[Trilinos-Users] Sparse matrix-vector multiplication with Tpetra::CrsMatrix and MKL

[Christopher Thiele]

Hello again,

I did some further measurements and it looks like some threads do not
participate in the matrix-vector multiplications at all. I found out that
in the code the number of threads to be used is dynamically determined. I
would like to override this and force a fixed number of threads. Is there a
way to do this?

Regards,
Christopher

On Fri, Oct 2, 2015 at 1:49 PM, Christopher Thiele <
christopher.thiele92 at gmail.com> wrote:

> Hello all,
>
> I ported a PDE solver to Trilinos and did some performance evaluations.
> The most demanding part are PCG solves, as expected. However, especially
> during the sparse matrix-vector products (SPMV) some OpenMP threads are
> spinning a lot and this is getting worse when the number of threads is
> increased. To my understanding, the SPMV does not require any
> synchronization or communication between the threads, so I can't see a
> reason for this behavior.
>
> Also, apparently the SPMV does not use the Intel MKL. Instead the method
> spmv_alpha_beta_no_transpose() from
> packages/tpetra/kernels/src/impl/Kokkos_Sparse_impl_spmv.hpp is being used.
> Is there a way to change that?
>
> Kind regards,
> Christopher
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://trilinos.org/pipermail/trilinos-users/attachments/20151005/9b4ec7e4/attachment.html>