1. View-Upload Decoupling: A Redesign of Multi-Channel P2P Video Systems Keith Ross Polytechnic Institute of NYU

2. Today’s Talk  Overview of P2P Video Streaming  View-Upload Decoupling (VUD): A Redesign of P2P Video Streaming  Queuing Models for P2P Streaming 2

3. 3 More Details…  Patent Pending “View-Upload Decoupling: A Redesign of Multi-Channel P2P Video Systems”, Di Wu, Chao Liang, Yong Liu and Keith Ross, To appear in IEEE Infocom, Mini-conference, 2009. “Queuing Network Models for Multi-Channel P2P Live Streaming Systems”, Di Wu, Yong Liu and Keith Ross, To appear in IEEE Infocom, 2009.

4. Today’s Talk  Overview of P2P Video Streaming  View-Upload Decoupling (VUD): A Redesign of P2P Video Streaming  Queuing Models for VUD Streaming 4

5. 5 Peer-Assited Video Streaming  Peers redistribute video to each other  exploit peer uploading/buffering capacity  reduces load on server  Large scale deployments on Internet  thousands of live/on-demand channels  millions of world-wide users daily  Leading P2P Video Companies  CoolStreaming  PPStream  PPLive  Sopcast  UUSee 3 6 5 4 1 2

6. PPStream (http://www.pps.tv) #1 P2P Video System in the World Developed by Liang Lei and Hongyu Zhang (China) in 2005. 350M installations ~12 Million active users each day Thousands of channels Num of Channels 6

7. Common Features  Multiple Channels  Channel Churn  Heterogeneous Streaming Rates  HDTV Channels, VCR-quality channels,…  Heterogeneous Channel Popularities  Very few viewers in less popular channels.  Isolated Channel Design: ISO  Viewer only redistributes channel it is viewing 7

8. Problems of Traditional ISO Design  Large Channel Switching Delay  Existing P2P video systems: 10-60 seconds  Large Playback Lag  Existing P2P video systems: 5-60 seconds  Poor Small-channel Performance  Inconsistent and poor performance in small channels.  Root causes: channel churn and resource imbalance 8

9. 9 Channel Churn in ISO Design A B C D F E 3 6 5 4 1 2 viewers CF 31 Channel 1Channel 2

10. 10 Drawback: distribution systems disrupted when peers switch channels Channel Churn in ISO Design A 6 5 4 2 B D E Channel 1Channel 2 CF 31 viewers after channel switching

11. Resource Imbalance in ISO Design  Recall the instantaneous resource index for a channel of rate r with n viewers:  Ratio of available upload rate to required download rate  Channel in trouble if  Resource index can be imbalanced across channels  Small channels particularly volatile. 11

12. Today’s Talk  Overview of P2P Video Streaming  View-Upload Decoupling (VUD): A Redesign of P2P Video Streaming  Queuing Models for P2P Streaming 12

13. 13 2 C 1 B 3 A Channel 1 Channel 2 A Redesign of Multi-Channel System: View-Upload Decoupling (VUD) 54 6 F E D B 3 A 2 C 1 ED F 6 5 4 channel 1 substream1 channel1 substream2 channel2 substream1 channel2 substream2 viewers distribution swarms New Rule: each peer is assigned to semi- permanent distribution groups; independent of what it is viewing.

14. 14 2 C B F A Channel 1 Channel 2 54 6 F E D B 3 A 2 1 1 ED 6 5 4 channel 1 substream1 channel1 substream2 channel2 substream1 channel2 substream2 viewers distribution swarms after channel switching Advantage: distribution swarms not modified when peers switch channels A Redesign of Multi-Channel System: View-Upload Decoupling (VUD) 3C

15. Advantages of VUD design  Channel Churn Immunity  Distribution swarms unaffected by channel churn  Cross-Channel Multiplexing  Distribution swarms can be provisioned and adapted to balance resource indexes across channels  Structured Streaming  Scheduling and routing can be optimized within the stable VUD swarms 15

16. Key Challenges of VUD design  VUD Overhead  In ISO, peer only downloads video it is watching.  In VUD, each peer downloads its assigned substreams as well as the video it is watching.  Solution: substreaming  Adaptive Peer Assignment  Bandwidth allocation  Peer reassignment 16

17. Simulation Experiments  Simulated features:  Channel switching  Peer churn  Heterogeneous upload bandwidth  Packet-level transmission  End-to-end latency  Zipf-like channel popularity  Comparison  ISO: using Push-Pull scheduling  VUD: using Push-Pull scheduling 17

18. Simulation Parameters  50 channels  Video rate 400 kbps each channel  Server rate 1 Mbps for each channel  2,000 peers  Peer upload rates 128-768 kbps  Avg peer system time: 67 minutes  Channel churn follows IPTV study  5 substreams per channel 18

19. Channel Switching Delay  VUD achieves smaller channel switching delay. 19 Switching delay = time to acquire 5 seconds of new channel

20. Playback Lag  VUD achieves smaller playback lag. 20

21. Today’s Talk  Overview of P2P Video Streaming  View-Upload Decoupling (VUD): A Redesign of P2P Video Streaming  Queuing Models for P2P Streaming 21

22. Motivation  Develop a tractable analytic theory for multi-channel P2P live video systems.  Use model to study how to optimize VUD performance and perform bandwidth dimensioning  PS = probability of universal streaming = fraction of time resource index > 1 for ALL channels 22

23. Queuing Network Model  Each channel can be thought of as a queue  Each viewer as a customer  When viewer changes channels, routed to new queue  Customers move about channels independently:  infinite server queues  Let p ij is probability of switching channel i to j. P = [p ij ]  Let average sojourn time in channel j  Can do all kinds of cool things with this model!  Inspiration from the queuing and loss network literature. 23

24. Closed Queuing Network Model  Peers never leave (e.g., set-top box peers)  Now just apply the standard closed Jackson network theory  Traffic equation  Relative channel popularity:  n is the total number of peers  M j = # of viewers in channel j. 24

25. Open Queuing Network Model  Applicable for systems with Peer Churn  Peers arrive at constant rate and join channel j with prob p 0j  Peer leaves system with probability p j0.  In this talk, we focus on Closed Queuing Network Model. 25

26. Analysis of VUD Design  Resource Index for substream s of channel j  Probability of system-wide universal streaming where and 26 M j = # of viewers in channel j

27. Asymptotic Analysis of VUD  Under what conditions do all channels perform well when number of peers becomes large?  Fix number of channels J.  Let number of peers  Assume for simplicity no substreaming  Asymptotic regime: n j = K j n  How to dimension K j for large n? 27

28. Asymptotic Analysis for VUD  Initially assume homogenous upload rates: u i = u.  Critical parameter:  Theorem: If α > 1, then PS goes to 0 for all choices of K j. If α < 1, then PS goes to 1 if 28

29. Asymptotic Analysis for VUD  Heterogeneous peer types: low u l and high u h.  f = fraction of low peers (fixed)  Can find optimal peer allocations by solving:  If the value < 0, then PS goes to 0. 29

30. Analysis of ISO Design  Let M j be the random set of nodes viewing channel j.  Once again:  Can be solved used Monte Carlo methods and importance sampling. 30

31. Asymptotic Analysis of ISO  Heterogeneous peer types: low u l and high u h.  f = fraction of low peers (fixed)  Critical Value:  PS goes to 1 if α ≤ 1 and goes to 0 otherwise. 31

32. Asymptotic Analysis: Example  u h = 4r, u l = 2r, f = ½  r 1 = 5r, r 2 = r, ρ 1 =.2, ρ 2 =.8  ISO : α > 1  PS goes to 0  VUD :  allocate high-bandwidth peers to channel 1 ; low bandwidth peers to channel 2.  PS goes to 1 32

33. Numerical results  Results from analytical equations  1,800 peers  20 channels  u l =.2 r and u h = 3 r  Use asymptotic heuristic to dimension substream swarms 33

34. Numerical Results  Probability of System-wide Universal Streaming (PS)  Vary Zipf parameter Many small channels 34

35. Numerical Results of VUD Design  Probability of Universal Streaming in each channel  VUD achieves higher probability of universal streaming (PU j ) in small channels. Least popular channel 35

36. Refined Heuristic for VUD  Basic idea: equalize probability of universal streaming for all substreams:  Assume normal distribution for M j  Use known mean and variance  Assume all streams of same rate r 36

37. Refined Heuristics for VUD Num. of low-bandwidth peers in substream s of channel j Num. of high- bandwidth peers in substream s of channel j 37

38. Refined Heuristic for VUD Streaming  Probability of universal steaming in each channel.  Refined VUD can achieve higher probability of universal streaming in small channels. 38

39. Summary  Introduced a new design of P2P Video systems: View- Upload Decoupling (VUD)  Developed a tractable analytic theory to study ISO and VUD streaming  More insights about ISO and VUD designs.  Guidelines and Efficient Heuristics for VUD provisioning 39