On-the-fly data race detection with the enhanced openmp series-parallel graph

Abstract

This paper proposes OMP-RACER, a dynamic apparent data race detector for OpenMP programs. Apparent data races are those races that manifest in a program considering the logical series-parallel relations of the execution. By identifying apparent races, OMP-RACER can detect races that occur not only in the observed schedule but also in other schedules for a given input. Our key contribution is a data structure to capture series-parallel relations between various fragments of an OpenMP program with both structured and unstructured parallelism directives, which we call the Enhanced OpenMP Series-Parallel Graph (EOSPG). OMP-RACER maintains information about previous accesses with each memory access and uses the EOSPG to check if they can logically execute in parallel. OMP-RACER detects more races with similar overheads when compared to existing state-of-the-art race detectors for OpenMP programs.