On Jul 22, 2005, at 8:18 PM, Lei_at_ICS wrote:
> Thanks a lot, Jeff! What you said all make perfect sense.
>
> One concern my client has, is that whether using LAM universe
> and published names to connect two MPI processes is
> portable or not to MPICH. Some of their clusters only provide
> MPICH but not LAM MPI (sad but true).
MPICH-1 does not have MPI-2 dynamic processes. I understand that
MPICH-2 does, but I have no direct experience to know. Hypothetically,
it should "just work" because what I suggested is fairly standard MPI-2
dynamic process usage (according to the standard). But I wouldn't
guarantee this (heck, they *are* a competitor, after all ;-) ) --
someone should give it a whirl.
> I know this is not a MPICH user group, but is the concept of a universe
> a standard MPI feature or LAM specific? Is the design you provided
> portable to all MPI implementations? If not, can you provide the
> necessary modifications please?
The "universe" that I've been talking about is fairly LAM specific --
it's our term for an established run-time environment based on the LAM
daemons. Other MPI-2 implementations have daemon-based environments as
well. As such, it could work in the same (or similar) way as LAM's
universe concept. The real question is how wide the scope of their
MPI-2 published names is -- the concept I outlined for you depends on
the ability to be able to lookup published names from within Matlab.
If that doesn't work in another MPI implementation due to limitations
of the scope of published names, then the scheme won't work.
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
