Comparing Checkpoint and Rollback Recovery Schemes in a Cluster System

摘要

Cluster systems play a central role to realize high performance computing with relatively low cost, and at the same time are necessary the fault-tolerance features for the practical use. In this paper we develop stochastic models to evaluate the expected total recovery overhead for a cluster computing system with three well-known checkpoint and rollback recovery schemes; checkpoint mirroring, central file server checkpointing and skewed checkpointing, where the fault latency time after a system failure is given by a random variable. In general, since the multi-node failure as well as single-node failure may occur in the cluster system, it is not so easy to obtain the closed form of expected total recovery overhead. Based on a simple failure model, we do this by listing up all the possible combinations of probabilistic events caused by the multi-node failure. Further we compare the respective expected total recovery overhead with different checkpoint and rollback recovery schemes, and evaluate quantitatively the effectiveness of these schemes.