A Lightweight Framework for Executing Task Parallelism on Top of MPI

摘要

This paper presents a directive-based programming and runtime environment that provides a lightweight framework for executing task parallelism on top of MPI. It facilitates the development of message-passing applications that follow the master-slave programming paradigm, supports multiple levels of parallelism and provides transparent load balancing with a combination of static and dynamic scheduling of tasks. A source-to-source translator converts C and Fortran master-slave programs, which express their task (RPC-like) parallelism with a set of OpenMP-like directives, to equivalents programs with calls to a runtime library. The result is a unified programming approach that enables the efficient execution of the same code on both shared and distributed memory multiprocessors. Experimental results on a Linux-cluster demonstrate the successful combination of ease of programming with the performance of MPI.