Reducing Correlated Failures Impact in Peer-to-Peer Storage Systems Using Mobile Agents Flocks

This paper introduces MinCor (Minimum of Correlations), a decentralized simulated annealing algorithm designed for the data placement in peer-to-peer networks. Its goal is to reduce the correlated failures impact in such data storage systems. This data placement is realized using a multi-agent system which turns the documents into mobile agents flocks. After a network clustering step where highly correlated peers are regrouped together, the flocks executing MinCor are able to find a placement minimizing the number of agents on the same clusters. This placement is obtained in a decentralized way thanks to the environment exploration capabilities of the flocks. A set of experiments are performed on this system in presence of correlated failures. They show that, in practice, the expected placement is well obtained. They also show that, flocks using the MinCor algorithm suffer less simultaneous losses in presence of correlated failures than a mere random placement.