dolfin team mailing list archive

[Johan Hake <hake@xxxxxxxxx>]

On Monday 06 April 2009 22:06:19 Johannes Ring wrote:
> On Mon, April 6, 2009 21:05, Anders Logg wrote:
> > On Mon, Apr 06, 2009 at 06:45:27PM +0200, Johannes Ring wrote:
> >> On Mon, April 6, 2009 17:19, Anders Logg wrote:
> >> > As far as I understand, the problems we see now are related to some
> >> > intricate problems loading MPI in Python on Mac with Trilinos.
> >>
> >> Yes, that's correct. The last failure was because of an issue with
> >> Trilinos 9.0.2 on Mac (got a Bus error in some ML tests). I just
> >> reverted
> >> to an earlier version (8.0.3) on the mac-osx slave and now the Trilinos
> >> demo worked fine so hopefully the mac-osx slave will turn green after
> >> the
> >> next build.
> >
> > The buildbot is currently red even for Hardy due to a timeout when
> > running the waveguide demo. I tried it here and it takes a very long
> > time to finish (waiting for SLEPc/LAPACK to converge). I changed the
> > resolution so it finishes in just a few seconds. Strange that this
> > shows up now. The buildbot has been green before.
>
> It's not that strange considering you increased the size of the mesh in
> the commit before the Hardy and Intrepid slaves failed:
>
> http://www.fenics.org/hg/dolfin?cs=f6478f948adbbdd29a659708473c08266cc7f07a
>
> >> > There are quite a few items in the queue both for DOLFIN and FFC that
> >> > will break both functionality and interfaces and it would be good to
> >> > release now so we can get started.
> >>
> >> The only remaining problem now is with Open MPI 1.3. The fix we did
> >> earlier today didn't work because the call to 'import cpp' made Python
> >> exit after the exception was raised in this code:
> >>
> >> try:
> >>     import cpp
> >> except ImportError, msg:
> >>     ...
> >>
> >> We need to come up with a better test.
> >>
> >> Johannes
> >
> > Does this always happen with Open MPI 1.3?
>
> Yes, unless we run something like
>
> import ctypes
> ctypes.CDLL('libmpi.so', ctypes.RTLD_GLOBAL)
>
> before importing the cpp module (as we do in dynamic_load_mpi.py). The
> problem is what to do if a user doesn't have mpi? Maybe we can do it like
> this:
>
> try:
>     import dynamic_load_mpi
> except OSError:
>     pass
> import cpp

I do not know the ctypes module, and the kind of errors that potentially could 
rise from a call to CDLL. We also need to make the call platform independent, 
that's probably a minor problem.

The whole fix also caused a segfault on a cluster, I tried to run some jobs on 
today. Reverting the fix and the jobs went smooth. So I have a fishy fealing 
about this "fix".

Johan