On Jan 27, 2009, at 3:36 PM, Spenser Gilliland wrote:
> We currently have a set of lam-mpi programs we need to get running
> and we haven't spent much time researching OpenMPI.
Fair enough.
> Would our lam-mpi programs be compatible with OpenMPI?
Source compatible, yes. They should "just work" with Open MPI, but
sometimes there are subtle differences in run-time behavior. If
shortest-time-to-runnning is the issue, then stick with LAM/MPI. But
be aware that no one is working on LAM any more; all future work is
occurring on Open MPI. It would be good to spend some cycles after
you meet your deadline to investigate switching to Open MPI.
> And if so would it make the task of cross-compiling for the nodes
> easier? BTW: I'm headint to the openMPI website right now.
I honestly don't remember the cross-compile behavior in LAM/MPI. I
know that it is possible in Open MPI if you specify a cache file so
that some of configure's tests are just given the answer (vs. trying
to compile and then run a program to determine a test result). I know
that some Cray owners compile OMPI in this way (head node is one
architecture, back-end nodes are a different architecture).
I think the Right way to do it in Open MPI is to specify
--with-platform=<your_platform_filename>
Where <your_platform_filename> is a file containing the cache value
answers to several of OMPI's tests. See https://svn.open-mpi.org/trac/ompi/browser/trunk/contrib/platform/redstorm
for an example. Most of the tests that you need to provide answers
for are Fortran datatype size and alignment tests, IIRC. But anything
with "cv" in it is something you need to provide manually so that
configure doesn't have to figure it out.
--
Jeff Squyres
Cisco Systems
|
