An open MPI extension for supporting task based parallelism in heterogeneous CPU-GPU clusters

Abstract

In this work we identify and analyze some of the patterns appearing in the development and deployment of scientific applications over clusters equipped with heterogeneous computing resources. The main contributions of this work are the identification of the patterns aforementioned, as well as the design and implementation of an Open MPI extension that supports the development and deployment of applications programmed using a task approach. In order to illustrate how to use our extension, we provide the implementation and performance evaluation of two sample applications: the N-Body problem and the general matrix multiplication.