Co-ordinating Combined Parallel Vector and Scalar Computation

A new parallel programming approach based on the structured co-ordination of parallel behaviour is applied to programming the Fujitsu AP1000 for mixed parallel vector and scalar computation. With such a structured parallel programming model resource organisation as well as data and computation distribution can be programmed at a high level. The performance of a program can also be systematically predicted. In this paper various parallel implementations of the iterative conjugate gradient algorithm have been investigated. The experiments indicate that the structured co-ordination method of programming allows sufficient control over the allocation of resources to make efficient use of heterogeneous machines. Furthermore, these experiments suggest that, even in an experimental machine such as the AP1000, the most efficient program comes from the co-ordinated use of both vector and scalar processors. This indicates that a well balanced heterogeneous architecture would have great potential and that the methods we have developed would be an appropriate means to program such machines.