Parallel methods for least squares problems
216.
L. T. Yang and R. P. Brent,
Parallel MCGLS and ICGLS methods for least squares problems on distributed
memory architectures,
Proc. ISPA 2003 (Aizu-Wakamatsu, Japan, July 2003),
Lecture Notes in Computer Science, Vol. 2745,
Springer-Verlag, Berlin, 2003, 197-208.
Revision in The Journal of Supercomputing, 29 (2004), 145-156.
Paper:
LNCS version available from
Springer Link.
Abstract
In this paper, we describe a modified CGLS method (MCGLS)
that improves parallel
performance by assembling the results of a number of inner products
and by overlapping communication with computation.
We also propose an improved CGLS method (ICGLS) that halves the number of
global synchronization points required for inner products.
For a longer abstract, see the
paper.
Go to next publication
Return to Richard Brent's index page