Rate control for ABR service in ATM Networks

1. Rate control for ABR service in ATM Networks 98/9/30 Multimedia & Comm. Lab 정승훈

2. Contents • Introduction • ABR Service • Congestion Control Mechanisms • Source-level Rate Adaptation • Examples • Open Issues

3. Introduction • Classes of Service • Why need Congestion Control • Traffic Management Functions • What is Expected from Congestion Control

4. Classes of Service • ABR (Available bit rate) • Follows feedback instructions. • Network gives max throughput with minimum loss. • UBR (Unspecified bit rate) • User sends whenever it wants. • No feedback mechanism, No guarantee. • Cells may be dropped during congestion. • CBR (Constant bit rate) • Throughput, delay, and Jitter guaranteed. • VBR (Variable bit rate)

5. Why need Congestion Control • Will the congestion problem be solved when: • Memory becomes cheap (infinite memory)? • Links become cheap (very high speed links)? • Processors become cheap? • Congestion is a dynamic problem • Static solutions are not sufficient • Bandwidth explosion • More unbalanced networks. • Buffer shortage is a symptom, not the cause.

6. Traffic Management Functions • Connection Admission Control (CAC) • Verify that the requested bandwidth and QoS can be supported • Traffic Shaping • Limit burst length, Space-out cells • Usage Parameter Control (UPC) • Monitor and control traffic at the network entrance • Network Resource Management • Scheduling, Queueing, Virtual path resource reservation

7. Traffic Management Functions • Priority Control • Cell Loss Priority (CLP) = 1 cells may be dropped • Selective Cell Discarding • Frame discard • Feedback Controls • Network tell the source to increase or decrease its load • Explicit forward congestion indication (EFCI) • Explicit rate (ER) • Backward explicit congestion notification (BECN)

8. What is expected ? • Objectives • Support a set of QoS parameters and classes for all ATM services • Minimize network and end-system complexity while maximizing network utilization • Selection Criteria • Scalability • Fairness • Robustness • Implementability

9. ABR Service • The Nature of the ABR Service • Some Early Debates • The Role of the Network • The Role of the End Systems

10. The Nature of the ABR Service • ABR Service • ABR connections will share the available bandwidth • The share of available bandwidth for each ABR connection is dynamic and may diminish down to a specified minimum cell rate (MCR) • The dynamic nature of the ABR service can be seen from the feedback model • The ABR service is appropriate only for applications which can adapt their rates to the time-varying available bandwidth and tolerate unpredictable cell delays • a low or zero cell loss rate is guaranteed to users who adapts their rates properly

11. Some Early Debates • Open-Loop vs. Close-Loop • Credit-based vs. Rate-based • Binary Feedback vs. Explicit Feedback

12. Open-Loop vs. Close-Loop • Open-Loop • not need end-to-end feedback • prior-reservation and hop-to-hop flow control • Close-Loop • the source adjust its cell rate in responding to the feedback information from the network. • Too slow in high-speed networks • But, ABR service is designed to use any bandwidth • ATM Forum specified that feedback is necessary for ABR flow control

13. Credit-based vs. Rate-based • Credit-based • hop-by-hop per-VC window • Static : Full round-trip worth of credit per VC • Adaptive : Credits depend upon activity • Rate-based • End-to-end rate control • Binary : Feedback via congestion bit in cells • Explicit : Feedback via resource management (RM) cells

14. Credit vs. Rate Debate : Issues • Per-VC queueing • Switch complexity, Nonscalable • Switch vs. end-system complexity • Zero cell loss • Isolation and misbehaving users • Buffer requirements • Full round-trip per VC

15. Binary vs. Explicit Rate • Binary Feedback • One-bit Feedback • Explicit forward congestion indicator (EFCI) set to 0 at source • Congested switch set EFCI to 1 • Every nth cell, destination sends a RM cell to the source indicating increase amount or decrease factor

16. Binary vs. Explicit Rate • Explicit Rate Feedback • Every Nrm cells, the sources send a control cell • The switches measure load over a period • The destination returns the cell to the source • The switches specify explicit rate in cell • The source adjusts the transmission rate

17. Binary vs. Explicit Rate • ER feedback schemes have several advantages • The switches know more information along the flow path • Faster to get the source to the optimal operating point • Policing is straight forward • Two ways for ER feedback • Forward feedback • Backward feedback

18. The Role of The Network • The network might provide information directly to the users • No information • A binary congestion indication • EFCI • Detailed congestion indication • RM cells with queue levels and severity level • Explicit bandwidth (or rate) information • RM cell with the current available bandwidth that can be adjusted by nodes along the connection in the forward direction • The destination returns the RM cell to the source with either and absolute rate or a relative rate adjustment

19. The role of the End systems • How the source and destination end systems work with Feedback information to adapt the source rate • Negative Feedback • source increments its rate by default • Positive Feedback • source decrements its rate by default • Explicit Feedback • source maintains its rate by default

20. Congestion Control Mechanisms • Fairness • Binary Feedback • EFCI • PRCA • Explicit Rate Feedback • EPRCA • ERICA

21. Fairness • Max-Min • available bandwidth = C / N • MCR plus equal share • available bandwidth = MCR + (C - MCR) / N • Maximum of MCR or Max-Min share • available bandwidth = max{MCR, Max-Min share} • Allocation proportional to MCR • The bandwidth allocation for a connection is weighted proportional to its MCR • Weighted allocation

22. EFCI • Mechanism • The network uses EFCI to convey congestion information (in the forward direction) • Feedback is returned via RM cells from the destination end system to source. • The sources adjust their rates by additive increase and multiplicative decrease (at periodic update intervals). • The feedback is negative, the source increase their rates by default and decrease only if an RM cell is received

23. Nrm EFCI=1 EFCI=0 Source Dest ATM node ATM node If EFCI=0 cell received RM cell PRCA • Proportional rate control algorithm • positive feedback • RM cells are generated at a rate proportional to the source rate • End system requires a means to discover when to generate an RM cell. • Every Nrm cells, only one cell with EFCI=0

24. Nrm User cell EFCI=0 RM cell CI=1 Source Dest ATM node ATM node CI=0 if no congestion CI=1 otherwise RM cell EPRCA • Enhanced proportional rate control algorithm • Source behavior • The source sends data cells with EFCI set to 0 and sends RM cells every n data cells. • The RM cells contain desired explicit rate(ER), current cell rate (CCR) and congestion indication(CI). • The source initializes CCR to the allowed cell rate(ACR) and CI to 0.

25. EPRCA (cont’d) • Switch behavior • computes a mean allowed cell rate(MACR) for all VCs using: MACR = (1 - a) * MACR + a *CCR • and the fair share as a fraction of this average • The ER field in the returning RM cells are reduced to fair share. • May set the CI bit in the cells passing. • Destination behavior • monitors the EFCI bits and mark the CI bit in the RM cell if the last seen data cell had EFCI bit set. • Problems • congestion detection is based on the queue length. • This method is shown to result in unfairness. • Sources that start up late may get lower throughput than those start early

26. ERICA • Explicit Rate Indication for Congestion Avoidance • Switch behavior • Set target rate at 95% of link bandwidth • Monitor input rate and number of active VCs k • Overload = Input rate / Target rate • VC’s share = VC’s current cell rate / Overload • Fairshare = Target rate / k • ER = Max(Fairshare, This VC’s share) • ER in Cell = Min(ER in Cell, ER) • Features • Measured overload/load at switch. • Small queue lengths during steady state. • Fast response.

27. Source level rate adaptation • Architecture • Encoder-level rate shaping • Rate shaping for precoded video

28. Rate control architecture MPEG Codec Rate Shaper Output buffer ATM Networks Quantizer Rate shaper Rate control Quantization level index Feedback

29. Rate shaping for precoded video • Frame discarding • Selective Block dropping • Eliminate some DCT coefficients • Block dropping with Error concealment • Feature-based block dropping

30. Examples of Rate control • Explicit Backward Congestion Notification • Composite Rate Control Scheme • Weighted Max-Min Fairness • Shaped VBR

31. EBCN • Explicit Backward Congestion Notification • M. Ghanbari - Essex Univ. • GLOBECOM ‘96 Tc Tn Video Sources Server Buffer occupancy max 0 • Using queue occupancy of the VP buffer of an ATM switch • RM cells : increase / decrease quantizer step size • qnew = max[qold + qdiff, qmax]

32. CRCS • Composite Rate Control Scheme • From • S. Karademir - Garleton Univ in Canada • GLOBECOM ‘96 • Congestion Notification • Explicit Feedback mechanism • Feedback cell • Traffic prediction • Prediction parameter • feedback info. • Average transmission rate during the last two cycles. • Prediction Model • TES : nonlinear auto-regressive model

33. WMMF • Weighted Max-Min Fairness • from • T.V. Lakshman - Bell Labs. • INFOCOM ‘97 • Design Goals • simple admission control • high statistical multiplexing gain • frequent bandwidth negotiation • adaptation of source rates to match available bandwidth • maintain low end-to-end delays • Key Idea • RCBR • associate a weight with each flow

34. WMMF (cont’d) • RCBR • Renegotiated CBR • hybrid of the CBR and VBR • the simplicity of admission control for CBR • the greater statistical multiplexing gains of VBR. • Key Points • short-term fluctuations are absorbed in local source buffers • long-term changes make the source renegotiate the bandwidth • Weighted fair share • Key • Using difference of flow activity

35. WMMF (cont’d) • Source Adaptation Mechanism • Demand Prediction • Discrete Auto-regressive model • Xn+k = m + rk(Xn-m) • r : correlation efficient • m : mean number of cels per frame • Gamma-Beta Auto-regressive Model • Heymann ‘96 • Encoder Rate Adaptation • rate adaptation function • lavg = Travg - [a * (Bp - SETPOINT) / Thorizon ] • Travg : Transmission Rate • Bp : Predicted Buffer

36. SVBR • Shaped VBR • From • M. Hamdi - ENST • IEEE JSAC Aug. ‘97 • Key Idea • CBR의 장점과 VBR의 장점을 혼합한 형태 • 비디오 전체에 대해 VBR로 인코딩 • CBR의 단점인 buffer delay를 제거 • Bursty한 부분의 영역에 대해서는 CBR을 적용 • VBR의 burstiness 감소 • Shaped Variable Bit Rate

37. SVBR (cont’d) • Rate Shaping • Principle • 하나의 GOP에 할당되는 비트수의 최대값을 설정 (leak rate) • GoP단위로 비트수를 계산하여 leak rate를 넘으면 Quantization parameter Q를 증가 • GoP scale rate prediction • GoP단위로 rate prediction 을 적용 • 다음 GoP의 크기를 예측한 후, 해당 GoP를 위한 Q를 재조정 • Qk+1 = QkRk/Rk+1 • Qk : k번째 GOP의 quantization parameter • Rk: k번째 GOP의 bit수

38. Open Issues • Policing • dynamic UPC control • time lag estimation • Point-to-Multipoint Connections • Branchpoint behavior • Priority service for RM cells • Virtual Source / destination

39. Conclusion • Rate control for ABR Services • Congestion Control algorithm • Source-level rate adaptation