libKOMP, an Efficient OpenMP Runtime System for Both Fork-Join and Data Flow Paradigms

International audienceTo efficiently exploit high performance computing platforms, applications currently have to express more and more finer-grain parallelism. The OpenMP standard allows programmers to do so since version 3.0 and the introduction of task parallelism. Even if this evolution stands as a necessary step towards scalability over shared memory machines holding hundreds of cores, the current specification of OpenMP lacks ways of expressing dependencies between tasks, forcing programmers to make unnecessary use of synchronization degrading overall performance. This paper introduces libKOMP, an OpenMP runtime system based on the X-Kaapi library that outperforms popular OpenMP implementations on current task-based OpenMP benchmarks, but also provides OpenMP programmers with new ways of expressing data-flow parallelism