[GE users] subordinate queues and MPI jobs

Bogdan Costescu bogdan.costescu at iwr.uni-heidelberg.de
Wed May 5 16:39:37 BST 2004

On Wed, 5 May 2004, Sean Dilda wrote:

> Has anyone here played around with subordinate queues and MPI jobs?

In principle suspending MPI jobs should _not_ work. The reason is the
communications between different nodes which most likely have some
timeouts to decide when one node has gone down. This is not something
that can be easily modified by an external agent (like SGE in this
case). So, the operation of suspending itself might work, but upon
resuming the job will abort with communication timeouts - the
suspension is in this case not better than killing the job altogether.  
Suspending MPI jobs might in fact not be desirable at all, as some
resources are not freed (say, GM ports for a Myrinet card) that will
deny any chance of another MPI job running on the same node.

However, if you add checkpointing in the mix, things might work. I 
know that LAM-MPI supports BLCR (on Linux as it needs a kernel patch):


The case might be different for MPI jobs that do not span several 
nodes (running as several processes on a SMP node). This should work, 
as the communication protocol in this case is based on direct access 
to memory and should not include any timeouts.

Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu at IWR.Uni-Heidelberg.De

To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net

More information about the gridengine-users mailing list