Peer-to-peer replication with preferences

Abstract

A P2P system can be viewed as a system that provides replication services. Unlike conventional structured replication systems (CDN, RAID), peers in an unstructured P2P system may have heterogeneous, sometimes low, online availability. Therefore, we formulate the problem with the objective to achieve good system level file availability, and study distributed algorithms for autonomous peers to accomplish that. In this paper, we emphasize the need to provide a differentiated replication service, since files are accessed with different frequency and have different importance. We quantify file preference in terms of weight and formulate the objective as to maximize a weighted sum of file availability. A bi-weight model is studied and then applied to a decentralized random replication algorithm through a statistical rounding policy. This algorithm is easily implementable by autonomous peers with partial information about the resources of the system, and yet yields favorable results in delivering the differentiated replication service while maintaining the system level replication goal.