An Evaluation of Data Race Detectors Using Bug Repositories

Multithreaded software is subject to data races. A large number of data race detectors exists, but they are mainly evaluated in academic examples. In this paper we present a study in which we applied data race detectors to real applications. In particular, we want to show, if these tools can be used to locate data races effectively at an early stage in software development. We therefore tracked 25 data races in bug repositories back to their roots, created parallel unit tests and executed 4 different data race detectors on these tests. We show, that with a combination of all detectors 92% of the contained data races can be found, whereas the best data race detector only finds about 50%.