Dynamic Clustering-Based Query Answering in Peer-to-Peer Systems

P2P computing has been employing in more and more application domains as the technology becomes mature. One popular and successful application area is file sharing. However, current file sharing systems support only or mainly keybased exact matching (e.g., Chord [27], CAN [25]) and keyword-based searching (e.g., Napster, Gnutella) for files discovery and location, which is not enough to meet the requirements of more advanced applications such as information retrieval and data management. In this paper, we propose a new query answering model for P2P applications, which is termed as clustering-based query answering (CBQA). In our definition, CBQA will retrieve the data objects that are in the same cluster of the query from the global dataset distributed over peers of a P2P system. Generally, CBQA may obtain more correct answers than similarity based query can, which means higher recall may be achieved. To implement the new query model, we first present a framework that support clustering based query answering, including general algorithms, lemmas and system architecture. Then we give three concrete algorithms for different clustering criteria, namely k-nearest-neighbor, distancebased, and density-based clustering, along with detailed analyses and discussions. Finally, implementation issues, especially dynamic neighbors selection and caching techniques to enable the scalability of our method are addressed. Theoretical analysis and preliminary experiments show that our method can guarantee to find desirable objects in the interested cluster with modest bandwidth overhead.