Doctor of Philosophy

dissertationHigh Performance Computing (HPC) on-node parallelism is of extreme importance to guarantee and maintain scalability across large clusters of hundreds of thousands of multicore nodes. HPC programming is dominated by the hybrid model "MPI + X", with MPI to exploit the parallelism across the nodes, and "X" as some shared memory parallel programming model to accomplish multicore parallelism across CPUs or GPUs. OpenMP has become the "X" standard de-facto in HPC to exploit the multicore architectures of modern CPUs. Data races are one of the most common and insidious of concurrent errors in shared memory programming models and OpenMP programs are not immune to them. The OpenMP-provided ease of use to parallelizing programs can often make it error-prone to data races which become hard to find in large applications with thousands lines of code. Unfortunately, prior tools are unable to impact practice owing to their poor coverage or poor scalability. In this work, we develop several new approaches for low overhead data race detection. Our approaches aim to guarantee high precision and accuracy of race checking while maintaining a low runtime and memory overhead. We present two race checkers for C/C++ OpenMP programs that target two different classes of programs. The first, ARCHER, is fast but requires large amount of memory, so it ideally targets applications that require only a small portion of the available on-node memory. On the other hand, SWORD strikes a balance between fast zero memory overhead data collection followed by offline analysis that can take a long time, but it often report most races quickly. Given that race checking was impossible for large OpenMP applications, our contributions are the best available advances in what is known to be a difficult NP-complete problem. We performed an extensive evaluation of the tools on existing OpenMP programs and HPC benchmarks. Results show that both tools guarantee to identify all the races of a program in a given run without reporting any false alarms. The tools are user-friendly, hence serve as an important instrument for the daily work of programmers to help them identify data races early during development and production testing. Furthermore, our demonstrated success on real-world applications puts these tools on the top list of debugging tools for scientists at large