On 11 October 2011 23:33, Gerald Ragghianti <[email protected]> wrote:
>
>> Like the OP mentioned, one could use a consumable complex for 6.1. If you 
>> add "complex_values network=16" to the queue, and "load_thresholds 
>> network=15" it will be pushed to alarm state automatically and you can avoid 
>> the load sensor. When you add a default consumption of 1, it works 
>> out-of-the-box (it's only subtracted if it's attached to a queue).
>>
>> I.e. the other queue for normal jobs don't have it attached, and you select 
>> the special multi-node queue by the requested PE.
> Unfortunately, I think there are two problems with this suggestion.
>
> 1. If I set network=16, then only 16 processors out of 48 will be usable
> by parallel jobs.
>
> 2. The use of a load threshold seems to prevent fill_up from working
> correctly, so even if I have network=48 for the queue complex and
> network=47 for the load threshold it will not use up all 48 slots before
> moving on to the next host.  This seems to be due to the alarm state
> becoming active on the queues at inconsistent times during a single
> scheduling iteration.  This would also affect the use of a custom load
> sensor, so I'm abandoning that idea.
>
> If we were to update to 6.2u5, what options would we then have?
My suggestion of a queue with an exclusive resource should work with 6.2u5.

Assuming that each multi node job running on a node consumes 1 context
then you should be able to generalise
the solution by adding one such multi-node/exclusive queue per context
plus a single queue for serial/single node PEs.
>
> --
> Gerald Ragghianti
>
> Office of Information Technology - High Performance Computing
> Newton HPC Program http://newton.utk.edu/
> The University of Tennessee, 2309 Kingston Pike, Knoxville, TN 37919
> Phone: 865-974-2448
>
> /-------------------------------------\
> | One Contact       OIT: 865-974-9900 |
> | Many Solutions         help.utk.edu |
> \-------------------------------------/
>
>
> _______________________________________________
> 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