Andreas Pitsillides

1. Andreas Pitsillides Congestion Controlin communication networks: A control theory based approach Department of Computer Science University of Cyprus 23-04-2007 http://www.NetRL.cs.ucy.ac.cy

2. Talk outline • preliminaries • sketch own areas of research – focus on CC • causes and effects of congestion and identification of control difficulties • networks as complex dynamical systems and control theory approach • outline networking trends and CC techniques over last 20 years; own contributions • two illustrative examples of our approach: FEM/FIO and ACP • overview of contributions in other areas of networking research • concluding remarks and future work

3. A recent remark ‘Networks are very complex. Do not kid yourselves otherwise.’ Debasis Mitra, Senior VP Research, Bell Labs Panel discussion at Infocom 2001 Modelling of the Shrew (beast): Quest for a ‘Model’ Network Model(Organiser: Ariel Orda)

4. Rapid changes in Networking technologies Networking technology and techniques are in constant state of change bits/ sec Gbits /sec

5. Rapid changes in Networking technologies

6. 4. Research Networking e-Infrastructures Supports own research activities, plus wider research community Research challenges we have contributed 1. Congestion control and flow control (CC) 2. Resource Allocation (RA) 3. Networked Applications: e-services and Telehealth Care

7. 1. Congestion control and flow control (CC)

8. CC for Packet-based Networks • Active research area for over 30 years • Still remains one of most importantresearchchallenges • demands on Internet, e.g. QoS, multiservice, multimedia (existing TCP only Best Effort/Elastic traffic) • high speed internet (existing TCP does not scale well) • wireless networks (existing TCP does not perform well) • new networking techniques / technologies: diff-serv, peer-to-peer, ad-hoc and sensor networks, etc… • A multi-faceted problem with many different approaches, not only on how we affect control, but also on the control structure, as well as the theoretical development • even a definition cannot be universally accepted by research community • Internet Congestion Control Research Group, ICCRG, debate in 2006 • an Internet Draft on metrics only now issued • Transport Modeling Research Group, TRMG, March 07: submitted to IRTF to be considered as Informational • research challenges are still debated • ICCRG February 2007

9. Causes and effects of congestion • To control congestion we must understand: • what is that we are trying to control and how we can model and quantify it (what metrics) • how and where we can control it • Quantifying and modelling congestion not easy • congestion is caused by constrained resources • felt by users as • delay • loss • throughput • But these metrics are subjective and conflicting, as seen by applications • If we are going to use feedback • how and where can we sense and control it? • what type of feedback? Inherent control difficulty: control elements are also within network system FB control system elements

10. Different approaches toward CC • diverse range of control approaches adopted • open loop, closed loop, proactive, reactive, adaptive • flow control, admission control, bandwidth control, routing • control structure • different feedback schemes and combinations and placement of CC elements • end-to-end, network assisted (AQM) • implicit, explicit, single-bit (e.g. ECN), multi-valued (ER) • at sender, at sender+receiver, at sender+receiver+routers • diverse range of mathematical tools adopted • From fuzzy logic to game theory, from feedback control theory to utility functions and linear and nonlinear programming • focus only on feedback based control approaches, regulating flow from sources

11. Early approaches • early methods for CC were not very effective. • open loop approaches • lack scalability and robustness properties • small inaccuracies in their model could easily lead to deterioration of performance and even instability • engineering intuition and ad-hoc techniques (often non-linear, open-loop or feedback-based) • some work remarkably well (e.g. TCP CC) • but little is known about their behaviour at large, and if and when they may fail

12. Our view: Networks behave as complex dynamical systems • Internetis largest and most complex artificially deployed system • possesses similar structural properties to other complex systemspervading science • heterogeneous subsystems (sources and routers) performing complex functions • interconnected by heterogeneous links (wired, wireless), often incorporating complex dynamics themselves • many other factorscontributeto immense complexity • large-scale and size • fragmented nature of underlying infrastructure • semi-hierarchical organization • extreme heterogeneity - diverse network technologies and communication services • distributed management of resources • complex structures which arise in implementation of the layered protocols

13. Our view: Networks behave as complex dynamical systems • identified control difficulties in networks include • non-linear time varying dynamics • non-linear interacting subsystems • large delays in relation to system time scales • tight control demands • largeness of network system • lack of appropriate measurements • placement of controls and lack of adequate local controls • uncertain feedback path • uncertain responsiveness of the system in-build control structures were never engineered in the network system

14. Our view: Networks behave as complex dynamical systems • introduced advanced control theoretic techniques, exploiting dynamic nature of network to come up with effective solutions • solid theoretical foundation in effectively dealing with complex dynamical systems • remarkable success in many diverse systems • demonstrated • via analysis and simulation that despite reported control difficulties, control theory provides effective control with tight QoS provision and good properties • proposed control theoretic techniques are effective in delivering efficient solutions in highly complex network systems, such as ATM and Internet, outperforming existing approaches

15. CC design objective • goal of designed CC protocols is to guide network to a ‘stable’ equilibrium • ‘tight’ QoScontrol featuring • high utilization • tolerable queue sizes • tolerable packet drops • good transient and steady state response • smooth responses with no or minimal oscillations • fast convergence • settles to fair allocations (max-min sense) • robust and scalable with respect to changing network dynamics • bandwidths and delays • number of users • network size • simple to implement, effective design

16. Performance control in new networking topologies and techniques Large-scale Networks Mathematical Formulation ‘frenzy’ Global stability results Internet Congestion Collapse ‘Reality check’ Turn to internet TCP Congestion Control New approach to CC for high speed internet. Robust control, Network Assisted, Multivalued Feedback, e.g. (XCP, ACP) Early to mid 2000’s Mid 2000’s Complex systems and networks and Nature inspired techniques perspective Early 1990’s ATM CC – from open loop controls to full network assisted feedback Per connection QoS support Late 1990’s early 2000’s QoS drive with IntServ and Diff-Serv architectures Jacobson TCP CC implicit feedback , end systems involved in control, late 1980’s past research is on fundamental design issues, little research on absolute performance in WSNs Back to Internet. Existing TCP limitations start to surface. Some attempts to introduce AQM (RED, ECN), RTP/RTCP protocols for real time services Mid-to-late 1990’s Approaches include control theoretic and resource optimisation. Fairness, Global view and important stability results. Early 2000’s Self-properties in large scale networks, robustness, and resiliency Highspeed Internet demands and Wireless, including ad-hoc and sensor networks Internet ‘QoS’ (AQM, Int-Serv, Diff-Serv) ATM ‘Era’ QoS ‘frenzy’ Our historical ‘CC lane’ Mid to Late 2000’s Early to late 1980’s

17. CC contributions 2006 1990 Large-scale Networks Mathematical formulation ‘frenzy’ global stability results Internet Congestion Collapse Performance control in Sensor networks ‘Reality check’ Turn to internet TCP Congestion Control expand work in related areas guide local deigns for better global performance scalability late 1990’s Early 1990’s mid 2000’s Early to late 1980’s Mid to Late 2000’s mid 1990’s Early 2000’s expand work in related areas effectively control the local loop Simple and robust Resiliency and robustness in large scale communication and other networks (complex systems, nature inspired techniques, transportation networks) Highspeed Internet demands and Wireless, ACP, estimate number of users AQM, Internet ‘QoS’ (Int-Serv, Diff-Serv) IDCC, FEM, FIO ATM ‘Era’ CC: Predictive Adaptive Fuzzy FERMNon-linear Contributed in many areas. Focus on two of our later research contributions new work

18. 2. Resource Allocation

19. Resource Allocation • Effective and fair resource allocation is necessary in any constrained system • In our study of resource allocation we provided solutions to a number of problems • BW allocation in ATM/MPLS • RRMs in UMTS, mainly for MBMS • capacity planning in UMTS • MBMS CAC • MBMS UE counting • MBMS Handover • RAN selection in 4G • Wireless Networks • Streaming video • Mobile-IP and mobility events • extended ns-2 network simulator to build a system level simulator for UMTS (over 70 download requests) • MBMS OPNET based simulator

20. Control structures Hierarchical BW allocation Sensor Networks VANETS Higher layer only UMTS MBMS Handover Hybrid CAC Power Counting Genetic Algorithm based Aggregated BW Allocation Multilevel optimal BW control (VP level) UMTS system level simulators Mobile Nets Video distribution and Streaming, Handover: MIPv6 HMIPv6-MPLS RA in 4G RAN selection in 4G Networks using Game Theory Sensor, VANETs Resource Allocation contributions All layers Late 1990’s Mid 2000’s Early 1990’s Early 2000’s Middle layers ATM Wireless and Mobile Networks (3rd and 4th G)

21. 3. Applications: e-services and Telehealth Care

22. 3. Applications: e-services and Telehealth Care • research challenges are driven by application demands • thorough understanding essential for designing next generation networks and protocols • some generalised work to study m-services • new methodology for designing and developing m-Commerce services and applications • develop roadmap for m-Business research • main focus is on e-health • only highlight DITIS, an award winning research project which is being commercialised K. G. Fouskas, G. M. Giaglis, S. Karnouskos, P. E. Kourouthanassis, A. Pitsillides, M. Stylianou, A Roadmap for Research in Mobile Business, International Journal of Mobile Communications (IJMC), Volume 3 - Issue 4 – 2005, pp. 350-373. A. S. Andreou, C. Leonidou, C. Chrysostomou, A. Pitsillides, G. Samaras, C. N. Schizas, S. M. Mavromoustakos, Key issues for the design and development of mobile commerce services and applications, International Journal of Mobile Communications (IJMC), Volume 3 - Issue 3 - 2005, pp. 303 – 323.

23. m-health: DITIS • DITIS: Networked Collaboration supporting Home Healthcare Teams • focus on mobile workers in home-based care • provide collaboration, secure, easy and timely exchange of information, and coordination of team activities • novel part introduction of virtual healthcare teams • introduces new and more specialized requirements and network dependant computational models • Accomplishments include • finalist 2003 eHealth Ministerial Conference • 7th in 2003 World Summit Award competition • paperless operation at PASYKAF • commercilaisation • funded for six years: RPF, Microsoft Cambridge Research Labs, HS24 e-TEN, LinkCare e-TEN • Current research incorporate sensor networks for unobtrusive monitoring patients in home environment (MPOWER) A. Pitsillides, G. Samaras, B. Pitsillides, D. Georgiades, P. Andreou, E. Christodoulou, Virtual Collaborative Healthcare Teams for Home Healthcare, Journal of Mobile Multimedia, special issue on Advanced Mobile Technologies for Health Care Applications, Vol.2, No.1, 2006, pp. 023-036.

24. 4. Research Networking e-Infrastructures

25. 4. Research Networking e-Infrastructures • building research e-Infrastructures for new networking technologies • essential for good understanding of research issues and experimentation, especially for research students • Since mid 90’s EC initiated number of projects to build • Pan European Network for research & education through NRENs • GEANT and GEANT2 initiated as part of 5th & 6th FP to provide • gigabit data transfer speeds • offers capabilities to build advanced networking and applications • CYNET, Cyprus NREN, established by council of ministers in 2000. [co-founder and chairman] • As part of Networks Research Laboratory • CISCO and LINUX-based networking equipment • fixed (IPv4/IPv6 and Diff-Serv) and mobile/wireless networks • video streaming • sensor network for performance control/QoS support for real time services and mobility

26. Contributions

27. Contributions • research activities focused on design and performance aspects of networks. In particular • introduced advanced control theory and nonlinear dynamics tools • solve challenging problems related to CC problems • provided solutions to open resource allocation and Radio Resource Management (RRM) problems • contributed to Internet technologies and their application in Mobile e-Services • e-Healthcare and security issues • My work led to 180 scientific publications, which include • 1 edited book • 4 editorials • 12 book chapters • 22 journal papers • 100 refereed conference papers • 41 other conference papers

28. Contributions • funding • European Commission IST program • National Research Promotion Foundation (RPF) • Cambridge Microsoft Research Labs • University of Cyprus • Swinburne University of Technology • Australian Government research grants board • total funding at UCY exceeds 9.5 million Euro • Funding obtained as PI or co PI > 3.09 million Euro • Funding obtained as Participant > 1.81 million Euro • Funding obtained through CYNET (support advanced networking infrastructure and networking services) > 4.75 million Euro • established Networks Research Laboratory and graduated 4 PhDs and 18 Masters students • lab currently supports 3 postdocs, 10 PhD and 7 Master’s students • collaborating with colleagues and their students, from United States, Australia, and European countries

29. Current and future research directions • Congestion and overload control • remains challenging, both from a theoretical, as well as a practicalperspective • CC in power constrained sensor networks and intermittently connected wireless networks (e.g. VANETS) largely unexplored. • global Stability of max-min CC schemes in networks of arbitrary topology incorporating delays and queuing dynamics is an open challenging research problem. • new topologies place many demands: ad-hoc and sensor networks. • advanced control techniques play a significant role in coming up with practical solutions to congestion and overload control and plan to continue making contributions in this area • Resource allocation and RRM • continue our work in Mobile Networks • video streaming over wireless and mobile networks with tight QoS support, combining network and user adaptation strategies • 4G continue work on connection admission to competing radio networks within game theoretic setting • new technologies, such as WiMax, pose many RA problems • new ideas, like mobile terminals acting as sensors

30. Current and future research directions • sensor networks andVANETS • Pose new research challenges, as compared to traditional wired Internet • Performance control in critical environments, • understand fundamentals of congestion and overload • Cross layer feedback • Self-organising abilities, co-operation, and adaptability • High mobility and power constraints often lead to intermittently connected networks and necessitate the development of delay tolerant applications • networked control systems, complex adaptive networks andnature inspired techniques • networked control systems such as Intelligent transportation systems and coordinated vehicles, pose major research challenges. • developing mathematical foundations to design and evaluate such systems can lead to a unified theory of complex adaptive systems. • will help explore other types of networks, such as biological networks, financial networks, social networks. • lessons learned can reveal new nature inspired techniques useful in designing robust, adaptable and survivable engineering networks. • mobile e-health applications • commercialisation of DITIS and its adoption in the field pose new opportunities to identify new areas of research • sensor technologies for monitoring of vital signals and patient well being

31. MANY THANKS to all NetRL people and collaborators previous students at UCY and elsewhere NetRL Andreas Pitsillides (Director) Vasos Vasiliou (Associate Director) Marios Lestas (PhD) Chrysostomos Chrysostomou (PhD) George Papageorgiou (PhD) Josephine Antoniou (PhD cand) Christoforos Christoforou (PhD Cand) Pavlos Antoniou (PhD Cand) Eliana Stavrou (PhD cand) Yiannos Mylonas (PhD Cand) George Hadjipollas (PhD Cand) Andreas Xeros (PhD cand) Andreas Panayides (PhD cand) Panayiotis Andreou (PhD Cand) Marios Koutroulos (PhD Cand) Marinos Stylianou (MSc) Eleni Themistocleous (MSc) Zenonas Zenonos (MSc) Antonis Antoniou (MSc) Haralambos Segiou (MSc) Christos Charalambous (MSc) Nicos Ioannou (MSc) George Samaras (CS, UCY) Costas Pattichis (CS, UCY) Andreas Andreou (CS, UCY) Marios Dikaiakos (CS, UCY) Marios Polycarpou (ECE, UCY) Christos Panayiotou (ECE, UCY) Maria Andreou (ECE, UCY, visiting) Petros Ioannou (USC, USA) Ahmet Sekercioglu(Monash Uni., Australia) N. Binucci, (3G, UK) G. Ramamurthy (NEC Research Labs, USA) Tran Gia, S. Kohler (Uni.Wurzburg, Germany) Athanasios Vasilakos (FORTH, Greece) Niovi Pavlidou (AUTh, Greece) David Tipper (Uni. of Pittsburgh, USA) Greg Egan (Monash Uni., Australia) Bob Warfield, (Telecom Res. Labs, Australia) Many European partners in EC funded projects

32. End

33. Fuzzy slides

34. Fuzzy Logic based AQM Congestion Control in TCP/IP Networks A. Pitsillides, A. Sekercioglu, G. Ramamurthy, "Effective Control of Traffic Flow in ATM Networks Using Fuzzy Explicit Rate Marking (FERM)", IEEE Journal on Selected Areas in Communications (JSAC), Volume 15, Issue 2, February 1997, pp. 209-225. C. Chrysostomou, A. Pitsillides, L. Rossides, M. Polycarpou, A. Sekercioglu, “Congestion Control in Differentiated Services Networks using Fuzzy-RED”, Special Issue on "Control Methods for Telecommunication Networks" in IFAC Control Engineering Practice (CEP) Journal, Vol. 11, Issue 10, pp. 1153-1170, September 2003. C. Chrysostomou, A. Pitsillides, Fuzzy Explicit Marking: A Unified Congestion Controller for Best Effort and Diff-Serv Networks, submitted.

35. “Quote 1” ‘So far as the laws of mathematics refer to reality, they are not certain. And so far as they are certain, they do not refer to reality.’ Albert Einstein

36. “Quote 2” ‘As complexity rises, precise statements lose meaning and meaningful statements lose precision.’ Lotfi Zadeh (“father” of Fuzzy Logic)

37. Precision and significance in real world

38. Our FLC-based Research Study • investigated the complex, but challenging, concepts of • ATM QoS aware flow control • TCP/AQM (Active Queue Management) CC • BE (best-effort) • Diff-Serv (differentiated services) environments • FOCUS OF PRESENTATION IS TCP/AQM CC

39. Problem Statement • Main Aim: provide effective control for high link utilization with low loss and queuing delay • focus on AQM mechanisms with ECN support, thus keeping TCP’s CC mechanisms unchanged • Internet standards track protocol RFC3168 • most routers support ECN • furthermore, address Diff-Serv congestion controlat core for aggregated QoS support

40. Motivation • Current CC solutions based on AQM/ECN areineffective to meet diverse needs of today’s Internet • they have serious limitations and drawbacks, as identified in literature • Extremely difficult for traditional modeling techniques to capture the network’s essential dynamics • even if they do resulting model is overly complex • Common approach in classical control theory is to • ignore such complex parameters in mathematical model • simplify model, often making overly conservative with restrictive stability bounds • Given need for effective control methodology • tocapture system behaviour under widely differing operating conditions • investigate usefulness of fuzzy logic control to meet such objectives

41. Why Use Fuzzy Logic Control • Fuzzy Logic Control • particularly appealing in nonlinear complex systems where • satisfactory analytic models are impractical • their behavior is well understood and can be captured by linguistic models • Has solid theoretical foundation (at times controversial) • achieves ‘inherent’ robustness and reduces design complexity • Part of what is termed Intelligent Control, or Computational Intelligence Control • Remarkable success demonstrated in research literature and commercialproducts in many diverse disciplines

42. Contributions • Offer significant improvements • achieve high link utilization and regulated queues • fast system response and robustness to varying system dynamics (differing topologies and traffic conditions) • in Dif-Serv, adequate and effective differentiation among different priority classes • Demonstrate that Fuzzy Logic based AQM control methodology better handles nonlinearities and dynamics, in contrast with existing, well-known, conventional counterparts

43. Proposed Mechanisms: BE • Proposed AQM scheme for BE (Best-Effort) environments • Fuzzy Explicit Marking (FEM) • regulates queues of IP routers at predefined levels • by achieving a specified target queue length (TQL) • in same spirit as RED • Fuzzy inference engine (FIE) operates on router buffer queues

44. Proposed Mechanisms: BE (cnt’d) • Feedback system model of FEM queue length Mark probability desired queue length Feedback signal • based on two network state inputs • error on instantaneous queue length for two consecutive sampling intervals like RED FEM dynamically calculates mark probability p(kT)

45. deep structure surface structure Membership functions linguistic values • Knowledge-base (linguistic rules) generated from IF-THEN control rules, e.g.: • IF e(kT) is NVB AND e(kT – T) is NB, THEN p(kT) is H • IF e(kT) is PVB AND e(kT – T) is PB, THEN p(kT) is Z Proposed Mechanisms: BE (cnt’d) • System model of FEM • mark probability responsive due to human reasoning and inbuilt nonlinearity

46. Proposed Mechanisms: Diff-Serv • Diff-Serv Fuzzy Logic Control Design (FIO) • goal to achieve same performance as BE • provide effective congestion for Diff-Serv, plus • differentiated treatment of traffic aggregates • built onfuzzy controllerdesigned for BE environments • two identical FEM controllers used • one for each differentiated traffic aggregate (FIO – FEM In-Out) • high-priority (low drop precedence / In packets) • low-priority (high drop precedence / Out packets) hence offering (differentiated) QoS in traffic aggregates

47. Proposed Mechanisms: Diff-Serv (cnt’d) • Two different TQLs, one for each FEM controller TQL for low-priority < TQL for high-priority • objective:regulate queue at TQLlow, where mark probability for high-priority traffic is closer to zero • but, if high-priority traffic >> low-priority traffic, at least regulate queue at TQLhigh • not enough low-priority traffic to maintain TQLlow • in this case mark probability for low-priority traffic is closer to 1 • accomplish both differentiation and bounded delay, by regulating queue between two TQLs, depending on network traffic high- and low-priority traffic share a FIO queue

48. Simulative Performance Evaluation • Use extensive simulative evaluation to demonstrate effectiveness and robustness, in both BE and Diff-Serv environments • Comparison made with other published results with well-known, AQM schemes • A-RED, PI, REM, and AVQ for BE networks • RIO and TL-PI for Diff-Serv networks • Performance of AQM schemes evaluated using most widely used simulator, NS-2, indifferent topologies andwide varying network conditions • In all cases, our approach outperforms all others in all scenarios and network conditions additional results FEM, FIO

49. FEM AVQ PI REM A-RED A-RED FEM AVQ PI REM PI REM AVQ A-RED FEM FEM Evaluation Effect of traffic load(increase flows from 100- 500) provides some time-varying dynamics and scalability. Single-bottleneck link (TQL = 200 packets ~ 100 msec) delay • FEM outperforms other AQM schemes • high link utilization,low delay and delay variation • exhibits more stable, robust behavior with bounded delay • lowest drops, over large traffic load delay variation loss

50. PI FEM AVQ REM A-RED FEM Evaluation (cnt’d) queue length evolution: sudden change in traffic conditions 600 users at t=0 300 leave network t=40 300 re-enter network t=70 multiple bottleneck links