Exploring Proximity Based Peer Selection in a BitTorrent-like Protocol
Asfandyar Qureshi · 2004
BitTorrent [1] is popular file-sharing tool, accounting for a significant proportion of Internet traffic. Files are divided into fragments and transferred out of order among nodes trying to download them. Individual nodes penalize and reward other nodes, adjacent in the BitTorrent overlay, depending on how willing others are to share data. This incentives scheme and the subsequent enforcement of sharing is credited with making BitTorrent outperform other contemporary file-sharing systems. Nonetheless, BitTorrent builds its overlays by randomly selecting peers, a fact that has the potential to seriously handicap both individual performance and waste global network resources. This paper investigates a scheme which attempts to build a more intelligent overlay network, particularly using synthetic network coordinates to select overlay peers that are close by in the underlying network. We evaluate our techniques both from the network’s perspective (resources used) and from the individual’s perspective (average time to complete a download). In both cases, our results show that better peer selection can lead to improved performance with no major changes to the basic BitTorrent protocol. Our evaluation is based on real-world experiments performed over Planet-Lab [2].