dolfin team mailing list archive

[Johan Jansson <johanjan@xxxxxxxxxxxxxxxx>]

On Fri, Sep 02, 2005 at 06:34:22PM -0400, Faheem Mitha wrote:
> 
> 
> On Fri, 2 Sep 2005, Anders Logg wrote:
> 
> >On Fri, Sep 02, 2005 at 12:24:39PM -0400, Faheem Mitha wrote:
> >>
> >>
> >>On Fri, 2 Sep 2005, Anders Logg wrote:
> >>
> >>>I tried a couple of things:
> >>>
> >>>1. Installing both mpich and lam4-dev:
> >>>
> >>>Works fine for me, no problems and /etc/alternatives/mpi points to
> >>>/usr/include/lam, but maybe I'm missing the point?
> >>
> >>I had both mpich-dev and lam4-dev installed. I think the problem is with
> >>the alternatives system, yes.
> >
> >Maybe this has been fixed in Debian, since I didn't have a problem?
> 
> Unlikely. I'm using Sarge, which released some time ago, and I'm certain 
> that I had this problems since is released. It is possible that is was 
> fixed as part of a security release, but I didn't notice anything 
> involving mpi.
> 
> I can check, though.
> 
> Johan, can you comment, please?

Hi!

I still get link errors when compiling a DOLFIN demo if I both have
mpich and lam4 installed, and lam4 active in
/etc/alternatives/mpi. What happens is that PETSc (and DOLFIN in turn)
links directly against the libraries in /usr/lib/mpich/, while DOLFIN
includes the headers in /usr/include/mpi, which is controlled by the
alternatives system. So when lam4 becomes active, DOLFIN is including
the lam4 headers and linking against the mpich libraries, which
produces errors.

_However_, if I reconfigure and rebuild PETSc after making lam4 active
(I didn't do this the last time we discussed this), it chooses lam4
instead and everything compiles ok. I still can't run the demos,
apparently because lamd isn't running, but that's a lam-specific
problem and nothing we have to worry about.

I think we should consider this issue resolved and closed now. If both
PETSc and DOLFIN are reconfigured and rebuilt, then they use the
actual mpi implementation. The expected situation is that both PETsc
and DOLFIN will have been installed as Debian packages, and then they
should have been configured to use the alternatives system properly by
the package maintainers (I don't think the DOLFIN package has to do
anything wrt. this). If PETSc is installed manually, then problems can
appear due to the alternatives system, but then you're basically on
your own I guess.

So to summarize:

Reconfigure and reinstall PETSc and DOLFIN if you switch around
libraries using the alternatives system.

The PETSc package (as well as the DOLFIN package, but I think it
already does) should be configured to use the alternatives system for
mpi (preferrably for all libraries they link against).

  Johan