No label defined (Q4035550)
From MaRDI portal
!
This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:
scientific article; zbMATH DE number 176399
| Language | Label | Description | Also known as |
|---|---|---|---|
| default for all languages | No label defined |
||
| English | No label defined |
scientific article; zbMATH DE number 176399 |
Statements
18 May 1993
0 references
biconjugate gradient-type algorithm
0 references
non-Hermitian linear systems
0 references
massively parallel architectures
0 references
parallel computation
0 references
conjugate gradient method
0 references
quasi-minimal residual method
0 references
sparse linear systems
0 references
Lanczos algorithm
0 references
look-ahead strategy
0 references
numerical examples
0 references
0.91938573
0 references
0.89037967
0 references
0.8882996
0 references
0.8864283
0 references
0.8815774
0 references
0.88007194
0 references
0.8780887
0 references
0.87722343
0 references
[For the entire collection see Zbl 0753.00017.] Except the well known conjugate gradient (CG) method for the special case that the nonsingular \(N\times N\) matrix \(A\) is Hermitian and positive definite, there has not been known any satisfactory algorithm for iterative solution of very large but sparse linear systems \(Ax = b\). In fact, an algorithm, for general non-Hermitian matrices, that is able to compete in efficiency with the above CG method was proposed by \textit{R. W. Freund} and \textit{N. M. Nachtigal} [Numer. Math. 60, No. 3, 315-339 (1991; Zbl 0754.65034)] and called the quasi-minimal residual (QMR) method. This constitutes the last step in a series of successive improvements proposed during a decade by several authors. However, the QMR method is not very suitable when the modern massively parallel machines are used. To enforce the biorthogonality condition contained in the method, the inner products of some vectors need to be computed. In the original QMR these inner products are calculated within each iteration step. As the authors state, on massively parallel machines such a sequential computation usually constitutes a bottleneck. The present paper gives a version of the QMR method that overcomes this problem. It is crucial in this version, called the \(s\)-step Lanczos algorithm with look-ahead, to keep the block size \(s\) variable and combine the process with a suitable look-ahead strategy. Such an approach has the advantage that all inner products arising in the same block can be computed in parallel. A sketch for the whole algorithm is given. The advantages of the method are illustrated by three numerical examples, all concerning the solution of partial differential equations.
0 references