A methodology for automatic generation of executable communication specifications from parallel MPI applications

Abstract

Portable parallel benchmarks are widely used for performance evaluation of HPC systems. However, because these are manually produced, they generally represent a greatly simplified view of application behavior, missing the subtle but important-to-performance nuances that may exist in a complete application. This work contributes novel methods to automatically generate highly portable and customizable communication benchmarks from HPC applications. We utilize ScalaTrace, a lossless yet scalable parallel-application tracing framework to collect selected aspects of the run-time behavior of HPC applications, including communication operations and computation time, while abstracting away the details of the computation proper. We subsequently generate benchmarks with nearly identical run-time behavior to the original applications. Results demonstrate that the generated benchmarks are in fact able to preserve the run-time behavior (including both the communication pattern and the execution time) of the original applications. Such automated benchmark generation is without precedent and particularly valuable for proprietary, export-controlled, or classified application codes.