Hi,

> Am 09.12.2015 um 16:56 schrieb Jerome Poitout <[email protected]>:
> 
> Hi all,
> I have a small compute farm on which users are fighting to get some cpu
> time. I have almost 650 CPU (grid view) and one user can take all the
> CPUs available for his jobs. I've tried to set share policies, but
> whatever I set, I can't handle to get another result than FIFO when
> users are submitting jobs.
> I have not been very far inside configuration and share policies, but is
> there a way to get 2 teams to take 1/4 the CPUs and 1 team to take half
> of the CPUs when the node are charged, and if only one team is working,
> then it can take all the CPUs?
> 
> I have tried to use the reprioritize feature in the cluster
> configuration, but the result is that the nodes are  used 70% for nice
> and renice processes... waste of time

As long as there are less active processes than cores in a machine, the nice 
value has no effect. There is simply no one to be nice to.

The reprioritization was mostly used in cases where you want to oversubscribe a 
machine and fine tune the functional policy. As nowadays it's often set up that 
number of cores = slot count, it's used less often.


> My goal is to have a fair use of the cpu resources, based on cpu time
> consumption, so each team will get a ratio of cpu usage and incoming
> jobs will be submitted with different priorities regarding how many
> resources were already consumed. And, I also want one team to be much
> prioritized than others...

Do you want to honor the running jobs only (functional policy) or also the past 
usage over a moving timeframe of the last 30 days or alike (share tree policy)?

-- Reuti


> I'm trying be be clear but I'm not sure to be...
> Another is that I read most of grid engine documentation, but did not
> understand a lot about share policies (tree or tickets) as I'm not
> english native speaker... Is there some documentation I could find in
> french?
> 
> Thanks for your support,
> 
> Jérôme
> _______________________________________________
> 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