[Mpi-forum] Discussion points from the MPI-<next> discussion today

Jeff Hammond jhammond at alcf.anl.gov
Sat Sep 22 08:02:54 CDT 2012

> I haven't investigated implementations in any depth recently, but recent
> quick looks confirms that quite a few MPI features are NOT available in
> all of the major ones.  I know that is true for heterogeneous systems,
> and the more extreme threading options, but there are quite a few other
> areas where I should be flabberghasted if it were not.  And, of course,
> at least some of the almost-unused features don't actually work quite
> right, but don't get fixed because nobody gives a damn ....

MPICH2 supports essentially everything except heterogeneous (in the
32/64-bit sense, for example; GPUs are a different issue) as far as I
know.  I don't know anyone running such a system anyways, since all
modern Linux systems are 64-bit.  Mixing endianness is even more rare;
I don't know anyone that cares about this scenario.

Microsoft MPI, MVAPICH2, Cray MPI, Intel MPI and IBM Blue Gene MPI are
all based upon MPICH2 so they should have similar feature support.
Cray and IBM Blue Gene do not support dynamic processes yet.  Blue
Gene will never support them in a nontrivial way, although I know how
to make the implementation standard compliant to the letter of the
standard (it's not worth doing because no one cares).

Can you list some of the "quite a few" features that are not available
in MPICH2 and OpenMPI, which I contend are the "major ones"?

It's somewhat frustrating that you write to this list with statements
about all the things about MPI that don't work but never actually
follow through and enumerate them in such a way that someone like me
could determine the veracity of your statements.

In the future, can you please provide such lists in the first message
rather than going around and around without ever posting them.  You
still haven't posted a specific instance of C99-MPI incompatibility
that one can verify...


Jeff Hammond
Argonne Leadership Computing Facility
University of Chicago Computation Institute
jhammond at alcf.anl.gov / (630) 252-5381

More information about the mpi-forum mailing list