Automatic Parameter Tuning of Hierarchical Incremental Checkpointing

As future HPC systems become larger, the failure rates and the cost of checkpointing to the global file system are expected to increase. To solve this problem, this paper proposes a hierarchical incremental CPR mechanism that utilizes a hierarchical storage system of local storages and global storages. Then, to adjust the parameters of the proposed mechanism, a runtime autotuning technique is presented. Evaluation results show that the timing overhead can be significantly reduced if the storage hierarchy can be exploited with appropriate checkpoint intervals.