[Mauiusers] soft and hard limits question

Gareth.Williams at csiro.au Gareth.Williams at csiro.au
Mon Jul 5 16:05:10 MDT 2010


Hi Arnau,

> -----Original Message-----
> From: Arnau Bria [mailto:arnaubria at pic.es]
> Sent: Monday, 5 July 2010 8:08 PM
> To: mauiusers at supercluster.org
> Subject: Re: [Mauiusers] soft and hard limits question
> 
> On Thu, 01 Jul 2010 11:42:09 -0400
> Chris Hunter wrote:
> 
> Hi Chris,
> 
> > We are also experimenting with hard/soft limits. Some suggestions:
> > - most of you cluster nodes are busy so jobs might be memory-bound &
> > not processor-bound; try MAXPE instead of MAXPROC
> 
> Mmmm... I don't remeber if most of them were busy. I remeber some free
> nodes. It started with ~170 jobs,  and went down till 30, and some
> other jobs where kept as blocked.
> 
> I'll try to repeat tests in test server.
> 
> 
> > - try using hard/soft limits for USERCFG instead of GROUPCFG (ie.
> > create a testuser)
> Wow, we'd like to user this feature with ~300 users ...

You can use a single DEFAULY setting to match all users that you don't have a more specific match for eg. we have for a ~300 proc cluster (in moab):
USERCFG[DEFAULT] MAXPROC=100,150 MAXIJOB=10 MAXIPS=8640000

'Special' users can get custom settings.

GROUPCFG limits apply to the aggregate usage of the group rather than to each of the members which can be awkward as one rogue user can stop the whole group from getting work done.  On the other hand, then it is the groups problem to discipline the offender.  We have not found GROUP limits useful but can see how they might be deemed politically necessary by some managers.

Cheers,

Gareth

> > - can you paste the output of "diagnose -g" & "diagnose -u" ?
> if I'm able to repeat the situation on test, i'll do it.
> 
> > Good Luck!
> >
> > Chris Hunter
> > Yale HPC
> > chris.hunter at yale.edu
> 
> Thanks for your reply Chris,
> Cheers,
> Arnau



More information about the mauiusers mailing list