Evaluating Sparse Data Storage Techniques for MPI Groups and Communicators

This paper explores various sparse data storage techniques in order to reduce the amount of memory required for MPI groups and communicators. The idea behind the approach is to exploit similarities between the objects and thus store only the difference between the original process group and the resulting one. For each technique, it details the memory saved compared to the currently used implementations, and present a runtime decision routine capable of choosing dynamically the most efficient technique for each scenario. Furthermore, the paper evaluates the performance impact of the new structures using point-to-point benchmarks as well as an application scenario over InfiniBand, Myrinet and Gigabit Ethernet networks.