Network and Systems Laboratory nslab.ee.ntu.edu.tw Yipeng Zhou, Dah Ming Chiu, and John C.S. Lui...

download Network and Systems Laboratory nslab.ee.ntu.edu.tw Yipeng Zhou, Dah Ming Chiu, and John C.S. Lui Information Engineering Department The Chinese University.

If you can't read please download the document

description

Network and Systems Laboratory nslab.ee.ntu.edu.tw Outline Introduction Model Chunk selection strategy Simulation results Conclusion

Transcript of Network and Systems Laboratory nslab.ee.ntu.edu.tw Yipeng Zhou, Dah Ming Chiu, and John C.S. Lui...

Network and Systems Laboratory nslab.ee.ntu.edu.tw Yipeng Zhou, Dah Ming Chiu, and John C.S. Lui Information Engineering Department The Chinese University of Hong Kong Presenter: Ben Presenter: Ben Network and Systems Laboratory nslab.ee.ntu.edu.tw Outline Introduction Model Chunk selection strategy Simulation results Conclusion Network and Systems Laboratory nslab.ee.ntu.edu.tw Outline Introduction Model Chunk selection strategy Simulation results Conclusion Network and Systems Laboratory nslab.ee.ntu.edu.tw Introduction Video streaming under a large amount of peers Unicast Multicast Overlay network P2P File-sharing protocol Streaming protocol Contributions Network and Systems Laboratory nslab.ee.ntu.edu.tw Unicast Directly from server to clients Network and Systems Laboratory nslab.ee.ntu.edu.tw Multicast Routers manage the distribution and duplication of contents Network and Systems Laboratory nslab.ee.ntu.edu.tw Overlay Network Ignore the realistic topology The most scalable one: P2P Network and Systems Laboratory nslab.ee.ntu.edu.tw P2P Each client is also a server File-sharing protocol No real-time demand Higher reliability demand BT, emule, Streaming protocol High real-time demand Lower reliability demand PPLive, PPStream, CoolStreaming, Network and Systems Laboratory nslab.ee.ntu.edu.tw Contributions of this paper Analytical models on p2p streaming system to better understand. Basic chunk selection strategies study. A new strategy is proposed. Trade off between continuity and scalability. Network and Systems Laboratory nslab.ee.ntu.edu.tw Outline Introduction Model Chunk selection strategy Simulation results Conclusion Network and Systems Laboratory nslab.ee.ntu.edu.tw Assumptions Mathematical model assumptions: M homogeneous peers with the same playback requirement. Each has a playback buffer that can cache up to n chunks. How does buffer work? Server sends out chunks sequentially. Peer downloads one chunk per time slot. Buffer shifts ahead one position per time slot. Network and Systems Laboratory nslab.ee.ntu.edu.tw Model Network and Systems Laboratory nslab.ee.ntu.edu.tw Mathematics Mathematical model In each time slot, the server randomly selects one peer and uploads one chunk. Users metric is the continuity, defined as p(n), the probability chunk n available. To compute p(n), recursively compute p(i). p(i) is defined as: p(i) = prob(position i filled). Network and Systems Laboratory nslab.ee.ntu.edu.tw Evaluation Without P2P Without P2P, server is the only video source: Performance: p(n)=p(n-1)= =p(1)=1/M Network and Systems Laboratory nslab.ee.ntu.edu.tw Evaluation With P2P With P2P: Each peers buffer is a sliding window. In each time slot, each peer downloads a chunk from server or its neighbor. q(i): the probability buffer(i) gets filled at this time slot. Performance: p(i+1) = p(i) + q(i) P2P technology effect Network and Systems Laboratory nslab.ee.ntu.edu.tw P2P Technology Effect Whats the value of q(i)? Think about this: w(i) = probability peer wants to fill buffer(i) w(i) = 1 p(i) h(i) = probability the selected peer has it h(i) = p(i) s(i) = buffer(i) selected by chunk selection strategy q(i) = w(i) * h(i) * s(i) Network and Systems Laboratory nslab.ee.ntu.edu.tw Outline Introduction Model Chunk selection strategy Simulation results Conclusion Network and Systems Laboratory nslab.ee.ntu.edu.tw Chunk Selection Strategy How to select what chunk we need now? Greedy strategy Try to fill the empty buffer closest to the playback. Rarest first strategy Try to fill the empty buffer for the newest chunk since p(i) is an increasing function. Consider this: Rarest first strategyGreedy strategy Network and Systems Laboratory nslab.ee.ntu.edu.tw Evaluation Greedy p(i+1)=p(i)+ (1-p(i)) * p(i) * (1-p(1)-p(n)+p(i+1)) w(i) buffer(i) is empty. h(i) selected peer has buffer(i). s(i) any particular chunk is neither downloaded into buffer positions between n to (i+1), nor is it downloaded directly from server. Network and Systems Laboratory nslab.ee.ntu.edu.tw Evaluation Rarest First p(i+1)=p(i)+ (1-p(i)) * p(i) * (1-p(i)) w(i) buffer(i) is empty. h(i) selected peer has buffer(i). s(i) any particular chunk is not downloaded into buffer positions between 1 to (i-1). Network and Systems Laboratory nslab.ee.ntu.edu.tw Evaluation Continuity Peer population and continuity Greedy strategy: RF strategy: where n = buffer size; M = population size = 1/p(1) = discontinuity = (1 p(n)) Network and Systems Laboratory nslab.ee.ntu.edu.tw Comparison Greedy strategy: Achieve better continuity for small number of peers. Rarest first strategy: More scalable as peer population increases. Network and Systems Laboratory nslab.ee.ntu.edu.tw A NEW Strategy Mixed of Greedy and Rarest first Cut the buffer into two halves. Difference equations: p(1)=1/M p(i+1)=p(i) + p(i) (1-p(i)) (1-p(i)) for i = 1,,m-1 p(i+1)=p(i) + p(i) (1-p(i)) (1-p(m)-p(n)+p(i+1) ) for i = m,,n-1 Network and Systems Laboratory nslab.ee.ntu.edu.tw Outline Introduction Model Chunk selection strategy Simulation results Conclusion Network and Systems Laboratory nslab.ee.ntu.edu.tw Definitions What do you mean by better? Playback continuity: p(n) as large as possible Start-up latency for new members: wait until its buffer has reached steady state where R is the average downloading rate. Network and Systems Laboratory nslab.ee.ntu.edu.tw Simulation Setup M = 1000 N = 40 In simulation number of neighbors = 60 Uploads at most 2 chunks in each time slot for one peer Network and Systems Laboratory nslab.ee.ntu.edu.tw Simulation for Validation Validate the models!! Network and Systems Laboratory nslab.ee.ntu.edu.tw Comparison between Strategies Modified Network and Systems Laboratory nslab.ee.ntu.edu.tw Performance Comparison 1000 peers, buffer length varies from 20 to 50 Mixed achieves best continuity than both RF and Greedy, and has better start-up latency than RF Network and Systems Laboratory nslab.ee.ntu.edu.tw Performance over Time AMAZING! Network and Systems Laboratory nslab.ee.ntu.edu.tw Sensitivity Study With fixed peers (40), greedy strategy is better. With fixed continuity (0.93), Rarest first strategy is better. Network and Systems Laboratory nslab.ee.ntu.edu.tw Outline Introduction Model Chunk selection strategy Simulation results Conclusion Network and Systems Laboratory nslab.ee.ntu.edu.tw Conclusion Given buffer size and relatively large peer population, and based on the models: Rarest first strategy is better in continuity. Greedy strategy is the best in start-up latency. Mixed strategy is the best in continuity and better in start-up latency than RF. Network and Systems Laboratory nslab.ee.ntu.edu.tw Discussion Propose a modified strategy for P2P networks. Propose an analytic model for P2P networks. But not very realistic, assume too many Paper filled with mathematics Network and Systems Laboratory nslab.ee.ntu.edu.tw The End Thank you! Q&A Network and Systems Laboratory nslab.ee.ntu.edu.tw Q&A Parameter m