openmpi versus mpich2

[Jason Swails]

On Mon, Aug 8, 2011 at 2:19 PM, Rodolfo Aramayo <raramayo at gmail.com> wrote:

> The developer of the package I want to run (Velvet) told me I should
> just install 'libgomp'
>
> Can you guys translate this for me?
>
> Is he talking about openmpi or mpich2?? or something else??
>

Something else.  libgomp is the GNU implementation of the OpenMP
parallelization scheme.  Note that OpenMPI and mpich2 are both message
passing interface (MPI) implementations designed for shared and distributed
memory systems (oversimplified; multi-core computers and many multi-core
computers networked together) by passing messages between the different
threads.  OpenMP is a shared memory parallelization scheme that has
different syntax and design philosophy to MPI.

OpenMP is (from what I've heard) simpler to write and implement than MPI,
yet it is limited to shared-memory machines and thus cannot scale as high as
MPI.

However, if you're just trying to compile code that someone else has already
written, you need to know which one they used to parallelize their code and
link with those libraries.  If the developer of the package you're trying to
use says to install libgomp, then just do that (or some other OpenMP
implementation if it's available) and leave MPI alone ;).

Hope this helps,
Jason


>
> Thanks
>
> On Mon, Aug 8, 2011 at 09:21, Eric A. Borisch <eborisch at macports.org>
> wrote:
> > On Mon, Aug 8, 2011 at 8:57 AM, Jason Swails <jason.swails at gmail.com>
> wrote:
> >> I personally use mpich2 myself, but tend to compile my own MPIs, as the
> >> multiplicity can get you in trouble (if you link serial libraries to MPI
> >> code, you can run into lots of problems if the compiler version used to
> >> build the MPI libraries, and therefore the ones you're using to build
> the
> >> MPI code if you use mpicc/mpif90, is not the same as the compiler
> version
> >> you used to build the serial libraries).
> >> HTH,
> >> Jason
> >
> > FWIW, The mpich2 port supports a number of underlying compilers
> > through the variants... Hopefully your favorite in the list.
> >  -Eric
> > --
> > Eric A. Borisch
> >
>



-- 
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.macosforge.org/pipermail/macports-users/attachments/20110808/0911aaa9/attachment.html>