A linear algebra formulation for optimising replication in data parallel programs

Abstract

n this paper, we present an effcient technique for optimising data replication under the data parallel programming model. We propose a precise mathematical representation for data replication which allows handling replication as an explicit, separate stage in the parallel data placement problem. This representation takes the form of an invertible mapping. We argue that this property is key to making data replication amenable to good mathematical optimisation algorithms. We further outline an algorithm for optimising data replication, based on this representation, which performs interprocedural data placement optimisation over a sequence of loop nests. We have implemented the algorithm and show performance figures.