Date of Award

2005

Level of Access

Open-Access Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Advisor

James L. Fastook

Second Committee Member

Phillip M. Dickens

Third Committee Member

David Hiebler

Abstract

Two current software packages for solving large systems of sparse simultaneous l~neare equations are evaluated in terms of their applicability to solving systems of equations generated by the University of Maine Ice Sheet Model. SuperLU, the first package, has been developed by researchers at the University of California at Berkeley and the Lawrence Berkeley National Laboratory. UMFPACK, the second package, has been developed by T. A. Davis of the University of Florida who has ties with the U. C. Berkeley researchers as well as European researchers. Both packages are direct solvers that use LU factorization with forward and backward substitution. The University of Maine Ice Sheet Model uses the finite element method to solve partial differential equations that describe ice thickness, velocity,and temperature throughout glaciers as functions of position and t~me. The finite element method generates systems of linear equations having tens of thousands of variables and one hundred or so non-zero coefficients per equation. Matrices representing these systems of equations may be strictly banded or banded with right and lower borders. In order to efficiently Interface the software packages with the ice sheet model, a modified compressed column data structure and supporting routines were designed and written. The data structure interfaces directly with both software packages and allows the ice sheet model to access matrix coefficients by row and column number in roughly 100 nanoseconds while only storing non-zero entries of the matrix. No a priori knowledge of the matrix's sparsity pattern is required. Both software packages were tested with matrices produced by the model and performance characteristics were measured arid compared with banded Gaussian elimination. When combined with high performance basic linear algebra subprograms (BLAS), the packages are as much as 5 to 7 times faster than banded Gaussian elimination. The BLAS produced by K. Goto of the University of Texas was used. Memory usage by the packages varted from slightly more than banded Gaussian elimination with UMFPACK, to as much as a 40% savings with SuperLU. In addition, the packages provide componentwise backward error measures and estimates of the matrix's condition number. SuperLU is available for parallel computers as well as single processor computers. UMPACK is only for single processor computers. Both packages are also capable of efficiently solving the bordered matrix problem.

Share