mpicroscope: Towards an MPI Benchmark Tool for Performance Guideline Verification

Abstract

We discuss a new, stand-alone MPI benchmark program (for now called mpicroscope) for assessing the performance of collective communication patterns, in particular of the standard MPI 2.2 collective operations. The benchmark is intended to facilitate comparisons between different ways of expressing the same functionality towards performing automatic detection of violations of self-consistent MPI performance guidelines. The benchmark can be used by MPI library developers for assessing the relative quality of new algorithms and implementations. It can also be used to make users aware of aspects of their MPI library where performance (portability) problems may lurk. The current version of the benchmark automatically detects, for any measured communication pattern, two universal, self-consistent guidelines that encourage monotone and split-robust performance. The benchmark aims to employ sound benchmarking procedures and is controlled via command-line options. It covers the MPI 2.2 collective operations, and a number of alternative patterns that express MPI collective functionality. In contrast to many other benchmarks data can be structured as described by an MPI derived datatype. We present results from a small InfiniBand cluster with a vendor MPI library, showing performance guideline violations that were detected and highlighted with mpicroscope.