Abstract

High performance computing and communication (HPCC) requires that space and time usage are kept to a minimum when solving very large scientific problems. These problems often are represented by sparse systems, and in particular banded systems. In this paper, a new banded array index algorithm is presented that uses substantially less memory than the BLAS (Basic Linear Algebra Subroutines) representation. The algorithm presented here works for multi-dimensional arrays and on all bandwidths. It is visualized as a layer between the applications and the data, thus shielding the caller from the complexities of the sparse representation. Information is returned to the caller so that the indexed item may be correctly operated on. To consider the time requirement, large problems are distributed, and solved in parallel by several processors. Automatic partitioning and processor mapping is performed using primitives that are based on the shape of the array, the upper half-bandwitdh and lower half-bandwidth. These inherent structural parameters are used in the index function to efficiently generate requested addresses.

Department(s)

Computer Science

Comments

The first Author is a Graduate Student.

Report Number

CSc-94-23

Document Type

Technical Report

Document Version

Final Version

File Type

text

Language(s)

English

Rights

© 1994 University of Missouri - Rolla, All rights reserved

Publication Date

1 June, 1994

Download

Share

 
COinS