University of Houston

OpenMP 3.0 introduced the concept of tasks with the aim of handling unstructured parallelism. The performance of certain algorithms stand to benefit substantially with the runtime detection and resolution of data dependencies among tasks. Such an approach embodies a dataflow model, exploiting parallelism, with the resolution of explicit data dependence relationships among tasks at runtime. We propose language extensions to the current OpenMP task directive, aiming to provide task-level granularity for synchronization of tasks, sharing the same parent, within the OpenUH OpenMP RTL. The application of the extensions on two algorithms, LU Decomposition and Smith-Waterman, demonsrated significant performance improvement over the standard tasking versions of the two algorithms using the GNU, Intel, PGI, Oracle/Sun and the Mercurium compilers. We compared our implementation with those of similar dataflow models - OmpSs and QUARK runtime API, and observed an average speedup of 2 - 6X.