Cross-thread critical sections and efficient dynamic race prediction methods

TL;DR

This paper introduces the concept of cross-thread critical sections to improve dynamic data race prediction, demonstrating their practical occurrence and impact on analysis accuracy and performance.

Contribution

It presents a theoretical framework for cross-thread critical sections and evaluates their effect on existing race prediction methods, enhancing accuracy with manageable overhead.

Findings

Cross-thread critical sections occur in real-world programs.

Inclusion of cross-thread sections reduces false positives in race detection.

Tracking these sections incurs 10-20% performance overhead.

Abstract

The lock set method and the partial order method are two main approaches to guarantee that dynamic data race prediction remains efficient. There are many variations of these ideas. Common to all of them is the assumption that the events in a critical section belong to the same thread. We have evidence that critical sections in the wild do extend across thread boundaries even if the surrounding acquire and release events occur in the same thread. We introduce the novel concept of a cross-thread critical section to capture such situations, offer a theoretical comprehensive framework, and study their impact on state-of-the-art data race analyses. For the sound partial order relation WCP we can show that the soundness claim also applies to cross-thread critical sections. For DCtp the occurrence of cross-thread critical sections invalidates the soundness claim. For complete partial order relations such as WDP and PWR, cross-thread critical sections help to eliminate more false positives. The same (positive) impact applies to the lock set construction. Our experimental evaluation confirms that cross-thread critical sections arise in practice. For the complete relation PWR, we are able to reduce the number of false positives. The performance overhead incurred by tracking cross-thread critical sections slows down the analysis by 10%-20%, on average.