Performance analysis of built-in parallel reduction's implementation in OpenMP C/C++ language extension

Abstract

Parallel reduction algorithms are frequent in high performance computing areas, thus, modern parallel programming toolkits and languages often offer support for these algorithms. This article discusses important implementation aspects of built-in support for parallel reduction found in well-known OpenMP C/C++ language extension. It shows that the implementation in widely used GCC compiler is not efficient and suggests usage of custom reduction implementation improving the computational performance.