Large-scale Peer-to-Peer Network Statistical Analysis, Characterization And Its Applications

The population of peers grows much large than ever before in today's peer-to-peernetwork, i.e., the number of online peers of a popular file sharing P2P system oftenreaches several million, while instance message system has much more users. However,many problems emerge during this period, including:Researchers do not precisely know the current status of peer-to-peer network, whichalso makes some research aimless.Lack of the information of peer-to-peer network structure as well as user activities,many simulation studies can only be performed based on heuristic models, whichare very likely to be unrealistic.The design for future P2P system can not be started or be beneficial withoutdetailed performance statistics.As the above phenomenon clearly suggested, P2P network statistical study andcharacterization is indispensable. To address the problems mentioned above, we per-formed a statistical study over a popular file sharing system, Gnutella 0.6, and proposedsome related applications. Our work mainly includes:(1) Designing and implementing an e?cient distributed collaborative peer-to-peercrawling system, which provides a solution for collecting statistic information ofGnutella 0.6.(2) Analyzing the characteristics of Gnutella's query message stream from a time seriesaspect, and evaluated the predictability of Box-Jenkins models for the number offuture query messages.(3) Proposing two applications based on Box-Jenkins models, one is collaborative re-sult caching in Gnutella, the other is enhanced dynamic load balancing in Chord.The contribution of our work mainly lies in the following areas:(1) Discovering new and important phenomena in Gnutella 0.6, such as peer roleselection, which provides target for future research.(2) Providing reliable statistic data for modeling and simulating of peer-to-peer net-work.