Workshops Archive

MPI-CCOB: Benchmarking MPI Overlap on CPU and GPU Systems


Workshop: ExaMPI26: Workshop on Extreme-Scale MPI

Authors: Ehab Saleh (Leibniz Supercomputing Centre), Josef Weidendorfer (Technical University of Dresden), and Martin Schulz (Technical University of Munich)

Abstract: In this paper, we introduce the MPI Communication–Computation Overlap Benchmark, or MPI-CCOB, a configurable benchmark for evaluating overlap in MPI applications on CPU and GPU systems. Motivated by domain-decomposition applications, the benchmark organizes processing units into one-, two-, or three-dimensional Cartesian topologies, producing different numbers of concurrent neighbor-exchange requests. It supports both compute-bound and memory-bound workloads to examine how computational characteristics affect overlap. Since not all MPI implementations support asynchronous progress, MPI-CCOB also provides an optional user-level progress thread per processing unit. By controlling workload characteristics and communication-progress mechanisms, our benchmark reveals the conditions required for effective overlap. Experiments with MPICH and Open MPI show that active progress enables nearly complete overlap during compute-bound workloads. However, overlap decreases substantially under memory-bound workloads on both CPU and GPU systems, even when communication is continuously given the option to progress.


Back to ExaMPI26: Workshop on Extreme-Scale MPI Archive Listing

Back to Full Workshop Archive Listing