Copa: Practical Delay-Based Congestion Control

Transcription

1 Copa: Practical Delay-Based Congestion Control Venkat Arun and Hari Balakrishnan MIT, CSAIL web.mit.edu/copa

2 The Internet is more challenging than ever (Why are we still talking about congestion control in 2018?)

3 The Internet is more challenging than ever (Why are we still talking about congestion control in 2018?) Higher bandwidth-delay product Greater bandwidth Lower tolerance for non-congestive loss Greater flow-churn

4 The Internet is more challenging than ever (Why are we still talking about congestion control in 2018?) Higher bandwidth-delay product Greater bandwidth Lower tolerance for non-congestive loss Greater flow-churn Large flows (e.g. video streaming) co-exist with short-flows

5 The Internet is more challenging than ever (Why are we still talking about congestion control in 2018?) Higher bandwidth-delay product Greater bandwidth Lower tolerance for non-congestive loss Greater flow-churn Large flows (e.g. video streaming) co-exist with short-flows Wireless links with variable bandwidths are commonplace

6 The Internet is more challenging than ever (Why are we still talking about congestion control in 2018?) Higher bandwidth-delay product Greater bandwidth Lower tolerance for non-congestive loss Greater flow-churn Large flows (e.g. video streaming) co-exist with short-flows Simultaneously, users are more sensitive to performance! Wireless links with variable bandwidths are commonplace

7 Loss-based schemes have long-standing problems Buffer-filling Vulnerable to non-congestive loss Loss is a coarse signal Worsens with increasing bandwidth

8 Delay-based congestion control? Benefits Challenges

9 Delay-based congestion control? Benefits Maintain low delay Challenges Not competitive with buffer-filling schemes

10 Delay-based congestion control? Benefits Maintain low delay Challenges Not competitive with buffer-filling schemes Robust to misleading loss Delay can mislead too!

11 Delay-based congestion control? Benefits Maintain low delay Challenges Not competitive with buffer-filling schemes Robust to misleading loss Delay can mislead too! Rich signal

12 Delay-based congestion control? Benefits Maintain low delay Challenges Not competitive with buffer-filling schemes Robust to misleading loss Delay can mislead too! Rich signal Finding true minimum RTT is hard

13 Delay-based congestion control? Challenges Our Solution Not competitive with buffer-filling schemes Delay is noisy too! Finding true minimum RTT is hard

14 Delay-based congestion control? Challenges Not competitive with buffer-filling schemes Our Solution If buffer-fillers are present, give up on low delay Delay is noisy too! Finding true minimum RTT is hard

15 Delay-based congestion control? Challenges Not competitive with buffer-filling schemes Delay is noisy too! Our Solution If buffer-fillers are present, give up on low delay MIN: a more robust statistic for queuing delay Finding true minimum RTT is hard

16 Delay-based congestion control? Challenges Not competitive with buffer-filling schemes Delay is noisy too! Finding true minimum RTT is hard Our Solution If buffer-fillers are present, give up on low delay MIN: a more robust statistic for queuing delay Empty queues periodically

17 Basic Goals Avoid congestion collapse + Efficient and Fair allocation of bandwidth + Low delay

18 Target rate = r ( = ) *+, - Adjustable Parameter default = 0.5 Queuing delay

19 Target Rate Nash Equilibrium Selfishly optimize for: "#$%$#& ' = log #,-# / ' log(1 2 ) Assuming Poisson arrivals (more details in paper)

20 Target Rate Nash Equilibrium Unique and Efficient Selfishly optimize for: "#$%$#& ' = log #,-# / ' log(1 2 ) Assuming Poisson arrivals (more details in paper)

21 Computing the Target Rate

22 Estimating queuing delay from RTT RTT RTT Queuing delay True minimum RTT RTT min Time

23 Estimating queuing delay from RTT RTT RTT Queuing delay RTT min RTT RTT min = queuing delay True minimum RTT Time

24 Estimating queuing delay from RTT New flow starts here RTT RTT min Wrong! True minimum RTT Time

25 Estimating queuing delay from RTT True minimum RTT = RTT min for new flows! RTT True minimum RTT Time

26 A noisy cellular link: Stanford to AWS Delay experienced by individual packets (ms) Time (s)

27 Decoupling queuing delay from other delay variation Queue length Measured RTT Wireless links Cross traffic ACK compression Bursty transmission Time

28 Decoupling queuing delay from other delay variation Queue length Positive Additive Noise! Measured RTT Take Min over last 0.5 RTT of samples Wireless links Cross traffic ACK compression Bursty transmission Time

29 A noisy cellular link: Stanford to AWS Delay experienced by individual packets Time (s) Using the MIN delay estimator improves throughput from 0.5 Mbits/s to 3.9 Mbits/s

30 Attaining the Target

31 The Copa Algorithm Calculate target rate = ' ( = ) *+, If current rate < r. : additively increase by 0 * pkts/rtt Else: additively decrease by 0 * pkts/rtt

32 The Copa Algorithm Calculate target rate = ' ( = ) *+, Velocity for faster convergence If current rate < r. : additively increase by 0 * pkts/rtt Else: additively decrease by 0 * pkts/rtt

33 The Copa Algorithm Calculate target rate = ' ( = ) *+, Velocity for faster convergence If current rate < r. : additively increase by 0 * pkts/rtt Else: additively decrease by 0 * pkts/rtt Cwnd (pkts) Cwnd RTT RTTstanding RTTstanding (ms) RTT (ms) Time (s)

34 Steady-State Dynamics of Copa 2.5 $ %& Period: 5 RTTs Amplitude: 2.5 $ %& Queue Length (in pkts) 1 $ %& Equilibrium queue length 0 0 RTT 0.5 RTT. 1.5 RTT 2.5 RTT 3 RTT 4 RTT Time

35 Steady-State Dynamics of Copa 2.5 $ %& Switch point Period: 5 RTTs Amplitude: 2.5 $ %& Queue Length (in pkts) 1 $ %& 0.5 RTT window Queue length corresponding to RTT standing at the switch point Equilibrium queue length 0 0 RTT 0.5 RTT. 1.5 RTT 2.5 RTT 3 RTT 4 RTT Time

36 Steady-State Dynamics of Copa 2.5 $ %& Switch point Period: 5 RTTs Amplitude: 2.5 $ %& Queue Length (in pkts) 1 $ %& 0.5 RTT window Queue length corresponding to RTT standing at the switch point Equilibrium queue length 0 0 RTT 0.5 RTT. 1.5 RTT 2.5 RTT 3 RTT 4 RTT Time

37 Steady-State Dynamics of Copa 2.5 $ %& Switch point Period: 5 RTTs Amplitude: 2.5 $ %& Queue Length (in pkts) 1 $ %& 0.5 RTT window Queue length corresponding to RTT standing at the switch point Equilibrium queue length 0 0 RTT 0.5 RTT. 1.5 RTT 2.5 RTT 3 RTT 4 RTT Time

38 Steady-State Dynamics of Copa 2.5 $ %& Switch point Period: 5 RTTs Amplitude: 2.5 $ %& Queue Length (in pkts) 1 $ %& 0.5 RTT window Queue length corresponding to RTT standing at the switch point 0.5 RTT window Equilibrium queue length 0 0 RTT 0.5 RTT. 1.5 RTT 2.5 RTT 3 RTT 4 RTT Time Switch point

39 Estimate true minimum RTT Queue empties every 5 RTTs! Detect buffer-filling TCP

40 TCP-Competitiveness

41 Mode switching for TCP competitiveness Delay sensitive (! = 0.5) Throughput (Mbits/s) TCP Cubic Copa Queuing Delay (ms)

42 Mode switching for TCP competitiveness Delay sensitive (! = 0.5) Throughput (Mbits/s) TCP Cubic Copa Low Delay Low Throughput Low Delay Queuing Delay (ms)

43 Mode switching for TCP competitiveness Delay sensitive (! = 0.5) Throughput (Mbits/s) TCP Cubic Copa Low Delay Low Throughput Low Delay Queuing Delay (ms) TCP Competitive (AIMD on! &' ) Throughput (Mbits/s) Copa TCP Cubic High Delay High Throughput High Delay Queuing Delay (ms) Time (s)

44 Mode switching for TCP competitiveness Delay sensitive (! = 0.5) Throughput (Mbits/s) TCP Cubic Copa Low Delay Low Throughput Low Delay Queuing Delay (ms) TCP Competitive (AIMD on! &' ) Throughput (Mbits/s) Copa TCP Cubic High Delay High Throughput High Delay Queuing Delay (ms) Time (s)

45 Mode switching for TCP competitiveness Best of both worlds! Throughput (Mbits/s) Copa TCP Cubic Low Delay High Throughput Low Delay Queuing Delay (ms) Time (s) When queue doesn t empty once every 5 RTTs, switch to TCP Competitive mode!

46 Copa gets higher throughput without hurting TCP Cubic! Avg. (Tpt / Ideal tpt) Scheme under test's throughput Cubic's throughput 0 Vegas Reno Cubic Copa PCC BBR

47 Copa gets higher throughput without hurting TCP Cubic! Avg. (Tpt / Ideal tpt) Scheme under test's throughput Cubic's throughput 0 Vegas Reno Cubic Copa PCC BBR

48 Copa gets higher throughput without hurting TCP Cubic! Avg. (Tpt / Ideal tpt) Scheme under test's throughput Cubic's throughput Copa gets higher throughput Cubic flows are not hurt 0 Vegas Reno Cubic Copa PCC BBR

49 Limitations Cannot ignore low frequency noise Queues don t empty periodically if: Propagation delay is much smaller than queuing delay Flows with very different propagation delays share a bottleneck queue Needs precise RTT measurements

50 Consistent Performance on Real Paths Cellular Networks Wired Networks 1 1 Avg. Normalized Throughput Cubic Vivace loss BBR Verus Vivace LTE Vegas PCC Remy Vivace latency Better LEDBAT Copa Sprout Avg. Normalized Throughput BBR Remy Copa Cubic Vivace loss PCC Verus Vivace LTE Vegas LEDBAT Sprout Better 2 Vivace latency 1 Avg. Queuing Delay (ms) Avg. Queuing Delay (ms)

51 Satellite link: High BDP, high loss Throughput (Mbit/s) PCC Better Copa Remy BBR Cubic Vegas Median Queuing Delay (ms)

52 Fairness during flow-churn 10 Copa 10 Cubic Throughput (Mbit/s) Avg ± Std. Dev BBR PCC Time (s) Time (s)

53 Fairness during flow-churn 10 Copa 10 Cubic Throughput (Mbit/s) Avg ± Std. Dev BBR Better Jain index PCC BBR Cubic Copa PCC Time (s) Time (s)

54 Summary A practical delay-based congestion control algorithm Estimate queuing delay and compute target! Network

55 Summary A practical delay-based congestion control algorithm Estimate queuing delay and compute target! Network Outer control loop Mode switching and Competitive mode

56 Summary A practical delay-based congestion control algorithm Rate/cwnd Move toward target (AIAD-variant) Estimate queuing delay and compute target! Outer control loop Mode switching and Competitive mode Network