Exploring Fault-tolerance and Reliability in a Peer-to-peer Cycle-sharing Infrastructure

Abstract. The partitioning of a long running task into smaller tasks that are executed separately in several machines can speed up the exe-cution of a computationally expensive task. This has been explored in Clusters, in Grids and lately in Peer-to-peer systems. However, transpos-ing these ideas from controlled environments (e.g., Clusters and Grids) to public environments (e.g., Peer-to-peer) raises some reliability chal-lenges: will a peer ever return the result of the task that was submitted to it or will it crash? and even if a result is returned, will it be the ac-curate result of the task or just some random bytes? These challenges demand the introduction of result verification and checkpoint/restart mechanisms to improve the reliability of high performance computing systems in public environments. In this paper we propose and analyse a twofold approach: i) two checkpoint/restart mechanisms to mitigate the volatile nature of the participants; and ii) various flavours of replication schemes for reliable result verification