parallel algorithms for the iterative solution of large sparse linear systems
abstract
in this thesis we are concerned with iterative parallel
algorithms for solving finite difference equations arising
from boundary value problems. we propose various new
methods based on approximate inverses. the jacobi iterative
method for the solution of linear equations is highly
parallel. its slow convergence rate, however, has initiated
the study of a variety of acceleration techniques. we
introduce the “diagonal-block” and the "least squares"
techniques for calculating approximate inverses which lead
naturally to generalized jacobi methods that are well suited
to parallel processing. in addition, the calculations
required to establish the iterative process can be done in
parallel. a convergence theorem for a one-dimensional model problem is given. gauss-seidel versions of these methods
are also considered. these versions converge faster but
often at the cost of decreased parallelism.
preconditioning has been successfully used in
conjunction with the conjugate gradient method. many such
techniques use an approximation m to the matrix a of the
linear system under consideration and solve a system of the
form m z = d at each iteration. if, however, an
approximate inverse m[superscript -1] is used, the solution z = m[superscript -1] d only involves a matrix-vector multiplication v/hich is well
suited to parallel computation. we examine the diagonal-block and least-squares techniques for preconditioning the conjugate gradient method.
numerical results are presented. the proposed jacobi
like and gauss-seidel like methods are compared with
parallel gauss-seidel methods in a class of parallel
algorithms proposed by evans and barlow (1982). the
preconditioned conjugate gradient methods are compared with
the plain conjugate gradient method. in all cases the speed
of convergence increased substantially, but often at the
expense of increased computational work. in some cases it
was necessary to assume the number of processors to be on
the order of n ( the number of unknowns ) in order to gain
an advantage from parallel processing.
collections
- retrospective theses [1604]