Improving Performance of CAPE Using Discontinuous Incremental Checkpointing

Originally, OpenMP was designed to develop parallel applications on shared-memory architectures. One of the advantages that made the success of OpenMP is the simplicity of the associated programming model. Check pointing Aided Parallel Execution (CAPE) is a paradigm that uses check pointing techniques to run parallel programs on distributed-memory architectures. In order to show its effectiveness, it has been used to develop a compiler to run OpenMP programs on distributed-memory architectures. The first prototype we developed proved the feasibility of the paradigm but the use of complete checkpoints led to poor performance. This was mainly due to the large amount of data to transfer and process. This paper presents the new prototype we developed for CAPE based on the discontinuous incremental check pointing technique and an analysis its performance.