Investigating and Improving Bittorrent’s Piece and Neighbor Selection Algorithms

In this thesis, we examine two important factors in the design of BitTorrent: how it chooses pieces and neighbors. We present a measurement study on the distribution and evolution of the pieces in BitTorrent. The data is collected by multiple administrated clients distributed in different parts of the network. Our results validate that the downloading policy of BitTorrent is effective, yet enhancements are still possible to achieve the ideal piece distribution. We also consider the topologies of multiple complex networks formed by neighbor selection in BitTorrent. Our results demonstrate that the networks exhibit fundamental differences during different stages of a swarm, and we discover the presence of a robust scale-free network in the network of peer unchokings. However, unlike previous studies, we find no evidence of persistent clustering in any of the networks. We therefore present a first attempt to introduce clustering, and verify its effectiveness through simulations and experiments.