[MOAB-dev] Parallel iZoltan file for MOAB?

[Mark Miller]

Hi Tim,

Yeah, I know we really haven't made progress on defining a common,
parallel file format for ITAPS yet.

But, I did get us one step closer to being able to interchange the two
iMeshP test codes with iMeshP implementations.

Previously, MOAB_iMeshP_unitTest worked only via creating a mesh on the
fly. I played with it to optionally skip on-the-fly mesh creation and
instead read the mesh from a file. I then had it run in on-the-fly mode,
saved the resultant file (which it was doing anyways and that is the
file to which I was referring in my response to Jim P.) and can now run
in file mode using that file and it passes.

Once I got that working, I think tried linking MOAB_iMeshP_unitTest to
FMDB and run it in read-the-file mode using an FMDB sms file that
Seegyoung says is semantically identically to the on-the-fly mesh the
test was creating. The load indeed did work but the test failed later on
for other reasons and I gave up there.

I also tried linking iMeshP_unitTest with MOAB and then loading the .h5m
file I created (in the steps two paragraphs above). That also proceeded
past the load but failed for other reasons.

So, I took things as far as I could in interchanging the two tests and
two implementations. I agree a common file would take us the final step
two closure here. At the same time, if these two tests cannot run under
the conditions I've already outlined above, I am not sure having a
single, common parallel file format is going to improve that situation.

Given the above explanation, what do you think of things at this point?

Mark

On Wed, 2012-01-25 at 07:22 -0800, Tim Tautges wrote:
> As we've discussed before, there's currently no spec-approved file format for parallel loading.  The current parallel 
> itaps tests are hardwired to load implementation-specific files.  Until we have a spec-approved format, the same will 
> have to be done for the iZoltan test.
> 
> Also, at one point I requested an option in iZoltan to write the partitioning results to sets instead of entity-based 
> tags.  That can be a big memory savings for MOAB, and is in some ways a more appropriate use of the data model (that's 
> my opinion, anyway).  Until we have that, we'll have to continue to maintain and use exclusively MOAB's mbpart tool.
> 
> - tim
> 
> On 01/24/2012 06:13 PM, Mark Miller wrote:
> > Yeah, I think it is in the ITAPS repo under interfaces/iMeshP/data.
> > There should be a .h5m file there generated from the
> > MOAB_iMeshP_unitTest.
> >
> > Mark
> >
> > On Tue, 2012-01-24 at 14:20 -0800, James Porter wrote:
> >> At the ITAPS telecon last week, we discussed the weekly test results (as
> >> usual). One thing that came up was the parallel iZoltan tests, which
> >> we're currently failing[1]. We think the issue is that we need a file
> >> which MOAB can understand. Any ideas on the file we should be using for
> >> this?
> >>
> >> - Jim
> >>
> >> [1] https://itaps.mcs.anl.gov/resultsp.20120118/IZOLTANP_MOAB.results
> >>
> 
-- 
Mark C. Miller, Lawrence Livermore National Laboratory
================!!LLNL BUSINESS ONLY!!================
miller86 at llnl.gov      urgent: miller86 at pager.llnl.gov
T:8-6 (925)-423-5901    M/W/Th:7-12,2-7 (530)-753-8511