[gridengine users] Setting priority on a per-user basis

[Reuti]

Am 02.07.2012 um 03:03 schrieb Gowtham:

> 
> Thank you.
> 
> I started reading more about setting up a
> share tree policy with portions of the pie
> as % of granted use:
> 
> First I did:
> 
>  qconf -mconf
> 
>    enforce_user      auto
>    auto_user_fshare  100
> 
>  qconf -msconf
> 
>    weight_tickets_functional  10000
> 
> 
> Then, using Qmon, I did:
> 
>  qmon 
>  --> Policy Configuration 
>  --> Share Tree Policy
>  --> Add node
>  --> Add Leaf (repeatedly)
>  --> I enter a user's name & the share
> 
> When I press 'Apply', I get the error message
> 
>  denied: share tree contains reference to unknown 
>  user/project "john"

All users need to be added first, as these objects will also store the past usage. If they are created automatically, they will be removed at some time ("auto_user_delete_time" in SGE's configuration).

$ qconf -auser


> This happens for all users except for amy. When I
> type
> 
>  qconf -suserl
> 
> the output is 'amy'. 

Maybe she is just running a job right now and the entry was therefore automatically added.


> I'm pretty sure I'm making several things incorrectly
> here. Could you please guide me to a suitable reference
> material (or man page) with sequence of commands that
> I can use to set up this share tree policy? 

http://arc.liv.ac.uk/SGE/howto/geee.html

http://docs.oracle.com/cd/E19080-01/n1.grid.eng6/817-5677/i1004072/index.html

-- Reuti


> 
> Also, it'd be nice to work from command line, 
> without having to use Qmon. Then the script that
> generates these % per user can do so in a format
> that's readily acceptable by a qconf -Mxxx command.
> 
> Thank you for your time and help.
> 
> Best,
> g
> 
> --
> Gowtham
> Information Technology Services
> Michigan Technological University
> 
> (906) 487/3593
> http://www.it.mtu.edu/
> 
> 
> On Sun, 1 Jul 2012, Reuti wrote:
> 
> | Am 01.07.2012 um 21:28 schrieb Gowtham:
> | 
> | > 
> | > Suppose that I have a list of all users in a cluster
> | > with their respective priorities calculated using
> | > some random criteria (good behavior, consistent
> | > usage of resources, etc.):
> | > 
> | >  amy     20.00
> | >  john    15.00
> | >  karen   25.00
> | >  laura   30.00
> | >  tom     10.00
> | > -----------------
> | >  total  100.00
> | > -----------------
> | 
> | You can setup a share tree policy with these percentages of granted usage.
> | 
> | 
> | > The numbers next to the usernames indicate the
> | > colloquial 'portion of the pie' they are entitled
> | > to.
> | > 
> | > How would I go about implementing this priority
> | > scale in SGE? Also, it'd be nice if scale was
> | > applied only to two queues - long.q and short.q -
> | > whose nodes are funded by the University.
> | 
> | This is not possible in SGE. The defined policy is always for the complete cluster. Maybe you could create a second account for these users?
> | 
> | 
> | > That is to say, if amy decided to submit jobs
> | > to a queue spanning her contributed nodes, amy.q,
> | > she would have the highest priority.
> | 
> | Even with a share tree policy, she is only entitled to use 20%  by the `halftime` decreasing past usage (in the scheduler configuration). So even on her own machines she might having the last jobs being scheduled as she is well above 20% already right now.
> | 
> | What could help (but would need her asistance):
> | 
> | - make her an "operator" in SGE
> | - define two queues on her machines: one common.q, one for her amy.q
> | - add amy to the "owner_list" in common.q for her machines
> | - as she is owner, she can drain the common.q on her machines
> | 
> | As you have now more than one queue per machine, the overall slot limit needs to be limited on these machines.
> | 
> | -- Reuti