On Apr 4, 2006, at 10:01 AM, imran shaik wrote:
> I want to run around 200 MPI processes. I have two p4,256MB nodes
> connected via cross cable. I have installed redhat9 on it. Now i am
> able to run 90 simple "hello world" processes. If i exceed it says,
> no of clients for lamboot exceeded. It says 72 is the limit. But i
> get 90, dont know how.Anyways i need to execute more processes
> (around 200 ) on these nodes i have. Can any one suggest any way to
> do that?
>
> Also can anyone say whats the problem i am getting. Is it any
> unusual thing ?
You are running out of internal resources in our process control
daemons. The limits are fairly fixed size, somewhere around 70 or so
(I'm not sure how it launched 90, but it's possible we bumped up the
limits a bit somewhere along the line). LAM/MPI really wasn't
designed to try to start hundreds of processes on a single node, as
it was more designed for clusters with small numbers of CPUs / node
and oversubscribing nodes by orders of magnitude is generally
considered bad form in the HPC community. I'm afraid this isn't
something with an easy work-around.
You may get better results trying Open MPI (the successor to LAM/
MPI), although you might not. Resources are going to get really
constrained - you're looking at less than 1.5MB memory / process,
which isn't going to make most modern apps very happy.
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
