Submission for Sc97 a Predictably Portable Process Farm Skeleton Extended Abstract

Abstract

Communications libraries like MPI enable architecture independent parallel program development but cannot alone guarantee performance consistency across diierent architectures. Here an algorithmic skeleton for a process farm is discussed and an architecture independent performance model is developed for it. The skeleton has been ported from a Meiko Computing Surface to a Cray T3D and a Fujitsu AP1000, demonstrating consistent behaviour. Preliminary performance prediction accuracy on the Cray T3D is around 90%, on a test application based on arbitrary length arithmetic. (The performance model is currently being instantiated for the other systems. This will be reported in the nal paper.)

Cite this paper

@inproceedings{Hamdan1997SubmissionFS, title={Submission for Sc97 a Predictably Portable Process Farm Skeleton Extended Abstract}, author={Mohammad Hamdan}, year={1997} }