Congestion Control. COSC 6590 Week 2 Presentation By Arjun Chopra, Kashif Ali and Mark Obsniuk

Transcription

1 Congestion Control COSC 6590 Week 2 Presentation By Arjun Chopra, Kashif Ali and Mark Obsniuk

2 Topics Congestion control TCP and the internet AIMD congestion control Equation Based congestion control Comparison of AIMD/Equation Based

3 Congestion Control Data is being sent by many sources, too fast

4 Congestion Control (Cont) Costs of Congestion Wasted bandwidth from retransmits Possible poor resource utilization (ie fairness)

5 Congestion Control (Cont) Goals of Congestion Control Maximize throughput Fairness Equal Share Maximize minimum rate

6 Model for Congestion Control Simple model for congestion control Congestion - lower rate No Congestion - increase rate

7 AIMD Additive Increase, Multiplicative Decrease AIMD(a, b) a - increase, b - decrease Certain values of a and b provide good performance

8 AIMD (Cont) AIMD(1,1/2) - TCP Ratios used in AIMD should be TCP friendly

9 AIMD HI'>?&"#- F'>?&"#- C*+,D"#;E'-"$?.9."& ".E#

10 AIMD (Cont) Performance is good with files but... What about Streaming media (Audio/Video)?

11 Streaming Media AIMD provides choppy throughput Smooth throughput yet TCP friendly

12 Alternative Rate Control Do not aggressively seek out available bandwidth Do not drop rate in half on a single packet loss Receiver sends feedback within round-trip time Sender reduces rate when there is no feedback

13 Equation Based Congestion Control Utilizes different criteria for rate adjustments then TCP (AIMD(1,1/2)) Maintains a TCP like performance over time but smoothes out transitions from low-high/high-low congestion

14 Equation-Based (cont) Determine R, p, s, t RTO Transmission rate based on T

15 Sender Functionality Determine R (round trip time) using numbered data packets t = 4R RTO T < T, increase rate actual T actual > T, decrease to T

16 Receiver functionality Provides feedback to measure RTT. Provide loss event rate p at least once per RTT.

17 The Loss Event Rate - p Obvious way: loss fraction = number lost /number transmitted The TCP friendly way: Ignore losses within a round trip time that follow an initial loss Reduce congestion window at most once for notifications in one window of data

18 Criteria for Calculating p Should track relatively smoothly in a stable steady-state loss event rate environment Measure loss event rate rather than packet loss rate Should respond strongly to loss events in several successive round -trip times

19 Criteria for Calculating p (cont) Increase only in response to a new loss event rate Decrease only if loss interval is longer than previous average interval since last loss event is sufficiently long

20 The Full Average Loss Interval Method Computes the average loss rate over the last n loss intervals. Most recent loss interval s ignored, unless it is 0 large enough to increase the average.

21 Full Average Loss Interval Method (Cont.)

22 Weighted intervals for loss probability calculation Sequence Number Interval since most recent loss Packet lost Packet Arrival interval 1 interval 2 weight 1 weighted interval 1 weighted interval 2 interval n weighted interval n Time now Time weight n

23 weights for the older loss intervals are discounted b the following discount factor: Full Average Loss Rate (Cont.) The lower bound of 0.5 on the discount factor ensu past losses will never be completely forgotten, regard History Discounting the number for of packet arrivals since the last loss. faster response to When history discounting is invoked, this gives sustained decrease lowing in estimated loss interval: congestion S 0 > 2 S (i>1) When loss occurs and the old interval is shifted to the discount factors are also shifted, so that once an is discounted, it is never un-discounted, and its discou tor is never increased. In normal operation, in the abs history discounting, = 1 for all values of. We do

24 Average Loss Interval method with idealized periodic loss Loss Interval Loss Rate TX Rate (KBytes/s) current loss interval (s0) estimated loss interval estimated loss rate 0.25 square root of estimated loss rate Transmission Rate Time (s) Time (s) Time (s)

25 Protocol Features Increasing the Transmission Rate Method provides sufficient damping to avoid oscillatory behavior Maximum Increase 0.14 packets per RTT with no discounting 1 packet per RTT with maximum history discounting

26 Protocol Features (Cont. ) Response to persistent congestion Requires three to eight RTTs to reduce sending rate to half

27 Protocol Evaluation Performance with long term duration background traffic TFRC responds only to aggregate loss rate then individual loss event

28 Protocol evaluation (cont.) TFRC changes the queue dynamic due to its nature Various metric such as link utilization, queue occupancy and packet drop rates were used

29 TCP Deterministic model Determinstic model suggests that TCP(1/5, 1/8) should compete fairly with standard TCP. Why? Using simulation results it is concluded that deterministic assumption and retransmit timeout is key factor in this misleading suggestion.

30 TCP/TCP(a,b)

31 TFRC/TCP(a,b)

32 Advantages of TCP(a,b) and Equation-based congestion AIMD(a,b) congestion control is reasonably well understood and familiar in terms of fairness, stability, oscilation and other properties. Whereas equation-based congestion control has less abrupt changes in the sending rate.

33 Key factors for comparison Transient Response Smoothness

34 Transient Response Responsiveness is number of round-trip to decrease the sending rate by half. Aggressiveness is the max. increase in sending rate in one round-trip time. Smoothness as largest reduction of the sending rate in one round-trip time.

35 Tradeoffs: responsiveness, smoothness and aggressiveness

36 TCP/TCP(a,b) Smoothness

37 TCP/TFRC Smoothness

38 Throughput

39 Conclusion Congestion control is very important in order to keep today Internet going Equation-based approach provides an alternate to those application that requires smooth traffic while still being TCP-compatible Various AIMD schemes can perform well while still maintaining TCP-compatibility

40 Conclusion TCP(1/5, 1/8) and TCP(2/5, 1/8) compete fairly equally with TCP and TFRC while avoiding reducing the sending rate to half TFRC changes its sending rate more smoothly then TCP, while maintaining similar transient response to congestion

41 References Equation-Based Congestion Control for Unicast Applications: Sally Floyd, Mark Handley A Comparison of Equation-Based and AIMD Congestion Control: Sally Floyd, Mark Handley fall2004/lectures/lectures13-congestion.pdf