The Gnutella file sharing system allows a large number of peers to share their local files. However, it does not coordinate the way by which these shared objects are named or how they are searched by other users; such decisions are made independently by each peer. In this work, we investigate the practical performance implications of this design. We collected the shared filenames and user generated queries over a three-year period. We show the mismatch between these naming mechanisms. We show the fundamental limitations of Gnutella performance that cannot be addressed by improvements in overlays or by varying the search mechanisms alone. Based on our observations, we describe two practical approaches to improve Gnutella performance. We describe a mechanism to build the file term synopsis using the observed popularity of queries routed through the ultrapeer. We also describe a query transformation mechanism that improves the success rates for failed queries.
展开▼
