[GE users] User-Fairshare-Policy

Daniel Templeton Dan.Templeton at Sun.COM
Fri Sep 7 06:19:32 BST 2007


    [ The following text is in the "ISO-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

Graham,

Answers inline.

Daniel

Graham Jenkins wrote:
> In an attempt to curb the enthusiasm of those users who have been running several long-running
> MPI jobs on multiple-nodes (thereby tying up the entire cluster), we've implemented a
> "fair-share" policy thus:
>
> 1. Make 2 changes in the main SGE configuration ('qconf -mconf'):
>        * enforce_user auto
>        * auto_user_fshare 100
>
> 2. Make 1 change in the SGE scheduler configuration ('qconf -msconf'):
>        * weight_tickets_functional 10000
>
> And I'm wondering:
>
> a) What's the significance of the '100' and '10000' values above?
>   

The 100 is the number of tickets assigned to every user.  Since you're 
assigning every user the same number of tickets and you're not using 
project, department, or job tickets, the number is completely 
irrelevant.  It will only matter if you assign different users different 
amounts of tickets, or if you start using the other ticket categories 
(projects, departments, jobs).

The 10000 is the number of functional tickets as compared to share tree 
and override tickets.  Since you're not using share tree or override 
tickets right now, the number is irrelevant.

> b) Does the policy take into account a user's recent resource-usage history?
>   

No.  You're using the functional ticket policy, which is non-historic.  
If you want usage history taken into account, you'll need to use the 
share tree policy.

> c) If so, how far back does it go?  and
>   

0.  If you were using the share tree policy, how far back it goes is 
governed by the halftime and halflife_decay_list values in the scheduler 
configuration.

> d) Does it use the 'accounting' file for this? and
>   

No, not even if you were using the share tree policy.  The scheduler 
keeps it's own usage record that gets spooled by the qmaster.

> e) What happens when the 'accounting' file gets rolled and compressed?
>   

Nothing in particular.

> f) How can we tell if it's all working?
>   

qstat -ext

> Regards ..
>   

---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net




More information about the gridengine-users mailing list