[Beowulf] hpl size problems

[Greg Lindahl]

On Tue, Oct 04, 2005 at 09:01:32AM -0400, Joe Landman wrote:

> There are some mpich versions out there that let you dynamically link at 
> run time (this is *good*) such as the Scali.

At least 6 commercial MPIs do that: Scali, Scyld, HP, Intel, ParTec, MPI Pro.

Free MPIs that do that: LAM, and eventually OpenMPI. And, of course,
OpenIB allows this to be done among hardware that supports OpenIB.

This plethora of versions is part of the argument for an MPI ABI. BTW,
the status of this effort is that there is a lot of interest from MPI
interconnect vendors and from ISVs, but no neutral person has stepped
up to run the effort. I don't count since I work for an interconnect
vendor. I was hoping one of the MPI committee heavyweights was going
to volunteer, but I don't seem to have convinced them that it's
worthwhile. But I do hear that it was discussed at the Euro PVM/MPI
meeting, so people are thinking about it.

ISVs are working around this problem in a couple of ways. Some provide
.o files so you can relink the app as long as mpi.h is compatible
enough. Some provide some wrapper functions so that you can rebuild
a single .o file with your mpi.h and then relink.

I was very pleased that Mathworks' forthcoming MPI port of Matlab
is going to be relinkable right from the start.

-- greg