A Prototype Implementation of OpenMP Task Dependency Support

OpenMP 3.0 introduced the concept of asynchronous tasks, independent units of work that may be dynamically created and scheduled. Task synchronization is accomplished via the insertion of taskwait and barrier constructs. However, the inappropriate use of these constructs may incur significant overhead owing to global synchronizations for specific algorithms on large platforms. The performance of such algorithms may benefit substantially if a mechanism of specifying finer gained point-to-point synchronization between tasks is available. In this paper we present extensions to the current OpenMP task directive to enable the specification of dependencies among tasks. A task waits only until the explicit dependencies as specified by the programmer are satisfied, thereby enabling support for a dataflow model within OpenMP. We evaluate the extensions implemented in the OpenUH OpenMP compiler using LU decomposition and Smith-Waterman algorithms. By applying the extensions to the two algorithms, we demonstrate significant performance improvement over the standard tasking versions. When comparing our results with those obtained using related dataflow models OmpSs and QUARK, we observed that the versions using our task extensions delivered an average speedup of 2-6x.