Exploring Composable Parallelism in Computational Modelling

The increase in multiprocessor cores offers significant potential for enhancing parallel application performance, but programmers face challenges such as increased work distribution overhead and reduced workload per thread, which can degrade performance. One solution is to leverage composable parallelism within applications, though this may require additional support in parallel programming models to address oversubscription issues. This study evaluates the performance of different programming models—OpenMP, BOLT, oneTBB, and Argobots—when applied to exploit composable parallelism in the multizone versions of NAS Parallel Benchmarks, with experiments conducted on an Arm-based system.