dolfin team mailing list archive

[Johan Hake <661342@xxxxxxxxxxxxxxxxxx>]

On Friday October 15 2010 15:49:58 Garth Wells wrote:
> On 15/10/10 19:28, Johan Hake wrote:
> > On Friday October 15 2010 10:35:31 Garth Wells wrote:
> >> Yes.
> >> 
> >> I think that the problem is with which library is initialising and
> >> finalising MPI.
> > 
> > I do not see that the import sequence makes any difference. If you have
> > created a PETSc object _and_ you load PyTrilinos in what ever order, I
> > get the error when ening the Python session.
> 
> That's half the story. When not creating a PETSc object, the order is
> important because the PyTrilinos module initialises MPI when imported.
> DOLFIN checks first if MPI has been intialised.

Ok.

> > Why would you like to use both PyTrilinos and a PETSc object?
> 
> I did it at first by accident. The point is that it shouldn't crash. 

Agree.

> I
> sent an email to the PyTrilinos developers. The problem seems to be that
> PyTrilinos does not check if another part of a program has taken
> responsibility for MPI.

Ok.

Johan

> Garth
> 
> > Also I know to little about the initialization/finalization of MPI to
> > suggest any solution. But I guess that when the PETScVEctor is created,
> > MPI is initialized by DOLFIN. DOLFIN then also finalize MPI which might
> > interfere with PyTrilinos doing the same?
> > 
> > Johan

-- 
Importing PyTrilinos can cause MPI crash
https://bugs.launchpad.net/bugs/661342
You received this bug notification because you are a member of DOLFIN
Team, which is subscribed to DOLFIN.

Status in DOLFIN: New

Bug description:
The import order appears to be very important with PyTrlinos, and can cause some MPI problems. 
In the below code, order doesn't help. The program just crashes.

from PyTrilinos import Epetra
from dolfin import *
parameters["linear_algebra_backend"] = "Epetra"
xx = PETScVector(20)