On Nov 3, 2005, at 5:05 PM, Sebastian Forsman wrote:
Are there any "hard coded" limits in the size of a Open MPI pool?
Which "pool" are you referring to? The number of nodes, size of
memory, etc.?
How about the maximum amount of nodes running a single job?
Open MPI jobs have been run on a few thousand nodes (2k, I believe?) on
Lawrence Livermore machines. We've still got some scalability issues
to work out after 1.0, but this first release series should be able to
handle jobs that large. Future versions will handle more.
Which factors should be
taken in to consideration when building large clusters? For an example
what
kind of limits does the OS set, etc.
Most commercial parallel computing environments seem to have a
restriction
of 100-256 nodes. Our project requires more. Any solutions? It's not
always
possible to "split" the data into x starting points and use a batch
system
to launch x jobs on several independent pools and then wait for the
results
and combine them.
You might want to bounce these questions over to the Beowulf list --
they talk much more about clusters, hardware, and specific setup over
there (dig through there list archives; this kind of stuff is covered
frequently).
--
{+} Jeff Squyres
{+} The Open MPI Project
{+} http://www.open-mpi.org/
