[torqueusers] Torque configuration for single node -single job
Coyle, James J [ITACD]
jjc at iastate.edu
Thu Apr 29 15:35:19 MDT 2010
Here are two suggestions for the case of pbs_sched (Maui and MOAB may have more sophisticated mechanisms)
The first is for any user, the second can only be implemented by the admin, and does not do exactly what you want,
but is automatic.
1) For user:
I am going to assume that all the node are of type cluster (not time-shared) you c an check this with the command:
pbsnodes -a | grep ntype
all lines should look like:
ntype = cluster
Assuming that the file /var/spool/torque/server_priv/nodes has lines like:
where they are all 4, then andy user can get nodes to himself/herself just by reserving full node, e.g. for the
above ppn=4 submit with
Even if you only use one or two processors. Wasteful, yes, but it works,.
You have reserved the entire node so no other jobs can run on this node. (You will likely get charged for 4 nodes also if charging is done.)
If you do not have access to /var/spool/torque/server_priv/nodes , then issue:
pbsnodes -a | grep np
and hopefully you see the same number after np =
np = 4
as for the case above. (This is not a typo, it is np= in pbsnodes -a and ppn= in the nodes file.)
2) For manager
If you are a manager for the cluster, you can issue
qmgr -c 'set server node_pack = False'
This will attempt to always start a new job on an empty node, so if there are free node, the jobs will spread out.
This will not prevent jobs from sharing a node, but will delay it.
I don't use this nor recommend it unless you are running the cluster like a farm, that is all the jobs are single processor
Jobs and you want to spread the load as much as possible.
If you are trying to run multi-processor jobs, it is best to pack them so that there are lots of fuly free nodes.
James Coyle, PhD
High Performance Computing Group
115 Durham Center
Iowa State Univ.
Ames, Iowa 50011 web: http://www.public.iastate.edu/~jjc
From: torqueusers-bounces at supercluster.org [mailto:torqueusers-bounces at supercluster.org] On Behalf Of alap pandya
Sent: Thursday, April 29, 2010 11:46 AM
To: torqueusers at supercluster.org
Subject: [torqueusers] Torque configuration for single node -single job
How can we avoid node sharing by multiple job in torque (i.e. we do not want multiple jobs to run on same node at same time). Please let me know what all configuration changes are required and how to do them.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the torqueusers