The development of a new preconditioner by modifying the simply sparse compression matrix to solve electromagnetic method of moments problems
No Thumbnail Available
Date
2010-02-12T12:26:00Z
Authors
Dreyer, Renier Lambertus
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The aim of this research was to improve the matrix solution methods for SuperNEC MoM
problems, which is an electromagnetic simulation software package used to model
antennas, and develop a new preconditioner for the iterative method BICGSTAB(L).
This was achieved by firstly implementing the ATLAS BLAS library optimised for a
specific computer architecture. The ATLAS code primarily makes use of code generation
to build and optimise applications. Comparisons show that the matrix solution times
using LU decomposition optimised by ATLAS is improved by between 4.1 and 4.6 times,
providing a good coding platform from which to compare other techniques.
Secondly the BICGSTAB iterative solution method in SuperNEC was improved by
making use of an alternative algorithm BICGSTAB(L). Systems of equations that
converged slowly or not at all using BICGSTAB, converged more quickly when using
BICGSTAB(L) with L set to 4, despite the high condition numbers in the coefficient
matrices.
Thirdly a domain decomposition method, Simply Sparse, was characterised.
Investigations showed that Simply Sparse is a good compression technique for SuperNEC
MoM matrices. The custom Simply Sparse solver also solves large matrix problems more
quickly than LU decomposition and scales well with increased problem sizes. LU
decomposition is still however quicker for problems smaller than 7000 unknowns as the
overheads in compressing the coefficient matrices dominate the Simply Sparse method
for small problems.
Lastly a new preconditioner for BICGSTAB(L) was developed using a modified form of
the Simply Sparse matrix. This was achieved by considering the Simply Sparse matrix to
be equivalent to the full coefficient matrix [A] . The largest 1% to 2% of the Simply
Sparse elements was selected to form the basis of the preconditioning matrix. These
elements were further modified by multiplying them by a large constant i.e. 7 1×10 . The
system of equations was then solved using BICGSTAB(L) with L set to 4.
The new preconditioned BICGSTAB(L) algorithm is quicker than both LU
decomposition and the custom Simply Sparse solution method for problems larger than
5000 unknowns.