The Research On Trust Model And Replication Scheme In P2P Environments

Nowdays Peer-to-Peer (P2P) computing has become a hot research issue. The goal of P2P computing is to pool end-systems in Internet-scale as more as possible to construct large-scale cooperating and resource-sharing environments.Because the following two problems directly decide the usability and pracitcabiltiy of P2P system, they have been in dire need of a solution. (1) In decentralized P2P environments, due to the anonymous and autonomous nature of peers, they have to manage the risk involved with the interactions without prior experience and knowledge about each other's reputation. Trust models can be used to quantify and to evaluate the trustworthiness of peers so as to reduce the risk. (2) P2P systems may operate in a highly dynamic environment in which the popularity of the stored content varies with time, nodes enter and leave the system at will, and content can be added and deleted at any time. These facts significantly raise the level of complexity involved in managing all available resources so as to provide efficient access to the stored content. Replication is used by most P2P systems as a basic means to achieve availabity and fault-tolerance.The dissertation focuses on these two main problems, and makes a number of contributions:1. A layered architecture is presented, which cleanly separates the functional components of a structcured P2P system. This architecture enables us to tailor the P2P infrastructure to the specific needs of various Internet applications, without having to devise completely new storage management and index structures for each application.2. A comprehensive survey of trust models in P2P environments is made. The newly-proposed models are analyzed, classified and compared.3. SWRTrust, a global trust model, is presented, which includes a mathematical description and a distributed implementation. Trust models in P2P environments have two main challenges: various malicious behavior of peers such as providing fake or misleading recommandation about other peers; decentralized and secure trust data management. Previous global trust models are based on the assumption that the peers with high trust value will givethe honest recommendation. We argue that this assumption may not hold in all cases. In SWRTrust, each peer is assigned a unique global trust value, computed by aggregating similarity-weighted recommendations of the peers who have interacted with it. Theoretical analyses and experiments show that the scheme is still robust under more general conditions where malicious peers cooperate in an attempt to deliberately subvert the system, converges more quickly and decreases the number of inauthentic files downloaded more effectively than previous schemes. Furthermore, a decentralized trust data management scheme is given and the anonymity requirements of peers for securing trust data are also considered.4. Two algorithms are proposed to improve the method for measuring the similarity between peers' rating vectors. (1) In a large-scale P2P environments, the extreme sparsity of peers' ratings leads to undermine the accuracy of similarity assessment. The C_Similarity algorithm is given to exploit the sparisity of vetors. By aggregating the ratings, the ratings to a class of peers instead of these to a peer are used to measure the similarity. (2) To compute cosine-based similarity between two n-dimensional rating vectors involves 3n multiplications. To update a peer's trust value requires O(n~2) multiplications. The Simplified_Simi algorithm is designed to simplify cosine-based similarity. The similarity between two vectors is then measured by comparing each element in them and counting the number of elements that have the same sign. Thus the similarity computation is scalable in a large-scale P2P environment. Simulation results show that these two algorithms work.5. A general-purpose replication scheme, Genre, is proposed for structured P2P systems. The drawbacks of the two typical replication schemes are analyzed: not achieving load-balance between replicas and requiring O(k) messages for each join and leave event to maintain a replica degree of size k. We present the full algorithmic specification of Genre which only needs O(1) messages for every join and leave operation to maintain the replicas, independent of the number of the replicas. And for any given data object, the load-balance between its replicas can be achieved. By building a binary-tree for each data object, Genre keeps track of the locations of replicas and then propagates update notifications. Theoretical analyses and experimental results show that Genre can effectively maintain replicas while requiring the fewer messages than the typical schemes, and its update operation is scalable.