1. 1 A Framework for Lazy Replication in P2P VoD Bin Cheng 1, Lex Stein 2, Hai Jin 1, Zheng Zhang 2 1 Huazhong University of Science & Technology (HUST) 2 Microsoft Research Asia (MSRA) NOSSDAV 2008, Braunschweig, Germany, May 30, 2008

2. 2 Background VoD, popular Internet service -Youtube, Hulu P2P, useful technology -File sharing, live streaming -BitTorrent, PPLive GridCast with caching -36% decrease -43% departure misses Replication in P2P VoD Can P2P help VoD? -Feasibility -Performance improvement

3. 3 Outline Replication algorithms 2 Conclusions 4 Performance evaluation 3 3 Motivation 3 1

4. 4 Motivation -what does GridCast look like? http://www.gridcast.cn

5. 5 Motivation -GridCast system overview Hybrid architecture (client-server + P2P) ―Tracker: indexes all joined peers ―Source Server: stores a complete copy of every video ―Peer: fetches chunks from source servers or other peers ―Web Portal: provides the video catalog tracker Source Server Web portal

6. 6 Motivation -trace collection GridCast has been deployed on CERNET since May 2006 ―Network (CERNET) 1,500 Universities, 20 million hosts Good bandwidth, 2 to 100Mbps to the desktop (core is complicated) ―Content 2,000 videos 48 minutes on average 400 to 800Kbps, 610 Kbps on average

7. 7 Motivation -trace analysis Classify misses by their causes Chunk X does not hit in the peer cache, Why?  New content ―Never fetched by any peer  Peer departed ―Fetched by some peers, but all of them are offline  Peer evicted ―Fetched by an online peer, but evicted  Can not connect ―Cached by some online peer that is not in the neighborhood  Insufficient bandwidth ―Cached by some neighbor, but cannot retrieve it Departure misses become a big issue 43%

8. 8 Motivation -challenges and chances Replication Caching is not enough. Can we do better? Challenges Short user sessions Depart at any time Chances Unused network resource 72% (DOWN), 81% (UP) Disk space 37% available disk

9. 9 Replication - three key questions Framework When ? Where? What ?

10. 10 Replication –fundamental tradeoff Benefit: Reduce departure misses Reduce some eviction misses if the cache is not full Cost: Increase network traffic Increase bandwidth misses Increase some eviction misses if the cache is full

11. 11 Replication -eager replication x x neighborhood A B C  Replicate all missed chunks  Use all of unused bandwidth

12. 12 Replication -lazy replication neighborhood A B C  Based on two predictors ―Peer departure predictor ―Chunk request predictor ―Lazy-oracle and lazy-simple  Lazy factor ―How much remained bandwidth can be used  Target peer selection ―Random, Sequentially, File locality first the increasing of chunk requeststhe increasing of online time

13. 13 Replication -peer departure predictor Based on the observation of online time -50% of user session, less than 10 minutes -the peer with higher online time is likely to stay longer Simple departure predictor -online time <= 10 minutes, leave -online time > 10 minutes, stay

14. 14 Replication -chunk request predictor Chunks requested recently are more likely to be requested earlier in the near future Simple chunk request predictor -use the chunk access history in the last several hours -give higher weight to the recent requests t 1234 futurehistory now 3 4 8 6 popularity

15. 15 Performance Evaluation -simulation setup  Trace-driven ―1GB ―Realized bandwidth ―Last 1 hour history for chunk request predictor ―10 minutes interval for peer departure predictor ―Use the existing neighborhood  Metrics ―Benefit: decrease of chunks served by the source servers ―Cost: increase of chunks replicated between peers ―Efficiency: Benefit / Cost

16. 16 Performance Evaluation -exploring configurations File locality first achieves the best performance

17. 17 Performance Evaluation -lazy factor -More chunks are delayed to be replicated when the peer leaves -Smaller lazy factor, more efficient Lower lazy factor is better

18. 18 Performance Evaluation -comparison  Lazy-simple is close to lazy-oracle, in terms of benefits  Lazy-simple is better than eager, in terms of efficiency  Lazy-simple, 15% decrease of server load

19. 19 Conclusions 1 We identify that departure miss is a major issue for P2P VoD with caching 2 With two simple predictors, lazy replication can decrease server load by 15% 3 Lazy replication is more efficient than eager replication

20. 20 Thank you! Any questions…… Bin Cheng, Lex Stein, Hai Jin and Zheng Zhang HUST and MSRA Huazhong University of Science & Technology Microsoft Research Asia NOSSDAV 2008, Braunschweig, Germany