[Trilinos-Users] BLAS

[Heroux, Mike]

For a typical use of Epetra with preconditioned Krylov solvers, an optimized BLAS implementation will have little impact.  With the exception of one kernel in GMRES (a call to DGEMV), standard Krylov methods contain only BLAS1-type kernels (a single "for" loop over 1D arrays) which typically don't see much improvement from a vendor BLAS library vs what the compiler can do.  Furthermore, BLAS1 kernels do not perfectly match the needs of Krylov methods, so many BLAS1-type kernels cannot be replaced with BLAS1 library calls.

Use of optimized BLAS can have a significant impact on Epetra and related solvers if you are using:

  *   GMRES via AztecOO or Belos and have a large Krylov subspace size.
  *   Any block solvers in Belos or Anasazi.
  *   Intrepid.
  *   Any other part of Trilinos that calls BLAS 2 or 3 kernels.

Mike


From: "mrh1390 at gmail.com<mailto:mrh1390 at gmail.com>" <mrh1390 at gmail.com<mailto:mrh1390 at gmail.com>>
Date: Friday, October 11, 2013 10:19 AM
To: "trilinos-users at software.sandia.gov<mailto:trilinos-users at software.sandia.gov>" <trilinos-users at software.sandia.gov<mailto:trilinos-users at software.sandia.gov>>
Subject: [Trilinos-Users] BLAS

Dear All Trilinos Users,

I am wondering whether an optimized BLAS lib (like MKL or OpenBLAS) may significantly affect the performance of Epetra and consequently the linear solvers!?

I’ve tested OpenBLAS in windows (x64) and it shows no improvement for Belos, while it was incompatible with trilinos in parallel (MPICH) computations.

Please let me know if you have any experience.

Regards,
Mohammad

Sent from Windows Mail