A Proposal to Extend the OpenMP Tasking Model with Dependent Tasks

Abstract

Tasking in OpenMP 3.0 has been conceived to handle the dynamic generation of unstructured parallelism. New directives have been added allowing the user to identify units of independent work (tasks) and to define points to wait for the completion of tasks (task barriers). In this document we propose extensions to allow the runtime detection of dependencies between generated tasks, broading the range of applications that can benefit from tasking or improving the performance when load balancing or locality are critical issues for performance. The proposed extensions are evaluated on a SGI Altix multiprocessor architecture using a couple of small applications and a prototype runtime system implementation.

The TERAFLUX Project: Exploiting the DataFlow Paradigm in Next Generation Teradevices

“…At runtime, this information about the directionality of the task data is used to build a task data-dependence graph that exhibits the inherent data dependences of the application as well as its potential task parallelism. Current efforts at the Barcelona Supercomputing Center are focused on the OmpSs [9][46] implementation of StarSs which extends the OpenMP explicit tasks [10] with dependence clauses that indicate the directionality of the tasks arguments, and on the Task Superscalar design [59] [60]. Fig.…”

Preliminary Experiments with XKaapi on Intel Xeon Phi Coprocessor

“…The XKaapi code illustrates the C++ interface that does not require a specific compiler: in the notation A(ri,rj), ri and rj denote a range of indexes so that A(ri,rj) is a sub-matrix of matrix A. Thanks to this finer knowledge of tasks dependencies, the runtime system can schedule ready tasks between two main iterations in k [15], [16]. The OpenMP version is similar to the CilkPlus version by replacing cilk_spawn by #pragma omp task and cilk_sync by #pragma omp taskwait.…”

“…In particular, several studies [15], [16] show that the strong synchronizations imposed by both OpenMP and Cilk execution models artificially limit the available parallelism: in these programming models, the synchronization construct block the running thread until previously spawned tasks have completed their execution. For instance, in a classical nested loops formulation of matrix factorization [15], [16], both OpenMP and Cilk enforce periodic synchronizations that block the execution of tasks from the next iteration. These studies emphasize data-flow approaches that are able to expose fine-grain one-to-one synchronizations between tasks: the runtime system can detect concurrent tasks as soon as their inputs are produced.…”

A Comparison of Task Parallel Frameworks based on Implicit Dependencies in Multi-core Environments

“…Since the inclusion of dependencies for tasks in OpenMP 4.0 [26] this standard is the most widespread option that supports this paradigm, and thus the first we study. A very active and related project, which in fact pushed for the adoption of dependent tasks in OpenMP [11], is OmpSs, which integrates features from the StarSS family of programming models [27]. While OpenMP and OmpSs cover well the space of compiler directives in our scope of interest, the area of libraries is much more sparse.…”