Offhand I do not know the answer to your question, but perhaps  
creating three additional partitions (one for each user) that overlap  
the nodes in "sade1" would solve the problem.

Quoting Guglielmi Matteo <[email protected]>:

> Hello everyone.
>
> Here is something i find difficult to debug:
>
> 15 identical nodes "sade[01-15]" are part of a partition
> called "sade1"
>
> 3 users were granted to run jobs on partition "sade1"
>
> the 3 users have decided to "manually subdivided" sade1
> among themselves using the "--exclude=nodelist" option:
>
> user1: #SBATCH --exclude=sade[06-15]
>
> user2: #SBATCH --exclude=sade[01-05,11-15]
>
> user2: #SBATCH --exclude=sade[01-10]
>
> 2 qos were defined:
>   "qos1" adds 5000 pts to the total priority count (default)
>   "qos2" adds 10000 pts to the total priority count
>
> total job's priority includes also a bit of job aging coeff:
>   PriorityWeightAge=1000 vs PriorityWeightQOS=10000
>
> Problem:
>
> if all the three users do use the default qos (qos1) there
> is always one user having all of his jobs in pending state
> with reason "priority".
>
> if the stuck user goes for "qos2", jobs do enter immediately.
>
> Please note that the --exclude options do make sure to have
> three disjoint "sub partitions".
>
>
> How can I find if this is a bug or some error in my slurm.conf?
>
> I'm also using slurmdbd + cgroups... I'll post anything useful if needed.
>

Reply via email to