Using locality of reference to improve performance of peer-to-peer applications

Abstract

Peer-to-peer, or simply P2P, systems have recently emerged as a popular paradigm for building distributed applications. One key aspect of the P2P system design is the mechanism used for content location. A number of different approaches are currently in use. In particular, the location algorithm used in Gnutella, a popular and extensively analyzed P2P file sharing application, is based on flooding of messages in the network, which results in significant processing overhead on the participant nodes and thus, poor performance.In this paper, we provide an extensive performance evaluation of alternative algorithms for content location and retrieval in P2P systems, in particular, the Freenet and Gnutella systems. We compare the original Freenet and Gnutella algorithms, a previously proposed interest-based algorithm and two new algorithms which also explore locality of interest among peers to efficiently allow content location. Unlike previous proposals, the new algorithms organize the peers into communities that share interests. Two peers are said to have common interest if they share some of the locally stored files.In order to evaluate the performance of these algorithms, we use a previously developed Freenet simulator and build a new Gnutella simulator, which includes several realistic system characteristics. We show that the new community-based algorithms improve the original Gnutella content location latency (and thus the system QoS) and system load by up to 31% and 30%, respectively. Our algorithms also reduce the average Freenet request and response path lengths by up to 39% and 31%, respectively. Furthermore, we show that, compared to the previously proposed interest-based algorithm, our new algorithms improve query latency by up to 27% without a significant increase in the load.