A Performance Analysis of GPU-Aware MPI Implementations Over the Slingshot-11 Interconnect

This study evaluates the performance of three GPU-aware MPI implementations-Cray MPICH, MPICH, and Open MPI-on the Slingshot-11 interconnect using the Perlmutter supercomputer. We use the OSU microbenchmarks to evaluate the bandwidth and latency of specific MPI routines, and two Exascale Computing Projects - LAMMPS and WarpX to evaluate application performance. Results confirm that Cray MPICH outperforms the others over Slingshot-11, significantly influencing HPC application efficiency by leveraging proprietary hardware mechanisms. This research assists domain scientists in selecting MPI libraries, enhancing application performance on Slingshot-11 systems and contributing to future software and hardware optimization studies.