Am 15.10.2013 um 08:39 schrieb Txema Heredia:

> El 10/10/13 15:17, Reuti escribió:
>> Am 09.10.2013 um 11:22 schrieb Txema Heredia:
>> 
>>> Hi all,
>>> 
>>> Is it possible to define in any way max_reservations per user basis?
>>> 
>>> I have some users with limited access to the cluster that have their slots 
>>> limited with an RQS, but they are still able to submit an arbitrary number 
>>> of jobs using -R y. This locks slots well beyond their RQS-defined limit 
>>> and leads to some slot bottlenecks as almost no one is using h_rt and 
>>> backfilling cannot kick-in.
>> I would say it's a "feature" in the scheduler when it reserves slots for a 
>> job while the job these slots were reserved for can't be started at all. But 
>> if it wouldn't do so, you could face job starvation on a user level (instead 
>> of a global level) where always serial jobs of the same user slip in.
>> 
>> I.e.: if the reservation would only start once enough slots are available 
>> again, then as soon as maybe a serial job slips in and the slot count drops 
>> again under the required ones and the reservation would stop again.
>> 
>> Looks like a strategy to get it implemented in a nice way regarding other 
>> users' jobs is not easy. Maybe a "soft" and "hard" reservation could work: 
>> the RQS count for slots is the soft limit for the reservation and before any 
>> serial job is started (used + already reserved) slots (or any other 
>> resource) must fall below this value, and (used + reserved) itself may only 
>> go up to another set amount like a "hard" limit.
>> 
>> -- Reuti
> 
> Thus would you reccomend to build some scripts to control this from "outside" 
> of sge?
> Maybe something that checks if the user has any job running and then check 
> how many waiting jobs have -R y, compare that against a user-threshold and 
> set the excess to -R n? With some mechanisms to set back -R y once more jobs 
> start working?

Yes, either this or put the job on hold.


> Is there any "easy" way to check a user's slot limit without having to check 
> every single RQS?

I fear this needs to be hardcoded in your custom scheduler.


> And to check how many slots has currently reserved?

No. It's necessary to parse the monitor file for reservations.

-- Reuti


> Thanks,
> 
> Txema
>>> Thanks in advance,
>>> 
>>> Txema
>>> 
>>> 
>>> PS: Obviously, I don't trust them enough to grant them access to Advanced 
>>> Reservations.
>>> _______________________________________________
>>> users mailing list
>>> [email protected]
>>> https://gridengine.org/mailman/listinfo/users
> 


_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to