Parallel processing: the power and the portability: experiments with “reusable toolkits”

Abstract Serial architectures and conventional languages (especially Fortran) have been the basis for almost all current industrial applications of high performance computing. Whilst the future of computing is certain to be ‘increasingly parallel’, and some successful applications of today's parallel processing technology have already been demonstrated, adoption by the main industrial users of conventional supercomputing has been limited. One reason for this is the perceived risk of investing in a rapidly developing (or, alternatively, immature) technology where specific systems (hardware or software) may become obsolete in a short period of time. It is thus reasonable to develop programming techniques, or software, which exploit available parallel power, but limit dependence on the details of specific parallel architectures. Such methods should also provide a measure of portability between machines from different manufacturers. This article discusses two examples (cellular automata and imaging) and endeavours to illustrate the ‘state of the art’ by giving an account of work done at, or in collaboration with, the Parallel Computing Centre at the University of Edinburgh.