Data race detection based on execution replay for parallel applications

Abstract

We describe an integrated approach to support debugging of nondeter-ministic concurrent programs. Our tool provides reproducible program behavior and incorporates mechanisms to identify synchronization bugs commonly termed data races or access anomalies. Both features are based on partially ordered event logs captured at run time. Our mechanism identifies a race condition that is guaranteed to be unaffected by other races in the considered execution. Data collection and analysis for race detection has no impact on the original computation since it is done in replay mode. The race detection and execution replay mechanisms are integrated in the MOSKTTO operating system.