Optimized Sound And Complete Data Race Detection In Structured Parallel Programs

Task parallel programs that are free of data race are guaranteed to be deterministic, serializable, and free of deadlock. Techniques for verification of data race freedom vary in both accuracy and asymptotic complexity. One work is particularly well suited to task parallel programs with isolation and lightweight threads. It uses the Java Pathfinder model checker to reason about different schedules and proves the presence or absence of data race in a program on a fixed input. However, it uses a direct and inefficient transitive closure on the happens-before relation to reason about data race. This paper presents Zipper, an alternative to this naive algorithm, which identifies the presence or absence of data race in asymptotically superior time. Zipper is optimized for lightweight threads and, in the presence of many threads, has superior time complexity to leading vector clock algorithms. This paper includes an empirical study of Zipper and a comparison against the naive computation graph algorithm, demonstrating the superior performance it achieves.