Empirical Comparison of Race Detection Tools for OpenMP Programs

Abstract

Data races or races which occur in parallel programs such as OpenMP programs must be detected, because they may lead to unpredictable results of program executions There are three representative tools which detect races which can occur in OpenMP programs Thread Analyzer of Sun Inc Thread Checker of Intel Corporation, and RaceStand of GNU Two of these tools, Thread Checker and RaceStand, are known with their power for race detection through empirical analysis But Thread Analyzer has not been analyzed empirically about its power for race detection and we cannot know exactly Its race verification capability as well as its efficiency This impel empirically analyzes the verification capability of Thread Analyzer using a set of synthetic programs winch has nested parallelism or inter-thread coordination, and compares three race detection tools in an aspect. of efficiency through OpenMP Micro-benchnmarks of EPCC that measures loop scheduling and array operations We predicted that Thread Analyzer verifies the existence of races in most OpenMP programs And we found that Thread Analyzer is faster than Thread Checker about 5 tunes and 2 times slower than RaceStand in arraybench of EPCC Also, it, shows similar tune variation with Thread Checker and is about 20 times slower than RaceStand in schedbench of EPCC