Extending CRAFT Data-Distributions for Sparse Matrices (1996)
BibTeX
@MISC{Bandera96extendingcraft,
author = {G. Bandera and E.L. Zapata and Gerardo Bandera and Emilio L. Zapata},
title = {Extending CRAFT Data-Distributions for Sparse Matrices},
year = {1996}
}
Bookmark
 |
 |
 |
 |
 |
 |
OpenURL
Abstract
The existing methods for distributing data across the processors are not very useful for irregular problems. However, there are some new distribution techniques for this kind of problems, as MRD (Multiple Recursive Decomposition) and BRS (Block Row Scatter), which could be applicated by a data-parallel compiler. In this paper we present a performance comparison of regular (BLOCK and CYCLIC) versus pseudo-regular (BRS) data distribution for sparse matrices on the Cray T3D. We will show results for one of the most useful nonstationary iterative methods for solving symmetric positive definite systems, the Sparse Conjugate Gradient algorithm, using CRAFT with its data-distributions and using two parallelized by-hand codes written in C and Fortran with the BRS sparse distribution. 1 Overview In literature, there are several methods for programming parallel machines: automatic parallelizers (Polaris [6], Parafrase [16]), data-parallel compilers (VFCS [10], Adaptor [8]) and using message pas...
