A Block Orthogonalization Procedure with Constant Synchronization Requirements (Q4785908)
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 1848723
| Language | Label | Description | Also known as |
|---|---|---|---|
| default for all languages | No label defined |
||
| English | A Block Orthogonalization Procedure with Constant Synchronization Requirements |
scientific article; zbMATH DE number 1848723 |
Statements
A Block Orthogonalization Procedure with Constant Synchronization Requirements (English)
0 references
5 January 2003
0 references
Gram-Schmidt orthogonalization
0 references
parallel computation
0 references
Householder QR factorization
0 references
orthonormal basis
0 references
singular value decomposition
0 references
algorithm
0 references
stability
0 references
error analysis
0 references
numerical experiments
0 references
0.7864781618118286
0 references
0.7853095531463623
0 references
0.7834243178367615
0 references
The authors propose a new method for computing an orthonormal basis for a given skinny matrix from its right singular vectors. Advantages of this algorithm are that it uses only level 3 BLAS kernels and it requires only one synchronization step in parallel implementations. The stability of the algorithm is analysed. In the second part the authors discuss the problem where a set of vectors (inner set) is orthogonalized against a previously orthonormal set of vectors (external set) and then among themselves. The problem is solved by alternating iteratively between an external and an internal phase. In the external phase the internal vectors are orthogonalized against the external ones and in the internal phase internal vectors are orthogonalized among themselves. For the external phase a block Gram-Schmidt algorithm and for the internal phase the new algorithm is applied. An error analysis is presented. Numerical experiments performed on a Cray T3E, on an IBM SP-2 and on a cluster of SUN workstations show the effectiveness of the new method.
0 references