Hierarchical composition of coordinated checkpoint with pessimistic message logging

Grid computing mutualizes more computing resources working in a calculation or a common task. The increase in the number of components in the system leads also increases the number of fault. These failures result in a loss of several cycles of running applications. It is therefore essential to be able to tolerate faults so that the computation can continue to execute and finish despite failures, all while maintaining maximum performance. One advantage of coordinated checkpoint is its capacity to have a very low overhead as long as the execution stays fault free. On the contrary, due to the fact that uncoordinated checkpoint requires being complemented by a message log protocol, this adds a significant penalty for all message transfers, even in case of fault-free execution. With message log, these problems do not arise simply because it processes checkpoint and restart autonomously. These differences suggest that the best approach depend on the fault frequency. In this paper, we propose a hierarchical composition of algorithms: Chandy-Lamport protocol and pessimistic message logging protocol. We have implemented and compared the performance of these protocols in grid computing using the Omnet++ simulator [3].