The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Workshop: 7th International Workshop on Software Correctness for HPC Applications (Correctness '23)
Authors: Simon Schwitanski, Joachim Jenke, Sven Klotz, and Matthias S. Müller (Chair for High-Performance Computing, IT Center, RWTH Aachen University)
Abstract: Parallel programming models with Remote Memory Access (RMA), such as MPI RMA, OpenSHMEM, and GASPI, allow processes to modify the memory of other processes directly. Special care is needed to avoid concurrent conflicting accesses that lead to data races across processes with undefined behavior. Although only some tools exist that can detect RMA races, there needs to be a possibility to compare their effectiveness systematically. We present RMARaceBench, a microbenchmark suite, to evaluate the detection capabilities of current and future RMA race detection tools for MPI RMA, OpenSHMEM, and GASPI. It consists of about 100 synthetic race test cases for each programming model, aiming to cover all possible race scenarios. Using RMARaceBench, we evaluate two MPI RMA race detectors implemented in the correctness tools PARCOACH and MUST. The evaluation shows that RMARaceBench can pinpoint the strengths and weaknesses of RMA race detectors.