It can also tell the driver of the location of nearby pet clinic. The transportation company and pet dog owner can know the result of the action taken after the instruction had been given.
5
Conclusion
This study suggests a system that can enable the safe transportation of pet dogs by resolving the problems during transportation using IP-RFID technology. In accordance with the growth of the pet market and the increase in pet demand, all pet related markets are growing; however, only the market of pet transportation service is not growing. The reason is that there is no system at present that can monitor the status of pets during transportation and enable the necessary action immediately. Accordingly, a system that can monitor the status of pets from remote
328
H.-R. Choi et al.
location using IP-RFID and take proper and quick actions if something happens during transportation was developed. This system is expected to create new value by directly providing various users such as pet dog owner, the transportation company and vets with information occurring during the pet dog transportation through the IP communication function of IP-tag. It is possible to confirm the location and status of pet dog at real-time and take immediate action caused by the environment change or status change. With such functions, this system will achieve more safe and improved transportation of the pet dogs. This study has its contribution value in the suggestion of possibility to safely transport the pet dog using the bidirectional communication function of IP-RFID. The limitation of this study is that, at present, only the photograph data can be transmitted and the transmission of video data cannot be transmitted because of the limitation in the transmission capacity of the IP-tag. Technology development increasing the transmission capacity of IP-tag in the future is required so that video data can be transmitted and this system can be actually realized. In addition, a business model development would be also required so that this system can be actually applied to the commercial market. Acknowledgments. We inform that this study is a result of the Ministry of Land, Transport and Marine Affairs; Active IP-RFID Technology Development Project for Marine Logistics and thank for the support with the research fund.
References 1. Park, J., Lee, G., Cho, J.: uPetCare: Ubiquitous Pet-CareSystem using Web2.0. Computer programmer qualification 15(4), 260–264 (2009) 2. Park, J., Lee, G., Cho, J.: uPetCare: Design and Implementation Ubiquitous PetCareSystem using Web2.0. Korea Computer Congress 2010 35(1), 39–40 (2008) 3. Choi, H.-R., Park, B.-K., Lee, C.-S., Park, Y.-S., Choi, K.-N.: Study on Valuable Cargo Transportation Service Using Smart Box. In: The 2nd International Conference on Logistics and Transport (2010) 4. Hong, S.-H., Park, H.-C., Lee, H.J., Park, J.-H.: UPMS: U-Pet Management System in smart Environments. In: The 2008 Fall Conference of Korea Multimedia Society, pp. 284– 287 (2008) 5. Kim, M.-G., Jung, B.-H., Juhn, J.-B., Lim, H.-G., Park, H.-H.: Development of Hybrid RFID Reader to Recognize Pet. The Institute of Electronics Engineers of Korea 3(1), 1848–1849 (2010) 6. Kwon, G.: Development of Service Information System (SIS) with Real-Time Using RFID in Ubiquitous Environment. National IT Industry Promotion Agency (2008) 7. American Pet Products Association Annual Report (2011), American Pet Products Association (2010)
The Security Process Model of Embedded Systems Shin-Hyeong Choi The Department of Control & Instrumentation Engineering, Kangwon National University, 1 Joongang-ro Samcheok-si, Gangwondo, Korea [email protected]
Abstract. Embedded systems for specific purposes perform certain roles in various environments. Security of embedded systems is becoming a paramount concern, and once a problem occurs in embedded systems, can lead to fatal accidents. Therefore, we need security policy for embedded systems. However, due to the limited size and performance, security techniques developed for enterprise and desktop computing might not satisfy embedded application requirements. In this paper, we analyze existing security policies and divide malicious attacks into three sections, and then present the security process model of embedded systems. Keywords: Security, Embedded Systems, Malicious Attacks.
In this section we briefly describe the embedded system environments common to the wired and wireless worlds. There is a difference between embedded systems and personal computer in terms of size, performance and utilization. But, security and privacy issues, which can occur on a personal computer, can cause similar in embedded systems. In particular, the spread of the Internet and the emergence of embedded devices equipped with wired and wireless communications capabilities, shows that IP addresses don't belong to personal computer any more. The advent of embedded devices such as smart phones resulted in a shortage of IP addresses. However, as more embedded systems are connected to the Internet, the potential damages from such vulnerabilities scale up dramatically. Internet connections expose applications to intrusions and malicious attacks[2, 3]. Today, security in one form or another is a requirement for and increasing number of embedded systems, ranging from smart phone to web servers. It has been observed that the cost of insecurity in embedded systems can be very high. Unlike desktop or server, embedded systems often must process within real-time deadlines. For example, navigation system should not inform past location information after a time. Embedded systems have traditionally optimized for speed, size, power and time to market. Recently however, feature size is decreased and functionality is increased. An embedded system is a specialized computer system that is part of a larger system or machine. Embedded systems are comprised of hardware and software for efficient management of hardware resources, which performs specific functions. Typically, an embedded system is housed on a single microprocessor board with the programs stored in ROM. The hardware-dependent software is connected with the physical hardware and network. Real-time operating system and device drivers are closely coupled with hardware platform. The embedded systems don't have enough software or hardware resources such as software testing tools and memory. The storage space that can store information is small, and CPU speed is relatively slow[4].
3
The Security Process Model
In this section, we have presented issues, architecture of embedded systems security process model to counter three section attacks. Although embedded systems with limited resources, are able to connect to the Internet through wired or wireless network. Like desktop and server, a problem may occur via a network connection. However, unlike a general-purpose system and personal computer, embedded systems are used for specific purpose and have communication functions, but because of somewhat limited access from the outside enemy, security issues has not been a significant highlight[5]. Due to this kind of importance as well as the pervasive deployment of embedded devices from home to big enterprises, security of embedded device has become a big issue. Many research initiatives have been undertaken to counter the issues of security in embedded
The Security Process Model of Embedded Systems
331
systems. Security for embedded systems is a more difficult problem than security for desktop and server computing. However, as more embedded systems are connected to the Internet, the potential damages from such vulnerabilities scale up dramatically. Internet connections expose applications to intrusions and malicious attacks. Unfortunately, security techniques developed for server and desktop computing might not satisfy embedded application requirements. Typically embedded systems have low computing power and finite energy supply based on a battery, and these factors are at odds with the computationally intensive nature of the cryptographic algorithms underlying many security protocols. In addition, embedded systems are vulnerable to attacks, like physical tampering, malware and side-channel attacks. To realize the full potential of embedded system data while providing secure communications, a more robust and flexible security mechanism is needed in embedded devices. This security mechanism of embedded systems has to be lightweight, low-powered[6, 7, 8, 9]. 3.1
Security Difficulties of Applying
However, these still exists significant difference between requirements of security processing and the capability of an embedded processor. In this paper, three difficulties are as follows: • Cost and CPU Performance Embedded systems are quite sensitive in terms of production cost, thus the CPU is 4bit or 8-bit usually, and 16 bit has been adopted recently. These CPUs are difficult to process with strong encryption keys and security programs, and the use of highperformance CPU leads to very high production costs. • Real-time Processing In addition, even if unexpected invasion or malicious code damages an accounting and information management systems that are running on the server, these systems can be repaired after a certain times. But, accidents caused by signal processing or trains control systems are often impossible to repair. Unlike transaction-oriented enterprise computing, embedded systems must run in real-time. Therefore, if the control part is paralyzed due to a very short time delay, embedded systems are vulnerable to attack from the outside. • The Absence of Security Manager Unlike server or PC, there is no security administrator or user in most embedded systems. Therefore, embedded systems are a difficult environment to take action in real time to external attacks. 3.2
The Types of Attacks on Embedded Systems
An adversary may attack through multiple paths. In this paper, we group these attacks into three categories(cf. Fig. 1), which are described as follows:
332
S.-H. Choi
Fig. 1. Attacks to embedded systems
• Long-distance attacks This is an attack to embedded systems that connected with the outside world through the wired or wireless networks, which through malicious code injection performs unwanted function such as a large amount of messages sent or automatic shut-off, and changes or disclosures illegally information stored in an embedded system. • Short-distance attacks This is an attack through short-range contacts such as RFID or Bluetooth. Adversary can change settings or know personal information. • Direct Attacks Embedded systems have limited hardware resources, so most of embedded systems don’t have a built-in security program. Therefore, adversary can trespass on embedded systems.
The Security Process Model of Embedded Systems
3.3
333
The Security Process Model for Embedded Systems
In the previous section, attacks on embedded systems from outside were classified into three types. This paper presents an embedded systems security process model, as shown in Fig.2, the specific security process is as follows: Embedded systems security activities should be started in approach distinctions stage.
Fig. 2. The Security Process Model
• Direct Approach Direct approach for embedded systems is to access through touch screen or keypad. In this way, an attacker can make several attempts to log in embedded systems. In general, authentication stage permits a few log-in. If number of login exceeds limit or adversary access without administrator permission, embedded systems automatically are locked. • Wired or Wireless Approach This approach is to access through LAN port or wireless LAN port such as 10/100 Mbps Ethernet port and 54Mbps, IEEE802.11g/b WLAN port. An attacker sends malicious codes for wrong purpose, and this may influence on embedded systems seriously. In this stage, an agent analyzes packets received from two ports and compares with normal packet patterns. If abnormal patterns are detected, block access to embedded systems. • Short-distance Approach This approach is to access through RF port or Bluetooth port, countermeasure is similar to long-distance attacks.
334
4
S.-H. Choi
Conclusion
Embedded systems for specific purposes perform certain roles in various environments. Security of embedded systems is becoming a paramount concern, and once a problem occurs in embedded systems, can lead to fatal accidents. Therefore, we need security policy for embedded systems. However, due to the limited size and performance, security techniques developed for enterprise and desktop computing might not satisfy embedded application requirements. In this paper, we analyze existing security policies and divide malicious attacks into three sections, and then present the security process model of embedded systems. This paper analyzes the necessity and problems of security on embedded systems and shows a security process model which is based on embedded systems.
References 1. Gupta, V., Gupta, S.: Experiments in Wireless Internet Security, 860–864 (2002) 2. Ashley, P., Hinton, H., Vandenwauver, M.: Wired versus Wireless Security: The Internet. WAP and iMode for E-Commerce, 296—306 (2001) 3. KayMiller, S.: Facing the Challenge of Wireless Security, 16–18 (2001) 4. Gao, F., Li, F., Bao, S., Wang, X.: Analysis and Implementation of Secure Console Server Based on Embedded Linux (2008) 5. Qian, H.-m.: A embedded software testing process model (2009) 6. Ravi, S., Raghunathan, A., Potlapally, N., Shankaradass, M.: System design methodologies for wireless security processing platform. In: Proc. Design Automation Conf., pp. 777–782 (2002) 7. Ukil, A., Sen, J., Koilakonda, S.: Embedded security for Internet of things. Emerging Trends and Applications in Computer Science (NCETACS), 1–6 (2011) 8. Henkel, J., Narayanan, V., Parameswaran, S., Ragel, R.: Security and Dependability of Embedded Systems. A Computer Architects’ Perspective, 30–32 (2009) 9. Kang, B., Kwon, Y.-J., Lee, R.Y.: A Design and Test Technique for Embedded Software. In: SERA 2005 Proceedings of the Third ACIS Int’l Conference on Software Engineering Research, Management and Applications (2005)
Electromagnetic Compatibility for U-Healthcare Remote Diagnosis System Eui-Bung Jeoung1, Je-Ho Song2, and You-Yub Lee1 1
Dept. of Auto & Mechanical Engineering, Howon Univ., Howondae 3 Gil, 64, Impi, Gunsan, Jeonbuk, Korea 2 Dept. of IT Applied System Engineering, Chonbuk National Univ., 567 Baekje-daero, Deokjin-gu, Jeonju-si, Jeollabuk-Do, Korea [email protected]
Abstract. In this paper, u-healthcare remote diagnosis system is proposed for chronic disease and medical vulnerable groups check the health systematically and support for the most optimal environment to improve the quality of life. U-healthcare remote diagnosis system using wireless measure the thoracic sound in the chest. And this is demonstrated that the system using radio frequency is not be affected by the electromagnetic wave with the use of an experiment and by confirming that this u-healthcare remote diagnosis system can not affect the doctors and the patients. Keywords: U-Health care, Diagnosis, Lung Sound, radio frequency, patients.
1
Introduction
Some lung and bronchial problems can be solve by using the thoracic sound that has been widely know as a good method for detecting diseases. A doctor who puts a stethoscope on his ear can get an ear pain while treating the patients. Therefore, it is necessary to make equipment consisting of radio transmitters and receiver modules that are suited for short range data transfer purposes.[1] In this paper, they want to show that people are not be affected by the electromagnetic wave with the use of an experiment and by confirming that this uhealthcare remote diagnosis system can not affect the doctors and the patients. Acceptable reference test of electromagnetic wave confirms through electromagnetic wave radiation and electromagnetic wave conduction.
E.-B. Jeoung, J.-H. Song, and Y.-Y. Lee Table 1. Lung Sound Nomenclature English Discontinuous Fine (high pitched, low amplitude, short duration)
French
Germa Japanese
Fine Rales Feines crackles crepitants Rasseln
捻髮音
Coarse: Coarse Rales Grobes (low pitched, high amplitude, crackles bulleuxou Rasseln long duration)
水泡音
Continuous High pitched Low pitched
Wheeze Rhonchus
Rales sibilants
Pfeifen
Rales Brummen ronflants
ふえ 音 (
)
いびき 音 (
)
Fig. 1. Tx, Rx Test Diagram
Acceptable references of electromagnetic interference are shown in Table 2 and Table 3. Table 2. Acceptable Reference of Electromagnetic Conduction
Electromagnetic Compatibility for U-Healthcare Remote Diagnosis System
337
Table 3. Acceptable Reference of Electromagnetic Radiation
U-healthcare remote diagnosis system provide one analog channel. The output was fed directly to the microprocessor that continuously monitors the input. In this levels when a change occurs, the radio transmitter is passed through a serial data consisting of the digitized. A single 3.6(V) lithium cell is used to power the unit. Thus, completely independent in use and suffers no external interference from other equipment. Also, the radio receiver monitors all the activity on the specific radio channel to which it is tuned and converts the presence or absence of the radio frequency carrier to the logic high or low respectively. This signal is monitored by the microprocessor which determines the validity of the data stream. This signal is converted by a line driver and passed it to the PC via a USB port. Power is supplied to the receiver via a standard main power adapter, the output of which is regulated within the receiver to 5(V), and l.e.d. was used to show the status of the power. The modulation of the radio frequency channel is performed entirely by the microprocessor in real time.
3
Test Specification, Methods and Procedure
To perform the relevant tests and assess the system for compliance with the under specification. (Table 4) Table 4. Equipment Test and Test Specification Description Supply Voltage
U-healthcare Diagnosis System Tx: Adaptor input AC230V, 50Hz Output DC 3.0V Rx: USB DC 5.0V
Adaptor : MNK Inc type LU-30 Support Earphone : N/A type Leemax Equipment or Peripherals Note PC : JOOYON type JYN07G Required Note PC Adaptor : LI SHIN type 0335C2065 Standard EN60601-1-2:2001 (IEC 60601-1 -2 : 2001)
Medical electrical equipment Part 1-2 :General requirements for safety Electromagnetic compatibility
338
E.-B. Jeoung, J.-H. Song, and Y.-Y. Lee
The conduction test method is carried out by following the steps below. 1. Environmental condition : Ambient temperature 20(oC), Ambient humidity 42(%) 2. Law: Tested by Radio Research Laboratory Notice No. 2008-5. 3. Transmitter and receiver are tested in operational status. Test status is shown in Fig. 1. 4. Tx and Rx ground device is grounded certainly. Also, power line put to earth. 5. Experiment value selects maximum. The radiation test method is carried out by following the steps below. 1. Environmental condition : Ambient temperature 12(oC), Ambient humidity 42(%) Field 10m. 2. Law: Tested by Radio Research Laboratory Notice No. 2008-5. 3. Radiation test method is the same as the conduction test condition. Test status is shown in Fig. 1.
4
Results and Discussion
Operating mode and condition is the Table 5. The results of conduction test is shown the Table 6. And measured graph in the hot line and neutral line is shown in Fig. 2. Table 5. Operating mode and Condition Port Basic Standard
AC Mains EN 61000-3-2:2000+A2:2005
Limit Table
Class A
Power
2.454W
Operating Mode
Window in normal operating status
Table 6. The Results of Conduction Test
Electromagnetic Compatibility for U-Healthcare Remote Diagnosis System
339
Fig. 2. Graph of Measured Hot Line(above) and Neutral Line(below) Table 7. The Results of Radiation Test
Fig. 3. shows a conduction test configuration and Table 7. is represented the result of radiation test for electromagnetic wave. Fig. 4. shows harmonics emission test configuration. Conduction and Radiation test environmental conditions and test equipment used are as follows. - Power supply : AC 230V, 50Hz - Temperature : 20oC(conduction), 12oC(radiation) - Relative Humidity : 42 % - EMI Test Receiver : Model ESHS10(R/s) - AMN : NLSK 8128(Schwarzbeck) - Test Receiver : Model ESVD(R/s) - Spectrum Analyzer : 8568B(Hewlett Packard) - Antenna Master : MA 240(HD GmbH) - Position Controller : HD 100(HD GmbH) - Biconical Antenna : 3110(EMCO) - Log Periodic Antenna : 3146(EMCO)
340
E.-B. Jeoung, J.-H. Song, and Y.-Y. Lee
Fig. 3. Conduction Test Configuration [Front(above)/Back(below)]
Table 2. is represented acceptable reference of electromagnetic conduction that is A class or B class is limited under minimum 60 or 50 in the frequency range(0.15~30 MHz). The conduction test results are shown in Table 6. and Fig. 2. The result values are displaying below the reference of electromagnetic conduction value. Also electromagnetic radiation test results are shown in Table 7. Similarly, the result values are displaying below the reference of electromagnetic radiation value.
5
Conclusion
In the experiment of electromagnetic compatibility for u-healthcare remote diagnosis system, the results are showed above Table 6, 7 and Fig. 2. Table 6 through Table 7 are shown result data for conduction and radiation of electromagnetic wave are harmless to person. Also, Fig. 2 is the result data for conduction of hot line and neutral line which are represented in low level. Considering the above results, it was shown that electromagnetic wave is not influential about people through an experiment. We confirmed that u-healthcare remote diagnosis system is harmless in electromagnetic wave for a doctor and a patient. Acknowledgment. This work was supported by Howon University.
References 1. Park, K.Y.: Electromagnetic Compatibility for Wireless Diagnosis System. Journal of Jeonju Vision University 46, 65–70 (1987) 2. Cugell, D.W.: Lung Sound Nomenclature. Northwestern Univ. Medical School, Chicago (1987)
Electromagnetic Compatibility for U-Healthcare Remote Diagnosis System
341
3. Park, K.Y., Lee, D.T.: Wireless Diagnosis System for Chest Disease using Thoracic Sound. Iee Honam Section 13(1), 57–62 (2004) 4. Beers, M.H., Berkow, R.: The Merck Manual of diagnosis and therapy, vol. 5. Merk Research Laborarories Division of Merck & Co. INC, Whithouse Station (2000) 5. Choi, J.S.: Respirstory sound, pp. 67–70 (1994) 6. Loudon, R., Murphy, P.: Lung sounds. Am Rev. Respir. Dis., 663–673 (1984) 7. Papoulis, A.: Probability, radndom variables, and stochastic processing, 3rd edn. McGraw Hill (1991) 8. SchöIkopf, B., Smola, A., Möller, K.R.: "Nonlinear Components Analysis as a Kernel Eigenvalue Problem", Technical Report No. 44, Max Planck Institut för biologische Kybernetik (1998) 9. Vapnik, V.N.: The Nature of Statistical Learning Theory, 2nd edn. Springer, Heidelberg (2000) 10. Sohara, K., kotani, M.: Application of Kernel Principal Components Analysis to Pattern Recognitions. In: SICE 2002, pp. 5–7 (August 2002)
Workload Prediction and Weighted Rule-Based Task Scheduling for Face Certification System on Distributed Parallel Computing Tae Young Kim and Jong Sik Lee School of Information Engineering, Inha University #253, YongHyun-Dong, Nam-Ku Incheon 402-751, Republic of Korea [email protected], [email protected]
Abstract. This paper presents a workload prediction and weighted rule-based task scheduling for face certification on distributed parallel computing. To compose a large-scale certification system, such as a criminal surveillance system for a public security, the system requires an enormous processing power. Thus a grid and distributed parallel computing is an essential approach for a large scale certification system. However his kind of approach is generally comprised of heterogeneous resources. And differential characteristics of each resource have influence on a performance of system. Therefore, an efficient task distribution and scheduling is necessary to improve a performance of system. There are various kinds of scheduling for task distribution. However existing methods cannot provide a suitable task distribution for a face certification system. Therefore, this paper proposes a task scheduling which includes a queue management policy with workload-prediction and weighted rule-based resource selection. The proposed method predicts the volume of certification task for a task queue management policy and selects the suitable certification server using performance weighted rules. Simulation result shows that the proposed method has better performance than other scheduling methods. Keywords: Task Scheduling, Workload Prediction, Rule-based System, Distributed Parallel Computing.
Workload Prediction and Weighted Rule-Based Task Scheduling
343
a public security or emigration and immigration control system on airport, requires multiple processing unit environments like a distributed parallel computing to process a large scale transaction of face certification request. In a distributed parallel computing environment, there are a lot of distributed resources. To utilize a various distributed resource efficiently, a suitable task distribution and scheduling policy is an essential method. Much study has been done on the topics of efficient task scheduling method in distributed parallel computing environment [3]. And every scheduling method provides a reasonable performance with its own policy. For instance, some kinds of scheduling method use a calculated cost of performance or a threshold of some factors [4]. However, a face certification system needs a higher reliability and a lower response time. It means, a face certification system requires some task distribution policy to reduce a certification job loss and some efficient resource selection with a certificate on performance. To satisfy requirements of task scheduling for a face certification system, we proposed two kinds of approaches. First, we use a workload prediction based queue management policy to reduce a certification task loss. This policy controls a capacity space of queue with predicted workload. Second, we use a resource selection policy with weighted selection rule. This policy can find and select a most suitable resource which records a high performance. The rest of this paper is organized as follows: In section 2, we briefly review related works. In section 3, we describe two main ideas. Section 4 shows simulation design and results. Finally we conclude in section 5.
2
Related Works
Existing task scheduling policy for a grid and distributed computing can classify into 2 groups. First group did not consider a performance factor of resources. Random or Round-robin method [5] is a best example of this group. These kinds of approach use simple information for resource selection. Random approach selects a resource using generated random value and round-robin approach selects a resource using predetermined sequence [5]. These kinds of approach can provide a basic policy for task distribution. However, previously mentioned, these approaches did not consider a performance of resources. So we cannot utilize a best performance of resources. Second group consider a performance factor of resources. Optimal cost scheduling [4] is a best example of this group. These kinds of approach represent performance of each resource as some specific value like a cost [4]. Each cost shows a profitable level of each performance. So, cost-based selection policy can find a most suitable resource with high probability. However, performance level is not a variable factor. If some resource possesses a high processing power and has an advantage of target area, the resource always evaluates a high performance level. So in this case, performance based selection is an obstacle policy for a task distribution. Thus additional approach is needed for a balanced distribution such as a load-balancing policy [6].
344
T.Y. Kim and J.S. Lee
In FCS, the system requires a reasonable certification result. It means the system requires high performance of certification success rate and response time. And the system also needs a high reliability. Thus the system does not want a frequent situation of task loss and task congestion situation on some specific resource.
3
Proposed Scheduling Method
In this section, we describe two main key ideas of our study. First, we describe a workload prediction based queue management policy. And second, we describe a weighted priority rule based resource selection. 3.1
Workload Prediction Based Queue Management Policy
In our study, proposed queue management policy consists of two kinds of sub policy. One is a predicted workload based variable capacity policy of task queue and the other is a threshold-based load balancing policy. These two kinds of sub policy are separated processes. But there is close correlation between two sub policies to improve a performance of FCS. More detailed description is as follows: Task Queue of Resource 1
R1 T1
Task Queue of Resource 1
R1 T2
R1 T1
Task Queue of Resource 2
Task Queue of Resource 2
R2 T1
R2 T1
Task Queue of Resource 3
R3 T1
R3 T2
R3 T3
R3 T4
Qu eu e
R3 T5
Task Queue of Resource 4
R1 T2
R3 T3
Task Queue of Resource 3 Fu ll
!
R3 T1
R3 T4
Task Queue of Resource 4
R3 T2
R3 T5
Fig. 1. Example of how to load balancing progress works
First, we describe a load balancing policy. As previously mentioned, our study uses a threshold value to determine a start time of load balancing policy. When a capacity of task queue of some resource reaches a given threshold value, the system starts a load balancing progress. Fig. 1 shows a simple example of load balancing progress. In Fig. 1, the threshold value is a maximum capacity of each resource. So threshold value is 5 in Fig. 1. And as you can see from left side of Fig. 1, a task queue size of
Workload Prediction and Weighted Rule-Based Task Scheduling
345
resource #3 reaches a given threshold value 5. Thus system starts a load balancing progress. Detailed progress is as follows: First, the system checks a total number of waiting tasks. As shown in Fig. 1, there are 8 waiting tasks in the system. Next, the system gets an average number of tasks per total number of resources. As shown in Fig. 1, there are 4 resources in the system. So 2 is an average number of tasks. It means, each resource should take 2 waiting tasks in its own queue. So finally, the system reassigns some waiting task of resource #3 into resource #2 and resource #4. In this load balancing progress, most waited task has a high priority. Thus as you can see from right side of Fig. 1, the waited task #2 of resource #3(R3T2) reassigns into a head position of task queue of resource #4. Therefore, we can make a balanced workload assignment and reduces a waiting time of tasks with this policy. In other word, this load balancing policy decreases a response time of certification tasks. Next, we describe a workload prediction based queue management policy. This policy is a determine progress of threshold value of task queue. Commonly, a fixed capacity of task queue is a general policy of the queue management. However we apply a variable capacity policy of task queue in our study. As previously mentioned, our study uses a threshold value to determine a start time of load balancing policy. And we use the threshold value as a maximum capacity of task queue. So a fixed capacity policy affects adversely when a workload is not light and heavy. It means, we should control a threshold value with a workload level. However, due to the workload is a continuously changed factor, it is hardly find a standard pattern of the workload. Besides, to determine a threshold value with workload, we need information of future workloads. It means, we need a prediction method for future workloads. There are various types of prediction method. However, statistical prediction is an essential approach for a workload which has an irregular pattern. In many statistical prediction methods, we choose an exponential smoothing method as an example method [7]. Although an exponential smoothing method is a simple predication method, we can get a close prediction result using historical data easily [7]. So the exponential smoothing is a suitable method to verify our idea. Pw(t) = (1 – α) × Pw(t-1) + α× Aw(t-1)
(1)
To find future workloads, the system continuously measures actual workloads and periodically records into historical data sets. And the system predicts a next workload using exponential smoothing equation (1) [7]. Pw(t) is a new predicted workload for next t round, Pw(t-1) is an old predicted workload for last t-1 round, Aw(t-1) is an old actual workload for period last t-1 round and α is a weighted factor to determine a reflection rate between Pw(t-1) and Rw(t-1). Q(t) = Q(t-1) + ( Aw(t-1) – Pw(t-1) ) + Rw(t-1)
(2)
After the prediction is finished, the system determines a threshold value using equation (2). Q(t) is a new threshold value for next t round, Q(t-1) is an old threshold value for last t-1 round, Aw(t-1) is an old actual workload for last t-1 round, Pw(t-1) is an old predicted workload for last t-1 round and Rw(t-1) is carried forward task from next t-1 round to next t round.
346
T.Y. Kim and J.S. Lee
This capacity control policy provides the ability of adapt for workload changes. And this policy also controls a timing of load balancing policy. Therefore, workload prediction based queue management policy can avoid a task loss and task congestion situation. 3.2
Weighted Priority Rule Based Resource Selection Policy
A certification system requires a high reliability. And a certification success rate and response time are essential factors of high reliability in FCS. However, a certification task contains facial information of some individual. And facial information contains various features. Those features have a positive or negative influence on a reliability of certification results.
F = {Glasses}
F = {RtoL}
F = {Glasses, Beard,
F = {ClosedEye}
Mustache} Fig. 2. Sample of features in facial information
Fig. 2 shows a sample of features in facial information [10]. There are various kinds of features in facial information. However we classify these features into two groups. First group is a frequently and continuously changed feature on every photograph. A direction of faces is a best example of this feature group. But on the other hand, second group is a less frequently changed feature on ever photograph. Glasses or beard is a best example of this feature group. So we consider a second group as a considerable feature group for reliability. A general certification result of every resource is an average result for all kinds of certification task. If some certification task contains a specific feature, the task should assigns into resource which records a high reliability for that feature. Thus we also extract and classify certification results by facial features. Then we can find a merits and demerits of each resource for every facial feature. Fig. 3 shows evaluating rules for resource selection progress. There are four major factors for evaluation. Most highly considered factor is a reliability of some features in facial information. And next highly considered factors are average certification success rate and the number of waited task. And most lowly considered factor is a response time. Every resource get or loss a selection priority score by evaluation rules. In Fig. 3, “drastically” means a high weighted priority score, “moderately” mains a normal weighted priority score and “slightly” means a low weighted priority score. After the evaluation progress is finished, the system selects a resource which evaluates a most high selection priority score. And then the system assigns a task into selected resources.
Workload Prediction and Weighted Rule-Based Task Scheduling
347
Fig. 3. Evaluating rules for resource selection progress
This weighted priority rule policy can find and select a most suitable resource for some certification task with facial features. Therefore weighted priority rule based resource selection policy can guarantee a high reliability of certification results. Thus the workload prediction based queue management policy and the weighted priority rule based resource selection policy can improve a performance of FCS on distributed parallel computing.
4
Simulation Design and Results
To verify a performance of our proposed method, we designed and implemented the simulation model with DEVS methodology [8, 9]. As shown in Fig. 4, our simulation model consists of 3 different components. A task generator generates a certification task, a coordinator schedules and assigns a task to certification resources with given scheduling policy, and a certification resource performs a certification progress.
Certification Resource
Certification Resource Task Generator
Coordinator Certification Resource
Certification Resource
Fig. 4. Designed simulation model
348
T.Y. Kim and J.S. Lee Table 1. Configuration for simulation
Type Total Simulation Round # of Task Generator # of Resource Processing Time Maximum Queue size Initial Threshold
α
Setting 100 2 16 10.0 25 20 0.69
Notes 1 Simulation round = 100 simulation times Each resource has same conditions
For our proposed scheduling policy Experimentally optimized value for our workload prediction.
When the simulation round is finished, the system predicts a workload for the next round using equation (1). And next the system changes a threshold value of task queue capacity using equation (2). However, an exponential smoothing method requires a large volume of historical data set to get a better prediction workload result. So the system only uses a resource selection policy in 1 through 10 round. When the
(a) Result graph of task loss
(b) Result graph of throughput
(c) Result graph of average TA time
(d) Result graph of certification accuracy
Fig. 5. Result graph of simulation (RD: Random scheduling policy; RR: Round-robin scheduling policy; TL: Threshold-based Load balancing policy;WR: Workload prediction and weighted rule-based policy)
Workload Prediction and Weighted Rule-Based Task Scheduling
349
10 round is finished, system uses a full scheduling policy with accumulated historical data sets. In addition, the system also changes a task generation time of the task generator to make a heavy or light workload situation. Each resource randomly classified and configured facial information database from ORL face database [10]. So every resource has a different facial information database and shows a different certification results. We choose a random policy, a round-robin policy and a threshold-based load balancing policy as comparison models. And Fig. 5 shows a simulation result graph. We measured a number of task losses, a throughput, an average turn-around time and a certification success rate. Fig. 5 (a) is a graph chart for a number of task losses. Our proposed policy loses only one task during 100 simulation rounds. Fig. 5 (b) is a graph chart for a throughput. Our policy records about 0.77997 and that is most high throughput. It means our policy solves more certification task than other scheduling policy. Fig. 5 (c) is a graph chart for an average turn-around time. Our policy records most fast turn-around time. And last, Fig. 5 (d) is a graph chart for a certification success rate. Our policy records most high certifcation success rate. Therefore, our workload prediction based queue management policy and weighted priority rule based resource selection policy improve performance reliability of FCS on distributed parallel computing.
5
Conclusion
In this paper, we proposed the scheduling method using workload prediction based queue management policy and weighted priority rule based resource selection policy. In order to improve performance reliability, we predict a certification workload and control a threshold value of task queue capacity. And we also apply a load balancing technique to reduce a task response time and a number of task losses. In addition, we use a performance based weighted priority rule for resource selection policy. In simulation results, our proposed method shows most high efficiency in overall performance factors. Therefore, workload prediction and weighted rule-based task scheduling can provide a high reliability for face certification system on distributed parallel computing.
References 1. Jain, A., Ross, A., Pankanti, S.: Biometrics: A Tool for Information Security. IEEE Transactions on Information Forensics and Security 1(2), 125–143 (2006) 2. Rodriguez, Y., Marcel, S.: Face Authentication Using Adapted Local Binary Pattern Histograms. In: Leonardis, A., Bischof, H., Pinz, A. (eds.) ECCV 2006. LNCS, vol. 3954, pp. 321–332. Springer, Heidelberg (2006) 3. Buyya, R., Abramson, D., Giddy, J.: Grid Resource Management, Scheduling and Computation Economy. In: 2nd International Workshop on Global and Cluster Computing (WGCC 2000), Japan (2000)
350
T.Y. Kim and J.S. Lee
4. Zhang, H., Ma, H., Ming, A.: Grid-Based Multi-scale PCA Method for Face Recognition in the Large Face Database. In: Shen, H.T., Li, J., Li, M., Ni, J., Wang, W. (eds.) APWeb Workshops 2006. LNCS, vol. 3842, pp. 1033–1040. Springer, Heidelberg (2006) 5. Hamscher, V., Schwiegelshohn, U., Streit, A., Yahyapour, R.: Evaluation of JobScheduling Strategies for Grid Computing. In: Buyya, R., Baker, M. (eds.) GRID 2000. LNCS, vol. 1971, pp. 191–202. Springer, Heidelberg (2000) 6. Shah, R., Veeravalli, B., Misra, M.: On the Design of Adaptive and Decentralized Load Balancing Algorithms with Load Estimation for Computational Grid Environments. IEEE Transactions on Parallel and Distributed Systems, 1675–1686 (2007) 7. Kim, I.K., Jang, S.H., Lee, J.-S.: QLP-LBS: Quantization and Location Prediction-Based LBS for Reduction of Location Update Costs. In: Thulasiraman, P., He, X., Xu, T.L., Denko, M.K., Thulasiram, R.K., Yang, L.T. (eds.) ISPA Workshops 2007. LNCS, vol. 4743, pp. 69–76. Springer, Heidelberg (2007) 8. Zeigler, B., Moon, Y., Kim, D., Ball, G.: The DEVS environment for high performance modeling and simulation. Computational Science and Engineering, IEEE CS&E, 61–71 (1997) 9. Zegler, B.: Hierarchical, modular discrete-event modeling in an object oriented environment. Simulation 49, 219–230 10. AT&T Lab. Cambridge: The Olivetti Research ltd database of faces, http://www.cl.cam.ac.uk/research/dtg/attarchive/ facedatabase.html
A Robust Energy-Efficient Framework for Heterogeneous Datacenters Kittituch Manakul1, Simon Chong Wee See2,3, and Tiranee Achalakul1 1
Department of Computer Engineering, King Mongkut’s University of Technology, Thonburi, Thailand 2 Department of Mechanical & Aerospace Engineering, Nanyang Technological University, Singapore 3 Nvidia Corporation, Singapore [email protected], [email protected], [email protected]
Abstract. Datacenters are facilities used to house computer systems. These facilities generally consume a large amount of energy. In recent years, many researches proposed datacenter management frameworks that allow energy to be utilized more efficiently. However, most of these frameworks were limited by constraints related to unpredictable behaviors of applications in both the perspectives of execution time and power consumption. In order to provide an efficient task scheduling in datacenters, this paper proposes a preliminary concept called a robust energy-efficient framework. In this framework, a software system is deployed on top of a datacenter middleware to oversee process migrations among heterogeneous machines with various configurations. Moreover, the framework integrates additional subsystems for tracking behavioral changes of scheduled processes. During runtime, these subsystems periodically generate profiles from monitored performance metrics of processes and machines. Process profiles represent resource-usage behavior of an application, while machine profiles represent resource-provisioning behaviors. Processes can be moved around on the fly based on information provided in these profiles. The proposed framework takes advantage of heterogeneity along with process migration to improve energy efficiency of a datacenter without prior knowledge on process behavior and resource usage fluctuation in users’ applications. Keywords: Energy-efficient, Heterogeneity, Process Migration.
Various techniques have been developed in order to make datacenters utilize energy more efficiently. Xiaofei et al. [1] classified energy-efficient techniques into two groups – local and cluster-wide techniques. Local techniques are concerned with lowering energy used in an individual machine or device. Examples of these are dynamic voltage frequency scaling (DVFS) and DLS [2]. Nevertheless, these local techniques are less efficient than turning off the machine. Unlike the local techniques, cluster-wide techniques consider a global image of the system. Because of modern virtualization technology, this class of energy-efficient techniques is sub-divided into two types – physical-machine-based (PM-based) and virtual-machine-based (VM-based) techniques. Both types aim at reducing the number of physical machines by defining a minimal set of machines that could handle a given workload. However, for PM-based techniques, it is possible that the minimal set could not handle fluctuation in applications’ resource demands at runtime. For instance, if a machine cannot support the increasing workload, the system will have to start up more machines to serve the increment. Consequently, this datacenter will consume more energy. VM-based techniques help address this problem as illustrated by Anton’s work [3]. Recent literatures proposed many frameworks for more energy-efficient datacenters. These frameworks combined both local and cluster-wide techniques together, and also integrated policies in order to gain control over energy utilization of the system. Examples are work of Andreas et al. [4] and “EnaCloud” [5]. It is noticeable that both frameworks share some common steps that help datacenter utilizes energy more efficiently. The first step is to examine characteristic of applications’ workload. In this step, the frameworks learn applications’ resource usage behavior so that, later, it will be possible to optimize utilization of components in the datacenter system. This workload characterization is widely studied and implemented as in [6-9]. The second step is to evaluate energy efficiency. The mentioned frameworks use energy-efficiency metrics to determine minimum number of machines and active components in a datacenter. Among many kinds of these metrics, “Performance per Watt” [10] and energy delayed product (EDP) [11] are used. Not only both metrics can evaluate energy-efficiency of the system, but also they extend flexibility of the framework to support performance or energy policies. For instance, in the situation that service performance has to be maintained, energy utilization can be scaled to stabilize the fixed “Performance per Watt”. The last step is to move tasks (i.e. processes, jobs or applications) to more optimized set of machines. The task migration is a key to improve the energy-usage efficiency because the system is able to consolidate workload by moving separated tasks to run on the same machine. However, task migration can be performed in many abstraction levels, e.g. process-level, VM-level. There are tradeoff issues between different levels of migration; nonetheless, [12] provides some evidence that process migration is faster than VM migration. Obviously, frameworks in [4, 5] will become more important in the upcoming energy-efficient datacenters. However, they miss two important factors, which
A Robust Energy-Efficient Framework for Heterogeneous Datacenters
353
directly affect energy efficiency of the datacenter system. One is that they assume that resource demand is known prior to the application’s runtime, and resource fluctuation is not taken into account. This may not be true for every application in the datacenter because some users, e.g. doctors or scientists may not know these technical details. Another is that performance of an application depends on specifications of the machine on which it is running. For instance, a program, which is data-intensive, will run more slowly on a machine with high-speed processor than on a machine with high-speed memory subsystem. Additionally, Byung-Gon et al. [13] emphasized this fact and showed advantages of having more than one set of similar machines in the datacenter. Despite the advantages from having more than one class of machines in the datacenter, none of the former energy-efficient frameworks takes benefits of this heterogeneous structure. It is obvious that heterogeneity together with process migration will improve energy efficiency of datacenters. Both techniques together help relax the constraint of knowledge required before runtime, and at the same time, improve energy efficiency of datacenters because the adjusted datacenter can dynamically change process placement at runtime. In this paper, a new framework, which allows applications to run on any available machine at their arrival time, is proposed. Later, the framework dynamically generates placement scheme for all processes in the system. This scheme should map a process to a certain machine whose configuration is suitable for the process’s activities. After executing the placement plan via process migration, the datacenter will be in the more energy-efficient state. With this workflow, machine configurations are taken into account for the improvement of energy-efficiency. The rest of this paper is organized as followed. Section 2 defines architectures and explains major workflow of the proposed framework. Then, Section 3 and 4 describe profiling techniques and placement scheme generation, respectively. Section 5 describes open challenges, and Section 6 summarizes main points of the work.
2
Framework Architecture
This paper introduces a robust framework for energy utilization improvement of a heterogeneous datacenter. The main idea is to adopt process migration when a more energy-efficient placement scheme is discovered. During runtime, process resource utilization and machine configuration are periodically captured, analyzed, and used to construct a predictive model for energy-efficient process scheduling scheme. The predictive model consists of two set of mathematical equations: “Process Profiles” describing how each process consumes machines’ resources, and “Machine Profiles describing how the performance and power consumption of a certain machine would change when the machine is assigned for a specific set of processes. When a process migration is required, these profiles together with the energy policy are used to generate a new placement scheme for the system. Fig. 1 shows modules within our energy-efficient framework.
354
K. Manakul, S.C.W. See, and T. Achalakul
Fig. 1. Energy-efficient Framework Design
The framework consists of five components. Monitoring module captures performance metrics related to process resource utilization and machine configurations. Example of the measurement metrics are “Percent User Time”, “Percent System Time” and “Number of context switches”, “Working (or Resident) Set Size”, “Number of page faults”, “Number of Read/Writes to/from I/Os” and “Number of Cancelled Writes”. Once these raw data are captured, Monitoring module analyzes those data and feeds both raw data and analyzed results to profiling module. Then, profiling module synthesizes process and machine profiles. Event Detection module keeps track of profiles and compares the new profiles to the older ones. When the difference between profiles is sufficiently large, the generation of a new placement scheme is triggered. Once completed, the new scheme is sent to Process Migration module, which is responsible for handling actual migrations. In addition, the framework attempts to prevent the Ping-pong effect occurs when processes are migrated around indefinitely, by dividing the simulation into time slots. Time slots allow metrics to be analyzed using control theory such as PI controller. Therefore, as soon as possible, the processes will work steadily at the anticipated performance and power consumption. For each time slot, the workflow described in Fig. 1 is executed only once. Moreover, since processes in the datacenter may change their behavior during their runtime, accuracy of the process profiles may be decreased with time. The proposed framework is then configured to regenerate each process profile in every time-slot. The following sections provide the details of the two main modules in the proposed framework, namely, profiling and placement scheme generation.
3
Process and Machine Profiling
Profiling is the key to success of the proposed framework. In order to predict performance of a particular process and power consumption of a machine, mathematical models must be constructed. In addition, not only these models must be precise, but also the modelling process should be done within a certain period of time. We thus proposed the use of primitive processes, called dwarfs [7], for profiling purposes.
A Robust Energy-Efficient Framework for Heterogeneous Datacenters
355
Each dwarf is a parallel process representing unique patterns of computation and communication. These patterns are observed from characteristics and types of data, communication between processors, frequency and density of communication. Currently, there are 13 dwarfs covering scientific, games, and machine learning applications [7]. The concept of the proposed profiling method based on Dwarfs and a profiling example are described next. 3.1
Profiling with Dwarfs
The Dwarf concept allows behaviors of various types of applications to be quickly profiled based on a combination of primitive processes.
Fig. 2. Prediction and profiling steps
Fig. 2 shows steps in deriving predicted performance and power consumption metrics using dwarfs. In the first step, resource usage patterns of a particular process (e.g. how much memory the process wants to allocate, how frequent the process writes to disk) are determined from its profile. Then, the predicted performance and the power consumption of a given process are obtained from a machine profile of a targeted machine. Extracting runtime behaviors of a process can be cumbersome. However, behaviors of a larger process can be described using smaller processes. Therefore, an application can be profiled based on a combination of some primitive processes. In order to combine primitive processes, mathematical models reflecting behavior of these tiny processes have to be derived. Fortunately, each of these processes is small enough that deriving the models from its source code is feasible. To construct a process profile, the performance according to a predefined set metrics of a targeted process, are measured at runtime. A combination of primitive processes that provides similar predicted performance is then identified and used as a process profile. Once the profile is determined, mathematical models that describe behavior of each selected primitive process can be merged and used to define behaviors.
356
K. Manakul, S.C.W. See, and T. Achalakul
Machine profiles, on the other hand, describe resource-provisioning behaviors and limitations of an individual machine. A machine profile can be constructed following three main steps as described below. First, to determine maximum effective resource capacity of the machine, a certain type of primitive processes that has outstanding behavior in one physical component, such as a pure processor-bound process, will be assigned to run on the machine in one time slot. Maximum number of the chosen processes, which the machine can handle, is the limitation of the machine capacity in the chosen resource type (machine’s component). Second, to determine resource-provisioning behavior of the machine, selected set of combinations of mixed primitive processes types is assigned to run on the machine, performance metrics of each running process will be captured. The different in performance metrics compared to metrics obtained from running each individual process defines interactions between processes and the machine itself. In this step, there are many alternatives to determine these interactions, such as partial factorial statistical experiment, neuron networks, etc. Third, to determine how much throughput a certain type of primitive processes obtained when running on the machine, one single process of each type of primitive processes is assigned to run on the machine one after another. The throughput is, then, calculated from the total amount work done by the running primitive process in a time slot divided by time in one slot. This throughput will be used to approximate power consumption of the new processes’ placement as mentioned earlier. Once determined, the obtained knowledge together with resource usage behavior of each primitive will be used as parameters in mathematical models, which describe machine resource provisioning and power consumption of any process when running on this machine. Moreover, running different numbers and types of primitive processes on a machine will determine biases occurred in resource provisioning by one or more types of primitive processes. These biases may result in probabilities of resource provisioning which allow us to calculate steady-state quantities via Markov’s chain.
3.2
Profiling Example
This section presents an example of how to profile processes and machines with mathematical models. Suppose an interesting Process X is currently scheduled to run on Machine A. At this moment, Process X consumes resources as shown in Fig. 3(a). Note that only three resource usage parameters, i.e. percentage processor usage, memory usage and I/O bandwidth, are chosen to simplify this explanation. Additionally, for simplification, suppose there are only two primitive processes – Primitive I and J. Each primitive can be configured to adjust its resource usage. Fig. 3(b) and Fig. 3(c) show amount of resources used by Primitive I and J, respectively. Let RP,C(M) denote resource usage vector of Process P with configuration C running on Machine M. Equation 1 draws a conclusion from Fig. 3. RX (A) = 2 * R I,α (A) + RJ,β (A) .
(1)
A Robust Energy-Efficient Framework for Heterogeneous Datacenters
357
Fig. 3. Resource utilization of example processes (a) Process X, (b) Primitive I and (c) Primitive J
Note that the linear relationship in the equation above is for simplifying subsequent equations. Suppose VX represents the number of primitive components for Process X. In this situation, VX is [2 1]. This vector represents the process profile of Process X. To predict resource usage behavior of Process X, if DP,C (M) denotes resource usage behavior (demand) vector of process P with configuration C running on machine M, resource demand of Process X can be written as Equation 2. DX (M) = 2 * DI,α (M) + DJ,β (M) .
(2)
However, performance of a process depends on real resource utilization not the resource usage behavior predicted by the above equation. Thus, machine profiles play significant roles to convert these demands into real performance metrics. Let PQ,C (R) denote performance metric vector of process Q with configuration C utilizing R amount of resource on a particular machine, and suppose that a Machine profile determines that primitive component I and J of Process X should use resource RI and RJ. Real performance metrics of the Process X is shown in Equation 3. PX (RX) = 2 * PI,α (RI) + PJ,β (RJ) .
(3)
Furthermore, let RM (VTotal,VP ) be machine profile of machine M taking total number of each primitive on that machine VTotal and the profile of Process P VP. Therefore, for Process X to run on another machine B, resource utilization of each component in the profile of Process X can be written as Equation 4. RX = [RI RJ] = RB (VTotal,VX) .
(4)
For power consumption, according to the work of Ricardo et al. [14], power consumption of an application is linearly dependent with its throughput. Therefore, linear relationship between each primitive process’s throughput and its power consumption is determined during the machine’s inspection period. Thus, let QP,C,M (T) denote power consumption of process P with configuration C resulting in throughput T when running on machine M, performance of the Process X running on machine B is defined by Equation 5. QX,B (TX ) = 2 * QI,α,B ( TI ) + QJ,β,B (TJ) .
(5)
358
K. Manakul, S.C.W. See, and T. Achalakul
In Equation 5, TI is a vector of throughput components in RI, and TJ is a vector of throughput components in RJ. These prediction equations will be used in Placement Scheme Generation module to determine the more energy-efficient placement scheme for the system. To sum up, primitive processes allow both process and machine profiling to bypass complexity in source-code-based analysis of larger applications. They also bypass the need to explore the underlying hardware and operating system. Thus, primitive processes help balance the trade-off between precision and meeting time-constraint in profiling module of the proposed framework.
4
Placement Scheme Generation
Placement Scheme Generation module is responsible for determining where to place each process in the system. After receiving trigger from Event Detection module, this generator will solve optimization problem in order to find an optimized placement scheme that transfers the system to the approximately best energy-efficient state. Fig. 4 illustrates inputs and outputs of the optimization problem. The algorithm will be based on a metric for energy efficiency called “Performance per Watt”. This metric allows administrators to specify energy and/or performance policies of a datacenter. For example, an administrator may want every machine in the system to utilize 300 Watts, or an application to consume at most 80% of available resources. These policies are taken into accounts when optimizing a system. The proposed framework takes profiles generated in Profiling module to estimate performance metrics of processes and power consumption of machines in the datacenter. Moreover, it also takes into account cost of power loss and performance degradation due to process migration. When the problem solving is done and the resulted scheme for relocation of processes is generated, the module passes the output to Process Migration module.
Fig. 4. Optimization problem of process placement
5
Open Challenges
Although this framework is designed to overcome the datacenter’s problems where variation in process’s resource usage behavior is unpredictable, there are some design issues that affect efficiency of the whole system.
A Robust Energy-Efficient Framework for Heterogeneous Datacenters
359
One of the challenges is identifying an appropriate time slot. The purpose of having time divided into slots is to lessen chance of triggering placement scheme generation, which will results in system inconsistency. For example, there may be an event that one process is migrated indefinitely or an event that a process bounces between two machines. However, over-extending time period per slot will make the framework lose track of changes in behaviors of processes. In contrast, overshortening the time will result in precision loss of profiling performance. Another challenge is the speed of a placement scheme generation. It is obvious that this multiple-objective optimization problem compares between more than one aspect of performance metrics. Not only taking into account the profiles and degradation due to migrations, datacenter usually operates under SLA with customers. In some cases, performance will be more significant than less energy consumption. The optimization algorithm must be able to customize its solution due to these administrative requirements.
6
Conclusion
This paper provides a concept for developing a robust energy-efficient framework for datacenters. The framework emphasizes the optimal energy-efficient placement scheme generation through process migration in a heterogeneous datacenter. First, the framework monitors resource usage of processes in the system. Monitored parameters are used for process profiling through combinations of well-study primitive processes, called Dwarfs. Each combination represents the resource usage behavior of a particular process in the system. Determining resource usage behaviors allow the additional subsystems to detect changes in resource demand and significant events such as application arrival, process termination, etc. By using the profiles, if the framework detects changes, it will estimate values of performance metrics and power consumption in order to determine suitability of process migration. As a result, the designed framework will generate a new placement scheme for processes in the system based on the predicted values. Placement scheme generation also considers other aspects such as overhead of process migration and performance and energy policies from system administrators. Once generated, the placement scheme can be used for actual process migration.
References 1. Liao, X., Hue, L., Jin, H.: Energy Optimization Schemes in Cluster with Virtual Machines. Cluster Computing 13(2), 113–126 (2010) 2. Kim, E.J., Yum, K.H., Link, G.M., Vijaykrishnan, N., Kandemir, M., Irwin, M.J., Yousif, M., Das, C.R.: Energy Optimization Techniques in Cluster Interconnects. In: Proceedings of International Symposium on Low Power Electronics and Design (ISLPED 2003), pp. 459–464. ACM, New York (2003)
360
K. Manakul, S.C.W. See, and T. Achalakul
3. Beloglazov, A., Buyya, R., Lee, Y.C., Zomaya, A.: A Taxonomy and Survey of EnergyEfficient Data Center and Cloud Computing Systems. In: Technical Report, CLOUDS-TR2010-3, June 30. Cloud Computing and Distributed Systems Laboratory, The University of Melbourne, Australia (2010) 4. Merkel, A., Stoess, J., Bellosa, F.: Resource-conscious Scheduling for Energy Efficiency on Multicore Processors. In: Proceedings of the 5th European Conference on Computer Systems (EuroSys 2010), pp. 153–166. ACM, New York (2010) 5. Li, B., Li, J., Huai, J., Wo, T., Li, Q., Zhong, L.: EnaCloud: An Energy-saving Application Live Placement Approach for Cloud Computing Environments. In: IEEE International Conference on Cloud Computing (CLOUD 2009), pp. 17–24 (2009) 6. Merkel, A., Bellosa, F.: Task Activity Vectors: A New Metric for Temperature-aware Scheduling. In: Proceedings of the 3rd ACM SIGOPS/ EuroSys Conference on Computer Systems (EuroSys 2008). ACM, New York (2008) 7. Asanovic, K., Bodik, R., Catanzaro, B.C., Gebis, J.J., Husbands, P., Keutzer, K., Patterson, D.A., Plishker, W.L., Shalf, J., Williams, S.W., Yelick, K.A.: The Landscape of Parallel Computing Research: A View From Berkeley. In: Technical Report, UCB/EECS-2006183. Electrical Engineering and Computer Sciences, University of California at Berkeley, USA (2006) 8. Che, S., Boyer, M., Meng, J., Tarjan, D., Sheaffer, J.W., Lee, S.H., Skadron, K.: Rodinia: A Benchmark Suite for Heterogeneous Computing. In: Proceedings of the IEEE International Symposium on Workload Characterization (IISWC 2009), pp. 44–54 (2009) 9. Hoste, K., Eeckhout, L.: Microarchitecture-independent Workload Characterization. IEEE Micro. 27(3), 63–72 (2007) 10. Laudon, J.: Performance/Watt: The New Server Focus. In: ACM SIGARCH Computer Architecture News - Special issue: dasCMP 2005, vol. 33(4). ACM, New York (2005) 11. Gonzalez, R., Horowitz, M.: Energy Dissipation in General Purpose Microprocessors. IEEE Journal of Solid-State Circuits 31(9), 1277–1284 (1996) 12. Wang, C., Mueller, F., Engelmann, C., Scott, S.L.: Proactive Process-Level Live Migration in HPC Environments. In: Proceedings of the 2008 ACM/IEEE conference on Supercomputing (SC 2008). IEEE, New Jersey (2008) 13. Chun, B.G., Iannaccone, G.: An Energy Case for Hybrid Datacenters. ACM SIGOPS Operating System Review 44(1), 76–80 (2010) 14. Koller, R., Verma, A., Neogi, A.: WattApp: An Application Aware Power Meter for Shared Data Centers. In: Proceedings of the 7th International Conference on Autonomic Computing (ICAC 2010). ACM, New York (2010)
A Study on the Electrical Properties of the BaTiO3 Ceramic Thin Films Ik-Tae Im1 , Hyun Seob Jo2, Sung Hoon Oh3, and Byung Moon So4 1
Division of Mechanical Design Engineering, Chon-buk National University, Dukjin-dong, Jeon-ju city, Jeollabuk-do, 561-756 [email protected] 2 Department of Electronics Engineering, Chung Woon University 29, NamJang-Ri, HongSung-Ub, HongSung Kun, ChoongChungnam-do, 350-701 [email protected] 3 Department of Mechanical System Engineering, Chonbuk National University, Dukjin-dong, Jeon-ju city, Jeollabuk-do, 561-756 4 Department of IT Applied System Engineering, Chon-buk National University, Dukjin-dong, Jeon-ju city, Jeollabuk-do, 561-756 [email protected] Abstract. Polycrystalline BaTiO3 thin films have been prepared on slid grass substrate by RF/DC magnetron sputtering equipment. The films were deposited at evoporator system and were annealed at heattreatment. The films had a dense microstruture with fine grains. The electrical properties of the films were dramatically controlled with annealing. Samples Preparation were analyzed in term of PTCR(positive temperature coefficient of Resistivity) Samples were made in the substrate tempera-true of 400oC deposition time of 10 hours,and forward power of 210watt. R-T(resistivity-tempera-true) Characteristics of the samples were investigated as a function of the substrate type and the ambient temperature. The resistivity of the thin film specimens was compared with that of the bulk type specimens. By using RF/DC magnetron sputtering system, we obtained lower resistivity in the thermistor with thin BaTiO3 film than that in the bulk type thermistor. Keywords: BaTiO3, Thin Film, Thermistor.
Lots of studies have been devoted to find the possibility of thin film thermistor which uses the grain boundary in principle but it has not been fully understood. Thin film positive temperature coefficient (PTC) thermistor was first proposed by Feldman[4] in 1954. He used evaporation in vacuum to deposit BaTiO3 thin films. After his study, BaTiO3 thin films were deposited from various techniques such as using solution[5], DC sputter[6] and RF sputter[7]. PTC semiconducting film was achieved by adding Y2O3 into BaTiO3 in 1972. Thin films having PTC characteristics could be obtained from the optimized process conditions in case of RF sputter. However, it is difficult to use the PTC films in mass production chips because of its high electrical resistivity. In order to obtain good electrical characteristics, SiO2 is added to BaTiO3 for fabricating PTC thermistor in this study. The optimum process conditions of RF/DC magnetron sputtering were found at first, and then thin films of BaTiO3 with SiO2 PTC thermistor were deposited from the optimized conditions. The electrical characteristics and dielectric constants according to the wave length were investigated. The films were heat treated before measuring the characteristics.
2
Experiments
2.1
Targets and Thin Films
Targets were prepared to obtain desired composition of BaTiO3 films. The final composition of the target, Ba0.907Y0.003Ti1.01O3.02+0.5wt%SiO2, was determined from the trial-and-error method by examining the contents of the deposited thin films. Raw materials used in the target are listed in table 1 with the suppliers. Manganese, which was usually added to increase the change of resistance, was not considered in this study. Targets were prepared using normal PTC ceramic process as shown in Fig. 1 using high purity TiO2 and BaCO3. Uniform size particles, obtained from sieve shaker of 45mm mesh after the reaction, were pressed with high pressure of 7000kg/cm2 and 3000kg/cm2 in a cylindrical mold with two different diameters. Polyvinyl alcohol was used as a binder. Targets were sintered at 1350 for two hours in ZrO2 surrounding powders. Sputtered target was heated for 20 minutes at 800 after silver paste had been printed to the both sides of the target. This target, 5.08cm diameter was used as specimen and an electrode. Silver plates as electrodes which is needed to measure the electrical characteristics was deposited on both sides of the thin film disc in a vacuum under 5.7×10-5torr.
℃
℃
Table 1. Purity and suppliers of raw materials for the target
Raw materials
Purity(%)
Manufacturer
BaCO3
99.999
Junsei chemical (Japan)
TiO2
99.9
Junsei chemical (Japan)
Y2O3
99.9
Strem Chemical (USA)
SiO2
99.99
Hayashi Pure Chemical (Japan)
A Study on the Electrical Properties of the BaTiO3 Ceramic Thin Films
363
Fig. 1. Diagram for the processing of a target
2.2
Sputtering
Thin films of the PTC BaTiO3 were deposited using a RF-magnetron sputter. Distance between the target and the substrate was set as 45mm during the process. Characteristics of the PTC thermistor are very sensitive to the amount of additives, temperature and surrounding gases. Sputter conditions such as a working gas, pressure, temperature of a substrate, sputter time and electrical power also affect the characteristics of the thin films. Sputtering conditions used in the experiment are listed in Table 2. Table 2. Sputtering conditions used for a preparation of specimens
Parameters Vacuum (torr) Chamber pressure (torr) Target(T) diameter (cm) Substrates(S) (mm) T-S distance (mm) Pre-sputtering (hr) Forward PWR (W) Reflect PWR (W) Gas flow rate, O2 (sccm) Gas flow rate, Ar (sccm) Sputtering time (hr)
Conditions Less than 8×10-6 5×10-3 5.008 25×75 for slide glass and Si wafer 22×35 for Al2O3 45 1.0 150 2 10 10 10
364
2.3
I.-T. Im et al.
Electrodes
In order to measure the resistivity – temperature characteristics on a substrate, the electrodes were prepared using thermal evaporation. Fig. 2 shows the structure of the electrodes. Gold was coated up to 200Å to see the surface structure using DC sputter at 140V, 6.0mA for 4 minutes.
(a) the horizontal direction and (b) the vertical direction. Fig. 2. Structure of the electrodes
3
Results and Discussion
Bulk specimen of a target was put into the closed bath maintained at constant temperature in order to measure the resistivity – temperature characteristics. Temperature of surrounding of the specimen was heated by Joule heating. The same technique was used to the thin films except using a shield box to remove noise. Fig. 3 shows the current values when the step voltage from -9.0V to 9.0V was applied to the specimen. The applying voltage was first increased from 0.0 to 9.0V then decreased to -9.0V and increased again to the initial value. The specimen was heat treated at different temperature, 600 , 800 , 1000 and 1200 before measurement. The output currents were very similar even at the different heat treatment temperatures as shown in Fig. 3(a) 600 and Fig. 3(b) 1000 . It can be seen that the current to the voltage makes a Hysterisis loop for all temperatures. The current-voltage curve for the bulk shown in Fig. 3(c) also shows the Hysterisis loop below 0.0V.
℃ ℃
℃
(a) 600
℃
℃
(b) 1000
℃ ℃
℃
(c) bulk
Fig. 3. Current-voltage curves for the specimen with different heat treatment temperature and the bulk
A Study on the Electrical Properties of the BaTiO3 Ceramic Thin Films
365
Fig. 4. Dielectric constant variation according to the photon energy
Fig. 5. Refractive index and loss according to the variation of the wave length
Fig. 6. Dependency on frequency for the absolute impedance
Fig. 4 shows the dielectric constants and dielectric loss function according to the variations of the photon energy. The dielectric constant showed increasing tendency when the photon energy was increased from 1.6eV to 2.5eV. The constant was then decreased when the photon energy was more increased over 2.5eV. However, it was rapidly increased when the photon energy was larger than 3.5eV. Fig. 5 shows the refractive index when the photon energy was given to the specimen with the same incident angle of 70 degree as of the dielectric constant measurements. The results shown in Fig. 5 are identical to those of Fig. 4 when ħ=h/2=1.054×10-34J⋅s.
366
I.-T. Im et al.
Fig. 7. Dependency on frequency for the admittance
Fig. 6 and Fig. 7 show the variations of impedance, admittance and phase angle according to the frequency. The frequency was varied from 100Hz to 14.0MHz, and the bias and oscillation levels were 1.0V and 1.0Vrms. The data were obtained by the impedance/grain-phase analyzer (4194A, HP). Impedance curve showed an abrupt change near 1.0MHz and this was attributed to the reactance change. Admittance variation shown in Fig. 7 shows the inverse values of the impedance. Conductance and susceptance were measured by the same analyzer to the impedance and shown in Fig. 8. Like the impedance and admittance, we can also see the discontinuities near 1.0MHz in case of the susceptance. Considering the results shown in Figs. 6, 7 and 8, reactance of the film seem to change abruptly over the particular frequency.
Fig. 8. Dependency on frequency for the conductance and susceptance
4
Conclusion
In this study, thin films having composition of (Ba0.997 Y0.003 Ti1.03 O3.02 + 0.5wt%SiO2) were deposited using RF/DC magnetron sputter and the electrical characteristics were analyzed. From the analyses results, we can conclude the followings.
A Study on the Electrical Properties of the BaTiO3 Ceramic Thin Films
367
1) Dielectric constant varies largely depending on the photon energy. It shows increasing-decreasing-increasing tendency. 2) Refractive index also varies according to the wave length. It shows rapid drop in short wave length region. 3) Impedance, admittance and susceptance have discontinuities around 1.0MHz and this is attributed to the abrupt change of the reactance.
References 1. Anderson, W.A.: Electrical and Dielectric Properties of Thin Film BaTiO3 Capacitors deposited by Ratio Frequency Magnetron Sputtering. J. Vac. Sci., Technal. A10(4), 733–736 (1992) 2. Anderson., W.A., et al.: Effect of Barrier Layer on BaTiO3 Thin Film Capacitors on si substrates. J. Elece, Mat. 23(1), 53–56 (1994) 3. Shintani, Y., et al.: Behaviors of High-Energy Electrons and Neutral Atom in the Sputtering of BaTiO3. Jpn. J. Appl. Phy. 14(12), 1875–1879 (1975) 4. Felderman, C.: Formation of Thin Films of BaTiO3 by Evaporation. Rev. Sci. Instrum. 12(6), 463–466 (1955) 5. Dervies, R.C.: On the Preparation of the Single-Crystal Films of BaTiO3. J. Am. Ceramic. Soc. 45, 225–228 (1962) 6. Shintani, Y., et al.: Preparation of the BaTiO3 Films by DC Diode Sputtering. J. Appl. Phys. 14, 2376–2380 (1970) 7. Nagatomo, T., et al.: Fabrication of BaTiO3 Films by Planar Magnetron Sputtering. Ferroelectrics 37, 681–684 (1981)
Gait Recognition by Multidimensional Representation for Accumulated Silhouette Seungdo Jeong1 and Jungwon Cho2, 1
Research Institute of Electrical and Computer Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul, 133-791 S. Korea [email protected] 2 Department of Computer Education, Jeju National University, 102 Jejudaehakno, Jeju-si, Jeju-do 690-756 S. Korea [email protected]
Abstract. Gait is defined as a style of walking and the gait recognition is to recognize individual by using gait image sequence. Most studies about gait recognition use silhouettes which are extracted from gait image sequence because shape information included in the silhouette is more useful for recognition than others. In this paper, we propose gait recognition method using multidimensional representation for gait silhouettes. This paper focuses on the cyclic characteristics of gait. Thus we propose the method to form the accumulated silhouette regarding the cyclic characteristics and then describe those as multidimensional representation. In order to recognize individual using the multidimensional representation for the accumulated silhouette, we adopt tensor decomposition. We verify the superiority of the proposed approach via experiments with real gait sequences. Keywords: Gait Recognition, Accumulated Silhouette, Multidimensional Representation, Tensor, N-mode SVD.
1
Introduction
Gait is a term used in definition for walking style [1]. Many researches to recognize individual using the gait have been reported because gait images are obtained more easily comparing to other biometrical information such as fingerprint, iris, and so on [2]. Most of these researches use silhouettes which are extracted from image sequences. Motion information is also used, but, shape information is more useful for recognition because the silhouette is shape information itself, and silhouette sequence includes motion characteristics [3][4]. However, most of silhouette-based gait recognition only use simple features extracted from silhouettes fragmentarily. To reflect integrated characteristics, more sophisticated analysis is required. Tensor analysis has already been used to analysis more complex characteristics in face recognition [6][7]. Tensor decomposition is powerful algebraic analysis
Corresponding author.
T.-h. Kim et al. (Eds.): GDC 2011, CCIS 261, pp. 368–375, 2011. c Springer-Verlag Berlin Heidelberg 2011
Gait Recognition by Multidimensional Representation
Fig. 1. Edge sequence of gait silhouettes
369
Fig. 2. Variation of stride length
in the recognition problem where several factors influence on the performance [5]. Gait recognition problem is also included in this category. This paper takes attention to the cyclic characteristic and respective features in certain interval, and thus employs tensor algebra to analyze shape itself and the respective features in certain interval simultaneously.
2 2.1
Multidimensional Representation for Gait Silhouettes Accumulated Gait Silhouettes
It is seldom to use frames separately in silhouette-based gait recognition because simple feature extracted from specific pose is insufficient to recognize individual. Instead, many researches used integrated information extracted from all of gait sequence. In this paper, we compose the accumulated gait silhouette to reflect entire characteristics of gait sequence. The accumulated silhouette means a frame of silhouette, which is made by cumulating several silhouettes to a single frame. Fig. 1 shows edge sequence of silhouettes extracted from gait image sequence. As shown in the figure, gait sequence has a certain periodicity. That is identified definitely from checking the variation in stride length as shown in Fig. 2. Such characteristics of periodicity and/or variation caused by motion may be represented as features by analyzing frames within certain interval all together. Thus, in this paper, we divide gait sequence as several sections according to stride length and then cumulate silhouette on single frame per each section to form the accumulated gait silhouette. Criterion to split sequence into sections is based on peaks of the stride length variation in Fig. 2. In one cycle of gait, there are two pairs of peaks; two high-peaks and two low-peaks. Therefore, we divide the cycle into four sections depending on peak points as presented in Fig. 3 and then cumulate all of silhouettes in each section separately. We also cumulate all of silhouettes within one cycle to reflect average characteristic of gait cycle. The number of frames in a section varies according to walking speed. In order to reduce influence on the accumulated silhouette caused by difference number of frames, pixel depth of accumulated silhouette is scaled into the predefined interval. Fig. 3 shows sections, totally accumulated silhouette, and the accumulated silhouettes in each section for one cycle of gait all together. Size of the accumulated silhouettes are normalized as 80 × 120 in pixels.
370
S. Jeong and J. Cho
Fig. 3. Accumulated silhouettes in each section
2.2
Multidimensional Representation
Through the process mentioned in Section 2.1, five accumulated silhouettes for individual are created. To describe these accumulated silhouettes for all people at once, we adopt multidimensional representation or tensor. Tensor used in this paper for these accumulated silhouettes consist of three independent axes; pixels, segments, and people. Pixel axis represents the accumulated silhouette image as a column vector. We show the accumulated silhouette as two-dimensional image for convenience in Fig. 4. Segment axis is to classify the accumulated silhouette of each section. The last one is people axis which represents individual. As a result, tensor used in this paper is three-dimensional tensor D defined as equation (1). Size of D is 9600 × 5 × 16, that is, the size of the accumulated silhouette is 9600 and 5 segments per each person are created for 16 people.
People
Segments
Pixels
Fig. 4. Multidimensional representation for the accumulated silhouettes
2.3
Tensor Analysis
Tensor data D for training gait sequence is defined as equation (1). D = Z ×1 Upixels ×2 Usegments ×3 Upoeple .
(1)
Where, Un means n mode matrix and ×n represents mode-n product. Z is known as core tensor which governs the interaction between the mode matrix U1 , U2 , . . . , Un . The gait image ensemble proposed in this paper is composed
Gait Recognition by Multidimensional Representation
371
of factors corresponding to pixels, segments, and people, respectively. Un spans D(n) matrix which is result by mode-n flattening of D. We use N-mode SVD algorithm to compute core tensor Z [6]. D(1) is mode 1 fattening matrix whose size is 9600 × 80. It is hard to apply the SVD algorithm directly to the matrix due to large size of column vector. Thus we compute DT(1) D(1) in the SVD as equation (2), and then compute U1 using equation (3), where Σ+ means the pseudo inverse of Σ [6]. DT(n) D(n) = Vn ΣUTn Un ΣVnT = Vn Σ2 VnT .
(2)
Un = D(n) Vn Σ+ .
(3)
U1 computed through above computation is 9600 × 80 matrix which is composed of coefficients reflecting influence of each pixel. D(2) and D(3) are 5 × 153600 and 16 × 4800 matrixes, respectively. As similar to the case of D(1) , it is also hard to compute them in the SVD directly. However, U1 and U2 are obtained by computation of D(2) DT(2) and D(3) DT(3) in the SVD as equation (4) [6]. D(n) DT(n) = Un ΣVnT Vn ΣUTn = Un Σ2 UTn .
(4)
U2 and U3 are set of coefficients determining segments and people with size of 5 × 5 and 16 × 16 matrixes, respectively. Core tensor Z is 80 × 5 × 16 tensor as a result of mode-n product in equation (5). Z = D ×1 UTpixels ×2 UTsegments ×3 UTpeople .
3 3.1
(5)
Proposed Gait Recognition Method Composing Basis for Recognition
The purpose of this paper is to recognize the individual based on unknown gait sequence. From thorough observation for the equation (1), mode matrix related to the people factor is Upeople , thus, Z ×1 Upixels ×2 Usegments is construed as the basis. In other words, tensor data D is equal to mode-3 product between basis tensor B whose size is 9600 × 5 × 16 and U3 that is 16 × 16 people factor as shown in equation (6), where row vector of U3 is 16 coefficients to represent individual. D = B ×3 Upeople .
(6)
B = Z ×1 Upixels ×2 Usegments.
(7)
In recognition phase, first, the accumulated silhouette d is extracted from inputted unknown gait sequence. The coefficient matrix C of people factor for unknown d is computed as equation (8), where C is 1 × 5 × 16 tensor or 5 × 16 matrix. C = P ×pixels d.
(8)
372
S. Jeong and J. Cho
P is the pseudo inverse tensor of B. P(1) is computed as equation (9) and then P is obtained through making tensor from P(1) . T P(1) = (B+ (1) ) .
3.2
(9)
Similarity Computation
C is 5 × 16 matrix as mentioned in Section 3.1. Similarity computation for gait recognition is accomplished with row vectors of 5 × 16 coefficient matrix extracted from unknown gait sequence and row vectors of Upeople as equation (10). As a result, similarity vector si has 5 values and s is 5 × 16 matrix. si = sim(C, Upeople (i)),
where Upeople (i) is i−th row vector of Upeople . (10)
The row vectors of C are interpreted as the respective responses according to 5 segments. However, Upeople means characteristics related to individual regardless of segments. Thus, we choose two types of values from similarities between unknown gait and user coefficient for recognition. The first is the maximum value among elements of si . If maximum among elements of sk is the largest value comparing to ones of all similarity vectors except sk , we can identify the unknown gait is person index k. This is the recognition result as nearest neighbor search. The second choice is mean value of si because responses related to segment of input gait to the same person is relatively higher than the others due to similar shape of silhouette even though segments is not the same. The similarity employed in this paper is the inverse of the Euclidean distance between vectors.
4
Experimental Results
To evaluate the performance of the proposed method, we compare the results with the one using the PCA. We obtained gait sequence for 16 people two times at different days; one set is used for tensor analysis and PCA, the other is for test. The preliminary experiment to evaluate the performance of the PCA is executed with only totally accumulated silhouette. The input data for the recognition test is also the accumulated silhouette extracted from the other gait set. As a result, the recognition rate through the nearest neighbor search for 16 people is 68.75%. In the regular experiment, all of segments extracted from another gait set are used to evaluate the performance. Fig. 5 and Fig. 6 show the similarity responses which are the inverse of the Euclidean distance in the recognition experiments employing the PCA and the tensor analysis, respectively, for person #5.
Gait Recognition by Multidimensional Representation
(a) Response of segment 1
(c) Response of segment 3
373
(b) Response of segment 2
(d) Response of segment 4
(e) Response of segment 5
Fig. 5. Similarity responses under the PCA for person #5
(a) Response of segment 1
(c) Response of segment 3
(b) Response of segment 2
(d) Response of segment 4
(e) Response of segment 5
Fig. 6. Similarity responses under the tensor analysis for person #5
As shown in Fig 5, it is hard to say that responses by using the PCA are dominant under the corresponding segments or person. Whereas, from responses in Fig 6, most of segments have the maximum response under the exact same segments of person #5 in the experimental results using the tensor analysis.
374
S. Jeong and J. Cho
(a) Result of the PCA
(b) Result of the tensor analysis
Fig. 7. Values of similarity vector with segments 5 for person #5
In Fig. 6-(c), the maximum response is presented under the first segment in spite of experiment for segment 3; however, it is not a matter of this paper to detect which segment is. Therefore, we inspect people index according to the similarity response for identification of individual. Fig. 7 shows the experimental results with segment 5 for person #5. In the figure, horizontal axis is person index. The max denotes maximum values among sj which is similarity vector computed between unknown segments and saved coefficients of each person and the mean denotes average of sj . All of four cases in the figures are identified as person index #5 that is the correct answer. However, the case related to the tensor analysis is more dominant than the one of the PCA. It means that the proposed method outperforms the one based on the PCA in the separability. Fig. 5 and Fig. 6 are only the experimental results for one person, thus we perform experiments for all 16 people and then compare the final recognition performance. Table 1 presents the recognition rate. Table 1. Experimental results in the recognition rate Similarity max mean
The proposed method 86.25% 92.5%
The PCA-based method 81.25% 55.0%
As shown in the table using the maximum similarity score among each segments, the proposed method outperforms the method using the PCA by 5%. The proposed method has more high performance when the mean values of the similarity. It is caused by similar shape of the accumulated silhouettes between the same individual regardless the segments. Thus, the proposed method outperforms the one based on the PCA by 37.5%.
Gait Recognition by Multidimensional Representation
5
375
Conclusion
In this paper, we have proposed a novel gait recognition method based on the multidimensional representation for the accumulated silhouettes. First of all, one gait cycle is divided to four intervals according to the peaks of the stride length variation, and then, all of silhouettes within the each interval are cumulated to one frame to form the accumulated silhouettes. All of silhouettes in the one cycle are also cumulated to make average accumulated silhouettes. As a result, each person is represented with five segments. Using N-mode SVD algorithm, we compute the basis tensor composed of the core tensor, pixels and segments, which interacts with people space. From experiment with 16 people, the proposed method outperforms the method based on the PCA with by up to 37.5%. Especially, the proposed method has superior separability to the one of the PCA-based method with no regard to the segments. It is noticeable that the proposed method is quite desirable for on-line recognition system due to high separability regardless divided interval in which the accumulated silhouette has been created.
References 1. Cunado, D., Nash, J.M., Nixon, M.S., Carter, J.N.: Gait Extraction and Description by Evidence-Gathering. In: Proc. Int’l. Conf. AVBPA, pp. 43–48 (1999) 2. Lee, L., Grimson, W.E.L.: Gait Appearance for Recognition. In: Proc. ECCV Workshop. Biometric Authentication, pp. 143–154 (2002) 3. Lee, L., Grimson, W.E.L.: Gait Analysis for Recognition and Classification. In: Proc. IEEE Conf. Face and Gesture Recogntion, pp. 155–161 (2002) 4. Little, J.J., Boyd, J.E.: Recognizing People by Their Gait: The Shape of Motion. J. Computer Vision Research, 2–32 (1998) 5. Dunlavy, D.M., Kolda, T.G., Kegelmeyer, W.P.: Multilinear Algebra for Analyzing Data with Multiple Linages. Tech. Report, SAND2006-2079 (2006) 6. Vasilescu, M.A.O., Terzopoulos, D.: Multilinear Analysis of Image Ensembles: Tenssorfaces. In: Heyden, A., Sparr, G., Nielsen, M., Johansen, P. (eds.) ECCV 2002. LNCS, vol. 2350, pp. 447–460. Springer, Heidelberg (2002) 7. Vasilescu, M.A.O., Terzopoulos, D.: Multilinear Projection for Appearance-Based Recognition in the Tensor Framework. In: Proc. ICCV, pp. 1–8 (2007)
The Method of Generating Assessment for Intelligent Tutoring System Youngseok Lee1, Jungwon Cho2,*, and Byung-Uk Choi3, 1
Department of Electronics Computer Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul, 133-791 South Korea [email protected] 2 Department of Computer Education, Jeju National University, 102 Jejudaehakno, Jeju-si, Jeju-do, 690-756 South Korea [email protected] 3 Division of Computer Science & Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul, 133-791 South Korea [email protected]
Abstract. Recent developments in content processing technology and the widespread diffusion of wired and wireless Internet mean that users can now learn by means of a computer, anytime and anywhere. English learning that involves multimedia content can increase the interest of learners and lead to the development of their communication ability. Although using computers to teach English in a conventional educational environment provides motivation and effective learning on the part of the students, the method still has problems, which include the provision of learning materials without consideration of teaching methods, and evaluation without provision for differences in individual student levels. This paper introduces the Intelligent Tutoring System (ITS) for English learning, using web-based technologies. Using this system, the above problems are solved at the same time as the benefits of computer-based learning are retained. Its design and implementation are based on an intelligent tutoring system that provides content suitable for specific levels of ability. We used the contents of the Korean Elementary school 300-certification program for English conversation and an estimation of students’ abilities using Item Response Theory (IRT) to evaluate the proposed system.
1
Introduction
Recent developments in content processing technology and the widespread diffusion of wired and wireless Internet mean that users can now learn by means of a computer anytime and anywhere English learning that involves multimedia content can increase the interest of learners and lead to the development of their communication ability [1]. Although using computers to teach English in a conventional educational environment provides motivation and effective learning on the part of the students, the *
The Method of Generating Assessment for Intelligent Tutoring System
377
method still has problems, which include the provision of learning materials without consideration of teaching methods, and evaluation without provision for differences in individual student levels. In addition there has, so far, been little research focused on learning or education. Related research has investigated various factors, including the importance of particular subjects and the degree of difficulty must be considered when seeking to provide content suitable for specific student levels; it is necessary to first estimate each student's level and then provide appropriate learning materials [2, 3, 4]. Learning materials must be ranked according to the degree of difficulty, as well as with regard to levels of distinction and weighting values, something that is not necessary for other methods of educational evaluation, such as Item Response Theory (IRT) [4, 5]. In order to solve these problems while retaining the advantages of webbased learning, this paper sets out an Intelligent Tutoring System (ITS) for learning English. We describe the subject contents and create a learner model that determines the provision of material appropriate to specific student levels, as assessed by an inference engine based on IRT.
2
Review of Pertinent Literature
2.1
Intelligent Tutoring System (ITS)
Generally, ITS consists of an interface module, a learner module, a teaching module, and an expert module. The expert module builds and applies a knowledge base obtained through pertinent professional advice. It also manages information about suitable learning objectives, material, and specific instruction conditions[5]. The learner module consists of a learner model and a diagnostic processor. The learner model illustrates the current state of a student’s knowledge, while the diagnostic processor infers the learner model. A diagnostic function applies statistics or probability theory to diagnose a student’s level of knowledge and classifies it into one of several grades[6]. The teaching module decides what, when, and how to teach given the student’s learning status[5]. The interface module provides an interface through which student and system can interact; the interface module must offer a friendly and appropriate interface and respond to student interactions. 2.2
Item Response Theory (IRT)
Estimating a student’s ability initially requires tests; use of item response theory enables assessment of parameters such as item difficulty, item discrimination and item surmise [3, 7]. IRT provides a mathematical characterization of what happens when an individual is presented with a question. Each individual is characterized by a proficiency parameter denoted θ; each question is characterized by a collection of parameters, one of which is the item difficulty. Briefly, the IRT model presents the probability of answering a question correctly in terms of interaction between b and θ. If a student is much more proficient than the question is difficult, then this probability will be large. Conversely, if the question is much more difficult than the student is proficient, this probability will be small [8, 9].
378
Y. Lee, J. Cho, and B.-U. Choi
P (θi ) =
1 −1.7 a (θ − b ) e 1+
(1)
i
a: Item discrimination parameter b: Item difficulty parameter c: Guessing parameter θ: The individual ability level P(θ ) : The probability of a correct response is a function of the ability θ IRT derives the probability of each response as a function of the latent trait and some item parameters. The same model is then used to obtain likelihood of ability as a function of actually observed responses and, again, the item parameters. The ability value with the highest likelihood becomes the estimated ability. To estimate a learner’s ability, we used an item characteristic function in conjunction with the 2PL logistic model [6, 9].
3
System Design
3.1
System Architecture
This paper proposes a system consisting of an expression section that presents learning materials to the student. The ITS core consists of an inference engine that chooses materials appropriate for each student and monitors the student’s learning progress. Students are presented with learning content and materials via computer. The ITS core consists of an inference engine that chooses materials appropriate for each student and monitors student learning progress. The ITS core was designed using a logical learning formulation to measure a learner’s traits and knowledge and to act based on rules of inference [2, 3]. Expert module Teacher Module Interface Module
Input Interpreter
Output Creater
Diagnosis Module
User
Item Generator
Result generator
Instruc tion De sign Ge nera tor
Item Type
Instructi on
Domain knowled ge
Learner Module Learner Level
Learner Traits
Learner Diagnostic
Diagnosis Result File
1
I n f e r e n c e E n g i n e
Learner Model
Knowledge Base for Contents
Knowledge Base for Item Knowledge Base for Instruction
Fig. 1. Core mechanism of the proposed ITS
Figure 1 shows the core mechanism for the proposed ITS and Figure 2 shows the adaptive tests that consists of detailed diagnosis module and expert module. The teacher module establishes an instruction strategy by estimating a learner’s level and using the learner model database. The strategy of instruction uses learner ability levels and preferences to select appropriate items and methods of instruction.
The Method of Generating Assessment for Intelligent Tutoring System
379
Examinee Adaptive Tests Interface
Result Generator Module
Dialog Recommendation Module
Sheet Generation Module
Item Generation Module Rule Base
Item Profile
Inference Engine
Results
Student Models
Contents
Fig. 2. Detailed architecture of the adaptive testing system
3.2
Procedure for Estimating Student Level
To estimate individual learner ability level, we developed a method using a 2 PL logistic model and incorporated a user profile consisting of a domain area, item type, and studying result. Whenever a learner solves an item, his or her ability is estimated in relation to this domain area, item type, and studying result. If the learner completes one assessment, then the learner’s ability level will be estimated based on an assessment of results for about 10 items. Figure 3 presents examples of item content. Item ItemType Type No
Item
Item ItemDifficulty Difficulty
Ty Domain Di Diffic Disc pe al. ulty rim.
Domain Domain
Dialogue Dialogue
Item ItemDiscrimination Discrimination θ
P( θ ) I (θ )
Corr Sum. ect Result
Item ItemInformation Information
Sum. I (θ )
θs
Standard Error
Learner Learnerability ability
Fig. 3. Results of estimated learner ability level
3.3
Recommendation of Learning Contents and Generation of Assessment Sheets as Learning Model
After the learning model has been created through a diagnosis module, the system analyzes the learner’s weak areas of knowledge. The system uses item dialogue marks indicating the areas to be analyzed, which are saved in the learner model, and the system generates dialogue content suitable for the recommended learning. To determine individual ability level and generate customized learning, we designed and implemented a method to generate assessment sheets not for the purpose of assessment, but as learning tools. Figure 4 presents the procedure for generating assessment sheets.
380
Y. Lee, J. Cho, and B.-U. Choi Item extraction step that considers the weak domain area & dialogue
Item extraction step that considers the weak item type
Student Model
Student Model
Selection of weak domain area Selection of weak dialogue
Selection of weak item type
Item extraction as individual level
Item extraction as individual level
Unsolved items
Unsolved items
Is individual level suitable?
Is individual level suitable?
no
no
yes
Item Generation Module
yes
Item Generation Module
Extracted items
Extracted items
Fig. 4. Procedure for generating assessment sheets
Because this system is based on a learner’s weak areas of knowledge, it analyzes the learner’s weak areas, and selects the domain; the analysis enables the system to select the domain area in which most dialogue content was answered incorrectly. It then selects four types of items from the selected dialogue and calculates the probability of each item being solved, based on learner level. The system selects the optimal items and, most importantly, calculates probability of selecting 10 items with at least a 0.6 probability of being solved correctly. Of these 10 items, 6 are selected at random. This system analyzes weak areas for each learner, and presents two items for each area of weakness and one item as a reminder. This means that far more items from weaker areas are presented to learners. The 0.6 probability of being solved correctly is applied to ensure learners do not feel overwhelmed by the difficulty level. This method provides solvable items by probability, enabling learners to improve these areas without feeling inadequate. It effectively improves the weak areas of learner knowledge.
4
System Evaluation : Item Difficulty Analysis
This system was applied to about 20 students in Sin-ka elementary schools in Korea. The experiment was carried out in two sessions. Figure 5 lists results of item analysis. Item Difficulty
1.20
1.00
0.80 A Ty pe B Ty pe 0.60 C Ty pe D Ty pe 0.40
0.20
0.00
1 0 0 0 1 0 6
2 0 0 0 1 0 6
1 0 0 0 2 0 6
2 0 0 0 2 0 6
1 0 0 0 3 0 6
2 0 0 0 3 0 6
1 0 0 0 4 0 6
2 0 0 0 4 0 6
1 0 0 0 5 0 6
2 0 0 0 5 0 6
1 0 0 0 6 0 6
2 0 0 0 6 0 6
1 0 0 0 7 0 6
2 0 0 0 7 0 6
1 0 0 0 8 0 6
2 0 0 0 8 0 6
1 0 0 0 9 0 6
2 0 0 0 9 0 6
1 0 0 0 0 1 6
2 0 0 0 0 1 6
1 0 0 0 1 1 6
2 0 0 0 1 1 6
1 0 0 0 2 1 6
2 0 0 0 2 1 6
1 0 0 0 3 1 6
2 0 0 0 3 1 6
1 0 0 0 4 1 6
2 0 0 0 4 1 6
1 0 0 0 5 1 6
2 0 0 0 5 1 6
1 0 0 0 6 1 6
2 0 0 0 6 1 6
1 0 0 0 7 1 6
2 0 0 0 7 1 6
1 0 0 0 8 1 6
2 0 0 0 8 1 6
1 0 0 0 9 1 6
2 0 0 0 9 1 6
1 0 0 0 0 2 6
2 0 0 0 0 2 6
1 0 0 0 1 2 6
Fig. 5. Result of Analysis for Item Difficulty
2 0 0 0 1 2 6
1 0 0 0 2 2 6
2 0 0 0 2 2 6
1 0 0 0 3 2 6
2 0 0 0 3 2 6
1 0 0 0 4 2 6
2 0 0 0 4 2 6
1 0 0 0 5 2 6
2 0 0 0 5 2 6
1 0 0 0 6 2 6
2 0 0 0 6 2 6
The Method of Generating Assessment for Intelligent Tutoring System
381
Table 1. Classification result of analysis for 208 Items
Difficulty Type
High
Medium
Low
Total
4 3 17 8 32
45 38 32 38 153
3 11 3 6 23
52 52 52 52 208
A B C D Total
Table 1 uses Cangelosi’s criterion to classify results from Figure 3[6, 9]. Preliminary tests (random) yielded a basic assessment of each student; If a user solved 208 items, our system would produce fewer sample items from this item bank. All students were required to study and evaluate without using an inference engine. When fewer than one-third of results changed, the system recommended an inference engine to learners. Table 2. User-defined table
Preliminary tests (random) yielded a basic assessment of each student; Table 2 presents the number of sheets and mean values. Each sheet included ten items, and included sample items from the same item bank. If a user solved 208 items, our system would produce fewer sample items from this item bank. All students were required to study and evaluate without using an inference engine. When fewer than one-third of results changed, the system recommended an inference engine to learners.
382
5
Y. Lee, J. Cho, and B.-U. Choi
Conclusion
Existing web-based education systems focus on providing study materials without considering the knowledge status or ability of individual students. This paper proposes an intelligent tutoring system that provides materials appropriate for specific student. Installation of such a system in elementary schools, as part of the English teaching program, showed that the inference engine can provide materials appropriate to the level of each individual student.
References 1. Davidovic, A., Warren, J., Trichina, E.: Learning Benefits of Structural Example-Based Adaptive Tutoring Systems. IEEE Trans. on Education 46, 241–251 (2003) 2. Lee, Y., et al.: An Intelligent Tutoring System Based on a Multi-modal Environment for the 300-Certification Program of English Conversation. In: Ikeda, M., Ashley, K.D., Chan, T.-W. (eds.) ITS 2006. LNCS, vol. 4053, pp. 778–780. Springer, Heidelberg (2006) 3. Lee, Y., et al.: A Personalized Assessment System Based on Item Response Theory. In: Luo, X., Spaniol, M., Wang, L., Li, Q., Nejdl, W., Zhang, W. (eds.) ICWL 2010. LNCS, vol. 6483, pp. 381–386. Springer, Heidelberg (2010) 4. Murase, T., Isomoto, Y.: e-Learnig System to Provide Optimum Questions Based on Item Response Theory. In: Ikeda, M., Ashley, K.D., Chan, T.-W. (eds.) ITS 2006. LNCS, vol. 4053, pp. 695–697. Springer, Heidelberg (2006) 5. Okamoto, T.: The Current Situation and Future Directions of Intelligent. IEICE Trans. on Information & System E77-D, 143–161 (1994) 6. Gin-Fon, N.J., Alfred, B.: The Implementation of an Adaptive Test on the Computer. In: Proceedings of the Fifth IEEE International Confernce on Advanced Learning Technologies (ICALT 2005), pp. 822–823 (2005) 7. Feng-Hsu, W.: Application of Componential IRT Model for Diagnostic Test in Standardconformant eLearning System. In: Proceedings of the Sixth International Confernce on Advanced Learning Technologies (ICALT 2006), pp. 237–241 (2006) 8. Johns, J., Mahadevan, S., Park Woolf, B.: Estimating Student Proficiency Using an Item Response Theory Model. In: Ikeda, M., Ashley, K.D., Chan, T.-W. (eds.) ITS 2006. LNCS, vol. 4053, pp. 473–480. Springer, Heidelberg (2006) 9. Crocker, L., Algina, J.: Introduction to Classical and Modern Test Theory. Rinehart and Winston, Inc., New York (1996)
A Global Time Synchronization Scheme for Wireless Sensor Networks Bongkyo Moon1 and Jonggoo Bae2 1
Abstract. All the clocks in wireless sensor networks (WSNs) need to have approximately the same reading at a global time point, irrespective of their relative distance in order that a sensor system with global clock may be capable of coordinated operation and data synthesis for future predictions. In this paper, we propose a fast asynchronous diffusion (FAD) scheme in order to improve the performance of the localized and fully distributed asynchronous diffusion method, which achieves global synchronization by spreading the local synchronization information to the entire system. We present the simulation results for the FAD scheme and asynchronous diffusion method and show that FAD scheme converges faster than the asynchronous diffusion method. Keywords: WSN, global synchronization, asynchronous diffusion.
allowing each node to perform its operation at random. In this paper, we investigate the asynchronous diffusion method introduced by Li and Rus and then propose a fast asynchronous diffusion (FAD) scheme for improving the performance of this method.
2
Asynchronous Diffusion Method
2.1
Randomized Node Operation
The asynchronous averaging algorithm (Algorithm 1) was proposed for overcoming the constraint of the existing synchronization diffusion method which requires a set order for all node operations. In this algorithm, a node can synchronize with its neighbors at any time in any order as long as each node always has the chance to be involved in the execution. Each node can run the asynchronous operations on the fly without knowing what other nodes are doing. That is, each node executes the average operation once although the order of the operations of all the nodes is randomized. Fig. 1 shows the example of randomized average operation groups. G1
G3 G2
Fig. 1. Example of randomized average operation groups
2.2
Diffusion of Average Clock Values
The main idea of Algorithm 1 [4] is to average all the clock time readings and set each clock to the average time. It gives a very simple average operation of a node over its neighbors. Each node tries to compute the local average value directly by asking all its neighbors about their values; it then sends out the computed average value to all its neighbors so they can update their values. After a certain number of rounds of diffusion, the clock in each sensor converges to a global average value, and eventually some error threshold can be achieved. Here, a round is defined to be the time for each node to finish the average operation in Algorithm 1 exactly once, so the number of rounds required for the network to achieve some error threshold signifies the convergence speed.
A Global Time Synchronization Scheme for WSNs
385
In this method, a node may have several neighbors for average operation since the network is assumed to be connected. If a node is involved in two or more average operations, therefore, these operations must be sequenced due to the assumption that the average operation is atomic. This algorithm, which has been mathematically proved by Li and Rus [4], can typically adapt to the changing network topology, limited node failure, adverse communication conditions, and node mobility.
Algorithm 1. Asynchronous Averaging Algorithm in WSN [4] 1: for each sensor ni with uniform probability do 2: Ask its neighbors the clock readings (read values from ni and its neighbors) 3: Average the readings (compute) 4: Send back to the neighbors the new value (write value to ni and its neighbors)
3
Fast Asynchronous Diffusion (FAD) Scheme
3.1
Most Recently Received Average Clock Value
In asynchronous diffusion algorithm (Algorithm 1), each node might have a series of adjusted clock values sent from several neighbors through average operations since all sensor nodes are assumed to be connected. According to the Algorithm 1 and its assumptions, therefore, a node ni in each round consequently adjusts its local clock with the most recently received value among a series of average clock values. Thus, this adjusted clock value is eventually diffused over the whole network. In equation (1), Ci-adjust presents an average clock value sent back to the ni’s neighbors for adjusting clock in Algorithm 1, where N is the number of the neighbors of ni and Cj is a clock value among neighbor sensor nodes. C i _ adjust =
3.2
1 N
Ci +
N −1
C j =1
j
(1)
Fast Convergence of Diffusion Round
We now propose a fast asynchronous diffusion (FAD) scheme (Algorithm 2) for improving the convergence speed of asynchronous diffusion method (Algorithm 1). In Algorithm 2, each node takes the mean of a series of average clock values. More specifically, a node ni adjusts its clock value at the end of each round with the mean of average clock values sent back from all the neighbors. Moreover, since the average operation in this algorithm is not atomic, the average operations do not have to be
386
B. Moon and J. Bae
sequenced even though a node is involved in two or more operation groups. Hence, the eventual adjusted clock value in a round can be expressed with equation (2), where M is the number of the nodes for average operation within each neighbor group including ni. C i − adjust
1 C + 1 = N i M
Ck + k =1
N −1
C j j =1
M −1
(2)
In this algorithm, each node has to keep in the buffer the average clock values sent back from all the neighbors in a round. Then, at the end of a round, every node computes the mean with the average clock values which are already compensated with the offsets according to the time elapsed in the buffer. Algorithm 2 rather seems to have less performance in the aspect of space complexity and time complexity since it may require more operations and storages than the Algorithm 1 in getting average value. However, the number of rounds required for convergence achievement in this scheme is less than that in the asynchronous diffusion method. That is, the overall time for synchronizing all the nodes with this scheme becomes shorter since this scheme converges faster than the asynchronous diffusion method. In the following sections, the results are presented with NS-2 simulation.
Algorithm 2. Fast Asynchronous Averaging Algorithm in WSN 1: for each node ni with uniform probability do 2: Ask its neighbors the clock readings (read values from ni and its neighbors) 3: Average the readings (compute) 4: Send back to the neighbors the new value (buffer the value to ni and its of neighbors instead of writing) 5: Each node ni performs average operation with all adjusted buffered values received from its neighbors (write value to itself ni)
4
Performance Evaluations
We implemented the FAD scheme (Algorithm 2) and asynchronous diffusion scheme (Algorithm 1) in simulation with NS-2 simulator (version 2.30) based on IEEE 802.15.4 module. We ran a series of scenarios with different network parameters. For each experimental set of parameters, the simulation was executed several times using a randomly generated network topology. In each experiment, a stimulus was generated at a randomly chosen node and propagated to the whole network until the relative error (0.01) was achieved. The detail simulation parameters are summarized in Table 1.
A Global Time Synchronization Scheme for WSNs
387
Table 1. The Simulation Parameters
Parameter Number of Nodes Sensor Field Transmission Range Physical Layer & MAC Layer Routing Protocol Relative Error Uniform Probability (Mean) Threshold (Percentage of Drift)
In this simulation, the round is counted as the number of the given algorithm performed at once. The number of operation is the number of average operation performed from all nodes in a round. Thus, the number of operation ranges with uniform probability between zero and the number of nodes participating in a round. The threshold is the drift rate between the received clock value and local clock value in a tick. In Fig. 2 and Fig. 3, each point(*) represents a running result on randomly generated network. That is, the markers are the number of rounds and the number of total operations when relative error becomes 0.01 which is a convergence condition in each experiment. The error rate for convergence decreases exponentially with the increase of the number of rounds [4]. The plotted curve is the average number of rounds and the average number of total operations for one suite of network parameters. A sparse network with fewer nodes undergoes large variation in terms of convergence speed. Fig. 2 shows the convergence speed with the number of nodes. The two figures represent the comparison between asynchronous diffusion method and FAD scheme in the number of rounds with threshold value 40%. In these figures, the number of rounds decreases with the increase of the number of nodes. The reason is that the increase of the number of neighbors in each node makes the network more connected and eventually makes the diffusion better expedited. Fig. 2 also shows that the number of rounds has exponential shape as the number of sensor nodes increases. It means that, when the number of nodes is especially over the specific value, the number of rounds required for achieving global time synchronization is not related with the number of nodes. In this simulation, when the number of sensor nodes is over 150, we can see that the convergence speed of FAD scheme is faster than that of asynchronous diffusion method. That is, when the number of nodes increases from 150 to 500, the number of rounds required for achieving global time synchronization in the FAD scheme is kept below compared to that of asynchronous diffusion method.
388
B. Moon and J. Bae
100 av erage v alue
90
data point
80
number of round
70 60 50 40 30 20 10 0
0
50
100
150
200 250 300 350 number of nodes
400
450
500
(a) Asynchronous Diffusion Method 100 av erage v alue
90
data point
80
number of round
70 60 50 40 30 20 10 0 0
50
100
150
200 250 300 350 number of nodes
400
450
500
(b) Fast Asynchronous Diffusion Method (FAD) Fig. 2. Comparison between Asynchronous Diffusion Method and FAD in the Number of Rounds with Threshold Value 40%
Fig. 3 shows the total number of average operations conducted by all the nodes in each network. The two figures show the comparison between asynchronous diffusion method and FAD scheme in the number of operations with threshold value 40%. These figures represent that the number of total operations increases when the number of nodes increases under the fixed sensor field. The reason is because the number of neighbors increases linearly with the number of nodes with other network parameters fixed. In this simulation, when the number of sensor nodes is 500, we can see that the number of total operations required for achieving global time synchronization in the FAD scheme less than that of the asynchronous diffusion method.
A Global Time Synchronization Scheme for WSNs
389
Fig. 4 represents the comparison between asynchronous diffusion and FAD in the number of operations and the number of rounds with threshold value (log scale). Fig. 4 depicts the number of average operation and Fig. 4 shows the number of rounds in this simulation. When the number of nodes is over 175, we can see that FAD requires fewer operations and fewer rounds than asynchronous diffusion. When the number of nodes is under 175, it might not be possible to compare FAD with asynchronous diffusion from this simulation. When the number of nodes is over 175, however, we can also see that FAD has better performance than asynchronous diffusion. 4
x 10
2
average value
1.8
data point
number of operation
1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0
0
50
100
150
200 250 300 350 number of nodes
400
450
500
(a) Asynchronous Diffusion Method 4
x 10
2
average value
1.8
data point
number of operation
1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0
0
50
100
150
200 250 300 350 number of nodes
400
450
500
(b) Fast Asynchronous Diffusion Method (FAD) Fig. 3. Comparison between Asynchronous Diffusion Method and FAD in the Number of Operations with Threshold Value 40%
390
B. Moon and J. Bae
9.8 average diffusion
number of operations(in log scale)
9.6
FAD(100%) FAD(80%) FAD(60%) FAD(40%)
9.4 9.2 9 8.8 8.6 8.4 8.2 8 7.8 50
100
150
200
250 300 350 number of nodes
400
450
500
(a) Comparison of the number of operations 6 av erage dif f usion
FAD(100%) FAD(80%) FAD(60%) FAD(40%)
number of rounds(in log scale)
5.5
5
4.5
4
3.5
3 50
100
150
200
250 300 number of nodes
350
400
450
500
(b) Comparison of the number of rounds Fig. 4. Comparison between Asynchronous Diffusion and FAD in the Number of Operations and Rounds with Threshold Value (log scale)
6
Conclusions
In the asynchronous diffusion method for the global synchronization, each node can perform its operation locally, but still achieve the global clock value over the whole network. In this paper, we propose a fast asynchronous diffusion (FAD) scheme in order to improve the performance of the asynchronous diffusion method. We present the simulation results for the FAD scheme and the asynchronous diffusion method and show that FAD scheme converges faster than the asynchronous diffusion method.
A Global Time Synchronization Scheme for WSNs
391
Acknowledgements. This work was supported by the Korea Research Foundation Grant funded by the Korean Government (KRF-2008-313-D00683).
References 1. Sivrikaya, F., Yener, B.: TimeSynchronization in Sensor Networks: A Survey. IEEE Network 18(4), 45–50 (2004) 2. Maroti, M., Kusy, B., Simon, G., Ledeczi, A.: The flooding time synchronization protocol. In: Proceedings of the ACM Conference on Networked Sensor Systems (SenSys 2004), pp. 39–49. ACM Press, New York (2004) 3. Su, W., Akyildiz, I.F.: Time-Diffusion Synchronization Protocol for Wireless Sensor Networks. IEEE/ACM Transactions on Networking 13(2), 384–397 (2005) 4. Li, Q., Rus, D.: Global Clock Synchronization in Sensor Network. IEEE Transactions on Computer 55(2), 214–226 (2006) 5. Wang, B., Fu, C., Lim, H.B.: Layered Diffusion-based Coverage Control in Wireless Sensor Networks. Computer Networks 53(7), 1114–1124 (2009) 6. Chinh, H.D., Tan, Y.K.: Smart Wireless Sensor Networks. Intech (2010)
The Correlations between Addicted Cell Phone Use of Adolescents and Mental Health: The Case of Middle School Students in Korea Choong-sook Ko1,*, Dong-uk Cheong2,*, Tae-su Park1, Young-sik Kang3, and Jung-hwan Park1,∗∗ 1 Dept. of Education, Jeju National University Ara 1-dong, 102 Jejudaehakno, Jeju-si, Jeju Special Self-Governing Province, 690-756, Republic of Korea [email protected] 2 Dept. of Education, Korea National University of Education Inmun-kwan #441, Gangnae-myeon Cheongwon-gun Chungbuk, 363-791, Republic of Korea [email protected] 3 Dept. of Social Welfare, Myungshin University 212-1, Kumchi-ri, Byulyang-myon Suncheon-si, JeonNam, 540-871, Republic of Korea [email protected]
Abstract. The purpose of this research was to analyze four factors of addicted cell phone use affecting mental health. The participants were 639 second graders, who were selected among eight middle schools in Jeju Province in Korea (Republic) by the multiple cluster sampling. The questionnaire was administered to measure participants’ addicted cell phone use and mental health, and the collected data were analyzed by Regression and Correlation Analysis. The results were that the four factors (Tolerance-dependency, Cling, Life-distress, and Obsession) were significantly correlated with mental health. The results indicated the more addictively students use their cell phones, the lower levels of mental health they have in the case of Korean middle school students. This research suggests that the guidance from teacher and parents for adequate cell phone use would improve adolescents’ mental health. Keywords: Mental health, addicted cell phone use, Korea, correlation.
1
Introduction
Modern society has rapidly been changed by the development of scientific technology. Especially the development of information communications has had a considerable impact on overall our society. Cell phones might be considered as one of * ∗∗
First authors: Two co-first authors who made equivalent contributions. Corresponding author.
The Correlations between Addicted Cell Phone Use of Adolescents and Mental Health
393
media with the biggest impact. Korea has 50,325,191 cell phone users [1]: Almost one cell phone per person has been used. However, imprudently using cell phones leads to negative aspects on adolescents with insufficient self-control. Firstly, one of the negative aspects is involved in the health problems. According to WHO, using cell phones causes the increase of risk of brain tumors by excessive electromagnetic waves exposure, headaches, pains in eardrums, dizziness and so on. Secondly, another negative aspect is involved in the problems of the destruction and changes of language. The excessive use of shortened forms of language, which is used to overcome the limit of transmission capacity for various emoticons and text message and to type faster text message, might bring about the severe destruction and change of language [2]. The most important negative aspect considered by this research is involved in mental health problems. Addicted cell phone use can cause abstinence symptom [3]. That is, they feel anxious and nervous while they don't use their cell phones. They feel as if they were ringing and experience the pounding of the heart and headaches, awaiting the ringtones while they don't use their cell phones. A research suggested adolescents using addictively cell phones get more significant scores than usual cell phone users in adolescence in scales of schizophrenia and hypomania, and tend to show high depression and impulsiveness in MMPI [4]. Moreover, the group of addicted cell phones users has more anxiety disorders as well as lower self-esteem than that of usual users due to lower emotional stability [5]. [6] suggested the nine scales in SCL-90-R (Symptom Checklist-90-Revision) showed the significant correlation with cell phone addiction and showed the addicted cell phone use influence meaningfully anxiety, depression and impulsive symptom. Adolescents are unstable and go through much stress as well as the exposure of mental health problems since they are under vulnerable period in the aspect of mental health due to their developmental stages. We can predict the level of mental health will lower as they use their cell phones addictively and the addicted cell phone use will have a significant effect on mental health. However, we have few studies on the relations between addicted cell phone use and mental health. The purpose of this research is to reveal the correlations between addicted cell phone use and mental health. Research questions are as follows. 1. What differences exist in actual conditions of cell phone use between addicted cell phone use group and non-addicted cell phone use group? 2. What correlations exist between mental health scores and addicted cell phone use?
2
Method
2.1
Participants
The participants of this research were 639 second graders, who were selected among the second-year pupils of eight middle schools in Jeju Special Self-Governing
394
C.-s. Ko et al.
Province by the multiple cluster sampling. These participants were made of 309 males and 330 females, and they have lived in 351 Jeju city, 109 Suburban of Jeju city, 122 Seoqwipo city, and 57 Suburban of Seoqwipo city. 2.2
Instruments
The Scale of Addicted Cell Phone Use. Korean-Type Mobile Phone Addiction Scale developed by [7] was used for the test on addicted cell phone use. This scale with a four-factor (Tolerance-dependency, Cling, Obsession, and Life-distress) self-report measure is made of a five-point Likert type scale, and 20 questions (5 questions per subscale). The total score of the scale ranges from 20 to 100 and means the higher score, the more addicted cell phone use. The group faring at the highest 25% of total score (20-100) was classified as the addicted cell phone use group and the group faring at the lowest 25% was classified as the non- addicted cell phone use group. 167 participants were classified as addicted use group and 160 as non-addicted group among 639 participants. The reliability (Cronbach's α) was .91 in the study done by [7] and [8] and .92 in this research. The Scale of Mental Health Symptom. SCL-90-R was used for the scale of mental health. It is consisted of 90 questions including nine-factor (Somatization, Obsessivecompulsive, Interpersonal-sensitivity, Depression, Anxiety, Hostility, Phobic-anxiety, Paranoid-ideation, Psychoticism) and represents each psychological diagnosis [9]. Korean Type SCL-90-R re-standardized by [10] was used for this research. The reliability (Cronbach's α) was .97 and .98 in each study done by [6] and [8] and was .98 in this research. 2.3
Data Analysis
Data were collected from Nov. 8th, 2010 to Nov, 19th, 2010. The researcher visited six schools and explained the purpose of the survey to participants. In two schools where the researcher couldn't visit, their homeroom teachers did the surveys instead of the researcher. Data analysis was carried out by using SPSS WIN 12.0 program as follows. First, t-test was carried out for the gap of mental health scores between two groups. Second, Pearson's correlation coefficient was done for the relation between addicted cell phone use and mental health.
3
Results
3.1
Differences in Actual Conditions of Cell Phone Use
The result of difference of the total duration of cell phone use, the frequency of text message use per day, and whether using cell phone in class or not between the two groups are as follows. When it comes to of the total duration of cell phone use, the duration ranges from 1-2 years in non-addicted cell phone use group whereas it is
The Correlations between Addicted Cell Phone Use of Adolescents and Mental Health
395
over 3 years in addicted cell phone use group. The difference was statistically significant (χ2=17.89, p<.01). With regard to the frequency of text message use per day, the most is under 10 times in non-addicted cell phone use whereas the most is over 40 times in addicted cell phone use. The difference was also statistically significant (χ2=105.61, p<.001). When it comes to whether using cell phone in class or not, non-addicted cell phone use group never use their cell phones in class (68.9%) whereas addicted group sometimes use them. 3.2
Differences in Mental Health Scores
Addicted cell phone use group’s mental health value (M=90.29, SD=64.23) was higher than non-addicted group’s mental health value (Mean=34.81, SD=39.77), and the difference was statistically significant (t=-9.34, p<.001) (Table 1). Table 1. Difference of mental health scores between two groups
Sub-factors health
of
mental
Non-addicted group (n=160) M
Addicted group (n=167)
SD
M
t
SD
Somatization
4.53
6.08
12.71
9.74
-9.06***
Obsessive-compulsive
6.02
6.04
13.29
8.12
-9.14***
Interpersonal-sensitivity
5.47
6.17
11.26
8.00
-7.31***
Depression
5.95
8.08
15.12
11.81
-8.16***
Anxiety
2.98
4.96
9.88
8.37
-9.03***
Hostility
3.01
3.96
7.30
5.53
-8.03***
Phobic-anxiety
1.51
3.28
4.89
5.36
-6.85***
Paranoid-ideation
2.34
3.57
6.27
5.71
-7.43***
Psychoticism
3.00
4.67
9.57
8.38
-8.71***
Mental health total
34.8 1
39.77
90.29
64.23
-9.34***
***p<.001 3.3
Correlations between Addicted Cell Phone Use and Mental Health
The 4 sub-factors of addicted cell phone use were examined to determine those that contributed to the 9 sub-factors of mental health, as summarized in Table 2. In general, addicted cell phone use was significantly correlated with mental health (r=.39, p<.01). Remarkably, all of the 4 sub-factors of addicted cell phone use were correlated statistically significantly with all of the 9 sub-factors of mental health as shown in Table 2.
396
C.-s. Ko et al. Table 2. Correlation between addicted cell phone use and mental health (n=639)
**p<.01 1.Tolerance-dependency, 2.Cling, 3.Life-distress, 4.Obsession, 5.Cell-phone addiction total, 6.Somatization, 7.Obsessive-compulsive, 8.Interpersonal-sensitivity, 9.Depression, 10.Anxiety, 11.Hostility, 12.Phobic-anxiety, 13.Paranoid-ideation, 14.Psychoticism, 15.Mental health total
4
Conclusions
The things from this research are as follows. First, this research suggests the total duration of cell phone use in addicted cell phone use group is longer than that in nonaddicted group. Most students in addicted cell phone use group have had their cell phones since fifth or sixth grader. We need to teach how to use cell phone right from the elementary school period. Addicted cell phone use group use text message more excessively than non-addicted cell phone use group. We need to let adolescents recognize problems resulted from excessive text message uses as well as a lot of problems caused by addicted cell phone use and to instruct ways to use cell phone right. Moreover, addicted cell phone use group has more cases of using cell phone personally in class than non-addicted cell phone use group. Therefore, rather than banning students from using their cell phones in class we had better make ways to restrict cell phone use in class through enough talks with them. Second, this research suggests the more addictively students use their cell phone, the lower levels of mental health they have. In addition, Addicted cell phone use is relevant to mental health as addicted cell phone use and mental health have positive correlation. Accordingly, adolescents must recognize addicted cell phone use affect
The Correlations between Addicted Cell Phone Use of Adolescents and Mental Health
397
their mental health and we need to provide adolescents with differentiated programs by levels of mental health. Besides we need to give them the individual counseling to strengthen vulnerable parts of their mental health by getting them in advance. The results of this research is useful to understand adolescents using their cell phones addictively and to set counseling aims for them. The number of adolescents using their cell phones addictively will continue to increase and the problems of their mental health get more serious. Therefore, we should develop prevention programs for adolescents using their cell phones additively, which are considered the characteristics of addicted cell phone use group. We also need to make use of the prevention program in school, home and counseling field.
References 1. Korea Communications Commission.: The present status on members of on- and off-line communication service. Seoul, Korea (end of October 2010) 2. Daily Seoprise.: Todrack-Todrack. Internet News Article, Seoul, Korea (August 25, 2006) 3. Han, J., Hur, G.: Construction and Validation of Mobile Phone Addiction Scale. Korean Journal of Journalism & Communication Studies 48(6), 138–165 (2004) 4. Yang, S., Park, Y.: A Prediction Model of Cellular Phone Addiction Tendency among Adolescents. Journal of the Korean Home Economics Association 43(4), 1–16 (2005) 5. Jang, H.: The Psychological Characteristics of Adolescents Addictive Using Cellular Phone. Unpublished master’s thesis. Sungshin Women’s University. Seoul, Korea (2002) 6. Song, E.: The Relationship between the using Mobile Phone and Mental Health State of High School Students. J. Korean Acad. Psych. Mental Health Nurs. 15(3), 325–333 (2006) 7. Lee, J.: An Exploration of Socio-environmental and Individual-psychological Variables affecting the Adolescent’s Cellular Phone Addiction. Unpublished master’s thesis. Korea National University of Education. Chungbuk, Korea (2006) 8. Yu, M.: A Study on High School Students’ Cellular Phone Addiction, Mental Health and Impulsiveness. Unpublished master’s thesis. Kangwon National University, Kangwon, Korea (2009) 9. Kwon, K., Lee, H.: Study on the relationship between adolescent’ Internet use and health of mind. Social Business and Mind Health 13, 59–86 (2002) 10. Kim, K., Kim, J., Won, H.: Implementation Outline of Simple Mind Diagnostic Test. Central Research Institute on Aptitude, Seoul, Korea (1992)
Security Requirements of Certificate Validation in Web Security Hee Bong Choi, Yong Dae Kim, San Yun Han, and Hyuk Joong Yoon The Attached Institute of ETRI, Korea [email protected]
Abstract. As Internet has been used widely, the enterprise is participated in the web service business as Internet advertisement, home shopping, home banking, and so on. Web security threats as the immoral, the cyber-crime which threaten the electronic commerce are increased in the reversed side of the convenience of WWW. Therefore, it is necessary to establish and apply the countermeasure in electronic commerce activities. It is very important to manage the key safely in order to provide the security web services which are the confidentiality, the authentication, the repudiation and so on. It needs the convenient key management system to provide the proper security service to electronic commerce activities. PKI-based key management system which a certificate is applied in is used widely for HTTPS. We can find the vulnerabilities of certificate application in a lot of web security products. This paper studies the vulnerability cases of the certificate application which we have found frequently in the real environment or at web security products. And this paper proposes the security requirements of the certificate when it uses in practical application. These security requirements which are proposed by this paper can be useful in the area of the validation of a certificate when the PKI based key management system is established. These security requirements can be used as a development guideline of a web security product developer in the area of the validation of a certificate. Also, they can be used as an evaluation criteria of a web security products in the area of the validation of a certificate when an evaluator assessments the web security products. Keywords: Web Security, PKI-based key management system, certificate, vulnerability of certificate, Security requirement of certificate.
Security Requirements of Certificate Validation in Web Security
399
Therefore, it is necessary to establish and apply the countermeasure in electronic commerce activities. It is very important to manage the key safely in order to provide the security web services which are the confidentiality, the authentication, the repudiation and so on. It needs the convenient key management system to provide the proper security service to electronic commerce activities. PKI-based key management system which a certificate is applied in is used widely for HTTPS. Now a certificate is made frequent use X.509 standard [9] and it is issued by Certificate Authority (CA). A certificate is an electronic document to relate public key with its owner. We can find the vulnerabilities of certificate application in a lot of web security products. There may be some vulnerabilities in operating environment of a certificate which depends on the storing place of a private key or the key computing place. And, there may be some vulnerabilities in the validation process of a certificate which depends on the user's carelessness of certificate validation when the user accesses to a web site, the contents of a certificate not to be appropriate to operating environment, the incorrect implementation of the web security products, the limitation to operating environment. And, there may be some vulnerabilities about the generation of a certificate which is used with the weak cryptographic algorithm by CA. This paper studies the vulnerability cases of the certificate application which we have found frequently in the real environment or at web security products. The vulnerabilities which we study in this paper are the user's carelessness of certificate validation, the inconsistence contents of a certificate to operating environment, the incorrect implementation of the web security products, the case to use one certificate in order to perform both of the digital signature and the key distribution, the generation of a weak certificate. An attacker tries a victim pc spoofing, exchanges a correct certificate to a forged certificate and can capture the information of the victim as password by using these vulnerabilities. In order to validate a correct certificate when the an entity want to use it there are the verification of the owner, the check of the validate period, the verification of the certificate signature, the verification of CRL(Certificate Revocation List). And this paper proposes 6 security requirements(SR1, SR2, SR3, SR4, SR5, SR6 and SR7) of the certificate when it uses in practical application. These security requirements which are proposed by this paper can be useful in the area of the validation of a certificate when the PKI based key management system is established. These security requirements can reduce the vulnerability of the certificate application in web security services These security requirements can be used as a development guideline of a web security product developer in the area of the validation of a certificate. Also, these security requirements can be used as an evaluation criteria of web security products in the area of the validation of a certificate when an evaluator assesses the web security products.
400
H.B. Choi et al.
This paper is organized as follows: section 2 introduces the key management system in web system environment and describes to validate a certificate. Next, we introduce the vulnerability cases of certification when it is used in real world in section 3. The security requirements of certificate application are shown in section 4. Finally, we conclude this paper in section 5.
2
PKI-Based Management System in Web System
The activities which are done in electronic commerce space are various kinds. Security threats are increased in the reversed side of the convenient web system. In order to perform these electronic commerce activities safely in web system, it is necessary to establish and apply the countermeasure to the threats. It is very important to manage the key safely in order to provide the security services which are the confidentiality, the authentication, the repudiation and so on in web system. It needs the convenient key management system to provide the proper security service to web system and the PKI-based key management system which uses a certificate authority (CA) has been implemented widely.
Fig. 1. Network which contains PKI-based key management system
Fig. 1 shows network that consists certificate authorities, web servers, entities which have various kinds of web browser. The certificate authority can generate and distribute a certificate under consistent and proper key management policy in electronic commerce and then this can support security services in the area of electronic commerce activities. Now, PKI-based key management system makes frequent use a certificate in real environment. A certificate generally is generated with X.509 certificate standard and is signed digitally by certificate authority. A certificate is an electronic document to relate public key with its owner. Fig.2 shows the example of certificate X.509 structure [8].
Security Requirements of Certificate Validation in Web Security
401
Version: 1 (0x0) Serial Number: 7829 (0x1e95) Signature Algorithm: md5WithRSAEncryption Issuer: C=ZA, ST=Western Cape, L=Cape Town, O=Thawte cc, OU=Certification Services Division, CN=Thawte Server CA/[email protected] Validity Not Before: Jul 9 16:04:02 1998 GMT Not After : Jul 9 16:04:02 1999 GMT Subject: C=US, ST=Maryland, L=Pasadena, O=Brent Baccala, OU=FreeSoft, CN=www.freesoft.org Subject Public Key Info: Public Key Algorithm: rsaEncryption RSA Public Key: (1024 bit) Modulus (1024 bit): 00:b4:31:98:0a:c4:bc:62:c1:88:aa:dc:b0:c8:bb: ........................................................................... e8:35:1c:9e:27:52:7e:41:8f Exponent: 65537 (0x10001) Signature Algorithm: md5WithRSAEncryption 93:5f:8f:5f:c5:af:bf:0a:ab:a5:6d:fb:24:5f:b6:59:5d:9d: ...................................................................................... 68:9f
Fig. 2. Certificate X.509 structure
If the trusted certificate authority is established in network, there are advantages and disadvantages. The advantages are made by the standard and the consistency policy of key management as following. The first is to save implementation cost of security products which are used in web system. The second is to perform key distribution conveniently among the entities in network. There are the disadvantages as following. The first is to need considerable cost in order to establish and maintain a key management system. The second is that it is complex to validate a certificate. It is very important to store and manage a private key which corresponds to a certificate in order to perform PKI-based key management safely. It is necessary to store the encryption of a private key in PC or store a private key in a crypto token. If security products perform the operation of a private key in the opened operating system, there are some vulnerabilities which can disclose the key. To validate a certificate authenticates the public key which is certificated by a trusted certificate authority. The process of the validation consists of the verification of the owner of a certificate, the check of the validity period, the verification of a certification digital signature, and the verification of a certificate revocation list (CRL) from LDAP server. There are vulnerabilities that the fraud of a certificate can be existed potentially at the cases as followings. The cases are a user's carelessness of certificate validation, the inconsistence contents of a certificate to operating environment, the incorrect implementation of the web security products, the skip of certificate validation because of IT environment limitation.
402
H.B. Choi et al.
There is vulnerability that the fraud of a trusted certificate authority can be existed potentially at the case as following. The case is a weak certificate which is generated with a weak cryptographic algorithm by a trusted certificate authority. There is vulnerability at the case that one certificate is used for both digital signature and key distribution. This paper studies vulnerability cases which we have found in the real environment or in a lot of web security products. And this paper supposes security requirements of certificate validation which can be applied in web security system.
3
Vulnerability Cases of Certificate Application
3.1
User Carelessness Vulnerability to Validation of Certificate
It is assumed that the entity of the web client has been implemented to perform the completed validation of a certificate of the web server when they perform the authentication, the secrete communication, or the digital signature communication between the web server and the web client. Here, the validation of a certificate means to authenticate a public key of the web server. The web client receives a certificate of the web server to begin the SSL communication, generates the session key SKcs, encrypts SKcs by the public key PUs of the web server and transmits the encrypted session key C to the web server as formulas (1). And then the web server decrypts the encrypted session key C by the private key PRs of the web server and gains the session key SKcs as formulas (2) C = E(PUs, SKcs)
(1)
SKcs = D(PRs, C) = D[PRs, E(PUs, SKcs)]
(2)
If an attacker captures the certificate of the web server, he transmits an attacker's forged certificate instead of it to the web client through spoofing. In other words, the attacker captures PUs and transmits his forged public key PUa to the web client. And then the web client validates an attacker's forged certificate and the result of the validation fails and the web client displays that the certificate is unreliable to a user as Fig 3. There are many cases in real world which a carelessness user approves the unreliable one. If a user approves an attacker's forged certificate, the attacker captures PUs and transmits his forged public key PUa to the web client. The web client receives PUa of the attacker to begin the SSL communication, generates the session key SKca, encrypts SKca by the public key PUa of the web server and transmits the encrypted session key Cca to the attacker as formulas (3). And then the attacker decrypts the encrypted session key Cca by the private key PRa of the attacker and gains the session key SKca as formulas (4). The attacker shares the session key Cca with the client.
Security Requirements of Certificate Validation in Web Security
403
Fig. 3. Unreliable certificate
And then the attacker generates the session key SKas, encrypts SKas by the public key PUs of the web server and transmits the encrypted session key Cas to the web server as formulas (5). And the web server decrypts the encrypted session key Cas by the private key PRs of the web server as formulas (6). The attacker shares the session key Cas with the web server. Cca = E(PUa, SKca)
(3)
SKca = D(PRa, Cca) = D[PRa, E(PUa, SKca)]
(4)
Cas = E(PUs, SKas)
(5)
SKas = D(PRs, Cas) = D[PRs, E(PUs, SKas)]
(6)
Fig. 4. Certificate sniffer in SSL
404
H.B. Choi et al.
If a user approves an attacker's forged certificate, an attacker can capture or modify sensitive information. Fig. 4 shows that an attacker can gain session key between the web client and the attacker and between the attacker and the web server in the case of the secrete communication as SSL which distributes session key encrypted by a certificate. Because the session key is used to encrypt user data as user password, an attacker can capture and modify the user data. We have attacked a web client by using a hacking tool as Ettercap[9] and succeeded in capturing a victim's password. 3.2
Certificate Owner Verification Skip Vulnerability
There is an implementation case to validate the signature of a certificate but not to verify the owner of a certificate. And also there is a case not to verify the owner of a certificate because a certificate contents doesn't suit real environment. If the information of the subject which is contained in a certificate as Fig. 2 is not enough to identify the subject when considerable servers which support same services use one certificate, it may be difficult to verify the owner of a certificate at the client. This case can show that an attacker can gain session key between servers and an attacker and between an attacker and a client in the case of the secrete communication which distribute session key encrypted by a certificate as Fig. 3. Here, an attacker uses a certificate which is issued by CA which is the same as CA of server and client. 3.3
There is an implementation case to verify the owner of a certificate but not to validate the signature of a certificate when it is difficult to gain a certificate of CA in real environment. In this case, an attacker creates a certificate that the information of the subject is normal and the public key is modified in order to capture or modify the user data. As a receiver receives a transmitter's certificate to begin the communication, an attacker captures a transmitter's certificate, transmits this attacker's forged certificate instead of it to the receiver through sniffing. And then the receiver validates an attacker's forged certificate and the result of the validation succeeds. An attacker can capture or modify the user data as Fig. 3. 3.4
CRL Verification Skip Vulnerability
There is a case not to verify CRL of a certificate when it is difficult to gain CRL (Certificate Revocation List) which is managed by CA in real environment. In this case, an attacker can reuse a revoked certificate in order to capture or modify the user data. 3.5
Certificate Misuse Vulnerability
There is a case to use one certificate in order to perform the digital signature, encryption and key distribution. In this case, there is vulnerability which the private key can be disclosed potentially.
Security Requirements of Certificate Validation in Web Security
3.6
405
Validate Period Check Skip Vulnerability
There is a case not to check the validate period of a certificate. In this case, an attacker can reuse a revoked certificate in order to capture or modify the user data. 3.7
Weak Certificate Vulnerability
There is vulnerability that the fraud of a trusted certificate authority can be existed potentially at the case as following. The case is a weak certificate which is generated with a weak cryptographic algorithm as MD5 by a trusted certificate authority [2]. An attacker can disguises himself as a trusted certificate to a client or a web server in this case.
4
Security Requirements of Certificate Validation
We designed 6 security requirements of certificate validation from the vulnerabilities which were studied in the above section. These security requirements can reduce the vulnerability of the certificate application in web security services The security requirements to be proposed by this paper are user carelessness prevention(SR1), certificate owner verification skip prevention(SR2), Certificate signature verification skip prevention(SR3), CRL verification skip prevention(SR4), certificate misuse prevention(SR5), validate period check skip prevention(SR6), weak certificate prevention(SR7) as followings. 4.1
SR1: User Carelessness Prevention
When the result that a web client validates certificate of a web server fails, it shall be implemented to disconnect the communication with a web server automatically without showing the result to a user at web browser. If web browser must show the result of the validation to a user and a user must decide certificate acceptance when the result that a receiver validates a transmitter's certificate fails, it is necessary for a user to take security training about certificate application in order to understand that his secrete data are able to be captured. 4.2
The process to verify the owner of a certificate shall be contained in web security products. In order to help the verification of the owner of a certificate, CA sometimes issues a certificate which contains hash code to digest the owner's identification information. A simple method to verify the owner of a certificate is that an entity stores the other's certificate at local storage beforehand and compares it with a certificate which is received from the other part on communication. At the case not to verify the owner of a certificate because a certificate contents doesn't suit to real environment, the contents of the certificate shall be changed.
406
H.B. Choi et al.
If web servers which support same services use one certificate, the information of the subject which is contained in a certificate as Fig. 2 is not enough to identify the subject. It may be difficult to verify the owner of a certificate at the client in the case. If the information of the subject which is contained in a certificate of the web server consists of URLs or IP addresses which correspond to each web server, a client can verify the owner of a certificate as web server. When the client connects to the web server, the verification about the owner of a certificate is to compare connecting URL or IP address with URLs or IP addresses contained in a certificate which is received from the web server. 4.3
The process to verify the signature of a certificate shall be contained in web security products in order to prevent forging a certificate. The process is that a certificate of a web server is verified by a certificate of CA and the certificate of CA is verified by a certificate of Root CA. Here, the web client receives the certificate of CA and the certificate of Root CA from CA with on-line. When it is difficult to receive a certificate from CA in real environment, the web client receives a certificate of Root CA beforehand and stores it at local storage of the web client. And the web client receives a certificate of CA and a certificate of the web server in order to communicate with a web server. And then the web client performs the process to verify the signature of a certificate. 4.4
SR4: CRL Verification Skip Prevention
The process to verify CRL of a certificate shall be contained in web security products in order to prevent reusing a revoked certificate. When it is difficult to gain CRL (Certificate Revocation List) which is managed by CA in real environment, there is other method which a web server transmits CRL which the web server gains it from CA if it is in better environment. 4.5
SR5: Certificate Misuse Prevention
Web security products shall be implemented to use each certificate in digital signature, encryption and key distribution. The use of a certificate may be contained in a certificate. 4.6
SR6: Validate Period Check Skip Prevention
Web security products shall be implemented to check the validate period of a certificate. The validate period of a certificate is be contained in a certificate. 4.7
SR7: Weak Certificate Prevention
When a trusted CA generates a certificate, CA must not use with a weak cryptographic algorithm as MD5. Because an attacker can disguises himself as a trusted certificate to a client or a web server.
Security Requirements of Certificate Validation in Web Security
5
407
Conclusion
This paper proposed security requirements of certificate validation in web security. We studied the vulnerabilities of certificate application which are existed in web security services. There are the vulnerabilities about the validation of a certificate which depend on user's carelessness of certificate validation, the inconsistence contents of a certificate to operating environment, the incorrect implementation, the limitation to operating environment. These security requirements can reduce the vulnerabilities of the certificate application in web security services. These security requirements which are proposed by this paper can be useful in the area of the validation of a certificate when the PKI based key management system is established. These security requirements can be used as a development guideline of a web security product developer in the area of the validation of a certificate. Also, these security requirements can be used as an evaluation criteria of web security products in the area of the validation of a certificate when an evaluator assesses the web security products.
References 1. Housley, R., Ford, W., Polk, W., Solo, D.: Internet X.509 Public Key Infrastructure: Certificate and CRL Profile, RFC 3280 (April 2002) 2. Wang, X., Yu, H.: How to break MD5 and other hash functions. In: Cramer, R. (ed.) EUROCRYPT 2005. LNCS, vol. 3494, pp. 19–35. Springer, Heidelberg (2005); ISBN 3540-25910-4 3. Wilson, S.: The importance of PKI today. China Communications (2005) 4. Burkholdder, P.: SSL Man-in-the-Middle Attacks. The SANS Institute (February 2002) 5. Chomsiri, T.: HTTPS Hacking Protection. In: Proc. of Advanced Information Networking and Applications Workshops (AINAW), Mahasarakham, Thailand (May 2007) 6. Wagner, R.: Address Resolution Protocol Spoofing and Man-in-the-Middle Attacks, The SANS Institute (August 2001) 7. Cahill, V., Shand, B., Gray, E.: Using for Secure Collaboration in Uncertain Environments. Pervasive Computing 2(3), 52–61 (2003) 8. http://www.ietf.org//dyn/charter/pkix-chater.html 9. http://ettercap.sourcforge.net/index.php
The Measurement for the Underwater Electric Field Using a Underwater 3-Axis Electric Potential Sensor Jae-Hyun Kim, Jeong-Chay Jeon, and Jae-Geun Yoo Korea Electrical Safety Research Institute, 27, Sangcheon-ri, Cheongpyeong-myeon, Gapyeong-gun, Gyeonggi-do 477-814, Korea {azalea,cameleon,jgyoo}@kesco.or.kr
Abstract. As underground distribution lines are increasing, the accidents in those lines are also increasing. For example, 2 pedestrians were killed by an electric shock on a hand-hole flooded by localized torrential downpours in Korea in 2005. In order to prevent such disasters, we have devised a electric leakage alarm system for a flooded hand-hole. To use underwater sensors is one of the most effective ways in detection of underwater electric leakage. We have experimented a 3-axis underwater electric potential sensor to verify whether it can be used to detect underwater electric leakage. In this paper, it was confirmed to be able to detect underwater electric leakage using 3-axis underwater electric potential sensors. Keywords: underground power utilities, underwater electric leakage, underwater sensors, leakage alarm systems, a hand-hole.
torrential downpours in Korea in 2005. Before that accident 19 pedestrians were killed by an electric shock from street lights and traffic lights flooded by localized torrential downpours in Korea in 2001. After those accidents several safety measures were took and carried out. Measures for a hand-hole were replacing metal covers with FRP covers, inserting a insulation board between a cover and cables, using waterproof cable connectors and grounding a hand-hole frame. Measures for street lights and traffic lights were using a water-proof ELB(Earth Leakage Breaker), the distribution panel height over 1m and the ballast height over 0.6m. But the fault rate of underground distribution lines is still high despite such measures. Especially electric leakage fault rate remains very high compared to general facilities. Following the safety inspection for street lights and traffic lights in Korea, over 30,000 of those are exposed to danger of electric shock. The percentage of bad insulation resistance among total faults for street lights and traffic lights was 95.6% in Korea in 2010. For general low-voltage(under 600V) electric facilities included street lights and traffic lights the percentage of bad insulation resistance was 22.8% in the same year. Although any fatal accident was not occurred, we are confident that additional measures are needed. In this respect we have devised and experimented a electric leakage alarm system for a flooded hand-hole. To measure underwater electric field was used to detect electric leakage in a flooded hand-hole in the devised system. If electric leakage occurs in a flooded hand-hole due to a fault point such as bad connections between cables or a peeled point of a cable sheath, underwater current flows from the fault point to the ground. So it can be detected by measuring underwater leakage current whether electric leakage occurs in a flooded hand-hole. Table 1. The percentage of underground distribution lines of primary cities in the world City (Year) Percentage[%]
2
New York (2005) 72.2
London (2003) 100
Paris (2006) 100
Tokyo (2005) 86.3
Seoul (2007) 51.3
The Electric Leakage Alarm System for a Hand-Hole
A underground junction box is used for connection between underground distribution cables. The underground junction box for low voltage is called as a hand-hole. And the underground junction box for high voltage is generally called as a man-hole. Electric leakage faults are rarely occurred in underground distribution lines for high voltage due to the concentric neutral conductor layer in a CNCV distribution cables. But those faults can be easily occurred in low voltage underground distribution lines because of the absence of the concentric neutral conductor layer. If frequent excavations are done around a hand-hole, underground cables can be damaged. It can be hardly possible to detect those faults by measuring the insulation resistance of the cable, if the damaged point on the cable is not contacted any structure. But when it rains heavily and the hand-hole is flooded, electric leakage may break out. So these
410
J.-H. Kim, J.-C. Jeon, and J.-G. Yoo
faults are called as hidden faults or potential faults. A hand-hole is located mostly near a sidewalk where pedestrians can approach frequently. So we consider that the possibility of the accident, which is caused by electric leakage, in a hand-hole is higher than a man-hole. For this reason we have invented the electric leakage alarm system for a hand-hole. The diagram of underground distribution system for low-voltage in Korea is depicted in Fig. 1. The frame of each hand-hole must be connected with the neutral line and earthing for safety. The hand-hole electric leakage alarm system, which is invented, is depicted in Fig. 2. The underwater electric potential sensor is used for detecting electric leakage in a hand-hole. If a hand-hole is flooded, electric leakage current flows through
Fig. 1. The diagram of low-voltage underground distribution system in Korea
Fig. 2. The hand-hole electric leakage alarm system
The Measurement for the Underwater Electric Field
411
various paths from a leakage point to the ground. The electric field distribution is formed in underwater by the leakage current. By checking the underwater electric field we can verify if electric leakage occurs in a flooded hand-hole. If electric leakage occurs in a flooded hand-hole, the information of electric leakage is communicated to alarm devices by using the zig-bee technology. Then alarm is activated in the electric leakage alarm system.
3
The 3-Axis Electric Potential Sensor for the Electric Leakage Alarm System
We use the 3-axis electric potential sensor, which is invented by EMF Safety, to detect underwater electric field. EMF Safety is a venture business founded in 2000 and has the patent of the 3-axis electric field sensor. The configuration of the 3-axis electric potential sensor is depicted in Fig. 3. 6 metal plates are attached on 6 sides of the insulation cube in the sensor. 6 resistive lines are connected to 6 metal plates of the sensor. 3 pairs of electric potential differences(up and down, front and back, right and left) can be measured in a flooded hand-hole. The size of the sensor cube is 5 cm x 5 cm x 5 cm
Fig. 3. The configuration of the 3-axis electric potential sensor
We can calculate electric field approximately by using of the relationship between electric field and electric potential difference. If electric potential difference only along the x-axis exists and magnitude of that potential, which is detected in the 3axis electric potential sensor, is 5V, the magnitude of electric field can be calculated to 1 [V/cm]. (5 V / 5 cm)
E = −∇V = −
∂V G ∂V G ∂V G . x− y− z ∂x ∂y ∂z
(1)
412
J.-H. Kim, J.-C. Jeon, and J.-G. Yoo
Fig. 4. The picture of the 3-axis electric potential sensor
4
Experiments of the 3-Axis Underwater Electric Potential Sensor
Experiments were carried out with a 3-axis underwater electric potential sensor and a cylindrical flooded hand-hole(diameter: 65 cm, height: 65cm) in order to verify if this underwater sensor can be used to detect underwater electric leakage in a flooded hand-hole. The configuration of the experiment is shown in Fig. 5.
Fig. 5. The configuration of the experiment
If coordinate system is set up like in Fig. 5, the position of the sensor is (-32.5 cm, 0 cm, 0 cm). The sensor was submerged with 10 cm in depth. As a leakage point was moved, potential differences of the sensor were measured. We used low voltage source(220 V, 60 Hz) to make electric leakage in the flooded hand-hole. The substance of the hand-hole for experiments is concrete. And the hand-hole is well buried as shown in Fig. 5.
The Measurement for the Underwater Electric Field
413
Table 2. Potential Differences measured in the experiment Symbol Px1 Px2 Px3 Px4 Py1 Py2 Py3 Py4 Pz1 Pz2 Pz3 Pz4 Pz5 Pz6 Pz7
It is observed that x-axis potential difference polarity is reversed near the origin when moving the leakage point on the x-axis.(Px1~Px4) And the magnitude of y-axis potential difference is bigger when moving the leakage point toward the sensor closer.(Py1~ Py4) When the sensor is in same depth as a leakage point, the magnitude of horizontal potential difference is bigger as a leakage point source is closer and polarity of horizontal potential difference is reverse from front to back and side to side. So we can use the 3-axis underwater electric potential sensor to find out approximate horizontal directions of leakage point which is dipped in same depth with the sensor. When moving the leakage point on the z-axis in this experiment, z-axis potential difference polarity is not reversed near the origin. And z-axis potential difference magnitude is very small when the leakage point is deeper than the sensor. We inferred that the current tends to flow downward in a flooded hand-hole with a electric leakage point because grounding path does not exist on upper side. In other words upward current is much smaller than downward current because the underwater current flows toward lateral side and lower side in a flooded hand-hole owing to well formed grounding path. Through this experiment we can determine that lower part of the flooded hand-hole is adequate to install the 3-axis underwater electric potential sensor in finding electric leakage with only one sensor. If several 3-axis underwater electric potential sensors are used at later side and lower side, it can be easily possible to detect electric leakage in a flooded hand-hole. Generally the size of a hand-hole is smaller than 1m×1m×1m. So we expect that voltage over about 100 mV can be detected in 3-axis underwater potential difference sensors for inspection electric leakage of a flooded hand-hole having single or multiple leakage points.
414
5
J.-H. Kim, J.-C. Jeon, and J.-G. Yoo
Conclusion
We looked into the risk of flooded underground distribution line. And we proposed the electric leakage alarm system for a hand-hole comprising 3-axis underwater electric potential sensor, controller and alarm device. We chose 3-axis underwater electric potential sensor for the hand-hole electric leakage alarm system and test it in a flooded hand-hole. Through the experiment we can determine that lower part of the flooded hand-hole is adequate to install the 3-axis underwater electric potential sensor in finding electric leakage. If several 3-axis underwater electric potential sensors are used at later side and lower side, it can be easily possible to detect electric leakage in a flooded hand-hole.
References 1. Kim, H.-S., Kim, K.-C.: A study on the safety assessment and countermeasure on the lowvoltage handhole installed in the road (2007) 2. Coi, C.-S., Lee, B.-Y., Kwak, H.-R.: A study on the electric safety investigation in the electrical facilities on the road and temporary power installation (2005) 3. Kim, Y.-M.: Patent.: Detecting device of leakage current in the flooded water (2006) 4. Joung, J.-M., Lee, B.-S., Choi, J.-G., Joung, Y.-H., Park, C.-B., Song, I.-K.: Analysis of Transformer Fluid Aging from Overload Operation. In: KIEE Spring Conference, pp. 284– 287 (2006) 5. Lee, J.-H.: Current Status and Future Prospects for undergoround distribution. Journal of The Electrical Word, 22–25 (2008) 6. Lee, B.-H., Choi, J.-H., Bae, S.-B.: Make-up of Equivalent Circuit of Grounding System using Water Resistivity in Hemispherical Electrode System. Journal of The Electrical Word, KIIEE 22(8), 109–115 (2008)
Implementation of Web Services for ILBS* Changyong Han1, Jaegeol Yim2,**, and Gyeyoung Lee2 1
Dept. of Management, Dongguk University, GyeongJu, Gyeongbuk, 780-714 Korea [email protected] 2 Dept. Of Computer Engineering, Dongguk University, GyeongJu, Gyeongbuk, 780-714 Korea {yim,lky}@dongguk.ac.kr
Abstract. LBS (Location Based Service) prevails throughout the everyday life. Navigation, logistics, fleet control, POI (Point of Interest), and so on are all types of LBS. LBS is so useful that it should be available everywhere including inside of a building, subway stations, underground shopping malls, and downtown of a big city. Such LBSs provided inside of a building is called ILBS (Indoor LBS). A web service is a piece of software or a database published on Internet so that it can be used in web application development in the standard manner. Therefore, web service is deemed to be a promising technique of developing a web application. This paper introduces our implementation of web services of elementary functions of ILBS system. Indoor positioning, rendering drawings, playing contents are a few of such functions. Keywords: LBS, ILBS, Web Services, Indoor Positioning, Rendering Drawings.
1
Introduction
We can much more efficiently develop a new web application system by reusing existing programs running on various platforms all over the world. Web service [1-3] is one of the techniques of reusing programs developed by other organizations on their proprietory platforms. In web service, a published program is described in WSDL (Web Services Description Language) [4] and registered on UDDI (Universal Description, Discovery and Integration) [5]. Web application developers access a UDDI and search for a program that provides the function they want. They can use a published web service by inserting a few instructions of invoking it in their *
**
This research was supported by ‘Development of Global Culture and Tourism IPTV Broadcasting Station’ Project through the Industrial Infrastructure Program for Fundamental Technologies funded by the Ministry of Knowledge Economy(10037393). Yim’s research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(2011-0006942).
applications. The application communicates with the service provider in SOAP (Simple Object Access Protocol) [6]. Web service is deemed to be a key leading to efficiently developing a web application. There are so many web service providers. Sabre, Datalex, and Galileo provide web services for traveling and tourism business. Amazon web service provides web services for commerce. LBS (Location Based Service) prevails throughout the everyday life. Navigation, logistics, fleet control, POI (Point of Interest), and so on are all types of LBS. LBS is so useful that it should be available everywhere including inside of a building, subway stations, underground shopping malls, and downtown of a big city. Such LBSs provided inside of a building is called ILBS (Indoor LBS). This paper introduces our implementation of web services of elementary functions of ILBS system. Indoor positioning, rendering drawings, playing contents are a few of such functions.
2
Related Work
ILBS cannot be realized without an indoor positioning program. Therefore, an indoor positioning program should be the first web service we are implementing. Many algorithms of indoor programming have been published. Among them, the wireless LAN (Local Area Network) based trilateration and K-NN methods are discussed in this section. A few of the pioneers of indoor positioning are [7, 8]. [7] used the K-NN method to determine the location of a moving object. In K-NN, we build a look-up table in the first phase, or off-line phase. The entire area is covered by a rectangular grid of points called candidate points. The distance between two adjacent candidate points was 1 meter. At each of candidate points they measured the UDP packet's signal strength many times. To do this, they placed 3 desktop PCs on the floor of their test bed and called them base stations. A laptop computer was used as their moving object. At each of the candidate points, they sent UDP packet from the laptop to the base stations many times so that the base stations can record the average RSS (Received signal strength) as shown in Table 1. There is one-to-one mapping relation between the rows of the look-up table and the candidate points. A row of the look-up table is an ordered pair of (coordinate, a list of RSSs). A coordinate is an ordered pair of integers (x, y) representing the coordinates Table 1. An example look-up table of K-NN BS BS1 C.P CP1 CP2 CP3 …
BS2
BS3
-39
-55
-56
-40 -44 …
-56 -42 …
-55 -62 …
Implementation of Web Services for ILBS
417
of a candidate point where the laptop sent UDP packets. A list of signal strengths consists of three integers, RSS1, RSS2 and RSS3, where RSSi is the average signal strength scanned by the i-th base station. In the second phase, or on-line phase, the positioning program gathers RSSs the base stations receive at the moment. If the positioning program is running on the user’s handheld terminal, then the terminal itself will collect RSSs. Let X=(-40, -56, 54) be the vector of the collected RSSs. K-NN, then look up the look-up table and finds the closest candidate point, CP2 in the case of Table 1, and returns it as the user’s current location. If K equals 2, then it will find two closest candidate points and return the average of them as the user’s current location. If we measure N ranges, r1 , r2 ,", rN from N base stations, T n1 = ( X 1Y1Z1 ) ,", nN = ( X nYn Z n )T to a mobile terminal, m = ( x y z )T as shown
in Fig. 1, then we can estimate the coordinate of squaring,
m by using trilateration. By
2
ri can be expressed as the following:
(x − X i )2 + ( y − Yi )2 + (z − Zi )2 = ri2 , (for i = 1,2", N )
Fig. 1. A diagram to illustrate trilateration 2
By subtracting r1
from ri
2
G G = (i = 2," N ) , we have Ax = b , where
(X 2 − X1 ) A = 2 # ( X N − X 1 )
(Y2 − Y1 ) (Z 2 − Z1 ) #
(YN − Y1 )
x , xG = y z (Z N − Z1 ) #
(
) (
) (
) (
)
(
) (
) (
) (
)
X 2 − X 12 + Y22 − Y12 + Z 22 − Z12 − r22 − r12 G 2 b = # X N2 − X 12 + YN2 − Y12 + Z N2 − Z12 − rN2 − r12
When the coordinates are 3 dimensional, we need to have at least 4 base stations.
G Applying MMSE (Minimum Mean Square Error) method, we can obtain xˆ , our estimation of m with the following position estimates: G −1 G xˆ = AT A AT b
(
)
(1)
418
C. Han, J. Yim, and G. Lee
The trilateration cannot be applied for positioning, unless the relationship between RSS and distance is known. Once we know the relationship, we can calculate the distances (ranges) with RSSs. One of the ways of finding the relationship is making use of actually measured RSSs. In other words, we measure RSS every certain distance, 1 meter or 0.5 meter, from a base station many times and perform the regression analysis on the measurements as shown in Fig. 2. The X-coordinate is the distance from the base station to the portable laptop computer and the Y-coordinate is the signal strength measured at the distance. From the results, we obtained the following equation. Given a signal strength (y), we can estimate the distance (x) from the base station using this equation. y = -12.11Ln(x) – 33.66
(2)
Relation of distance and RSSI -20
Mean of measurements Propagation model
RSSI (dBm)
-30 -40 -50
y = -12.1ln(x) - 33.66
-60 -70 -80 1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20
Distance (m)
Fig. 2. Relation between distance and RSSI
[8] uses infrared ray and can determine in which room a moving object is located. Both radio and audio signals can also be utilized in indoor positioning. With the difference between the arrival times of these signals they calculate the distance between the sender and the receiver.
3
Implementation
It is obvious that reusing software significantly improves productivity. Web service is one of the most convenient ways of reusing software. Therefore, this paper introduces a few web services which perform essential functions of ILBS. 3.1
C# Web Service
We have implemented our web services using C#. The first step was installing IIS (Internet Information System). We have done this by the following steps on Windows server 2003: - Click "Start" menu and click "Control panel". - Double-click "Add or remove Programs" - Click "Add/Remove Windows Components" - Click "Application Server".
Implementation of Web Services for ILBS
419
- Click "Details" - Click "Internet Information Services Manager" - Click "Details" - Select optional components The procedure of creating a web method with Microsoft Visual Studio 2005 is as follows: - Click "file" on the main menu - Select "New" followed by click "Web Site" - Select "ASP.NET Web Service" among the installed templates. - [WebMehtod] template appears. A web method is placed under inetpub/wwwroot by default. In order to put our web methods under a meaningfully named directory we have built a virtual directory. The following is the procedure of creating a virtual directory: - create a directory, K_NN for example - click "start" -> "Control Panel" -> Performance and Maintenance" -> "Adminirtrative Tools" -> "Internet Information Services" - Keep clicking "+" under "Internet Information Services" until "Default Web Site" appears - Right-click "Default Web Sites" - Select "New Virtual Directory" - Specify "alias", KNN for example, on the wizard window - Browse to the directory, K_NN in this case, and press "NEXT" - Click "Finish" Once we have put our web method file in the virtual directory, we can use the method in our web program by the following procedure. - Right-click the name of the project on the Solution Explorer panel. - Click "Add Web Reference" - type in the URL, http://203.247.239.92/mhKnn for example. 3.2
Web Services for Trilateration
We have discussed two methods of indoor positioning, trilateration and K-NN in section 2. This section describes our web methods for trilateration. As we discussed before, the trilateration can determine the location of the mobile terminal if at least three distances from base stations to the mobile object are given. We also showed that we can estimate the distance with RF (Radio Frequence) signal strength. If we use existing equipments to obtain RF signal strength, then we can implement indoor positioning without any additional cost. WLAN (Wireless Local Area Network) provides such equipments and many WLAN based indoor positioning methods have been published [9-10]. WLAN based trilateration method scans signal strengths of APs (Access Point) and estimates the distances by evaluating Expression 2 with the signal strengths. Scanning signal strength should be done on a mobile terminal. Therefore, we have implemented a dynamic link library which reads signal strengths of APs. There are many ways of representing RF signal strength. Among them WLAN card uses RSSI (Received
420
C. Han, J. Yim, and G. Lee
Signal Strength Indicator). So, our library is a RSSI reader. We have implemented our RSSI reader in C# with Microsoft Visual .NET 2005 on a Samsung laptop SENS M40 equipped with an Intel(R) PRO/Wireless 2200BG Network Connection. We named the RSSI reader WirelessManager.dll. Using our WirelessManager, a mobile terminal can scan RSSIs and send them to the service provider. By evaluating expression 2, the service provider can estimate the distances between the mobile terminal and the APs. If the locations of the APs are known to the service provider, the service provider can perform the trilateration procedure to estimate the location of the mobile terminal. Therefore, we also have implemented a web method, uploadApLocation(), with which a user can upload the location information of APs. We named our web method which performs the trilateration procedure TrilaterationPositioning(). 3.3
Web Services for K-NN
Deployment of K-NN consists of off-line phase and real-time phase. It builds a lookup table during its off-line phase. If a client sends its location and RSSIs to the server then the server can add the data to the look-up table to make the table one step closer to perfect. Therefore, we have implemented a web method called uploadRSSIs. During the real-time phase, if a client sends RSSIs to the server, then the server returns the location of the client by looking up the look-up table. For real-time phase, we have implemented a web method called KnnLocation. 3.4
Web Services for RFID Positioning
Deployment of RFID positioning is also consists of off-line phase and real-time phase. This also builds a look-up table during the off-line phase. A row of the look-up table is a (tag-ID, location) pair. This information should be provided by a client. Therefore, we have implemented a web method called uploadTagId. During the realtime phase, if a client sends a tag-ID to the server, then the server returns the location of the client by looking up the look-up table. To do this, we have implemented a web method called RfidLocation. 3.5
Rendering Drawings
One of the most important elements of ILBS's user interface is a drawing. Rendering a drawing should be done on the client system. So, we have implemented a DLL which renders a drawing given a list of object types (called objectList) and another list of data (called dataList). The i-th element of dataList contains the necessary and sufficient information to draw a figure of the type specified by the i-th element of objectList. For example, if the i-th element of objectList is Line, then the i-th element of dataList consists of two points. Our rendering function also provides zoom-in, zoom-out, up, down, move, and so on. Fig. 3 shows experimental results of the rendering program. It is hard to imagine an ILBS that does not render drawings and a drawing should be provided by a client. Therefore, we have implemented a web method called
Implementation of Web Services for ILBS
421
Fig. 3. Execution results of our rendering program
Fig. 4. List of implemented web services
uploadDrawing with which a client can upload a drawing. We have also implemented a web method with which a client can select a drawing from a list of drawing names. Fig. 4 shows the web methods we have implemented. [4] proposed a ”generalized Petri net,” a type of modified Petri net for modeling rule-based systems. There are many formalized ways to represent Petri nets, including the metamodel and UML (Unified Modeling Language). [5] introduced a Petri net ontology written in UML, RDFS, and OWL (Web Ontology Language). This work shares many similarities with [6], which introduced a method of transforming OWL DL (Description Logic) into predicate/transition nets. Because the RDF model in the present paper is a subset of OWL DL, there is an overlap between the objectives of [6] and those of this paper. However, our paper proposes a method that can transform RDF models into CPNs, not into predicate/transition nets. Furthermore, we actually built a CPN model with CPNTools to verify the proposed method.
4
Implementation of an Application
Making use of those web methods and DLLs, we have implemented another DLL named "ViewControl.DLL". Then we have created an ASP page as follows. Fig. 5 shows a typical test result of the ASP page.
422
C. Han, J. Yim, and G. Lee
Fig. 5. A typical test result of our web service
5
Conclusions
Web service is one of the easiest ways of publishing our programs to the world so that anybody can call the program as if it is a function. Therefore, web service can improve the productivity of programming. This paper introduced several web methods which provide essential functions of ILBS. This paper also introduced a web program which uses the web methods we have published. We are planning to develop more web methods and develop a web service providing all the functions need for ILBS. We are also planning to expand the web methods to make them more versatile.
References 1. Alonso, G., Casati, F., Kuno, H., Machiraju, V.: Web Services: Concepts, Architectures and Applications. Springer, Heidelberg (2004) 2. Stal, M.: Web Services: Beyond Component-Based Computing. Comm. ACM 55(10) (2002) 3. Vogels, W.: Web Services Are Not Distributed Objects. IEEE Internet Computing 7(6) ( November/December 2003) 4. Booth, D.: Web Service Description Language (WSDL) Version 2.0. Technical report (2006) 5. W3C: UDDI technical white paper (2000) 6. Box, D.: Simple Object Access Protocol (SOAP) 1.1, W3C note (2000) 7. Bahl, P., Padmanabhan, V.: RADAR:An In-Building RF-Based User Location and Tracking System. In: INFOCOM 2000, pp. 775–784 (2000) 8. Harter, A., Hopper, A.: A New Location Technique for the Active Office. IEEE Personal Communications 4(5), 43–47 (1997) 9. Yim, J.: Introducing a decision tree-based indoor positioning technique. Expert Systems with Applications 34(2), 1296–1302 (2008) 10. Yim, J., Park, C., Joo, J., Jeong, S.: Extended Kalman Filter for Wireless LAN Based Indoor Positioning. Decision Support Systems 45, 960–971 (2008)
Petri Net Representation of Ontologies for Indoor Location-Based Services* Jaegeol Yim1, Jaehun Joo2, and Gyeyoung Lee1 1
Dept. of Computer Engineering, Dongguk University, GyeongJu, Gyeongbuk, 780-714 Korea {yim,lky}@dongguk.ac.kr 2 Dept. of Information Management, Dongguk University, GyeongJu, Gyeongbuk, 780-714 Korea [email protected]
Abstract. This paper presents a Petri net representation of an ontology for indoor location-based services. The ontology is described within the Resource Description Framework (RDF). Therefore, this paper proposes a method for transforming the RDF model into a CPN (colored Petri net). The proposed method represents the semantics of the RDF model in the CPN by mapping the classes and properties of the RDF model onto CPN places and representing the relationships between those classes and properties as token transitions in the CPN. To represent an RDF statement in the CPN, the method introduces a transition that produces a (complex) token composed of two ordered (simple) tokens: one corresponding to the subject and the other corresponding to the object of the statement. Applying the proposed method, we build a sample CPN for an RDF model and perform simulations using the model to answer RDF queries. This paper also introduces a simple ontology for an indoor locationbased service. Using the proposed transformation method, we transform the ontology into a CPN. Finally, we introduce our inference algorithm for the CPN and a prototype database system for demonstrating the practicality of our method. The results indicate that the proposed database system can resolve the semantic ambiguities in the query by using the ontology. Keywords: Petri net, colored Petri net, ontology, location-based services, indoor moving objects database.
1
Introduction
This paper introduces a method that can transform an RDF (Resource Description Framework) model into a colored Petri net (CPN) model. Performing a simulation on the CPN, we can obtain the answers of an RDF query. The study demonstrates the practicality of the proposed method by introducing a simple ontology for an indoor *
“This research was supported by ‘Industry, Academy and Research Institute Co-development of Techniques’ funded by the Small and Medium Business Administration”.
location-based service described in the RDF and showing how the proposed method can be used in the moving object database system’s query processing for the service. The RDF is a family of World Wide Web Consortium (W3C) specifications that can be used as a general method of modeling information. An RDF model consists of an RDFS (RDF Schema) layer and an RDF layer. The vocabulary used in RDF models is defined in the RDFS layer. A user defines the vocabulary, specifies the properties to be applied to each object, and describes the relationships between objects in the RDFS. That is, a user represents his knowledge in the RDFS. This implies that a wellformed RDF model is a type of ontology. The basic component of an RDF layer is a statement consisting of a subject, a predicate, and an object. A statement expresses a known fact The Petri net was first introduced in 1962, and since then, it has been widely used in computer science and engineering for system performance tests, communication protocol consistency and validation tests, and so forth. The Petri net has been popular because constructing a Petri net is easy and because mathematical methods of analyzing Petri net models are readily available [1]. The colored Petri net is a modified Petri net; it was developed so that the explosive increment of the size of the Petri net can be avoided when a complex phenomenon is modeled. There are several software tools that can provide convenient environments for constructing a Petri net model and for performing simulations. Among such tools, CPNTools [2] is designed to run on a desktop computer. This paper proposes a method that can transform an RDF model into a Petri net model. Applying the proposed method, we build a sample CPN for an RDF model and perform simulations using the model to answer RDF queries. An RDF model describes web resources in a machine-readable form. That is, it is possible to develop a software tool that interprets RDF models and inferences valuable information implied in those models. In other words, developing an RDF model is one thing, and developing an inference software tool is another. On the other hand, both the RDFS and RDF layers are mixed in a CPN model; the RDFS layer is reflected in the topology of the CPN model, whereas the RDF layer is reflected in the dynamics of the CPN model. As a result, we can extract inferential information by simulating CPN dynamics. To demonstrate the practicality of the proposed method, a simple indoor locationbased service system and an ontology for the system are considered. Applying the abovementioned transformation algorithm to the ontology, we construct a Petri net model. Then we introduce an inference engine for the Petri net model based on the ontology. We show that by using the inference engine, the moving object database of the indoor location-based service resolves ambiguities in queries.
2
Related Work
2.1
RDF
XML (Extensible Markup Language) is a general language that can be used for defining markups. Therefore, XML documents are typically used to exchange data between
Petri Net Representation of Ontologies for Indoor Location-Based Services
425
applications. However, XML cannot represent data semantics. The RDF is a data model in which this shortcoming of XML is reinforced. That is, a user can define his or her own vocabulary, properties of objects, and relationships between the objects by using RDFS language. The purpose of the RDF is to describe web resources, and the basic component of RDF is a statement consisting of a subject, a predicate, and an object. For example, “The name of the resource whose ID is ‘352’ is ‘David’” can be specified as follows: David, Where “uni” is a predefined name space. The subject of this statement is “352,” the predicate is the name, and the object is David. We assume here that the reader is familiar with the RDF. Please refer to References [3] for more information on the RDF. The sample RDF model shown in Fig. 1 is from [3]. We refer to this model frequently in the present paper
Fig. 1. RDF and RDFS layers
2.2
Colored Petri Net
Petri (1962) was the first to introduce the concept of the Petri net. Since then, the Petri net has been widely used for modeling and analyzing concurrent, nondeterministic, asynchronous, parallel, and/or distributed information processing systems. A Petri net is both graphical and mathematical. That is, we can visually represent the information and data flow of systems and establish state equations and mathematical models governing the behavior of systems by using the Petri net. A number of researchers have examined the applications and theory of Petri nets. Petri nets have been successfully applied in performance evaluation and communication protocols. Promising Petri net application areas include the modeling and analysis of distributed database systems, distributed software systems, flexible manufacturing control systems, concurrent and parallel programs, and so forth. We assume here that the reader is familiar with Petri nets. Please refer to References [1, 2] for more information on Petri nets.
426
2.3
J. Yim, J. Joo, and G. Lee
Other Related Research Results
CPNTools [2] is a software tool that can be used to easily build CNP models and perform simulations. CPNTools has been used to develop a number of Petri net models and validate their performance. In this study, we built a CPN model by using CPNTools and performed simulations to validate the proposed method. Petri nets have been used in a wide range of areas. [4] proposed a ”generalized Petri net,” a type of modified Petri net for modeling rule-based systems. There are many formalized ways to represent Petri nets, including the metamodel and UML (Unified Modeling Language). [5] introduced a Petri net ontology written in UML, RDFS, and OWL (Web Ontology Language). This work shares many similarities with [6], which introduced a method of transforming OWL DL (Description Logic) into predicate/transition nets. Because the RDF model in the present paper is a subset of OWL DL, there is an overlap between the objectives of [6] and those of this paper. However, our paper proposes a method that can transform RDF models into CPNs, not into predicate/transition nets. Furthermore, we actually built a CPN model with CPNTools to verify the proposed method.
3
Our Method of Transformation
We propose a method that can transform an RDF model into a CPN. The proposed method consists of two phases: the transform of the RDFS layer and that of the RDF layer. In the “transform of the RDFS layer” phase, we identify the core classes, map them onto CPN places, and identify the core properties. For each property, we introduce a transition so that the tokens at the place corresponding to the subject of the property can be transferred to the place corresponding to the object of the property. We then identify the domains and ranges of properties, and for each pair of the domain and range of a property, we introduce a transition that transfers an ordered pair of tokens: the first one from the place corresponding to the domain to the place corresponding to the property at a time and the second one from the place corresponding to the range to the same. This completes the topology of the CPN model. During the “transform of the RDF layer” phase, RDF statements are reflected in the CPN. For each statement, we identify its subject, predicate, and object and then do the following. First, if the predicate is ref:type, then we introduce a transition that generates a token representing the subject and adds it to the place corresponding to the object. Second, if the object is a container (e.g., rdf:Bag, rdf:Seq, or rdf:Alt), then for each case, we do the following. For rdf:Bag, we separate the statement into many statements and go to the third step. For rdf:Seq, we put all the elements of rdf:Seq into a list in order. For rdf:Alt, by using the random number generation function, we make the transition to produce one of the tokens corresponding to the elements of rdf:Alt. Third, we finally map the subject, predicate, and object onto places ps, pp, and po, respectively, and map the sentence onto the transition ts. Then we put a guard on ts so that it consumes a token x from ps and a token y from po and adds a token of (x, y) to pp when it fires. In this manner, we can also transform rdf:resource attributes and reifications into a CPN. Our transformation method is summarized in Table 1.
Petri Net Representation of Ontologies for Indoor Location-Based Services
427
Table 1. Algorithm for Transforming an RDF Model into a CPN
Input: an RDF model Output: CPN model 1. Identify the core classes, resources, classes, literals, properties, and statements of the RDFS and map them onto CPN places. 2. Represent the properties of RDFS:subClassOf and subPropertyOf by introducing a transition that transfers a token from the place corresponding to the subject class/property to the place corresponding to the object class/property when it fires. 3. Identify the domains and ranges of RDFS properties. For each property and its domain and range, we introduce a transition that consumes a token x from the place corresponding to the domain and a token y from the place corresponding to the range and adds a token of (x, y) to the place corresponding to the property. 4. For each RDF statement consisting of a subject, a predicate, and an object, we map its subject, predicate, and object onto places and the statement onto a transition. Then we put a guard to the transition so that it consumes tokens x and y from the places corresponding to the subject and the object, respectively, and adds a token of (x, y) to the place corresponding to the predicate. (A) In case of the predicate rdf:type, instead of mapping the predicate onto a place, we introduce a transition that transfers a token representing the subject to the place corresponding to the object class. (B) In case of the “Container,” we represent the Container by using the approach explained earlier in the discussion about the “transform of the RDF layer” phase.
4
Example of Transformation
A CPN model for the RDF model shown in Fig. 1 is shown in Fig. 2. The topology of the CPN in Fig. 2 is constructed during the “transform of the RDFS layer” phase (Steps 1 to 3 of the algorithm shown in Table 2), except for the source transitions (t1 and t2). In Step 1, we identify all the classes (e.g., Professor, Academic_staff, Literal, etc.) and properties (e.g., ID, Phone, isTaughtBy, etc.) and introduce a place for each. In Step 2, we introduce the transitions for subClassOf (t3, t4, t5, …) and subPropertyOf (t6). In Step 3, we introduce t7 and t8 to represent the domains and ranges. In Step 4, we introduce t1 and t2 to represent that David is a professor and DisMath is a course. We then put a guard to represent “DisMath isTaughtBy David.”
428
J. Yim, J. Joo, and G. Lee
Fig. 2. A CPN model for the RDF in Fig. 1
Drawing an inference is easy with a CPN model and we can extract answers for all RDF queries in CPN.
5
Ontology for Indoor Location-Based Services
Consider an indoor location-based service (ILBS) provided by a small, imaginary museum exhibit area shown in Fig. 3. In the area, there are 7 booths (Booth1 to 7), 2 vending machines (one for hot coffee (d2) and the other for soft drinks (d1)), and 12 chairs (c1 to c12). One of the main parts of the ILBS is the moving object database (MODB) in which the spatio-temporal information of moving objects is recorded. The ILBS translates questions such as “Who is visiting Booth 1?” into a query of the MODB and displays the answer to the query on the user interface of the ILBS. The MODB records the information in a relational database. A sample ontology for the imagined exhibit area is shown in Fig. 4. Three classes “people,” “place,” and “object” are used for the ontology for the museum’s LBS. The class “object” has various subclasses such as “vending-machine,” “chair,” “lounge,” “corridor,” and so on. C1, C2, …, Cn are the instances of “chair,” and their type is “chair.” The object property C-locatedIn relates the class “chair” to the class “sittingplace.” As shown in Fig. 5, the domain of “c-loacatedIn” is “chair,” and its range is “sitting-place.” It can indicate that people can sit on a chair; a person sitting on a chair is taking a break and not moving; a booth is a place where an exhibit is displayed and can be viewed by people; a booth is something a person cannot walk through, and so forth.
Petri Net Representation of Ontologies for Indoor Location-Based Services
429
Fig. 3. A Museum Exhibition Room
Fig. 4. An ontology for the ILBS on Fig. 3
Fig. 5. Object property c-locatedIn
As discussed in Section 3, the ontology can be transformed into the Petri net shown in Fig. 6. All places (e.g., staff, visitors, chairs, and booths) have only two input transitions in the figure, but in the real world, they would have many input transitions.
430
J. Yim, J. Joo, and G. Lee
For example, if there are 10 chairs, then the place chair has 10 input transitions. An object property in an ontology is represented as exactly one place in the Petri net. There are many object properties in the ontology, and there should be that many places representing properties; however, in Fig. 6, only one property is represented, and others are omitted. We have integrated the Petri net model of the ontology with our ILBS on the museum. For example, our system interprets “Find places to sit” as “Find all available chairs” by referring to the ontology.
Fig. 6. A Petri net representation of Fig. 5
6
Conclusions
This paper proposed a method that can transform RDF models into CPN models. To test the practicality of the proposed method, we introduced an ontology for a small, imaginary exhibit area in a museum. For the location-based service, we built a prototype Indoor Moving Objects Database (IMODB) system for the exhibit area. Our test results indicated that the IMODB can successfully resolve ambiguities in queries by referring to the proposed ontology.
References 1. Murata, T.: Petri nets: Properties, analysis and applications. Proceedings of the IEEE 77(4), 541–580 (1989) 2. http://wiki.daimi.au.dk/cpntools/cpntools.wiki 3. Antoniou, G., Harmelen, F.: A Semantic Web Primer. MIT Press (2004) 4. Shih, D., Chiang, H., Lin, B.: A Generalized Associative Petri net for Reasoning. IEEE Transactions on Knowledge and Data Engineering 19(9), 1241–1251 (2007) 5. Gasevic, D.: Petri Nets on the Semantic Web Guidelines and Infrastructure. ComSIS 1(2), 127–151 (2004) 6. Zhang, G., Meng, F., Jiang, C., Pang, J.: Using Petri Net to Reason with Rule and OWL. In: IEEE International Conference on Computer and Information Technology (CIT 2006), vol. 42 (2006)
A Platform for Real Time Brain-Waves Analysis System* Eun-Gyu Jeong1, Byungin Moon2, and Yong-Hwan Lee1,** 1 School of Electronic Engineering, Kumoh National Institute of Technology, Gumi, Korea {kyou4848,yhlee}@kumoh.ac.kr 2 Gyeongbuk National University, Daegu, Korea [email protected]
Abstract. There are many methods for human-computer interface. Recently, many researchers are studying about brain-signal. This is because not only the disabled can use a computer by their thought without their limbs but also it is convenient to general people. In this paper, we propose a platform for the real time electroencephalogram(EEG) and magnetoencephalography(MEG) analysis system. The signals on the scalp are so weak. Thus, the signals must be amplified by the amplifier. Filters are used to filter the incoming continuous signal before it is sampled by the analog-to-digital converter(ADC). The filtered EEG and MEG are analyzed and compacted by using FFT and JPEG2000, respectively. The platform is easy to reuse because of its simple block architecture and is implemented as a single chip on the Xilinx Virtex5 FPGA development board. Keywords: Electroencephalogram, EEG, Magnetoencephalography, MEG, Brain-computer interface, BCI.
1
Introduction
The brain-waves test is obtained by recording the brain's electric activities from the scalp electrodes. The electrodes are placed on the scalp to detect the electroencephalogram(EEG) and magnetoencephalography(MEG) in order to detect the brainwaves[1]. Since the frequency of EEG is different depending on a human state[2], the value of EEG is determined as delta, theta, alpha, beta and gamma by its frequency. Alpha brain-waves are typically the dominant cycle during dream states and light meditation. The alpha waves are oscillations with between 8 and 13 cycles per second. Beta brain-waves are dominant during normal waking consciousness, and range from 14 to 30 cycles per second. Theta brain-waves are considered to be 4 *
to 7 cycles per second and dominate during normal sleep and meditation. Delta brainwaves range from less than 1 cycle per second to about 3 cycles per second. These slow brain-waves tend to be dominating during the deepest sleep or deepest meditation states. Gamma brain-waves typically dominate during periods of advanced problem solving or critical thinking. They oscillate at between 30 and 60 cycles per second. The Table 1 shows the state of brain according to the frequency of EEG[3]. Table 1. Main EEG indicators Indicator Alpha wave(α) Beta wave(β) Gamma wave(γ) Theta wave(θ) Delta wave(δ)
Frequency Definition 8-13Hz 14-30Hz 30-60Hz 4-7Hz 0.5-3.5Hz
State Awake Concentration, Activity Excitement Sleep Deep Sleep
MEG is unique and effective diagnostic tool for evaluating brain function in a variety of surgical planning applications. This, when pooled with other information from the patient, forms the basis for determining whether surgery is the best option for treatment and, if so, how to plan it. MEG can determine exactly which bit of your brain controls actions such as speaking or moving arms and legs. In this paper, the architecture of the real time EEG and MEG analysis system is proposed. The architecture consists of an analog part and a digital part. The analog part is composed of a band-pass filter and amplifiers to process analog signal. The digital consists of analog-to-digital converter (ADC) which converts EEG from time domain to frequency domain and FPGA.
2
Electroencephalogram Analysis System
2.1
System Architecture
The block diagram of real time EEG and MEG analysis system is shown in Fig. 1. The sensor modules are on the scalp to measure EEG or MEG depending on the external mode selection. Since the amplitude of both signals is small, the amplifier of which gain is large enough and the band-pass filter are needed to measure them before signals are converted to digital codes though ADC. The digital codes are serialized to be transmitted to the chip that is shown in Fig. 1. The serialized data are transformed to parallel signals which are connected to JPEG2000 and FFT in order to be analyzed and be compacted for transmission at Data Transmitter module. The data are stored at FIFO. And the data are sent to the host PC through RS232 and USB (Universal Serial Bus).
A Platform for Real Time Brain-Waves Analysis System
433
Fig. 1. Block diagram of EEG and MEG analysis system
Fig. 2. Sensor module structure and digital MPW structure
2.2
Filter
High pass filter and low pass filter structure is shown in Fig. 3. The frequency of EEG signal of scalp is from 0.5Hz to hundreds of Hz with dozen of milli-voltage (mV) and the frequency of MEG signal is from 0.1Hz to several KHz with several micro-voltage(uV). The signals are different frequencies and voltages. Thus, EEG and MEG signal analyzed circuit get each different Amp and filter[4], as shown in Fig. 2.
434
E.-G. Jeong, B. Moon, and Y.-H. Lee
Fig. 3. High pass filter and low pass filter structure
2.3
Fourier Transform
EEG signal analysis FFT(Fast Fourier Transform) processor IP(Intellectual Property) for the diagnosis of brain function, EEG signals delta(δ), theta(θ), alpha(α), beta(β) and gamma(γ) frequency bands are analyzed. EEG signals using a Fourier transform is converted to the frequency band. The Fourier transform is defined by the equation(1). N
G k
Gn
cos
j2πnk N
j sin
j2πnk N
,k
0, 1, … , N
1
(1)
Fig. 4. 8-channel EEG signal analysis FFT processor IP structure
We use 256 point decimation-in-frequency as radix-2 fast Fourier transform algorithm[5] to improve the speed of the calculation. The EEG signal as input data has been broken into eight parallel data stream. In our implementation, input data is stored to a single port RAM whose size is 8x256x12. To avoid read and write conflict, we control data flow in memory using double buffering with two single ports RAMs. The output data is stored to a single port RAM whose size is 1x256x12 and sent to USB or UART FIFO.
A Platform for Real Time Brain-Waves Analysis System
435
The Hamming window function as shown in equation (2) is applied to the samples before calculating the FFT. It's basically a smoothing function that will make our samples look a bit nicer. The FFT class has a function by which we can enable Hamming window smoothing. ωn 2.4
0.54
0.46 cos 2πn/N ,
0
n
N
1
(2)
JPEG2000
64-channel MEG data from the sensors are for interface use serial to parallel interface module connected to JPEG2000 to be MEG data compacted for transmission at the data transmitter module.
Fig. 5. JPEG2000 IP for MEG compression structure
JPEG2000 IP structure system is shown in Fig 5. 8-bit MEG data is entered 64 times consecutively. The MEG data from DCT blocks divided up into four frequency bands and EBCOT tier-1 block through the EEG data compression, EBCOT tier-2 compressed stream of data to be packetized[6]. And then, EEG data is compressed in frames. 2.5
Brain-Signal Analysis Embedded Platform
The block diagram of embedded hardware architecture is shown in Fig. 6. (a). ARM11-based embedded real-time brain-waves monitor system configuration using analog brain-waves signal analysis chip and digital brain-waves analysis chip. ARM11-based embedded board is Embedded Linux to configure system to use for displays brain-waves at 7 inches TFT-LCD. The brain-waves are analyzed and compressed. The data stored in an external storage device. The block diagram of embedded software architecture is shown in Fig. 6. (b). Embedded software be modular, each functions. Data Collector data received from the external interface of the packet by analyzing the collected packets will be put to a temporary buffer stack. The received data is stored through the database. It also stores measured information for experimenter and environment.
Embedded brain-waves monitoring system is EEG signals the delta, theta, alpha, bbeta, and gamma power able to see the size as shown in Figure 7 is a graph. EEG signnals the delta, theta, alpha, beta,, and gamma is displayed red, yellow, green, light blue and blue. Absolute power of eacch wave is displayed with the color of the area.
Fiig. 7. EEG signal frequency spectrum
MEG signals are composed of 64 channels. The size of each component represeents the EEG activity. The sizee value is displayed in a row, as shown in Fig. 8. (a) for analysis should be possiblle. Alternatively, Fig. 8. (a) MEG signal brain activiities graph showed according to each part of the brain activity.
A Platform for Real Time Brain-Waves Analysis System
(a)
437
(b)
Fig. 8. (a)MEG sig gnal raw graph (b) MEG signal brain activities graph
4
Conclusion
We designed the platform for real time brain-waves analysis system in this papper. This platform is composeed of the analog part and digital part. The analog ppart processes brain-waves sign nals that swings within few microvolt at particular frequuency. The digital part analyzes that the brain-waves signal using FFT and JPEG20000. was This platform is easy to reeuse due to its simple block structure. The platform w implemented as FPGA deveelopment board with Virtex5 LXT FFG FPGA board. Real time brain-waves an nalysis system will contribute greatly to health care, eduucation and forensic science in n the future. Later, the brain-waves analysis system cann be utilized for that gaming, auttomotive and portable devices. Acknowledgements. This work was sponsored by ETRI System Semiconducctor SoC Industry Promotion Centter, Human Resource Development Project for S Convergence.
References 1. Casson, A., Yates, D., Sm mith, S.: Wearable Electroencephalography. IEEE Engineerinng in Medicine and Biology Maagazine: The Quarterly Magazine of the Engineering in Mediccine & Biology Society 29(3), 44–56 (2010) 2. Samei, S., Chamber, J.: EEG signal Processing. John Wiley & Sons, Ltd. (2007) 3. Oregon Health & Sciencee University Information, http://www.ohsu.edu 4. Haifeng, C., Donghee, Y., Y Jungtae, L.: Development of a Portable EEG Monitooring System based on WLAN. IEEE Circuits and Systems Society, 460–465 (2010) 5. Budiardja, R.D., Cardalll, C.Y.: Parallel FFT-based Poisson solver for isolated thhreedimensional systems. Com mputer Physics Communications 182(10), 2265–2275 (2011) 6. Auli-Llinas, F., Bilgin, A., A Marcellin, M.W.: FAST Rate Allocation through Steeepest Descent for JPEG2000 Video Transmission. IEEE Signal Processing Society 200(4), 1166–1173 (2011)
Design of FCM-Based Fuzzy Neural Networks and Its Optimization for Pattern Recognition Keon-Jun Park1 , Dong-Yoon Lee2, and Jong-Pil Lee3 1
Department of Information and Communication Engineering, Wonkwang University, 344-2, Shinyong-dong, Iksan-si, Chonbuk, 570-749 South Korea [email protected] 2 Department of Electrical Electronic Engineering, Joongbu University, 101, Daehak-ro, Chubu-myeon, Geumsan-gun, Chungnam, 312-702 South Korea [email protected] 3 Korea Electric Institute, 147-11 Sachang-dong, Heungdeok-gu, Cheongju-city, Chungbuk, 361-831 South Korea [email protected]
Abstract. In this paper, we introduce a new category of fuzzy neural network with multi-output based on fuzzy c-means clustering algorithm (FCM-based FNNm). The premise part of the rules of the proposed network is realized with the aid of the scatter partition of input space generated by FCM clustering algorithm. The number of the partition of input space is composed of the number of clusters and the individual partitioned spaces describe the fuzzy rules. Due to these characteristics, we may alleviate the problem of the curse of dimensionality. The consequence part of the rule is represented by polynomial functions with multi-output. And the coefficients of the polynomial functions are learned by BP algorithm. To optimize the parameters of FCM-based FNNm we consider real-coded genetic algorithms. The proposed network is evaluated with the use of numerical experimentation. Keywords: Fuzzy Neural Networks, FCM clustering, Scatter partition, Optimization, Genetic Algorithms.
consequence, there is a need for an optimization environment to construct and/or adjust a collection of linguistic rules. In this paper, we present the structure of fuzzy neural networks with multi-output by means of fuzzy c-means clustering algorithm [3] (FCM-based FNNm). The premise part of the rules is realized with the aid of the scatter partition of input space generated by FCM clustering algorithm. The consequence part of the rule is represented by polynomial functions with multi-output for pattern recognition. And the coefficients of the polynomial functions are learned by BP algorithm. We also optimize the parameters of the networks using real-coded genetic algorithms [4]. The proposed network is evaluated through the numeric experimentation.
2
Design of FCM-Based FNN with Multi-output
In this section, the form of fuzzy clustering if-then rules along with their development mechanism is discussed. More specifically, we elaborate on the three types of fuzzy inference and present the learning algorithm. 2.1
FCM Clustering Algorithm
The premise part of the FCM-based FNNm is developed by means of the Fuzzy CMeans clustering algorithm [3]. This algorithm is aimed at the formation of ‘c’ fuzzy sets (relations) in Rn. Consider the set X, which consists of N data points treated as vectors located in some n-dimensional Euclidean space. In clustering we assign patterns xp∈X into c clusters, which are represented by its prototypes vi∈Rn. The assignment to individual clusters is expressed in terms of the partition matrix U = [uip] where c
u
N
ip
i =1
= 1, 0 < uip < N , 1 ≤ p ≤ N , 1 ≤ i ≤ c
(1)
p =1
The objective function Q guiding the clustering is expressed as a sum of the distances of individual data from the prototypes v1, v2, …, and vc, c
N
Q = uipmi x p − v i
2
.
(2)
i =1 p =1
Here ⋅ denotes the Euclidean distance; ‘m’ stands for a fuzzification coefficient, mi>1.0. The resulting partition matrix is denoted by U = [uip]. The minimization of Q is realized through successive iterations by adjusting both the prototypes and entries of the partition matrix, that is min Q(U, v1, v2, …, vc). The corresponding formulas used in an iterative fashion read as follows. N
vi = u x p mi ip
p =1
N
u p =1
mi ip
x p − vi , u =1 ip xp − v j j =1 c
2
mi −1 .
(3)
440
2.2
K.-J. Park, D.-Y. Lee, and J.-P. Lee
The Structure of the FCM-Based FNNm
The structure of the FCM-based FNNm involves FCM clustering algorithm in the premise part and neural networks present in the consequence part of the rules. The overall topology of the network is illustrated in Fig. 1.
yˆ1
yˆ q
Fig. 1. The structure of FCM-based FNNm
FCM-based FNNm is implied by the fuzzy scatter partition of input spaces. In this sense, each rule can be viewed as a certain rule of the following format
Then ysj = f ( x1 , … , xd ) .
R j : If x1 and " and xd is F j
(4)
As far as inference schemes are concerned, we distinguish these cases Case 1 (Simplified Inference):
f = w sjo
(5)
f = w sj 0 + k =1 w sjk xk
(6)
Case 2 (Linear Inference): d
Case 3 (Modified Quadratic Inference):
f = w sj 0 + k =1 w sjk xk + k =1 i =k +1 w sjz xk xi , d
d
here, z = d + 1," d ( d + 1) / 2.
d
(7)
Design of FCM-Based Fuzzy Neural Networks
441
To be more specific, R j is the j-th fuzzy rule, while Fj denotes j-th membership grades using FCM clustering algorithm. w sjk , k = 0, " , d ( d + 1) / 2 are consequent parameters of the rule and s is the number of output. The functionality of each layer is described as follows. [Layer 1]. The nodes in this layer transfer the inputs. [Layer 2]. The nodes here are used to calculate the membership degrees using FCM clustering algorithm. [Layer 3]. The nodes in this layer realize a certain inference process. n
hsj = fˆ j y sj
(8)
j =1
[Layer 4]. The nodes in this layer compute the outputs. n
yˆ s = hsj
(9)
j =1
2.3
The Learning Algorithm
The parametric learning of the network is realized by adjusting connections of the neurons and as such it could be realized by running a standard Back-Propagation (BP) algorithm. The performance index is based on the Euclidean distance. As far as learning is concerned, the connections are adjusted in a standard fashion,
w sj 0 ( p + 1) = w sj 0 ( p) + Δw sj 0
(10)
where this update formula follows the gradient descent method. Quite commonly to accelerate convergence, a momentum coefficient α is being added to the learning expression. Then the complete update formula reads as follows
(
Δw sj 0 = η ( y ps − yˆ ps ) fˆ j + α w sj 0 ( p ) − w sj 0 ( p − 1)
)
(11)
(
)
(12)
(
)
(13)
Δw sjk = η ( y ps − yˆ ps ) fˆ j xk + α w sjk ( p ) − w sjk ( p − 1)
Δw sjz = η ( y ps − yˆ ps ) fˆ j xk xi + α w sjz ( p ) − w sjz ( p − 1) .
3
Optimization of FCM-Based FNNm
It has been demonstrated that genetic algorithms (GAs) [4] are useful in a global population-based optimization. GAs are shown to support robust search in complex search spaces. Given their stochastic character, these methods are less likely to get trapped in local minima in comparison to the performance offered by gradient-descent techniques. GAs start with a randomly generated population of l chromosomes positioned in solution (parameter) space. The population is evolved repeatedly toward achieving a
442
K.-J. Park, D.-Y. Lee, and J.-P. Lee
better overall fitness value. The search in the solution space is completed with the aid of several genetic operators. Let us briefly elaborate on the essence of these operators. Reproduction is a process in which the mating pool for the next generation becomes selected. Individual strings are copied into the mating pool according to their fitness function values. Crossover usually proceeds in two steps. Firstly, members from the mating pool are mated at random. Secondly, each pair of strings undergoes crossover as follows; a position l along the string is selected uniformly at random from the interval [1, l-1], where l is the length of the string. Swapping all characters between the positions k and l creates two new strings. Mutation is a random alteration of the value of a string position. In real coding, mutation is defined as an alternation at a random value in special boundary. Mutation occurs with a small probability. Those operators, combined with the proper definition of the fitness function, constitute the main body of the genetic computing. In order to optimize the parameters of the FCM-based FNNm, we determined the fuzzification coefficients associated with the corresponding clusters, the learning rate, and the momentum coefficient as the parameters.
4
Experimental Studies
In this paper, we use the Iris dataset. The Iris dataset is a collection of 150 Iris flowers of 3 kinds, with four attributes, leaf and petal width and length in cm. Three classes are the setosa, versicolor, and virginica. For the evaluation of the performance of the network, the random sub-sampling method was applied. The random sub-sampling was performed with 5 data splits of the data set. Each split was randomly selected from the training examples and the test examples with the ratio of 7:3. We experimented with the proposed network using the parameters outlined in Table 1 with the weight factor θ = 0.5 [5]. Table 2 summarizes the performance for FCM-based FNNm before optimization and Table 3 shows the performance for FCM-based FNNm using genetic optimization. From these tables we know that the optimized FCM-based FNNm is better than before optimization. From Table 3 we select the network that has ten fuzzy rules and linear inference engine. This network exhibits CR=98.29±0.43, PI=0.031±0.00 for training datasets and CR=99.56±0.99, PI=0.026±0.00 for testing datasets. Table 1. Initial parameters for GAs and FNNm
GAs
FNNm
Parameter Generation Population size Crossover rate Mutation rate Fuzzification coefficients Learning rate Moment coefficient
Value 100 50 0.65 0.1 1.0<mi≤2.5 0.0≤η≤0.01 0.0≤α≤0.001
Design of FCM-Based Fuzzy Neural Networks Table 2. Performance of the FCM-based FNNm No. of Inference Clusters (Case) 1 5 2 3 1 10 2 3 1 2 15 3 1 20 2 3
Fig. 2. Optimization process for the selected network
80
100
443
444
K.-J. Park, D.-Y. Lee, and J.-P. Lee
Fig. 2 presents the optimization procedure for the CR and PI when using ten rules with Case 2 (Linear Inference) obtained in successive generations of the genetic optimization. These figures depict the average values using the random sub-sampling.
5
Conclusions
In this paper, we have introduced fuzzy neural networks based on fuzzy c-means clustering algorithm and discussed its optimization using real-coded genetic algorithms. The input spaces of the proposed networks were divided as the scatter form using FCM clustering algorithm to generate the fuzzy rules. From this method, we could alleviate the problem of the curse of dimensionality and have designed fuzzy neural networks compact and easy. And genetic algorithms were also used for parametric optimization of the proposed networks. From the result in the previous section, we were able to design good networks. Through the use of performance we were able to achieve a balance between the approximation and generalization abilities of the resulting network. Finally it could be possible to apply to many fields.
References 1. Yamakawa, T.: A Neo Fuzzy Neuron and Its Application to System Identication and Predition of the System Behavior. In: Proceeding of the 2nd International Conference on Fuzzy logic & Neural Networks, pp. 447–483 (1992) 2. Buckley, J.J., Hayashi, Y.: Fuzzy neural networks: A survey. Fuzzy Sets Syst. 66, 1–13 (1994) 3. Bezdek, J.C.: Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press, New York (1981) 4. Golderg, D.E.: Genetic Algorithm in search, Optimization & Machine Learning. Addison wesley (1989) 5. Oh, S.K., Yoon, K.C., Kim, H.K.: The Design of Optimal Fuzzy-Neural Networks Structure by Means of GA and an Aggregate Weighted Performance Index. Journal of Control, Automation and Systems Engineering 6(3), 273–283 (2000)
A File Assignment Strategy for Parallel I/O System with Minimum I/O Contention Probability Bin Dong, Xiuqiao Li, Limin Xiao, and Li Ruan School of Computer Science and Engineering, Beihang University, Beijing 100191, P.R. China {bdong,xiuqiaoli}@cse.buaa.edu.cn, {xiaolm,ruanli}@buaa.edu.cn
Abstract. Many solutions have been proposed to tackle the problem of assigning files in a parallel I/O system. The primary objective of the existing solutions is either to balance the load among disks or to minimize the service time variance at each disk, whereas the dynamic characteristics of the file requests which would access these files are ignored. The studies on the dynamic I/O behaviors of applications show that the file requests targeted on the different popular files situated in the same disk may temporally compete with each another for the same disk. Consequently, the performance gained from the parallelism of multiple disks is degraded because this type of I/O contention turns the parallel I/O into sequential one. Hence, how to minimize the I/O contention among the file requests should become one of the new objectives which the file assignment strategy should take into consideration. In order to address this issue, this study proposes a new static file assignment algorithm named MinCP for parallel I/O system. Through assigning files sorted in their access rates onto multiple disks in round-robin fashion, the MinCP aims to minimize the I/O contention probability among file requests, thereby optimizing the mean response time of these requests. The experiment results show that the MinCP achieves optimal performance on mean response time among the existing schemes for comparison. Keywords: Parallel I/O system, file assignment algorithm, I/O contention probability, heuristic algorithms.
1
Introduction
The volumes of data that need to be stored grow rapidly in the recent past [1], whereas the rate at which the program can access data from single disk is still increasing much slowly [2]. As a result, the I/O has become the performance
The work described in this paper is supported by the fund of the National Natural Science Foundation of China under Grant No. 60973007 and No. 61003015, the Doctoral Fund of Ministry of Education of China under Grant No. 20101102110018, and the Fundamental Research Funds for the Central Universities under Grant No. YWF-10-02-058.
T.-h. Kim et al. (Eds.): GDC 2011, CCIS 261, pp. 445–454, 2011. c Springer-Verlag Berlin Heidelberg 2011
446
B. Dong et al.
bottleneck of dedicated applications when they read or write files [3,4]. Meanwhile, since the parallel I/O system permits the accesses from applications to these data executing concurrently, the parallel I/O systems like Redundant Array of Inexpensive Disks (RAID) [5] have received a lot of attention. While the partitioning of a file and then spreading the partitions onto different disks allow the single requests of the file to be serviced by multiple disks, the allocation of all files onto the disks plays an equally important role in the parallel I/O system performance. The reason is that proper allocation of all files permits the independent requests targeted on different files ( and even partitions ) to be concurrently served by the parallel I/O system. Thus, in order to reap the benefits of parallel I/O system, the files should be carefully assigned onto multiple disks before they are accessed. The file allocation problem (FAP) of the parallel I/O system has been extensively investigated in the previous literature [6,7,8,9,10] and the solutions to this problem fall into two camps: optimal and heuristic. The optimal file allocation problem actually is a NP-complete problem and can be transferred to other resource allocation problem [7]. The heuristic file assignment algorithm, on the other hand, owns low computational complexity and therefore becomes an effective and practical file allocation approach. The well-known heuristic file allocation algorithms include Greedy[6], Sort Partition (SP)[8], Hybrid Partition (HP)[8], Static Round-robin (SOR) [9], Balanced Allocation with Sort (BAS) [10], and Balanced Allocation with Sort for Batch (BASB) [10]. Typically, these heuristic file assignment algorithms try to optimize the mean response time of file requests, either through balancing the load among disks or via minimizing the service time variance at each disk, whereas most of them ignore the dynamic characteristics of the file requests targeted on the assigned files. Several previous studies [11,12,13,14] which focus on the dynamic I/O behaviors for both the scientific and enterprise applications show that the I/O requests from these applications own certain correlation. Specifically, the I/O requests from applications may temporally compete with each other for the same disk but leaving other disks idle [15,16,17]. Consequently, the load among the disks becomes imbalanced and the parallelism of multiple disks is underutilized [17]. One of the important reasons for the I/O contention is that the popular files are assigned onto the same disk and thus more requests from the applications are directed to this disk. Because the I/O contention turns the parallel I/O into sequential one [6], the performance gained from the parallelism of multiple disks would be reduced. Meanwhile, the file assignment algorithm which is responsible for allocating files onto disks decides the location and distribution of files among the disks [8]. Thus, it is necessary to redesign and implement a new and more efficient file assignment strategy which can take the I/O contention among the file requests into consideration. To achieve this target, this study proposes a new static heuristic file allocation algorithm named minimum I/O contention probability (MinCP). Firstly, we explore the relationship between the file access rate and I/O contention probability. The I/O contention probability theory proven in this study states that the I/O
File Assignment Strategy
447
contention probability between the popular files is higher than that between the unpopular files when the file size follows the Zipfian distribution [18] or the Uniform distribution [9,11]. In other words, the file requests targeted on the popular files compete with each other with higher probability than the requests targeted on the unpopular files do. The I/O contention probability theory also indicates that using the file access rate as the file allocation metric can help minimize the I/O contention probability among the file requests. Then, we design and implement the MinCP file assignment algorithm. The idea behind the MinCP is to assign all the files sorted according to their access rates onto disks in roundrobin fashion. As a result, the popular files should be distributed evenly among the disks and thus the parallelism of the multiple disks is fully utilized. Finally, we compare MinCP with the state of arts under the synthetic. When the file size follows Zipfian distribution, MinCP and SOR [9] own equal performance and they outperform other algorithms for comparison. When the file sizes obeys Uniform distribution, the MinCP consistently performs best among the existing algorithms for comparison. The rest of the article is organized as follows. The MinCP algorithm, including its mathematic model, design and analysis, is presented in Section 2. In Section 3, we evaluate the MinCP algorithm with the synthetic workload. Section 4 concludes this article and presents the future work.
2 2.1
The MinCP Algorithm Mathematic Model Description
The file allocation algorithm can be incorporated into the logical of disk array controller, a file system or the logical volume manager (LVM) [19] . Suppose that there are m non-partitioned files F = f1 , f2 , . . . , fm and n disks d1 , d2 , . . . , dn . The access information for each file includes the access rate λ, the file size s and the expected service time E(s). The expected service time of a file is assumed to be proportional to the file size because the rotation latency and seek time are negligible compared to the transfer time when the file is big and entirely accessed [9]. The previous studies [20,9,11] show that the file accesses among files obeys either the Zipfian ( also called Zipf-like ) distribution or the the Uniform distribution. Suppose that there are overall p requests which would access this parallel I/O system and the R = r1 , r2 , . . . , rp denotes the request set. During the time duration t, the expected total number of requests submitted to the system is t×( m i=1 λi ), and the expected number of requests targeted on the kth file is denoted with E(Ckt ): E(Ckt ) = t × λk .. Take ri , the ith request in the R and fk , the kth file in F , as example. We use ri = fk to denote the event that the request ri accesses the file fk . During the time duration t, the probability with which the request ri accesses the kth file is:
448
B. Dong et al.
E(Ckt ) λk P (ri = fk ) = m = m . t) E(C h=1 h h=1 λh In order to define the I/O contention probability and explore the relationship between the file access rate and the contention probability among I/O requests, we firstly define the following two access probability: Consecutive access probability, denoted with P s , is the chance that two files are accessed by two adjacent requests in the R. For example, the P s for s the fk and fj is: Pk,j = P (ri = fk , ri+1 = fj ). Overlapping access probability, denoted with P o , is the likelihood at which the expected service time of the first request covers the expected average arrival interval of the second request when the two files are accessed by two o o consecutive requests. For example, the Pk,j for the fk and fj is: Pk,j = 1 2
×
E(sk ) E(aj )
+
1 2
× 1.
Thus, the I/O contention probability, denoted with P c , can be defined as the file consecutive access probability multiplied with the I/O overlapping probability. In terms of the fk and fj , it can be computed with the following equation: s o s o P c (k, j) = Pk,j Pk,j + Pj,k Pj,k .
Theorem 1 (I/O Contention Probability Theorem). F is one file set and the file size obeys the Uniform or Zipfian distribution. Let f1st be the file with the most significance access rate and f2nd be the second. For two random files fr1 and fr2 , exclusive the file f1st and the file f2nd , in F , P c (ff1st , ff2nd ) > P c (ffr1 , ffr2 ). Proof (of theorem). Firstly, we prove the theory when the file size follows the Uniform distribution. For the f1st and the file f2nd , we have P c (f1st , f2nd ) = P sf1st ,f2nd Pfo1st ,f2nd + Pfs2nd ,f1st Pfo2nd ,f1st According to definition of the P s and P o , we can obtain Pfs1st ,f2nd > Pfsr1 ,fr2 , Pfs2nd ,f1st > Pfsr2 ,fr1 , Pfo1st ,f2nd > Pfor1 ,fr2 , Pfo2st ,f1nd > Pfor2 ,fr1 . Therefore, P c (f1st , f2nd ) >Pfsr1 ,fr2 Pfor1 ,fr2 + Pfsr2 ,fr1 Pfor2 ,fr1 , and the right of this inequality is equal to P c (fr1 , fr2 ). Thus, we obtain P c (f1st , f2nd ) >P c (fr1 , fr2 ), which completes the proof when the file size follows Uniform distribution.
File Assignment Strategy
449
Secondly, we prove the theorem when the file size obeys the Zipfian distribution. For the file f1st and the file f2nd , we have P c (f1st , f2nd ) = P sf1st ,f2nd Pfo1st ,f2nd + Pfs2nd ,f1st Pfo2nd ,f1st . According to the definition of the Pfs1st ,f2nd and Pfo1st ,f2nd , we have Pfs1st ,f2nd Pfo1st ,f2nd
When the file size obeys Zipfian distribution, we can assume that λf1st E(sf1st ) = C and λfr1 E(sfr2 ) = C . So Pfs1st ,f2nd (λf2nd )2 E(sfr1 )λfr1 s o m Pf1st ,f2nd Pf1st ,f2nd = + . 2 ( h=1 λh )2 × 2 Combined with the inequalities: λf2nd > λfr2 and Pfs1st ,f2nd > Pfsr1 ,fr2 , we obtain Pfs1st ,f2nd Pfo1st ,f2nd
the right part of which is equal to Pfsr1 ,fr2 Pfor1 ,fr2 . Thus, Pfs1st ,f2nd Pfo1st ,f2nd >Pfsr1 ,fr2 Pfor1 ,fr2 . By employing the same approach, we can prove Pfs2nd ,f1st Pfo2nd ,f1st >Pfsr2 ,fr1 Pfor2 ,fr1 . Combined with these two inequalities, we obtain P c (ff1st , ff2nd ) >P c (ffr1 , ffr2 ), which completes the proof under the Zipfian file size distribution. 2.2
MinCP File Assignment Strategy Description
The MinCP file allocation algorithm is outlined in the Fig. 1. It tries to minimize the I/O contention probability among file requests through assigning file sorted in descending order of their access rates among multiple disks. Be different from most existing solutions [8,9,10] which employ the expected service time or the heat of the files as the allocation metric, the MinCP allocates files according to their access rates. This is because, according to the theorem 1, the I/O contention probability between the popular files is much larger than that between the unpopular files and assigning file with access rate as metric can help reduce the I/O contention probability.
450
B. Dong et al.
Input:
m = number of files n = number of disks λi = access rate of file i si = expected service time of file i Output: {A1 , A2 , . . . , An } = assignment of m files to n disks 1: Compute the height hi ← λi × si for each file 2: Compute the average disk load ρ : ρ ← n1 m i=1 hi 3: Sort files in descending order of their access rate λi 4: d ← 1 5: for each file fi sorted by access rate do 6: if loadd + hi < ρ then 7: Ad ← Ad ∪ fi // assign file fi to disk d 8: loadd ← loadd + hi 9: d ← (d + 1)%(n − 1) // next disk 10: else 11: Search a disk t with enough free load space to store file fi 12: if Successful then 13: At ← At ∪ fi 14: loadt ← loadt + hi 15: else 16: An ← An ∪ fi // assign file fi to last disk n 17: loadn ← loadn + hi 18: end if 19: end if 20: end for Fig. 1. MinCP File Assignment Algorithm
After computing the average load of all the files in Step 2 and 3, the MinCP algorithm sorts files in descending order of their access rates. Then, all files are assigned onto disks in the round-robin way. When the popular files are allocated evenly among disk, the total load and requests targeted on these files would be distributed among disks evenly. As the step 6 indicates, when the load of the target disk plus the load of the being allocated file is less than the average load ρ, the file is assigned onto this disk. Otherwise, the MinCP would try to find one disk with enough free load space to hold this files. But, if there is no disk owning enough free load space, the file should be very large and own low access rate [8,9,10] and thus this file is allocated onto the last disk. The reason why we use the last disk to hold the large files is that the larges files would block the response for the requests for small files when larges files and small files are stored together. Table 1. Configurations of the disk Attributes Value
Attributes
Value
Attributes
Value
Interface SCSI Average seek time 5.4 msecs Revolutions per minute 10,000 rmp # of platters 1 Transfer rate 31Mb/sec Average rotation latency 3 msecs
File Assignment Strategy
3 3.1
451
Evaluation Simulation Setup
To compare the MinCP algorithm with the existing solutions for the nonpartitioned file assignment problem, we develop an event-driven parallel I/O system simulator with OMNet++ [21]. The disk simulated is the Cheetah disks [9], and the main configurations of it are presented in the Table 1. Table 2. Simulated synthetic workload characteristics Parameters
Value (Fixed) – (Varied)
# Files # Disks File access distribution File size distribution
In this section, the MinCP is compared with the existing static file algorithms under the workload with the Uniform file size distribution. When the file size obeys the Uniform distribution, the SOR outperforming both the SP and Greedy has been proven in the previous work [9]. So, to the best of our knowledge, SOR should be the most optimal static file assignment strategy under the Uniform file size distribution. In summary, four static file assignment algorithms, including SOR [9], SP [8], BAS [10], and MinCP, are compared in this test. The average response time with different aggregated access rates for the evaluated four algorithms are presented in the Fig. 2(a). Among the four algorithms, MinCP offers optimal average response time in all cases. This is because the I/O contention probability between the file requests is minimized when the MinCP assigns the files sorted in descending order of their access rates onto disks in round-robin fashion. In such case, the requests from the clients is distributed more evenly among disks, which can be verified by the Fig. 2(b). As a result, the parallelism of multiple disks is effectively utilized and the performance of the parallel I/O system can be efficiently reaped. Although SOR also allocates files in the round-robin way, it assumes the file access rate and service time are correlated. Thus, different popular files may be assigned onto same disk by SOR. As a result, the performance offered by SOR is degraded. The performance of both SP and BAS is worse than that of the MinCP or the SOR. This reason is that the SP and BAS group files with same service time on the same disk and therefore
452
B. Dong et al. 1.0
Variance of request number per server
Mean response time
1.0
BAS SP SOR MinCP
0.8
0.6
0.4
0.2
0.0
BAS SP SOR MinCP
0.8
0.6
0.4
0.2
0.0 25
50
100
200
300
400
500
600
700
800
900 1000
25
50
Aggregate access rate (1/second)
(a) Access rate
200
300
400
500
600
700
800
900 1000
(b) Request number
1.0
1.0
BAS SP SOR MinCP
0.8
Mean response time
Mean response time
100
Aggregate access rate (1/second)
0.6
0.4
0.2
0.0
BAS SP SOR MinCP
0.8
0.6
0.4
0.2
0.0
1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0
8
16
24
Maximum file size (1000M)
32
40
48
56
64
Number of disk
(c) File size
(d) Scalability
Fig. 2. Uniform file size distribution
more requests from applications would be directed onto same I/O server. When the file size follows Uniform distribution, this would make the I/O content at this server happen with high probability. Thus, the performance gained from the parallelism of multiple disks is reduced. Furthermore, we also explore the impact of file size range (Fig. 2(c)) and the number of disk (Fig. 2(d)) on these algorithms. Among the four algorithms for comparison, the MinCP always achieves optimal performance in all cases. Thus, the MinCP can offer performance improvement for a wide range of file size and own high scalability.
0.6
0.4
0.2
0.0
0.6
0.4
0.2
0.0 25
50
100
200
300
400
500
600
700
800
900 1000
Aggregate access rate (1/second)
(a) Access rate
1.0
BAS MinCP
0.8
50:50
60:40
70:30
80:20
Mean response time
1.0
BAS MinCP
0.8
Mean response time
Mean response time
1.0
90:10
Skew degree
(b) Skew degree Fig. 3. Zipfian file size distribution
BAS MinCP
0.8
0.6
0.4
0.2
0.0
8
16
24
32
40
48
Number of disk
(c) Scalability
56
64
File Assignment Strategy
3.3
453
Workload with Zipfian File Size Distribution
In this section, we evaluate the MinCP under the workload with the Zipfian file size distribution. Note that, the performance MinCP algorithm should be equal to that of SOR when the file size obeys Zipfian distribution. Thus, we only compare the proposed MinCP with BAS in this test. The experiment results on the average response time of all the I/O requests are presented in the Fig. 3(a). As the results show, the MinCP offers better performance than BAS in all cases. This is because that the SOR allocates file in the round-robin fashion and therefore popular files are more evenly distributed among disks. As a result, the I/O contention probability among the file requests is reduced. For the unpopular file, the BAS allocates one file with the most unpopular file onto one dedicated disk, but the MinCP proposed in this study employs one dedicated disk to hold more than one unpopular files. Thus, MinCP inherits all the advantages of SOR when file size follows Zipfian distribution. Furthermore, we also explore the impact of skew parameter (Fig. 3(b)) and the number of disk (Fig. 3(c)) on these algorithms. Among the four algorithms for comparison, the MinCP always achieves optimal performance in all cases. Thus, the MinCP can offer performance improvement for for a wide of skew parameter θ of the Zipfian distribution and own high scalability.
4
Conclusions
This study proposes a static nonpartitioned file assignment strategy called MinCP for parallel I/O system. The idea behind the MinCP to assign the files sorted in the descending of their access rates onto disks in the round-robin way. The experimental results show that the MinCP outperforms the existing static file allocation schemes when the file size follows the Uniform distribution or Zipfian distribution. Thus, the MinCP proposed in this study is one effective strategy to allocate the non-partitioned files in the parallel I/O system. Future studies can be performed in the following directions. Firstly, the MinCP should be extended to support the intra-request parallelism of single request [6]. Secondly, since some file statistic information is not known in advance in dynamic environment, one dynamic version of MinCP should be developed to assign files without further file access characteristics known and work online.
References 1. Short, J.E., Short, J.E., Bohn, R.E., Baru, C.: How Much Information? 2010 Report on Enterprise Server Information. Technical report, Global Information Industry Center (2011) 2. Deng, Y.: What is the future of disk drives, death or rebirth? ACM Comput. Surv. 43 (2011) 3. Isaila, F., Garcia Blas, J., Carretero, J., Latham, R., Ross, R.: Design and Evaluation of Multiple-Level Data Staging for Blue Gene Systems. IEEE Transactions on Parallel and Distributed Systems 22, 946–959 (2011)
454
B. Dong et al.
4. Lang, S., et al.: I/O performance challenges at leadership scale. In: SC 2009: Conference on High Performance Computing Networking, Storage and Analysis, pp. 1–12. ACM, New York (2009) 5. Chen, P.M., Lee, E.K., Gibson, G.A., Katz, R.H., Patterson, D.A.: RAID: highperformance, reliable secondary storage. ACM Comput. Surv. 26, 145–185 (1994) 6. Scheuermann, P., Weikum, G., Zabback, P.: Data partitioning and load balancing in parallel disk systems. The VLDB Journal 7, 48–66 (1998) 7. Dowdy, L.W., Foster, D.V.: Comparative Models of the File Assignment Problem. ACM Comput. Surv. 14, 287–313 (1982) 8. Lee, L.W., Scheuermann, P., Vingralek, R.: File Assignment in Parallel I/O Systems with Minimal Variance of Service Time. IEEE Transactions on Computers 49, 127–140 (2000) 9. Xie, T., Sun, Y.: A file assignment strategy independent of workload characteristic assumptions. Trans. Storage 5, 10:1–10:24 (2009) 10. Zhu, Y., Yu, Y., Wang, W., Tan, S., Low, T.C.: A Balanced Allocation Strategy for File Assignment in Parallel I/O Systems. In: 2010 IEEE Fifth International Conference on Networking, Architecture and Storage (NAS), pp. 257–266 (2010) 11. Wang, F., Xin, Q., Hong, B., Brandt, S.A., Miller, E.L., Long, D.D.E., Mclarty, T.T.: File System Workload Analysis for Large Scale Scientific Computing Applications. In: 12th NASA Goddard Conference on Mass Storage Systems and Technologies, pp. 139–152 (2004) 12. Nisar, A., Keng Liao, W., Choudhary, A.: Scaling parallel I/O performance through I/O delegate and caching system. In: SC 2008 International Conference for High Performance Computing, Networking, Storage, pp. 1 –12 (2008) 13. Feng, D., Zou, Q., Jiang, H., Zhu, Y.: A novel model for synthesizing parallel I/O workloads in scientific applications. In: 2008 IEEE International Conference on Cluster Computing, pp. 252–261 (2008) 14. Zoll, Q., Zhu, Y., Feng, D.: A study of self-similarity in parallel I/O workloads. In: IEEE / NASA Goddard Conference on Mass Storage Systems and Technologies, pp. 1–6 (2010) 15. Reddy, A.L.N., Wyllie, J.C.: I/O issues in a multimedia system. Computer 27, 69–74 (1994) 16. Carretero, J., No, J., Chen, P.: COMPASSION: A parallel I/O Runtime System Including Chunking and Compression for Irregular Applications. In: Ferreira, A., Rolim, J.D.P., Teng, S.-H. (eds.) IRREGULAR 1998. LNCS, vol. 1457, pp. 262– 273. Springer, Heidelberg (1998) 17. Dong, B., et al.: Self-acting load balancing with parallel sub file migration for parallel file system. In: Third International Joint Conference on Computational Science and Optimization, vol. 2, pp. 317–321. IEEE Computer Society, Washington (2010) 18. Cunha, C., Foster, D.: Characteristics of WWW client-based traces. Technical report, Boston University (1995) 19. Verma, A., Anand, A.: General store placement for response time minimization in parallel disks. J. Parallel Distrib. Comput. 67, 1286–1300 (2007) 20. Kwan, T., Mcgrath, R., Reed, D.: Ncsas world wide web server design and performance. Computer 28, 67–74 (1995) 21. Omnest Global, Inc. (OMNeT++ User Manual for Version 3.2)
Replication and Checkpoint Schemes for Task-Fault Tolerance in Campus-Wide Mobile Grid* SookKyong Choi1, JongHyuk Lee1, HeonChang Yu1, and Hwamin Lee2 1
Dept. of Computer Science Education, Korea University, Anam-Dong, Seongbuk-Gu, Seoul, Korea {csukyong,spurt,yuhc}@korea.ac.kr 2 Dept. of Computer Software Engineering, Soonchunhyang University, 336-745, Asan-si, Korea 1 [email protected]
Abstract. Mobile grid computing is a computing environment that incorporates mobile devices to an existing grid environment and supports users’ mobility. But this environment is not stable, so methodologies to cope with the reliability issue are needed. Fault tolerance approaches for task execution in grid computing can be categorized into replication and checkpoint. We apply these techniques to a SimGrid simulator to provide a fault tolerance for a mobile environment and show the results in this paper. The results demonstrate that the best solution for fault tolerance in mobile grid computing depends on the situations of the network. The contribution of this paper is the use of real-life trace data to simulate fault tolerance in a mobile grid computing. Keywords: mobile grid, replication, checkpoint, fault tolerance, reliability.
1
Introduction
Grid computing is a technology to offer high scalability, multitasking, and multitenancy. Grid computing provides an environment that effectively execute tasks by sharing and integrating computing resources in heterogeneous computing systems. Traditional grid environment is mostly implemented with resources in physically fixed locations. But this environment is evolving and extending the scope of resources into movable resources, by using mobile devices, to take advantage of the exploding population of mobile gadgets such as laptops, netbooks, and even smartphones. More recently, the mobile devices are equipped with high-performance computing engines with inexpensive price tags, attracting more and more users. Mobile grid computing is a computing environment that incorporates mobile devices to an existing grid environment and supports users’ mobility. In mobile grid computing environment, mobile devices can act as both a service provider and a service requester. *
This work was supported by National Research Foundation of Korea Grant funded by the Korean Government (2009-0070138)
As a service provider, it can provide its computing resources as a part of a mobile computing infrastructure. That is, mobile devices can execute tasks in mobile computing environment, like desktops or servers in a traditional wired computing. As a service requester, a mobile device can provide an interface to request services and use resources in mobile grid computing. The mobile computing environment exhibits its own characteristics, unlike the wired computing. Users of mobile devices may freely move in and out of a mobile network, which causes unstablility in network connection. Even worse, mobile devices may experience sudden power-offs by user or dead battery. We regard these cases that may results in outcome loss as faults. Therefore, the environment of mobile grid is not stable, so it should provide methodologies to cope with the reliability issue. Fault tolerance approaches for task execution in grid computing can be categorized into replication and checkpoint. Replication is used to improve reliability, fault tolerance, or accessibility using redundant resources. If the same data is duplicated and stored on several storages, it is referred to as data replication. If the same task is duplicated and executed several times, it is referred to as computation replication. Computation task can be executed on several resources simultaneously (replicated in space) or can be executed many times on one resources (replicated in time). Most of studies for replication have been on how to replicate data to improve accessibility and minimize access latencies. Checkpoint is used to minimize efforts of reexecuting long task from the beginning when the task is interrupted by faults such as a hardware failure, a software failure or resource unavailability. Checkpointing is taken periodically at specified intervals. Some information like a machine status, intermediate results and logs, are stored at every checkpoint interval on non-volatile storage. In case a fault occurs, information stored at the latest checkpoint is retrieved and used to execute task from the checkpoint. In the checkpoint scheme, it is crucial to figure out the optimal checkpoint interval (frequencies). Frequent checkpoints affect to the system performance due to overhead. Infrequent checkpoints may not provide fault tolerance. Especially, there are several studies for minimizing overhead in checkpoint. Compared to the traditional grid computing, a special care should be taken in mobile environment to consider characteristics of mobile devices, especially when mobile devices are utilized as resources for task processing. So, we apply these replication and checkpoint techniques to a SimGrid simulator and show the proper methods for fault tolerance when mobile devices are used as resources in this paper. To do this, we consider two incomplete cases caused by faults in mobile environment. One is the case that a task cannot be completed due to faults, and the other is the case that a task is completed but its outcome cannot be returned to requester due to faults. It is imperative to provide methods to withstand faults in a mobile grid computing.
2
Related Works
To withstand faults and provide efficient grid services, it is essential to investigate on task replicas that do not affect the process of other tasks, minimize waste of resources, and improve task execution time with replication. Early studies focused on fault tolerance methods using a fixed number of task duplicates [1,2,3,4]. Recent studies
Replication and Checkpoint Schemes for Task-Fault Tolerance
457
shifted to the dynamic decision of task replicas, that is, optimal number of task replica is dynamically determined according to system status [5,6,7]. Kang et al [1] researched on meeting users' QoS requirements and fostering balanced resource consumption by task replication using under-utilized and underpriced resources in computational service market composed of less reliable desktop PCs. They have a point of view that under-utilized and cheap resources can be exploited to build a high quality resource and hence can facilitate balanced resource usage. Katsaros et al [2] suggested installment-based scheduling and task replication to overcome intermittent connectivity in mobile grid. They defined that installment is consecutive fragments of a divisible task. And they used mobile device as a resource. However, the study assumed that every mobile device has the same performance and same network environment. Silva et al [3] proposed a dynamic replication approach, Workqueue with Replication (WQR). The study has not a fixed number of task replication and dynamically replicates tasks to idle resources when there is no other task to schedule. Moreover, replication is not used when system load is high. Dobber et al [4] proposed a method combining dynamic load balancing (DLB) and job replication (JR) to cope with unpredictable and dynamic nature of grid computing. This approach is a scheduling scheme that selects DLB or JR by comparing statistic time of expectation to threshold value set in advance. The study used four job replicas in JR. Limaye et al [5] proposed smart failover approach that is a job-site recovery to proactively handle failures and use backup server to overcome situation where the execution of client’s job is failed due to the failure of primary server. This approach supports transparent recovery by storing job states in local job-manager queue and transferring those states to the backup server. Priya et al [6] proposed task level fault tolerance in grid environment. With the technique, the study claimed that their checkpoint technique achieved the optimal load balance across different grid sites. Katsaros et al [7] considered independent checkpoint activities, proposed a statistical decision-making approach, and defined response time metrics of fault-tolerance performance and effectiveness. Paul J. Darby III et al [8] suggested checkpoint arrangement based on reliability of consumer and provider to maximize recovery probabilities of checkpoint data in mobile grid computing, which concerned about host mobility, dynamicity, less reliable wireless links, frequent disconnections, and variations in mobile environments. A mobile host simply sends its checkpoint data to neighboring mobile host, or saves checkpoint data for neighboring mobile host. There are some comparison studies to investigate the optimal solution for fault tolerance between checkpoint and replication in grid computing. However, studies in mobile grid environment have not been reported to the best of our knowledge. Wu et al [9] compared four different fault tolerance scheduling strategies based on genetic algorithm to improve reliability of grid system. These scheduling strategies are compared in terms of the performance metrics such as makespan, average turnaround time and job failure rate. The study reports that the checkpoint provides the best solution and replication is not suitable due to the overhead. Chtepen et al [10] introduced some heuristics using parameter dynamically based on grid status information and proposed a hybrid scheme based on both checkpoint and replication to improve job throughput against failures. The study reports that the dynamic
458
S. Choi et al.
adjustment of checkpoint frequency according to resource stability and remaining job execution time minimizes the checkpoint overhead. It also reports that postponing replication minimizes the replication cost. Moreover, the study reports that the hybrid scheme is the best approach when system information is not provided in advance.
3
Task-Fault Tolerance Approaches in Mobile Grid Environment
3.1
Architecture for Campus-Wide Mobile Grid Computing
In this paper, we assume campus-wide mobile grid computing environment similar to [11]. In Fig. 1, Mobile_Network indicates the entire campus that is composed of several Sub_Mobile_Networks. Sub_Mobile_Network indicates small network of each building in campus. MN(Mobile Node; mobile device)s are connected to the Sub_Mobile_Network by AP. There is a Global_Scheduler in a Mobile_Network. There are Local_Schedulers in Sub_Mobile_Networks and there are some Intermediate_Scheduler among Sub_Mobile_Networks. High level scheduler controls low level schedulers and supports load balancing among them. For this environment, we assign Proxies to act as schedulers in networks, so a proxy receives jobs from MNs and delivers outcomes to MNs. Because MNs act both an interface to mobile services and computational resources for task processing, an MN can submit tasks and process the submitted task. Submitted task from user is transmitted to a Local_Scheduler. The Local_Scheduler selects an MN in a Sub_Mobile_Network that including the Local Scheduler to process user’s task. Then the Local_Scheduler allocates the task to the selected MN. If Local_Scheduler has not enough resources, namely MNs, to process the task, the task is transmitted from the Local_Scheduler to an Intermediate_Scheduler or a Global_Scheduler recursively. Reversely, high level scheduler selects low level scheduler to return the task outcome, in turn. Therefore, task outcome is transmitted to the user submitting the task. Because we assume that tasks are divisible and independent in this paper, if big sized task is submitted and single mobile device cannot execute the task, then scheduler can divide the task. Fault includes network disconnection and power-off of mobile device. Further, we regard all types of cases that cause task failures in mobile grid as fault. But, we don’t consider Byzantine failure in this paper. When a fault is occurred, a task cannot complete due to the fault, and user submitting the task cannot receive task outcome even thought the task is completed. 3.2
Status of Mobile Device (MN)
The status of mobile device for processing a job is ‘Available’ or ‘Not Available’. An ‘Available’ status means that an MN can submit, process a job, and receive a job outcome. And a ‘Not Available’ status means that MN can’t be used for resource to provide grid service, because of battery shortage, network disconnection, or other physical faults. We don’t care ‘Not Available’ status in this paper, so we consider only if the MN is available or not for some time duration to process a job.
Replication and Checkpoint Schemes for Task-Fault Tolerance
459
Fig. 1. Architecture for Campus-wide Mobile Grid Computing
To select a proper MN for processing a job, a probability that a mobile device is available can be calculated by equation (1). n
Availabili tyij =
Pers k =1
ijk
Tavail + Tnotavail
(1)
n
where
Pers k =1
ijk
is the User Persistence(Pers) meaning a time duration during
which the ith MN stays at the jth Sub_Mobile_Network until the MN moves to other Sub_Mobile_Network or network link is down[12]. K is the number of available status. And Tavail and Tnotavail mean a time duration during which a MN can or cannot process a job respectively. 3.3
Task Scheduling Scheme
We apply 3 types of task scheduling schemes for fault tolerance in this paper; No replication and no checkpoint, task with Replication, and task with Checkpoint. The first refers to task execution without replication and checkpoint, the second and the third refer to task execution according to several replication options and checkpoint options. Basic principle in this paper is that a submitted task is processed in the Sub_Mobile_Network that includes the MN first. During a task execution, various situations that task outcome cannot be returned to user can be happened.
460
S. Choi et al.
3.3.1 No Replication and No Checkpoint When a task is submitted to a Sub_Mobile_Network, Local_Scheduler in the Sub_Mobile_Network selects an MN among MNs connected to the Sub_Mobile_Network. And the Local_Scheduler allocates the task to the MN to process the task. An MN with the best availability is selected first by equation (1). If there is no MN to process the task in the Sub_Mobile_Network, the request for task processing is transmitted to a high level scheduler. A high level scheduler selects an MN for task processing by using informations of the MN from IS(Information Service). Finally, if there is no MN to process the task in whole network, the task cannot be processed. And if the MN fails due to faults, other MNs can process the task from the beginning. We call this scheduling scheme No-No scheme in section 4. 3.3.2 Replication An original task submitted to a Sub_Mobile_Network can be replicated according to the replication options by a proxy as follows. • option_1 : replicate necessarily • option_2 : replicate when length of the task is long • option_3 : replicate when the size of remaining task is not much • option_4 : replicate when system load in a mobile network is low • option_5 : replicate when the ratio of replicated task in a mobile network is low The default number of replica for a job is one in order to reduce the waste of resources and overhead for replication in whole Mobile Network. System load increases with the number of MN processing a job. If a task is replicated, original task and replicas are executed on different MN simultaneously. Remaining size of the task, system load, and the ratio of replicated task are dynamically changed while the task is executed. So, replicated or not, which is determined dynamically by a proxy. When the faster task between original task and replica is completed, the other task has to be canceled. Fig. 2 shows the algorithm for replication. 3.3.3 Checkpoint An original task can be checkpointed according to the checkpoint options by a proxy as follows. • option_1 : checkpoint necessarily • option_2 : checkpoint when length of task is long • option_3 : checkpoint when the size of remaining task is not much • option_4 : checkpoint when system load in a mobile network is low • option_5 : checkpoint when the ratio of checkpoint task in a mobile network is low
Replication and Checkpoint Schemes for Task-Fault Tolerance
461
if (option_1_flag is true) { // replication process create same task. select another available MN to put the task. dispatch the task on the MN. } if (option_2_flag is true) { check task size. if (task is long) { start replication process of option_1 } } if (option_3_flag is true) { check system load. if (system load is low) { start replication process of option_1 } } if (option_4_flag is true) { check the size of remaining task. if (the size of remaining task is not much) { start replication process of option_1 } } if (option_5_flag is true) { check the ratio of replicated task in a mobile network. if (the ratio of replicated task is low) { start replication process of option_1 } } when one task out of the two, original task or replica, is over, kill the other task, and set the completion time by own completion time.
Fig. 2. Algorithm for replication
The default MN number of checkpointing for a job is one, like replication. And remaining size of the task, system load, and the ratio of checkpoint task are dynamically changed while the task is executed. So, checkpointed or not, which is determined dynamically by a proxy. Fig. 3 shows the algorithm for checkpoint. 3.3.4 Example Assume option_1_flag is false and option_2_flag is true, but length of task is short in each replication and checkpoint algorithm. Then a task will not be replicated or checkpointed, and the task is restarted from zero when faults occur.
4
Simulation
4.1
Data Analysis
To show the optimal methods for fault tolerance in a mobile grid computing, we analyzed a real-life trace: WLAN trace of Dartmouth campus[13]. The trace is a syslog produced by APs from September 1, 2005 to October 4, 2006[14]. We selected some part of trace as of June 6, 2006 for input of Simulation, which includes 987 APs and 3,367 mobile devices. After analyzing the trace, network information was extracted. The rate of the sessions maintained less than 2 hours is about 80% of all.
462
4.2
S. Choi et al.
Setup for Simulation
In this paper, we used Java based SimGrid simulator[15,16]. SimGrid is a toolkit that provides core functionalities for the simulation of distributed applications in heterogeneous distributed environments. To consider dynamic characteristics of mobile devices in a mobile grid computing, i.e. unstable and unpredictable characteristics, we setup simulation environment according to the results of real-life trace. And we used MSG of SimGrid that can support a proxy and MN clients. Configurations for simulation are presented in Table 1. The type of a task is a computational intensive task. The number of tasks is classified 10, 100, and 500, namely 3 types. And their length type of a size is categorized as short and long, namely 2 types. We assumed that all MN(mobile devices) has same performance in this paper. They can freely move around the Sub_Mobile_Network or other Mobile_Networks. The number of MN to submit, process, and receive a job is classified 10, 100, 1000, and 4500, namely 4 types. There are 3 types of task scheduling schemes; No replication and no checkpoint, task with Replication, and task with Checkpoint.
check the status of MN executing a task. if (MN fail) { if (option_1_flag is true) { // checkpoint process get the checkpoint information. select another available MN to put the task. restart the task from the checkpoint. } if (option_2_flag is true) { check task size. if (task is long) { start checkpoint process of option_1 } } if (option_3_flag is true) { check system load. if (system load is low) { start checkpoint process of option_1 } } if (option_4_flag is true) { check the size of remaining task. if (the size of remaining task is not much) { start checkpoint process of option_1 } } if (option_5_flag is true) { check the ratio of checkpoint task in a mobile network. if (the ratio of checkpoint task is low) { start checkpoint process of option_1 } } }
Fig. 3. Algorithm for checkpoint
Replication and Checkpoint Schemes for Task-Fault Tolerance
463
Table 1. Configurations for simulation •
task
• • •
Mobile devices
• • •
Methods for Task Fault-tolerance
4.3
• • •
type of task : computational intensive the number of tasks : 10, 100, and 200 length for task (task size) : short and long the number of mobile devices : 10, 100, 1000, and 3300 availability : 0 ~ 100% network connectivity : 0 or 1 (0:disconnection 1:connection) whole system load : high and low no replication and no checkpoint (No-No scheme) task with Replication task with Checkpoint
Results of Simulation
Fig. 4 shows the results of No-No scheme, namely No replication and no checkpoint, which is the basis of other scheduling scheme, Replication and Checkpoint. Completion time of tasks, the number of completed tasks, the number of failed tasks, and the number of used hosts are presented in Fig. 4. Y-axis means classification of tasks like 10/10/L that is composed of the number of tasks, the number of hosts, and size of the tasks(Long or Short). In addition, we analyze results of simulation for 4 evaluation metrics, namely average execution time of all tasks, completion rate of tasks, the number of completed tasks, and the utilization rate of resources according to the 5 options of each task scheduling scheme.
Fig. 4. No replication and no checkpoint
● Average Execution Time Fig. 5 and 6 show the average execution time in replication and checkpoint scheme, compared with No-No scheme represented by dotted line. Average execution time in replication scheme is shorter than in checkpoint scheme. It is due to checkpoint overhead that decision of checkpointing is made during processing, unlike replication.
464
S. Choi et al.
Fig. 5. Average Execution Time in Replication
Fig. 6. Average Execution Time in Checkpoint
● Completion rate of Tasks Fig. 7 and 8 show the completion rate of tasks in each scheme. In replication scheme, completion rate of tasks is getting higher according to the number of hosts. This means that if there are plenty of hosts, then using replication can increase the completion rate. In checkpoint scheme, the completion rate is mostly higher than in No-No scheme. This means that checkpoint scheme basically increase the probability of job completion to a certain extent.
Fig. 7. Completion Rate of Tasks in Replication
Replication and Checkpoint Schemes for Task-Fault Tolerance
465
Fig. 8. Completion Rate of Tasks in Checkpoint
● The number of Completed Tasks Fig. 9 and 10 represent the number of completed tasks in each scheme. The number of completed tasks in checkpoint scheme is higher than in No-No and replication scheme. That’s because a checkpoint scheme deal with the faults of resources more actively
Fig. 9. The number of Completed Tasks in Replication
Fig. 10. The number of Completed Tasks in Checkpoint
● Utilization Rate of Resources Fig. 11 and 12 show the utilization rate of resources in each scheme. There is little difference in two scheme. That seems an extra resource for replication and another resource for checkpoint are the same sense.
466
S. Choi et al.
Fig. 11. Utilization Rate of Resources in Replication
Fig. 12. Utilization Rate of Resources in Checkpoint
5
Conclusion and Future Work
In this paper, we present fault-tolerance methods for mobile computing environment. We apply replication and checkpoint scheduling scheme to real-life trace data. The results demonstrate that the best solution for fault tolerance in mobile grid computing depends on the situations of the whole network. Average execution time is shorter in replication scheme, and the number of completed tasks is bigger in checkpoint scheme. So if there are plenty of resources in a network and they are comparatively reliable, then replication scheme is the better, else checkpoint scheme is the better. We have a plan to apply the real-life trace to mobile cloud computing environment using GloudSim simulator. And we will conduct provisioning of reliable services based on SLA and QoS in mobile cloud computing.
References 1. Kang, W., Huang, H.H., Grimshaw, A.: A highly available job execution service in computational service market. In: 8th IEEE/ACM International Conference on Grid Computing, September 19-21, pp. 275–282 (2007) 2. Katsaros, K., Polyzos, G.C.: Evaluation of scheduling policies in a Mobile Grid architecture. In: Proc. International Symposium on Performance Evaluation of Computer and Telecommunication Systems (SPECTS 2008), Edinburgh, UK (June 2008) 3. Silva, D., Cirne, W., Brasileiro, F.: Trading Cycles for Information: Using Replication to Schedule Bag-of-Tasks Applications on Computational Grids. In: Kosch, H., Böszörményi, L., Hellwagner, H. (eds.) Euro-Par 2003. LNCS, vol. 2790, pp. 169–180. Springer, Heidelberg (2003)
Replication and Checkpoint Schemes for Task-Fault Tolerance
467
4. Dobber, M., van der Mei, R., Koole, G.: Dynamic Load Balancing and Job Replication in a Global-Scale Grid Environment: A Comparison. IEEE Transactions on Parallel and Distributed Systems 20(2), 207–218 (2009) 5. Limaye, K., Leangsuksun, C.B., et al.: Job-Site Level Fault Tolerance for Cluster and Grid environments. In: The 2005 IEEE Cluster Computing, Boston, MA, September 27-30 (2005) 6. Baghavathi Priya, S., Prakash, M., Dhawan, K.K.: Fault Tolerance-Genetic Algorithm for Grid Task Scheduling using Check Point. In: Sixth International Conference on Grid and Cooperative Computing, GCC 2007 (2007) 7. Katsaros, P., Angelis, L., Lazos, C.: "Performance and Effectiveness Trade-Off for Checkpointing in Fault-Tolerant Distributed Systems. Concurrency and Computation: Practice and Experience 19(1), 37–63 (2007) 8. Darby III, P.J., Tzeng, N.-F.: Decentralized QoS-Aware Checkpointing Arrangement in Mobile Grid Computing. IEEE Transactions On Mobile Computing 9(8), 1173–1186 (2010) 9. Wu, C.-C., Lai, K.-C., Sun, R.-Y.: GA-Based Job Scheduling Strategies for Fault Tolerant Grid Systems. In: IEEE Asia-Pacific Services Computing Conference (2008) 10. Chtepen, M., Claeys, F.H.A., Dhoedt, B., De Turck, F., Demeester, P., Vanrolleghem, P.A.: Adaptive Task Checkpointing and Replication: Toward Efficient Fault-Tolerant Grids. IEEE Transactions on Parallel and Distributed Systems 20(2), 180–190 (2009) 11. Katsaros, K., Polyzos, G.C.: Optimizing Operation of a Hierarchical Campus-wide Mobile Grid for Intermittent Wireless Connectivity. In: 15th IEEE Workshop on Local & Metropolitan Area Networks, LANMAN 2007, June 10-13, pp. 111–116 (2007) 12. Balazinska, M., Castro, P.: Characterizing Mobility and Network Usage in a CorporateWireless Local-Area Network. In: Proceedings of the First International Conference on Mobile Systems, Applications, and Services (2003) 13. Henderson, T., Kotz, D.: CRAWDAD trace dartmouth/campus/syslog/05_06 (February 8, 2007), http://crawdad.cs.dartmouth.edu 14. Lee, J.H., Choi, S.J., Suh, T., Yu, H.C.: Mobility-aware Balanced Scheduling Algorithm in Mobile Grid Based on Mobile Agent. The Knowledge Engineering Review (2010) (accepted for publication) 15. Buyya, R., Murshed, M.: GridSim: A Toolkit for the Modeling and Simulation of Distributed Resource Management and Scheduling for Grid Computing. J. Concurrency and Computation: Practice and Experience 14, 13–15 (2002) 16. Sulistio, A., Cibej, U., Venugopal, S., Robic, B., Buyya, R.: A toolkit for modelling and simulating data Grids: an extension to GridSim. Concurrency and Computation: Practice & Experience 20(13), 1591–1609 (2008)
Dynamic Arrangement of Control in a Personalized Learning System Based on User Competency Youngseok Lee1, Jungwon Cho2,*, Sungjae Han3, and Byung-Uk Choi4 1
Department of Electronics Computer Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul, 133-791 South Korea [email protected] 2 Department of Computer Education, Jeju National University, 102 Jejudaehakno, Jeju-si, Jeju-do, 690-756 South Korea [email protected] 3 Department of Electronics Computer Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul, 133-791 South Korea [email protected] 4 Division of Computer Science & Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul, 133-791 South Korea [email protected]
Abstract. This paper proposes the e-Space Manager to reconstruct a learning environment and describes the method used to modify components. The proposed system analyzes user competencies, and the results influence which components the users can select. That is, it provides not only for the rearrangement of components but also for the modification of components according to the user’s competencies. To verify the effectiveness of the proposed system, the e-Space Manager was applied to an SQL-based learning system. Users chose the optional learning contents that they wanted. It was confirmed that information about the learner's competency influenced the learning components. A survey of users showed in a higher satisfaction rating for the proposed study system than for the previous system. Using this system, tutors not only could analyze the learners’ abilities precisely based on their competencies but also could recommend study areas and provide a customized environment to each individual. Keywords: Customized Learning, User Competency, e-Learning, Learning Environment, Web 2.0.
1
Introduction
With the advent of Web 2.0 technology has come an increasing interest in the personalization of web sites. In the early days of the web, web sites specified content
Dynamic Arrangement of Control in a Personalized Learning System
469
in a uniform way, delivering essentially the same content to all users according to the web technology of the day. Today, a major concept driving web development is the concept of personalized services [1]. This paper introduces the concept of Web 2.0 to traditional hypertext markup language in a manner that is not possible with traditional rich applications. This concept makes possible a variety of new services that require implementation of a new type of interface, which is the subject of an active area of current research [2]. A larger variety of services is possible with the personalization of sites based on the users’ needs. An example of a typical personalized service is Google's personalized page, which permits the relocation of services such as widgets. However, this relocation is limited to existing components; it requires storing relocation information and does not provide for the creation of new features [3, 4]. Personalized service applied to an e-learning system offers convenience [1]. In this paper, we propose a learning environment using a module that we call e-Space Manager, which provides personalized service, as system users can select components to design their own learning spaces. The learning content as well as an analysis of the user's competency output format, selected by the user, changes the system directly. Thus, the selection of competency elements is user-centered, and the system can take advantage of the respective component to alter the overall system configuration. The learning environment is modified by the user through competency modification and reconfiguration of components. To evaluate the proposed system, learner surveys were conducted. The survey results indicated a high level of student satisfaction, which illustrates the superiority of this new system.
2
Review of Pertinent Literature
2.1
Customized Services
Since the advent of the Internet as a marketing tool, customer relations management (CRM or eCRM) software has boosted interest in the "personalization" of web services. Although the concept of personalization has long been a part of traditional offline marketing, the concept re-emerged during the 2000s with respect to web development [5]. Personalized services based on recommendations from the user profile allow modification of the selection of services and user-based customization of services [6, 7]. Web 2.0 permits personalization of a web site from an existing structure according to the changing needs of users, and it enables users to receive desired information quickly, allowing for intensive services to be provided [1, 2]. As shown in Fig. 1, a typical portal site such as Google can provide distinctive and personalized services. These personalized services are based on user recommendations. Recent advances in technology, combined with customization services, permit the continual development of personalized services.
470
Y. Lee et al.
Fig. 1. Customized service of iGoogle
2.2
Competency
Currently, there is considerable interest in competency-based learning models. Competency is defined as possession of a specific combination of complex knowledge, skills, and attitudes necessary to achieve a specific performance level at a task or job [8]. In competency-based learning, the competencies required are first defined using specific terminology to describe the knowledge, work, and learning outcomes desired. Both competencies and taxonomies, which are key means by which the competencies are classified, are presented in various ways. Fig. 2 illustrates the concept of using a skill gap analysis to target competencies that need to be acquired. The current competence of a learner is determined based on a pre-defined competency model, and the target competencies are derived. Capacity data, reusable competency definitions, and proof of competence can all be categorized by target date or complexity. For example, an estimate of predicted technological advances in the quantitative component can be defined [8, 9].
Fig. 2. Competency-based skill gap analysis
Dynamic Arrangement of Control in a Personalized Learning System
3
471
System Design
In this paper, we propose a learning space, called the e-Space Manager, that allows the user to reconfigure the core engine. A personalized service provides help for the overall system configuration. Through competency modification, the user is able to reconfigure components. 3.1
System Architecture
The proposed system architecture is shown in Fig. 3. Learners access the learning and evaluation modules through the interface module. We use the phrase “learning space representation” of the learning content to refer to the output area of the webpage. A component that is placed in the learning space is called a learning component. A component of the basic study and recommendations is divided into learning components. A basic learning component for a general learning content is used for the purpose of representation. The representation of content can be changed by the user (i.e., is vulnerable to the user). The recommended learning component is the vulnerable content combined with the core functionality of the component. The learning content presented in the learning component is selected from the vulnerable areas according to the learner’s target competencies.
Fig. 3. System architecture
User-selectable configurations are provided in the learning environment. The user's goals, determined through an analysis of the competency model chosen by the learner, are reflected in the components' variable settings, and tasks are performed automatically. The program can generate a page showing the learner's target competency model. Based on the information generated in the existing set of components, changes are made to the program’s settings. When the information is insufficient to update the system, the learning system can use comparators to recommend reorganization.
472
3.2
Y. Lee et al.
Competency Model
The competency model uses SQL as the data manipulation language (DML). A prerequisite course is set to determine a student’s problem-solving abilities. After unpacking all the problems, the student’s answers are compared against the competency model and tokenized. From the results, the evidence records are derived, and the corresponding competency tables are stored in appropriate fields.
Fig. 4. Competency gap analysis process for SQL query
The proposed method first determines the learner’s existing competencies, then determines which competencies need to be acquired, and finally derives the steps necessary to acquire the goal competencies. Data collected from students for this competency record are compared against the competency model. As a result, the learner’s acquired skills and competencies are determined. Goals that the student failed to accomplish indicate the need to acquire essential information or the need to derive competencies. This record is created by performing a comparative analysis against the competency model. The “SELECT” query is used to determine the competencies required in each of the “Grammar” and “Parameter” fields and determines the lack of any item. Fig. 4 shows how SQL queries are used to perform a competency gap analysis. Components that reflect the learner's competence are selected and combined into a single component of the process shown in Fig. 5; weak learners are shown content separately as a new component.
Dynamic Arrangement of Control in a Personalized Learning System
473
Fig. 5. Competency gap analysis process for SQL query
4
System Evaluation
Fig. 6 shows the competency analysis calculated from the learner's answers to the competency model, and a competency map determined by comparing the learner's screen representation to the competency model. Professors can target students who are identified as lacking certain skills and capacity. After the student has completed the problem-solving section, the records management module, which correlates each item with the student's competency information, can be viewed directly.
Fig. 6. Analysis result of user competency
Fig. 7. Analysis result of survey
“Existing system” is the existing learning system, which consists of standardized content and learning components. “Update system” means to reset the competency fields to reflect the results from the learning components. “Proposed system” means to update the system and create new components. Personalized content is created in a controlled environment, based on the user’s recommendations, content selection, and
474
Y. Lee et al.
evaluation of content. The system examines which areas the student chooses to study to evaluate the readability of vulnerable areas and to rate the learners' satisfaction. The results of an assessment survey are shown in Fig. 7. The proposed system received higher student satisfaction scores than did the existing system with respect to both learning effectiveness and learning experience.
5
Conclusion
In this paper, we have proposed an SQL-Tutor Learning System. At the heart of the system is user-controlled customizability that allows personalized service and maximizes user satisfaction. In this system, users can place a random component of their learning environment into their learning space and have the ability to make it optionally available. Learners are enabled to change settings and reassemble components. This vulnerability of the content to user change makes possible the creation of new learning components. Flexibility is ensured by allowing advances in the learner's competence to continually be reflected in the learning environment.
References 1. Papanikolaou, K.A., Grigoriadou, M.: An Instructional Framework Supporting Personalized Learning on the Web. In: The 3rd IEEE International Conference on Advanced Learning Technologies, pp. 120–124 (2003) 2. Google, iGoogle, http://www.google.com/ig?hl=en 3. Wu, D., Im, I., Tremaine, M., Instone, K., Turoff, M.: A Framework for Classifying Personalization Scheme Used on e-Commerce Websites, System Sciences. In: Proceedings of the 36th Annual Hawaii International Conference (2003) 4. Good, N., Schafer, J.B., Konstan, J.A., Borchers, A., Sarwar, B., Herlocker, J., Riedl, J.: Combining collaborative filtering with personal agents for better recommendations. In: Proceedings of the AAAI 1999 Conference on Artificial Intelligence, Orlando, FL, pp. 439–446 (1999) 5. Thomas, J.P., Thomas, M., Ghinea, G.: Modeling of Web services flow. In: IEEE International Conference on E-Commerce (CEC 2003), pp. 391–398 (June 2003) 6. Papazoglou, M.P., Georgakopoulos, D.: Service-Oriented Computing. CACM 46(10) (2003) 7. Booth, D.: Web Services Architecture W3C Working Group Note 11 (2004), http://www.w3.org/TR/ws-arch/ 8. IMS RDCEO, IMS Reusable Definition of Competency or Educational Objective, http://www.imsglobal.org/competencies/index.html 9. Sayar, A., Pierce, M., Fox, G.: Integrating AJAX Approach into GIS Visualization Web Services. In: IEEE AICT-ICIW 2006, pp. 169–169 (2006)
Design and Implementation of Deduplication Storage Server Using Stride Scheme Ho Min Jung1, , Jin Kim1 , Jin Haeng Cho2 , and Young Woong Ko1 1
Dept. of Computer Engineering, Hallym University, Chuncheon, Korea {chorogyi,jinkim,yuko}@hallym.ac.kr 2 Boston Design Center AMD, MA, USA [email protected]
Abstract. In this paper, we propose a deduplication storage system exploiting stride scheme to minimize storage space and network bandwidth. The key idea of this paper is to provide duplication checking mechanism based on stride scheme of fixed-length block algorithm. Our approach can minimize computation time for detecting duplicated region of blocks while decreasing computation time. We made several experiments to show performance enhancement of the proposed system. In experiments, stride scheme can diminish storage space and efficiently manages duplicated data blocks. Keywords: Stride scheme, Hash, Deduplication, Storage Server.
1
Introduction
Storage administrators constantly strive to deal with increasing proliferation of data, such as, text, audio, video, image, and backup data. In the face of exponentially growing data volumes, redundant data elimination techniques have assumed critical significance in the design of modern storage systems. According to the appeal of a majority of clients who are using storage systems, traditional storage systems typically divide data objects such as files into fixed-sized blocks and then store these blocks. The key goal that due to the inability to efficiently identify those portions of the object that are actually new in the latest update, a large part of existing data must get necessarily rewritten to storage. In particular peer to peer systems, back-ups systems, ftp mirrors, and virtualization system [1] always have high ratio duplication.[2] For instance, in the Linux ftp mirror storage server, there is variety media has been format in a same file that has duplication data blocks more than 50%. In p2p system, there are large size media files with identical content where only title had been changed.
This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(2011-0002439) and by the Ministry of Education, Science Technology (MEST) and National Research Foundation of Korea(NRF) through the Human Resource Training Project for Regional Innovation.
T.-h. Kim et al. (Eds.): GDC 2011, CCIS 261, pp. 475–483, 2011. c Springer-Verlag Berlin Heidelberg 2011
476
H. Min Jung et al.
Data deduplication is a way to reduce storage space by eliminating redundant data to ensure that only a single instance of the data is stored in a storage medium. If implemented properly, data deduplication can reduce the substantial demand for storage space, thus reducing the cost of the disk. In this paper, we propose a novel deduplication scheme using stride approach. The key idea of this paper is to minimize searching time of duplicated blocks, which can diminish disk I/O and hashing time. The key idea is to adapt slide approach for fixed-length block algorithm, called DBC-SS(Duplicated Block Check with Stride Scheme). We try to improve the performance of fixed-length block algorithm while decreasing computation time. Because the running time of traditional byte shifting approach takes severely long and managing a high-capacity file is difficult in a storage system. The remainder of this paper is organized as follows: Section 2 introduces the related works for deduplication systems. Section 3 describes the design principles of deduplication storage systems and the operation of the deduplication algorithm. Section 4 gives a description of implementation and performance evaluation of the proposed system. Conclusions and future works are given in Section 5.
2
Related Work
Rsync[3] is a software application for Unix systems which synchronizes files and directories from one location to another while minimizing data transfer using delta encoding when appropriate. An important feature of Rsync not found in most similar programs/protocols is that the mirroring takes place with only one transmission in each direction. Rsync can copy or display directory contents and copy files, optionally using compression and recursion. Venti[4] is a network storage system that permanently stores data blocks, where 160-bit SHA1 hash used as the address of the data. This enforces a write-once policy since no other data block can be found with the same address. The addresses of multiple writes of the same data are identical. So duplicate data is easily identified and the data block is stored only once. Data blocks cannot be removed, making it ideal for permanent or backup storage. Venti is typically used with Fossil to provide a file system with permanent snapshots. LBFS[5], a network file system designed for low bandwidth networks. LBFS exploits similarities between files or versions of the same file to save bandwidth. It avoids sending data over the network when the same data can already be found in the servers file system or the clients cache. Using this technique, LBFS achieves up to two orders of magnitude reduction in bandwidth utilization on common workloads, compared to traditional network file systems. In IDE[6], modifying the block size of frequent anchor space is too small to resolve the growing hierarchical hash data structures, rarely matrix (sparse matrix) and designing of the block size evenly. In addition to HPs TTTD(Two Thresholds, Two Divisors) algorithm[7] designed to minimize the deviation of the block size of the anchors space and the overhead required for communication between the client and the server has confirmed the decline. DRED system[8][9] can efficiently remove the duplicate data which use delta
Deduplication Storage Server Using Stride Scheme
477
encoding technique in Web pages, and e-mail. Delta encoding is a way through sequence associated with a time of two data sets, the file name, size, and data compression.
3 3.1
Design Principle of Proposed System Overview of Deduplication Stage
Figure 1 shows the general architecture of proposed system. it is made up of the client and the deduplication file server.
Fig. 1. Overall architecture of the proposed system
In file/block hash generation module, it checks duplicated files by comparing file hash and then it process block-level deduplication work. In file deduplication stage, hash data of each file is sent to the server. In server, it checks duplicated file by comparing existing file hashes on a DBMS. The server sends hash list of duplicated files to the client. With this approach, we can prevent duplicated files are transferring to the server. In data deduplication module, block-level data deduplication is processed. The system divides data stream into blocks with chunking function. And then, we can get each data block hash using hash function such as SHA1[10], MD5[11] and SHA256. Generally, a chunking function can be dived into a fixed length chunking method and variable length chunking method. In our work, we adapted a fixed length chunking method because it is much more simple and easy to implement. The chunking size of data block is varying from 4Kbyte to several Mega Byte. In our work, we fixed 4Kbyte chunking size for increasing the performance of data deduplication. By choosing small chunking block, we can increase the possibility of finding duplicated block. Hash retrieval
478
H. Min Jung et al.
also very important because it causes frequent comparison, insert and delete operation. So, we adapted a red-black tree data structure for high performance hash operation. To process file and block deduplication, all the metadata have to be efficiently managed in a database module. The metadata includes file and block information, file and block hash data, file and block location, etc. Moreover, each file is composed of several blocks with and without duplicated blocks. To build file from blocks, we have to carefully manage each block index. 3.2
Previous Deduplication Algorithms
Deduplication schemes are divided into fixed-length blocks, variable-length blocks and delta encoding. Fixed-length blocks let files be partitioned in fixed-sized blocks, and then apply hash functions to create the hash value of the blocks. The main limitation of this approach is data shift. For example when adding a data to a file, all subsequent blocks in the file will be rewritten and are likely to be considered as different from those in the original file. Therefore, it’s difficult to find duplicated blocks in the file, which makes the compression effect less significant. In this approach, smaller blocks will get better deduplication than large one, however, it needs more computation power for deduplication. Fixedlength blocks can be implemented with two approaches. One is duplicated block check with fixed scheme (DBC-FS) and the other is duplicated block check with byte shift (DBC-BS). DBC-FS divides a file into fixed blocks and compare each block with a block in a deduplication server. In this case, if there is data shift, we cannot find duplicated block in a deduplication server. However, DBC-BS compares duplicated block by shifting one byte, therefore this approach can find all the duplicated blocks in a deduplication server. Generally, DBC-BS scheme can detect overall duplicated blocks with the overhead of computation time. In variable-length blocks, each block size is partitioned based on variable approach by anchoring based on their data patterns. This scheme can prevent the data shifting problem of the fixed-size block approach. Delta encoding stores data in the form of differences between sequential data. Lots of backup system adopts this scheme in order to give their users previous versions of the same file from previous backups. This reduces associated costs in the amount of data that has to be stored as differing versions, moreover, those costs in the uploading of each file that has been updated. DRED system use delta encoding approach to implement deduplication service. 3.3
Fixed-Length Block Algorithm Using Stride Scheme
In this paper, we propose a novel fixed-length block algorithm using stride scheme, called DBC-SS. We try to improve the performance of fixed-length block algorithm while decreasing computation time. Because the running time of DBCBS is severely long, managing a high-capacity file is difficult in a storage system. In this work, we mixed DBC-BS approach and DBC-FS. It only search fixed size of region for detecting duplicated blocks, then skip stride size of region. We assume that duplicated data has spatial locality, therefore, if we can find one
Deduplication Storage Server Using Stride Scheme
479
Fig. 2. Conceptual diagram of DBC-SS
duplicated block with DBC-BS then we can find lots of duplicated blocks around that position with DBC-FS approach. Figure 2 shows the conceptual diagram of the proposed system. In our work, we slice each file into fixed blocks and calculate Rabin[12] hash and SHA1 hash. Therefore, each data blocks have two hash data. Rabin hash data is used for fast and simple duplication check for data blocks. However, Rabin hash has high probability of hash collision, so we need another hash data (SHA1) that has no potential for hash collisions. First we check duplicated blocks with Rabin hash by shifting one byte through a chunking. If there is no duplication, DBC-SS skips fixed size of stride region and continue deduplication. Second, if there are duplicated blocks with Rabin hash, we compare the block with SHA1 hash data. If SHA1 hash data is exactly same with the block of deduplication server, then we change duplication check into DBC-FS because this region has high spatial locality. Algorithm 1 shows how DBC-SS works.
4
Performance Evaluation
For experiment, we used Linux Distribution data(CentOS 5.2 i386 package) of CD and DVD image, which contains about 7 GByte of data. We also conducted experiment with several VMware image data which contains Fedora Core 4, Fedora Core 5 Linux system. The system image size of Fedora Core 4 and Fedora Core 5 is 5,235,474,432 bytes and 8,232,566,784, respectively. In this work, we compared computation time and duplication data size for each algorithm (DBCFS, DBC-BS and DBC-SS). All block size in the experiment fixed as 8KByte and the stride size is gradually increased by 8KByte, maximum stride size limit is 56Kbyte. Figure 3 shows deduplication result for Linux distribution data. We can see that DBC-FS method has good performance in running time but shows poor
480
H. Min Jung et al.
Algorithm 1: Fixed-length block algorithm using stride scheme Input: FileStream fd, StrideSize Output: MetaStruct begin offset ← 0; checksize ← 0; fdsize ← legth(fd); while offset < fdsize do if checksize < StrideSize ) then Byte ← ReadByte(fd); checksize ← checksize + 1; if CompareRabinHash( Byte ) > 0 then StopCondition ← 0; repeat Block ← SubString(fd, offset, BlockSize); StopCondition ← CompareSha1( Block ); if StopCondition = 0 then MetaStruct ∪ MS(offset, block); end until StopCondition > 0 ; end else offset ← offset + StrideSize; checksize ← 0; end end return MetaStruct; end
Fig. 3. Performance result for Linux distribution data
Deduplication Storage Server Using Stride Scheme
481
Fig. 4. Performance result for VMware data
performance result for duplication data size. DBC-FS can find 2.6 GByte duplicated data in 2000 seconds. DBC-BS takes very long time (350000 seconds) to complete deduplication but it can find 5.1 GByte of duplicate data. The proposed scheme, DBC-SS, takes about 6000 seconds and finds 4.2 GByte duplicate data. Figure 4 shows performance result for VMware data. It also shows similar result with Linux distribution data. In this result, DBC-SS shows almost same performance on duplication data size.
Fig. 5. Varying stride size for Linux distribution data
Figure 5 shows the result of deduplication for Linux distribution data varying stride size. If we increase the stride size from 8KByte to 56KByte then
482
H. Min Jung et al.
Fig. 6. Varying stride size for VMware Data
deduplication performance is decreased slightly (50Mbyte). The computation time is very high with 8 Kbyte stride up to 2400 seconds but decreased to 1100 seconds when we increased the stride size to 56 Kbyte. In this experiment, we can conclude that broad stride size can be effective for computation time but cause slight performance degrades. Figure 6 shows performance result for VMware data. The running time in VMware data is not efficient compared with Linux distribution data.
5
Conclusion
In this paper, we introduced an algorithm and structure for a deduplication method which can be efficiently used in a storage system. The key points are to exploit stride scheme that supports DBC-FS’s fast and simple approach, and DBC-BS’s deduplication capability. We try to improve the performance of fixedlength block algorithm while decreasing computation time. Because the running time of DBC-BS is severely long, managing a high-capacity file is difficult in a storage system. In this work, we mixed DBC-BS approach and DBC-FS. It only search fixed size of region for detecting duplicated blocks, then skip stride size of region. Experiments result shows that the proposed system can minimize storage space effectively.
References 1. Robin, J.S., Irvine, C.E.: Analysis of the Intel Pentium’s ability to support a secure virtual machine monitor. In: Proceedings of the 9th Conference on USENIX Security Symposium, vol. 9, p. 10. USENIX Association (2000) 2. Cox, L.P., Murray, C.D., Noble, B.D.: Pastiche: Making backup cheap and easy. ACM SIGOPS Operating Systems Review 36, 285–298 (2002)
Deduplication Storage Server Using Stride Scheme
483
3. Tridgell, A.: Efficient algorithms for sorting and synchronization. vol. PhD. Doktorarbeit, Australian National University (1999) 4. Quinlan, S., Dorward, S.: Venti: a new approach to archival storage. In: Proceedings of the 1st USENIX Conference on File and Storage Technologies, pp. 7–7. USENIX Association, Monterey (2002) 5. Muthitacharoen, A., Chen, B., Mazieres, D.: A low-bandwidth network file system. ACM SIGOPS Operating Systems Review 35, 174–187 (2001) 6. Bobbarjung, D.R., Jagannathan, S., Dubnicki, C.: Improving duplicate elimination in storage systems. ACM Transactions on Storage (TOS) 2, 424–448 (2006) 7. Eshghi, K., Tang, H.K.: A framework for analyzing and improving content-based chunking algorithms. Hewlett-Packard Labs Technical Report TR, vol. 30 (2005) 8. Douglis, F., Iyengar, A.: Application-specific delta-encoding via resemblance detection. In: Proceedings of the USENIX Annual Technical Conference, pp. 1–23 (2003) 9. Kulkarni, P., Douglis, F., LaVoie, J., Tracey, J.M.: Redundancy elimination within large collections of files. In: Proceedings of the annual conference on USENIX Annual Technical Conference, p. 5. USENIX Association (2004) 10. Eastlake, D., Jones, P.: US secure hash algorithm 1 (SHA1). RFC 3174 (September 2001) 11. Rivest, R.: RFC 1321: The MD5 message-digest algorithm. Internet activities board 143 (1992) 12. Rabin, M.O.: Fingerprinting by random polynomials. Center for Research in Computing Techn., Aiken Computation Laboratory, Univ. (1981)
Robust Gait Recognition Using Planar Homography and Shape Sequence Descriptor Seungdo Jeong1 , Youngseok Lee1 , Keun-Wang Lee2 , and Jungwon Cho3, 1
Research Institute of Electrical and Computer Engineering, Hanyang University, 17 Haengdang-dong, Sungdong-gu, Seoul 133-791 S. Korea {sdjeong,yslee38}@hanyang.ac.kr 2 Department of Multimedia, Chungwoon University, San 29 Namjang-ri, Hongseong-eup, Hongseong-gun, Chungnam 350-701 S. Korea [email protected] 3 Department of Computer Education, Jeju National University, 102 Jejudaehakno, Jeju-si, Jeju-do 690-756 S. Korea [email protected]
Abstract. In the gait recognition, the dependency to the direction of walking is very serious problem. To reduce this dependency, we propose a view synthesis method based on the planar homography. However, even though the synthesized gait is used, shape information is not enough to recognize individual. Thus, in this paper, we use shape sequence descriptor for recognition, which describes shape information and variation according to motion at a same time. Our experiments show that the proposed method efficiently reduces the dependency to directional variations of gait. Keywords: Gait Recognition, Biometrics, Canonical View, Shape Sequence, View Synthesis.
1
Motivation
In the shape-based gait recognition, the direction of walking influences shape features extracted from gait sequence [1]. That is, difference exists between features obtained from gait which is perpendicular to camera optical axis and ones from oblique-directional gait. To resolve this problem, reconstruction method which synthesizes gaits to the canonical-viewed ones is required [2]. The perfect image synthesis needs exact camera parameters; furthermore, those should be fixed in any environment [3]. However it is great restriction to apply the method to general-purpose recognition system. Therefore, in this paper, we propose view synthesis method through simple operation to minimize shape distortion of gait silhouette caused by the direction of walking. It is impossible to restore shape information without accurate threedimensional reconstruction. Thus, the recognition method using only distorted
Corresponding author.
T.-h. Kim et al. (Eds.): GDC 2011, CCIS 261, pp. 484–488, 2011. c Springer-Verlag Berlin Heidelberg 2011
Robust Gait Recognition Using Planar Homography and SSD
(a) perpendicular-directional sequence
485
(b) oblique-directional sequence
Fig. 1. Cumulated gait silhouettes according to the direction
shape information degrades the performance seriously. To compensate the degradation, we adopt the feature description algorithm in the gait recognition to consider shape information and variation due to the motion, simultaneously.
2 2.1
Proposed Method View Synthesis Using Planar Homography
Gait sequence has a cue for the direction of walking. Fig. 1 shows the direction that is perpendicular to the camera optical axis and oblique direction, respectively. The cumulated silhouette comprises one plane. From this observation, we adopt the planar homography H which maps specific point Px to the infinite P∞ , where specific point Px is the cross of two lines composed of respective head points and foot points. The planar homography H mapping a point Px to the infinite is represented as equation (2) [4]. P∞ = HPx .
(1)
H = GRT.
(2)
In equation (2), T is the translation matrix which translates center of image x0 to the origin. R is the rotation matrix which rotates px to the point (f, 0, 1)T lying on the x-axis. G maps (f, 0, 1)T onto the infinite point (f, 0, 0)T . However, if H is used to synthesize all of silhouettes without any modification, distortion of body in synthesized silhouette increases as gap between silhouette and the center of image increases. To compensate the distortion, the center of transformation should vary according to the position of each silhouette.
Fig. 2. The center of transformation
Fig. 2 shows the proposed center of transformation, where V L denotes vanishing line which is parallel to the x-axis and goes through the cross point px .
486
S. Jeong et al.
As shown in the figure, the center of transformation of each silhouette is the cross point between V L and center line of silhouette. As a result, R becomes identity matrix I because y coordinate value of px is identical to the one of the center of transformation. To be coincided scale of all silhouettes, we insert scale factor s = hn /h to G as equation (3), where h and hn are real height and normalized height of silhouette, respectively. Therefore, the proposed planar homography Hp is represented as equation (4), where Tp translate each center of transform to the origin. ⎡ ⎤ s 00 Gp = ⎣ 0 s 0 ⎦ . (3) − f1 0 1 Hp = Gp ITp . 2.2
(4)
Shape Sequence Descriptor
Even though oblique-directional gait is synthesized to canonical-viewed one, there are some distortions in shape information, thus, it causes the degradation in recognition performance. To improve the performance, we describe gait feature using the Shape Sequence Descriptor(SSD) which represents variation caused by the motion as well as shape itself [5]. The point of the SSD is the Angulr Radial Transform(ART) [6]. nm-th coefficients of the ART Fnm is computed as equation (5). Vnm denotes basis function computed as equation (6), where Am is angular component and Rn is radial component. ART coefficient is complex because angular component is complex. Degrees of the ART coefficients n and m which are recommended in the MPEG7 standard are 3 and 12, respectively, thus, dimension of ART is 36 [6]. However, in this paper, we use n = 4 and m = 9 because radial component is more dominant to discriminate gait silhouette. 2π 1 ∗ Fnm =< Vnm (ρ, θ), f (ρ, θ) >= Vnm (ρ, θ)f (ρ, θ)ρdρdθ. (5) 0
The SSD is created from the matrix which is a set of the ART coefficients arranged with time axis. Thus, starting points of silhouette sequences of respective individual should be coincided with themselves to preserve consistent pattern. For this, we set the starting point at frame of silhouette with the maximum stride length, and the following 30 frames comprise the shape sequence for respective description of each gait. The shape sequence formed with 30 frames is not same comparing to the one of other people due to the different speed of walking. However, it could be also respective characteristic of individual, thus, is reflected to
Robust Gait Recognition Using Planar Homography and SSD
487
Fig. 3. Matrix of the ART coefficients for shape sequence
the description automatically. Fig. 3 shows 3-tuple of 36 × 30 matrix composed of the ART coefficients for shape sequence of single gait. The final step in the SSD is selection of the spatial frequency contents about the ART coefficient matrix in the frequency domain. For this, this paper uses the 2D FFT algorithm. The low frequency components represent the outline of content while the high frequency components represent the detail of content. In the case of this paper, the DC component is useless to separate individual because it denotes average silhouette which might be almost same regardless each person. The high frequency components have high possibility to be influenced by noises due to imperfect segmentation for the silhouette, light condition and so on. Thus, we use only top 5 lines in frequency component except the DC and L1 distance is used the similarity measure.
3
Experimental Results
To evaluate the performance of the proposed method, we perform experiment with gait sequences of 16 people. Gait sequences for 16 people are obtained through three times at different days with different view point. Gait sequences for database are with direction almost perpendicular to the camera optical axis. One set for the test is almost same direction comparing to one for database, the other set is oblique-directional gait sequences. We compare experimental results using the proposed method and using simple normalization method. Table 1 presents the recognition rate for all experiments. Table 1. Experimental results in the recognition rate Direction Perpendicular Oblique
The proposed method 93.75% 68.75%
Normalization 87.5% 43.75%
As shown in the table using the perpendicular-directional gait sequence, the proposed method failed to recognize for only one person. Even though, the method using simple normalization missed only two people. It denotes that the
488
S. Jeong et al.
shape sequence descriptor is useful to apply the shape-based gait recognition. The reason about a little improvement is because the direction of gait is not exact perpendicular but almost perpendicular to the camera optical axis, thus, the proposed method compensates such a little variation. In the case of experiments using the oblique-directional gait, recognition rate based on the normalization method dropped seriously down with over 40%. It denotes that the direction of walking influences bed effect to the recognition, seriously. However, the proposed method decreases such dependency, thus, outperforms the one using normalization by 25%.
4
Conclusion
This paper has proposed the gait recognition method based on planar homography and shape sequence descriptor. The dependency to the direction of walking is very serious problem in the shape-based gait recognition. To reduce the dependency, this paper have proposed view synthesis method based on the planar homography which not only transform the oblique-directional gait to canonical-viewed one but also normalize the size of silhouette with simple operation regardless the position. Even though the synthesized silhouette is used, distortion related to shape information for the gait might give the bed effect in recognition. To compensate that, this paper has also proposed the recognition method based on shape sequence descriptor which is able to describe shape information and variation caused by the motion, simultaneously. From experiment with 16 people, we have proof experimentally that the SSD is effectively used for the shape-based gait recognition. Furthermore, it has noticed that the proposed method has efficiently reduced the dependency to directional variation of gait.
References 1. Lee, L., Grimson, W.E.L.: Gait Appearance for Recognition. In: Proc. ECCV Workshop, Biometric Authentication, pp. 143–154 (2002) 2. BenAbdelkader, C., Cutler, R.: View-invariant Estimation of Height and Stride for Gait Recognition. In: Proc. ECCV Workshop on Biometric Authentication, pp. 155– 167 (2002) 3. RoyChowdhury, A.K., Kale, A., Chellappa, R.: Video Synthesis of Arbitrary Views for Approximately Planar Scenes. In: Proc. IEEE Conf. on Acoustics, Speech and Signal Processing, vol. 3, pp. 497–500 (2003) 4. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press (2000) 5. Choi, M.S.: An Efficient Representation and Matching of a Moving Object in Silhouette, Ph.D. Thesis, Hanyang University (2004) 6. Bober, M.: Mpeg-7 Visual Shape Descriptors. IEEE Trans. Circuits System for Video Technology 11, 716–719 (2001)
A Real-Time Distributed Architecture for RFID Push Service in Large-Scale EPCglobal Networks Min Sung Kang1 and Do-Hyeun Kim2,* 1
Dept. of Research, Gaunuri Co., Ltd., Jeju, Republic of Korea [email protected] 2 Dept. of Computer Engineering, Jeju National University, Jeju, Republic of Korea [email protected]
Abstract. Recently, a large-scale EPCglobal network rapidly increasing, the vision of a large-scale EPCglobal network is close to becoming a reality. However, there is neither a research for efficient accessing EPCIS(EPC Information Service) repository and distributing to the business application. In this paper, we propose a real-time architecture of effective accessing EPCIS repository based on RFID push services in distributed large-scale EPCglobal network. Proposed architecture provides the real-time monitoring and transferring EPC data to the business application when happen the event. Keywords: EPCglobal, Distributed architecture, RFID.
1
Introduction
Radio Frequency Identification (RFID) is a technology that can be used for the efficient tracking of materials as they move though the supply chain[1]. RFID is expected to significantly increase the amount of data available to decision makers who are in charge of steering the supply chain. Current research on RFID among other things focuses on efficient distributed system architectures for making contextualized RFID data usable for decision support software. EPCglobal, an international industry consortium, has provided a specification for the Electronic Product Code (EPC) which can be used to uniquely identify single products. The consortium has also initiated the standardization of infrastructure components that are supposed to work together in order to collect, filter, store, and retrieve EPC related data in a supply chain context[2]. Traditional databases support one-time queries over stored data, and stream processing engines focus on processing push-based data continuously. However, there is not a research related efficient accessing EPCIS repository and distributing to the business application. This paper proposes a real-time architecture of effective accessing EPCIS repository and distributing to the business application for business application services in large-scale EPCglobal network [3].
The rest of this paper is structured as follows. In Section 2, we detail the related work of the EPCglobal network. In Section 3, we describe our proposed architecture and show how our design addresses. Finally we conclude in Section 4.
2
Related Works
The EPCglobal Network is a network providing a shared view of the disposition of EPC-bearing objects between EPCglobal subscribers, within a relevant business context. In the EPCglobal Network, product data is distributed to several EPCISes (EPC Information Services) via movement of the product. The ONS (Object Naming Service) and the EPCIS DS (EPCIS Discovery Service) are used to identify the distributed data for tracing the product [3].
Fig. 1. EPCglobal architecture framework
A Real-Time Distributed Architecture for RFID Push Service
491
The EPCglobal architecture framework can be seen in figure 1. The standards can be divided into three parts; identity, capture and exchange. The identity portion contains standards for the identification of tags and the translation of tag data. The capture portion contains the standards for filtering and collection of the tag data. The exchange portion contains standards for storing and sharing collected and filtered EPC product data, and also has the main role for the tracing of products. We will therefore focus on EPCIS, EPCISDS and ONS in this paper. The EPCglobal architecture framework is a collection of interrelated standards for hardware, software and data interfaces, together with core services that are operated by EPCglobal and its delegates, all in service for a common goal of enhancing the supply chain through the use of EPCs. Figure 1 shows the role and the interface of each part of the EPCglobal standards.
Fig. 2. Use of the EPC in EPCglobal architecture framework
The role of EPCIS in the EPCglobal Network is to provide a repository for EPC event and master data. Event data refers to data arising when business processes are carried out and master data is additional data providing the necessary context for interpretation of the event data [3]. Due to the role of EPCIS in the EPCglobal Network, querying to EPCISes is essential for obtaining EPC data. EPCIS Standard has been defined by EPCglobal. This contains EPCIS interface specification, event data schemes and master data. However, it lacks specifications for the master data attributes and the master data management interface [4-5].
492
M. Sung Kang and D.-H. Kim
The EPCglobal Architecture Framework includes software standards at various levels of abstraction, from low-level interfaces to RFID reader devices all the way up to the business application level. The different forms of the EPC specified in the EPC Tag Data Standard are intended for use at different levels within the EPCglobal architecture framework. The primary representation of an Electronic Product Code is as an Internet Uniform Resource Identifier (URI) called the Pure Identity EPC URI. The EPC memory bank of a Gen 2 RFID Tag contains the EPC plus additional “control information” that is used to guide the process of data capture from RFID tags. The EPC memory bank of a Gen 2 RFID Tag actually contains a compressed encoding of the EPC and additional “control information” in a compact binary form. Use of the EPC in EPCglobal Architecture Framework can be seen in figure 2[2].
3
A Proposed Real-Time Distribution Architecture for RFID Push Services
RFID applications are numerous and far reaching. The most interesting and widely used applications include those for supply chain management, security, and the tracking of important objects and personnel. In supply chain management, RFID tags are used to track products throughout the supply chain—from supplier delivery, to warehouse stock and point of sale. New applications target tracking from checkout through customer billing. A central database records product movement, which manufacturers or retailers can later query for location, delivery confirmation, or theft prevention. Security and personal identification applications are a major and broad application of RFID. A common use of RFID is in identification cards to control
Business
Business
Business
Application 1
Application 2
Application N
Business Application Interface
Real-time EPC Data Distributor
EPCIS Query Interface
EPCIS
EPCIS
EPCIS
Monitoring 1
Monitoring 2
Monitoring N
EPCIS 1
EPCIS 2
EPCIS 2
Fig. 3. Proposed real-time distribution architecture for RFID push services
A Real-Time Distributed Architecture for RFID Push Service
493
building access. Many organizations use RFID tags embedded in ID cards, which are readable at a building entrance. Some schools are requiring children to wear tagembedded bracelets or wrist bands while on school grounds, to monitor attendance and to locate lost children. We propose real-time distribution architecture for supporting push services of RFID applications in figure 3. The architecture consist of the business application interface, the real-time EPC data distributor, EPCIS query interface, and EPCIS monitoring part. The role of EPCIS monitoring part is checking using trigger method and forwarding by EPCIS query interface when we add and modify and delete EPC data in EPC repository. EPCIS query interface supports to deliver EPCIS events. The role of the real-time EPC data distributor receives and handles the EPC event data, and transmits these to necessary business applications.
Real-time EPC Data
Graphic User Interface Repository Monitoring Module
Business Application Interface Business Application
Business Application Interface Business Application
Business Application
Fig. 4. Proposed real-time EPC data distributor architecture
The real-time EPC data distributor setups to control for connection of database. And this gets the table information back of database to support user interface. This distributor is consisted of the user interface part, distributed data process part; EPCIS query interface, and business application interface. Also, the distributed data process has the data communication module, XML parser module, data queue management module, control module, data filter module, repository management module, thread pool, configuration management module. XML parser module is analyzed the received XML data and filtering. And there is Database administration module that
494
M. Sung Kang and D.-H. Kim
connects to database and sets necessary table information and Forwarding module. The control module supports the distributed processing and buffeting with thread pool. The forwarding service setup module supports to make lists of business applications the defined IP address, port number, and transmission interface. The role of data filter is to select the transferred data for each business application. The proposed real-time EPC data distributor architecture can be shown in figure 4. [Configuration Part]
[Control Part]
Configuration Management Module
Process Controller
Thread Controller
[Service Process Part] EPC Data
XML Serialization / Deserialization
[Data Process Part]
Data Queue Management Module
Data Communication Module
EPC Data
XML Parser Module
····
···· Data Filter Module
EPC Data Data Communication Module
XML Dom Tree
Thread Pool
Fig. 5. Configuration of real-time EPC data distributor
The figure 5 show a data flow in the real-time EPC data distributor for push services. Then data communication module receives from EPCIS repository using Web services and stores the message queue for buffering. XML parser disassembles XML EPC data received from EPCIS. And data filter classifies EPC data using configuration setup information. Also data queue management module supports temporary to storage for transmitting to each business application using many message queues.
A Real-Time Distributed Architecture for RFID Push Service
4
495
Conclusions
In this paper, we propose a real-time distributed architecture for RFID push services in large-scale EPCglobal network. This architecture supports to distribute more quick, saleable, and effective. Our approach has two benefits for existing business application based on RFID: (a) it is possible a real-time push services based on RFID due to the centralized distributor, (b) it reduce the processing and storage overhead due to distributed processing and storage architecture. It is a main research area of distribution of EPC data in business application based on RFID. Yet, many RFID issues are still unsolved and require more efforts from industrial practitioners. Acknowledgments. This work was supported by the Industrial Strategic Technology Development Program funded by the Ministry of Knowledge Economy (MKE, Korea). [10038653, Development of Semantic based Open USN Service Platform]. This research was supported by the project of regional technology innovation of the Ministry of Knowledge Economy (MKE, Korea).
References 1. Niederman, F., Mathieu, R.G., Morley, R., Kwon, I.-W.: Examining RFID Applications in Supply Chain Management. Journal of Communications of the ACM 50(7) (2007) 2. EPCglobal.: EPCglobal Architectural Framework v. 1.4, http://www.epcglobalinc.org/standards/architecture/ architectarc_1_4-framework-20101215.pdf 3. EPCglobal: EPCIS Standard v. 1.0.1, http://www.epcglobalinc.org/standards/epcis/ epcis_1_0_1-standard-20070921.pdf 4. EPCglobal.: EPC Information Services (EPCIS), Standard v. 1.0.1., http://www.epcglobalinc.org/standards/epcis/ epcis_1_0_1-standard-20070921.pdf 5. Weinstein, R.: RFID: a technical overview and its application to the enterprise. IT Professional 7(3) (2005)
Heuristic Scheduling Strategies for Linear-Dependent and Independent Jobs on Heterogeneous Grids Min-Yi Tsai, Ping-Fang Chiang, Yen-Jan Chang, and Wei-Jen Wang* Department of Computer Science and Information Engineering, National Central University, Taoyuan 320, Taiwan [email protected]
Abstract. Grid computing came into being an active research area because of the advances in wide-area network technologies and the low cost of computing resources. One motivation of grid computing is to aggregate the power of distributed resources and integrate the resources into a unified platform. To minimize the total completion time of the submitted computing jobs to a grid platform, people employ various scheduling algorithms to dispatch the jobs to the resources. However, it has been proved that the optimal scheduling algorithm is NP-hard. Therefore, many people turn to use heuristic approaches for grid scheduling. In this paper, we introduce ten common scheduling heuristics to schedule a combination of job-chains (linear-dependent jobs) and independent jobs on a heterogeneous environment. We implemented these methods on a grid simulator to evaluate their performance under different circumstances. The results of scheduling job-chains and independent jobs on a heterogeneous environment are quite different from previous studies, and we provide our explanations for the differences. We also propose a hybrid method based on our observation, and the simulation results show that it has good performance in terns of makespan. Keywords: Job scheduling, Grid computing, Heuristics, Job dependency.
1
Introduction
With the development of science and technology, the computing power of a CPU has become more powerful; moreover the number of cores in the CPU keeps increasing in the past decade as well. Grid computing technology [1] can be used to integrate these distributed resources into a more powerful, unified system. By coordinating the scheduler, a grid system automatically and efficiently finds appropriate computing nodes. Then, it can assign and transfer jobs that are waiting for execution to available nodes. In a typical grid system, a scheduler may schedule dependent jobs and independent jobs. A set of independent job can be executed without any constraints, while a series of dependent jobs must follow some order to complete those jobs.
Heuristic Scheduling Strategies for Linear-Dependent and Independent Jobs
497
According to the dependency relationship, dependent jobs can form job-chains (linear-dependent jobs) or DAG (directed-acyclic-graphical). According to the above considerations, a great deal of algorithms and mechanisms for job scheduling had been proposed in the past [2-6]. Job scheduling on parallel and distributed computing systems is a hard problem. Under multi-processor systems, a typical job scheduling problem is proved to be NP-complete [7]. As a result, many heuristics has been proposed in the past decades to improve efficiency of the job schedule problems. These heuristics may focus on different considerations of situations and constraints. For example, some heuristics, such as the longest-job-first heuristic, have good completion time (makespan) while scheduling job-chains [4] on a heterogeneous environment. Sahu et al. [6] found out that the min-max method has the best makespan while scheduling independent jobs on a homogeneous environment. To minimize the makespan of the system, this paper proposes a hybrid scheduling method that achieves good makespan under most circumstances. The proposed scheduling method combines the max-min method and the longest-job-first method to schedule a combination of job-chains and independent jobs. This Hybrid method can reduce the constraints that single heuristic may face under some specific situations. Therefore, the proposed method can adapt to heterogeneous environments while scheduling a set of jobs. We have implemented several heuristic scheduling methods on grid simulator, namely GridSim [8], conducted experiments to evaluate the performance of several commonly-used heuristics, and compared them with the proposed hybrid method using a set of job-chains and independent jobs on a heterogeneous environment. The rest of the paper is organized as follows. Section 2 describes the background of this study and some related work. Section 3 shows the simulation results of ten heuristic scheduling methods on a heterogeneous environment to schedule job-chains and independent jobs. Section 4 describes the proposed hybrid method and Section 5 provides the experimental results. Finally, Section 6 gives concluding remarks.
2
Background
This section will briefly introduce the concept of the objective of grid scheduling, and the related work of this research. 2.1
Scheduling Objectives
The essence of grid scheduling can be formulated into an optimization problem involving several objective functions, such as makespan (completion time), flowtime, resource utilization, matching proximity, and algorithm computation time [6]. Among those objectives, makespan is the most popular objectives to optimize, and this paper focuses on the problem of makespan optimization on a heterogeneous environment. Makespan represents the time from the first executing job to the last finished job. Let Cj represent the makespan of 1 ≤ j ≤ n. Then, the makespan optimization problem is to minimize the following equation:
498
M.-Y. Tsai et al.
Cmax = MAX {Cj | 1 ≤ j ≤ n} 2.2
(1)
Heuristic Methods for Grid Scheduling
Many heuristic scheduling algorithms [4-6] have been employed to improve the performance of grid scheduling. The subsection will describe some common heuristic algorithms. Note that most heuristic methods are originally proposed for independentjob-scheduling. Those heuristic strategies must be modified to support job-chain scheduling. • Minimum Completion Time (MCT): The MCT algorithm randomly chooses a job that is ready for execution, and computes the finish time of the job on each machine. The job is then assigned to the machine with the smallest finish time. • Min-Min: The min-min algorithm is based on the MCT algorithm. The MCT algorithm considers one job at a time while min-min considers all unscheduled jobs. The purpose of the min-min algorithm is to assign as many jobs as possible to the fastest machine. Thus, it can minimize the overall makespan. • Max-Min: The max-min algorithm is a variation of min-min. The min-min algorithm selects the job with the minimum completion time. However, max-min chooses the job with the largest minimum-completion-time. The goal of max-min algorithm is to reduce the cost of executing the job with long minimum completion time. • Min-Max: The min-max heuristic calculates the minimum completion time and the minimum execution time (MET) for every unassigned job, and then selects the job with the largest ratio of MCT to MET. • Longest-Job-First (LJF): The goal of the longest-job-first method is to reduce the total makespan by preventing the job with longest execution time from being the late-time bottleneck. To achieve this goal, the longest-job-first method schedules the longest job to the earliest machine that can process the job. If many machines are available at the same time, the method always picks the fastest machine. • Shortest-Job-First (SJF): The shortest-job-first method schedules the shortest job to the earliest available machine that can process the job. If many machines are available at the same time, the method always picks the fastest machine. • Opportunistic-Load-Balancing (OLB): The method chooses the earliest idle machine, and randomly picks an unassigned job for execution. • Suffrage: Suffrage for a job is the difference between second-best minimum completion time and the best minimum completion time. The job with the maximum suffrage value is assigned to the machine that gives the best minimum completion time. • Work Queue (WQ): The method randomly selects a job and assigns it to the machine with the minimum workload. • Average-Execution-Time (AET): The method is dedicated for job-chains. It assigns an entire job-chain to a machine for execution. It calculates the averagejob-size of each job-chain, and then uses LJF to schedule the job-chain that has the largest average-job-size.
Heuristic Scheduling Strategies for Linear-Dependent and Independent Jobs
2.3
499
Related Work on Analysis of Heuristic Scheduling Methods
Many researchers have been conducted many experiments and simulations to evaluate the performance of scheduling heuristics. Braun et al. [5] evaluated eleven heuristic methods for independent-job-scheduling on a heterogeneous environment. They found that the genetic-based algorithm, min-min. and duplex (a hybrid of min-min and min-max) are among the best scheduling strategies. Lin et al. [4] pointed out that, to schedule job-chains (linear-dependent jobs) on a heterogeneous environment, longest-job-first is the best strategies among average-execution-time, and shortestjob-first. Sahu et al. [6] evaluated twelve heuristic methods for independent-job scheduling on a homogeneous environment. They indicated that, in the category of makespan comparison, min-min, min-max, suffrage, and RC (Relative Cost that considers both the load balancing of machines and the execution time of jobs) are among the best. The results also showed that min-max is the best among the min-min, min-max, suffrage, and RC.
3
Preliminary Analysis for Ten Scheduling Heuristics
In this section, we will show the simulation results for the ten scheduling heuristics described in Subsection 2.2. We implemented these scheduling heuristics on a grid simulator, namely GridSim [8, 11] to evaluate their performance. We devised several cases for scheduling job-chains and independent jobs on a heterogeneous environment, simulated those cases, and compared the results of using different scheduling strategies in those cases. 3.1
Supporting Scheduling for Job-Chains
Nine of the ten scheduling heuristics described in Subsection 2.2 are not able to schedule job-chains directly, and only the AET strategy can schedule both job-chains and independent jobs. To support the ability to schedule job-chains in the remaining nine scheduling heuristics, we enforce dependency checking in each of them. The concept of the procedure is shown as follows: 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
let k be a small constant value for each unscheduled job J that is ready to be executed calculate the cost of scheduling J to a machine M using the heuristic strategy find the best pair of J and M using the heuristic strategy if J has a preceding job I then if I.finish_time > J.start_time then J.start_time = I.finish_time + k set J to be executed at J.start_time in M update the earliest available time of M goto Line 2 until all jobs are scheduled
500
M.-Y. Tsai et al. Table 1. Simulation parameters
Parameter Total amount of jobs Job size Total amount of nodes Computing power of a node Times of simulation
3.2
Value 512 Random( 1~999999 ) million instructions 10 Random (1000, 2000, …, 10000) MIPS 100
Preliminary Simulation Results on Heterogeneous Environments
We have designed three simulation cases for scheduling independent jobs and jobchains in a heterogeneous environment. The goal is to find the best scheduling strategy among the ten scheduling heuristics under different circumstances. The parameters of our simulations are shown in Table 1. The total amount of jobs is fixed at 512, and each job has a random size. The job size represents the number of instructions to be executed, and it ranges from 1 to 999999 million instructions. The jobs are scheduled on a heterogeneous environment consisting of 16 nodes. Each computing node has random processing power. The computing power of a grid node is set to 1000n MPIS, where n ranges from 1 to 10. To obtain the result of a particular simulation case, we simulate it a hundred times and take the average of the simulated makespan values as the result. The first simulation case (Case 1) is to schedule 512 independent jobs on a cluster consisting of heterogeneous computing nodes. Figure 1 shows the makespan values of different heuristic strategies. The simulation result indicates that LJF and AET are the best for scheduling independent jobs on a heterogeneous environment; min-min, SJF, and WQ are the relatively bad scheduling strategies. Note that AET has the same behavior as LJF while scheduling independent jobs. The max-min strategy also works well in this simulation case since its behavior is similar to the LJF strategy. The result shows that the longest-job-first strategy, used by LJF, AET, and maxmin, works well for scheduling independent jobs on a heterogeneous environment. On the contrary, min-min and SJF schedule the shortest job to the earliest available machine. Thus, they may increase the chance of running bigger jobs on slower machines, which results in a larger makespan. The WQ always allocates jobs to the machine with the minimum workload and doesn’t consider the completion time for each job on every machine, So, a slow machine has a higher chance to execute a large job and greatly increases the makespan. The second simulation case (Case 2) is to schedule 64 job-chains, each of which consists of eight jobs, and the jobs are executed on a heterogeneous environment. According to Figure 2, when jobs are linear-dependent, LJF no longer performs well; the max-min strategy and the suffrage strategy become the best among the ten heuristics. Note that max-min does not always schedule the longest job first. It prefers the job that has the largest minimal-completion-time on the heterogeneous environment. The suffrage strategy has similar behavior to the max-min strategy since it utilizes the information of the largest minimal-completion-time of a job. Therefore,
Heuristic Scheduling g Strategies for Linear-Dependent and Independent Jobs
501
its performance is relatively good in Case 2. The min-max strategy and the min-m min strategy do not work welll in this case. The result of Case 2 indicates that, whhile scheduling job-chains in thee heterogeneous environment, a scheduling method shoould first consider the job with th he largest minimal-completion-time. The third simulation case (Case 3) is to schedule a set of job-chains onn a heterogeneous environmentt. Each job-chain consists of a normally distributed randdom number of jobs, given the average to be 3 and the standard deviation to be 1. T The simulation result is shown n in Figure 3. The simulation result of Case 3 show ws a similar result to Case 2. Th hat is, max-min and suffrage are the better choices, whhile min-min and min-max are the t worse choices.
Fig. 1. Case 1: th he simulation result of scheduling independent jobs
Fig. 2. Case 2: the result of o scheduling job-chains, each of which consists of eight jobs
502
M.-Y. Tsai et al.
Fig. 3. Case 3: the result off scheduling job-chains, each of which consists of a norm mally distributed random number of jobs j
The results of Cases 1-3 3 show that a heterogeneous environment favors the L LJFbased strategies, especially the max-min strategy. This finding is consistent with Lin et al.’s study [4]. On the contrary, c a homogeneous environment favors the SJF-baased strategies, especially the min-min m strategy [5]. Our simulation results, which are not provided the paper, also co onfirm the results by Sahu et al [6]. However, the studyy by Braun et al. [5] is quite diffferent from our observation since they found that min-m min is a better solution. Our ex xplanation is that, the simulation models are different and thus the simulations producced different results. The simulation model in Braun et aal.’s study assumes task hetero ogeneity, making the execution time of a particular job almost unrelated to the com mputing power of a machine. For example, the executtion time on Machine 1 for Job b A and Job B can be 25137.5 and 399562.1 respectively, while the execution time on n Machine 2 can be 30802.6 and 58987.9 respectively. T This assumption favors SJF-baased strategies since longer minimum completion tiime usually implies wasting of time. t
4
The Proposed Hybrid H Method
Base on the simulation resu ults described in Section 3, we can conclude that LJF is the best while scheduling indeependent jobs on a heterogeneous environment, and m maxmin is the best while sched duling job-chains. Thus, we propose a new hybrid methhod using these two heuristic methods. m The strategy is simple. First, the proposed methhod examines whether a job is linear-dependent or not. Second, this hybrid method uuses LJF to schedule independent jobs and the uses max-min to schedule job-chains. T The algorithm can be expressed as follows: 1. 2. 3. 4. 5.
let k be a small constaant value for each unscheduled d independent job J that is ready to be executed use LJF to find thee best pair of J and M set J to be executed in nM update the earliest avaailable time of M
Heuristic Scheduling g Strategies for Linear-Dependent and Independent Jobs
6. 7. 8. 9. 10. 11. 12. 13. 14. 15.
5
503
goto line 2 until all in ndependent jobs are scheduled for each unscheduled d dependent job J that is ready to be executed calculate the cost of o scheduling J to a machine M using max-min find the best pair of o J and M if J has a precedin ng job I then if I.finish_time > J.start_time then J.start_time = I.finish_time + k set J to be executed at J.start_time in M update the earliest avaailable time of M goto line 7 until all jo obs are scheduled
Preliminary Exp perimental Results
We used the same configurration described in subsection 3.2 for simulations. We hhave designed several schedulin ng scenarios to evaluate the performance of the propoosed hybrid method.
Fig. 4. Case 4: the simulation result of normal distributed job chain length with hybrid heuriistic
Fig. 5. Case 5: the t simulation result of scheduling 20% job-chains
504
M.-Y. Tsai et al.
Fig. 6. Case 6: the t simulation result of scheduling 50% job-chains
Fig. 7. Case 7: the t simulation result of scheduling 80% job-chains
First, we use the settin ng of Case 3 (each job-chain has a normally distribuuted length) to compare the pro oposed hybrid method with other heuristics. The propoosed hybrid method has the best makespan, as shown in Figure 4. It can save up to 2.5% % of the makespan compared with w the second best, suffrage. We also designed thhree different simulation scenarrios, Cases 5~7 to compare the proposed hybrid methhod with other heuristic methods. Case 5 is to schedule 20% of job-chains and 80% % of independent jobs, Case 6 is to schedule 50% of job-chains and 50% of independdent jobs, and Case 7 is to sched dule 80% of job-chains and 20% of independent jobs. E Each job-chain consists of five jobs in Cases 5~7. The simulation results are shownn in Figures 5~7 respectively. The T results indicated that the proposed hybrid method is a relatively better strategy. However, H the gain of using the proposed hybrid methodd is limited.
6
Conclusions
This paper used GridSim to t simulate and to demonstrate the results of ten comm mon heuristics for scheduling lin near-dependent jobs (job-chains) and independent jobs oon a
Heuristic Scheduling Strategies for Linear-Dependent and Independent Jobs
505
heterogeneous environment. From the simulation results, we figured out that LJF is the fastest method for scheduling independent jobs, and max-min is the best method for scheduling job-chains. The observation is different from previous studies that are conducted on homogeneous environments [6] or heterogeneous environments [5] to schedule independent jobs. We provided our explanations for the difference. We also proposed a hybrid heuristic method using LJF and max-min. The simulation results showed that the hybrid method demonstrates good performance values under most circumstances. When a job-chain has a random number of jobs or has a normally distributed number of jobs, the proposed hybrid method can still save up to 3% of makespan compared to the other ten scheduling method. Acknowledgments. This work was partially supported by the Nation Science Council of Republic of China under Grant No. 100-2218-E-008-005-. We are grateful to the National Center for High-performance Computing, Taiwan, for computer time and facilities.
References 1. Foster, I., Kesselman, C.: The Grid 2: Blueprint for a New Computing Infrastructure. Morgan Kaufmann Publishers Inc., San Francisco (2003) 2. Yu, J., Buyya, R.: A Taxonomy of Workflow Management Systems for GridComputing. Journal of Grid Computing 3(3), 171–200 (2005) 3. Du, J., Leung, J.Y.-T., Young, G.H.: Scheduling chain-structured tasks to minimize makespan and mean flow time. Information and Computation 92, 219–236 (1991) 4. Lin, P.-Y., Liu, P.: Job Scheduling Techniques for Distributed Systems with Temporal Constraints. In: Bellavista, P., Chang, R.-S., Chao, H.-C., Lin, S.-F., Sloot, P.M.A. (eds.) GPC 2010. LNCS, vol. 6104, pp. 280–289. Springer, Heidelberg (2010) 5. Braun, T.D., Siegel, H.J., Beck, N., Bölöni, L., Maheswaran, M., Reuther, A.I., Robertson, J.P., Theys, M.D., Hensgen, B.Y.D., Freund, R.F.: A comparison of eleven static heuristics for mapping a class of independent tasks onto heterogeneous distributed computing systems. J. Parallel Distrib.Comput. 61(6), 810–837 (2001) 6. Sahu, R., Chaturvedi, A.: Many-Objective Comparison of Twelve Grid Scheduling Heuristics. International Journal of Computer Applications 13(6), 9–17 (2011) 7. Ullman, J.D.: NP-complete scheduling problems. Journal of Computer and System Sciences 10, 384–393 (1975) 8. GridSim, http://www.cloudbus.org/gridsim/ 9. Al-ali, R.J., Amin, K., Laszewski, G., Rana, O.F., Walker, D.W., Hategan, M., Zaluzec, N.: Analysis and provision of QoS for distributed grid applications. Journal of Grid Computing 2(2), 163–182 (2004) 10. Krauter, K., Buyya, R., Maheswaran, M.: A taxonomy and survey of grid resourcemanagement systems. Software Practice and Experience 32, 135–164 (2002) 11. Sulistio, A., Cibej, U., Venugopal, S., Robic, B., Buyya, R.: A toolkit for modelling and simulating data Grids: an extension to GridSim. Concurr. Comput.: Pract. Exper. 20, 1591–1609 (2008)
Fusion Context Model Based on User Scenario for Smart Service Svetlana Kim1, HagYoung Kim2, and YongIk Yoon1 1
Department of Multimedia Science, Sookmyung Women’s University, Chungpa-dong 2ga, Younsan-Gu, 140-742, Seoul, Korea 2 Electronics and Telecommunications Ressearch Institute(ETRI), 218 Gajeong-ro, Yuseong-gu 305-700, Daejeon, Korea [email protected], [email protected], [email protected]
Abstract. Today, smart phones are used continuously and repeatedly in daily life. In addition, smart phones can be equipped with various sensors. These features take advantage of the functions of the smart phone. The information collected is made to the scenario depending on time, location, action based on the Fusion process. This paper is proposes a service recommendation model based on user scenario using fusion context-awareness. The scenarios can help predict a user’s situation and provide the services in advance. Also, content categories as well as the content type are determined depending on the scenario. The scenario is a method for providing the best service as well as a basis for the user’s situation. Using this method, proposing a new service model with context –awareness is the goal of this paper. Keywords: Context-Awareness, Smart phone, Scenario, User Behavior, Smart Service.
Fusion Context Model Based on User Scenario for Smart Service
507
Fusion elements are making a collection of all relevant information of the user. The information includes the location, user’s action and the user’s areas of interest. The fusion of data is a useful collection of continual user patterns in the form of relevant data. The whole process is undertaken intelligently and automatically, so users are able to use the required services in the relevant situations and events. In this paper, through of fusing the variety of collected user information and pattern analysis, we generate contextual scenarios. Scenarios consist of elements such as user’s patterns, time, location, place and information action. Action information determines the content type selection for the appropriate user’s situation and needs. Services that are provided to users can be optimized into a kind of service and content type based on Scenarios.
2
Related Work
2.1
Context-Awareness Technology
Context awareness technology is finding the optimal solution by input information and user’s information provided at the user’s device interface and sensors. Future, user interface (UI) has the intelligence to implement the required technology, like context awareness technology. In particularly, the technology is recognized as an important technology that affects all fields and industries for a ubiquitous computing environment (health, education, disaster relief, shopping etc.). This technology has also been introduced as being one of four trends / technologies leading 2010 in the ‘Gartner forecast 2010’. In addition, the smart phone is a device that can be possible to exact contextawareness that is an indispensable technology for providing high quality service to smart phone users on either side. So the study of context awareness in smart phones is being extensively conducted by many parties [1][2][3]. In this paper, a custom service model is proposed by gathering information from smart phones and identifying potential patterns and creating scenarios. 2.2
Status of Mobile Context-Awareness
Studies in user situations are actively progressing in many methods and fields. In a study of Service-oriented architecture (SOA) published in 2009 [4] low-level context data was extracted by applying an emphasis on studying events for a long period of time. The SOA model is in order to analyze and reinforce some solid statistics on many or any actions on collected information. Based on SOA, the overall, broad perspective on the study of the user is available and presented here. In addition, the study was performed by using context-awareness by extracting sub context using the built-in sensor in the mobile device and then extracting a higher level of context from a logic inference engine, a logical proposition is proposed and used. In this study, data was collected and processed according to the user’s schedule and timeline using a built-in GPS and Bluetooth mobile sensor. Context Aware Browser (CAB) of SMDC Lab at the University of Udine R&D [5] shows appropriate web content based on using a built-in sensor data from the smart phone. However, CAB provides the content from understanding on the user’s
508
S. Kim, H. Kim, and Y. Yoon
location, so an argument is not available for what the user is performing or doing at each particular location. SenSay is a context-aware model mobile prototype being researched at the Carnegie Mellon University. In this study SenSay recognizes situations of users using input data from around the phone and user, automatically changing the reception mode of the mobile phone depending on the state and situation of the user. But SenSay considers only the time as a factor and cannot apply context-awareness precisely [6-9]. The current studies presented above are not enough to provide sufficient reflection and analysis for the user’s living patterns and situation. This is because most of these studies are limited to time and location. To ensure proper application of this paper, we created scenarios that focused on the time and location but also more importantly identified a regular pattern to the user and use it effectively.
3
Fusion Context Information Model
The Context Information is by all means information on the user’s situation taken from the smart device. Ordinarily all the collected information is sorted and stored depending on a flow of time in a chronological order. The time stamp is a reference point for all of the analysis, processing and convergence on all information. In addition, the location information is the typical user’s context, so location-based research is an area that received the primary focus of research. The Fusion context that is proposed in this paper also focuses on time and location, but it is the concept of combining a variety of other information that is its focus. This other information can consist of web access, IP address, and application category. Fig 1 shows the user information acquisition model.
Fig. 1. Collection information architecture
Fusion Context Model Based on User Scenario for Smart Service
509
Context information is collected throughout the three stages. The user situation in the architecture of the whole system is the most important state and all of the user’s information exists in this main step. The Context Monitoring (CM) monitors the various stages of data continuously, and this acquires the sources of information for collecting and processing. The CM can obtain time, place, environment and weather information with sensors on a smart phone. Using GPS to extract time and place is not difficult. Also, this is able to obtain temperature, humidity, illumination and weather information. The smart phone has benefits of internet access which allows it to install and run various applications within the phone using its capabilities. This feature allows for immediate action from the user and of the user and is suitable to identify the user’s preferences. For this reason the application types collected provide data such as media type, run time, the number of executions on smart phone and internet sites. The information about the using, existing device is also very important as well as the environment and preference. Services should be provided according to the device specifications. These specifications such as screen size, quality, storage capacity, features of hardware devices are required to be checked as a final stage in order to provide personalized services. Additionally, device information about Bluetooth or other information connected to the network is expected to be useful for further N-screen services. This entire user’s information will always be in constant change, the changing situation needs to be updated constantly in Context Monitoring. Finally, information obtained through continual observation is processed and stored in the database in the Context Collection step. The information collected is processed by filter combinations according to needs and is classified into four categories. After this step is generated, renewal information for creating scenarios. “Fusion” means to combine similar categories in the first place. Thus, the user information will then be fused with accurate and quality by being classified and matched according to the salient features and relevant patterns (see Fig 2.)
Fig. 2. Fusion context
510
S. Kim, H. Kim, and Y. Yoon
Analysis of actions, that represents the movement, depending on the flow at the time including stopping, walking or running, taking a bus or a subway. In combination with the environmental data such as current location including brightness, noise level, weather etc. The frequently used application, web accessed sites, SMS and phone reception history can determine user preferences that the user often wants and needs as well as the kind of services required. This data variety has a great influence on the subject matter, variety and type for further recommendations. Additional services can be provided with the distribution through using physical data such as device hardware, internet access, IP and whether or not it is connected to other devices So that the stored data is basis for creating scenarios for providing personalized service. The Fig 3 shows the details of belonging to each category of information that is stored in any entry. According to following classification the situation is determine then scenario will be created.
Fig. 3. Information of Service Category
4
Fusion Context Model Based on Scenario
4.1
Process of Scenario
Information collected does not have any rules or method. It is firstly not easy to identify the user’s situation. The collected information of the user is then stored in the
Fusion Context Model Based on User Scenario for Smart Service
511
specified order and categories. In proposing scenarios (S) the user’s location and time is used to record the collected information (see Fig. 4). To create scenarios a number of factors are used the Location (L) Time (T) Actions (A) based on the user’s collected data. The user’s actions(A) includes four kinds of typical behavior: stop/wait, walk, run and possible transportation by car, bus or subway. Depending on the place and time through the sensor information can determine the user’s behavior. The Table 2 and the Fig 5 shows a simple example scenario.
Fig. 4. Service Building Algorithm Table 1. Scenario Time T1 = morning T2 = morning T3 = morning
Location L1 = home L2 = street L3 = street
Action A1 = wait A2 = wait A3 = walk
Fig. 5. Example of scenario
Scenario S1 S2 S3
512
S. Kim, H. Kim, and Y. Yoon
The scenario S1 is an example of the waiting action at home in the morning. Example S3 shows a scenario of walking on the street in the morning. This scenario is not a common scenario generated every time; it is representative of the exact behavior of the information collected during a certain period. The point of this scenario depends on the location, time of the user and an ability to know in advance the expected users behavior. Time, place and action information have more detailed subcategories. But in this paper, we propose the primary sub categories. The final step is creating scenarios that are identified in the action; this allows for the offering of information on the best content on the service models and provides better services. The form of content rating and symbols from users for example device information can be provided using the most up to date and typical scenarios. The user does not need to research and select information themselves because this system recognizes the recommendation automatically. 4.2
Service Scenario Model
To provide a behavior service to the user uses “Fusion” information. The Fusion information describes and extracts information of various multimedia materials; the fusion contents. These contents have the same information with one scenario and stored for quick and efficient access when needed. In this section we introduce how to utilize “Fusion” information in a scenario. In the [Table 2] shows a predicted service depending on the scenario. In S1, the time is in the morning, the place, home and situation of the user is stationary. The expected situation here is that the user is eating or drinking coffee for breakfast. Table 2. Scenario Scenario
Kind of Content
Type
S1
Schedule, news, breakfast, sport, music, weather inform Schedule, news, sport, music. Movie, traffic info, bus station, metro Schedule, news, sport, music. Movie, traffic info, bus station, metro
Audio , Smartphone image, text Video, audio, Smartphone graphic image, text Audio, text Smartphone
S2
S3
Device
Connected State TV, Tablet PC null
null
The content can be provided in the morning, news, and weather or breakfast recommendation. The users use the main application or service information available on the basis of the existing user preference; which is prepared in advance, whom is
Fusion Context Model Based on User Scenario for Smart Service
513
able to enjoy the service. In addition, the scenario S1 can be modified according to the user behavior and determined content type. In the case of the motion situation the video content is not appropriate, it is hard to use. Therefore, the video content is provided to the user in only wait/stop situations. In scenario S3, the user is walking in the street. In S3, there is provided more necessary information in the street; street advertising, bus arrival times. Considering the walking scenario the content is recommended as a simple text or voice.
5
Conclusion
The proposed scenario and the resulting service recommendation models are expected to give much greater satisfaction to the users due in part to receiving data and information quickly and conveniently in each individual’s busy and hectic lives and lifestyles. In this paper the process of collecting information on the user and how to create scenarios are introduced. The user can be provided with a fitted content type dependent on user situation and through adding action information to the scenarios. To satisfy operator position and interests in providing appropriate services and relevant user position content. The service is applied through using created scenarios in order to easily understand user situations and for performing context-awareness efficiently and effectively. In future research, we have a plan to study the detail conditions and concrete service domain information, method and analysis. Acknowledgment. This work was supported by the Ministry of Knowledge Economy through GLORY project and the (R00044941) Business for Cooperative R&D between Industry, Academy, and Research Institute funded Korea Small and Medium Business Administration in 2011.
References 1. Siewiorek, D., Smailagic, A., Furukawa, J., Moraveji, N., Reiger, K., Shaffer, J.: SenSay: A Context-Aware Mobile Phone. In: Proc. of IEEE International Symposium on Wearable Computers (ISWC) (2003) 2. Zheng, P., Nio, L.M.: The rise of the smart phone. IEEE Distributed Systems Online 7(3) (2010) 3. Raento, M., Oulasvirta, A., Petit, R., Toivonen, H.: ContextPhone - A prototyping platform for context-aware mobile applications. IEEE Pervasive Computing, 51–59 (2005) 4. Yu, L., et al.: Applying Context-awareness to Service-oriented Architecture. In: Proceedings of IEEE International E-Business Engineering Conf (ICEBE 2009), pp. 397– 402 (2009) 5. http://smdc.uniud.it/smdc/en/projects/cab 6. Korpipaa, P., Mantyjarvi, J., Kela, J., Keranen, H., Malm, E.-J.: Managing context information in mobile devices. IEEE Pervasive Computing 2, 42–51 (2003) ISSN 15361268
514
S. Kim, H. Kim, and Y. Yoon
7. Cheung, G., Tan, W., Yoshimura, T.: Double feedback streaming agent for real-time delivery of media over 3G wireless networks. IEEE Trans. Multimedia, Special Issue on Streaming Media (April 2004) 8. Yoon, Y., Kim, S.: Mobile Cloud Computing Technology. NIPA National IT Industry Promotion Agency 1439, 28–39 (2010) 9. Yoon, Y., Kim, S.: Mobile OS. Communication of the Korea Institute of Information Science and Engineers 28/6(253), 79–87 (2010)
A Data Processing Framework for Cloud Environment Based on Hadoop and Grid Middleware* Hyukho Kim1, Woongsup Kim2, Kyoungmook Lee2, and Yangwoo Kim2 1 R&D Innovation Center, SAIT, Maetan-3dong, Yeongtong-gu, Suwon, 416 – South Korea 2 Dept. of Information and Communication Engineering, Dongguk University, Seoul, 100-715 – South Korea [email protected], {woongsup,mooky8099,ywkim}@dongguk.edu
Abstract. Owing to performance improvement of mobile devices, number of mobile applications and their variety has increased exponentially in recent years. However, many of these mobile applications are not executed alone and need server-side Internet services which require computing functions such as processing, networking, and storage. The server-side Internet services are usually provided using computing resources at Cloud data center because mobile applications are rapidly increasing in number and they tend to be more and more complex in nature. In addition, the conventional data managing framework, like 3-tier architecture, face additional problems such as heterogeneous external data to import and the vast amount of data to process. In this paper, we propose a data processing framework for mobile applications based on OGSA-DAI for heterogeneous external data import and MapReduce for large data processing. We designed and implemented a data connector based on OGSA-DAI middleware which can access and integrate heterogeneous data in a distributed environment, supporting various data management functions. And then we deployed a data processing framework (we call this data connector) into a Cloud system for mobile applications. We also used MapReduce programming model for data connector. Finally, we conducted various experiments and showed that our proposed framework can be used to access heterogeneous external data and to process large data with negligible or no system overhead. Keywords: Cloud computing, hadoop, HDFS, mapreduce, grid computing, Globus, OGSA-DAI.
1
Introduction
Owing to performance improvement of mobile devices, number of mobile applications and their variety has increased exponentially in recent years. However, many of these applications are not executed alone and need server-side Internet *
This research was supported by the MKE, Korea, under the ITRC(Information Technology Research Center) support program supervised by the NIPA(NIPA-2011-C1090-1101-0008). This work was supported by the Dongguk University Research Fund of 2011.
services which require computing functions such as processing, networking, and storage. Due to limited processing capacity and small screen of mobile devices, mobile computing works like an old client/server model with mobile clients and remote servers with server-side Internet services. The server-side Internet services are usually provided using Cloud computing resources at Cloud data center because mobile applications are rapidly increasing in number and they tend to be more and more complex in nature. In addition, the conventional data processing methods like 3tier architecture face additional problems such as managing heterogeneous external data to import/export and process the vast amount of data. Cloud computing is a new style of web based computing [1], whereby shared resources, software, and information are provided on demand to personal computers, Smartphones [2], and other similar devices. In cloud computing, scalable resources are provided dynamically through virtualization technology [3]. This allows service providers and users to adjust their computing capacity depending on how much is needed at a given time or for a given task. Typically the implementation of cloud service is composed of storage devices clustered by network, distributed file systems, a set of applications domains, and a middleware supporting programmable service interfaces. And hence cloud computing requires effective cooperation among multiple devices, application domains, and middleware in order to provide scalable and virtualized services, and the performance is heavily affected by various factors such as the amount of data exchanged, response time of applications and middleware, the status of network traffic, and the effectiveness of cooperation among client, middleware and application services. For example, Facebook provides various plug-in based services, such as user login analysis service that analyzes user access and usage patterns. These kinds of services gathered and analyze Facebook’s data, and then posted the analyzed results to Facebook clients. So Cloud systems require storage servers holding user data, network infrastructure enabling user data exchange, applications supporting database access and usage information analysis, and a persistent access point where client applications interact with. Existing cloud implementations have shown that the level of performance, in terms of throughput and predictability, does not satisfy user expectations when there is high volume of data exchange or heavy network traffic. The read and write throughput is small and highly variable causing significant fluctuation in client performance [4][5]. And they also do not guarantee the acceptable response time when various types of data are used [6]. In this paper, we propose a method to reduce the data processing time and hence enable guaranteed QoS for the Cloud system. We designed and implemented a data processing framework (we call this data connector) extending OGSA-DAI middleware [3] which can access, import and integrate heterogeneous data in a distributed environment. And we deployed our data connector into a Cloud system for external application usage. We used HDFS [7][8] in Hadoop [9] framework to managing distributed file system. To provide programmable service interface, we adopted MapReduce programming model [10][11][12] which is applied for processing distributed large volume of data.
A Data Processing Framework for Cloud Environment
517
Our data processing framework aims at integrating and managing heterogeneous data sources such as database and distributed file systems using Hadoop and grid based middleware OGSA-DAI. Through our proposed framework, files in user file system are located first using single query, then transferred to local storage in the Cloud data center, and finally uploaded to HDFS. We used HDFS for the primary file system and MapReduce for complex data analysis and processing tasks. Our data process framework is based on OGSA-DAI, and we extended and customized OGSADAI to support various file management functions like data ‘search’, ‘transfer’, and ‘upload’ because current OGSA-DAI implementation focuses on simple unified data access only to heterogeneous database sources not to heterogeneous distributed file system. Therefore our approach will provide more diverse data resource management functions using simple unified query that is possible in OGSA-DAI middleware. This paper is organized as follows. Chapter 2 introduces background technologies, such as OGSA-DAI, MapReduce, and HDFS, which we used to implement our data processing framework. We also mentioned the limitations in current OGSA-DAI and HDFS and improvement we obtained. In chapter 3, we briefly describe our system architecture and our primary strategy to implement data connector. Chapter 4 illustrates our experimental results to show our approach has benefits in handling large volume of data while our approach does not give significant overhead to existing approach.
2
Background Technologies
Hadoop [9] consists of two major components - Hadoop's Distributed File System (HDFS) [7][8] and MapReduce framework [9][10][11]. These two components working together allow for Hadoop to promise ease of programming and high reliability. MapReduce is a software framework for easily writing applications which process vast amounts of data in-parallel on large clusters of commodity hardware in a reliable, fault-tolerant manner. A job submitted to MapReduce usually splits the input data-set into independent chunks which are processed by the map tasks in a parallel. Then the framework sorts the outputs of the maps, which are then inputs to the reduce tasks. (Fig. 1)
Fig. 1. Map Reduce Data Flow [9]
518
H. Kim et al.
The HDFS is modeled very closely on the Google file system [13]. The approach to this file system assumes that failure in a large scale computing environment is commonplace, rather than a unique event. HDFS uses a three way replication scheme to ensure that the files stored are always intact in three separate places across a Hadoop cluster. Fig. 2 shows a user application accessing the file system in HDFS.
Fig. 2. HDFS architecture [9]
OGSA-DAI (Open Grid Service Architecture – Data Access and Integration) is developed as an extensible framework accessed via web services that executes datacentric workflows involving heterogeneous data resources for the purposes of data access, integration, transformation and delivery within a grid and is intended as a toolkit for building higher-level application-specific data services. OGSA-DAI provides three types of Grid services: data resource registration and discovery, Grid proxies for data resources, and data access and manipulation [3]. Using OGSA-DAI middleware implementation has a benefit that only one single query is needed to access multiple various data resources and only one single result is returned as a query. However, the usage of function is currently limited to some data resource types. Moreover, data management functions are mainly focused on data access rather than data transfer, so that there are limitations to cover diverse client application needs.
3
System Design
3.1
System Architecture
In this section, we present the overall system architecture which is composed of job scheduler, allocator, MapReduce-based (MR-based) data processing service, OGSADAI data processing components and interfaces. As you can see Fig. 3, we used HDFS as a primary distributed file system. And we implemented data connector (noted three ‘C’ in Fig 3.) for the access and loading of
A Data Processing Framework for Cloud Environment
519
data which is located in the external data of HDFS. The data resource connector can use all of data resources which are supported by OGSA-DAI middleware, such as, RDBMS (DB2, MySQL, Oracle, PostgreSQL, SQL Server), XMLDB, eXist, and user file systems.
Fig. 3. System architecture
We used two kinds of the data handling strategies: one is Load-and-Copy method. This strategy loads the external data and then, makes a copy of the loaded data to HDFS. The other is Load-Only method. It loads the external data and then, the loaded data is deleted right after data processing is completed. In the Load-and-Copy method, it has the disadvantage that it takes a long time to load and create replication data into HDFS. However, if same data is re-used frequently, it has the advantage that it can analyze faster than first processing time due to HDFS replication mechanism. On the other hand, the Load-Only has the advantage for faster analysis time when new data is used as the method does not require time for data replication. It, however, has the disadvantage that it is slower than the Load-and-Copy method when there are frequent data reuses. The performance comparison of two data handing methodologies is shown in Chapter 4. Job Scheduler, Allocator, and MR (MapReduce)-based Data Analysis Service are running based on HDFS. Job Scheduler uses FIFO scheduling algorithm. Allocator is responsible for checking if the system uses the external data, allocates tasks to task nodes depends on which data is used. Finally, MR-based Data Analysis Service is responsible for carrying out data processing service. 3.2
System Configuration
We used 5 nodes for our experiments. Each computing node has two 3.00 GHz CPUs, 1GB Memory, 120 GB local disk space, and connected to the others by Ethernet. We used only a 1Gbps Ethernet for the network communication. Every node uses Fedora Linux Core 8. The home directory is shared by NFS, and each node has own local disk. For implementation, we used Apache Hadoop 0.20, Grid middleware – Globus
520
H. Kim et al.
Toolkit [14][15], and OGSA-DAI component [16] for system configuration. We also designed and implemented Data Analysis Service based on Android platform for measuring of data processing performance. It is running on mobile devices. And we used additional tools for experiments such as Apache Ant 1.6.5, J2SDK 1.6, and Apache Log4j-1.2.8.
4
Evaluation
4.1
Analysis of Data Connector Performance
In this paper, we used HDFS as basic file system. We conducted experimental comparison of the two data handling strategies for the external data handling. The external data is located in user file system which is supported by OGSA-DAI. We used the same size and type of the data which is located in HDFS and OGSA-DAI supported file system. And we measured the analysis time by increasing the data size.
Fig. 4. The performance results of two data connectors
The performance result of two data handling strategies implementing data connector is shown in Fig. 4. We compared Load-and-Copy data connector with Load-Only data connector. And job requests are sent by mobile device. So we measured the request time (Req. time) to send jobs to MR-based Data Analysis Service. Both of two data connectors do not heavily affect the data analyzing time because the Req. time takes quite a short duration. And the analysis time of Load-Only connector is a little faster than Load-and-Copy connector. However as you can see graphs in Fig 4, performance differences of two data processing strategies are quite small. Therefore we integrated Load-and-Copy strategies in our framework and used it in all the rest experiments. We also conducted additional experiments for comparison of file download time with our data connector, OGSA-DAI, and HDFS.
A Data Processing Framework for Cloud Environment
521
Fig. 5. The analysis results of file download time
HDFS supports various functions like file upload function (put) and file download function (get), list function (ls), etc. HDFS shows the best performance because it was designed to be optimized the data I/O operations. OGSA-DAI middleware, however, performed the worst because it focuses to access and integrate the heterogeneous data resources. And OGSA-DAI middleware doesn’t work alone. It deploys into Globus Toolkit (GT) or Axis/Tomcat container and then, it works on the container. In other words, OGSA-DAI takes a long time because it has additional processes like property parsing of Data resource, service location and service behavior confirmation, etc. So data downloading of OGSAI-DAI middleware had poor performance. Lastly, Data Connector is worse than HDFS because it was designed and implemented with OGSA-DAI libraries for access and loading the data which is located in the remote file systems. But it is better than OGSA-DAI middleware because it was implemented to be optimized the transmission function by removing unnecessary parsing processes. 4.2
Analysis of MapReduce and Multi-thread with Data in HDFS
This experiment is aimed at comparing the performance of typical data analysis using multi threading (thread-based analysis) and MR-based data analysis with data in HDFS only. And we do not include data request time in the result because we believe data request time is quite small considering total data analysis time and does not significantly affect the comparison results.
Fig. 6. The experiment results with data in HDFS
522
H. Kim et al.
We conducted experiments for performance analysis of those two methodologies with a varying data size. As you can see in Fig. 6, two methods have similar experimental results. MR-based data analysis, however, is better than thread-based data analysis, though the performance difference is small. That is, when MR-based analysis method applied to mobile Cloud system, it means that MR-based data analysis can provide good performance. 4.3
Analysis of MapReduce and Multi-thread with Data in Data Resource
This experiment is designed to compare the performance of thread-based data analysis and MR-based data analysis with data under OGSA-DAI supported file system. The experiment results are shown in Fig. 7. The performance of two methods is almost same because two methods used the data which is located in the external file system. These results imply that the overhead to load and analyze the data using Data Grid middleware is quite small.
Fig. 7. The experiment results with data in OGSA-DAI supported file system
4.4
Analysis of MapReduce and Multi-thread with Data in HDFS and Data Resource
We conducted experiments with data which is located in HDFS and OGSA-DAI supported file system. The experiment results are shown in Fig. 8. The performance difference of two methods is not significant when we use small size of data. Therefore, we figure out that there is no noteworthy operational overhead though both HDFS and OGSA-DAI supported file system are used together. We, however, noticed significant performance difference when large volume of data is used. As a result, MR-based data analysis shows better performance than the typical thread-based method, when large volume of data is used and exchanged in distributed file system.
A Data Processing Framework for Cloud Environment
523
Fig. 8. The experiment results with data in HDFS and OGSA-DAI combined file system
5
Conclusions
In paper, we proposed the method for reducing of data processing overhead in Cloud systems. We also proposed application method extending Grid middleware for accessing and importing the large volume of external data. So we implemented Data process framework and then, deployed it into Cloud system. For experimental purpose, we made a test bed utilizing existing HDFS and OGSA-DAI frameworks supporting distributed file management and conducted comparative experiments to measure the effectiveness of our proposed method. From the experiment results, we conclude that MapReduce-based data processing method shows better performance than the existing method, especially when we use large volume of external data. And the overhead of using Grid middleware was quite small even though our approach uses both HDFS and OGSA-DAI supported file system together, and it does not affect overall the system performance. As a result, we present that Grid middleware can be applied to Cloud environment to improve data processing performance, and using with Hadoop framework together considering Grid middleware does not make significant overhead for the performance of MapReduce-based data connector.
References 1. Kovachev., D., Renzel., D., Klamma, R., Cao, Y.: Mobile community cloud computing: emerges and evolves. In: Proc. 1st Intl. Workshop on Mobile Cloud Computing (MDM 2010), pp. 393–395 (2010) 2. Smartphone, Wikipedia, http://en.wikipedia.org/wiki/Smartphone/ 3. Marinelli, E.: Hyrax: cloud computing on mobile devices using MapReduce. Master thesis, Carnegie Mellon University (2009) 4. Wang, G., Ng, T.: The impact of virtualization on network performance of amazon ec2 data center. In: Proc.of INFOCOMM, pp. 1–9 (2010) 5. Wang, J., Varman, P., Xie, C.: Avoiding performance fluctuation in cloud storage. In: Proc., of HiPC (2010)
524
H. Kim et al.
6. Brito, M., Kakugawa, F., Sato, L., Correa, P.: An Architecture for Integrating Databases with Replication Support Based on the OGSA-DAI Middleware. In: Proc. of International Conference on Computational Science and Engineering (2009) 7. Jiang., L., Li., B., Song, M.: THE optimization of HDFS based on small files. In: Proc. of 3rd IEEE International Conference on Broadband Network and Multimedia Technology, pp. 912–915 (2010) 8. Mackey, G., Sehrish, S., Wang, J.: Improving metadata management for small files in HDFS. In: Proc. of IEEE International Conference on Cluster Computing, pp. 1–4 (2009) 9. Hadoop, A.: http://hadoop.apache.org/ 10. Huang, L., Wang, X.-W., Zhai, Y.-D., Yang, B.: Extraction of User Profile Based on the Hadoop Framework. In: Proc. of IEEE Conf. on Wireless Communications, Networking and Mobile Computing, pp. 1–6 (2009) 11. Gunarathne, T., Wu, T.-L., Qiu, J., Fox, G.: MapReduce in the Clouds for Science. In: Proc. of IEEE Conf. on Second International Conference, pp. 565–572 (2010) 12. Mackey, G., Sehrish, S., Bent, J., Lopez, J., Habib, S., Wang, J.: Introducing map-reduce to high end computing. In: Proc. of Petascale Data Storage Workshop, pp. 1–6 (2008) 13. Dean, J.: Experiences with mapreduce, an abstraction for large-scale computation. In: Proc. of the 15th International Conference on Parallel Architectures and Compilation Techniques, New York, pp. 1–1 (2006) 14. Montero., R.S., Huedo., E., Llorente, I.M.: Dynamic deployment of custom execution environments in Grids. In: 2nd International Conference on Advanced Engineering Computing and Applications in Sciences, pp. 33–38 (2008) 15. Foster, I., Kesselman, C., Tuecke, S.: The Anatomy of the Grid: Enabling Scalable Virtual Organizations. International Journal of High Performance Computing Applications 15(3), 200–222 (2001) 16. Grant, A., Antonioletti, M., Hume, A.C., Krause, A., Dobrzelecki, B., Jackson, M.J., Parsons, M., Atkinson, M.P., Theocharopoulos, E.: OGSA-DAI: Middleware for Data Integration: Selected Applications. In: Proc. of IEEE Fourth International Conference on eScience, pp. 343–343 (2008)
CloudTSS: A TagSNP Selection Approach on Cloud Computing Che-Lun Hung1, Yaw-Ling Lin2, Guan-Jie Hua2, and Yu-Chen Hu3 1
Dept. of Computer Science & Communication Engineering, Providence University, 200 Chung Chi Rd., Taichung 43301, Republic of China, Taiwan [email protected] 2 Dept. of Computer Science & Information Engineering, Providence University, 200 Chung Chi Rd., Taichung 43301, Republic of China, Taiwan [email protected], [email protected] 3 Dept. of Computer Science & Information Management, Providence University, 200 Chung Chi Rd., Taichung 43301, Republic of China, Taiwan [email protected]
Abstract. SNPs are fundamental roles for various applications including medical diagnostic, phylogenies and drug design. They provide the highestresolution genetic fingerprint for identifying disease associations and human features. Genetic variants that are near each other tend to be inherited together; these regions of linked variants are known as haplotypes. Recently, genetics researches revealed that SNPs within certain haplotype blocks induce only a few distinct common haplotypes in the majority of the population. The existence of haplotype block structure has serious implications for associationbased methods for the mapping of disease genes. This paper proposes a parallel haplotype block partition and SNPs selection method under a diversity function by using the Hadoop MapReduce framework. The experiment shows that the proposed MapReduce-paralleled combinatorial algorithm performs well on the real-world data obtained in from the HapMap data set; the computation efficiency can be significantly improved proportional to the number of processors being used. Keywords: Hadoop, MapReduce, SNPs, Haplotype, cloud computing.
[1-6]. Many studies suggest that human genome maybe arranged into block structure, in which SNPs are relevant and only a small number of SNPs are sufficient to obtain most of haplotype structures, called tagSNP [7-13]. Several approaches have been suggested for defining block structure from which some are more commonly used. Four main criteria for haplotype block partitioning are based on haplotype diversity, LD, four gamete test and information complexity. In diversity-based methods [9, 14, 15], a block is defined as a region, in which a certain percentage of haplotypes are common haplotypes, haplotypes which are represented more than certain percent in the population. In LD-based methods [8, 16], a block is defined as a region with high pair-wise LD with in block and low pair-wise LD between blocks. In methods based on four gamete test [17, 18], a block is defined as a recombination-free region of consecutive SNPs. However, existing approaches that address these issues are generally complicated and computationally intensive. Hadoop [19] is a software framework intended to support data-intensive distributed applications. It is able to process petabytes of data with thousands of nodes. Hadoop supports MapReduce programming model [20] for writing applications that process large data set in parallel on Cloud Computing environment. The advantage of MapReduce is that it allows for distributed computing of the map and reduction operations. Each map operation is independent of the others and all maps can perform the tasks in parallel. In practice, the total number of the maps is limited by the data source and/or the number of CPUs near that data. Similarly, a set of reducers can perform the reduce operations. All outputs of the map operation which share the same key are presented to the same reducer, at the same time. In addition, one of the important benefits to use Hadoop to develop the applications is due to its high degree of fault tolerance. Even when running jobs on a large cluster where individual nodes or network components may experience high rates of failure, Hadoop can guide jobs toward a successful completion. Many applications of bioinformatics are often computation-consuming; sometimes it needs weeks or months to complete the jobs. The traditional parallel models, such as MPI, OpenMP and Multi-thread, are not suitable to such applications, where a fault occurred in some nodes leads the entire application into total failure. In these situations, the Hadoop platform is considered as a much better solution for these real-world applications. Recently, Hadoop has been applied in various domains in bioinformatics [21, 22]. In this paper, we propose parallel diversity-based haplotype block selection algorithms on Hadoop MapReduce framework. In particular, the map-stage is utilized to calculate the diversity of each block, while the reduce-stage is utilized to locate the blocks. As these experimental results indicate that these proposed algorithms are significantly faster than the corresponding sequential algorithms as the number of Map operations increased.
2
Method and Materials
Haplotype Block Partitioning. A SNP (Single Nucleotide Polymorphisms) is defined as a position in a chromosome where each one of two (or more) specific nucleotides is observed in at least 10% of the population [9]. The nucleotides involved in a SNP are called alleles. A SNP is said biallelic if it has only two different alleles. Almost all SNP are biallelic and we will consider exclusively biallelic SNP in this paper.
CloudTSS: A TagSNP Selection Approach on Cloud Computing
527
Diversity Function. Abstractly, the input to the haplotype blocking problem consists of m haplotype vectors. Each position in a vector is associated with a site of interest on the chromosome. Usually, the position in the haplotype vector has a value of 0 if it is the major allele or 1 if minor allele. Let the haplotype matrix A be an m×n matrix of m observations over n markers (sites). We refer to the jth allele of observation i by Aij. For simplicity, we assume that Aij ∈ {0, 1}. A block, or marker interval, [j, k] = {j, j + 1, …, k} is defined by two marker indices 1 ≤ j ≤ k ≤ n. A segmentation is a set of non-overlapping non-empty marker intervals. A segmentation is full if the union of the intervals is [1, n]. The data matrix limited to interval [j, k] is denoted by M(j, k); the values of the i-th observation are denoted by M(i, j, k), a binary string of length k － j + 1. Figure 1 illustrates an example of a 7×11 haplotype matrix. Given an interval [j, k], a diversity function, δ : [j, k] → (j, k) ∈ R is an evaluation function measuring the diversity of the submatrix M(j, k). We say an interval [j’, k’] is a subinterval of [j, k], written [j’, k’] ⊂ [j, k], if j ≤ j’ and k’ ≤ k. Note that δ-function is a monotonic non-decreasing function from [1..n, 1..n] to the unit real interval [0, 1]; that is, 0 ≤ ( j’, k’) ≤ (j, k) ≤ 1 whenever [j’, k’] ⊂ [j, k]. Given an input set of n haplotype vectors, a solution to the Haplotype Block Selection (HBS) Problem is a segmentation of marker intervals, revealing these nonoverlapped haplotype blocks of interest in the chromosome. In our previous work [23], we have proposed several efficient algorithms related to selecting interesting haplotype blocks under different evaluation (diversity) functions that generalizes many previous results in the literatures [8, 9, 12, 13, 16, 17, 18, 24, 25, 26]. In this paper, we focus on finding common haplotypes. Common Haplotypes. Two haplotypes are said to be compatible if the alleles are identical at all loci for which there are no missing data; otherwise the two haplotypes are said to be incompatible. As in Patil et al., [9] we define the ambiguous haplotypes as those haplotypes compatible with at least two haplotypes that are themselves incompatible. It should be noted that when there are no missing data, all of the haplotypes are unambiguous. We define the common haplotypes as those haplotypes that are represented more than once in a block. The haplotypes are called singleton if they are not compatible with any others.
Fig. 1. A haplotype matrix B and its corresponding submatrix M(8, 11)
528
C.-L. Hung et al.
We are mainly interested in the common haplotypes. Therefore we require that, in the final block partition, a significant fraction of the haplotypes in each block is common haplotypes. Patil et al. [9] require that at least α = 70%, 80%, and 90%, respectively, of the unambiguous haplotypes appear more than once. The parameter, α, is also referred to as the coverage of common haplotypes in a block. Ambiguous haplotypes are not included in calculating percent coverage. The coverage of block B can be mathematically formulated as a form of diversity: δ S ( B) = 1 −
U U = C S
(1)
Here U denotes the number of unambiguous haplotypes, C denotes the number of common haplotypes, and S denotes the number of singleton haplotypes. For example, the submatrix M(8, 11) of Fig. 1 can be viewed as a sample S = {0001, 0011, 0100, 1000, 1011}. It follows that δ(S) = δ(M(8, 11)) =1 4/7 = 3/7.
－
TagSNPs Selection. For each block, we want to minimize the number of SNPs that uniquely distinguish at least 80% (the α parameter) of the unambiguous haplotypes in the block. Those SNPs can be interpreted as a signature of the haplotype block partition. They are referred to as tagSNPs that are able to capture most of the haplotype diversity, and therefore, could potentially capture most of the information for association between a trait and the marker loci [28]. Our strategy for selecting the tagSNPs in haplotype blocks is as the following. First, the common haplotypes are grouped into k distinct patterns in each block. After the missing data are assigned, as explained in the next subsection, we decide the least number of groups needed such that haplotypes in these groups contain at least 80% ( α ) of the unambiguous haplotypes in the block. Finally, we select a loci set which consists of the minimum number of SNPs on the haplotypes such that each pattern can be uniquely distinguish. Exhaustive searching methods are used very efficiently here since the number of tagSNPs needed for each block is usually modest in the situation. The exhaustive searching algorithm enumerates next γ-combination in lexicographic order to generate the next candidate tagSNP loci set until each pattern can be uniquely distinguish. 2.1
Hadoop MapReduce Framework
Hadoop is a software framework for coordinating computing nodes to process distributed data in parallel. Hadoop adopts the map/reduce parallel programming model, to develop parallel computing applications. The standard map/reduce mechanism has been applied in many successful Cloud computing service providers, such as Yahoo, Amazon EC2, IBM, Google and so on. An application developed by Map/Reduce is composed of Map stage and Reduce stage (optionally). Input data will be split into smaller chunks corresponding to the number of Maps. Output of Map stage has the format of pairs. Output from all Map nodes, pairs, are classified by key before being distributed to Reduce stage. Reduce stage combines value by key. Output of Reduce stage are pairs where each key is unique.
CloudTSS: A TagSNP Selection Approach on Cloud Computing
529
Hadoop cluster includes a single master and multiple slave nodes. The master node consists of a jobtracker, tasktracker, namenode, and datanode. A slave node, as computing node, consists of a datanode and tasktracker. The jobtracker is the service within Hadoop that farms out Map/Reduce tasks to specific nodes in the cluster, ideally the nodes that have the data, or at least are in the same rack. A tasktracker is a node in the cluster that accepts tasks; Map, Reduce and Shuffle operations from a jobtracker. Hadoop Distributed File System (HDFS) is the primary file system used by Hadoop framework. Each input file is split into data blocks that are distributed on datanodes. Hadoop also creates multiple replicas of data blocks and distributes them on datanodes throughout a cluster to enable reliable, extremely rapid computations. The namenode serves as both a directory namespace manager and a node metadata manager for the HDFS. There is a single namenode running in HDFS architecture. 2.2
Haplotype Block Partitioning and Selection on Map/Reduce Framework
Figure 2 illustrates the MapReduce framework for the block partitioning and selection scheme. Assume that the number of map operations is N and the pattern length is L, the input N×L haplotype matrix is split into L/N chunks. Each map calculates the diversity scores of each block within the chunk where the map operationis responsible. Thus the output pairs for each Map are <(block start number, block end number), diversity score> pairs. The mapi calculates diversity scores of blacks {δ(i N/L, i N/L), δ(i N/L, i N/L+1), …, δ(i N/L+N/L, i N/L+N/L)}. Therefore, each map has (N/L)2 diversity scores. Reduce stage performs haplotype block selection algorithm. In our algorithm, just one reduce operation is needed in the reduce stage. Since the selection is a linear time algorithm, it is not necessary to perform the computation in parallel. The haplotype block selection algorithm is described in the previous section 2.1. The reduce operation finds the longest block by merging blocks with the interesting diversity scores.
•
•
•
•
•
•
Fig. 2. Haplotype block partitioning and selection on MapReduce framework
530
C.-L. Hung et al.
3
Experiment
3.1
Experimental Environment and Data Source
All of the experiments were performed on two IBM blade servers within our Cloud Computation laboratory. Each server is equipped with two Quad-Core Intel Xeon 2.26GHz CPU, 24G RAM, and 296G hard disk running under the operation system Ubuntu version 10.4 with Hadoop version 0.2 MapReduce platform. Under the current system environment, we control the server execution processes by up to 8 map operations and 8 reduce operations and the total number of the map/reduce operations are up to 16 respectively. The SNP haplotype data sources are gathered from the International HapMap Project (http://hapmap.ncbi.nlm.nih.gov/), which is a multi-country effort in order to identify and catalog genetic similarities and differences in human beings. This project collects many genetic sequences of different individuals. We downloaded the sequence data (Chromosome 1) from the HapMap3 Genome Browser release #2African ancestry in Southwest USA (ASW). ASW includes 136 Chromosome 1 (chr 1) sequences (patterns) and the length of SNP is 116,416. These sequences are treated the input data for our experiments. In the experiments, we applied formula (1) to as calculate diversity scores of blocks. 3.2
Experimental Results
To assess the performance of the proposed Hadoop MapReduce algorithm, we compare the computational time between various sequence data and various number of map/reduce operations. The sequential algorithm [23] has been proved as an efficient algorithm than other algorithms. Two factors, number of patterns and the length of patterns, affect the performance of sequential algorithm and the proposed algorithm. Patil et al [9] proposed that the haplotype block can be found within 300bp and 500bp. Therefore, the block size can be 300bp and 500bp. The diversity scores are calculated according to their corresponding block sizes; these scores are {δ(1, 1), δ(1, 2), …, δ(1, 500), δ(2, 2), …, δ(2, 501), δ(3, 3), …, δ(L, L)}. Figures 3 and 4 illustrate the comparisons between sequential algorithm and our proposed algorithm under the MapReduce framework. In Fig. 3 and Fig.4, the block sizes are 300bp and 500bp, respectively. In Fig. 3 and Fig. 4, it is observed that computational time increases corresponding to number of pattern and sequence length. The computational time for our algorithm with block size 300bp is less than that with block size 500bp. More patterns and longer sequence length lead to higher computational cost. These experimental results are corresponding to the algorithm analysis in previous section. The experimental results reveal that the computational time is effectively reduced when more map operations are deployed. Two and four map operations almost improve the computation time by factors of two and four times accordingly, comparing to the original sequential algorithm, respectively. Moderate enhancements between 8 and 16 map operations are observed in all experiments, since the size of data set split by 8 is similar to that by 16. Figure 5 illustrates the computation efficiency can be effectively improved proportional to the number of processors being used.
CloudTSS: A TagSNP Selection Approach on Cloud Computing
531
Fig. 3. Performance comparison between sequential haplotype block selection and MapReduce haplotype block selection with block size 300bp
Fig. 4. Performance comparison between sequential haplotype block selection and MapReduce haplotype block selection with block size 500bp
532
C.-L. Hung et al.
Fig. 5. Speed up comparisons for MapReduce haplotype block selection over equential haplotype block selection. (a) illustrats the speed up with block size 300bp. (b) illustrats the speed up with block size with 500 bp.
4
Conclusions
Studying on SNP and haplotype blocks can assist biomedical researchers to detect inherited diseases and contribute to classify the race of human and researching on species evolution. The traditional methods for detecting haplotype blocks are based on dynamic programming approach. With the abundance of bioinformatic data that are all too common these days, the time-consuming traditional sequential methods require imminent assistance of the emerging parallel processing methodology. Here in this paper we discuss how we develop the parallelized frame works improving our original dynamic programming algorithms, based on Hadoop map/reduce framework. The haplotype block partitioning copes with the problem of finding the minimum number of representative SNPs required to account for most of the haplotype block quality in each block. Due to the fault tolerance of Hadoop, the jobs are just re-submitted to other nodes if the node is failure. This property is useful for analyzing large amount of sequence data since the job will not be stopped by node’s fail. The experimental results show that the proposed algorithm can decrease the computational cost significantly. In this paper, we compared the performance between various sequence lengths and pattern numbers. We also compared the performance between different block sizes. In the future, we will apply more diversity functions to the parallel algorithms to provide more perspectives for biologists to analyze these SNP data and investigate the relations of haplotype block selection between various block sizes and diversity measurements.
References 1. Bonnen, P.E., Wang, P.J., Kimmel, M., Chakraborty, R., Nelson, D.L.: Haplotype and linkage disequilibrium architecture for human cancer-associated genes. Genome Res. 12, 1846–1853 (2002) 2. Gray, I.C., Campbell, D.A., Spurr, N.K.: Single nucleotide polymorphisms as tools in human genetics. Hum. Mol. Genet. 9, 2403–2408 (2000)
CloudTSS: A TagSNP Selection Approach on Cloud Computing
533
3. Indap, A.R., Marth, G.T., Struble, C.A., Tonellato, P.J., Olivier, M.: Analysis of concordance of different haplotype block partitioning algorithms haplotype tagging for the identification of common disease genes. BMC Bioinformatics 6, 303 (2005) 4. Mas, A., Blanco, E., Monux, G., Urcelay, E., Serrano, F.J., de la Concha, E.G., Martinez, A.: DRB1-TNF-alpha-TNF-beta haplotype is strongly associated with severe aortoiliac occlusive disease, a clinical form of atherosclerosis. Hum. Immunol. 66, 1062–1067 (2005) 5. Nowotny, P., Kwon, J.M., Goate, A.M.: SNP analysis to dissect human traits. Curr. Opinion Neurobiol. 11, 637–641 (2001) 6. Reif, A., Herterich, S., Strobel, A., Ehlis, A.C., Saur, D., Jacob, C.P., Wienker, T., Topner, T., Fritzen, S., Walter, U., Schmitt, A., Fallgatter, A.J., Lesch, K.P.: A neuronal nitri coxide synthase (NOS-I) haplotype associated with schizo-phrenia modifies prefront alcortex function. Mol. Psychiatry 11, 286–300 (2006) 7. Daly, M.J., Rioux, J.D., Schaffner, S.F., Hudson, T.J., Lander, E.S.: High-resolution haplotype structure in the human genome. Nat. Genet. 29, 229–232 (2001) 8. Gabriel, S.B., Schaffner, S.F., Nguyen, H., Moore, J.M., Roy, J., Blumenstiel, B., Higgins, J., DeFelice, M., Lochner, A., Faggart, M., Liu-Cordero, S.N., Rotimi, C., Adeyemo, A., Cooper, R., Ward, R., Lander, E.S., Daly, M.J., Altshuler, D.: The structure of haplotype blocks in the human genome. Science 296, 2225–2229 (2002) 9. Patil, N., Berno, A.J., Hinds, D.A., Barrett, W.A., Doshi, J.M., Hacker, C.R., Kautzer, C.R., Lee, D.H., Marjoribanks, C., McDonough, D.P., Nguyen, B.T.N., Norris, M.C., Sheehan, J.B., Shen, N.P., Stern, D., Stokowski, R.P., Thomas, D.J., Trulson, M.O., Vyas, K.R., Frazer, K.A., Fodor, S.P.A., Cox, D.R.: Blocks of limited haplotype diversity revealed by high-resolution scanning of human chromo- some 21. Science 294, 1719–1723 (2001) 10. Dawson, E., Abecasis, G.R., Bumpstead, S., Chen, Y., Hunt, S., Beare, D.M., Pabial, J., Dibling, T., Tinsley, E., Kirby, S.: First-generation linkage disequilibrium map of human chromosome 22. Nature 418, 544–548 (2002) 11. Mahdevar, G., Zahiri, J., Sadeghi, M., Nowzari-Dalini, A., Ahrabian, H.: Tag SNP selection via a genetic algorithm. J. Biomed. Inf. (2010), doi:10.1016/j.jbi.2010.05.011 12. Zhang, K., Calabrese, P., Nordborg, M., Sun, F.: Haplotype block structure and its applications to association studies: power and study designs. Am. J. Hum. Genet. 71, 1386–1394 (2002) 13. Wall, J.D., Pritchard, J.K.: Assessing the performance of the haplotype block model of linkage disequilibrium. Am. J. Hum. Genet. 73, 502–515 (2003) 14. Johnson, G.C.L., Esposito, L., Barratt, B.J., Smith, A.N., Heward, J., Di Genova, G., Ueda, H., Cordell, H.J., Eaves, I.A., Dudbridge, F., Twells, R.C.J., Payne, F., Hughes, W., Nutland, S., Stevens, H., Carr, P., Tuomilehto-Wolf, E., Tuomilehto, J., Gough, S.C.L., Clayton, D.G., Todd, J.A.: Haplotype tagging for the identification of common disease genes. Nat. Genet. 29, 233–237 (2001) 15. Zahirib, J., Mahdevar, G., Nowzari-dalini, A., Ahrabian, H., Sadeghic, M.: A novel efficient dynamic programming algorithm for haplotype block partitioning. J. Theor. Biol. 267, 164–170 (2010) 16. Greenspan, G., Geiger, D.: High density linkage disequilibrium mapping using models of haplotype block variation. Bioinformatics 20, i137 (2004) 17. Wang, N., Akey, J.M., Zhang, K., Chakraborty, R., Jin, L.: Distribution of recombination crossovers and the origin of haplotype blocks: the interplay of population history, recombination, and mutation. Am. J. Hum. Genet. 71, 1227–1234 (2002)
534
C.-L. Hung et al.
18. Hudson, R.R., Kaplan, N.L.: Statistical properties of the number of recombination events in the history of a sample of DNA sequences. Genetics 111, 147–164 (1985) 19. Hadoop - Apache Software Foundation project home page, http://hadoop.apache.org/ 20. Taylor, R.C.: An overview of the Hadoop/MapReduce/HBase framework and its current applications in bioinformatics. BMC Bioinformatics 11, S1 (2010) 21. Dean, J., Ghemawat, S.: MapReduce: A Flexible Data Processing Tool. Communications of the ACM 53, 72–77 (2010) 22. Schatz, M.: Cloudburst: highly sensitive read mapping with MapReduce. Bioinformatics 25, 1363–1369 (2009) 23. Lin, Y.L.: Efficient Algorithms for SNP Haplotype Block Selection Problems. In: Hu, X., Wang, J. (eds.) COCOON 2008. LNCS, vol. 5092, pp. 309–318. Springer, Heidelberg (2008) 24. Clayton, D.: Choosing a set of haplotype tagging SNPs from a larger set of diallelic loci. Nature Genetics 29(2) (2001) 25. Zhang, K., Qin, Z., Liu, J.S., Chen, T., Waterman, M.S., Sun, F.: Haplotype block partitioning and tag SNP selection using genotype data and their applications to association studies. Genome Res. 14, 908–916 (2004) 26. Anderson, E.C., Novembre, J.: Finding haplotype block boundaries by using the minimum-description-length principle. Am. J. of Human Genetics 73, 336–354 (2003) 27. Li, W.H., Graur, D.: Fundamentals of Molecular Evolution. Sinauer Associates, Inc. (1991) 28. Chapman, J.M., Cooper, J.D., Todd, J.A., Clayton, D.G.: Detecting disease associations due to linkage disequilibrium using haplotype tags: a class of tests and the determinants of statistical power. Hum. Hered. 56, 18–31 (2003)
The Study on Spatial Reuse Efficiency of Network Nodes Using Advanced Power Control Algorithm* Seung-dae Lee Electronic Engineering, Namseoul Univ., 21 Maeju-ri, Seonghwan-eup, Seobuk-gu, Cheonan-city, Choongnam, Korea [email protected]
Abstract. An algorithm to minimize power consumption while keeping network throughput such as distributed power control, PCM(power control MAC), and F-PCF(fragmentation based PCM) was suggested to reduce power consumption through power control up to the present. A potential spatial for communication of surrounding nodes was generated as such power control method was planned to make optimal transmit power in order to minimize power consumption. But, potential spatial for communication generated through power control had not yet been utilized appropriately as an algorithm focused on power consumption. This study suggested an algorithm that advanced lots of network throughput by using potential spatial for communication generated by power control. Keywords: Network node, PCM, Ad Hoc, MANET, Power control.
1
Introduction
There had actively been study on delay reduction, bandwidth fair distribution, power control and improvement of network throughput in the MAC layer of an ad-hoc network [1]. Especially, the study on power control related to an ad-hoc equipment using limited energy and network throughput had been regarded as the main area of concern. Multi-hop information transmission from a network source area to a destination with optimal power through power control has advantage of reducing transmit power required in proportion to the 5th power of the square of propagation distance [2],[3]. In other words, the whole lifespan of a network can be extended as optimal transmit power is made to be consumed through a multi-hop instead of consuming transmit power with one node high. There had been study on an algorithm that did not exert effect on network performance, reducing unnecessary power consumption as a power control algorithm suggested before now controlled transmit power of a source node with minimum optimal transmit power that a node of the next hop was able to receive. A MAC protocol of a mobile ad-hoc network based on wireless LAN limits communication interference as it transmits communication of surrounding nodes that *
Funding for this paper was provided by Namseoul university.
influence transmission and reception with the RTS/CTS/DATA/ACK maximum transmit output to avoid transmission collision between nodes through physical carrier sense and to solve problems between hidden nodes and exposed nodes through virtual carrier sense. Especially, even if a protocol is planned considering only a communication radius in the majority of cases when a MAC layer protocol is planned, a career sense area becomes twice the size of a communication radius because delay threshold about a carrier sensing threshold is set up as double as shown in the figure 1 [4].
Fig. 1. Carrier Detection Threshold and Delay Threshold [8]
Therefore, interference between much more nodes occurs not by a real carrier sensing threshold area, but by a defer threshold area in multi-hop communication and loss in network performance to that extent is brought about. It means that transmission error can occur as the president signal works as interference in case of locating within a defer threshold area of transmit node even if surrounding node is outside the radius of a carrier sensing threshold.
Fig. 2. Power control and carrier sensing zone
The section that interference explained earlier exists is as shown in the figure 2. The line source based on node A shows the communication radius, a carrier detect threshold of A and the dotted line source drawn at the same center is a defer threshold area that the maximum signal of node A can operate as interference. Node C can receive power more than a carrier detect threshold and node D receives defer threshold power lower than a carrier threshold. However, as this signal has power
The Study on Spatial Reuse Efficiency of Network Nodes
537
more than defer threshold, node D defers media approach. Collision is blocked as collision window value was back-off and media approach is made delayed during time defined by EIFS when EIFS (Extended Inter-Frame Space) value is defined to transmit defer threshold for this defer in IEEE 802.11 wireless LAN. The use of optimal transmission and reception power through power control algorithm can enable the improvement of performance of network as it makes area that spatial reuse is possible by reducing a carrier spatial area explained earlier and interference areas of nodes. This study attempted to suggest an algorithm for reuse of power control spatial that improved network performance as communication can be possible at the same time without interference between node in communication and surrounding nodes by using a spatial reuse area calculated after a real carrier detect area was reduced with optimal transmission and reception power through a power control algorithm. This study was organized as follows. Chapter two, mentioned power control technique used in the existing wireless ad-hoc network and chapter three explained a power control algorithm for spatial reuse suggested in this study. Chapter 4 conducted simulation evaluation to judge the suitability of an algorithm proposed in this study and chapter 5 stated study findings.
2
Power Control Techniques in a Wireless Ad-Hoc Network
The transmitting power Pd was made to be estimated by a transmitter side for physical carrier detect between nodes in a system following RTS/CTS/DATA/ACK mechanism basically used in IEEE 802.11 wireless LAN as a transmitter side transmitted RTS with the maximum transmit power Pmax and a receiver side that received this with Pr power decided power level Pd that was able to keep desirable signal to noise ratio considering receiving power Pr and the present channel condition to contain it in CTS and to transmit with Pmax power. A node at a transmitter side uses transmit power of Pd on transmitting data. The reason that CTS is transmitted with Pmax is to solve collision between nodes through physical carrier detect. However, interference can occur in a transmission and reception node as power asymmetry with transmission and reception power is caused when a node that exists in a defer threshold area that the EIFS period is done transmits RTS with power of Pmax on transmitting data with transmit power of Pd. Distributed power control is a method that transmissible power level is divided into 10 levels and power level that can reduce power consumption is selected as each node is made to manages the optimal transmit power level with surrounding nodes with a table. However, interference due to power asymmetry can be caused by a node that exists in a defer threshold area on transmitting link data like the basic power control [6]. In case of PCM, the possibility of interference caused by power asymmetry is decreased as data is periodically transmitted into a basic power control method with transmit power of Pmax to solve an asymmetry problem of transmit power shown in basic and distributed power control. This cycle is a smaller value than EIFS and nodes that existed in a defer threshold area keep backing off for communication time through a EIFS value [7]. It becomes a main cause that lowers network efficiency in a topology with high network node density as it stops nodal communication within a
538
S.-d. Lee
carrier detect area, caused by communication with static power even if power loss can be decreased as data is transmitted and received by decreasing transmit power as much as Pd in Pmax. F-PCM is an algorithm to solve a problem that brings about interference in a node at the receiver side as nodes that exist in a carrier detect area at the receiver side do not detect corresponding signals and then transmits RTS with Pmax because data is periodically transmitted with transmit power Pmax on transmitting data in PCM. This algorithm is divided into the size of data that can be transmitted at the shorter time than a EIFS period. As ACK is transmitted to each with the maximum transmit power Pmax, it makes nodes that exist in a carrier detect area at a receiver side back-off [8]. However, it is necessary to expand EIFS as much as a overhead is enlarged and transmit time can be covered because this algorithm divides data into many pieces to transmit if using EIFS value. Even if this algorithm makes nodes that exist in all the carrier detect areas not interfere communication of a transmission and reception node, much more power is consumed than PCM. Nodes within a carrier detect area can have a problem that specific nodes do not approach media for a long time due to periodic use of the maximum transmit power like PCM even if the channel is empty according to result of a carrier detect.
3
Power Control Algorithm for Space Reuse
No concern had been paid to network capacity loss of nodes caused by interference of surrounding nodes as a power control algorithm had been planned based on efficiency of power consumption on a mobile ad-ho network up to the present. However, use of optimal transmission and reception power through a power control algorithm reduces a carrier detect area and interference spatial of surrounding nodes to enable to make a spatial re-usable area. His chapter explained an algorithm that was able to improve network performance through spatial reuse in power control. The number of nodes that receive interference becomes smaller because a carrier detect area becomes smaller that mush in case that optimal transmit power Pd is acquired to transmit and receive through exchange of RTS/CTS. Surrounding nodes that are not interfered when real data is transmitted comes to maintain the resting state as the existing power control algorithm delays surrounding nodes using the maximum transmit power periodically to avoid interference caused by a problem of power asymmetry between surrounding nodes and transmission/reception nodes. A suggested algorithm suggested a plan that had a surrounding node enable to use a potential reuse spatial generated by power control as a surrounding node solves an asymmetry problem of power not to interfere communication of nodes at the transmitter- receiver side. A potential space on a network came to be consumed as much as a range area of the figure 3 as node interference occurred in a ( 4π + 3 3 / 2)( Pmax`s range ) 2 m 2 area related to a carrier detect area of Pmax even if a real carrier detect area was (4π + 3 3 / 2)( Pd `s range ) 2 m 2 when it was communicated with Pd's transmit power on data transmission as shown in the figure 3, as a node at the transmitter side transmitted RTS with transmit power Pmax and the receiver side that receives this transmitted with Pmax's power after deciding power level Pd to contain it in CTS.
The Study on Spatial Reuse Efficiency of Network Nodes
539
Fig. 3. Ideal carrier sensing zone
In an algorithm for spatial reuse based power control using potential spatial on this network, node A and B transmitted SRS(Spatial Reuse to Send) including length of data that would be exchanged and information of optimal power level Pd after a transmit node A in the figure 4 transmitted RTS and a receiver side that received this with Pr's power decided the power level Pd that was able to keep a desirable signal to noise ratio, considering a channel condition to transmit with CTS. Edge node C and D in a carrier detect threshold area that received this transmitted RPT that included power information of PSR (Pmax/2) and information about data length that would be exchanged of node A and B after an EIFS period as optimal power PSR (Pmax/2) in a spatial reuse area calculate based on Pd. Node E and F that do not exist in a carrier detect threshold and received RPT was able to improve performance on the whole network through spatial reuse as it exchanged information through transmission/reception power of PSR and RTS/CTS/ DATA /ACK for a data exchange period of node A and B. PSR had to be smaller than Pmax/2 to avoid interference of transmission/reception node A and B so that s surrounding node communicates without interference through spatial reuse. A node that receives RPT through PSR transmit power was able to reuse 95% spatial of ( 4π + 3 3 / 2)( Pmax`s range − Pd `s range) 2 m 2 as shown in the figure 5 as it communicated with transmission/reception power of PSR.
4
Simulation Performance Evaluation
A simulation for performance evaluation used NS-2 allinone-2.29. A parameter set up on simulation was set up as the value to calculate propagation distance at a two-ray model from frequency to loss factor as shown in the table 1. The received threshold(Rx Threshold) shows threshold received power to detect a idle state of a channel and the edge threshold is threshold received power to distinguish the edge within a communication radius from a suggested algorithm. The set-up of the present simulation was able to obtain expected effect only when the more proper difference had to be made the smaller a node density became even if it was set-up without any big difference with received threshold.
540
S.-d. Lee
Fig. 4. SR-PCM algorithm
Fig. 5. Spatial reuse zone
A directional antenna was used for an antenna and transmission speed was set up as 2Mbps. For the transmit power level, transmit power was divided into 10 levels to be used to have a transmit radius from the minimum 25m to the maximum 250m based on the value in the table 1. A simulation compared increased data throughput with PCM to evaluate performance as two nodes located at the 1/2 distance of a carrier detect area added a CBR traffic to surrounding nodes that existed at a grid in a carrier detect area generated while exchanging data.
The Study on Spatial Reuse Efficiency of Network Nodes
541
Table 1. Simulation Parameter Items Frequency Propagation Tx Ant. Gain Rx Ant. Gain Loss factor Rx threshold CS threshold Traffic Routing protocol Data size
Value 5.8GHz Two-ray Model Omni-Antenna 1 1 1 1.295999e-9W CBR/UDP DSR 210 bytes
(a) CBR traffic 2
(b) CBR traffic 3
(c) CBR traffic 4
(d) CBR traffic 5
Fig. 6. Throughput of CBR Traffic
Result measured as two nodes added more than two CBR traffics in a carrier detect area communicating shows that the whole data throughput was improved as shown in the figure 6. However, it appeared that the initial delay was had at the early of transmission compared with an existing PCM according to the additional treatment process for spatial reuse. The fact that data throughput was not increased at the multiple as much as a CBR traffic occurred due to packet loss transmitted through ARP message, routing table update, and multi-hop. The figure 7 shows result about improvement limit of throughput and CBR traffic crossing two existing nodes. There was little increase effect of throughput from 6 traffics in this scenario as shown in the figure 7(a). This shows that a carrier detect area that spatial reuse was available was almost used. This showed limitation that was able to bring limited performance improvement even if a node density was high and there were many traffics that intended to transmit. However, it was identified that performance improvement was brought through a multi-hop in a spatial reuse area even in CBR traffic of a surrounding node crossing an existing node like (b).
542
S.-d. Lee
(a) CBR traffic 5,6
(b) Cross CBR traffic
Fig. 7. Throughput of CBR Traffic
5
Conclusion
There had been lots of studies on efficient multi-hop transmission and an power control algorithm to reduce power loss in an ad-hoc network. This power control algorithm reduces a carrier detect area dn interference areas of surrounding nodes to make a spatial reusable area and data throughput on the whole network can be improved by using this space. This study introduced a power control algorithm that spatial reuse was available and identified performance improvement through simulation as spatial reuse power calculated base on power of a transmission and reception node was transmitted to surrounding nodes. As a result, throughput per 01. second of network was improved for 11.1% to the maximum 51% according to the traffic distribution. A power control algorithm for spatial reuse suggested in this study conducted evaluation focusing on principles of an algorithm based on a ideal case. Further study would be needed on route set-up of routing class for multi-hop of nodes that existed within a spatial reuse area and a limited algorithm as well as on solution related to a part assumed to evaluate principles of a further algorithm and an algorithm to select the outermost transfer mode to suggest a more improved algorithm.
References 1. Siva Ram Mupthy, C., Manoi, B.S.: Ad Hoc Wireless Network Architectures and protocols. Prentice Hall (2004) 2. Abramson, N.: The ALOHA System Another Alternative for Computer Communications. In: Proc. AFIPS Fall Joint Computer Conference, pp. 281–285 (1970) 3. Santi, P.: Topology Control in Wireless Ad Hoc and Sensor Networks. John Wiley& Sons (2005) 4. Kamerman, A., Monteban, L.: WaveLan-II: A High-Performance Wireless LAN for Unlicensed Band. Bell Labs Technical Journal 2(3) (1997) 5. Wireless, L.A.N.: Medium Access Control (MAC) and Physical Layer (PHY) Specifications. IEEE Standard 802.11 (2007)
The Study on Spatial Reuse Efficiency of Network Nodes
543
6. Agarwal, S., Krishnamurthy, S., Katz, R.H., Dao, S.K.: Distributed Power Control in Ad Hoc Wireless Networks. In: Proc. IEEE PIMRC, pp. 56–66 (2001) 7. Jung, E.-S., Vaidya, N.H.: A Power Control MAC Protocol for Ad Hoc Networks. Springer Science Business Media Wireless Network (2005) 8. Kim, D.: F-PCM: A Fragmentation Based Power Control MAC Protocol for IEEE 802.11 Mobile Ad Hoc Network. Wiley Inter-Science, Wireless Communications and Mobile Computing 6, 727–739 (2006)
Agents for Cloud Resource Allocation: An Amazon EC2 Case Study J. Octavio Gutierrez-Garcia and Kwang Mong Sim Gwangju Institute of Science and Technology, Gwangju 500-712, Republic of Korea [email protected], [email protected]
Abstract. Infrastructure-as-a-service consumers are presented with numerous Cloud providers with a wide variety of resources. However, consumers are faced with providers that may offer (even similar) resources at different hourly cost rates, and also that no single provider may have matching resource capabilities to fulfill a highly heterogeneous set of requirements. This work proposes an agent-based approach endowed with the well-known contract net protocol for allocating heterogeneous resources from multiple Cloud providers while selecting the most economical resources. The contributions of this paper are: (i) devising an agent-based architecture for resource allocation in multiCloud environments, and (ii) implementing the agent-based Cloud resource allocation mechanism in commercial Clouds using Amazon EC2 as a case study. The Amazon EC2 case study shows that agents can autonomously select and allocate heterogeneous resources from multiple Cloud providers while dynamically sampling resources’ allocation cost for selecting the most economical resources. Keywords: agent-based Cloud computing, Cloud computing, multi-agent systems, resource allocation.
Agents for Cloud Resource Allocation: An Amazon EC2 Case Study
545
foundation for an agent-based Cloud resource allocation mechanism to autonomously allocate Cloud resources in multi-Cloud environments. Finally, an agent-based system prototype was devised and implemented in which a case study using Amazon EC2 as Cloud provider was carried out. The significance and novelty of this paper is that, to the best of the authors’ knowledge, it is the earliest work in applying, developing, and deploying an agentbased approach to allocate Cloud resources from multiple and commercial Cloud providers (e.g., Amazon EC2) in an autonomous manner. The contributions of this work are as follows. (i) Devising an agent-based Cloud architecture for resource allocation in multi-Cloud environments (Section 2), and (ii) implementing and deploying the agent-based resource allocation mechanism in commercial Clouds using Amazon EC2 as a case study (Section 3). In addition, Section 4 includes a related work comparison, and Section 5 presents conclusion and future work.
2
An Agent-Based Cloud Resource Allocation Architecture
The components of the agent-based architecture that supports resource allocation in multi-Cloud environments (Fig. 1) are as follows:
1) A Cloud resource ontology (Fig. 2) is a formal specification of Cloud resources. Cloud resources are defined by their functional and non-functional capabilities as well as their location (URI address). Non-functional capabilities are (i) storage capacity, (ii) memory, (iii) architecture type, and (iv) processing capacity. Functional capabilities are (i) pre-installed operative system, and () pre-installed applications, e.g., database servers.
Fig. 2. Cloud resource ontology
2) A Cloud directory is a listing of Cloud participants either brokers or service providers and their capabilities. The capability of brokers is stated as providing resource allocation services to Cloud consumers. The capabilities of service providers are defined by the Cloud resource types they offer. The Cloud directory is handled by a system agent (see Section 3 for details) to which (i) service provider agents and broker agents register their services, and (ii) consumer agents and broker agents contact to request for broker and service provider agents’ addresses, respectively. 3) Consumer agents (CAs) act on behalf of Cloud consumers and are in charge of submitting resource allocation requests composed of a set of Cloud requirements. To do this, CAs adopt the initiator (manager) role of the CNP consisting of (i) sending a call-for-proposals (resource allocation request) to q broker agents (contractors). Then, from the p (p ≤ q) broker agents’ proposals, CAs select the best (e.g., cheapest) proposal and send an accept message to the broker agent with the winning bid, and reject messages to the remaining broker agents. 4) Broker agents (BAs) provide Cloud resource allocation services to CAs. To do this, BAs adopt the participant (contractor) role of the CNP consisting of (i) receiving resource allocation requests (call-for-proposals) from CAs, and (ii) replying with proposals based on the current hourly cost rates associated to the Cloud resources. If a BA is selected, the BA allocates the Cloud resources from service providers listed in the Cloud directory with matching resource capabilities. To do this, BAs adopt the initiator role of the CNP with service provider agents as participants. BAs’ call-forproposals are composed of single Cloud requirements. In doing so, CAs’ allocation requests consisting of multiple Cloud requirements are handled by BAs as independent and parallel resource allocations from the best (cheapest) service providers for each Cloud requirement. 5) Service provider agents (SPAs) offer and supply Cloud resources to BAs by adopting the participant role of the CNP with BAs as initiators. SPAs’ proposals consist of allocation costs based on the Cloud resource types. In addition, SPAs
Agents for Cloud Resource Allocation: An Amazon EC2 Case Study
547
handle Cloud resource provisioning by requesting (de)allocations of Cloud resources to resource agents. 6) Resource agents (RAs) (de)allocate and monitor Cloud resources by using specific vendor APIs (usually web services), e.g., Amazon AWS APIs [1]. RAs are described by the Cloud resource type they can allocate. RAs’ functions are: (i) receiving allocation requests from SPAs, (ii) creating private/public keys to access Cloud resources, (iii) extracting and decrypting passwords to access Cloud resources, (iv) extracting public IP addresses from recently allocated Cloud resources, and (v) forwarding Cloud resources’ access information to SPAs. An agent-based Cloud resource allocation scenario (Fig. 3) is as follows. A CA retrieves BAs’ addresses from the Cloud directory, and then the CA sends a call-forproposals (consisting of a set of p Cloud consumer requirements) to a set of BAs. The BAs reply with proposals containing the overall allocation cost. Afterwards, the CA sends an accept-proposal message to the selected BA and reject-proposal messages to the remaining BAs. Subsequently, the selected BA adopts p CNP executions with SPAs (listed in the Cloud directory) as participants in a parallel manner, i.e., the
selected BA adopts a CNP execution for each Cloud resource requested by the CA. The SPAs (selected by the BA) send allocation requests to appropriate RAs (i.e., RAs matching the Cloud consumer requirements). Finally, the RAs extract the public IP addresses and passwords from the recently allocated Cloud resources, and forward the results to their SPAs that forward the information to the BA, which hand over the information to the CA.
3
Agent-Based Cloud Resource Allocation in Amazon EC2
A case study using Amazon EC2 was carried out for which an agent-based system prototype was implemented using (i) the java agent development framework (JADE) [6], (ii) Bouncy Castle Crypto APIs [7] (for encrypting and decrypting Cloud resources’ passwords), and (iii) Amazon SDK for Java [5]. 3.1
Amazon EC2 Cloud Ontology
The Cloud resource ontology (Fig. 2) was provided with Cloud resource definitions based on (i) Amazon instance types and (ii) Amazon machine images (AMIs), for a total of 58 different Cloud resource definitions that resulted from the valid combinations (i.e., not all the AMIs can be run on a given instance type) between instance types and AMIs (Table 1). Table 1. Amazon instance types and Amazon machine images Amazon instance types (i) t1.micro, (ii) m1.small, (iii) m1.large, (iv) m1.xlarge, (v) m2.xlarge, (vi) m2.2xlarge, (vii) m2.4xlarge, (viii) c1.medium, (ix) c1.xlarge, (x) cc1.4xlarge, and (xi) cg1.4xlarge.
3.2
Amazon machine images (i) Basic 32-bit Amazon Linux, (ii) Basic 64-bit Amazon Linux, (iii) Red Hat Enterprise Linux 6.1 32 bit, (iv) Red Hat Enterprise Linux 6.1 64 bit, (v) SUSE Linux Enterprise Server 11 64-bit, (vi) SUSE Linux Enterprise Server 11 32bit, (vii) Microsoft Windows Server 2008 Base, (viii) Microsoft Windows Server 2008 R2 Base, (ix) Microsoft Windows Server 2008 R2 with SQL Server Express and IIS, (x) Microsoft Windows Server 2008 R2 with SQL Server Standard, (xi) Cluster Instances Amazon Linux, and (xii) Cluster Instances HVM SUSE Linux Enterprise 11.
Cloud Participants and Distributed Cloud Environment
The agents involved in the case study were 1 RA, 5 BAs, 5 SPAs, and 2500 RAs. Each agent, either CAs, BAs, or SPAs was deployed on a different JADE agent container (see Fig. 1 and Fig. 4), i.e., an instance of a JADE runtime environment. In addition, since RAs do not interact among themselves, all the RAs were deployed on a single container (Container-1 in Fig. 4). In doing so, SPAs had to contact RAs located at a remote location, and an unnecessary large number of containers was avoided in the system prototype. The Cloud resource type of the RAs was randomly
Agents for Cloud Resource Allocation: An Amazon EC2 Case Study
549
selected from the available 58 Cloud resource types (Table 1). Moreover, all the RAs were randomly assigned to the SPAs to simulate a scenario with highly heterogeneous Cloud providers. All the agent containers must be and were registered in a main JADE container that manages and supports the agent-based platform by (i) handling asynchronous message passing communication through Java RMI and IIOP, (ii) starting and killing agents, and (iii) providing services such as: a directory facilitator agent (Cloud directory), a sniffer agent, a remote management agent, etc., see [6] for details of JADE.
Fig. 6. AWS management console – My instances option
As shown in Fig. 4, agent-based platform CloudMAS had: (i) 1 Main-Container including a remote monitoring agent (RMA@CloudMAS), an agent management system (ams@CloudMAS), a directory facilitator (df@CloudMAS), i.e., a Cloud directory, and a sniffer agent (for illustrative purposes); and (ii) 12 basic containers (from Container-1 to Container-12). Container-1 included all the RAs. The SPAs and BAs were included in agent containers ranging from Container-2 to Container-11, one agent for each container. Finally, the CA was included in Container-12. Since all the agent containers were deployed on the same host, each container was provided with a different network port to simulate a fully distributed environment. 3.3
Cloud Resource Allocation Scenario
The CA was provided with a Cloud resource allocation request composed of 6 Cloud resources: 4 m1.small instances with an AMI ami-8c1fece5 (Basic 32-bit Amazon Linux AMI 2011.02.1 Beta) and 2 m1.large instances with an AMI ami-8e1fece7 (Basic 64-bit Amazon Linux AMI 2011.02.1 Beta).
Agents for Cloud Resource Allocation: An Amazon EC2 Case Study
551
Fig. 7. Console output for agent-based resource allocations using Amazon SDK
The CA submitted the allocation request to the 5 BAs by using the CNP. The selected BA executed 6 CNP (one for each Cloud resource to be allocated) with the 5 SPAs in a parallel manner. Finally, the selected SPAs requested the Cloud resource allocations to their RAs. Fig. 4 shows an extract of the messages exchanged among all the agents to carry out the Cloud resource allocation request, which was fulfilled by agent BA4 (a BA selected by consumer CA1). The messages received by agent BA4 (Fig. 4) came from the SPAs bidding for allocating a Cloud resource and/or providing data to access the recently allocated Cloud resources, e.g., the messages exchanged between agents SPA1 and BA4, see Fig. 4. In addition, as soon as allocation data (public IP address and password to access a given Cloud resource) was received, broker BA4 forwarded the data to consumer CA1, as shown in the bottom of Fig. 4. The interleaving of messages received by agent BA4 from all the SPAs (see Fig. 4) is the result of the parallel execution of CNPs for allocating Cloud resources. 3.4
Technical Aspects to Handle Cloud Resource Allocation in Amazon EC2
The RAs were provided with (i) Amazon EC2 API tools to handle Cloud resource allocations, and (ii) Amazon AWS security credentials to access Amazon EC2. It should be noted that although the RAs shared the same security credentials (i.e., all the RAs accessed Amazon EC2 using the same Amazon AWS account), sharing the credentials had no advantageous effects on the agent-based Cloud resource allocation approach.
552
J.O. Gutierrez-Garcia and K.M. Sim
When the RAs received the SPAs’ requests to allocate Cloud resources, the RAs created new RSA key pairs to access Amazon EC2 (Fig. 5). The key pairs were automatically named based on the identifiers of the RAs that allocated the Cloud resources, e.g., newKeyRA2462 (see the left side of Fig. 5). Right afterwards, the RAs proceeded to allocate the Cloud resources (Fig. 6) corresponding to the CA’s initial allocation request (consisting of 6 Cloud resources, see Section 3.3). The console output of the agent-based system (Fig. 7) corresponding to the CA’s allocation request shows: (i) JADE initialization messages displaying agent containers’ addresses and names, (ii) self-generated output messages derived from the creation of key pairs by using Amazon SDK, and (iii) self-generated output messages derived from the Amazon instance allocations by using Amazon SDK. In general, the self-generated output messages contained the following information: timestamp, key pair name, AWS access key, type of instance allocated, etc., see Fig. 7 for details. Since Amazon instances take some time to be fully functional (i.e., to start) and the delay time may vary due to the size of AMIs, number of instances to be allocated, among other factors [2], the RAs were continuously checking (every 200 s) whether Amazon instances were up and running by retrieving the console output of the recently allocated instances as indication of the start of the instances. Once the RAs detected an output in the instances’ console, the RAs proceeded to extract the public IP addresses and passwords (only possible when the instances are up and running), which were forwarded to their corresponding SPAs.
4
Related Work
Resource allocation mechanisms have been widely investigated (see [12]). However, little attention has been directed to (i) Cloud resource allocation in multi-Cloud environments, and (ii) to actual implementations of autonomous Cloud resource allocation mechanisms. Whereas current Cloud management systems (see [8], [10], and [14]) may allocate Cloud resources from different Clouds to execute consumers’ applications, no explicit consideration of autonomous Cloud resource selection based on fees associated to Cloud resources has been made. In contrast, this present work uses both the agent paradigm and the CNP to (i) sample Cloud resources’ hourly cost rates, and (ii) allocate Cloud resources in multi-Cloud environments in an autonomous and dynamic manner. In addition, the proposed agent-based Cloud resource allocation mechanism is fully distributed, in contrast to centralized allocation mechanisms (see [4]) that require a central control entity (allocator) that commonly becomes a system bottleneck.
5
Conclusion and Future Work
The contributions of this paper are as follows. (i) Devising the earliest (to the best of the authors’ knowledge) agent-based Cloud architecture for resource allocation in multiCloud environments, and (ii) implementing and deploying the agent-based resource allocation mechanism in commercial Clouds using Amazon EC2 as a case study.
Agents for Cloud Resource Allocation: An Amazon EC2 Case Study
553
In this work, autonomous agents equipped with the CNP to (i) dynamically sample hourly cost rates and (ii) support cost-based Cloud resource allocation among selfinterested Cloud participants were used to deal with Cloud resource allocation in multi-Cloud environments. By using the agent paradigm, Cloud consumers can efficiently (i.e., with the lowest allocation costs) allocate heterogeneous sets of Cloud resources from multiple, distributed Cloud providers in a dynamic and autonomous manner as shown in the Amazon EC2 case study. Since this work provides the foundations for a general-purpose agent-based multiCloud platform by providing an infrastructure-as-a-service solution (allocated from multiple Cloud providers) to Cloud consumers, future research directions include: (i) adding agent capabilities to schedule and execute both workflows and bag-of-tasks applications in multi-Cloud environments, and (ii) implementing access to more commercial Cloud providers, such as: GoGrid [9] and RackSpace [11]. Acknowledgments. This work was supported by the Korea Research Foundation Grant funded by the Korean Government (MEST) (KRF-2009-220-D00092) and the DASAN International Faculty Fund (project code: 140316).
References 1. 2. 3. 4. 5. 6.
7. 8.
9. 10.
11. 12. 13. 14.
Amazon EC2 API Tools, http://aws.amazon.com/developertools/351 Amazon EC2 FAQs, http://aws.amazon.com/ec2/faqs Amazon Elastic Compute Cloud (Amazon EC2), http://aws.amazon.com/ec2 Asmild, M., Paradi, J.C., Pastor, J.T.: Centralized Resource Allocation BCC Models. Omega 37(1), 40–49 (2009) AWS SDK for Java – A Java Library for Amazon S3, Amazon EC2, and More, http://aws.amazon.com/sdkforjava Bellifemine, F., Poggi, A., Rimassa, G.: JADE - A FIPA-Compliant Agent Framework. In: 4th International Conference and Exhibition on the Practical Application of Intelligent Agents and Multi-Agents, pp. 97–108 (1999) Bouncy Castle Crypto APIs, http://www.bouncycastle.org Buyya, R., Pandey, S., Vecchiola, C.: Cloudbus Toolkit for Market-Oriented Cloud Computing. In: Jaatun, M.G., Zhao, G., Rong, C. (eds.) Cloud Computing. LNCS, vol. 5931, pp. 24–44. Springer, Heidelberg (2009) GoGrid, http://www.gogrid.com Lee, K., Paton, N.W., Sakellariou, R., Deelman, E., Fernandes, A.A.A., Metha, G.: Adaptive Workflow Processing and Execution in Pegasus. Concurr. Comput.: Pract. Exper. 21(16), 1965–1981 (2009) RackSpace, http://www.rackspace.com Sim, K.M.: A Survey of Bargaining Models for Grid Resource Allocation. SIGecom Exch. 5(5), 22–32 (2006) Smith, R.G.: The Contract Net Protocol: High-Level Communication and Control in a Distributed Problem Solver. IEEE Trans. Comput. 29(12), 1104–1113 (1980) Yang, Y., Liu, K., Chen, J., Lignier, J., Jin, H.: Peer-to-peer Based Grid Workflow Runtime Environment of SwinDeW-G. In: 3rd IEEE International Conference on e-Science and Grid Computing, pp. 51–58. IEEE Computer Society, Washington (2007)
An Offline Quality Management Strategy to Improve Online Quality of Service for Next Generation Networks Calduwel Newton Pitchai1,*, Arockiam Lawrence2,**, and Gang-soo Lee3,*** 1
Computer Science, Bishop Heber College (Autonomous) Computer Science, St.Joseph’s College (Autonomous) Tiruchirappalli, India 3 Dept. of Computer Engineering, 133 Ojeong-dong Daedeok-gu Daejeon, Korea [email protected]
2
Abstract. The Next Generation Networks (NGN) create various challenges to deal with network management operations, like fault management, service restoration, performance management, traffic management, data gathering and analyses, and routine network tests. As the network is growing exponentially, the amount of data that has to be stored to improve Quality of Service (QoS) are also high. In other words, data about the past history has to be stored effectively in order to deal with above challenges. If there is no effective strategy to overcome these challenges then the NGN or the upcoming networks could not perform well. This will result in poor QoS. The word ‘Offline’ refers already stored data or past data. The word ‘Online’ refers the QoS for the current requirements. That is, a strategy is proposed to improve current QoS requirements by analyzing the past data. The objectives of this paper are to appropriately monitor or manage the above operations, effectively gather and analyze the voluminous of data, efficiently identify the reason behind the problem and successfully take remedial measures to improve QoS. Ultimately, the outcome of this paper will improve the performance of the above network management operations. Keywords: QoS, Next Generation Networks, Traffic, Network Management, Offline, Online.
1
Introduction
The upcoming networks are witnessing a drive to the convergence of wired and wireless networks. The challenges in the converged networks are increasing day by day. Improving or ensuring Quality of Service (QoS) is a great challenge for the upcoming networks as they grow rapidly. QoS is the ability to deliver network services according to the parameters specified in a service level agreement. Quality is *
Assistant Professor. Associate Professor. *** Professor. **
An Offline Quality Management Strategy to Improve Online QoS for NGN
555
characterized by service availability, delay, jitter, and throughput and packet loss ratio. At a network resource level, QoS refers to a set of capabilities that allow a service provider to prioritize traffic, control bandwidth, and network latency [1]. Frequently, quality is referred as customer satisfaction, first class, good quality rather than quantity, etc. The term QoS has become a very dominant and attractive notion in many aspects of our daily life. Since communication networks become a very essential part of our life, many efforts have been made towards improving quality of service in order to achieve more and more customer satisfaction which leads to strong loyalty and therefore more profit to the service providers and to achieve global efficiency in resource utilization. Moreover, recent advances in high speed networking technology have created opportunities for the development of multimedia applications. These applications integrate several media such as text, graphics, audio, video, image, etc. By having algorithms for QoS may not guarantee the requirements, because some of the nodes will violate the service requirements. In other words, flow must be monitored and appropriate measures must be taken in order to ensure the Quality of Service. Ultimately, ensuring the QoS will improve the performance of the network. QoS management can be done in two ways. They are, QoS Assurance Strategy (QAS) and QoS Control Strategy (QCS). The QAS is used to prevent QoS problem. It will give 100% guarantee for the QoS. The QCS is used to control the QoS problem. It is a kind of detection and correction method. In other words, identify the QoS problem(s) in the previous services and then correct them to improve future QoS. In all the cases, QoS problem cannot be prevented. Hence, there is a great need for the QCS. This paper concentrates on the second method. The rest of the paper is organized as follows. Section 2 emphasizes on the motivations behind this paper. Section 3 describes the strategy to improve QoS in the future. Section 4 shows the analysis and results of our strategy. Section 5 focuses on the conclusion and future works. Finally, references are listed.
2
Motivation
QoS management is an important discipline that should be addressed to deliver good services and can be considered by two different viewpoints. One is the infrastructure and another one is the techniques and tools for service delivery and management. Service Quality Management (SQM) is a general term to address QoS management under the second viewpoint [2]. SQM functions can be classified in three categories: collection and consolidation; inference and control; and presentation [3]. Collection and consolidation functions interact with external systems to obtain the necessary data. Inference and control functions are the core of SQM. These are used to derive action plan to improve the QoS in the future. Presentation functions construct service level reports. In a connection-oriented network, all the packets belonging to a flow follow the same route; in a connectionless network, they may follow different routes. There is a great need to guarantee Quality of Service through network and protocol design [4]. QoS routing offers serious challenges even for today’s Internet [5]. Different service types (e.g., voice, live video, and document transfer) have significantly
556
C.N. Pitchai et al.
different objectives for delay, bandwidth, and packet loss. Determining the QoS capability of candidate links is not simple for such scenarios; for multicast services, the difficulties are even larger. More than one QoS constraint often make the QoS routing problem NP complete [6]. Suboptimal algorithms such as sequential filtering are often used, especially for large networks, where an optimal path based on a single primary metric (e.g., bandwidth) is elected first, and a subset of them are eliminated by optimizing over the secondary metric (e.g., delay), and so on, until all the metrics have been taken into account. A random selection is made if there is still more than one choice after considering the network throughput as the last metric. As long as the QoS constraints are satisfied, the same route is used for all packets in the flow. Once a route has been selected for a specific flow, the necessary resources, (bandwidth, buffer space in routers, etc.) must be reserved for the flow. These resources will not available to other flows until the end of this flow. Consequently, the amount of remaining network resources available to accommodate the QoS requests of other flows will have to be recalculated and propagated to all other pertinent nodes as part of the topology update information. Analysis of data from number of sources clearly shows that the internet traffic has been almost doubling every year since 1997 [7]. As of now, the Internet has only supported best effort service; best effort in the sense that it will do its best to transport the user packets to their intended destination, although without any guarantee [8]. IP-based networks were never designed for realtime traffic, yet QoS support in such networks is needed to accommodate both global use and the more demanding applications now emerging [9]. There are various schemes used to enhance QoS [10]. The above discussions elucidates that QoS problem is a serious issue that needs effective strategy to provide good or better services. This issue can not be prevented in all the situations. So, a strategy is very much needed to control QoS problem. In other words, previous flow characteristics must be analyzed, reason for the QoS problem must be identified and necessary steps must be taken to reduce or eliminate QoS problem in the future.
3
The Strategy: QCS
The QCS refers the Quality of Service Control Strategy. The upcoming and future networks expect very high QoS requirements and it continues to grow. Unless, there is a proper monitoring and measuring mechanism to analyze the past flows, the future networks will create complex problems. So an effective strategy to serve the future QoS requirements is very much needed. The needs of each flow can be characterized by four primary parameters: reliability, delay, jitter and bandwidth [4]. These parameters are considered as binary attributes. If expected requirement is met, and then it is 1, otherwise, it is 0. For example, if E-mail application utilized high reliability, low delay, low jitter and low bandwidth, then the corresponding values will be 1,1,1 and 1 [11, 12]. In other words, the QoS is met for E-mail application. If it has utilized low reliability, then the value will be 0. Similarly, In Videoconferencing, only the
An Offline Quality Management Strategy to Improve Online QoS for NGN
557
Table 1. Actual QoS requirements
Application E-mail File Transfer Web Access Remote login Audio Video Telephony Video Conference
Reliability 1 1 1 0 1 1 1 1
Delay
Jitter
1 1 1 0 1 1 1 0
1 1 1 0 1 1 1 0
Band width 1 1 0 0 1 0 1 0
Status 1 1 0 0 1 0 1 0
reliability is high and all other parameters are low. It shows the poor QoS. Table 1 shows the actual performance of various applications such as E-mail, File transfer, etc. The last column ‘Status’ shows the final outcome of the QoS. To monitor and measure QoS effectively, the ‘Status’ column has been included. It shows the final outcome of the data transfer that is either QoS met or not met. If it has a value 1 then QoS is met. Otherwise, QoS is not met. The Status column is computed by doing Logical AND with the values of all the parameters such as Reliability, Delay, Jitter and Bandwidth. Table 1 reveals that 50% of the applications received QoS and the remaining 50% has not received QoS. This strategy is definitely effective because all the values are in binary. The ‘Status’ column is needed to identify whether QoS is met or not. This column can be used to monitor all the flows and measure the status of QoS and take remedial steps to improve the QoS. This separates the applications that enjoy the QoS from those that starve QoS. As the number data transfers are extremely high in today’s network and exponentially growing, an effective strategy to monitor and measure the flow is needed. This strategy will ensure future QoS requirements to be met.
4
Analysis and Results
This section uses the QoS management functions to analyze QCS strategy and shows the results using graph. As mentioned in section 2, QoS management functions can be classified in three categories. First, the collection and consolidation functions get the data from different nodes after the data transfer is over and consolidates the data in the master table. This table can be stored in a server or gateway. Table 3 shows the outcome of first function. Second, the inference and control functions deduct, what applications could not receive QoS, what parameters are responsible for the QoS problem. The problem deductions can be done in two ways depending on the requirements. They are application-wise deduction and parameter-wise deduction. Apart from this, it is also possible to deduct the exact reason(s) for the QoS problem. For example, if reliability is the problem then the reason(s) may be an error deduction or error correction
558
C.N. Pitchai et al.
mechanism or a cable problem. Now, it possible to control the QoS problem as the exact reason(s) has been deducted. Third, the presentation functions deal with service level reports. They contain the reason (s) behind the problem and the necessary steps to be taken to reduce or eliminate the problem. The value of the ‘Status’ column is ‘1’ then there is no QoS problem. Otherwise, there is QoS problem. Table 2 shows the sample data for the application-wise deduction (E-mail). The value of third data transfer in Table 2 is 0,1,1,1. The value ‘0’ indicates that the third data transfer is affected by the reliability. Table 2. Application-wise Deduction ( E-mail Application )
Reliability 1 1 0 0 1 1 1 1
Delay
Jitter
1 1 1 0 1 1 1 0
1 1 1 0 1 1 1 0
Band width 1 1 1 0 1 0 1 0
Status 1 1 0 0 1 0 1 0
Table 3 shows sample data for parameter-wise deduction (Bandwidth). The value of first data transfer in Table 3 is 1,1,1,0,1,1,1,1. Here, the value ‘0’ indicates that the remote login application could not receive the required bandwidth. Wherever the value is zero, it indicates there is a QoS problem in the corresponding attribute (i.e. either application or parameter).
Remote Login
Audio
Video
1 1 1 0 1 1 1 0
0 1 0 0 1 0 1 0
1 1 0 1 1 0 1 1
1 1 1 1 1 0 1 0
1 1 0 1 1 1 1 0
Status
Web Access
1 1 1 0 1 1 1 0
Video Conference
File Transfer
1 1 1 0 1 1 1 1
Telephony
E-mail
Table 3. Parameter-wise Deduction ( Bandwidth )
1 1 0 1 1 0 1 1
0 1 0 0 1 0 1 0
An Offline Quality Management Strategy to Improve Online QoS for NGN
559
Performance (High / Low)
E-mail Application
0
1
2
3
4
5
6
7
8
9
No. of Data Transfe r
Fig. 1. QoS Status of E-mail Application
Perform an ce (H igh / L ow )
Bandwidth
E-mail
File Transfer
Web Access
Remote Login
Audio
Video
Telephony
Video Conf.
Fig. 2. QoS Status of Bandwidth for all Applications
Fig. 1 shows the QoS status of E-mail application. It shows for four data transfers the QoS is met and for the remaining it is not met. Fig. 2 shows the QoS status of bandwidth for all the applications. It shows only three applications (File Transfer, Audio, Telephony) have received QoS with respect to bandwidth. QoS for the remaining five applications are affected by bandwidth. It is also possible to store the status of a particular application for a particular parameter, instead of a particular application for all parameters (Table 2) and a particular parameter for all applications (Table 3). Figure 1 clearly shows the data transfers that had QoS problem. They are
560
C.N. Pitchai et al.
data transfer 3, 4, 6, and 8. The data transfer 3 had a QoS problem due to reliability. This problem is due to error deduction or error correction mechanism or a cable problem. So, measures can be taken to check the above mechanisms or a cable. Similarly, the reasons for the other data transfers (4, 6 & 8) can also be identified using Table 2.
5
Conclusion and Future Works
As the upcoming networks are unpredictably expanding day by day, the QoS assurance can not be given. So, the Quality of Service Control Strategy (QCS) is very much needed. The QCS uses the three major functions such as collection and consolidation; inference and control; and presentation. Section 3 describes these functions with the sample data. The analysis and the results are shown in Section 4. This paper proposed a strategy that contributes to appropriately monitor or manage the above operations, effectively gather and analyze the voluminous amount of data, efficiently identify the reason behind the problem and successfully take remedial measures to improve QoS. The following are the advantages of our strategy. They are, 1. QCS is easy to implement. 2. All the values are in binary which speeds up computation and needs less memory. 3. Only one column (i.e. Status) is enough to check whether QoS is met or not. It also helps to identify which parameter(s) is / are responsible for QoS problem. Though this paper has many advantages, it also has few weaknesses as well. The number of tables used may require more spaces and if more parameters and more applications are used care must be taken to reduce them. In future, this strategy could be implemented to experience the above advantages. Acknowledgments. This work was supported by the Security Engineering Research Center, granted by the Korea Ministry of Knowledge Economy.
References 1. Westerinen, A., Schnizlein, J., Strassner, J., Scherling, M., Quinn, B., Herzog, S., Huynh, A., Carlson, M., Perry, J., Waldbusser, S.: Terminology for Policy-Based Management. IETF RFC 3198 (2001) 2. Menawda, S., Rich, R., Hawley, D.: Service Quality Management: An Emerging Source of Competitive Advantage. The Yankee Group (2002) 3. Penna, M.C., Wandresen, R.R.: On-line Techniques for Service Quality Management. In: Proceedings of the First International Conference on Quality of Service in Heterogeneous Wired/Wireless Networks. IEEE CS (2004) 4. Tanenbaum, A.S.: Computer Networks, 4th edn., pp. 397–417. PHI (2005)
An Offline Quality Management Strategy to Improve Online QoS for NGN
561
5. Crawley, E., et al.: A Framework for QoS-Based Routing in the Internet. RFC 2386 (1998), http://www.ietf.org/rfc/rfc.2384.txt 6. Chen, S.: Routing Support For Providing Guaranteed End-To-End Quality-Of-Service. Ph.D. thesis, Univ. of IL at Urbana-Champaign (1999), http://cairo.cs.uiuc.edu/papers/SCthesis.ps 7. Gupta, G.K.: Introduction to Data Mining with Case Studies, pp. 218–244. PHI (2006) 8. Chakrabarti, S., Mishra, A.: QoS Issues in Ad Hoc Wireless Networks. IEEE Communications Magazine, 142–148 (2001) 9. Bhatti, S.N., Crowcroft, J.: QoS-Sensitive Flows: Issues in IP Packet Handling. IEEE Internet Computing, 48–57 (2000) 10. Calduwel Newton, P., Arockiam, L.: Route Optimization Mechanisms for Internet Applications in Mobile Networks: A Survey. International Journal of Future Generation Communication and Networking 3(3), 57–69 (2010) 11. Calduwel Newton, P.: A Contemporary Technique to Guarantee Quality of Service (QoS) for Heterogeneous Data Traffic. In: Proceedings of the International Conference on Information Security and Assurance. IEEE CS, pp. 210–213 (2008) 12. Calduwel Newton, P., Arockiam, L.: A Novel Prediction Technique to Improve Quality of Service (QoS) for Heterogeneous Data Traffic. Journal of Intelligent Manufacturing (2009)
Hegemonic Power and Technology Advancement O.F. Bahrisch and Jin-Suk Kim Namseoul University, Department of International Management {Bahrisch,mktjskim}@gmail.com
Abstract. During the past 600 years the world has seen changing hegemonic powers, situations in which one country dominates the world economically, intellectually and militarily. Since 1945 this attribution is undoubtedly true for the USA. A hegemonic change always occurred after a period of global war. Is the current ‘global war of terror’ such a global war that might cause a new hegemon to emerge? This paper summarizes the existing literature on the debate over the emergence of a new hegemon and analyses two possible scenarios a) the USA stay the dominating power, b) a new hegemon arises and one of the aspirants of the hegemonic throne, take over the leadership position.
1
Theory of Hegemony
In order to understand the present day hegemonic situation in helps to draw on the field of history. Modern European history (the period after the Middle Ages beginning at the end of the 15th century) has seen four full hegemonic cycles so far, with the fifth cycle still in progress. The first superpower in modern history was Portugal. Growing out of the middle ages, a period of decline for almost one thousand years, Portugal was a the first nation that spearheaded the Age of Discovery and dominated the world from about 1494 on. Portuguese explores, like Vasco da Gama, opened trade routes to Africa and India and established trading posts protected by a strong navy. Portuguese dominance lasted until 1580, when The Netherlands took over the hegemonic throne. The Netherlands profited form mass immigration of highly skilled craftsmen and rich merchants. Protestants from Bruges, Ghent and Antwerp (main trading ports at that time) where forced out after Spain conquered the cities, and settled in Amsterdam. From Portugal and Spain arrived Sephardi Jews that fled from religious prosecution. A highly skilled workforce and a strong financial endowment combined with a cheap source of energy (the windmill) and efficient transportation (canals) brought the Dutch Golden Age. The sawmill enabled dutch shipbuilders to construct new ships in unprecedented numbers that where used for trading and military use. The purpose of the dutch navy was mainly to protect shipping lanes and securing uninterrupted trade. An impressive example of the dutch naval might was given in 1607, when the Dutch navy attacked the Spanish fleet in Gibraltar wiping it out within four hours. The Dutch opened up Japan for trade, secured trade with the
English colonies in North America, and dominated trade between European countries. Dutch domination lasted for about 110 years until 1690. A serious of wars and internal disputes weakened Holland to the advantage of two of its neighbors: France and England. France aimed to control Continental Europe while England focused on control of the seas. Holland was not strong enough to keep its colonies in North America and lost New Amsterdam (which is called New York from 1667 on) to England. In relative safety, protected by the English Channel, England grew in the slipstream of Holland's colonial efforts into the British Empire with control over India and North America. Until about 1792 Britain was successfully defending its position as the hegemon. In 1776 the Thirteen Colonies in North America declared its independence and drew Britain in the costly American War of Independence. The power of Britain in decline, France was able to regain its might and started to control the continent. Driven by an ambitious emperor, Napoleon, and his Grande Armée France was able to influence the power of balance to its advantage. But Britain was not finished. Fueled by the Industrial Revolution, Britain became the first industrialized country and was able to maintain its fleet to control its shipping lanes and most of its colonies. Between 1792 and 1914 Britain shared the status of a superpower with France. World War I and II changed the power balance a fifth time and left the US as the only aspirant of the hegemonic throne with an intact economy. As the Omaha WorldHerald aptly noted in 1986 "There were 20 million Russians dead, Europe was flattened, Asia was flattened. The reason why we were so fantastic as exporters for 30 years was because there wasn't anybody else. It's like having the Chicago Bears play against a Brownie troop and then being proud of the score we've run up."(Gauger, 1986, p. 43). 1.1
Realist School
The study of hegemony has produced four main theoretical directions that can be divided in 2 schools of thought. The first school of thought is the Realist School with its two theoretical approaches to hegemony: the Theory of Hegemonic Stability and the Power Transition Theory. The intellectual roots of the Realist School reaching back in the 16th century with the seminal book of Niccolo Machiavelli "The Prince". In the first sentence of the first chapter Machiavelli unveils the ultimate source of power: the state (Machiavelli, 1961, p. 8). Following that logic there can be no actor above the state and states are in free competition with each other with no referee defining or enforcing any rules. Therefore, from a Realist perspective the world is made up of antagonists (states) with the ultimate goal to enhance their position, if necessary by any means, under the condition of national survival. Therefore the focus of analysis of the Realist School is politics, especially politics between states that can be enforced be a strong military. The main idea of Hegemonic Stability Theory is that the best possible condition for economic prosperity, peace and stability only exists when it is enforced by a strong hegemon. One single dominant state dictates the rules of the game and punishes deviators. One reason for the superiority of a system with one strong a hegemon over
564
O.F. Bahrisch and J.-S. Kim
a system with numerous weak but equal countries is the successful provision of a solution for the public good problem (see Kindleberger, 1986 p. 205). This gametheoretic approach states that order in an international system can only be provided at certain cost. In an system with many small but weak states, no state would contribute to this cost. Free-riding would be the rational pattern of behavior. Without a strong hegemon no country could be excluded from using an orderly system and no country could be forced to contribute to this order. If everybody is using and nobody maintains, the system will ultimately collapse and fail. For example, international trade depends much on sea transportation. If international shipping lanes would be threatened by pirates (like off the Somali coast nowadays) a strong navy is required to secure sea bound trade. Every trading nation would have an advantage of pirate-free waters but no country would be willing to bear the cost to send its own navy. Only a strong hegemon could enforce order and distribute the costs. Interestingly, three of the main schoolers of Hegemonic Stability Theory argued that the U.S. lost its hegemonic status at the end of the 1960s (Kindleberger, 1981), in the mid 1970s(Gilpin, 1975), or are a declining hegemonic position since the 1970s (Webb & Krasner, 1989). Power Transition Theory, initiated by A.F.K. Organski (Organski, 1958), deals, unlike the name might suggest, less with hegemonic succession but "... is mainly concerned with the explanation of the onset international conflict" (Clark, 2011, p. 14). Power Transition Theory models the international system of states into a set of hierarchies with five degrees of power (Kugler & Organski, 1989, p. 174). The five degrees of power distinguish a dominant nation, great powers, middle powers, small powers and colonies. The first place and most powerful country is the dominant nation. Below the dominant nation are the great powers. The great powers are not a match for the dominant nation in any point of time during its reign but might do so in the future during a period of hegemonic change. The next dominant power will be among the great powers. Middle powers and small powers are accordingly down in this hierarchy. Colonies, the last place in this order, have disappeared and have no real corresponding country today. The important point in the Power Transition Theory is the level of dissatisfaction of states in different hierarchy levels. As lower in the hierarchy a country is, as higher is level of dissatisfaction with the present situation. The only country that is free from dissatisfaction is the dominating country; all other countries would like to see a change of the present situation and would opt for a different hegemon, preferably themselves. As lower in the hierarchy a country is as more it wishes to initiate change, but unfortunately, countries of a low place in this hierarchy have the least means to do so. Transition Theory provides a good tool to judge the stability of a given system. Instability is only likely when two potential rivals grow into the same size (Kugler & Organski, 1989, p. 175): "As a dissatisfied great nation approaches parity by growing in power more rapidly than the dominant nation, instability increases and so does the probability of conflict. The closure of the power gap engenders fear on the part of the leaders in the dominant nation that the challenger will (1) surpass the dominant country, (2) become increasingly unwilling to accept a subordinate position in the international order, and
Hegemonic Power and Technology Advancement
565
(3) challenge the leadership and rules of the international order. And this, in fact, is very likely what the challenger will do. Thus, power transition argues that competition for dominance in the international order is finally joined when the dissatisfied party anticipates greater benefits and privileges if a conflict is successfully waged than if the current status quo is preserved". It should be noted that Power-of-Balance Theory (Morgenthau, 1978) suggests a contrary outcome: only if power is balanced between at least two nations a stable system is possible. 1.2
Systemic School
The second school of thought is the Systemic School. The Systemic approach to the study of hegemony takes off with one important insight of the second half of the 20th century. That is that states might be the main actors on the world stage but the part of the director takes someone else: organizations. Therefore scholars of the Systemic School take their seat far from the stage in order to see as much as possible. They see the whole world as one system and they recognize that politics might trigger events of hegemonic importance but the driving force is the economy. The Systemic School "... renders political-military power merely a subordinate extension of economic power, and its use of a global unit of analysis - the capitalist world economy-relegates national, state actors to a backup role." (McCormick, 1990, p. 127). Carrying the spirit of the systemic approach, Long Cycle Theory came to popularity in the mid 80's by the works of George Modelski and George Thompson. One of their important insights is that the international system and most of the events in it (like war) move in recurring cycles (Modelski, 1987). Each cycle passes through four distinct phases (Modelski & Thompson, 1989, p. 24): 1. Macrodecision - a period of war in which the question for leadership is settled, 2. Implementation - a period with a dominating nation (global leader) that implements major new programs, 3. Agenda setting - a period in which questions about the legitimacy of the global leader its agenda are raised, 4. Coalitioning - a period of de-concentration in which leadership reaches a low point and challengers recognize opportunities to spring into action. The idea that the international system moves in recurring cycles parallels Nikolai Kondratieff's suggestion, that economies move in cycles, or waves, of about 50 years (Kondratieff & Stolper, 1935). This theory was developed further by Jospeph Schumpeter, who suggested that the trigger for a new cycle to begin is a 'leading sector' (Schwartz, 1994). A leading sector is a cluster of innovation that creates new products with high demand and new and cheaper forms of energy. Sectors at the periphery of that leading sector would profit from the boom, developing innovative new products and lifting the whole economy. Following the finding that the international system as well as global economic development move in cycles it comes to mind that both types of cycles might be
566
O.F. Bahrisch and J.-S. Kim
correlated. Combining Carlo Cipolla (1970, p. 5) and Robert Gilpin's argumentation (Gilpin, 1981, p. 156) it might be suggested the following causal chain: 1) A nation that ascends the hegemonic throne provides a higher standard of living for its people due to superior access to raw materials, trade surplus and the economic fruits of the exploitation of Schumpetererian leading sectors. 2) Nations at the rim of the hegemon display tendencies of emulation and threaten the high standard of living of the leading nation. 3) The Hegemon is forced to allocate more recourses from the private sector to the public sector to maintain a political and military superiority in fending off emulating nations. 4) Protection costs of the hegemon rise due to an internal high standard of living. 5) Allies of the hegemon tend to take free rider positions and benefit form measures that are necessary for the hegemon to keep its superiority. 6) The equilibrium of the costs to maintain a leading position is reached: costs increase exponentially compared to benefits. 7) Snowball effect sets in and more and more resources are allocated away from the efficient private sector toward military spending, which increases the (relatively inefficient) public consumption. 8) The gap of superiority decreases until the reigning hegemon is too weak to defend its position. Following that, long economic waves and the long cycles of the international system should be highly correlated. (Rosecrance, 1987), studying this relationship, found only a rather week correlation. He concluded that only "every other Kondratieff upswing is associated with Modelski's global war phase" (Modelski's initial phase in each cycle) (p.291). Consistency is fully missing from the 1970s on, when economic development and war periods are completely out of step (Rosecrance, 1987, p. 291). Similar to Long Cycle Theory, the focus of the World-Systems Theory is the global system. Nation states are part of the world system but not center of the analysis. The primary representative in the advancement of the World-Systems Theory is Immanuel Wallerstein. McCormick (1990, p. 126) gives a splendid summary of Wallerstein's World-Systems Theroy: "... Wallerstein posited the existence since 1500 of a world system or capitalist world economy. Its chief actors have been transnational business organizations and operations, driven by the premise that capitalism functioned most efficiently and thus most profitably when its universe of options was sufficiently large and fluid for capital, goods, services, and people to move from one place to another to secure greater returns, even if that place was both distant and foreign. Its chief process was the spatial expansion of European capitalism at the expense of autarkic empires in Russia, China, and the Near East, as well as the subsistence, village minisystems of Eastern Europe, Ireland, the Americas, Africa, and Asia. Its chief consequence was a system of three successive zones, each performing a specialized function in a complex, hierarchical, international division of labor. Core countries (the First World) monopolize high-tech, high-profit enterprises. The periphery (the Third World)
Hegemonic Power and Technology Advancement
567
specializes in primary production of agricultural commodities and raw materials: "hewers of wood and carriers of water." In between, the semiperiphery (the Second World) performs inter- mediate functions of transport, local capital mobilization, and less complex, less profitable forms of manufacturing". It is important to note that Wallerstein's analysis starts off from an sociological standpoint ("social or societal behavior" (Wallerstein, 1979, p. 665)), less form an economical perspective (Wallerstein, 1998, p. 103). Wallerstein approaches the topic through an interpretation of perception of a society (Wallerstein uses the term "Weltanschauung" (1979, p. 665)). Human greed has lead to "the capitalist world system..." which "...has been well established now for some 400-500 years. It covers the globe" (Wallerstein, 1992, p. 616). "Virtually everyone tends to see capitalism as the system in which humans seek to transform (or "conquer") nature in an eternally expansive way and to amass ever larger residues of this expansion. Whether it is David Landes speaking of "Prometheus unbound" or Carlyle deploring the "cash nexus" or Marx analyzing the search for an endless accumulation of capital or Keynes's references to the "animal spirits" of (Schumpeter's) entrepreneurs or, as we have already seen, the description of the achievement of intensive growth as a "miracle," the phenomenon that is being observed takes the form of a hyper-bolic curve that knows no social limit" (Wallerstein, 1992, p. 567). Nations are in Wallerstein's analysis only a vehicle for businesses to achieve their goals for growth (access to markets, technology, labor, raw materials and so forth).
2
Hegemonic Change
The aforementioned four theoretical approaches to hegemonic power, despite belonging to different schools of thought and very different foci, identify one common trigger that sets in motion hegemonic change: extended military conflict. Gilpin refers to it as 'hegemonic war' (1988), Modelski/Morgan named it 'global war' (1985). "Throughout the history of the modern international state system, ascending powers have always challenged the position of the dominant (hegemonic) power in the international system—and these challenges have usually culminated in war (Layne, 2008, p. 16). A global or hegemonic war does not necessarily mean one single war, but a period of warfare with the participation of major powers. World War I and II have been such a period of warfare which, and that is throughout the the different theoretical school undisputed, lifted the United States on the hegemonic throne. Modelski/Morgan (1985) estimate that World War II consumed possibly one-third of the total world product between 1939 and 1945. They further quote Rasler and Thompson (1983) and Thompson and Zuk (1982), which analyzed statistical time series since 1694 and found that all earlier periods of global warfare were marked by extraordinary abrupt and permanent discontinuities in public debt levels of the hegemon and other participating nations. Additionally they found evidence for
568
O.F. Bahrisch and J.-S. Kim
significant changes of price levels and economic growth. Interstate wars occurring outside a period of global warfare did not show similar changes. In order to determine whether the current hegemonic cycle has reached its end and a new cycle starts, it is necessary to identify the current period of war as a global war. To qualify as a global war two measures must be met: first, engagement of the major powers and, second, extreme and permanent changes of public debt levels. The first measure is probably met. The current period of war is named Global War on Terror (manly fought in Afghanistan since 2001 and in Iraq since 2003) and involves with varying intensity all major nations. As Long Cycle Theory suggests, the major burden is carried by the current hegemon, the US, while the other participating nations rather reluctantly show willingness to shoulder arms and move to the hegemon's aid in the War on Terror. Each of the participating nations, but the US, have a strong incentive to free ride and the recent withdrawal of allies from the battlefield (Great Britain from Iraq in 2009) shows that the allies are attempting to shift costs to the hegemon. The second measure, permanent changes of debt levels, is more difficult it estimate. Although the US is currently running an nominal record level of public debt, it does not appear to be historically outstanding. The War on Terror is costly but not yet costly enough for lasting debt level changes. Figure 1 shows the ratio of US government spending in relation to the GDP. The current level of government spending is still dwarfed be the level of public spending during the years 1943-1945, the entry of the US into World War II - the last part of the a period of global warfare. Other core countries (or First World countries, according to Wallerstein) do not yet show signs of permanent changes of their debt levels. It might be argued that Japan, with public debt about twice of its GDP - on the rise since the late 1990s - would qualify for signaling the end of an hegemonic cycle. But Japan still appears to be an isolated case. The question whether we reached the end of the current hegemonic cycle is difficult to answer, like all predictions of the future. But it is possible to discuss the possible developments with the highest likelihood to become reality. Therefore are in the following three scenarios discussed are most likely to theorize the future hegemonic development: the US stays the hegemonic power, a new hegemon arises, or the hegemonic concept becomes obsolete.
Fig. 1. US government spending in percent of gross domestic product (1929-2010) (Government consumption expenditures and gross investment/Gross Domestic Product). (own calculation).
Hegemonic Power and Technology Advancement
569
Scenario 1: USA continue to be Number One Despite the currently popular prediction of the decline of United States, consultation of the present facts might suggest that the US, nevertheless remains in the hegmonic seat. During past hegemonic cycle changes have other core countries (World-System Theory), nations at the rim (Long-Cycle Theory) or great nation (Hegemonic Stability Theory) challenged the hegemon by attempting to match his military power (e.g. German-British naval arms race until 1914). The US defense spending had 2009 an unparalleled share of the entire world defense spending of 43 percent (SIPRI, 2009, 349). But even with an defense budget of USD 661bn, this sum only accounts for 4.3 percent of the US GDP. That makes the US a military superpower "on the cheap" (Kennedy, 2002, p. 13). The Stockholm International Peace Research Institute estimates second largest defense spender after the US as China with about USD 100 bn. Especially, when paying attention to the US naval power it becomes clear that for the next two decades there is no close rival to fear. An aircraft carrier battle group is probably the strongest military instrument a nation may posses. It is deployable with great flexibility and enormous firepower. The US maintain all of the 11 aircraft carrier of the larges class (supercarrier). As Kennedy (2002, p. 8) puts it: "No equivalent of power to a US carrier task force exists in the world; the few UK, French and Indian carriers are minuscule by comparison, the Russian ones rusting away." Military power alone does not keep a hegemon on the throne. Considerable political power is necessary as well in order to stay in control. In 1999, Samuel Huntington aptly summarized the overwhelming political ability to bend others to its will, but it is still valid today (p. 39): "In the past few years the United States has, among other things, attempted or been perceived as attempting more or less unilaterally to do the following: pressure other countries to adopt American values and practices regarding human rights and democracy; prevent other countries from acquiring military capabilities that could counter American conventional superiority; enforce American law extraterritorially in other societies; grade countries according to their adherence to American standards on human rights, drugs, terrorism, nuclear proliferation, missile proliferation, and now religious freedom; apply sanctions against countries that do not meet American standards on these issues; promote American corporate interests under the slogans of free trade and open markets; shape World Bank and International Monetary Fund policies to serve those same corporate interests; intervene in local conflicts in which it has relatively little direct interest; bludgeon other countries to adopt economic policies and social policies that will benefit American economic interests; promote American arms sales abroad while attempting to prevent comparable sales by other countries; force out one U.N. secretary-general and dictate the appointment of his successor; expand nato initially to include Poland, Hungary, and the Czech Republic and no one else; undertake military action against Iraq and later maintain harsh economic sanctions against the regime; and categorize certain countries as "rogue states," excluding them from global institutions because they refuse to kowtow to American wishes".
570
O.F. Bahrisch and J.-S. Kim
This impressive account is only true for the US it is difficult to imagine that there could be any nation that could come close to it in the two decades to come. Besides its superiority in military and political matters, the U.S. is also still the world’s economic powerhouse. The total output of the U.S. economy amounts to about one fifth of the entire world economy. Scenario 2: A New Hegemon Arises Despite of the apparently strong looking hegemon there are certain indicators that might be interpreted as signs of decline. Militarily the U.S. reached its limit. Engaging in two costly wars (Iraq and Afghanistan) left the U.S. with little room to be involved in a third war without risking to be overstretched. Economically has the aftermath of the financial cries 2008 left its mark in the shape of the highest level of public debt since World War II. That limits the monetary and fiscal policy of the government and casts the dark shadow of inflation over the U.S. economy. Fergueson (2010) found that increasing debt is responsible for the decline of most empires. The US might be no exception. Reinhart and Rogoff (2010), analyzing the relation of public debt, economic growth and inflation with a huge dataset reaching back until 1790, found a level of debt that exceeds 90% of the annual GDP has a negative impact on economic growth. They stated that: "above the threshold of 90 percent, median growth rates fall by one percent, and average growth falls considerably more" (p. 2). That might be interpreted that the current fiscal policy has a high possibility of being unsustainable. Currently a large share of US public debt is financed by China. If for any reason, China reduces its level of holdings in US government bonds, other financiers have to be found. Very likely is that US households will have to step into the breach, crowding out capital that would have otherwise flown into more productive private projects. Especially for future growth perspectives it is decisive whether money is used for the government's debt service or if it would finance innovations. A further point is that a strong military and economy might be of less importance in the future. It seems that the ability to establish networks (military, science, diplomacy and economy) might be much more contributing to strong influence. The US have lost some of this ability while other nations have gained. Probably the most discussed heir to the hegemonic throne is China. Its growing interest in raw material rich Africa, its growing military spending and its high economic growth rates lifted China to the position of the first possible successor. Tkacik (2007) estimated that the Chinese defense spending is actually way higher than the reported one. He assumes that in PPP China is spending as much as the U.S.
3
Hegemony and Technology
In May 1890 Theodore Roosevelt red “The Influence of Sea Power Upon History, 1660-1793” by Alfred Thayer Mahan (Morris, 2001, p. 433), a new book that probably not only influenced the future President of the United States, but also Kaiser Wilhelm II and Japan before World War II at aiming to build not only a strong navy
Hegemonic Power and Technology Advancement
571
but a navy with fierce striking power that leaves competing nations far behind. Mahan (2004) suggested that a nations security and international greatness depended on the navy, the technological superiority of ships and weaponry, and the concentration of various controlling spots that would be essential to global commerce. Mahan argued that striking with a superior force against certain pressure points could paralyze whole regions more effectively than any army. The power of great forces on land would be left useless compared with the concentrated use of a technologically advanced navy. Mahan correctly identified a source of ‘greatness’ but delivered a less correct explanation. The concentration of a society’s recourses on the development of certain innovative military instruments does not necessarily lead to a hegemonic position. It is rather being first in creating a cluster of technologies in leading sectors with spillover effects to the whole economy that put a nation on the hegemonic throne (Schumpeter, 1939, 1942). The basic idea is that innovation changes one sector into a leading sector, a cluster of innovation that creates new products with high demand and new and cheaper forms of energy and innovation. Sectors at the periphery would profit from the boom as well and might generate another key innovation that starts the cycle anew (Schwartz, 1994, p. 69). A hegemonic nation is therefore always a technological hegemon as well.
References Cipolla, C.M.: The economic decline of empires. Methuen, London (1970) Clark, I.A.N.: China and the United States: a succession of hegemonies? International Affairs 87(1), 13–28 (2011) Ferguson, N.: Complexity and Collapse - Empires on the Edge of Chaos. Foreign Affairs (89), 18–32 (2010) Gauger, A.J.: U.S. Business must Revolutionize, vol. 43. Omaha World-Herald, Omaha (1986) Gilpin, R.: U.S. power and the multinational corporation: the political economy of foreign direct investment. Basic Books, New York (1975) Gilpin, R.: War and change in world politics. Cambridge University Press, New York (1981) Gilpin, R.: The Theory of Hegemonic War. Journal of Interdisciplinary History 18(4), 591–613 (1988) Huntington, S.P.: The Lonely Superpower. Foreign Affairs 78(2), 35–49 (1999) Kennedy, P.: The Greatest Superpower ever. New Perspectives Quarterly 19(3 (Winter)), 8–18 (2002) Kindleberger, C.P.: Dominance and Leadership in the International Economy: Exploitation, Public Goods, and Free Rides. International Studies Quarterly 25(2), 242–254 (1981) Kindleberger, C.P.: The world in depression, pp. 1929–1939. (Rev. and enl. ed.)University of California Press, Berkeley (1986) Kondratieff, N.D., Stolper, W.F.: The Long Waves in Economic Life. The Review of Economics and Statistics 17(6), 105–115 (1935) Kugler, J., Organski, A.F.K.: The power transition: A retrospective and prospective evaluation. In: Midlarsky, M.I. (ed.) Handbook of war studies, pp. 171–194. Unwin Hyman, Boston (1989) Layne, C.: China’s Challenge to US Hegemony. Current History 107(705), 13–18 (2008) Machiavelli, N.: The Prince (G. Bull, Trans.). Penguin Books, Harmondsworth (1961)
572
O.F. Bahrisch and J.-S. Kim
Mahan, A.T.: The influence of sea power upon history, pp. 1660–1783. Barnes & Noble Books, New York (2004) McCormick, T.J.: World Systems. The Journal of American History 77(1), 125–132 (1990) Modelski, G.: Long cycles in world politics. University of Washington Press, Seattle (1987) Modelski, G., Morgan, P.M.: Understanding Global War. The Journal of Conflict Resolution 29(3), 391–417 (1985) Modelski, G., Thompson, W.R.: Long Cycles and Global War. In: Midlarsky, M.I. (ed.) Handbook of war studies, pp. 23–54. Unwin Hyman, Boston (1989) Morgenthau, H.J.: Politics among nations: the struggle for power and peace, 5th edn. Knopf, New York (1978) Morris, E.: The rise of Theodore Roosevelt. Random House, New York (2001) Organski, A.F.K.: World politics. Knopf, New York (1958) Rasler, K.A., Thompson, W.R.: Global Wars, Public Debts, and the Long Cycle. World Politics 35(4), 489–516 (1983) Reinhart, C., Rogoff, K.: Dept and Growth Revisited. Munich Personal RePEc Archive, Munich (2010) Rosecrance, R.: Long Cycle Theory and International Relations. International Organization 41(2), 283–301 (1987) Schwartz, H.M.: States versus markets: history, geography, and the development of the international political economy. St. Martin’s Press, New York (1994) Thompson, W.R., Zuk, L.G.: War, Inflation, and the Kondratieff Long Wave. Journal of Conflict Resolution 26(4), 621–644 (1982) Tkacik, J.J.: A Chinese Military Superpower? The Heritage Foundation. Asian Studies Center, The Heritage Foundation, Washington (2007) Wallerstein, I.: Kondratieff up or Kondratieff Down? Review (Fernand Braudel Center) 2(4), 663–673 (1979) Wallerstein, I.: The West, Capitalism, and the Modern World-System. Review (Fernand Braudel Center) 15(4), 561–619 (1992) Wallerstein, I.: The Rise and Future Demise of World-Systems Analysis. Review (Fernand Braudel Center) 21(1), 103–112 (1998) Webb, M.C., Krasner, S.D.: Hegemonic Stability Theory: An Empirical Assessment. Review of International Studies 15(2), 183–198 (1989)
Adaptive Hand Color Extraction Using Wave Motion Detection Kyungboo Jung1, Il-Moek Kim1, and Byung-Uk Choi2 1
Dept. of Electronics and Computer Eng., Hanyang University Division of Computer Science & Eng., Hanyang University {kbjung,imkim,buchoi}@hanyang.ac.kr
2
Abstract. Correct hand segmentation in image sequences is the most important aspect of hand-gesture interfaces. Existing hand-segmentation methods may be classified into two general approaches. One uses predefined skin color and the other uses a learning process. The skin color-learning method gives good results but its excessive processing time precludes its use in real time. In the predefined color approach, automatically modifying the predefined color when the environment changes would increase convenience and efficiency. Identifying a change in the environment usually requires user intervention. We propose a method by which the user can easily modify the predefined skin color any time in any environment simply by waving his/her hand in front of the camera. The hand interface system detects the waving hand region and efficiently extracts the hand color for use in hand-gesture interfaces.
system segments the hand and extracts skin color from the detected hand region for use in the next frame [5]. However, the learning process is very complex and time consuming. Another approach is to train the system about the hand color before the user actually uses the hand interface. The user simply positions his/her hand at some specific location in the scene so the system can extract the skin color and process it before the system is used [6]. Even though the predefined skin color can be modified by the user whenever the environment changes, this is a complicated interruption. We propose a method that uses the skin color extracted from the user’s simple hand motion in video images. When the user realizes that the hand interface system is no longer able to detect hands, he/she simply makes a gesture that can be detected easily in any environmental condition. The proposed method makes modifying the predefined skin color easy and efficient under any conditions. In the following sections, we briefly describe an overview of our method. Section 2 presents waving hand-detection and robust color-extraction methods. Section 3 describes the results of hand segmentation and fingertip estimation using our method. Section 4 presents our conclusions.
Fig. 1. Flowchart of proposed method
Adaptive Hand Color Extraction Using Wave Motion Detection
2
575
Proposed Method
In this section, we describe our approach for obtaining hand color semi-automatically. 2.1
Flowchart of Proposed Method
The proposed method of setting the hand color consists of two steps. The first step includes detection of the hand-waving motion and necessary preprocessing. We use the difference between two successive images to detect motion in image sequences. A residual image is created before the estimation phase for detecting successive finger movements. This residual image is a blend of the current differentiated image and the previous residual image. We can estimate and detect hand-waving gestures using these differentiated images. The second step is color extraction. The hand-waving estimation results in a color extraction circle. Hand color is extracted and accumulated until the number of clustered color samples reaches a specific threshold. Figure 1 presents a flowchart of our method. 2.2
What Is Waving Hand?
We cannot recognize hands without initial information about the skin color. To provide this, the user simply waves his/her hand in front of the camera. The arrow in Figure 2 shows how a waving hand motion causes finger movement. The figure also shows how the hand region below the finger motion can be described. The hand region contains much information about hand color in that specific environment. The red circle in the figure represents the hand region and the red point represents the center point of the hand.
Fig. 2. Waving hand gesture and hand region
2.3
Residual Image
Analyzing the difference between two successive frames is an easy way to detect a moving object in image sequences. However, the method must deal with the constant unexpected background motion due to normal human movement with portable interface devices such as head-mounted cameras. To remove this unexpected motion
576
K. Jung, I.-M. Kim, and B.-U. Choi
in the image sequences, we estimate camera movement between two successive frames and calculate a transform matrix. The motion can be reduced by warping the previous image, making it possible to distinguish between large object movements and small camera movements. We use the global motion estimation technique to calculate the transform matrix. We use a residual image to detect stable movements such as the fingers of a waving hand in the image sequences. A residual image is a blend of the previous residual image and the current difference image calculated as shown in Eq. (1): R
1
r R
rF .
(1)
where and are the current and previous residual images, respectively; is the current difference image; and is the blend ratio. Figure 3 shows how residual images are generated; panels (a), (c), and (e) show the original images in the sequence, and panels (b), (d), and (f) show the residual images. If the movement to extract hand color is very stable in the image sequences and shaped in the residual image at time , the movement region is estimated in the residual image. Finally, the hand color is extracted from the appropriate region.
Fig. 3. Difference images are blended top to bottom to form Rt. Camera images (a), (c), and (e) are at times t0, t1, and t2, respectively, while the corresponding residual images are in (b), (d), and (f).
Adaptive Hand Color Extraction Using Wave Motion Detection
2.4
577
Estimation of Waving Hand
A waving hand can be estimated using Eq. (2), which is a formula known as kernel density estimation (KDE) [7]: D x
∑N
N
K
R x .
(2)
is the estimated function, which is a probability density function of the where is a residual image; is the number of waving hand; is the kernel function; pixels in the image; and is the width of the kernel. The rough two-dimensional function denotes how much movement occurred in the scene during the time in question. The maximum point in the density function is the most active movement positions in the scene. The center point of the function is formulated using Eq. (3): µ
max D x .
(3)
where µ denotes the center point of the moving object. The standard deviation σ is calculated to determine the radius of the color extraction circle using Eq. (4): σ
N
∑
x
µ
T
x
µ D x
/
.
(4)
Figure 4 shows a residual image and the result of waving hand estimation. Figure 4(a) shows a residual image for estimation and Figure 4(b) shows the estimated density function. The circle indicates the most probable region for the waving hand to exist. We experimented with several types of kernel such as uniform kernel, Gaussian kernel, and the difference between two Gaussian kernels. We determined experimentally that the difference between two Gaussian kernels defined in Eq. (5) is suitable for detecting only waving hand movement:
Fig. 4. (a) Residual (b) the result of waving hand estimation
x
N
e
e
.
(5)
where K is the kernel, N is the number of pixels in the kernel, and σ is the standard deviation of a Gaussian kernel.
578
2.5
K. Jung, I.-M. Kim, and B.-U. Choi
Keyframe Selection and Color Extraction
In extracting hand color from image sequences, frames that exhibit a stable central point with a certain size of extraction circle are selected as keyframes. These have the correct hand color extraction circle. The colors extracted from the circle have the highest probability of being hand colors. To determine which frames are keyframes, we use the probabilistic model defined by Eq. (6): ∑
P S
C .
(6)
where P is an accumulation histogram and S is a random value that is true if frame t is chosen to extract the color or false if the frame is not chosen. The C is 1 if frame t meets several conditions for extracting hand color; otherwise it is 0. Eq. (7) is used to determine whether frame t is suitable for color extraction: 1, C 0,
, ,|
|
∆,
∆
.
(7)
otherwise
If the estimated waving hand value at central point D µ is greater than the threshold d, the central point has not moved more than µ∆ , and the standard derivation is and σ and has not changed more than σ∆ , then this between specific sizes σ frame t has the correct extraction circle positioned inside the hand in the image. The color information inside the circle is used in the next steps. We use the HSV color space because of its convenient characteristics. We take the colors in the circle to be the hand color. However, the circle may also contain a small proportion of nonhand colors such as the background. Therefore, colors are clustered in the HSV color space. The group with the most samples represents the hand color. For greater reliability and correct results, the colors extracted from only one frame are not sufficient. The colors extracted in each keyframe are accumulated as one color histogram. This makes the ratio of hand color increase dramatically in the accumulation process so that the clustering result is reliable. In each accumulation process, one color group that represents hand color has a significantly higher number of samples than any other group. The accumulation step can be terminated by checking the ratio of hand color to the others.
3
Experiments
To segment hands, we use the region-growing method that grows segmented regions by checking for similar colors among the seed pixels. Figure 5 shows the results of segmentation using our method along with the generated histogram, where panel (b) shows the segmented hand and detected fingers by applying the result of our method to the original video image in panel (a). Panels (c) and (d) show samples extracted from video images and the full HSV histogram that we used, respectively.
Adaptive Hand Color Extraction Using Wave Motion Detection
579
Fig. 5. Hand segmentation result using the color that extracted by our method and generated histogram. (a) Original video image (b) Segmented image (c) HSV color histogram with extracted color samples. (d) A full HSV color histogram.
4
Conclusions
We proposed a method that chooses a hand color by extracting a color from the densest hand-waving movement region for use in hand-gesture interfaces in various environments. Our method includes using the difference between two Gaussian kernels for KDE. With hand-waving detection, color extraction occurs only when the user waves his/her hand in front of the camera. This results in robust and efficient hand color extraction that is suitable for use in hand-gesture interfaces in any environment.
References 1. Hardenberg, C., Bérard, F.: Bare-hand human-computer interaction. In: Proceedings of the 2001 workshop on Perceptive user interfaces. ACM International Conference Proceeding Series, vol. 15, pp. 1–8 (2001) 2. O‘Hagan, R., Zelinsky, A.: Finger Track – A Robust and Real Time Gesture Interface. In: Australian Joint Conference on Artificial Intelligence, pp. 475–484 (1997) 3. Quek, F., Mysliwiee, T., Zhao, M.: Finger mouse: A freehand pointing interface. In: International Workshop on Automatic Face and Gesture-Recognition, pp. 372–377 (1995) 4. Mistry, P., Maes, P.: SixthSense – A Wearable Gestural Interface. In: The Proceedings of SIGGRAPH, Asia (2009)
580
K. Jung, I.-M. Kim, and B.-U. Choi
5. Lee, T., Hollerer, F.: Handy AR: Markerless Inspection of Augmented Reality Objects Using Fingertip Tracking. Wearable Computers. In: 11th IEEE International Symposium on In Wearable Computers, pp. 83–90 (2007) 6. Xiaoming, Y., Dong, G., Ming, X.: Hand image segmentation using color and RCE neural network. Robotics and Autonomous Systems 34(4), 235–250 (2001) 7. Parzen, E.: On estimation of a probability density function and mode. Ann. Math. Stat. 33, 1065–1076 (1962)
A Dynamic Resource Allocation Model for Virtual Machine Management on Cloud* Chao-Tung Yang**, Hsiang-Yao Cheng, and Kuan-Lung Huang Department of Computer Science, Tunghai University, Taichung 40704, Taiwan [email protected], {hsycheng,peter760504}@gmail.com Abstract. It is current issue to achieve the goal of management multiple virtualization platforms and multiple virtual machine migrations across physical machines without disruption. We discuss that ensure load balance when multiple virtual machines run on multiple physical machines. In this paper, we present a system which is implementation of optimization with Dynamic Resource Allocation (DRA) dealing with virtualization machines on physical machines. And practice DRA method in this system. The results confirmed that the virtual machine which loading becomes too high, it will automatically migrated to another low loading physical machine without service interrupting. And let total physical machine loading reaching balance. Keywords: Virtualization, VM Migration, Dynamic Resource Allocation, Cloud Computing.
1
Introduction
A virtual machine was originally defined by Popek and Goldberg as ―an efficient, isolated duplicate of a real machine [1]. Virtual machine operates in a virtual platform can be considered an independent operating system. In addition, due to the independence of this individual, the gust operating system can operate from the original virtual platform, and to maintain the operation of the original schedule. And this move action, generally referred to as Migration [2-23, 30-38]. The migration is dynamic: if a move action, cause the system to pause time is extremely small enough, it cannot be aware of user using the system services. Setting up virtual machine cluster environment on physical machine can provide stable service, but this environment often includes unpredictable workloads. Currently most systems of virtual machine are loading balanced statically, as the systems where the load changes dynamically over time run it is inevitable that some physical hosts with higher load, so for throughput and response time of a system to be maximized it is necessary for load to be distributed to each part of the system in proportion to their computing/IO capacity [3, 20, 24-27]. *
In this paper, we develop an adaptive resource control system on a virtual machine cluster that dynamically adjusts the resource shares to individual tiers in order to meet DRA (Dynamic Resource Allocation) goals. There are three efforts in this paper: (1) supporting DRA mechanism; (2) implementation OpenNebula management tool on a web-based interface; and (3) efficiently isolating the cluster workloads.
2
Virtualization
Virtualization is simply the logical separation of the request for some service from the physical resources that actually provide that service. In practical terms, virtualization provides the ability to run applications, operating systems, or system services in a logically distinct system environment that is independent of a specific physical computer system. Obviously, all of these have to be running on a certain computer system at any given time, but virtualization provides a level of logical abstraction that liberates applications, system services, and even the operating system that supports them from being tied to a specific piece of hardware. Virtualization, focusing on logical operating environments rather than physical ones, makes applications, services, and instances of an operating system portable across different physical computer systems. Virtualization can execute applications under many operating systems, manage IT more efficiently, and allot resources of computing with other computers [1, 4-7, 9, 11, 12, 28].
3
Related Work
Recently, the dramatic performance improvements in hypervisor technologies have made it possible to experiment with virtual machines (VM) as basic building blocks for flexible computational platforms. Many research efforts have been introduced to reduce the overhead of the networking in virtualized environments. Data transfer between server nodes frequently occurs in parallel and distributed computing systems, the high overhead of networking may induce significant performance loss in the overall system. Jae-Wan Jang [19] use virtualized parallel and distributed computing systems are rapidly becoming the mainstream due to the significant benefit of high energyefficiency and low management cost. Processing network operations in a virtual machine incurs a lot of overhead from the arbitration of network devices between virtual machines, inherently by the nature of the virtualized architecture. Wang L et al.[19] propose a new methodology for Grid computing – to use virtual machines as computing resources and provide Virtual Distributed Environments (VDE) for Grid users. Paul Willmann [36] presents hardware and software mechanisms to enable concurrent direct network access by operating systems running within a virtual machine monitor. It is declared that employing virtual environment for Grid computing can bring various advantages, for instance, computing environment customization, QoS
A Dynamic Resource Allocation Model for Virtual Machine Management on Cloud
583
guarantee and easy management. A light weight Grid middleware, Grid Virtualization Engine, is developed accordingly to provide functions of building virtual environment for Grids. VMware DRS [20] achieves an on-demand resource scheduling scheme for virtual machine cluster via migrating virtual machines among physical machines. In our scheme, the two measures are used simultaneously while reallocating resource of virtual machines within same physical machine is the first choice to get higher efficiency. Additionally, R. S. Montero [21] proposes a performance model to characterize these variable capacity (elastic) cluster environments. The model can be used to dynamically dimension the cluster using cloud resources, according to a fixed budget, or to estimate the cost of completing a given workload in a target time. This paper focuses VMs running on physical machines and use DRA technology to Implementation a virtualization environment of HPC.
4
System Implementation
The purpose of our proposed Dynamic Resource Allocation (DRA) is to reach the best balance between each physical machine. To avoid computing resources centralized on some specify physical machines, how to balance the resources becomes the most important issue. To achieve the maximum efficiency the resource must be evenly distributed. DRA manages resources allocation to a set of virtual machines running on a cluster hosts to the goal of fair and effective use of resources. Virtual machine placement and migration recommendations serve to enforce resource-based service level agreements, user-specified constraints, and loading balance maintenance across the cluster as workloads change, as shown in Figure 1. Load balancing seeks to improve the performance of a distributed system by allocating the workload amongst a set of cooperating hosts. Such system may attempt to ensure the workload on each host is within a small tolerance of the workload on all other physical hosts, or may attempt to avoid congestion of individual servers. Load balancing can be either centralized or distributed [29]. The OpenNebula core orchestrates three different management areas: image and storage technologies (that is, virtual appliance tools or distributed file systems) for preparing disk images for VMs, the network fabric (such as Dynamic Host Configuration Protocol [DHCP] servers, firewalls, or switches) for providing VMs with a virtual network environment, and the underlying hypervisors for creating and controlling VMs. The core performs specific storage, network, or virtualization operations through pluggable drivers. Thus, OpenNebula isn’t tied to any specific environment, providing a uniform management layer regardless of the underlying infrastructure. Besides managing individual VMs’ life cycle, we also designed the core to support services deployment; such services typically include a set of interrelated components (for example, a Web server and database back end) requiring several VMs. Thus, we
584
C.-T. Yang, H.-Y. Cheng, and K.-L. Huang
can treat a group of related VMs as a first-class entity in OpenNebula. Besides managing the VMs as a unit, the core also handles the delivery of context information (such as the Web server’s IP address, digital certificates, and software licenses) to the VMs [35].
Fig. 1. Global load balancing algorithm concept
A separate scheduler component makes VM placement decisions. More specifically, the scheduler has access to information on all requests OpenNebula receives and, based on these requests, keeps track of current and future allocations, creating and updating a resource schedule and sending the appropriate deployment commands to the OpenNebula core. The OpenNebula default scheduler provides a rank scheduling policy that places VMs on physical resources according to a ranking algorithm that the administrator can configure. It relies on real-time data from both the running VMs and available physical resources. Figure 2 shows the system perspective. According to the previous works we build a cluster system with OpenNebula and also provide a web interface to manage virtual machines and physical machine. Our cluster system was built up with four homogeneous computers; the hardware of these computers is equipped with Intel i7 CPU 2.8 GHz, four gigabytes memory, 500 gigabytes disk, Debian operating system, and the network connected to a gigabit switch. The Dynamic Resource Allocation is an efficient approach to increasing availability of host machine. However, at present open source virtual machine management software merely provide a web interface for user managing virtual machine. Such as Eucalyptus [37] cannot accomplish load balance. When a part of virtual machines load increasing, it will affect all virtual machine on the same host machine. Our Dynamic Resource Allocation algorithm can overcome this obstacle, and improve host machine performance. Dynamic Resource Allocation works by continuously monitoring all virtual machines resource usage to determine which virtual machine have to migrate to another host machine. The goal is to make all host machine CPU and memory loading identically.
A Dynamic Resource Allocation Model for Virtual Machine Management on Cloud
585
Fig. 2. Our system architecture
The Dynamic Resource Allocation process is as follows. Assuming j host machines are in this pool. Every host machine loading ideal ratio is α=1⁄j. And i virtual machines are not running load balancing in these host machine. Each virtual machine resource usage is defined as “VMi Ratio” ሺܸ݁ݐ݈݈ܽܿܽܯܣܴ݅ܯܸ כ ݁ݏݑܷܲܥ݅ܯሻΤσୀଵሺܸ݁ݐ݈݈ܽܿܽܯܣܴ݅ܯܸ כ ݁ݏݑܷܲܥ݅ܯሻ(1)
Where VMi Ratio denotes virtual machine resource usage percentages in all allocate CPU and memory physical resource. When VMi CPUuse increases, VMi Radio is also increase too. In the next step, virtual machines resource usage ratio has been added up on different host machine. Each host machine current resource usage is defined “HOSTj Ratio”, is: ∑
(2)
Where HOSTj Rate must compare with ideal ratio α. When HOSTj Rate is bigger than α, it presents this host loading is too high, must migrate virtual machine to another host machine, and it also determine migrate source host. At the first, decide which host machine to be the migrated source host. Such as: max (HOST_(j Rate)-α), and decide which host machine to be the migrated destination host, such as: min (HOST_(j Rate)-α). And at last, the migrated virtual machine is defined VMk migrate. Such as: (3) The algorithm performs some calculations for monitoring physical resource information. It is follows: [Initialization] Defined virtual machine amount i and host machine amount j
586
C.-T. Yang, H.-Y. Cheng, and K.-L. Huang
Calculate ideal ratio α=1⁄j, virtual machine resource ratio VMi Rate, host machine resource ratio HOSTj Rate do determine migrate source host machine m max (HOST_(j Rate)-α) determine migrate destination host machine n min (HOST_(j Rate)-α) determine migrated virtual machine on HOSTj VMk migrate migrate virtual machine VMk of host m to host n while (VMk migrate value = 0)
5
Experimental Results
We focus on resource utilization of computing under DRA model. Therefore, we also used HPCC software to verify that DRA has a good performance and utilization on virtualization cluster. HPCC is an abbreviation of High Performance Computing Challenge, the HPC Challenge Benchmark is a set of benchmarks targeting to test multiple attributes that can contribute substantially to the real-world performance of HPC systems, co-sponsored by the DARPA High Productivity Computing Systems program, the United States Department of Energy and the National Science Foundation [37]. First of all, we focus on resource utilization of computing under DRA model. Therefore, we used HPCC [11, 16, 17, 20, 39-42] software to verify that DRA has a good performance and utilization on virtualization cluster. HPCC is an abbreviation of High Performance Computing Challenge, the HPC Challenge Benchmark is a set of benchmarks targeting to test multiple attributes that can contribute substantially to the real-world performance of HPC systems [31]. Figure 3 is shown HPCC computing time. The horizontal axis represented HPCC problem size and the vertical axis represented HPCC computing time. We noticed that while HPCC problem size growing up, the difference of HPCC computing finished time when DRA function enable or not will be more obviously. In this experiment, we run HPCC programs onto six virtual machines and calculate HPCC performance on these six virtual machines. It caused virtual machines cluster CPU usage jumped and affect HOST machine CPU usage relatively. When DRA function disable, virtual machines located on same HOST machine and proceeding HPCC computing simultaneously. It caused virtual machines snatch at physical resource each other. When DRA function enable, it will detect all host machine resource usage was balancing or not, therefore, virtual machines on same HOST machine were migrated to others automatically. Figure 4 is also shown DRA function effectiveness. The vertical axis represented virtual machine floating point calculation performance. With DRA function enabled will obtain good performance. It also proved our thesis is workable under this circumstance. In Figure 4, it shows better performance when virtual machines centralized on the same host than on distributed hosts. Because HPCC performance computing on virtual machines cluster transfer computing data to each virtual machine, so these virtual machines deliver message to each other by the host virtual switch. But we observed that when problem size reach 6000, DRA enabled virtual machines
A Dynamic Resource Allocation Model for Virtual Machine Management on Cloud
587
distributed to different hosts, the HPCC performance is better than DRA disabled virtual machines. Because problem size is too big, so virtual machines cluster on the single host cannot afford the computation.
Fig. 4. The performance of Floating Point per second
Figure 4 is shown HPCC computing time. The horizontal axis represented HPCC problem size and the vertical axis represented HPCC computing time. We noticed that while HPCC problem size growing up, the difference of HPCC computing finished time when DRA function enable or not will be more obviously. Figure 5 is also shown DRA function effectiveness. The vertical axis represented virtual machine floating point calculation performance. With DRA function enabled will obtain good performance. It also proved our thesis is workable under this circumstance.
588
C.-T. Yang, H.-Y. Cheng, and K.-L. Huang
Fig g. 5. The CPU usages of host machines
6
Conclusions
In this work we have pressented an optimization with dynamic resource allocattion model for clusters that allo ows a flexible management of these computing platforrms by: (1) supporting DRA mechanism; m (2) implementation OpenNebula managem ment tool on web-based interfacee; and (3) efficiently isolating the cluster workloads. Mooreover, this architecture is able a to transparently grow the cluster's capacity usingg an external cluster provider. Although A there is another Open Source virtualization proj oject like, Eucalyptus. But it is difficult to reach DRA goal, because lack life migrattion oose the OpenNebula solutions to hit our goal in this papper. function. Therefore, we cho Based on this model it is straightforward to plan the capacity of the cluster to, for o complete a given workload. We envision the use of thhese instance, meet a deadline to kinds of models by addition nal components to dynamically change the cluster capaccity according to a given budg get, performance policy or in conjunction with a run and queue wait time prediction service. Finally, the architecture presented in this workk is compatible with the use of physical resources. These resources can be divided eveenly by the mechanism.
References 1. Goldberg, R.P., Popek, G.J.: G Formal requirements for virtualizable third generation arrchitectures (1974) 2. Clark, C., et al.: Live migration of virtual machines. Presented at the Proceedings off the 2nd Conference on Symp posium on Networked Systems Design & Implementation, vool. 2 (2005)
A Dynamic Resource Allocation Model for Virtual Machine Management on Cloud
589
3. Yi, Z., Wenlong, H.: Adaptive Distributed Load Balancing Algorithm Based on Live Migration of Virtual Machines in Cloud. Presented at the INC, Fifth International Joint Conference on IMS and IDC, NCM 2009 (2009) 4. Hagen, W.V.: Professional Xen Virtualization (2008) 5. Uhlig, R., Neiger, G., Rodgers, D., Santoni, A.L., Martins, F.C.M., Anderson, A.V., Bennett, S.M., Kagi, A., Leung, F.H., Smith, L.: Intel virtualization technology. Computer 38, 48–56 (2005) 6. Whitaker, A., et al.: Rethinking the design of virtual machine monitors. Computer 38, 57– 62 (2005) 7. Rosenblum, M., Garfinkel, T.: Virtual machine monitors: current technology and future trends. Computer 38, 39–47 (2005) 8. Kivity, A., et al.: kvm: the Linux virtual machine monitor. In: OLS 2007: Proceedings of the Linux Symposium, Ottawa, Ontario, Canada, pp. 225–230 (2007) 9. Andrew Whitaker, M.S., Gribble, S.D.: Denali: Lightweight Virtual Machines for Distributed and Networked Applications. Technical Report 02, 10 (2002) 10. Barham, P., et al.: Xen and the art of virtualization. Presented at the Proceedings of the Nineteenth ACM Symposium on Operating Systems Principles, Bolton Landing, NY, USA (2003) 11. Lowe, S.: Mastering VMware vSphere 4, pp. 26–41 (2009) 12. Williams, K.B.D.E., Garcia, J.R., Rosen, R.: Virtualization with Xen Including XenEnterprise, XenServer, and XenExpress, pp. 23–117 (2007) 13. Srodawa, R.J., Bates, L.A.: An efficient virtual machine implementation. Presented at the Proceedings of the Workshop on Virtual Computer Systems, Cambridge, Massachusetts, United States (1973) 14. Xu Hui, M. H., Wang, X., Wang, Z.: Study on the dynamic model of leaf area of Eucalyptus camaldulensis. Yunnan Forestry Science and Technology, 20–22 (2000) 15. Nurmi, D., et al.: The Eucalyptus Open-Source Cloud-Computing System. In: 9th IEEE/ACM International Symposium on Cluster Computing and the Grid, CCGRID 2009, pp. 124–131 (2009) 16. Sempolinski, P., Thain, D.: A Comparison and Critique of Eucalyptus, OpenNebula and Nimbus. In: 2010 IEEE Second International Conference on Cloud Computing Technology and Science (CloudCom), pp. 417–426 (2010) 17. Dejan, F.D., Miloji, S., Paindaveine, Y., Wheeler, R., Zhou, S.: Process migration. ACM Comput. Surv. 32, 241–299 (2000) 18. Hansen, J.G., Jul, E.: Self-migration of operating systems. Presented at the Proceedings of the 11th Workshop on ACM SIGOPS European Workshop, Leuven, Belgium (2004) 19. Yang, C.-T., Tseng, C.-H., Chou, K.-Y., Tsaur, S.-C.: Design and Implementation of a Virtualized Cluster Computing Environment on Xen. In: Presented at the The second International Conference on High Performance Computing and Applications, HPCA (2009) 20. Resource Management with VMware DRS 21. Seo, E., Jang, J.-W., Jo, H., Kim, J.-S.: A low-overhead networking mechanism for virtualized high-performance computing systems. The Journal of Supercomputing (2010) 22. Willmann, P., et al.: Concurrent Direct Network Access for Virtual Machine Monitors. In: IEEE 13th International Symposium on High Performance Computer Architecture, HPCA 2007, pp. 306–317 (2007) 23. Sotomayor, B., et al.: Virtual Infrastructure Management in Private and Hybrid Clouds. IEEE Internet Computing 13, 14–22 (2009) 24. Greschler, D., Mangan, T.: Networking lessons in delivering ’Software as a Service’: part I. Int. J. Netw. Manag. 12, 317–321 (2002)
590
C.-T. Yang, H.-Y. Cheng, and K.-L. Huang
25. Greschler, D., Mangan, T.: Networking lessons in delivering ’Software as a Service’: part II. Int. J. Netw. Manag. 12, 339–345 (2002) 26. Hine, J.H., Bubendorfer, K.: A Compositional Classification For Load-Balancing Algorithms (1998) 27. Somani, G., Chaudhary, S.: Load Balancing in Xen Virtual Machine Monitor. In: Ranka, S., Banerjee, A., Biswas, K.K., Dua, S., Mishra, P., Moona, R., Poon, S.-H., Wang, C.-L. (eds.) IC3 2010. CCIS, vol. 95, pp. 62–70. Springer, Heidelberg (2010) 28. Smith, J.E., Ravi, N.: The architecture of virtual machines. Computer 38, 32–38 (2005) 29. Amazon Web Services home page, http://aws.amazon.com/ 30. Enomalism elastic computing infrastructure, http://www.enomaly.com/ 31. Salesforce Customer Relationships Management (CRM) system, http://www.salesforce.com/ 32. Logical Volume Manager, http://linuxconfig.org/Linux_lvm_-_Logical_Volume_Manager 33. Meth, K.Z., Satran, J.: Features of the iSCSI protocol. IEEE Communications Magazine 41, 72–75 (2003) 34. Anderson, E., et al.: Hippodrome: Running Circles Around Storage Administration. Presented at the Proceedings of the 1st USENIX Conference on File and Storage Technologies, Monterey, CA (2002) 35. OpenNebula, http://www.opennebula.org 36. Shafer, J., Willmann, P., Carr, D., Menon, A., Rixner, S., Cox, A.L., Zwaenepoel, W.: Concurrent Direct Network Access for Virtual Machine Monitors. In: The Second International Conference on High Performance Computing and Applications, HPCA (2007) 37. Eucalyptus, http://open.eucalyptus.com 38. Borja Sotomayor, R.S.M., Llorente, I.M., Foster, I.: Virtual Infrastructure Management in Private and Hybrid Clouds. IEEE Internet Computing 13 (2009) 39. Apache JMeter, http://jakarta.apache.org 40. You, J., et al.: JMeter-based aging simulation of computing system. In: 2010 International Conference on Computer, Mechatronics, Control and Electronic Engineering (CMCE), pp. 282–285 (2010) 41. HPCC, http://icl.cs.utk.edu/hpcc/ 42. Browne, S., et al.: The National HPCC Software Exchange. IEEE Computational Science & Engineering 2, 62–69 (1995)
Design of an Iterative Learning Controller of Nonlinear Dynamic Systems with Time-Varying In Ho Ryu1, Hun Oh2, and Hyun Seob Cho3 1
Dept. of IT Applied System Engineering, Chonbuk National University #664-14, 1-ga, Duck-jin Dong Duck-jin Ku, Jeonju, 561-756, Rep. of Korea 2 Dept. of Electrical Electronic and Information, WonKwang University Engineering #416, Jinbuk-dong, Deokjin-gu, Jeonju-si, Jeonbuk, Rep. of Korea 3 Dept. of Electronic Engineering, ChungWoon University #San29, Namjang-ri, Hongseong-eup Hongseong-gun, Chungnam, Rep. of Korea [email protected]
Abstract. Connectionist networks, also called neural networks, have been broadly applied to solve many different problems since McCulloch and Pitts had shown mathematically their information processing ability in 1943. In this thesis, we present a genetic neuro-control scheme for nonlinear systems. Our method is different from those using supervised learning algorithms, such as the backpropagation (BP) algorithm, that needs training information in each step. The contributions of this thesis are the new approach to constructing neural network architecture and its training.
algorithm. We applied reinforcement learning, a powerful machine learning mechanism, particularly Q-learning, to genetically optimized networks.
2
Systems and Control
2.1
Nonlinear Systems
In a conventional way, nonlinear control problems have been solved by using linearization technique. However, it provides a method which is valid for only a limited range of operation. An nth-order continuous-time system has the following general form:
(2.1)
Similarly, an nth-order discrete-time system has the following form:
(2.2)
for k=0, 1, 2, ... . Those can be expressed in the vector form (2.3) (2.4) The systems above depend upon both the state x and time, k or t. These systems are, in general, time varying. 2.2
Dynamics of Neural Networks
The activation function or transfer function, denoted by F[⋅], maps the unbounded junction value to a bounded output of neuron, and defines the activation level of node. There are three main classes of activation functions which have been developed and used: Binary: The output is hardlimited to binary [0, 1] or bipolar [-1, 1] values.
3
Genetic Algorithms
Genetic algorithms are a highly parallel mathematical algorithms that transform a set (population) of mathematical objects(typically fixed -length binary character strings), each with an associated fitness value, into a new set (population) of mathematical
Design of an Iterative Learning Controller of Nonlinear Dynamic Systems
593
objects. The transformation performed by GAs consists of naturally occurring genetic operations and the Darwinism of reproduction and survival of the fittest. 3.1
Genetic Reinforcement Learning Control
In the inverted pendulum problem, each real-valued string in the population is decoded to form a network with five input units, five hidden units, and one output units. The network is fully connected and this network configuration is same as that used by Anderson with the AHC algorithm. Since there are 35 links in the network, each string used by the genetic search includes 35+1 real values concatenated together. Before any input is applied to the network, the four state variables are normalized between 0 and 1. A bias unit fixed at 0.5 is also used as a fifth input to the net; a weight from the bias unit to a hidden node(or output node) in effect changes the threshold behavior of that node. The action of the neural network for a particular set of inputs is determined from the activation of the output unit. Learning is stopped when a network was found that was able to maintain the system without generating a failure signal for 120,000 time steps. One potential problem with such a simple evaluation criterion is that a favorable or unfavorable start state may bias the fitness ranking of an individual net. In other words, the evaluation function is noisy. We would like to assign a fitness value to a string based on its ability to perform across all possible start states.
Fig. 1. Genetic reinforcement learning control system
4
Simulation Results
Figure 2 and Figure 3 shows results for an AHC network and a genetically trained network using a failure signal occurring at 12 degrees during learning. These plots illustrate the tracking control behavior over time. During training, the output is determined probabilistically depending on the activation of the output unit. During testing, the action applied to the system is obtained by deterministically thresholding the activation value of the output unit. If the activation value is greater than 0, then
594
I.H. Ryu, H. Oh, and H.S. Cho
output 1 and push right; if it is less than or equal to 0, then output -1 and push left. If the pole is vertical and the cart is centered and the velocities are 0, then all state variables will have the normalized value 0. When the system is started in an ideal state, then a successfully trained network will maintain the state variables close to the 0 level. It is not possible to balance the pole and avoid the track terminals from all possible, a perfectly trained network should drive all state variables back to the 0 level representing the ideal state of the system. In Figure 2 and 3, the cart is at the far right end of the track with the pole learning 32degrees to the left; the 12-degree failure signal is not suitable for these tests. The cart velocity and pole velocity are initialized 0. This initial state constitutes a position from which it is difficult for the system to recover. Both the AHC network and the genetically trained network used to produce these graphs are the best networks obtained for the 12-degree problem. In the case of Figure 3, the genetically trained network gets all of the input variables into tolerable ranges fairly quickly, whereas the AHC network takes longer. The AHC network quickly damps pole velocity and reduces oscillation in the pole position, however at that time, the cart almost crashes into the opposite end of the track. The genetically trained network handles problems with starting pole angles beyond 32 degrees, but the AHC network does not. Figure 4 and 5 show results for an AHC network and a genetically trained network using a failure signal at 36 degrees during learning. These plots indicate that both AHC network and the genetically trained network exploit similar information to determine the output activation levels and that they employ similar control strategies. The networks trained at 36 degrees proved to be more similar across a wider range of start states, but as the difficulty of the initial start states is increased the AHC networks fail sooner than the genetically trained networks. In these plots, the system is started with cart in the same far right position and the pole learning 35 degrees to the left. Cart velocity and pole velocity are initially 0. These plots make it evident that both networks track pole velocity by varying the magnitude of the output value. The correlation between pole velocity and the output activation is not as discernible in the first 50 to 100 time steps because the system is recovering from a difficult initial situation; correlation between the pole velocity and the output activation is much more pronounced as the networks begin to bring the system under control. Also notable is that cart velocity and pole velocity tend to be negatively correlated. Given the input definitions used in our experiments, cart velocity and pole velocity have a similar, but opposite relationship.
Fig. 2. Control results by AHC network for 12°problem
Design of an Iterative Learning Controller of Nonlinear Dynamic Systems
595
Fig. 3. Control results by genetically optimized network for 12°problem
Fig. 4. Control results by AHC network for 36°problem
Fig. 5. Control results by genetically optimized network for 36°problem
5
Conclusions
In this thesis, we showed genetic algorithms can be used for optimizing neural network topology and connection weights. In addition, we presented the optimized neural network was good for solving nonlinear control problem. The performance of the proposed system was confirmed by applying it to the inverted-pendulum control problem. In our genetic approach, updates to the action network occur only after one or more failures: learning is not continuous. Another difference is that the genetic approach used in the experiments will assign an equal evaluation to two networks that
596
I.H. Ryu, H. Oh, and H.S. Cho
avoid failure for an equal number of time steps. However the AHC algorithm evaluates networks by the trajectory of states by that are experienced. The evaluation associated with any two networks would differ when the AHC algorithm is used, favoring the network that drives the cart-pole through more highly valued states. The restriction of the search to highly valued states may also explain why the performance of AHC networks did not improve when the stricter stopping criterion was used. The genetic algorithm, because it ranks each network based on performance, is able to ignore those cases where the pole cannot be balanced; only the successful cases will obtain the chase to engage in genetic reproduction. For the AHC evaluation network, however, the preponderance of failures may cause all states to overpredict failure. This problem can be corrected either by selectively sampling the space to extract a better balance of success and failure information or by tuning the AHC algorithm to place more emphasis on positive results and less on failure. Overall, proposed method can be used for nonlinear control problems. Improvement for more nonlinear and complicated problems is the future work.
References [1] Franklin, G.F., David Powell, J., Emami-Naeini, A.: Feedback Control of Dynamic Systems, 3rd edn. Addison-Wesley (1994) [2] Ogata, K.: Modern Control Engineering, 2nd edn. Prentice Hall (1990) [3] Hocking, L.M.: Optimal Control: An Introduction to the Theory with Applications. Oxford University Press (1991) [4] Fu, L.: Neural Networks in Computer Intelligence. McGraw-Hill (1994) [5] Hecht-Nielsen, R.: Neurocomputing. Addison-Wesley (1990) [6] Haykin, S.: Neural Networks. Macmillan (1994) [7] Kolk, W.R., Lerman, R.A.: Nonlinear System Dynamics. Van Nostrand Reinhold (1992) [8] Michalewicz, Z.: Genetic Algorithms + Data Structures = Evolution Programs, 3rd edn. Springer, Heidelberg (1996)
Adaptive Controller Design of the Nonlinear Dynamic Systems with a Neural Networks Compensator Hyun Seob Cho Dept. of Electronic Engineering, ChungWoon University #San29, Namjang-ri, Hongseong-eup Hongseong-gun, Chungnam, Rep. of Korea [email protected]
Abstract. Dynamic Neural Unit (DNU) based upon the topology of a reverberating circuit in a neuronal pool of the central nervous system. In this thesis, we present a genetic DNU-control scheme for unknown nonlinear systems. Our method is different from those using supervised learning algorithms, such as the backpropagation (BP) algorithm, that needs training information in each step. The contributions of this thesis are the new approach to constructing neural network architecture and its training. Keywords: DNU-Control, Genetic Algorithm, Nonlinear System.
1
Introduction
The conventional design methods of a control system often require the construction of a mathematical model describing the dynamic behavior of the plant to be controlled. When such a mathematical model is difficult to obtain due to uncertainty or complexity of systems, these conventional techniques based on a mathematical model are not well suited for dealing with. Artifical neural network teachiques have been suggested for identification and control of nonlinear plants for which conventional techniques of control do not give satisfactory performance, such as the accuracy in matching the behavior of the physical system. A good method of applying neural networks for control must have the following properties: 1. It must be applicable to nonlinear plants, since there are already good methods of control for linear plants. 2. It should not put too much restriction on the type of nonlinearity that it can handle. 3. It is preferable to have an Unsupervised Learning method for the neural network because the desired output form of a system for a given input may be known, but the input form of a plant that produces that desired output is not generally known. Unsupervised Training can avoid identification of the plant or its inverse model, which is generally not easy to obtain. 4. The system should be stable at least for the class of inputs it has been trained for.
5. In most cases open loop performance of a plant can be observed and a approximate controller can be devised for that. It would be desirable if we could put as much knowledge as possible in the design of this controller and only leave the extra fine tuning to the neural network controller. According to the above requirements, a direct auxiliary controller for nonlinear plants using neural network is presented.
2
Controller Design
The controller presented here is composed of an approximate controller and a neural network auxiliary controller. The structure of controller is shown in Figure 1. The approximate controller gives the approximate performance and the neural network auxiliary controller is used for the purpose of further fine tuning. The approximate controller can be a PID or any other conventional controller. It can be designed by using the known dynamics of nonlinear plant. The neural network employed in this scheme is an Radial Basis Function Network (RBFN). It produces the complementary signal to further reduce the error e between output y and the reference input r. The structure of RBFN is showed in Figure 2. It is a network with two layers. A hidden layer of radial basis neurons and an output layer of linear neurons. A common choice for the basis function is a Gaussion given by the equation:
x − ci Gi ( x) = exp − 2σ 2
2
, i = 1,2,...m
(1)
Where Cirepresents the center of the basis function and denotes its width. The norm |||| in equation can be expressed by Euclidean distance. The weights and biases of each neuron in the hidden layer define the position and width of a radial basis function. Each linear output neuron forms a weighted sum of these radial basis functions. With the correct weight and bias values for each layer, and enough hidden neurons, a RBFN can fit any function with any desired accuracy. The advantage of the RBFN is its rapid learning, generality and simplicity. RBFN finds the input to output map using local approximators. It can be trained faster than BP and have none of BP's training problems such as saturation and local minima.
Fig. 1. The structure of the controller
Adaptive Controller Design of the Nonlinear Dynamic Systems
599
In the RBFN training stage we first observe the performance of the system with the approximate controller for a certain period of time and measure the range of error between the output of the plant and desired output. Then we divide this error span into certain sections and for each section we perform a perturbation test: We increased the input to the plant by e whenever total square error between the output of the plant and desired output falls within a specified region. If this change of the input results in a lower value of the total square error, we modify the output weight of the neural network controller to work accordingly. This action is continued for all sections and the whole process is repeated until no modification can reduce the error. Taking an overlapping Gaussian activation function for kernel units supposedly provides a smoother response and better generalization. but in our case the amount of interference was so high and we obtained a better performance with non-overlapping regions. Nevertheless, smoothness of the output can be enhanced by dividing the correction for each section by modifying the cost function used for training from J=?e2 to J=?(enew2+k(eneweold)2) for k<1. During the training stage, each time only one kernel unit responds and one weight is adjusted, This results in a shorter training time compared with Multi-layer Perceptron (MLP) type networks. Since weights are adjusted by a small value each time, the number of necessary iterations depends on the size of error or the accuracy of the approximate controller. The point that makes this method different is the way in obtaining the necessary corrections, which is based on the effect of weight perturbation on the total square error for a certain period of time. Perturbing the weights of a network is used in Madaline Rule III (MR III) training for analog implementation of neural networks. Because this method does not need prior knowledge about the transfer characteristics of the computing devices, it is not affected by the effects of neuron to neuron variations. Training the network based on its instantaneous result of the error will cause instability when used in a feedback loop. The performance of the system here is observed for a certain period of time and if any adjustment for any given amount of error increases the total square error, it would not be accepted. This proves the stability of the system for the class of inputs it has been trained for.
Fig. 2. Radial Basis Function Network
600
H.S. Cho
Generally optimization methods based on parameter perturbation are bound to failure when many parameters are involved in perturbation and that is because of the moving target effect of the other parameters. It is not the case in the proposed system for which only one or few related parameters are active at each time. At present, methods for directly adjusting the control parameters based on the output error are not available. The work presented in this paper is a step toward direct adaptive control.
3
Simulation Study
Different nonlinear systems were considered to evaluate the effect of the proposed method. Figure 3 gives the effect of controller when it is used for three different nonlinear plants. In this figure, y(.) denotes the plant output without controller, ym(.) denotes the desired output, r(.) denotes the reference input and yc(.) denotes the output of the plant with controller. An RBF type network with 30 kernel units each sensitive to different ranges of error was used. Input weights and activation function of the units were fixed. Output weights were perturbed and adjusted according to the effect of perturbation on the total square error for 400 sampling time. For a sinusoidal input (r(k)=sin(k/50)) and perturbation amount of =0.1, the amount of reduction in aggregated square error for a full cycle of input is example 1:from 3.69 to 1.33, example 2:from 105.84 to 26.44 and example 3:from 32.11 to 4.56. To observe the learning and generalization capability of the system, the controller was trained on a step response and its performance was observed on the sinusoidal input. Training the network with different types of input will enchance the generalization capability of the system. Example 1
1.2 y ( k − 1) + 0.5u (k − 1) y (k ) = 1 + y ( k − 1) 2 y m (k ) = 0.3 y m (k − 1) + 0.2 y m (k − 2) + 0.5r (k − 1) Example 2
y (k − 1) y (k − 2)( y (k − 1) + 2.5) + 1.5u (k − 1) y (k ) = 1.3 + y (k − 1) 2 + y (k − 2) 2 y m (k ) = 0.7 y m (k − 1) + 0.1y m (k − 2) + r (k − 1) Example 3 y ( k ) = sat (0.5 y ( k − 1) + 0.2 y ( k − 2) + 2.1u ( k − 1)) y m ( k ) = 0.48 y m ( k − 1) + 0.2 y m ( k − 2) + r ( k − 1) ( sat ( x) ≡ −3 +
6 ) 1 + exp(− x)
Adaptive Controller Design of the Nonlinear Dynamic Systems
601
Fig. 3. Ym(.) is the output of the reference model, Y(.) is the output of system without NNC and Yc(.) is the output of system without NNC in the example 1,2 and 3
4
Conclusions
A neural network controller for nonlinear plants is used in combination with an existing conventional controller, which removes the need for a generalized training scheme. The controller is guaranteed to perform stably for the class of inputs that it has been trained for. Using this method of control does not require assumption of a model for the plant and it makes it different from conventional control methodologies. Further -more, since training of this network does not require backpropagation of error, it makes direct adaptive control possible, a structure beyond the capabilities of backpropagation based on neural networks. The structure used here can be view as a
602
H.S. Cho
fuzzy controller implementation for which the control actions or rules which depend on the error between the plant output and the desired output are deduced in training time. It can also be viewed as a gain scheduling adaptive controller which can work for any unknown plant with no attempt to linearize the system at each region.
References [1] Gupta, M.M., Rao, D.H.: Dynamic Neural Units with Applications to the Control of Unknown Linear System. Journal of Intelligent and Fuzzy Systems 1(1), 73–92 [2] Wang, P., Kwok, D.P.: Optimal Design of PID Process Controllers Based on Genetic Algorithms. In: Proc. of the 12th Triennial World Congress of the IFAC, Sydney, Australia, pp. 193–197 (1993) [3] Varsek, A., Urbancic, T., Filipic, B.: Genetic Algorithms in Controller Design and Tunning. IEEE Trans. on Systems, Man, and Cybernetics 23(5), 1330–1339 (1993) [4] Satyadas, A., Krishnakumar, K.: GA-optimized fuzzy Controller for Spacecraft Attitude Control. In: Proc. IEEE Intl. Conf. Fuzzy System, pp. 1979–1984 (1994) [5] Linkens, D.A., Nyongesa, H.O.: Genetic Algorithms for Fuzzy Control: Part 1: Offline System Development and Application. In: IEEE Proc. Control Theory Appl., vol. 142(3), pp. 161–176. Addison-Wesley, US (1989)
Effects of Lotus Root Extract on Osteoblast and Osteoclast Sang Hyun Moh1, Tae Hoon Kang1,2, Suk Hyung Cho3, and Young Jun Kim4 1
Anti-aging Research Institute of BIO-FD&C Co., Ltd, Phytomecca Co., Ltd, Pilot plant 7-50, Songdo-Dong, Yeonsu-Gu, Incheon, Korea 406-840 2 Phytomecca Co., Ltd, Phytomecca Co., Songdo-Dong, Yeonsu-Gu, Incheon, Korea 406-840 3 Department of Fire Safety and Management, Hyejeon University, Hongsung-up Chungnam, Korea 350-800 4 Department of Cosmetic Science, Chungwoon University, Hongsung-up Chungnam, Korea 350-800 [email protected]
Abstract. Lotus root is a commonly used folk herb for the treatment of hypertension and stomach diseases. In the present study we investigated the effect of its extract on osteoblast and osteoclast in vitro using human osteoblastlike Saos-2 cells and primary mouse bone marrow-derived macrophages, respectively. Lotus root extract (LRE) stimulated proliferation of Saos-2 cells and increased their alkaline phosphatase (ALP) activity in dose- and timedependent manner. In primary mouse bone marrow-derived macrophages, LRE significantly reduced the number of tartrate‐resistant acid phosphatase (TRAP)‐positive multinucleated cells and decreased their TRAP activity. In addition, LRE inhibited bone resorption of differentiated osteoclast cells. These results suggest that LRE has the potential to prevent bone metabolic disorders such as osteoporosis. Keywords: Lotus root, osteoblast, osteoclast, alkaline phosphatase, tartrateresistant acid phosphatase.
osteoporosis is high in women after menopause [4], [5]. There are osteoporosis medicines that have developed and used to the present: for instance, bisphosphonate, selective estrogen receptor modulator, parathyroid hormone and other hormones [6], [7], [8]. However, they have various side effects and give a difficulty to take them. Thus, studies on other medicines to replace these are being actively performed focusing on natural extract [9], [10], [11], [12]. Lotus root is the rhizome of Nelumbo nucifera Gaertn., a perennial plant which belongs to Nymphaeaceae. It has been widely used for food and medicine due to its functions, such as tonics, promoting digestion, antioxidant, stopping bleeding, detoxification and sedation [13]. Main ingredient of lotus root is sugar, mostly starch. It also contains amino acids such as asparagine, arginine, tyrosine, etc and lecithin, pectin, tannin, mucin and vitamin B12 and more. According to recent studies, methanol extract of lotus root reduced blood sugar in a rat model of diabetes mellitus [14] and was effective to restrict inflammation [15]. Also, Mukherjee et al [16] reported that lotus root extract booted immunity. Yang et al [17] suggested that lotus root extract promotes development of nerve cells in hippocampus so that learning and memory ability are improved. However, there have been no studies focused on the effect of lotus root on bone metabolism so far even though lotus root has a variety of biological activities. This study aims to investigate the effects of lotus root extract on osteoblast and osteoclast. Saos-2 cell lines of human osteosarcoma were used to find the effect of lotus root extract on osteoblast. In order to find the effect of lotus root extract on osteoclast, undifferentiated precusor cells were separated from mouse bone marrow and processed by M-CSF and RANKL. Then, the cells were differentiated into osteoclast.
2
Materials and Method
2.1
Sample
Lotus root used in this study, produced in Korea, was purchased in Daeyoung Pharmaceuticals, Bucheon, Gyeonggi, 100g of lotus root was washed with distilled water. 500mL of 30% Ethanol was added to the lotus root. Then, condensation, vaccum evaporation and freeze drying are performed and the powdered lotus root is melt by dimethyl sulfoxide (DMSO, Sigma Chemical Co., St. Louis, MO, USA). 2.2
Incubation of Osteoblast
Human osteosarcoma cell lines, Saos-2 cells used in this study was from Korean Cell Line Bank (Seoul National University College of Medicine, Seoul, Korea). It was cultured under 5% CO2 at 37°C using RPMI1640(Gibco BRL, Grand Island, NY, USA) that 10% FBS(Gibco BRL, Grand Island, NY, USA) medium and 1% antibiotics(Gibco BRL, Grand Island, NY, USA) were added. The medium was changed once every three days and subcultured once a week.
Effects of Lotus Root Extract on Osteoblast and Osteoclast
2.3
605
Measurement of Proliferation Rate of Osteoblast
Mosmann method [18] was used to investigate the effect of lotus root extract on the proliferation of osteoblast: MTT assay was performed using 3‐(4,5‐dimethylthiazol‐ 2‐yl)‐2,5‐diphenyltetrazolium bromide(MTT, Sigma Chemical Co., St. Louis, MO, USA). Saos‐2 cells were seeded in a 96-well plate at 1 x 104 cells/well and incubated for 24 hours. After having the cells adhere to the well, lotus root extract was diluted in various concentration: 0, 1, 5, and 20 μg/mL. Then, they were incubated for 7 and 14 days for each concentration. Change of medium during incubation was performed for each concentration once every three days. After incubation, 0.05mg/ml of MTT was added to the cells and they were cultured for three more hours. Then, medium was removed, DMSO was added to resolve unsoluble formazan crystal. ELISA reader(BIOTEK Instruments Inc., Power wave X340, Winooski, VT, USA) was used to measure optical density at 540 nm and the result was converted into a percentage to control group. 2.4
Measurement of Activity of Alkaline Phosphatase
Sao-2 cells were seeded in a 96-well plate at 1 x 104 cells/well and incubated for 24 hours. After having the cells adhering to the well, lotus root extract was diluted in various concentration: 0, 1, 5, and 20 μg/mL. Then, they were incubated for 7 and 14 days for each concentration. The change of medium during incubation was performed for each concentration once every three days. After incubation, medium was removed and 200 μL of Triton X-100 was added to resolve the cells for 30 minutes at 37°C. Suspension was centrifuged and supernatant was collected. 10 μL of 8 mM p‐nitrophenyl phosphate(pNPP) was added to 100 μL of supernatant and it was incubated for 30 minutes at 37 °C. Then, 100μL of 0.1 N NaOH was added to discontinue the reaction and optical density was measured at 405 nm. ALP activity was drawn by the level of p‐nitrophenol(pNP) created from pNPP and standard graph on pNP. ALP activity per unit protein content was calculated by dividing protein content, considering the effect of difference in cell count on ALP activity. 2.5
Incubation of Bone Marrow Cells to Differentiate into Osteoclast
Femur and tibia of 5-week old ICR mouse were separated, cut off the both ends and washed the empty space of bones inside with syringe. Then, bone marrow cells were collected. The cells were cultured in α‐minimum essential medium(α‐MEM, Gibco BRL, Grand Island, NY, USA) that 10% FBS, 1% antibiotics, and M‐CSF(30 ng/mL; PeproTech Inc., Rocky Hill, NJ, USA) were added for 24 hours. Unadhering cells were collected and cultured for two more days. After incubation, adhering cells were collected and seeded in 96 well plate at 1×105 cells/well. Then, the lotus root extract was cultured with M‐CSF(30 ng/mL), RANKL(50 ng/mL; PeproTech Inc., Rocky Hill, NJ, USA) for three days by each concentration: 0, 1, 5 and 20 μg/mL.
Cells that culture was completed were fixed with 10% formalin for 10 minutes and reacted with TRAP staining kit(Sigma Chemical Co., St. Louis, MO, USA) for an hour at 37 °C. After the reaction is completed, the cells were counter stained with hematoxylin solution. The number of multinuclear osteoclast developed per well were measured considering multinuclear TRAP positive cells that had more than three nucleus per cell, regarded as osteoclast. 2.7
Measurement of TRAP Activity
The activity of TRAP which was known as specific marker of osteoclast was measured. To do this, the cells cultured in 96 well plate were washed with PBS and fixed with 10% formalin. Then, 100 μL of substrate solution(1.36 mg/ml 4‐nitrophenyl phosphate disodium salt, 10 mM tartrate, 50 mM citrate buffer) was added and cultured for 30 minutes at 37°C. After incubation, the reaction was discontinued by 0.1 N NaOH and optical density was measured at 405 nm. 2.8
Resorption Pit Assay
Osteoclast was cultured in OAASTM plate(Oscotec Inc., Cheonan, Korea) coated by calcium phosphate to investigate the effect of lotus root extract on the restriction of bone resorption in osteoclast. Additionally, it was cultured for 24 hours in the medium that contained lotus root extract (0, 1, 5 and 20 μg/mL). After incubation, medium was removed and 5% sodium hypochlorite(Sigma Chemical Co., St. Louis, MO, USA) solution was added to take cells out. Then, image‐analyzing software, Image Pro PlusTM(version 3.0, Media Cybernetics Inc., Bethesda, MD, USA) was used to measure absorbed area. 2.9
Statistical Analysis
The quantitative result obtained in this study was indicated with average value and standard deviation. Student's t-test was used for statistical analysis. If *p<0.05 or **p<0.01, it was considered that there was significance between control group and lotus root extract group.
3
Results and Discussion
According to recent studies, some of natural food that people regularly consume may help bone metabolism. Muhlbauer et al [19] stated that some vegetables in rat restricted bone resorption; especially, onion extract had the strongest effect for that. Also, Kim et al [20] reported that edible mushroom, leurotus eryngii extract increased the activity of osteoblast while it restricted that of osteoclast, which indicated that it had an effect on strengthening bone metabolism of rat that had ovariectomy. Besides,
Effects of Lotus Root Extract on Osteoblast and Osteoclast
607
there are many studies trying to find effective natural product on prevention and treatment of osteoporosis [21]. 3.1
Effect of Lotus Root on Proliferation and Activity of Osteoblast
It is known that osteoblast mostly synthesizing ingredients of bone matrix is originated from undifferenciated mesenchymal cells and plays an important role in calcification process through active metabolism [22]. Alkaline phosphatase (ALP) the glucoprotein enzyme exists in the cell membrane of osteoblast, which plays a role in transfer of inorganic phosphorus during calcification process and as a regulator of cell division and differentiation as well as a marker for measurement of activity of osteoblast [23]. This study was focused on the effect of lotus root extract on the proliferation and activity of osteoblast using Saos-2 cells, human osteosarcoma cell line. To examine the effect of lotus root extract on the proliferation of osteoblast, Saos-2 cells were treated by lotus root extract and cultured for 7 and 14 days separately and MTT assay was performed. The proliferation rate of osteoblast was indicated as the ratio to control group as shown in Fig.1.
Fig. 1. Effect of Lotus root extract (LRE) on the proliferation of human osteoblast‐like Saos‐2 cells The Saos‐2 cells were cultured for 7 and 14 days in the presence or absence of Lotus root extract (LRE). All values were presented as mean±SD (n=4). Values with asterisks are significantly different at *p<0.05 and **p<0.01 by Student's t‐test.
After 7 days of incubation, 1, 5 and 20 μg/mL of lotus root extract groups indicated proliferation rate, 111.7±1.9, 183.0±3.9 and 207.6±8.2%, respectively. This proliferation induction effect of lotus root extract on osteoblast was shown exceptional proliferation 14 days after treating lotus root extract: treatment group of lotus root extract with doses of 1, 5, and 20 μg/mL shows the proliferation rate, 310.3±40.4, 326.2±20.5 and 348.6±10.3%, respectively. It proved that lotus root extract promoted the proliferation of osteoblast. Lee et al [24] reported that ethanol
608
S.H. Moh et al.
extract of Asparagi Radix significantly promoted the proliferation of Saos-2 cells within the range of 1-20 g/mL, which is similar to the level of lotus root extract in this study. To examine the effect of lotus root extract on osteoblast, the activity of ALP, a marker of osteoblast was measured. ALP activity of Saos-2 cells treated by lotus root extract was indicated as the ratio to the control group as shown in Table 1. In 7 days of incubation after treatment, lotus root extract group with doses of 1, 5, and 20 μg/mL shows ALP activity, 255.0±9.8, 409.6±35.4 and 761.5±261.1%, respectively. Thus, we found that lotus root extract not only induced the proliferation of osteoblast, but also increased the activity. However, further studies is necessary to find if the effect of lotus root extract on the promotion of proliferation and activity of osteoblast was whether by single material of active ingredients or by complicated synergism. Additionally, it is considered necessary to find the correlation between lotus root carboxyl extract and biomarker such as osteocalcin, osteopontin and procollagen terminal extension peptide, except ALP.
Ⅰ
Table 1. Effect of Lotus root extract (LRE) on the alkaline phosphatase activity of human osteoblast-like Saos-2 cells LRE (
ALP activity (% of control)
㎍/mL)
Control
1
5
20
7 days
100±2.4
255±9.8a
409.6±35.4b
761.5±261.1
14 days
100±4.2
401.7±16.2
1289.9±47.8c
1682.5±74.5d
The Saos‐2 cells were cultured for 7 and 14 days in the presence or absence of Lotus root extract (LRE). All values were presented as mean±SD (n=4). Values with different letters are significantly different at p<0.01 by Student's t‐test.
3.2
Effect of Lotus Root Extract on the Differentiation and Activity of Osteoclast
Osteoclast, a multinuclear cell, that resorbs mineralized bones is known to be differentiated from mononuclear cells/ macrophageprecursor cells originated from hematopoietic stem cells. Cytokines released from osteoclast, such as macrophage colony‐stimulating factor(M‐CSF) and receptor activator of nuclear factor‐kB ligand(RANKL) are known to participate in the differentiation process [25], [26], [27]. Also, osteoclast has activity of tartrate‐resistant acid phosphatase(TRAP), actively produces acids and absorbs bone matrix by actin ring formation [28]. It is known that if the proliferation and activity of osteoclast excessively increased due to any problem in this regulation process, bone metabolism diseases such as osteoporosis may occur. For these reasons, studies on development of medicine for osteoporosis have been mostly focused on the restriction of differentiation and activation of osteoclast. Thus, this study investigated if lotus root extract was effective for restriction of differentiation and activation of osteoclast. To do this,
Effects of Lotus Root Extract on Osteoblast and Osteoclast
609
mouse bone marrow cells were cultured into osteoclast precursor in the medium containing M‐CSF and RANKL. Only TRAP positive multinuclear cells, with more than three nucleus were regarded as osteoclast. To examine the effect of lotus root extract on the differentiation of osteoclast, Saos-2 cells were treated with lotus root extract by each level of concentration and cultured for three days. Then, the formation of TRAP positive multinuclear cells was observed. In control group, 375.3 TRAP positive multinuclear cells were observed while in the lotus root extract treatment group with 1, 5 and 20 μg/mL, TRAP positive multinuclear cells observed were 255.7, 66.0 and 51.0, respectively (Table 2). Table 2. Effect of Lotus root extract (LRE) on RANKL-induced osteoclast differentiation LRE (
TRAP-positive multinucleated cells number
㎍/mL)
Blank
Control
1
5
20
35.5±3.9
375.3±7.4
255.7±5.3a
66.0±4.2b
51.0±2.8c
Bone marrow macrophages were cultured for 3 days with M‐CSF (30 ng/mL) and RANKL (50 ng/mL) in the presence or absence of Lotus root extract (LRE). TRAP‐positive cells were counted as osteoclasts. All values were presented as mean±SD (n=4). Values with different letters are significantly different at p<0.01 by Student's t‐test.
Fig. 2. Effect of Lotus root extract (LRE) on TRAP activity of RANKL‐induced osteoclasts Bone marrow macrophages were cultured for 3 days with M‐CSF (30 ng/mL) and RANKL (50 ng/mL) in the presence or absence of Lotus root extract (LRE). Data are expressed as the percentage of control. All values were presented as mean±SD (n=4). Values with asterisks are significantly different at *p<0.05 and **p<0.01 by Student's t‐test.
Alice et al [29] reported that flavonoid ingredients such as quercetin, kaempferol and tannin induced apoptosis of osteoclast and restricted its differentiation. Thus, the
610
S.H. Moh et al.
differentiation restriction effect of lotus root extract on osteoclast is considered due to the ingredients largely contained in lotus root such as kaempferol and tannin. To find the effect of lotus root extract on the activity of osteoclast, Saos-2 cells were treated by each level of concentration, cultured for three days and their TRAP activity was measured. TRAP activity was indicated by the ratio to the control group as shown in Fig. 2. In lotus root extract treatment group with doses of 1, 5 and 20 μg/mL, TRAP activity was shown as 96.9±0.8, 23.8±1.5 and 10.6±0.1%, respectively. The result indicated that TRAP activity was restricted depending upon the level of lotus root extract. 3.3
Effect of Lotus Root Extract on the Restriction of Bone Absorption Function of Osteoclast
One of the remarkable features of osteoclast is that the activated osteoclast functionally reabsorbs the surface of mineralized bones. Thus, this study examined the effect of lotus root extract on bone resorption using OAAS™ plate coated by calcium phosphate. Bone absorption was indicated as the ratio to the control group as shown in Table 3. The lotus root extract treatment group of 1, 5, and 20 μg/mL shows bone absorption of 13.4±1.7, 5.4±0.6 and 1.0±0.1%, respectively. This proved that lotus root extract was effective to the restriction of bone resorption as well as that of differentiation and activity of osteoclast. These results indicate that lotus root extract promoted the proliferation and activity of osteoblast and restricted the differentiation and activity of osteoclast. Thus, lotus root is considered to have a potential to be developed as functional food to prevent bone metabolism disease. As mentioned above, however, further studies are considered necessary to investigate active ingredients and action mechanism of lotus root extract. Table 3. Effect of Lotus root extract (LRE) on bone resorption LRE (
Resorption area (% of control)
㎍/mL)
Blank
Control
1
5
20
0.8±0.1
100±5.7
13.4±1.7a
5.4±0.6b
1.0±0.1c
Mature osteoclasts were seeded on OAAS™ plate and incubated for 24 hr. Resorption pit areas were quantified using the image Pro‐plus program, version 3.0. Data are expressed as the percentage of control. All values were presented as mean±SD. Values with different letters are significantly different at p<0.01 by Student's t‐test.
References 1. Kalervo, V., Haibo, Z., Mika, M., Jussi, M.H.: The cell biology of osteoclast function. J. Cell Sci. 113, 377–381 (2000) 2. Suda, T., Nakamura, I., Jimi, B., Takahashi, N.: Regulation of osteoclast function. J. Bone Miner. Res. 12, 869–879 (1997)
Effects of Lotus Root Extract on Osteoblast and Osteoclast
611
3. Canalis, E., McCarthy, T., Centrella, M.: Proliferation factors and the regulation of bone remodeling. J. Clin. Invest. 81, 277–281 (1988) 4. Ryan, P.J., Evans, P., Gibson, T., Fogelman, I.: Osteoporosis and chronic back pain: A study with single-photon emission computed tomography bone scintigraphy. J. Bone Miner. Res. 7, 1455–1460 (1992) 5. Jilka, R.L.: Cytokines, bone remodeling and estrogen deficiency. Bone 23, 75–81 (1998) 6. Reginster, J.Y.: Treatment of bone in elderly subjects: calcium, vitamin D, fluor, bisphosphonates, calcitonin. Horm. Res. 3, 83–88 (1995) 7. Rodan, G.A.: Emerging therapies in osteoporosis. Ann. Rep. Med. Chem. 2, 275–285 (1994) 8. Stock, J.L.: Drug therapy in osteoporosis, in diagnostic and therapeutic principles. In: Rosen, C.J. (ed.), Totowa (1996) 9. Cryer, R., Bauer, D.C.: Oral bisphosphonates and upper gastrointestinal tract problem: what is the evidence? Mayo. Clin. Proc. 77, 1031–1043 (2002) 10. Recker, R.P.: Current theraphy for osteoporosis. J. Clin. Endocrinol. Metab. 76, 14–16 (1993) 11. Ochiuto, F., Pasquale, R.D., Guglielmo, G., Palumbo, D.R., Zangla, G., Samperi, S., Renzo, A., Circosta, C.: Effects of phytoestrogenic isoflavones from red clover (Trifolium pretense L.) on experimental osteoporosis. Phytother. Res. 21, 130–134 (2007) 12. Lee, J.W., Lee, I.S.: Effects of Rubus coreanus Miquel extracts on the activity and differentiation of MC3T3-E1 osteoblastic cell. J. Life Sci. 14, 967–974 (2004) 13. Bae, K.H.: The medicinal plants of korea. In: Hwang, J.S. (ed.), Seoul (2007) 14. Mukherjee, P.K., Saham, K., Paml, M., Saham, B.P.: Effect of Nelumbo nucifera rhizome on blood sugar level in rats. J. Ethnopharmacol. 58, 207–213 (1997) 15. Mukherjee, P.K., Saham, K., Dasm, J., Paml, M., Saham, B.P.: Studies in the antiinflammatory activity of rhizomes of Nelumbo nucifera. Planta. Med. 63, 367–2010 (1997) 16. Mukherjee, D., Khatua, T.N., Venkatesh, P., Saha, B.P., Mukherjee, P.K.: Immunomodulatory potential of rhizome and seed extracts of Nelumbo nucifera Gaetn. J. Ethnopharmacol. 128, 490–494 (2010) 17. Yang, W.M., Shim, K.J., Choi, M.J., Park, S.Y., Choi, B.J., Chang, M.S., Park, S.K.: Novel effects of Nelumbo nucifera rhizome extract on memory and neurogenesis in the denate gyrus of the rat hippocampus. Neurosci. Lett. 443, 104–2010 (2008) 18. Mosmann, T.: Rapid colorimetric assay for the cellular proliferation and survival; application to proliferation and cytotoxic assay. J. Immun. Methods. 65, 56–58 (1983) 19. Muhlbauer, R.C., Lozano, A., Reinli, A.: Onion and a mixture of vegetables, salads, and herbs affect bone resorption in the rat by a mechanism independent of their base excess. J. Bone Miner. Res. 1, 1230–1236 (2002) 20. Kim, S.W., Kim, H.G., Lee, B.E., Hwang, H.H., Baek, D.H., Ko, S.Y.: Effects of mushroom, Pleurotus eryngii, extracts on bone metabolism. Clin. Nut. 25, 166–170 (2006) 21. Seo, B.I., Ku, S.K., Cha, E.M., Park, J.H., Kim, J.D., Choi, H.Y., Lee, H.S.: Effect of Morindae radix extracts on experimental osteoporosis in sciatic neuroectomized mice. Phytother. Res. 19, 231–238 (2005) 22. Laurence, D., Maylis, D., Philippe, J., Dominique, H.: Embryonic stem cells: new tool to study osteoblast and osteoclast differentiation. Stem Cells 25, 544–552 (2007) 23. Stein, G.S., Lian, J.B., Owen, T.A.: Relationship of cell proliferation to the regulation for tissue specific gene expression during osteoblast differentiation. FASEB. J. 4, 3111–3123 (1990)
612
S.H. Moh et al.
24. Lee, S.Y., Kim, S.N., Kim, J.K.: Effects of Asparagus cochinchinensis on the stimulation of osteoblast differentiation and inhibition of osteoclast generation. J. Korean Soc. Food Sci. Nutr. 37, 16–19 (2008) 25. Teitelbaum, S.L.: Bone resorption by osteoclasts. Science 289, 1504–1508 (2000) 26. Takayanagi, H.: Mechanistic insight into osteoclast differentiation in osteoimmunology. J. Mo. l Med. 83, 170–179 (2005) 27. Lacey, D.L., Timms, R., Tan, H.L., Kelly, M.J., Dunstan, C.R., Burgess, T., Elliot, R., Colonboro, A., Elliot, G., Scully, S.: Osteoprotegerin ligand is a cytokine that regulates osteoclast differentiation and activation. Cell 93, 165–176 (1998) 28. Mundy, G.R., Roodman, G.D.: Osteoclast ontogeny and function. In: Peck, W.A. (ed.), Amsterdam (1987) 29. Alice, W., Said, K., Romuald, M., Florence, L., Christophe, P., Petit, J.P., Patrice, F., Michel, B.: Potent inhibitory effect of naturally occurring flavonoids ruercetin and kaempferol on in vitro osteoclastic bone resorption. Biochem. Pharmacol. 65, 35–42 (2003)
Park, Byoung-Kwon 317 Park, Chang-Hyun 317 Park, Chung-Yill 129 Park, Hyoung-Keun 187 Park, Hyung-sung 286 Park, Jae-Pyo 195 Park, Jea-Bum 102 Park, Jeong-Hyo 138 Park, Ji-Hye 67 Park, Ji Won 271 Park, Jong-Cheon 228 Park, Jung-hwan 286, 392 Park, Keon-Jun 438 Park, Sung-Hoon 25 Park, Tae-su 392 Park, Young-Jae 317 Parthiban, Rajendran 161 Pitchai, Calduwel Newton 554 Rho, Dae-Seok 59, 87, 95, 102, 110 Ruan, Li 445 Ryu, In Ho 591 Ryu, Kyung-Sang 87 Ryu, Min-Woo 42 See, Simon Chong Wee 351 Shaikh, Abdul Khalique 161 Shen, Jian 102 Shin, Jin 8, 49 Sim, Kwang Mong 544 So, Byung Moon 148, 361 Son, Joonho 59 Song, Bangwoon 95 Song, Inseong 238 Song, Je-Ho 335
Author Index Song, Keun-Bae 67 Song, Suk-hwan 87 Sung, Haeng Nam 300 Tsai, Min-Yi
