An Automatic Parallelization Tool for Sequential Programs

Abstract : In this project we have been trying to improve a system called ASPAR- the automatic, symbolic parallelization system. This is a software tool designed to automate the process of converting sequential C and Fortran programs for execution on parallel computers. Our strategy, as developed in our own internal research is to modify sequential programs by the addition of calls to the Express runtime library. This is a set of library utilities specifically designed to allow parallel programs to operate in a manner independent of the underlying hardware. A set of abstractions is provided in which the physical machine topology can be hidden by a programming interface in which the machine interconnections can be assumed to match the topology of the data being manipulated. This ability is of central importance to our automated techniques: multi-dimensional arrays can be mapped logically onto a similarly multi- dimensional parallel processing network regardless of that machine physical interconnectivity. A second important advantage of Express is its portability. Express programs can be executed on a wide range of different parallel processing systems including nCUBE and INTEL hypercubes, multi-headed Crays, transputer arrays and networks of UNIX workstations. This allows us to test our methods on many different parallel architectures rather than concentrating on an single type of system. As a result our tools will be of wide relevance in the parallel processing community. The preliminary version of ASPAR was able to generate parallel C programs from a standard ANSIC program contained in a single source file.