Modeling fetch-at-most-once behavior in peer-to-peer file-sharing systems

Abstract

Recent measurement studies show that the object popularity distribution in Kazaa file sharing systems deviates significantly from the Zipf distribution which is commonly seen for the World Wide Web. We measure a real BitTorrent network and we figure its object popularity distribution, which also shows, on a log-log scale, a non-Zipf curve with flattened head. The fetch-at-most-once behavior of peer-to-peer (P2P) client is responsible for such a non-Zipf distribution and we propose two mathematical models to describe this. The models are based on different probability assumptions, though both indicate flatter heads in object popularity distribution curves than Zipf would predict. Our models provide theoretic tools to analyze differences between P2P file-sharing system and Web systems. © Springer-Verlag Berlin Heidelberg 2006.