SEMINAR
PSOR Orderings for Parallel Conjugate Gradient Preconditionings
Dexuan Xie
Department of Mathematics
University of Southern Mississippi
ABSTRACT
The PSOR method is an efficient parallel SOR method by domain partitioning. Theoretically, it can be regarded as the SOR method using a new ordering of unknowns--the PSOR ordering. This talk will discuss the effect of the PSOR ordering on the convergence rate of the preconditioned conjugate gradient method with the symmetric SOR (SSOR) method as preconditioner. It first shows that the PSOR ordering based on the strip domain partitioning is equivalent to the SWOS (strip without separators) ordering, which is the best parallel ordering for SSOR preconditioner among what we knew so far. It then shows that the PSOR ordering based on the block domain decomposition is a new parallel ordering for SSOR preconditioner, and outperforms the classic red-black ordering in both the convergence rate and the interprocessor communication. Numerical experiments were made on two MIMD parallel machines (an SGI Origin 2000 and a cluster of workstations).
WHERE: TEC 205
WHEN(day): Friday, April 27th, 2001
WHEN(time): 2:00pm
EVERYBODY IS INVITED