1. On the Interactions Between Layered Quality Adaptation and Congestion Control for Streaming Video 11 th International Packet Video Workshop Nick Feamster Deepak Bansal Hari Balakrishnan MIT Laboratory for Computer Science http://nms.lcs.mit.edu/projects/videocm/

2. April 30, 2001MIT Laboratory for Computer Science2 Not Like Watching TV!

3. April 30, 2001MIT Laboratory for Computer Science3 Why Is This Happening? The Internet poses several problems for the delivery of data –Variable Bandwidth –Variable Delay –Packet Loss Very detrimental to interactive video delivery How do we transmit video on the Internet in the face of varying bandwidth?

4. April 30, 2001MIT Laboratory for Computer Science4 Video Client Streaming Video System Context This talk is about bandwidth adaptation Conclusion: The combination of smooth congestion control and clever receiver buffering can overcome the evils of bandwidth variation! Request for Video Streaming Video Loss/Latency Feedback Video Server

5. April 30, 2001MIT Laboratory for Computer Science5 Bandwidth Adaptation Available bandwidth varies with time Servers should adapt to varying bandwidth –Congestion Control: Transmission rate must correspond to available bandwidth be TCP-friendly –Quality Adaptation: Quality of video should correspond to transmission rate Limited capacity for buffering!

6. April 30, 2001MIT Laboratory for Computer Science6 Layered Video Simulcast: Each layer is independent Hierarchical: Higher depends on lower –Base/Enhancement layers –Linear granularity (C bits/layer) Hierarchical Layering Simulcast Layering

7. April 30, 2001MIT Laboratory for Computer Science7 Binomial Congestion Control Trade-off between increase aggressiveness and decrease magnitude K+L=1 implies TCP-friendly [Bansal, INFOCOM 2001] SQRT has a modest backoff (~R 1/2 ) => attractive for streaming media t w(t) AIMD Exponential Backoff AIMDBinomial Increase  /  w K Decrease  w  w L

8. April 30, 2001MIT Laboratory for Computer Science8 Reduced Oscillations In many cases, AIMD drops multiple layers in one backoff! This is not the case with SQRT. Rate oscillations in SQRT are much less pronounced than in AIMD. AIMD Bitrate SQRT Bitrate

9. April 30, 2001MIT Laboratory for Computer Science9 Layered Quality Adaptation Tailor video to available bandwidth! Can be immediate or receiver-buffered –Rejaie et al., SIGCOMM ‘99

10. April 30, 2001MIT Laboratory for Computer Science10 Receiver Buffering Allocate more buffer space to lower layers Add a layer when the following conditions are met: –Enough bandwidth is available –Enough video is buffered to sustain a backoff and continue playing all of the layers (including the new layer) Optimal L0 buffering R/2 C R Consumption Rate (n a +1)C Buffer Filling Buffer Draining

11. April 30, 2001MIT Laboratory for Computer Science11 Interaction of SQRT and QA: We Win! With SQRT: –Smaller Oscillations –Less buffering required for quality adaptation R R –  R 1/2 Consumption Rate (n a +1)C Total buffering to add an additional layer is O(R 3/2 ) rather than O(R 2 ) C

12. April 30, 2001MIT Laboratory for Computer Science12 Reduced Buffering SQRT requires less buffering to add layers! 56% less buffering to add 4 layers

13. April 30, 2001MIT Laboratory for Computer Science13 Conclusion Combination of SQRT congestion control with receiver quality adaptation enables smooth video delivery –Reduces rate oscillations –Reduces buffering/Increases interactivity Software is available –Includes selective reliability for packet loss –http://nms.lcs.mit.edu/software/videocm/

14. April 30, 2001MIT Laboratory for Computer Science14 Extra Slides

15. April 30, 2001MIT Laboratory for Computer Science15 Outline Problem Overview Background –Bandwidth Variation –Quality Adaptation –Binomial Congestion Control Approach Results Conclusion

16. April 30, 2001MIT Laboratory for Computer Science16 The Goal TCP-friendly congestion control Reduce rate oscillations: –Limit size of playout buffer –Smooth perceptual quality Limit receiver buffering for QA –Reach acceptable playout rate faster –More interactivity in certain cases (i.e., if RTT and RTT jitter are small)

17. April 30, 2001MIT Laboratory for Computer Science17 Results of SQRT Tested on emulated network conditions with Dummynet and SURGE toolkit SQRT reduces rate oscillations for: –Immediate adaptation –Receiver-buffered QA Also reduces buffering: –Less jitter due to rate oscillations –Backoffs less severe => less QA buffering –Can play out at higher layers more quickly –More interactivity

18. April 30, 2001MIT Laboratory for Computer Science18 Internet MPEG ServerMPEG Client RTP/RTCP CM data loss/RTT/requestsdata callbacks RTSP loss/RTT loss/RTT/requests SR-RTP

19. April 30, 2001MIT Laboratory for Computer Science19 Internet MPEG ServerMPEG Client RTP/RTCP CM data loss/RTT/requests data callbacks RTSP loss/RTT loss/RTT/requests SR-RTP Layering callbacks Bandwidth Adaptation System Architecture

20. April 30, 2001MIT Laboratory for Computer Science20 Reduced Oscillations In many cases, AIMD drops multiple layers in one backoff! This is not the case with SQRT. Rate oscillations in SQRT are much less pronounced than in AIMD. Layers Dropped AIMD SQRT Bitrate