Prediction-based Prefetching to Support VCR-like Operations in Gossip-based P2P VoD Systems Tianyin Xu, Weiwei Wang, Baoliu Ye Wenzhong Li, Sanglu Lu, Yang Gao Nanjing University Dislab, NJU CS

Nanjing University 2 Outline  Background  P2P VoD streaming; Gossip-based systems; VCR-like interactive behavior.  Motivation  Solutions  System architecture; Prefetching model; Data scheduling; VCR-like operation support. Data scheduling; VCR-like operation support.  Performance Evaluation  Conclusions Dislab, NJU CS

Nanjing University Background (1)  P2P media streaming  Everyone can be a content producer/provider.  Cache-and-relay mechanism: peers actively cache media contents and further relay them to other peers that are expecting them. 3 * P2P live streaming is very successful! -CoolStreaming (INFOCOM’05), -PPLive, Joost Dislab, NJU CS

Nanjing University Background (2)  P2P VoD streaming is challenging!  Provide free access to any segment in the video at anytime by VCR-like operations.  VCR-like (Video Cassette Recorder) operations  random seek, pause, fast forward/backward (FF/FB)  For VCR-like operations, “jump” process is the most important. Most VCR-like operations can be implemented by “jump”. –random seek & pause: 1 jump; FF/FB: series of jump; 4 Dislab, NJU CS

Nanjing University Motivation (1)  How to support the “jump”?  Optimizing the index overlay to realize fast segment relocation Jump => locate-and-download process; Necessary, but far more sufficient.  Prediction-based Prefetching Expect a zero jump delay; Proactively prefetch segments that are likely to be requested by future VCR- like operations; Rely on prediction accuracy. 5 Question: Is the prediction feasible? Dislab, NJU CS

Nanjing University User Access Patterns (1)  User rarely view the movie from the beginning to the end.  The total playing time of a user is quite limited and tends to be short.  Because some popular segments (called highlights) attract more user requests than non-popular segments. Brampton et al., NOSSDAV-2007 Zheng et al., P2PMMS Dislab, NJU CS

Nanjing University User Access Patterns (2)  Probability distribution of object and segment popularity  Log-normal distribution  Zipf distribution Brampton et al., NOSSDAV-2007 Yu et al., EUROSYS Dislab, NJU CS

Nanjing University User Access Patterns (3)  Fast Forward is more frequent than Fast Backward.  Short Jump is more frequent than Long Jump. Cheng et al., IPTPS-2007 Cheng et al., IPTPS-2007 Brampton et al., NOSSDAV-2007 Brampton et al., NOSSDAV Dislab, NJU CS

Nanjing University Motivation (2)  Our Objective: Effective Prediction-based Prefetching Scheme Effective Prediction-based Prefetching Scheme  Effective prediction model Based on user access patterns  Easy to be integrated in current P2P VoD systems  Practical data scheduling 9 Dislab, NJU CS

Nanjing University System Architecture (1)  Solution 1: Let the server do prediction for each user [1]  Pro: Server has large volumes of user viewing logs  Con: poor scalability  Solution 2: Let the client exchange user logs and do prediction [2]  Pro: scalable  Cons: 1. lack of large volumes of user logs 2. high computing cost & training time [1] Huang et al, “A User-Aware Prefetching Mechanism for Video Streaming”, WWW [2] He et al, “VOVO: VCR-Oriented Video-On-Demand in Large-Scale Peer-to-Peer Networks”, TPDS Our solution: Server side: offline pattern mining => prediction model Peer side: lightweight online prediction Dislab, NJU CS

Nanjing University System Architecture (2)  Take full advantage of tracker  Tracker has large volume of user viewing logs;  Every node have to contact the tracker to join the system initiate its neighbor & partner list 11 Dislab, NJU CS

Nanjing University Prediction Approach: Overview  Frequent Sequential Pattern Mining  PerfixSpan[1] : Mining Sequential Patterns Efficiently by Prefix- Projected Pattern Growth.  Splitting Video Segments into Abstract States  Mapping User Logs to Abstract States  Construct Contingency Table (CCT)  Model Utilization [1] Pei et al., “Mining Sequential Patterns by Pattern Growth: The PrefixSpan Approach”, TKDE Dislab, NJU CS

Nanjing University Prediction Approach (1) Frequent Sequential Patterns Dislab, NJU CS

Nanjing University Prediction Approach (2)  Sequential patterns found may be overlapped?  e.g. and  Splitting Approach  Filter out the sub-patterns e.g.,,,  Scan over the remaining sequential patterns Cut them into intervals without overlapping - e.g. and  [1,7],[8,12]  Take intervals not exist in the mined sequential patterns as separate intervals  Split the contiguous intervals into appropriate granularity intervals(States) - MIN, MAX 14 Dislab, NJU CS

Nanjing University Prediction Approach (3)  Map Raw User logs into State Transitions   e.g. map to [1,6]  [7,13]  Transition Table Construction  Simple Frequency Counting 15 Dislab, NJU CS

Nanjing University Data Scheduling  Two stage scheduling strategy:  Stage 1: fetch urgent segments into playback buffer Guarantee the continuity of normal playback Urgent line mechanism [1]  Stage 2: prefetch based on prediction Reduce jump latency Utilize residual bandwidth [1] Li et al., “ContinuStreaming: Achieving High Plackback Continuity of Gossip-based Peer-to-Peer Streaming”, IPDPS Dislab, NJU CS

Nanjing University VCR-like Operation Support  The jump process caused by VCR-like operations: Case 1. The jump segment is already prefetched on the local peerCase 1. The jump segment is already prefetched on the local peer => Just playback!! Case 2. The jump segment is cached on the partners’ bufferCase 2. The jump segment is cached on the partners’ buffer => download and playback! Case 3. Neither cached on the local peer nor cached by the partnersCase 3. Neither cached on the local peer nor cached by the partners => relocate, connect and download 17 Dislab, NJU CS

Nanjing University Simulation Settings  User Log Generation Modify GISMO [1] –Using log-normal distribution to let users trend to jump around hot scenes.  The simulation is built on top of a topology of 5000 peer nodes based on the transit-stub model generated by GT-ITM.  The streaming rate is S = 256 Kpbs, the download bandwidth is randomly distributed in [1.5S, 5S].  The default size of the playback buffer is 30Mbytes, i.e., each peer can cache 120 second recent stream (100 for playback, 20 for prefetching).  The arrival of peers follows the Poisson Process with λ = 5. [1] GISMO: A Generator of Internet Streaming Media Objects and Workloads 18 Dislab, NJU CS

Nanjing University Performance Evaluation (1) 19 Dislab, NJU CS

Nanjing University Performance Evaluation (2) 2 Dislab, NJU CS

Nanjing University Performance Evaluation (3) 3 Dislab, NJU CS

Nanjing University Performance Evaluation 4 Dislab, NJU CS

Nanjing University Conclusions  A practical architecture that can be used in almost all existing P2P VoD systems  A novel and simple prediction approach  State abstraction plays an important role  A two stage data scheduling 23 Dislab, NJU CS

Nanjing University 24 Dislab, NJU CS The End