Peer-to-peer systems provide the opportunity to pool large amounts of distributed resources to enable internetscale applications. However, the participant nodes are highly dynamic and unreliable. Thus, any shared resource such as file objects must incorporate redundancy to be useful. While many studies have proposed heuristics to determine redundancy levels based on object popularity, there has been little work in determining optimal or near-optimal resource allocation based on node reliability. In this paper, we present a strategy for the allocation of objects in the presence of dynamic and unreliable peers. We have built an availability model of peer-to-peer storage systems based on the bimodal and time-dependent availability characteristics of a P2P node. Using this model, we can select the size of a candidate node set for storage allocation and assign storage objects to maximize availability while still maintaining a balanced distribution of objects.
John A. Chandy