On Wed, Jul 24, 2013 at 11:27 AM, Stephen Wornom <[email protected]>wrote:

> <Yes, though it's much better if you use MPI-IO collective IO. The worst is
> <if every process reads a different file.
>
> The unstructured code I use collects the different core solution on
> processor 0 then writes the global solution file to be used at a restart.
> When I have a mesh > 2 MillionNodes
> I cannot execute as the RAM/core is 3.4Gb which is sufficient to execute
> but the virtual memory is too large and I get an error.
>
> Three  questions:
> 1- If I used mpi-io to write the file, would that reduce the virtual
> memory needed and permit  me to run larger meshes?
>

Yes, since you would not be using only the memory on one process


> 2- How would I use the mpi-io routines in the oetsc library to write the
> files? The mesh is partitioned using metis.
>

Make a Vec with your information, and use VecView(). I recognize that this
only writes doubles. We do not have infrastructure
for MPI/IO with other data types.


> 3- Is there a coding example of mpi-io writes that I could use to
> understand how petsc functions?
>

Should not be necessary.

   Matt


> Thanks if advance,
> Stephen
>
> --
> [email protected]
> 2004 route des lucioles - BP93
> Sophia Antipolis
> 06902 CEDEX
>
> Tel: 04 92 38 50 54
> Fax: 04 97 15 53 51
>
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

Reply via email to