Abstract FR:

We study a new approach for decentralized data storage in peer-to-peer networks. In this approach, the responsibility of data management is transferred from the peers to the documents. It means that documents are not passive data sets anymore but become autonomous and responsible for their own durability. Thanks to a multi-agent system modeling and bio-inspired algorithms, we transform each document into a mobile agents flock able to move into the network. Firstly, we assess the feasibility of this approach with several experiments done on a prototype deployed in a real peer-to-peer network. We note that, given some hypothesis, our motion algorithms are sound. We also note that, topological relationships between the agents are enough for the emergence of a global flocking behavior. Secondly, we focus on mechanisms required to ensure flocks durability. We note that those flocks are self-adaptive and that, this property can be used to find the accurate fragmentation parameters, given a network instance and a required level of availability. Finaly, we study this self-adaptation property in the context of correlated failures. We propose and we analyze a decentralized flock placement algorithm aimed at reducing the correlated failures impact on data storage systems.