skip to main content
OSTI.GOV title logo U.S. Department of Energy
Office of Scientific and Technical Information

Title: Maximizing sparse matrix vector product performance in MIMD computers

Abstract

A considerable component of the computational effort involved in conjugate gradient solution of structured sparse matrix systems is expended during the Matrix-Vector Product (MVP), and hence it is the focus of most efforts at improving performance. Such efforts are hindered on MIMD machines due to constraints on memory, cache and speed of memory-cpu data transfer. This paper describes a strategy for maximizing the performance of the local computations associated with the MVP. The method focuses on single stride memory access, and the efficient use of cache by pre-loading it with data that is re-used while bypassing it for other data. The algorithm is designed to behave optimally for varying grid sizes and number of unknowns per gridpoint. Results from an assembly language implementation of the strategy on the iPSC/860 show a significant improvement over the performance using FORTRAN.

Authors:
; ; ;
Publication Date:
Research Org.:
Front Range Scientific Computations, Inc., Boulder, CO (United States); USDOE, Washington, DC (United States); National Science Foundation, Washington, DC (United States)
OSTI Identifier:
219559
Report Number(s):
CONF-9404305-Vol.2
Journal ID: ISSN 0743-7315; ON: DE96005736; TRN: 96:002321-0007
Resource Type:
Conference
Resource Relation:
Journal Volume: 37; Journal Issue: 2; Conference: Colorado conference on iterative methods, Breckenridge, CO (United States), 5-9 Apr 1994; Other Information: PBD: [1994]; Related Information: Is Part Of Colorado Conference on iterative methods. Volume 2; PB: 261 p.
Country of Publication:
United States
Language:
English
Subject:
99 MATHEMATICS, COMPUTERS, INFORMATION SCIENCE, MANAGEMENT, LAW, MISCELLANEOUS; MATRICES; MATHEMATICS; PARALLEL PROCESSING; OPTIMIZATION

Citation Formats

McLay, R T, Kohli, H S, Swift, S L, and Carey, G F. Maximizing sparse matrix vector product performance in MIMD computers. United States: N. p., 1994. Web. doi:10.1006/jpdc.1996.0115.
McLay, R T, Kohli, H S, Swift, S L, & Carey, G F. Maximizing sparse matrix vector product performance in MIMD computers. United States. doi:10.1006/jpdc.1996.0115.
McLay, R T, Kohli, H S, Swift, S L, and Carey, G F. Sat . "Maximizing sparse matrix vector product performance in MIMD computers". United States. doi:10.1006/jpdc.1996.0115. https://www.osti.gov/servlets/purl/219559.
@article{osti_219559,
title = {Maximizing sparse matrix vector product performance in MIMD computers},
author = {McLay, R T and Kohli, H S and Swift, S L and Carey, G F},
abstractNote = {A considerable component of the computational effort involved in conjugate gradient solution of structured sparse matrix systems is expended during the Matrix-Vector Product (MVP), and hence it is the focus of most efforts at improving performance. Such efforts are hindered on MIMD machines due to constraints on memory, cache and speed of memory-cpu data transfer. This paper describes a strategy for maximizing the performance of the local computations associated with the MVP. The method focuses on single stride memory access, and the efficient use of cache by pre-loading it with data that is re-used while bypassing it for other data. The algorithm is designed to behave optimally for varying grid sizes and number of unknowns per gridpoint. Results from an assembly language implementation of the strategy on the iPSC/860 show a significant improvement over the performance using FORTRAN.},
doi = {10.1006/jpdc.1996.0115},
journal = {},
issn = {0743-7315},
number = 2,
volume = 37,
place = {United States},
year = {1994},
month = {12}
}

Conference:
Other availability
Please see Document Availability for additional information on obtaining the full-text document. Library patrons may search WorldCat to identify libraries that hold this conference proceeding.

Save / Share: