A deployable sampling strategy for data race detection

Dynamic data race detection incurs heavy runtime overheads. Recently, many sampling techniques have been proposed to detect data races. However, some sampling techniques (e.g., Pacer) are based on traditional happens-before relation and incur a large basic overhead. Others utilize hardware to reduce their sampling overhead (e.g., DataCollider) and they, however, detect a race only when the race really occurs by delaying program executions. In this paper, we study the limitations of existing techniques and propose a new data race definition, named as Clock Races, for low overhead sampling purpose. The innovation of clock races is that the detection of them does not rely on concrete locks and also avoids heavy basic overhead from tracking happens-before relation. We further propose CRSampler (Clock Race Sampler) to detect clock races via hardware based sampling without directly delaying program executions, to further reduce runtime overhead. We evaluated CRSampler on Dacapo benchmarks. The results show that CRSampler incurred less than 5% overhead on average at 1% sampling rate. Whereas, Pacer and DataCollider incurred larger than 25% and 96% overhead, respectively. Besides, at the same sampling rate, CRSampler detected significantly more data races than that by Pacer and DataCollider.