A Quantitative Analysis of OpenMP Task Runtime Systems

Abstract

Although OpenMP is heavily used to parallelize for-loops, it also supports task-parallel programming, which is important for parallelizing irregular applications. In this work, we focus on the performance of OpenMP runtime systems for task-based applications. In particular, we investigate the performance of different OpenMP runtime systems when scheduling a large set independent tasks of different granularity. To that end, we propose a new OpenMP benchmark, which features profiling and tracing options that help developers to reason about the observed performance differences. We compare the execution times measured for a variety of compilers, such as gcc, icc, clang, aocc, and pgcc, for both homogeneous and heterogeneous workloads. Our study shows that there are significant performance differences between the different OpenMP implementations. We also show that the performance attainable with a compiler strongly depends on the machine architecture, the number of threads, the thread-pinning strategy, and the task granularity.