Eliminating Costs for Crossing Process Boundary from MPI Intra-node Communication
Akio Shimada, Atsushi Hori, Yutaka Ishikawa · 2014
In many-core environments, the performance and memory footprint of MPI intra-node communications are both important issues. In this paper, we contend that the address space boundaries between MPI processes are detrimental to efficient intra-node communication and should be removed as we advance into the many-core era. We confirmed this contention by using the PVAS task model, which can remove the address space boundaries between MPI processes. The benchmark results show that removing address space boundaries enhances the intra-node communication performance without increasing the memory footprint.