Binary instrumentation support for measuring performance in OpenMP programs

Abstract

In parallel computations, evaluating the causes of poor speedup is an important development activity to reach the goal of creating the most efficient parallel computation possible. In our research on irregular parallel computations, especially graph algorithms, we had specific measurement needs for which a dearth of tools could be found. We created PGOMP, a small library-based profiling tool for the Gnu OpenMP implementation, and show its use here in discovering some of the causes of poor speedup in graph computations.