[petsc-users] MPI-IO

[Barry Smith]

On Jul 21, 2014, at 11:25 AM, Jed Brown <jed at jedbrown.org> wrote:

> Barry Smith <bsmith at mcs.anl.gov> writes:
> 
>>> but it would be nice to have a common way to
>>> address other viewer implementations.
>> 
>>   What do you mean by this?
> 
> That we should be able to tell the VTK viewer to use MPI-IO in the same
> way we tell the Binary viewer or any other viewer capable of using MPI-IO.

   Yes the correct PetscViewerFormat is limited in that there is only one supported at a time. We could go the horrible bit oring of options or we could collect multiple options instead of just one.

> 
>>>>> PETSc devs, do you suppose MPI-IO support is stable enough that we could
>>>>> make this a default?
>>>> 
>>>>    It is not so much a question of stability, it is more a question
>>>>    of (I believe) MPI IO is just loads slower for what might normal
>>>>    PETSc users use, it only pays off for large numbers of nodes.
>>> 
>>> Do we have current data demonstrating that this is still the case?  It's
>>> really a failure of the implementations if they can't fall back to the
>>> dumbest mechanism when it's faster.
>> 
>>   Oh, actually I just remembered what always drove me crazy with MPI
>>   IO is that it saves it in native format meaning it is not portable
>>   to all systems (which have different endian). Portability before
>>   performance :-)
> 
> Uh, MPI_Pack_external and MPI_Unpack_external?

   Ok, so maybe the finally implemented them. We don’t use this BTW because Rajev told me (a long time ago) that MPICH doesn’t support it.

  Barry