Evaluating performance of openmp tasks in a seismic stencil application

Abstract

Simulations based on stencil computations (widely used in geosciences) have been dominated by the MPI+OpenMP programming model paradigm. Little effort has been devoted to experimenting with task-based parallelism in this context. We address this by introducing OpenMP task parallelism into the kernel of an industrial seismic modeling code, Minimod. We observe that even for these highly regular stencil computations, taskified kernels are competitive with traditional OpenMP-augmented loops, and in some experiments tasks even outperform loop parallelism. This promising result sets the stage for more complex computational patterns. Simulations involve more than just the stencil calculation: a collection of kernels is often needed to accomplish the scientific objective (e.g., I/O, boundary conditions). These kernels can often be computed simultaneously; however, implementing this simultaneous computation with traditional programming models is not trivial. The presented approach will be extended to cover simultaneous execution of several kernels, where we expect to fully exploit the benefits of task-based programming.