A Study On Key Techniques Of Large Scale Peer-to-peer Resource Sharing

Compared with Client/Server (C/S) model, Peer-to-Peer (P2P) is an emerging computing model, which can exploit resource on the edge of network efficienty. P2P is characterized by self-organization, large scale, scalable, efficient, and low cost. Many robust and large scale Internet applications can be constructed with P2P architecture. Among them, the most important one is large scale P2P resource sharing which accounts for the majority of today's Internet traffic. Therefore, research on large scale P2P resource sharing techniques has significient theoretic and practical values.The key points of large scale P2P resource sharing are resource search and conent distribution. P2P is a fully distrubted system. Resources are distributed throughout the network. There are no global state and a central control node. These attributes challenge the large scale P2P resource sharing. Although many problems have been solved, new problems are emerging with the operations of real systems, which include efficient search algorithms for rare items and efficient content distribution paradigms.To address above problems, this work focuses on resource search and content distribution. Firstly, new replication strategies are employed to improve the search performance of rare items in unstructured P2Ps. Then, this work attacks problems in P2P content distribution in terms of mathematical model, architecture, and traffic optimization. The main contributions of this work are as follows:1. A novel item replication and search strategy is proposed, which is DRS. Unstructured P2P networks are characterized by power-law distribution of node degrees. Firstly, biased-random walk can cover high degree nodes rapidly in power-law random graphs, which are validated by both mathematical analysis and simulations. Then, replication and search strategy based on biased random walk is designed. By directional replication item to high degree nodes, it is very easy to search these items. Analysis and simulations show that the proposed search algorithm can achieve high search success rate with low message overhead and low delay. DSR is very suitable for searching rare items, for example system state information, in unstructured P2P systems and improves the performance of P2P resource sharing.2. A general stochastic equation-based model for P2P file distribution is presented. Nodes are divided into categories according real systems and the corresponding quality models for nodes are also constructed. Combining with the dynamics of nodes, a stochastic equation-based model is put forth. New model abstract neighbor selection strategies and node arrival rate. Therefore, it can support different neighbor selection algorithms and node arrival rate. Based on the proposed model, different parameter settings are analyzed. Furthermore, a real system trace dataset is employed to evaluate the proposed model. The results show that the proposed model can describe the dynamics of P2P file distribution systems accurately. The proposed model is a general and practical model which can be used to evaluate existing paradigms and design new ones.3. GnutellaBT is proposed to address the architecture problem of BitTorrent. P2P file sharing is one of the most important Internet applications. Among them, BitTorrent is the most efficient P2P collaborative protocol for file distribution and the de facto standard of this kind of systems. However, the architecture of current BitTorrent has some shortcomings which decrease its reliability. It requires WWW to publish Torrent files and a dedicated Tracker node for neighbor allocation. GnutellaBT mergs Gnutella and BitTorrent to remove above shortcomings. By exploiting the search functionality of a Gnutella network, Torrent files can be located efficiently without WWW. To allocation neighbors randomly, GnutellaBT employs random walk on a specific Torrent network. Two-tier architecture is designed to implement above functions. Simulations show that GnutellaBT can work as well as original BitTorrent without WWW and Tracker nodes.4. An ISP-aware BitTorrent traffic optimization scheme is proposed, which is called STracker. The STracker is composed of tracker proxies in different ISPs, which are connected with P2P architecture. Tracker proxies are collaborative to perform node information management and ISP-aware neighbor allocation. The node management and neighbor allocation algorithms are also designed and implemented. Analysis and simulations show that STracker reduces the cross-ISP traffic massively without increasing the download time. STracker can efficiently improve the efficiency of Internet and decrease the operation cost of ISPs. The results of this work solved the key issues in large scale P2P resource sharing. DSR can efficiently improve the search performance for rare items in unstructured P2P networks. The general model for P2P conten distribution is a more reasonable approach for algorithm design and evaluation. GnutellaBT can improve the reliability of P2P content distribution systems. STracker can massively reduce the cross-ISP traffic and improve the efficience of Internet.