LAPACK/ScaLAPACK Development

Hi, There:

I got two questions and hope you can help me. I am using PDSYEVX routine

1). What is the suggested block size in order to calculate the eigenvalue
and eigenvector for a symmetric matrix of size 20k X 20k, also what about 40k X 40k?

2). Given an eight-node cluster, can I set the prow and pcol to be 4 and 2
respectively for the aforementioned calculation?

Your help is higly appreciated!

Bruce

Based on ScaLapack User's Guide, which is BTW 10 years old, block size should be at least 64x64. For processor matrix, 4x2 is better than 8x1.

Just my 2 cents. I think real performance depends on your specific machine also. Better way is just to experiment.

The SEP problem (Symmetric Eigenvalue Problem) is not that much sensitive to the grid size. So basically take a grid as square as you can and you shoudl be good.
64x64 is a good idea. 80x80 when using ATLAS BLAS. Well, you need to try to know.
But think also to change of algorithms, if you want a few of the eigenvectors then PDSYEVX
is a good bet, but if you want them all, what about trying PDSYEVD.
Julien

Hi, Julien:

Thank you for your advice!

Bruce