On Thu, 28 Mar 2002, Senthil Kandasamy wrote:

> I fixed it. On the master node, the default shell was csh whereas on the
> slaves, it was bash. I changed the default shell to csh on all three
> nodes and recon worked. Lamboot also worked. I do not understand why
> this shoudl affect the boot, but hey....it worked. .I am not
> complaining. However, when I lamboot , it recognizes only one cpu per
> node though the nodes have dual processors. Is this a problem?

I'm not sure why the shell change would make a difference.  We (the LAM 
team) haven't really tested that configuration, so it looks like something 
to test for the next OSCAR release.  We don't do anything that should 
cause problems, but obviousl something is :).

AS for the CPU count, LAM doesn't actually detect how many CPUs you have, 
it requires you to tell it.  The only purpose in LAM's CPU count is 
scheduling - if you tell it all your nodes are dual CPU machines, you can 
do something like:

  mpirun C foo

and have LAM run two processes on each node.  More importantly, LAM will
guarantee that processes on the same machine will be "neighbor" ranks in
MPI_COMM_WORLD - ranks 0 & 1 might be on the first node, for instance.  
Since most communication occurs with "neighbors", this can actually make a
performance difference on some applications.

You might want to take a look at the LAM/MPI FAQ, available at:

 http://www.lam-mpi.org/faq/

for more information on the topic.  There is a detailed explination of the 
CPU count stuff under the "Booting LAM" section.

> I just submitted a job through command-line mpirun and it seems to work.
> 
> Now I have to see if I can submit jobs through pbs. If that works, the
> cluster wil be fully functional.

Hopefully, you have made it through most of the sticky points.  Getting 
LAM or MPICH to run under PBS isn't all that difficult.  Especially since 
PBS should already be running with a basic FIFO queue enabled.


Brian

-- 
  Brian Barrett
  LAM/MPI developer and all around nice guy
  Have a LAM/MPI day: http://www.lam-mpi.org/


_______________________________________________
Oscar-users mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/oscar-users

Reply via email to