parallel distribution of data

[Axel]

On Mar 11, 10:03 am, "Nichols A. Romero" <naro... at gmail.com> wrote:
> Axel,
>
> Is this related to OPENMPI gobbling up tons of memory?

how would you tell it is open-mpi?
it actually is the lowlevel libraries that have
these requirements. when using OpenMPI there are also
a bunch of flags that you can use the ease the memory
allocation load. particularly, putting

mpi_leave_pinned = 1

into

~/.openmpi/mca-params.conf

seemed to help a lot in my runs.

you can reduce the risk of crashes due to timeouts with:
btl_openib_ib_timeout = 12

and you can use the following flag to see the efficiency
of the "lazy pinning":
mpool_rdma_print_stats = 1

and finally, if you are on a machine with NUMA support
you want to try whether

mpi_paffinity_alone = 1

has some effect.


cheers,
   axel.

>
> On 3/10/08, Axel <akoh... at gmail.com> wrote:
>
>

> --
> Nichols A. Romero, Ph.D.
> DoD User Productivity Enhancement and Technology Transfer (PET) Group
> High Performance Technologies, Inc.
> Reston, VA
> 443-567-8328 (C)
> 410-278-2692 (O)