[GE users] SGE and MPICH2

Reuti reuti at staff.uni-marburg.de
Mon Jun 18 16:40:59 BST 2007


Am 18.06.2007 um 16:55 schrieb Maxime Kinet:

> I'm currently setting up a new cluster using SGE6.0 and running  
> Fedora Core 6.
> I have some troubles with the integration of MPICH2. I found the  
> usefull how-to by Reuti "Tight Integration of the mpich2 library  
> into SGE", and decided to use smpd as a process manager for mpich.  
> I succeeded in making everyting work, but I need to start an smpd  
> service on each node and for each user, which is a bit annoying.

it must be this way, to have complete control of the daemons by SGE  
and correct accounting. Otherwise there is no way to guaranty, that a  
qdel will really remove the job on all nodes.

> More precisely, the only way I founded is by adding the following  
> lines in the launching script :
> #Variables and parameters definition
> ...
> #starting process manager
> smpd -s
> #running the esecutable a.out
> mpiexec -np 4 ./a.out
> #shutting down the process manager.
> smpd -shutdown

This shouldn't be with my setup. What is necessary, is to specify a  
port for the mpiexec to use.

If you use "smpd -s", this will run only locally and out of the  
process tree (smpd -s will fork into daemon-land AFAIR).

To have it more transparent for the user, OpenMPI might be an option.

-- Reuti

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