high performance computing on graphics processing units: hgpu.org

Data race detection is an important problem in GPU programming. The paper presents a novel solution. It uses the compiler support to privatize shared data and then at run time parallelizes the race checking. It has two distinct features. First, there is no per access monitoring, so the race detection has a low overhead and does not affect the scalability. Second, race checking utilizes the massively parallel hardware of the GPU. Preliminary results show two orders of magnitude performance improvement over a previous method.