An OpenMP runtime profiler/configuration tool for dynamic optimization of the number of threads

摘要

This paper describes the implementation and the experimental results of a tool for dynamic configuration of the number of threads used in the OpenMP environment based on the current state of the runtime at the time of the call. For this purpose, we use a mix of profiling and machine learning techniques to determine the number of threads. The decision to set the number of threads is made at the time of the call. The proposed approach is designed to be cost effective in the scenario of a highly dynamic runtime primarily when running relatively long tasks consisting of a number of parallel constructs.