cp2k speedup on multicore machines

[Axel]

On Jan 30, 1:06 am, Fawzi Mohamed <fa... at gmx.ch> wrote:
> Hi Axel,
>
> nice numbers... just for the record
>
> > first off, contrary to fawzi's statement, cpu affinity in OpenMPI has
> > to be explicitely enabled (e.g. via setting mpi_paffinity_alone=1 in
> > ~/.openmpi/mca-params.conf).
>
> well I didn't say it was automatic :), what I meant was that I knew
> that both LAM MPI (on which I tested) and openMPI could do it, if

please check the LAM/MPI sources. to the best of my knowledge
there is no processor affinity in LAM/MPI. i just downloaded the 7.1.4
sources and there is no call to the respective system or library
calls.

> configured to do so, and then I told how to have it with LAM MPI...

sorry, but as i wrote, this only tells about locality, _not_
processor
affinity. schedulers are still kicking processes around.


> > however what both LAM/MPI and
> > particularly
> > OpenMPI have activated by default are algorithms that can take
> > advantage
> > of locality and that require the correct specification of nodes.
>
> yep, I think that Open MPI is the best choice to have a well
> performing MPI, especially with multiple cores.
> There is active development and they try to take advantage of the
> latest hardware.

yes. OpenMPI has the best collective algorithms
amongst all available open source packages.

cheers,
   axel.

>
> ciao
> Fawzi