2007
DOI: 10.1016/j.parco.2007.04.004
|View full text |Cite
|
Sign up to set email alerts
|

Parallel Arnoldi eigensolvers with enhanced scalability via global communications rearrangement

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
1
1
1
1

Citation Types

0
46
0
2

Year Published

2007
2007
2022
2022

Publication Types

Select...
6
3
1

Relationship

0
10

Authors

Journals

citations
Cited by 53 publications
(48 citation statements)
references
References 23 publications
0
46
0
2
Order By: Relevance
“…Replacing Modied Gram-Schmidt with Classical Gram-Schmidt, although considered to be less stable, requires fewer synchronization points in methods which require explicit orthogonalization [18].…”
Section: Related Work In Avoiding Communication In Ksmsmentioning
confidence: 99%
“…Replacing Modied Gram-Schmidt with Classical Gram-Schmidt, although considered to be less stable, requires fewer synchronization points in methods which require explicit orthogonalization [18].…”
Section: Related Work In Avoiding Communication In Ksmsmentioning
confidence: 99%
“…Second, for efficiency reasons, Algorithm 3.1 (lines 5 and 6) uses the classical Gram-Schmidt procedure instead of modified Gram-Schmidt, but in a practical implementation this process must be repeated twice to enhance numerical stability. See [10] for details on parallel implementation of iterated Gram-Schmidt. Finally, although full reorthogonalization is performed, all Gram-Schmidt coefficients are discarded except the diagonal element α j , as is done in practical Lanczos implementations, and similarly we only store the diagonal elements of Ω k , even if V * k BV k is not really diagonal in finite precision arithmetic (furthermore, ω j 's are rounded to ±1 when Lanczos vectors are B-normalized).…”
Section: Pseudo-lanczos Methodsmentioning
confidence: 99%
“…The default configuration (used in §2.4) employs iterative CGS (which yields better parallel scaling and higher floating point operations throughput than MGS), with a DGKS-like re-orthogonalization criterion. See [Hernandez et al, 2007] for details.…”
Section: Subspace Orthogonalizationmentioning
confidence: 99%