Scientific Computing Seminar

Date:

Tuesday, September 27, 2005

Time:

2:00pm-3:00pm

Location:

70-191

Seminar Speaker:

Stephane Pralet
CNRS - ENSEEIHT/IRIT, Toulouse

Title:

Improved scaling and pivoting strategies for sparse parallel L D L^T factorizations

Abstract:

We present recent advances for computing the $L D L^T$ factorization of a symmetric indefinite matrix.
We first describe the adaptation of maximum weighted matching and scaling techniques to the symmetric indefinite case. We show that our new scheme reduces the number of delayed pivots while still maintaining an accurate factorization. This improves the factorization time, the memory estimation and the memory requirements.

We then discuss new alternatives to numerical pivoting. We first cover the sequential factorization. Secondly, we propose strategies that are numerically robust and do not limit the scalability of a parallel distributed factorization. They are based on the estimation of element growth and do not require any supplementary messages. Finally, we show the positive influence of recently developed orderings on our static pivoting strategies.

Sponsor of Seminar:

Sherry Li

Scientific Computing

Contact Esmond G. Ng EGNg@lbl.gov