🚀 go-pugleaf

Amsterdam.


DEPARTMENT OF COMPUTER SCIENCE
UNIVERSITY OF WATERLOO
SEMINAR ACTIVITIES

SCIENTIFIC COMPUTATION SEMINAR

Professor    Rob   H.   Bisseling,   Koninklijke/Shell-
Laboratorium, Amsterdam will speak on ``Parallel Sparse
Linear System Solving on a Transputer Network.

TIME:                 3:30 p.m.

ROOM:                 DC 1304

DATE:           Monday, February 12, 1990 *NOTE DATE!*

ABSTRACT

Large  sparse systems of linear equations are important
in  many  areas,  such  as  e.g. PDE-solving and Linear
Programming.   This  talk presents a parallel algorithm
for  the  direct  solution  of  a  sparse linear system
Ax = b  on a local memory MIMD computer.  The algorithm
-- - -
consists  of the LU decomposition of A, followed by the
                                     -
solution of two triangular systems.

The  matrix  A  is  distributed  across  the processors
             -
according  to  the grid (i.e., scattered) distribution.
Each  processor represents the nonzeroes of its part of
the  sparse  matrix  by  a two-dimensional linked list.
The   pivot   elements  of  the  LU  decomposition  are
determined   by   a  Markowitz  criterion  to  preserve
sparsity, with threshold pivoting to maintain numerical
stability.  The algorithm has been implemented in Occam
2,  and  experimental  results  have been obtained on a
mesh of 36 transputers.

This  algorithm  is  part  of  the  PARPACK  library of
parallel   programs   for   dense   and  sparse  matrix
computations  which  is  currently under development at
KSLA.