LOCALITY-AWARENESS IN BITTORRENT-LIKE P2P APPLICATIONS R 黃琇琳 R 呂柏頡

Authors  IEEE Transactions on Multimedia, 2009  Bo Liu  Yi Cui  Yansheng Lu  Yuan Xue

Introduction  P2P applications introduce tremendous amount of traffic crossing the boundary of ISPs.  Such traffic often causes great financial loss to ISPs.  This paper presents a comprehensive study on various ways to embed locality-awareness into P2P applications and their impacts on the ISPs.

Two Facts Taken Into Account  ISPs themselves interconnect into a complex network of autonomous systems (AS).  A P2P application is usually composed of sophisticated semantics.  Take BitTorrent as an example.

- Obtaining AS-Level Map. - Evaluation Setup. Evaluation Methodology

Obtaining AS-Level Map  Construct on the PlanetLab testbed.

Evaluation Setup  Two application scenarios the evaluation covers.  Downloading V.S. On-Demand Streaming Downloading V.S. On-Demand Streaming  Optimal strategy as the theoretical baseline.  Minimum AS-Hop Strategy Minimum AS-Hop Strategy

- BitTorrent - Locality-Aware BitTorrent Locality-Aware BitTorrent.

BitTorrent  Neighbor Selection  Tracker randomly generates a list of peers.  Choking / Unchoking  Peer sends data to neighbors which have highest uploading rate.  Rarest First Piece Picking  Download the piece which is rarest among its neighbors.

Locality-Aware BitTorrent  Tracker locality.  Choker locality.  Piece picker locality.  Accommodation to streaming scenario.

Tracker Locality  Tracker sorts all other peers in the swarm by their distances to the requesting peer in terms of AS hop count.  Send the prefix of the sorted list to the requesting peer.

Choker Locality  Peers unchoke the four neighbors that are closest to itself in terms of AS hop count.  This policy will not result in the same selection of peers again and again.  A seed will keep unchoking four of its closest neighbors who still have not finished downloading.

Piece Picker Locality  Introduce a distance value to each piece, which is the mean value of the distances of all peers possessing this piece.  Download the piece closest to itself.

Accommodation to Streaming Scenario  Restrain the piece picking action within a window marching with the video playback.  The window is automatically pushed forward whenever its leftmost piece is downloaded.  A stream-watcher process is used when the downloading falls behind the play back.

Findings

 Downloading time in downloading scenario:  Choker and Picker locality perform better.  Downloading time are very uneven with Tracker locality. User-Perceived Performance

 Interruptions in streaming scenario:  Tracker locality makes the most number of peers suffering interruptions.

 AS hop count in downloading scenario:  Tracker locality makes the shortest AS hop count.  All policies in unlimited seeding get lower hop count. Locality-Related Performance

 Redundancy in downloading scenario:  Majority of them achieve the minimum value across all solutions, due to the fact each ISP only hosts one peer.

- Standard BitTorrent achieves similar disruption as Choker and Picker locality and less disruption than Tracker locality. - Tracker locality achieves the lowest AS hop count. - Standard BitTorrent achieves similar disruption as Choker and Picker locality and less disruption than Tracker locality. - Tracker locality achieves the lowest AS hop count. - Choker and Picker locality can significantly reduce downloading time. - Tracker locality achieves the lowest AS hop count, but suffers most unbalanced peer load. - Choker and Picker locality can significantly reduce downloading time. - Tracker locality achieves the lowest AS hop count, but suffers most unbalanced peer load. Downloading scenario On-Demand Streaming Conclusion This study suggests the necessity to consider, in the design of future P2P downloading and streaming solutions.

- Peers start viewing the video at different times. - Support the “viewing-while-downloading” feature. - A peer is more likely to download from an earlier-joined peer. - Video file must be downloaded in an approximately sequential fashion. - Peers start viewing the video at different times. - Support the “viewing-while-downloading” feature. - A peer is more likely to download from an earlier-joined peer. - Video file must be downloaded in an approximately sequential fashion. - All peers show the interest to the file at the same time. - Join the P2P network simultaneously. - All peers initiate downloading under the same condition. - All peers show the interest to the file at the same time. - Join the P2P network simultaneously. - All peers initiate downloading under the same condition. Downloading scenario On-Demand Streaming Most important metric Downloading time Most important metric Interruption time

Minimum AS-Hop Strategy  For downloading scenario:  Construct a complete graph, where each node represents a peer, and edge weight represents the AS hop count.  Finds the minimum spanning tree on this graph.  For on-demand streaming scenario:  The complete graph becomes directed, where at each edge the earlier-joined peer directs to the later-joined peer.