Hi,

Am 15.03.2013 um 16:32 schrieb Gowtham:

> I'm trying to experiment with the share tree policy on our test clusters 
> running Rocks 6.1 (CentOS 6.3) and Grid Engine 2011.11p1. I'd like this 
> policy to treat users as if they were all *not* created equally. To that 
> effect, I have done the following so far:
> 
> 
>  #1. SGE configuration (via qconf -mconf)
> 
>      http://sgowtham.net/downloads/rocks61_ge201111p1_qconf_sconf.txt
> 
> 
>  #2. Scheduler configuration (via qconf -msconf)
> 
>      http://sgowtham.net/downloads/rocks61_ge201111p1_qconf_ssconf.txt
> 
> 
>  #3. The cluster has 4 users: amy, bob, john, karen.
>      For each one of these users, I have added a user object
>      via the command
> 
>      qconf -Auser USERNAME.txt
> 
>      and the contents of USERNAME.txt look like
> 
>      name USERNAME
>      oticket 0
>      fshare 0
>      delete_time 0
>      default_project NONE
> 
> 
>  #4. Suppose that the 'share' of each one of these users is
>      as follows:
> 
>      amy     4000  40%
>      bob     1000  10%
>      john    1000  10%
>      karen   4000  40%
>      TOTAL  10000
> 
>      and is included in a file
> 
>      http://sgowtham.net/downloads/rocks61_ge201111p1_share_tree_policy.txt
> 
>      Then, I ran
> 
>      qconf -Mstree rocks61_ge201111p1_share_tree_policy.txt
> 
>      to implement the Share Tree Policy.
> 
> 
> I hope my work flow in Steps #1 - #4 are correct.
> 
> 
> With that, I do have the following questions:
> 
>  #1. Does the above set up *guarantee* that 'amy' will have
>      access to 40% of the cluster's resources, 'bob' will
>      have access to 10% of the cluster's resources, over a
>      period of time ('halftime'?)

It can't guarantee this. Suppose only 'amy' is using the cluster during these 
28 days. She got 100% then. Even if the other users step in on the very last 
day of the 28 days period: they can't catch up to get these distribution of 
resources which was intended (controlled by the "compensation_factor" in the 
scheduler configuration to catch up with the targeted resource distribution). 
There would be a machanism necessary to block 'amy' if her 40% were used up.


>  #2. Suppose that this policy has been in place 28 days
>      (equal to the 'halftime' parameter in scheduler configuration).
> 
> 
>  #3. For some inexplicable reason (e.g., bad behavior, not
>      paying the $$, etc.), suppose that bob's new share is 0.
>      That is,
> 
>      amy     4500  45%
>      bob        0   0%
>      john    1000  10%
>      karen   4500  45%
>      TOTAL  10000
> 
>      So, I update the rocks61_ge201111p1_share_tree_policy.txt
>      with these new shares and run
> 
>      qconf -Mstree rocks61_ge201111p1_share_tree_policy.txt
> 
>      to implement the new Share Tree Policy.
> 
> 
>  #4. If #1 is true, will the new 'share' layout mean that any newer
>      jobs submitted by 'bob' will just wait in 'qw' mode forever?
> 
>      If yes, awesome.

The priority is a result of many settings (`man sge_priority`). If you setup 
the overall priority to honor only the share tree (by adjusting the several 
"weight_*" entries), it will work as long as there are jobs in the cluster. But 
as soon as only jobs of the bad user are left, they will start.


>      If not, how do I make sure that it happens - without really
>      removing bob's account from the cluster OR making him part of
>      a 'bad dudes' users list that is not allowed to run anything
>      in any queue/host via RQS?

a) xuser_lists in SGE's configuration.

b) a JSV to block these user's jobs

-- Reuti


> Thanks for your time and help.
> 
> 
> Best regards,
> g
> 
> --
> Gowtham, PhD
> Information Technology Services
> Adj. Assistant Professor, Physics
> Michigan Technological University
> 
> (906) 487/3593
> http://it.mtu.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