[GE users] Allowing user jobs to take over entire nodes in a mixed cluster node environment

Reuti reuti at staff.uni-marburg.de
Fri May 11 10:44:17 BST 2007


Am 11.05.2007 um 10:57 schrieb Yogesh Bhanu:

>       I am using Grid Engine 6.0u8 on my opteron cluster.
>  I have a cluster queue which is a good mix of Dual Opterons (2  
> CPUs) and
> Dual Core opterons (4 CPUs). I have made host groups with slots=2  
> and slots
> =4 and combined them in one cluster queue .
> Recently we are getting jobs which are threaded and I would like to  
> bind
> these jobs to one node .
> I have looked into Seans solutions of specifying a consumable,  
> which works
> quite well If I have a only one type of nodes.
> My question
> How do I acheive same in a dynamic way in a mixed cluster node  
> environment.

to get complete nodes in SGE, it's always best to request a PE with  
the exact number of CPUs and allocation_rule=$pe_slots in a uniform  
cluster. This conforms to the internal bookkeeping of used slots and  
can easily be scanned later on in the accounting file.

If you have a mixed cluster, you need just two PEs with a slightly  
different setting:

for the 2-CPU nodes:

pe_name smp2
allocation_rule 2

for the 4-CPU nodes:

pe_name smp4
allocation_rule 4

and bind them to one queue:

pe_list [@nodes2cpus=smp2],[@nodes4cpus=smp4]

And if you don't care, on which node your job will start (your  
jobscript will just start 2 oder 4 threads depending on the node  
where it is scheduled to), you could request:

qsub -pe smp* 2- myjob.sh

In any case, you will get a complete node.

HTH - 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