[gpaw-users] MPI support without gpaw-python

[Gaël Donval]

On Tue, 2018-08-07 at 16:25 +0000, Gaël Donval via gpaw-users wrote:
> On Tue, 2018-08-07 at 15:01 +0000, Gaël Donval via gpaw-users wrote:
> > On Tue, 2018-08-07 at 16:55 +0200, Jens Jørgen Mortensen wrote:
> > > On 08/07/2018 04:45 PM, Gaël Donval via gpaw-users wrote:
> > > > Hi,
> > > > 
> > > > Is it still possible to get MPI support without compiling the
> > > > custom
> > > > interpreter?
> > > 
> > > I don't know if it still works.  Try to turn on this section:
> > > 
> > >      https://gitlab.com/gpaw/gpaw/blob/master/customize.py#L80
> > > 
> > > compile and then set the environment variable $GPAW_MPI to point
> > > to
> > > your 
> > > dynamic MPI library.
> > 
> > Hi Jens Jørgen,
> > 
> > That's what I did (on a standard ArchLinux distribution, nothing
> > fancy,
> > really). I get this:
> > 
> >    ----------------------------------------------------------------
> > ----------
> >    Open MPI has detected that this process has attempted to
> > initialize
> >    MPI (via MPI_INIT or MPI_INIT_THREAD) more than once.  This is
> >    erroneous.
> >    ----------------------------------------------------------------
> > ----------
> >    [hostname:25791] *** An error occurred in MPI_Init
> >    [hostname:25791] *** reported by process [2821259265,0]
> >    [hostname:25791] *** on a NULL communicator
> >    [hostname:25791] *** Unknown error
> >    [hostname:25791] *** MPI_ERRORS_ARE_FATAL (processes in this
> > communicator will now abort,
> >    [hostname:25791] ***    and potentially your MPI job)
> > 
> > Other MPI programs on that computer are working fine and a (very)
> > simple mpi4py test works. For the records, the version of openmpi
> > is
> > 3.1.0.
> > 
> > I'll have a look at the mpi.c sources and activate whatever macro
> > can
> > help.
> > 
> > Gaël
> > 
> > > 
> > > Jens Jørgen
> 
> Success (apparently) \O/
> 
> Race conditions that exist with python (multiple instances are
> spawned
> by mpiexec) but not in gpaw-python (since it initialises MPI right
> away). Once solution is to use MPI_Initialized().
> 
> I'll submit a patch tomorrow though I am not sure that this is
> enough.
> 
> Gaël
> 

I've created a new merge request:

https://gitlab.com/gpaw/gpaw/merge_requests/403

This is working with OpenMPI but I don't know for other vendors.

Best regards,
Gaël