dolfin team mailing list archive

["Johannes Ring" <johannr@xxxxxxxxx>]

On Mon, June 2, 2008 Nuno David Lopes wrote:
> Finally i could install it:
> _________________________________________________________________________
> petsc:
> $./config/configure.py --with-cc=gcc --with-fc=gfortran\\
> --download-umfpack=1 --download-hypre=1 --with-mpi-dir=/usr/lib/mpich
> --with-shared=1
> dolfin:
> $scons enablePetsc=1 --withPetscDir=$PETSC _ DIR --enableMpi=1

Note that we don't use dashes when specifying options to scons. The above
line should therefore be like this:

$scons enablePetsc=1 withPetscDir=$PETSC_DIR enableMpi=1

> the --download-mpich=1 in PETSc didn't work because scons::DOLFIN
> still didn't find the MPICH stuff.

Then the problem is probably that we are unable to locate a MPI C++
compiler on your system. As I said earlier, we only look for mpicxx,
mpic++, and mpiCC. Using --download-mpich=1 when configuring PETSc, you
will only get mpicc which is normally just a MPI C compiler. As far as I
can remember, the mpicc compiler you get should also work as a C++
compiler. You can try this by defining the CXX environment variable,
giving the complete path to the mpicc compiler that was downloaded and
installed by PETSc. An alternative is to configure PETSc with the
--with-clanguage=C++ option. This should produce a MPI C++ compiler when
using --download-mpich=1.

> And i've really change the PATHs to the
> downloaded mpich dir.
> Probably it should be good to have a scons option "withMpiDir= " like for
> the
> other packages.
>
> I've also tried with lam or openmpi...
>
> But still I can't see the PETSc parallel system solving in my codes?

Did you start your codes with mpirun or mpiexec? If you just do, e.g.,

  ./app

when starting your application, you won't be running in parallel. Try with

  mpirun -np 2 ./app

or

  mpiexec -np 2 ./app

This should run two copies of your program in parallel.

Johannes

> (I really don't know much about parallel algorithms...so sorry if i'm
> saying
> something that doesn't make sense)
> but what i get when i run top at the time the program is solving a system
> is
> still a
> --------------------
> Cpu0 : 0.0%
> Cpu1 :100.%
> -------------------
> it seems that petsc isn't allocating the two cores at the same time...so
> it
> isn't really parallel, and there aren't really any advantage on it.
> Is this a know issue? Like an mpich bug or something..
> Should i try again with opemmpi like was suggested?
>
> Thanks
>
>
>
> On Monday 02 June 2008, Jed Brown wrote:
>> On Mon 2008-06-02 08:29, Johannes Ring wrote:
>> > On Fri, May 30, 2008 Nuno David Lopes wrote:
>> > > Ok, i have mpich on my system, petsc (without hypre)  compiled
>> > > with --download-mpich=1, it all works, my code compiles and runs
>> > > etc.... Everything is running with the exception that MPI isn't
>> > > working...
>> > >
>> > > ..................scons: Reading SConscript files ...
>> > > Using options from scons/options.cache
>> > > MPI not found (might not work if PETSc uses MPI).
>> > > .......................................................................
>> > >.
>> >
>> > The reason you get this warning is that we are unable to locate mpirun
>> or
>> > an MPI C++ compiler (we look for mpic++, mpicxx, and mpiCC) on your
>> > system. This might be a bit limiting, so I added mpiexec and orterun
>> as
>> > alternatives for mpirun. Please try again with the latest from the hg
>> > repository.
>>
>> When PETSc installs mpich, it doesn't automatically put it in your path
>> so
>> you will have to specify the path explicitly when you configure Dolfin.
>>
>> In terms of Dolfin configuration, it's not really a good idea to just
>> use
>> any random MPI implementation that may be in your path.  You can get the
>> correct implementation from PETSc (CC,CXX,MPIEXEC in
>> $PETSC_DIR/bmake/$PETSC_ARCH/petscconf).  This would be especially nice
>> if
>> you have multiple PETSc builds with different MPI implementations (i.e.
>> OpenMPI with a debugging build for development/testing and vendor MPI
>> with
>> optimized build for production runs).
>>
>> Jed
>
>
>
> --
> Nuno David Lopes
>
> e-mail:ndl@xxxxxxxxxxxxxx        (FCUL/CMAF)
>            nlopes@xxxxxxxxxxxxxxx    (ISEL)
> http://ptmat.ptmat.fc.ul.pt/%7Endl/
> _______________________________________________
> DOLFIN-dev mailing list
> DOLFIN-dev@xxxxxxxxxx
> http://www.fenics.org/mailman/listinfo/dolfin-dev
>