[MOAB-dev] Performance regressions in MOAB

[Jason Kraftcheck]

On 02/19/2011 11:20 PM, James Porter wrote:
> This is a bit long after the fact, but I noticed when running PyTAPS
> tests locally that the iMesh checks were taking an awfully long time. I
> looked through the buildbot history and it looks like the biggest
> performance loss happened on the Jan 27 nightly, which makes me think
> that r4473 <http://trac.mcs.anl.gov/projects/ITAPS/changeset/4473> is
> culprit. 

Have you tried changing the constants back to 4k?

> I'm sure that the extremely small-scale tests I'm doing inflate
> the performance regression, but the tests originally took 1-2 seconds,
> and are now up to 70-80 seconds on the buildbot. 

Are you creating and destroying many MOAB instances?  If so, then the
culprit is likely time time spent initializing the memory allocated for the
SequenceDatas.  If that is the case, then we could rearrange things such
that the memory isn't initialized unless the corresponding entities are
allocated.  But I don't know that it is worth the time for what is an
unlikely use case.

> This may also be why I
> currently can't run "make check" on MOAB locally because it will eat up
> all available memory and start thrashing. I'm not sure if this is
> actually an issue when doing real work, but I wanted to give everyone a
> heads up.
> 

The most memory use I see when doing "make check" for MOAB is exodus_test
consuming 3 MB.

- jason