Ruchir Bindal, Pei Cao, William Chan, Jan Medval, George Suwala, Tony Bates, Amy Zhang
Using simulations, we show that biased neighbor selection maintains the nearly optimal performance of BitTorrent in a variety of environments, and fundamentally reduces the cross-ISP traffic by stopping it from growing linearly with the number of peers. A key reason for its performance is the rarest first piece replication algorithm used by BitTorrent clients.
Compared with existing locality-enhancing approaches such as bandwidth limiting, gateway peers, and caching, biased neighbor selection requires no dedicated servers and scales to a large number of BitTorrent networks. Furthermore, it can be combined with bandwidth limiting and caching to improve their performance further.
The paper in PDF is here.
The source code for the simulator can be found here.