Am 30.01.2013 um 23:30 schrieb Joe S:
> I tried limiting the slot usage all in one rule:
>
> limit queues lower.q hosts {@computenodes} to slots=$num_proc
You were speaking of three queues in total, so I thought of:
limit queues lower_one.q,lower_two.q hosts {@computenodes} to slots=$num_proc
-- Reuti
> then submitted some jobs with a hard resource request for lower.q. The
> scheduler ended up running both jobs on a host that was already at its slot
> limit (with jobs in a different queue).
>
>
> From: Reuti <[email protected]>
> To: Joe S <[email protected]>
> Cc: "[email protected]" <[email protected]>
> Sent: Wednesday, January 30, 2013 2:22 PM
> Subject: Re: [gridengine users] Slot limits and subrodinate queues
>
> Am 30.01.2013 um 23:19 schrieb Joe S:
>
>>
>> I am using RQS to limit slots per host. And I tried adding the queues to
>> the limit but it didn't seem to take as you mentioned.
>
> In a separate RQS or all rules in one (as the first matching will be used
> only)? For us the effect was that no jobs were scheduled at all, hence
> scheduling was blocked.
>
> -- Reuti
>
>
>>
>>
>> From: Reuti <[email protected]>
>> To: Joe S <[email protected]>
>> Cc: "[email protected]" <[email protected]>
>> Sent: Wednesday, January 30, 2013 2:01 PM
>> Subject: Re: [gridengine users] Slot limits and subrodinate queues
>>
>> Am 30.01.2013 um 22:31 schrieb Joe S:
>>
>>> Sorry if this has been asked before, but I couldn't find a answer in the
>>> mail archives. We currently have two queues setup. The first only a subset
>>> of users have access to and its configured to use all hosts/slots in our
>>> cluster. The 2nd queue is limited to only 25% of the cluster using a RQS
>>> slot limit on the queue. We then have per host slot limits to prevent over
>>> subscription since the queues
>>
>> Also as RQS or in the exechost definition?
>>
>>> share the same set of hosts. This works nicely in keeping the "other"
>>> class citizens from using too much of our resources. But since some jobs
>>> are "urgent" we want to add a third queue, "urgent"and have the other two
>>> queues be subordinate. That way users can submit a job and request a
>>> urgent resource to have their jobs run. It seems though that subordination
>>> won't work in this case since the jobs have to actually get scheduled and
>>> that won't happen because of the slot limit being in effect. Is there any
>>> way to do this?
>>
>> Your conclusion is correct. The suspension is the result of the scheduling.
>>
>> If the slot count per host is limited by an RQS, it could be set up to honor
>> only certain queues instead of all like it's working if it's limited in the
>> exechost definition. Unfortunately I found RQS sometimes not working in this
>> scenario and no jobs got scheduled any longer at all.
>>
>> -- Reuti
>>
>
>
>
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
