NWChem/web/doc/peigs/docs/performance.old.html
2006-01-12 21:18:59 +00:00

39 lines
1.4 KiB
HTML

<HTML>
<!-- DO NOT EDIT: this is a cleared public file -->
<HEAD>
<TITLE>Performance of PeIGS</TITLE>
</HEAD>
<BODY>
<H1>Performance of PeIGS</H1>
<P>
The numerical method used by PeIGS is bisection for eigenvalues
and repeated inverse iteration and orthogonalization for eigenvectors.
Accuracy and orthogonality are similar to
LAPACK's DSPGV and DSPEV.
Orthogonality of eigenvectors is enforced even for arbitrarily large
clusters that span multiple processors.
<P>
PeIGS is both fast and scalable -- on a single processor of a KSR-2 it takes
only 45% longer than LAPACK for a 1000-by-1000 generalized
eigenproblem, and parallel efficiency is good even for large
processor counts. For example, in one of our computational chemistry
applications, the standard eigenproblem Ax = kx was solved for all
eigenpairs of a 2053-by-2053 matrix. This computation required only 117
seconds on 150 processors of an Intel Paragon computer, a time-to-solution
estimated as 77 times faster than the 1 processor time.
<P>
PeIGS performance depends to some extent on how eigenvalues cluster
within the spectrum. Generally speaking, systems with many eigenvalues
close together take longer to solve due to increased computation and
communication costs for reorthogonalization. This is an important
issue because actual applications with large matrices often have
highly clustered spectra.
</BODY>
</HTML>