[GE users] "Nodes" as a consumable resource

reuti reuti at staff.uni-marburg.de
Fri Nov 19 13:24:41 GMT 2010

Am 19.11.2010 um 00:40 schrieb jagladden:

> We have a cluster composed entirely of eight core nodes.  We currently 
> have a resource quota that looks like this:
> {
>    name max_slot_per_user
>    description "resource quota for restricting execution slots"
>    enabled TRUE
>    limit users{@core_userset} to slots=48
>    limit users{@noncore_userset} to slots=24
> }
> This rule restricts any given user in the "core_userset" userset to 
> using a maximum of 48 processor slots and user in the "noncore_userset" 
> to 24 processor slots.  However, there is some dissatisfaction with the 
> way this works in practice.  Most of the jobs on the system are  eight 
> processor shared memory parallel jobs that fit neatly on one node.  
> Mixed in with these is a scattering of single processor jobs.  A single 
> processor job running on any given nodes effectively makes it 
> unavailable for use by a parallel job.

The was the discussion in the beginning to have besides "yes", "no" and "job" also the entries "host" (one time per job per used host) and "hostonce" (one time per job per host, but only one time). The latter was meant to support licenses which are floating and per node. Hence one job of one type on an exechost and other jobs of this type can run there for free.

In your case each user would get his own "license" which he would have to request or attached automatically by a JSV. But it's not there anyway.

> What we would like to do is create a resource quota that restricts the 
> total number of nodes that a users jobs can be running on.  Thus instead 
> of 48 slots, they would be restricted to running jobs on a maximum of 6 
> nodes, irrespective of how many slots they actually use on each node.  
> However, "nodes" does not appear to be one of the built in resource 
> complexes to which I can assign a limit.
> Is there a way to do this?  Can I somehow define a "nodes" complex that 
> I can then use in a resource quota?

What about a cron-job, which will adjust the xuser_list used lists per exechost? I.e. the cron-job checks, whether a user is using already 6 nodes. In this case the user is excluded from all other machines (well, means one dedicated ACL per exechost or each user has an ACL for himself to be attached to certain machines) - but he is free to run additional serial jobs on the machines he is already using.

-- Reuti


To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].

More information about the gridengine-users mailing list