[Pw_forum] Trying to find reasonable parallelization parameters

Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu
Thu Jan 22 23:49:36 CET 2009


oh... and one more thing.

_what_ MPI are you using? most of the time, the infiniband MPI implementations
are not optimized for large problems and allocate too many non-swappable RDMA
buffers. now, this can create the impression of running out of memory,
but in reality
all you'd need to do is to change the settings of your MPI lib to not
use fixed buffers,
but rather shared request queues. have a chat with the sysadmin folks
of the machine.

in any case, since there is no way to make any predictions, you should first do
some systematic scaling tests (perhaps with gamma only) and see whether it
would run better with less processors (at the numbers you are mentioning, this
is what i would definitely expect) altogether, and then less processors pernode
(i.e. allocate all cores of a node and then use only half of them).

cheers,
    axel.

On Thu, Jan 22, 2009 at 5:38 PM, Axel Kohlmeyer
<akohlmey at cmm.chem.upenn.edu> wrote:
> On Thu, 22 Jan 2009, J. J. Ramsey wrote:
>


-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
  Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.



More information about the users mailing list