[GE users] group resource quota

brs brs at usf.edu
Wed Nov 19 15:22:28 GMT 2008


Despite RQS not being able to set enforceable resource limits, its still 
a pretty sweet solution since, given my hodgepodge of different compute 
resources, I can dynamically set a value, such as h_vmem, as a 
percentage of total physical memory.  And since we make our users 
specify h_vmem in their jobs (if they go over 256M), it works as a 
pretty acceptable solution (granted they don't under-advertise their 
memory usage).  If we do get support for adjusting resource limits via 
RQS, that would be excellent. This has been informative as usual.  Thanks!

-Brian

andreas wrote:
> On Wed, 19 Nov 2008, reuti wrote:
>
>   
>> Am 19.11.2008 um 11:27 schrieb andreas:
>>
>>     
>>> On Tue, 18 Nov 2008, reuti wrote:
>>>
>>>       
>>>> Am 18.11.2008 um 17:48 schrieb andreas:
>>>>
>>>>         
>>>>> On Tue, 18 Nov 2008, brs wrote:
>>>>>
>>>>>           
>>>>>> Hi, Reuti,
>>>>>>
>>>>>> How about another example:
>>>>>>
>>>>>> Say I want to limit vmem usage for all users on any host to, say,
>>>>>> 95% of
>>>>>> the physical memory available (not exactly how I would set it
>>>>>> up, but
>>>>>> you get the idea).  Could I accomplish this like so?
>>>>>>
>>>>>> {
>>>>>>   name         vmem_limit
>>>>>>   description  "Limit users to only 95% of total physical memory
>>>>>> to avoid \
>>>>>>   swaping"
>>>>>>   enabled      FALSE
>>>>>>   limit        hosts {*} to h_vmem=$mem_total*0.95
>>>>>> }
>>>>>>
>>>>>> And how would this limit manifest itself?  Would the first job that
>>>>>> tried to allocate more than the quota allows bomb on a malloc() or
>>>>>> something?  Is this even possible within the rqs or should I
>>>>>> just use
>>>>>> limits.conf?
>>>>>>             
>>>>> This does not (yet) work. Reason is that resource quota limits are
>>>>> not enforced as resource limits. If they were malloc()
>>>>> would return NULL once the quota is exceeded.
>>>>>           
>>>> And the reason is not, that it's only working (the formula) for INT
>>>> and DOUBLE?
>>>>         
>>> No. It's because resource limitations that would stem from RQS are
>>> simply disregarded with the 'math exercise' that is done right before
>>> job delivery from qmaster to execd.
>>>       
>> Then I would suggest to add this to the documentation. It states only
>> DOUBLE and INT as allowed complexes, but not MEMORY (man
>> sge_resource_quota).
>>     
>
> Actually the point is not about DOUBLE/INT vs. MEMORY complexes. It's about
> resource quotas not being enforced as resource limits as one would expect
> based on the section queue_conf(5) where the impact of the queue resource 
> limit is explained.
>
> Either way you are right that man pages should reflect this:
>
>     http://gridengine.sunsource.net/issues/show_bug.cgi?id=2793
>
> Regards,
> Andreas
>
> Sitz der Gesellschaft: Sun Microsystems GmbH, Sonnenallee 1, D-85551 Kirchheim-Heimstetten
> Amtsgericht Muenchen: HRB 161028
> Geschaeftsfuehrer: Thomas Schroeder, Wolfgang Engels, Dr. Roland Boemer
> Vorsitzender des Aufsichtsrates: Martin Haering
>
> ------------------------------------------------------
> http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=89113
>
> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
>   


-- 
Brian Smith
HPC Systems Administrator
Research Computing, University of South Florida
4202 E. Fowler Ave. LIB618
Office Phone: +1 813 974-1467
Organization URL: http://rc.usf.edu

------------------------------------------------------
http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=89119

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



More information about the gridengine-users mailing list