Interesting idea, but...

Let's consider a farm of 24 slots with 5 users trying to get 8 slots each:

limit users !A to slots=16
limit users !B to slots=16
limit users !C to slots=16

That means D and/or E could submit a job and get up to 16 slots.  The
rules would have to be sequentially created as follows:

limit users !A to slots=16
limit users !A and !B to slots=8
limit users !A and !B and !C to slots=0

And then of course when A finishes before B you'd have to rejig all
the rules so I don't think that'd work in practise.

The second part though, where you grab a whole machine... that might work...

limit users !A hosts box1 to slots=0
limit users !B hosts box2 to slots=0
limit users !C hosts box3 to slots=0

and then remove once the job's done.

I guess that'd work.  I was just hoping for something a little less hacky :-)
Thanks for the help Reuti,

Stephen



On Mon, Sep 26, 2011 at 10:34 AM, Reuti <[email protected]> wrote:
> Am 24.09.2011 um 14:49 schrieb Stephen Willey:
>
>>>> An example:
>>>>
>>>> 10 machines with X slots each
>>>> 15 users (a,b,c,d,...o) submit array jobs with 1000 tasks, each of
>>>> which requires X slots.
>>>
>>> X is not the same X like above? Or are you using always machines exclusive
>>> per array task?
>>
>> In this particular case, the machines are being used exclusively but
>> that was really just to give a simple example.  It's definitely not
>> always the case.
>
> I could think of the following implementation, but it needs an external cron 
> job, checking the actual jobs therein. When it discovers that user A has an 
> array job running using 8 slots out of 216 in the cluster for each instance, 
> it will create a new RQS which reads:
>
> limit users !A to slots=208
>
> As the first rule in an RQS grants access, you will need one RQS per user 
> (not many lines in one and the same RQS). It could also be extended to 
> include the particular host, i.e. host Z having 12 cores:
>
> limit users !A hosts Z to slots=4
>
> Unfortunately there is no "jobs" rule implemented, which could implement it 
> even better.
>
> -- Reuti



-- 
Stephen

http://lensframephoto.com

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

Reply via email to