1. IEEE JSAC Special Issue Adaptive Media Streaming Submissions by April 1 Details at http://www.jsac.ucsd.edu/Calls/adaptivemediastreamingCFP.pdf

2. Packet Video Workshop 2013 San Jose, CA (@ Cisco) December 12/13, 2013 (Right after PCS 2013) Submissions by mid June http://pv2013.itec.aau.at/

3. Server-Based Traffic Shaping for Stabilizing Oscillating Adaptive Streaming Players Ali C. Begen Saamer Akhshabi, Lakshmi Anantakrishnan, Constantine Dovrolis

4. Briefly … Problem When multiple adaptive streaming players compete for bandwidth, we have several problems: Instability Unfairness Bandwidth underutilization Objective A server-based traffic shaping solution to mitigate the oscillation problem without significant loss in utilization 4

5. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Conclusions 5

6. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Conclusions 6

7. Adaptive Streaming over HTTP Media is split into “chunks” Each chunk corresponds to a certain amount of content Each chunk is encoded at multiple bitrates Clients request chunks based on their estimate of available bandwidth 7 From IIS Smooth Streaming Website

8. Typical Behavior of a Player One chunk per HTTP request Two states: Buffering Request chunks as fast as possible Build up the playback buffer Steady Request a new chunk every T seconds Keep buffer size constant ON-OFF download pattern Estimate avail-bw with running average of per-chunk TCP throughput measurements 8

9. Outline Overview of adaptive streaming over HTTP Multiple-player competition Instability Bandwidth underutilization Unfairness Stabilization method and demonstration Results Conclusions 9

10. Simple Model: Two Competing Players 10 Ideal TCP A single active connection gets entire capacity C Two active connections share the capacity fairly receiving C/2 each Two competing adaptive streaming players Steady-State Full buffers Shared link Capacity C Fair share = C/2 Based on the temporal overlap of the ON-OFF periods of the players three performance problems can arise Instability Unfairness Bandwidth underutilization

11. The Root Cause of Oscillations Both players measure per-chunk throughput of more than C/2 Overestimate the fair share (f=C/2) They will request bitrate greater than f, if available Oscillations Bandwidth underutilization Unfairness 11 ON

12. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Server-based shaping solution Experimental setup Stabilization method Results Conclusions 12

13. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Server-based shaping solution Experimental setup Stabilization method Results Conclusions 13

14. Traffic Shaping Solution: Basic Idea A server-side stabilizer module Client independent Basically, a reactive and adaptive traffic shaper The stabilizer Remains inactive if there is no client-side instability When instability is detected, stabilizer shapes requested video chunks at the rate of a lower profile Client will then measure lower throughput, and it will return (and hopefully stabilize) to a lower profile 14

15. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Server-based shaping solution Experimental setup Stabilization method Results Conclusions 15

16. Experimental Methodology 16 Clients Simpler player Logs internal parameters Does not render video Smooth Streaming player DummyNet Sets the capacity of the share bottleneck Wireshark Captures the traffic for offline analysis Server Hosts the video content in multiple bitrates Host for stabilizer DummyNet Stabilizer

17. Implementation Server: Smooth Apache module: http://smoothstreaming.code-shop.com/trac/wiki/Mod- Smooth-Streaming-Apache Shaping done with Apache mod_bw module: http://bwmod.sourceforge.net Shaping module is modified to implement the stabilizer 17

18. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Server-based shaping solution Experimental setup Stabilization method Results Conclusions 18

19. Relation between Shaping Rate and Chunk Encoding Rate Eliminate OFF periods by shaping each chunk so that its download takes about T seconds Shape the chunk to the average encoding rate for that chunk In practice, the shaping rate is set to a slightly higher value than the encoding rate (r/c) Shaping slack parameter c (0.7 in this study) 19 ON

20. Relation between Shaping Rate and Chunk Encoding Rate Eliminate OFF periods by shaping each chunk so that its download takes about T seconds Shape the chunk to the average encoding rate for that chunk In practice, the shaping rate is set to a slightly higher value than the encoding rate (r/c) Shaping slack parameter c (0.7 in this study) 20 ON

21. 21 Outline

22. Oscillation Detection Detecting direction changes Direction change defined as a change in the requested profile (upshift or downshift) that is different than the last such change When two or more direction changes occur within W successive chunks, an oscillation is detected and the player is flagged as unstable Parameter W is the detection window size 22

23. Oscillation Detection 23

24. Initial Shaping Rate Selection Goal is to find the highest sustainable profile Consider a set of candidate profiles Highest profile is obviously not sustainable Shaper starts by setting the initial shaping rate to the next highest profile 24

25. Experiment with the Simpler Player Six video profiles between 0.7 and 5 Mbps Bottleneck capacity 10 Mbps Four players are competing 25 Player-1 starts streaming Three more player join Two players leave

26. Shaping Rate Decrease During shaping the server distinguishes between oscillations Due to OFF periods Shaping rate should be decreased Due to short-term avail-bw reductions No reason to modify the shaping rate How to distinguish between the two cases? Based on the requested profiles in the last W chunks 26 Short-term available bandwidth drops detected

27. Shaping Rate Increase The server occasionally estimates the available bandwidth De-activates shaping for randomly chosen chunks Measure the connection’s throughput by looking at the ratio cwnd/rtt Increase the shaping rate if the estimated available bandwidth is higher than the shaping rate 27 Two players leave Avail-bw increases Shaping rate increases Abort procedure activated

28. Shaping Rate Increase 28

29. 29 Outline

30. Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Performance metrics Number of competing players In the presence of a TCP bulk transfer Mix of players Conclusions 30

31. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Performance metrics Number of competing players In the presence of a TCP bulk transfer Mix of players Conclusions 31

32. Performance Metrics Instability Fraction of successive chunks in which the requested bitrate is not constant For each client: Increase: 1, Decrease: -1, Constant: 0 The number of 1’s and -1’s divided by the total number of chunk requests Utilization Aggregate throughput of all players divided by avail-bw 32 0010 001

33. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Performance metrics Number of competing players In the presence of a TCP bulk transfer Mix of players Conclusions 33

34. Number of Competing Players Instability Unshaped players Instability metric peaks at some mid-range N value Shaped players Instability metric significantly lower We cannot reject the hypothesis that the mean instability is constant as N is increased 34

35. Number of Competing Players Utilization Between shaped and unshaped players For some values of N, the utilization metric does not show a statistically significant difference For other values of N, the difference in the actual utilization is not large 35

36. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Performance metrics Number of competing players In the presence of a TCP bulk transfer Mix of players Conclusions 36

37. In the Presence of a TCP Bulk Transfer Instability metric is much less in the shaped case Aggregate utilization is high in both cases The TCP flow tends to fill up the bottleneck TCP connection takes up the largest share of the bottleneck’s capacity 37 BW=10 Mbps3 players (shaped or unshaped) Greedy

38. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Performance metrics Number of competing players In the presence of a TCP bulk transfer Mix of players Conclusions 38

39. Mix of Players Shaped players have the lowest instability Their presence helps to also stabilize the competing unshaped players Utilization is slightly less in the mixed-player experiments 39 BW=12 MbpsTwo shaped and two unshaped players

40. Outline Overview of adaptive streaming over HTTP Multiple-player competition Stabilization method and demonstration Results Conclusions 40

41. Conclusions Competition between adaptive streaming players can lead to performance problems Instability, unfairness and bandwidth underutilization Root cause is the ON-OFF behavior of players during Steady-State Players overestimate the fair share if their ON periods are not perfectly synchronized (which is rarely the case) Traffic shaping can help mitigate instability (Upon oscillations) by shaping chunks at a lower bitrate pushing players to switch to the lower profile Without incurring significant loss in bandwidth utilization 41