COCONET: Co-Operative Cache driven Overlay NETwork for p2p VoD streaming Abhishek Bhattacharya, Zhenyu Yang & Deng Pan

Roadmap Introduction Basic Design Protocol Details Simulation Results Summary

Introduction Multimedia streaming applications such as live/on- demand are gaining popularity with millions of viewer every day in Internet P2P based approaches have been proved to be successful for reducing server load in scalable streaming Live Streaming systems already successfully deployed in Internet like PPLive, Coolstreaming, etc. Video On-Demand (VoD) systems can also benefit from P2P based approaches.

Introduction Fundamental challenges in VoD are in supporting VCR related operations such as Fast Forward(FF)/ Fast Backward(FB). Users can jump to any portion of the video stream unlike live streaming where always move forward. VCR operations lead to fundamentally lower levels of content overlap. Searching new supplier peers is more important to support random seek operations.

Basic Design Each peer contributes a part of its storage for building a global distributed co-operative cache which will also help others. This design philosophy is in accordance with P2P principle where each peer contributes some resource to the system in return of service. Storage cache at each peer is static and remains throughout its life in the system. A gossip based protocol is used which helps to efficiently disseminate content index among the peers.

Basic Design Video is divided into M equal segments where each segment corresponds to one minute length. Size of the storage cache is b segments where each peer randomly caches b segments from the server/peers after its entry into the system. Each peer maintains a SegmentMap(S M ) which contains the peer index information of all the M segments based on storage cache contents. Fault tolerance factor (c) entries are maintained for each S M entry to keep multiple supplier peers.

Protocol Details We exploit the following theorem from SCAMP: Thoerem: For a group size N and the partial view maintained at each peer as O(log N), the probability for a gossip to reach every member in the group converges to e -e -c provided the link/node failure probability is not greater than c/(c+1). Each COCONET peer needs to fill up c*M entries in its S M by gossiping its own b storage cache entries with other peers. The number of gossip partners required to effectively disseminate its own information and filling up S M is log(cM/b).

Protocol Details: Join/Leave Contacts the tracker to acquire a contact peer. New peer sends a join message to contact peer. Each peer maintains InView and OutView data structures. Contact Peer keeps the new peer in its Partial View and forwards the join message to all its OutView neighbors. At other peers on receiving the forwarded join, either it keeps the new peer in its OutView with a certain probability or forward it again to any one of its random OutView neighbor.

Protocol Details: Gossip Each peer exchange S M information with its gossip partners which should be a randomized subset of the entire system. The OutView neighbors are chosen as gossip partners with size log(cM/b) which will help for efficient information dissemination across the whole system. Excess gossiping is controlled by using ttl or with a certain probability. After a few rounds of gossiping, S M will contain c*M entries and the stale entries are removed periodically.

Protocol Details: Caching/Lookup Caching protocol is simple where each peer after joining randomly chooses b segments to cache. Download from peer if available or else from server. If the S M is maintained correctly then a lookup for any segment takes O(1) complexity with high probability. Suppose a peer performs a lookup for segment i, then it multicasts download request to all the c entries in S M [i]. Downloads segment i from the peer that replied first or from server as a last resort if all the c peers fail/shortage of upload bandwidth.

Simulation Results

Detailed results in the paper

Summary A distributed storage caching scheme helps to spread the query load uniformly through the overlay and organizes the overlay in a randomized fashion which makes the content distribution independent from the playing position unlike state-of-art P2P VoD systems. High content availability for any skewed viewing pattern. Lookup complexity of O(1) for continuous play/random seek to any position across the video stream with a high probability. Low control overhead even under high churn with good load balancing/ fault tolerant properties.