The NAS Parallel Conjugate Gradient Benchmark on the Cray T3D

Abstract

The NAS Parallel Benchmark kernels were developed to evaluate the performance of highly-parallel supercomputers. These benchmarks are unstructured in the sense that they give only an algorithm defining the benchmark; all implementation details are left to the programmer. We looked specifically at a kernel to solve an unstructured sparse linear system via the conjugate gradient method, the CG kernel. This kernel was implemented on one of the newest massively-parallel supercomputers, the Cray T3D. Currently, our implementation of the CG kernel on the T3D achieves 306 MFlops on 64 processors. In comparison, a Cray YMP single head gets 127 MFlops, a 128 processor iPSC/860 gets 181 MFlops, and a 32,768 processor CM-2 gets 105 MFlops.

@techreport{SRC-TR-94-192,  
author = {Tamara L. {Gibson (nee Kolda)}}, 
title = {The {NAS} Parallel Conjugate Gradient Benchmark on the {Cray} {T3D}}, 
number = {SRC-TR-94-192}, 
institution = {Supercomputing Research Center, Bowie, MD}, 
year = {1994},
}