Abstract
Peer-to-peer (P2P) networking is emerging as an important communication model which brings advantages of low deployment cost, high scalability, and high throughput for network applications. Unlike the conventional client-server model, P2P model allows services to be offered by all participating peers in a distributed and cooperative manner. The major application in P2P networking model is file sharing. Napster, eDonkey, LimeWire, and BitTorrent are famous and popular P2P-based file sharing systems where all peers share their files directly among themselves without the need of a vulnerable file repository. Retrieving a target file in the P2P file sharing system has two major phases: the target file is located during the discovery phase while it is transmitted during the delivery phase. In this thesis, we study the performance of file distribution in the delivery phase and focus on the BitTorrent system, which is designed for fast and efficient distribution of large files. The two issues concerned are the unfairness problem and the last piece problem. The former refers to the unbalance of data volume each peer serves, while the latter addresses download failures due to missing pieces. The unfairness problem discourages users from sharing their resources in most P2P-based systems while the last piece problem, which is prone to occur when the last seed leaves, degrades service availability and download speed. In this thesis, three schemes are proposed to resolve the two aforementioned problems in BitTorrent. One deal with peer fairness and the other two with the last piece problem. First, in order to enforce better peer fairness, Adaptive Optimistic Unchoking (AOU) suggests to maintain cooperation between similar peers and to restrict the use of optimistic unchoking. Second, Explicitly Acquiring Rare Piece (EARP) allows peers to explicitly acquire designated rare pieces outside its current peer set so as to rapidly distribute those rare pieces over the network. Finally, Interest-Intended Piece Selection (IIPS) tends to maintain piece diversity in a wider view of piece rareness as well as keep stable cooperation between peers for alleviating the last piece problem and improving overall download performance respectively. Simulation results show that the three proposed schemes successfully enhance the performance of BitTorrent. In one aspect, AOU achieves about 60% improvement on peer fairness while maintaining equivalent performance as original BitTorrent. In the other aspect, EARP outperforms BitTorrent's local-rarest-first (LRF) algorithm in terms of higher robustness as well as fewer occurrences of piece loss. Even under high peer dynamics, EARP still shows 69% less piece losses with comparison to LRF. Finally, IIPS successfully results in 28%-60% fewer occurrences of the last piece problem than LRF even under controlled tough environments.