Section: New Results

Running Compliant OpenMP Applications on top of StarPU with the Klang-Omp Compiler

Several robust runtime systems proposed recently have shown the benefits of task-based parallelism models. However, the common weakness of these supports is to tie applications to specific APIs. The OpenMP specification aims at providing a common parallel programming means for shared-memory platforms. It appears a good candidate to address this issue. We assessed the effectiveness and limits of this approach on the ScalFMM library developed by Inria HiePACS team, implementing fast multipole methods (FMM) algorithms. We showed that OpenMP dependent tasks allow for significant performance improvements over OpenMP loops and independent tasks for this application. We also demonstrated that suitable, targeted language extensions can further improve performances by a important margin in some cases.