Stochastic Forecasts Achieve High Throughput and Low Delay over Cellular Networks

Transcription

1 Stochastic Forecasts Achieve High Throughput and Low Delay over Cellular Networks Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan M.I.T. CSAIL April 5, 2013

2 Cellular networks are variable 2000 Verizon LTE uplink throughput throughput (kbps) :00 0:30 1:00 1:30

3 Cellular networks are too reliable 10 Round-trip time (s) Time (s) (Verizon LTE, one TCP download.)

4 Interactive apps work poorly We measured cellular networks while driving: Verizon LTE Verizon 3G (1xEV-DO) AT&T LTE T-Mobile 3G (UMTS) Then ran apps across replayed network trace: Skype (Windows 7) Google Hangout (Chrome on Windows 7) Apple Facetime (OS X)

5 Performance summary 8000 Verizon LTE Downlink Throughput (kbps) Better Skype Self-inflicted delay (ms)

6 Why is performance so bad? Exiting schemes react to congestion signals. Packet loss. Increase in round-trip time. Feedback comes too late. The killer: self-inflicted queueing delay. Throughput overshoot means a queue filling up.

7 Sprout s goal Most throughput Bounded risk of delay > 100 ms

8 Bounded risk of delay Infer link speed from interarrival distribution. Predict future link speed. Don t wait for congestion. Control: Send as much as possible, but require: 95% chance all packets arrive within 100 ms.

9 Infer: link speed from flicker noise process 100 (< 0.5) 10 t 3.27 Percent interarrivals interarrival time (ms) (Verizon LTE, phone stationary.)

10 Predict: future link speed Model evolution of speed as random walk. (Brownian motion) Cautious forecast: 5th percentile cumulative packets Receiver makes forecast; sends back to sender in ack Almost all precalculated

11 Sprout s model Poisson process drains queue Sender Queue Receiver Rate λ controls Poisson process λ Brownian motion of σ t varies λ σ If in an outage, λ z is escape rate. λ z

12 Parameters Volatility σ: Expected outage time 1/λ z : 1 s Tick length: 20 ms Forecast length: 160 ms Delay target: 100 ms Risk tolerance: 5% 200 pkts/s s All source code was frozen before data collection began.

13 Control: fill up 100 ms forecast window Cumulative packets sent Cautious forecast Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan time (ms) M.I.T. CSAIL

14 Control: fill up 100 ms forecast window Cumulative packets sent ms Cautious forecast Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan time (ms) M.I.T. CSAIL

15 Control: fill up 100 ms forecast window Cumulative packets sent Send 50 pkts 100 ms Cautious forecast Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan time (ms) M.I.T. CSAIL

16 Control: fill up 100 ms forecast window 70 Cumulative packets sent Send 50 pkts 100 ms 100 ms Cautious forecast Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan time (ms) M.I.T. CSAIL

17 Control: fill up 100 ms forecast window 70 Cumulative packets sent Send 50 pkts Send ms 100 ms Cautious forecast Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan time (ms) M.I.T. CSAIL

18 Control: fill up 100 ms forecast window 70 Cumulative packets sent Send 50 pkts Send ms 100 ms 100 ms Cautious forecast Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan time (ms) M.I.T. CSAIL

19 Control: fill up 100 ms forecast window 70 Cumulative packets sent Send 50 pkts Send 5 Send ms 100 ms 100 ms Cautious forecast Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan time (ms) M.I.T. CSAIL

20 Verizon LTE Downlink Throughput (kbps) Better Skype Self-inflicted delay (ms)

21 Verizon LTE Downlink Throughput (kbps) Facetime 2000 Better Skype Google Hangout Self-inflicted delay (ms)

22 Verizon LTE Downlink Throughput (kbps) Better LEDBAT Vegas Compound TCP Facetime 2000 Skype Google Hangout Self-inflicted delay (ms)

23 Verizon LTE Downlink Throughput (kbps) Cubic Better LEDBAT Vegas Compound TCP Facetime 2000 Skype Google Hangout Self-inflicted delay (ms)

24 Verizon LTE Downlink Throughput (kbps) Cubic Better LEDBAT Sprout Vegas Compound TCP Facetime 2000 Skype Google Hangout Self-inflicted delay (ms)

25 Verizon LTE Downlink Sprout-EWMA Throughput (kbps) Cubic Better LEDBAT Sprout Vegas Compound TCP Facetime 2000 Skype Google Hangout Self-inflicted delay (ms)

26 Verizon LTE Uplink 5000 Sprout-EWMA Throughput (kbps) Cubic Skype Facetime Sprout LEDBAT Vegas Compound TCP Google Hangout Self-inflicted delay (ms)

27 Verizon 3G (1xEV-DO) Downlink Cubic Throughput (kbps) Google Hangout Compound TCP Facetime Skype Sprout-EWMA LEDBAT Vegas Sprout Self-inflicted delay (ms)

28 Verizon 3G (1xEV-DO) Uplink 600 Sprout-EWMA Throughput (kbps) Cubic LEDBAT Skype Facetime Vegas Compound TCP Google Hangout Sprout Self-inflicted delay (ms)

29 AT&T LTE Downlink Sprout-EWMA Throughput (kbps) Cubic Sprout Facetime LEDBAT Compound TCP Vegas Skype Google Hangout Self-inflicted delay (ms)

30 AT&T LTE Uplink Throughput (kbps) Cubic LEDBAT Sprout-EWMA Compound TCP Vegas Skype Facetime Sprout 300 Google Hangout Self-inflicted delay (ms)

31 T-Mobile 3G (UMTS) Downlink Sprout-EWMA Throughput (kbps) Cubic LEDBAT Compound TCP Facetime Skype Google Hangout Vegas Sprout Self-inflicted delay (ms)

32 T-Mobile 3G (UMTS) Uplink Throughput (kbps) 1000 Cubic LEDBAT Sprout-EWMA Vegas Compound TCP Facetime Sprout Skype Google Hangout Self-inflicted delay (ms)

33 Overall results Sprout vs. Avg. speedup Delay reduction Skype Hangout Facetime Compound TCP Vegas LEDBAT Same 2.8 Cubic

34 Varying risk tolerance Throughput (kbps) 1000 Cubic % LEDBAT 95% 50% Vegas 25% Compound TCP Facetime Skype Sprout (5%) Google Hangout Self-inflicted delay (ms)

35 Competes with AQM even though end-to-end Sprout-EWMA Utilization (percent) Cubic Cubic-over-CoDel Sprout Self-inflicted delay (ms)

36 Competing traffic inside Sprout tunnel Direct via Sprout Benefit Cubic throughput 8336 kbps 3776 kbps 0.5 (= worse) Skype throughput 78 kbps 490 kbps 6 Skype 95% delay 6.0 s 0.17 s 35

37 Replication by Stanford students (February March 2013) Alterman & Quach reproduced some of our measurements Won best project award in Stanford networking class!

38 M.I.T contest (March April 2013) Turnkey network emulator, evaluation Sender, receiver run in Linux containers 4th prize: $20 3rd prize: $30 2nd prize: $40 (If beat Sprout) 1st prize:

39 M.I.T contest (March April 2013) Turnkey network emulator, evaluation Sender, receiver run in Linux containers 4th prize: $20 3rd prize: $30 2nd prize: $40 (If beat Sprout) 1st prize: Co-authorship on future paper

40 Baseline

41 Land of 3,000 student protocols

42 Sprout is on the frontier

43 Limitations Only evaluated long-running flows. All testing data from Boston. User should wrap competing flows inside Sprout. If queue is full of another user s packets, an end-to-end scheme can t help. Fortunately, cells have per-device queues but Wi-Fi generally doesn t. What about when the cell link isn t the bottleneck?

44 Our approach Pick a model, any model. All models are wrong, but they help anyway! See if it lands on the frontier.* * (On a large set of real network paths or newly-collected traces.) Kaizen for congestion

45 Thank you Lakshminarayanan Subramanian Shuo Deng Jonathan Perry Katrina LaCurts Andrew McGregor Tim Shepard Dave Täht Michael Welzl Hannes Tschofenig members ( NSF & Shannon family (fellowship)

46 Sprout for controlled delay over cellular networks Infer link speed from interarrival distribution Predict future link speed Control risk of large delay with cautious forecast Yields 2 4 throughput of Skype, Facetime, Hangout Achieves 7 9 reduction in self-inflicted delay Matches active queue management without router changes Code and directions at alfalfa@mit.edu