Slate performance on Generalized eigenvalues decomposition

31 views

Skip to first unread message

andrea chiariello

unread,

Oct 17, 2023, 8:50:37 AM10/17/23

to SLATE User, Salvatore Ventre

Dear Slate developers,
Salvatore Ventre and I are currently developing a code where a generalized eigenvalue decomposition is applied to full matrices with dimensions ranging from 150k to 350k unknowns..

We are currently using the pdsygvx routine and would like to understand the performance we can expect from SLATE. Do you have any benchmarks that indicate the potential improvements achievable?

Is there a Fortran interface for the eigenvalue decomposition routines?

Sincerely,

Andrea & Salvatore

Mark Gates

unread,

Oct 25, 2023, 9:45:00 AM10/25/23

to andrea chiariello, SLATE User, Salvatore Ventre

Hi Andrea,

The functionality for a generalized eigensolver, either with or without eigenvectors, should work now in SLATE. At the moment, it isn't scaling up well due to the 2nd stage, reduction from band to tridiagonal.

I'll look around to see what results we may have. The 2-stage eig algorithm prefers small block sizes, precisely for that 2nd stage, something like nb = 192.

The C and Fortran interfaces are a little out of date. It looks like they have eigenvalues but not yet vectors.