Sparse data representation for data-parallel computation

Abstract

Performance optimization has ben achieved by a transparent parallel sparse data representation in a data-parallel programming environment. In a sparse data representation, only the non-zero data elements of an array are stored and processed. The parallel sparse data representation is designed to efficiently utilize system resources on multicomputer systems for a broad class of problems; the main focus of this work is on the sparse situations that arise in dense data-parallel algorithms rather than the more traditional sparse linear algebra applications. A number of sparse data formats have been considered; one of these formats has been implemented in a high-level data-parallel programming environment called Paragon. Experimental results have been obtained with a distributed-memory multicomputer system.<>