In this paper we describe how performance models can be used as a
more structured approach to the problem of resource allocation when programming heterogeneous parallel systems. Functional skeletons are used in this paper to co-ordinate parallel computation in a heterogeneous system. An advantage of this approach is the ability to associate performance models with the implementations of a skeleton. We describe how these performance models can be used to predict the cost of of a particular resource allocation strategy over an entire program expressed using skeletons. Using a parallel conjugate gradient algorithm as a case study we investigate the approach on a mixed vector and scalar parallel machine when exploiting SPMD and simple MPMD parallelism in the algorithm.
pubs.doc.ic.ac.uk: built & maintained by Ashok Argent-Katwala.