A machine-independent approach to parallel programming

Abstract

A multitasking model for specifying parallelism in a machine-independent fashion is described. Parallel programs are decomposed into a set of concurrent tasks which work on local or shared data. The computation of a task is described as the computation of atomic actions or activities. Software tools translate multitasking specifications into a parallel representation of the program. A master-slaves execution model is used to provide machine independence, reliability, and load-balanced execution on a variety of parallel architectures. This model has been implemented on a heterogeneous network of workstations, a shared-memory multiprocessor, and a supercomputer. An example of the use of this methodology for a multiple-Doppler radar meteorology program is presented.<>