[petsc-dev] Multiple MPICH/fblaslapack Installs For Multiple Arches

[Matthew Knepley]

On Sun, Mar 22, 2020 at 5:46 PM Jacob Faibussowitsch <jacob.fai at gmail.com>
wrote:

> Wrt MPICH - My suggestion is to install with --download-mpich --prefix [as
> mentioned]. Primary reason is: we use --with-device=ch3:sock - which is
> good for 'oversubscribe' usage [running "mpiexec -n 8 ./exe" on a dual core
> box] - which is what we normally do during development. And also its
> valgrind clean.
>
>
> Prefix is where the source is downloaded to? Or where the libs/binaries
> are put?
>

Its where the install goes to (libraries and binaries).

  Thanks,

    Matt


> Ideally I would like my source to be in /my/custom/path and the resulting
> binaries and libraries to be in /usr/bin or /usr/local/bin where all my
> other downloaded packages are.
>
> Best regards,
>
> Jacob Faibussowitsch
> (Jacob Fai - booss - oh - vitch)
> Cell: (312) 694-3391
>
> On Mar 22, 2020, at 4:37 PM, Satish Balay <balay at mcs.anl.gov> wrote:
>
> On Sun, 22 Mar 2020, Jacob Faibussowitsch wrote:
>
> Yes on all points regarding MPI and BLAS/Lapack.  I recommend installing
> a current MPICH and/or Open MPI system-wide, preferably hooked up to
> ccache (see replies to this thread:
> https://lists.mcs.anl.gov/pipermail/petsc-dev/2020-January/025505.html <
> https://lists.mcs.anl.gov/pipermail/petsc-dev/2020-January/025505.html>),
> as well as BLAS/Lapack system-wide.  It's the other packages that are
> more likely to depend on int/scalar configuration, but even many of
> those (HDF5, SuiteSparse, etc.) aren't built specially for PETSc.
>
>
> Is the home-brew MPICH, openblas, lapack sufficient here? Or is it
> recommended to build all three from source?
>
>
> On Mac - configure defaults to using VecLib [Apple's default blas/alapck].
>
> Wrt MPICH - My suggestion is to install with --download-mpich --prefix [as
> mentioned]. Primary reason is: we use --with-device=ch3:sock - which is
> good for 'oversubscribe' usage [running "mpiexec -n 8 ./exe" on a dual core
> box] - which is what we normally do during development. And also its
> valgrind clean.
>
> And then we recommend xcode gcc/g++ with brew/gfortran - so MPICH can be
> built with this combination of compilers [using this mode]
>
> Satish
>
>
> Best regards,
>
> Jacob Faibussowitsch
> (Jacob Fai - booss - oh - vitch)
> Cell: (312) 694-3391
>
> On Mar 22, 2020, at 4:25 PM, Jed Brown <jed at jedbrown.org> wrote:
>
> Jacob Faibussowitsch <jacob.fai at gmail.com <mailto:jacob.fai at gmail.com
> <jacob.fai at gmail.com>>> writes:
>
> Hello all,
>
> As part of development, I have several arch folders lying around in my
> PETSC_DIR namely a 32-bit OSX, 64-bit OSX, 32-bit linux with valgrind,
> 64-bit linux with valgrind, and a 32-bit up to date with current master.
> All of these have a —download-mpich —download-fblaslapack and hence their
> own copy of each (so that’s 5 copies of each, plus other duplicated
> packages im sure). At this stage, even getting the bare minimum of these
> arches ready for dev work after a rebase/git pull takes decades as package
> versions, or conf settings change, forcing a rebuild of the same packages
> multiple times.
>
> My question(s):
> What petsc ./configure options are necessary to change the
> configuration of each library w.r.t. petsc? i.e. can my 64-bit arches
> use my 32-bit MPICH/fblaslapack and vice-versa?
>
>
> Yes on all points regarding MPI and BLAS/Lapack.  I recommend installing
> a current MPICH and/or Open MPI system-wide, preferably hooked up to
> ccache (see replies to this thread:
> https://lists.mcs.anl.gov/pipermail/petsc-dev/2020-January/025505.html <
> https://lists.mcs.anl.gov/pipermail/petsc-dev/2020-January/025505.html>),
> as well as BLAS/Lapack system-wide.  It's the other packages that are
> more likely to depend on int/scalar configuration, but even many of
> those (HDF5, SuiteSparse, etc.) aren't built specially for PETSc.
>
> Does this change when I have —with-debug on or off? If so, what other
> packages have a similar ability? Is there anywhere in ./configure
> —help where this kind of information would be documented?
>
> I suspect that this hasn’t been fully explored since its primarily a
> developer “problem” and not one the average user will run into/care about
> (since they usually aren’t building petsc multiple times). I’m sure
> everyone has their own ways of tackling this problem, I’d love to hear them.
>
> Best regards,
>
> Jacob Faibussowitsch
> (Jacob Fai - booss - oh - vitch)
> Cell: (312) 694-3391
>
>
>
>
>

-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20200322/d1041aac/attachment.html>