Fuzzy logic based congestion control

1. Fuzzy logic based congestion control Andreas Pitsillides, Department of Computer Science University of Cyprus Ahmet Sekercioglu, Department of Informatics Swinburne University of Technology

2. Congestion control problem • Generally accepted that • network congestion control remains critical issue and high priority, especially given growing size, demand, and speed (bandwidth) of the increasingly integrated services networks. • Despite research efforts spanning few decades and large number of different schemes proposed, • no universally acceptable control solutions in “fight” against congestion (a control strategy, a control system, or a “package” of control solutions). • Current solutions in existing networks • increasingly becoming ineffective, and • generally accepted that these cannot easily scale up – even with various proposed “fixes”.

3. congestion • Definition • network state in which performance degrades due to saturation of network resources, (communication links, network switches, processor cycles, and memory buffers). • E.g. if communication link delivers packets to queue at higher rate than its service rate, then size of queue grows. If queue space finite then in addition to delay, losses occur • Control • refers to set of actions taken by network to minimise intensity, spread, and duration of congestion. • optimal control of networks of queues • well-known, much studied, and notoriously difficult problem, even for simplest of cases. E.g. Papathemetriou and Tsitsiklis show problem of optimally controlling simple network of queues with simple arrival and service distributions and multiple customer classes is complete for exponential time (i.e. provably intractable).

4. Congestion control • effect of network congestion is degradation in the network performance. • user experiences long delays in delivery of messages, perhaps with heavy losses caused by buffer overflows. • Thus degradation in quality of delivered service, with need for retransmissions of packets (for services intolerant to loss). In event of retransmissions, drop in throughput, eventually leading to network throughput collapse. • Congestion is a complex process to define. • felt by degradation of performance. • loss, delay and throughput deterioration good indicators of congestion. • A “good” congestion control system should be preventive, if possible. Otherwise should react quickly and minimise spread of congestion and its duration.

5. Congestion can be sensed (or predicted) by: • packet loss • sensed by the queue as an overflow, • sensed by destination (through sequence number) and acknowledged to a user, • sensed by sender due to lack of acknowledgment (timeout mechanism) to indicate loss • packet delay • can be inferred by the queue size, • observed by destination and acknowledged to user (e.g. with time stamps in packet headers) • observed by sender, e.g. by packet probe to measure Round Trip Time (RTT) • loss of throughput • observed by the sender queue size(waiting time in queue) • other calculated or observed event through which congestion can be inferred • increased network queue length and its growth • calculated from measured data, e.g queue inflow and its effect on future queue behaviour.

6. potential problems of control: • Large scale • Distributed nature • Large geographic spread (at its limit it covers globe) • Increasingly processing delay at nodes gets smaller, in comparison to propagation delay in links. Large-bandwidth delay product makes control of congestion through feedback potentially difficult • Diverse nature and behaviour of carried traffic (voice, video, www, ftp, e-mail, ….) • Unpredictable and time varying user behaviour • Lack of appropriate dynamic models for control • Expectation of the need for guaranteed levels of performance to each user, which can be negotiated with the network

7. (lack of) modelling of network system • To design control system necessary to model input and output cause and effect of system. • difficulty of deriving such model • Several factors affect model, and some are time varying. Take model of traffic behaviour as example: • diverse user (human) behavior affects way traffic is generated (can be time varying and different for different humans, even for same interactive services). • inherent fuzzines, e.g, in • definition of “contract” between user and network and its policing, • in the controls (declared objectives of controls and observed behaviour of the system). • Data generation, organisation and retrieval (long range dependance shown for both source generation, as well as storage of data) • Traffic aggregation process is a very complex one—studies suggest self-similarity preserved under variety of network operations and network conditions • Network controls (speculation that fractal features in network traffic remain even after network controls). • Network evolution (self-similarity appears robust to network changes, eg. upgrades).

8. control theoretic concepts • Making use of control theoretic concepts has potential benefits, including: • Simpler an/or more effective algorithms with more predictable properties. • Better understanding of performance of controlled system (including dynamic behaviour). • Better understanding of existing non-linear algorithms, including need for any fixes (“jacketing software”). • Better analysis techniques for large systems of interacting algorithms. • A major difficulty in control system design is to reconcile the large-scale, fuzzy, real problems with the simple well-define problems that control theory can typically handle • good understanding of fundamental control theory (which can be sophisticated and complex), as well as deep understanding of system under control (not necessarily in form of accurate mathematical model).

9. For historical reasons, and due to fundamental philosophical differences earlier approach to congestion control, differed between TCP/IP and ATM However some convergence between classical TCP/IP and ATM approach evident, (RFC2309, RFC2481, ATM Forum). become clear that existing TCP congestion avoidance mechanisms (RFC2001), while necessary and powerful, not sufficient. Basically, there is a limit as to how much control can be accomplished from edges of network Some mechanisms needed in routers to complement endpoint congestion avoidance mechanisms. (Need for gateway control realised early; e.g. see [Jacobson, 1988], where for future work gateway side advocated as necessary). Evolutionary, for TCP/IP and ATM we see progressive shift of controls from edges of network (initially open loop then edge binary feedback) to inside network. Feedback also shifting from implicit to explicit, from pure binary to multivalued and explicit. Existing approaches to congestion control and trends

10. Network controls • network system • large distributed complex system, with difficult often highly non-linear, time varying and chaotic behaviour. • inherent fuzziness in definition of controls (declared objectives and observed behaviour). • Dynamic or static modelling of such system for (open or closed loop) control is extremely complex. • Measurements on the state of the network are incomplete, often relatively poor and time delayed. • Its sheer numerical size and geographic spread are mind-boggling. E.g. customers (active services) in 10s of millions, network elements in 100s of million, and global coverage. • Computational intelligence to handle complexity and fuzziness present in network system surely has an essential role to play here. We should exploit tolerance for imprecision and uncertainty to achieve tractability, robustness and low cost

11. Computational Intelligence (CI) • area of fundamental and applied research involving numerical information processing (in contrast to symbolic information processing techniques of Artificial Intelligence (AI)). • Nowadays, CI research is very active and consequently its applications are appearing in some end user products. • definition of CI can be given indirectly by observing the exhibited properties of a system that employs CI components: • “A system is computationally intelligent when it: deals only with numerical (low-level) data, has a pattern recognition component, and does not use knowledge in the AI sense; and additionally, when it (begins to) exhibit • computational adaptivity; • computational fault tolerance; • speed approaching human-like turnaround; • error rates that approximate human performance. • The major building blocks of CI are artificial neural networks, fuzzy logic, and evolutionary computation.”

12. Fuzzy Explicit Rate Marking (FERM) for congestion control • Basic idea: • measure queue length and queue growth rates (hence providing rudimentary prediction of future behaviour) at the output buffer of a switch, • calculate and send explicit rate control signals to traffic sources to avoid or alleviate congestion. • explicit rate control signals calculated periodically by fuzzy inference engines located in switches, and sent to traffic sources in resource management (RM) cells. • analyzed and compared performance of FERM with EPRCA in detail regarding fairness, responsiveness, resource utilization and cell loss in LAN and WAN environments. FERM has been further refined (FERM2) and as an adaptive scheme which has self tuning capabilities (A-FERM).

13. FERM2 explicit rate congestion control scheme Note: desired queue length is explicit, can be set by a higher level control module to provide more dynamic resource utilization across switches utilised by VC

14. FERM 2 Operation • Overall operation compliant with ATM Forum Traffic Management Specification, Version 4. • Source Cell rates adjusted by Explicit Rate (ER) carried by RM cells. • RM cells periodically generated by traffic sources, transmitted towards destination end systems, and initial ER information is set by the ICR. • Destination end systems bounce RM cells back to sources. • During return path, when RM cell passes through ATM switch, its ER value is examined and possibly modified. • Data source, upon receiving RM cell, adjust cell rate = ER. If ER>PCR, cell rate=PCR. Similarly, cell rate=MCR if ER<MCR. • ER provided to all active VCs at all time so congestion and undesired resulting behavior can be avoided. • does not need to keep state of current VC connections sharing switch. • Periodical ER calculations are performed by the Fuzzy Congestion Controllers (FCCs) located in each ATM switch.

15. implementation of FCC • Chosen most widely used and computationally lighter methods, which are • singleton fuzzification • t-norm algebraic product for the mathematical representation of the connective “and” • Larsen's product rule of implication • sup-product compositional rule of inference • weighted mean of maximums defuzzification.

16. Control surface • if queue length and queue then flow rate • too short decreasing fast increase sharply • `` `` slowly `` moderately • `` not changing `` `` • `` increasing slowly decrease moderately • `` `` fast `` `` • acceptable decreasing fast increase moderately • `` `` slowly `` `` • `` not changing no change • `` increasing slowly decrease moderately • `` increasing fast `` `` • too high decreasing fast no change • `` `` slowly `` • `` not changing decrease moderately • `` increasing slowly `` sharply • `` `` fast `` `` Set of linguistic rules defining the control surface of the FCC

17. network model for performance analysis: simulation of ATM WAN (1500 km between switches) backbone and LAN backbone (10 km) Real time sources ABR sources 3-hop 10/1500 km ABR sources 1-hop

18. simulated ATM LAN under FERM2 congestion control: average end-to-end ABR cell delay vs. useful throughput

19. simulated ATM WAN under FERM2 congestion control: average end-to-end ABR cell delay vs. useful throughput

20. Time evolution of Explicit Rate for case of LAN; calculated by the FCC

21. Time evolution of the queue length for the case of a LAN. Note that the reference value set at 400 cell places.

22. Time evolution of Explicit Rate for the case of the WAN; calculated by the FCC

23. Time evolution of the queue length for the case of a WAN. Note that the reference value is set at 400 cell places.

24. Conclusions and recommendations • congestion control in communication networks is a real challenge, especially supporting video, voice and data applications simultaneously. • Computational Intelligence techniques expected to play central role, especially in large scale, geographically distributed network systems. Hybrids also expected to supplement these techniques and prove useful, especially in optimising overall network objectives. • challenges include: • Agreement on structured approach to congestion control for network. Control theoretic concepts and techniques have essential role to play. • Engineer network system with network control system together in order to add another degree of flexibility. • Theoretical advances in handling large scale complex systems are required, including decomposition and organisation of controls. • Globally optimise overall network objectives. • Agreement in common simulative framework and common test bed framework for testing congestion control algorithms