Massively parallel data optimization

Abstract

An optimizing compiler for a data parallel programming language can significantly improve program performance on a massively parallel computing system by incorporating new strategies for allocating array elements to processors. We discuss techniques for automatic layout of arrays in a Fortran compiler supporting Fortran 8x array features and targeted to the Connection Ma, chine computer system. Our goal is primarily to minimize the costs of moving data between processors and secondarily to minimize memory usage. Improved array layout may allow communications operations to be eliminated or to be replaced by more specialized communications operations having lower cost. We exhibit and discuss thoroughly a typical example of a code fragment that can be improved by a factor of 2 in memory consumption and a factor of 20 in speed.

Architecture-independent scientific programming in data parallel C

“…partitioning of data among the processors of a distributed memory computer[10]. They have developed sophisticated heuristics through which the compiler attempts to determine the best mapping of data to processors to reduce the amount of data communication.Chatterjee et al have studied the problem of automatic data partitioning of data-parallel programs [1 1].…”

A framework for efficient execution of array-based languages on SIMD computers

“…The burden that we seek to diminish is that of having to write the program according to a fixed domain decomposition. This paper is concerned with a framework in which domain-decompositions may be expressed and executed, but we do not address here how to arrive at an optimum (or even good) domain decomposition, although we expect it will be possible to use this framework to help discover decompositions from analysis of the source program along the lines of [Kn88] or [Ch88].…”