Community Models In Peer-to-Peer Networks And Their Applications In Search

Nowadays, peer-to-peer applications are getting popular among all the Internet applica-tions. Although there are many users using peer-to-peer applications, it is still quite difficultperform do text search in peer-to-peer networks, because of the difficulty in acquiring globalstatistical information. This dissertation aims at improving the efficiency of text search insuch networks by forming user communities and using partial statistical information to esti-mate the global statistical information.First, we propose an effective meta data to describe peer interests. It can correctly por-tray user interest, be updated incrementally, be generated and updated by local informationonly, and be generated and updated implicitly as well. The meta data is built using topicmodel, which comes from the idea of language model, a popular method in informationretrieval. The calculation of peer interest similarity is based on fuzzy set theory. The experi-ments show that the size of the meta data is quite small, which makes it appropriate for beingtransferred across the network. In addition, it is also quite accurate.Using the proposed meta data, we also present an effective community formation mech-anism in unstructured peer-to-peer networks. The mechanism frequently exchanges the metadata describing peer interest and information that are useful for community formation whenthe network resource is available. It forms explicit user community. Compared with themechanism presented by Khambatti et al. for Small World networks, our mechanism isbetter because it is more appropriate for Internet whose degree of nodes follows Power Law.To form user community in structured peer-to-peer networks, there exists an importantproblem that is the unbalanced load. From the perspective of either forming community only,or improving text search efficiency in peer-to-peer networks, the situation of unbalanced loadis serious because of both the non-uniform distribution of text data and the requirement thatsome indexing mechanisms should store similar items on neighboring peers. To solve theproblem, we first analyze it in detail and design a system that can adjust load to be balanced.The system is named cuckoo ring. Different from the traditional distributed hash basedsystems, it adopts locality sensitive hash functions. It can move load from hot areas to coldones, to make the load balanced. Finally,our experiment platform is introduced. It is a general peer-to-peer networksimulator, named GRASPS. Since the special characteristics of peer-to-peer applications, itis impossible for us to ask a large population of users to use an experimental application totest the performance of various peer-to-peer mechanisms. So, a general simulator is neededto substitute the users. Considering these mechanisms can be quite different in their networkstructures, peer actions, resource distributions and the difference between peer behaviors likeonline/of?ine, sharing resources, the simulator should be quite general and effective enough.We divide the whole system into five layers to simulate all the details. They are networklayer, protocol layer, security layer, application layer and peer layer. We also adopt sometechniques to improve the simulation efficiency. Compared with the traditional peer-to-peersimulators, our simulator is more scalable and consumes less system resource.