photograph

Research Project:

An Efficient and Reliable Block GMRES Algorithm

Professor: Prof. Martin H. Gutknecht


Date: 20.10.2007

Project Title: An Efficient and Reliable Block GMRES Algorithm

Summary:

In the thesis of Vital (1990), the generalized minimal residual method (GMRES) of Saad and Schultz (1986) was extended to a block version for linear systems with multiple right-hand sides. But while the formal extension of CG or GMRES to block CG (BlCG) and block GMRES (BlGMRES) is quite easy, safeguards are required for a robust implementation. The problem is that the vectors in the blocks associated to the multiple right-hand sides may become linear dependent, in which case one should reduce the block size. This is often referred to as (implicit) deflation.
While the reduction of block size (deflation) in the block Lanczos and block conjugate gradient algorithms has been discussed in detail in the literature, this is only partially true for BlGMRES, despite the fact that deflation may strongly reduce the number of matrix-vector products needed. Moreover, the necessity to restart gives us the option to choose a more generous deflation tolerance than in other block methods or to restrict deflation to restarts. But there is also a risk that goes with such strategies. The aim of the project is to find the right compromise.

Refs.: Martin H. Gutknecht, Block Krylov space methods for linear systems with multiple right-hand sides: an introduction, In: Modern Mathematical Models, Methods and Algorithms for Real World Systems (A.H. Siddiqi, I.S. Duff, and O. Christensen, eds.), 420--447, Anamaya Publishers, New Delhi, India, 2007.
Contacts:
Prof. M.H. Gutknecht
Seminar für Angewandte Mathematik
ETH Zürich
ETH-Zentrum, HG
CH-8092 Zürich

Tel.: +41-1-632 34 64
FAX: +41-1-632 11 04
EMail: gutknecht@math.ethz.ch