Automatic generation of parallel unit tests

Multithreaded software is subject to data races. Currently available data race detectors report such errors to the developer, but consume large amounts of time and memory; many approaches are not applicable for large software projects. Unit tests containing fractions of the program lead to better results. We propose AutoRT, an approach to automatically generate parallel unit tests as target for data race detectors from existing programs. AutoRT uses the Single Static Multiple Dynamic (SSMD) analysis pattern to reduce complexity and can therefore be used efficiently even in large software projects. We evaluate AutoRT using Microsoft CHESS and show that with SSMD all 110 data races contained in our sample programs can be located.