Fast Distributed Algorithm for Replication in Unstructured P2P Networks

Replicating data in distributed systems is an important issue, which is often needed for reliability, availability and performance improvement. In unstructured Peer-to-Peer networks, in which random walks are utilized for query routing, replicating data items can improve the probability of successfully finding requested items. In such networks, nodes often donate resources, in particular a part of their storage, to improve the overall performance of the network. Taking into consideration limited storage capacity that each node possesses, it is significant to replicate items with highest ”worthiness”. The ”worthiness” of data items measures the popularity of each data item for each peer. Considering peer specific request rate of each data item, network topology, allocation of items to peers, we investigate the problem of computing the worthiness of each data item. In this thesis, we present a fast distributed algorithm for computing the worthiness of data items. We first propose an algorithm for this problem in a centralized setting (i.e. with complete knowledge of the network) and then turn the centralized algorithm into a fast distributed one exploiting the local information of each individual peer. Simulation results also verify better performance of our algorithm compared to the algorithm proposed in [SNW08]. As a result, applying our fast distributed algorithm, a performance improvement of the distributed replication algorithm (P2R2) presented in [SNW08] is achieved.