Improving Traffic Locality in BitTorrent via Biased Neighbor Selection Ruchir Bindal Jan Medved Pei Cao George Suwala William Chan Tony Bates Department of Computer Science Amy Zhang Stanford University Cisco Systems, Inc.
[email protected] Abstract transfers among randomly chosen sets of peers distributed over the Internet, it proves to be costly for ISP’s. Peer-to-peer (P2P) applications such as BitTorrent ig- ISPs often control BitTorrent traffic by “throttling”, or nore traffic costs at ISPs and generate a large amount of bandwidth limiting. Since BitTorrent traffic typically runs cross-ISP traffic. As a result, ISPs often throttle BitTor- over a fixed range of ports (6881 to 6889) [6] and is easily rent traffic to control the cost. In this paper, we examine decoded, traffic shaping devices such as [20, 13, 19, 23] are a new approach to enhance BitTorrent traffic locality, bi- deployed to limit the amount of bandwidth consumed by the ased neighbor selection, in which a peer chooses the ma- BitTorrent protocol. However, this mainly slows down the jority, but not all, of its neighbors from peers within the content transfer and worsens the user download experience, same ISP. Using simulations, we show that biased neighbor not addressing the fundamental concern of the ISP, which is selection maintains the nearly optimal performance of Bit- to improve the locality (i.e. reduce the cross-ISP traffic) of Torrent in a variety of environments, and fundamentally re- those transfers. duces the cross-ISP traffic by eliminating the traffic’s linear Many analytical and simulation studies [16, 1, 25, 22] growth with the number of peers.