[gridengine users] Implementing a credit system in a cluster

Reuti reuti at staff.uni-marburg.de
Fri Jan 27 15:08:32 UTC 2012


Hi,

Am 27.01.2012 um 15:01 schrieb Martin Gumbau:

> Hi,
> I don't known if it is possible and the best way to make it (if was possible):
> SCENARIO: 
> - 2 Cells (cell-A and cell-B)
> - Cell-A have 3 queues (q1-A, q2-A,q3-A)
> - Cell-B have 2 queues (q1-B, q2-B)
> ONLY IN CELL cell-A
> My pretension is to have  a credit system in that at any users will be assigned 'n' credits in one cluster-cell. 
>    - A job in a queue q1-A will cost 'x' credits
>    - A job in a queue q2-A will cost 'y' credits
>    - A job in a queue q3-A will cost 'z' credits
> where x < y < z 
> This means that q1-A is the best and sending a job in this queue cost more credits than sending jobs to other queues.
> Each time one job is sent to a queue, a number of credits is substracted of the 'n' 'total of user credits.
> RESULT:
> one user could have less jobs running in the best queue and more jobs running in the worst queue.

do you want to subtract the credit from a budget which needs to be refilled, or do you want to limit what is running at a time in the cluster only? The latter could be done:

1 - one consumable integer complex, which is defined in `qconf -me global` with an arbitrary high value, as it's only necessary to have something to subtract the credits from

2 - you don't need 3 queues, it would be more SGE like to have 3 boolean complexes, which the user could request for the type of job, but it can also be done with queues if you prefer it

3 - for each requested queue (or complex) you request unconditionally the credit for the job (i.e. the complex you define in step 1) by a JSV (i.e. the JSV has to check which queue/complex was requested and add a hard request for the complex), this way the user can't fool it - the credits are always requested for the job

4 - in an RQS you limit the overall credits per user, i.e. how many credits are active at a time; it can be limited by individual user/group or just all the same for all users

-- Reuti




More information about the users mailing list