[petsc-dev] buildsystem package dependencies

[Satish Balay]

On Tue, 5 Apr 2011, Lisandro Dalcin wrote:

> On 5 April 2011 13:29, Matthew Knepley <knepley at gmail.com> wrote:
> > On Tue, Apr 5, 2011 at 11:20 AM, Satish Balay <balay at mcs.anl.gov> wrote:
> >>
> >> Not sure I understand. If I say 'mpi depends on mpe' - then the link line
> >> is:
> >>
> >> -lhypre -lmpi -lmpe
> >>
> >> But thats incorrect usage - and not what I want.. The correct usage is:
> >> -lhypre -lmpe -lmpi
> >>
> >> The issue comes up with -llmpe - which traps MPI_xx functions - and does
> >> some
> >> logging - and then calls MPI_xx functions from -lmpi. i.e usage is:
> >>
> >> -lhypre -llmpe -lmpe -lmpi
> >
> > Right, but this MPE package is changing the MPI interface. All those other
> > packages,
> > like Hypre, depend on MPI. MPI depends on MPE, but the MPE module may modify
> > the MPI lib to be '-lmpe -lmpi'. I do not see another way to do this.
> >    Matt
> >
> 
> We could also build a custom libmpe.so, load it at runtime from
> petscmpiexec by passing a -mpe flag. This way, we can get MPE logging
> for free, using a regular PETSc build. This is more or less described
> at the end of this ticket:
> https://trac.mcs.anl.gov/projects/mpich2/ticket/573.
> 
> What do you think?

Hardly anyone uses petscmpiexec.

Satish