[torqueusers] scheduling cycle
Ronny T. Lampert
telecaadmin at gmail.com
Fri Aug 3 02:00:58 MDT 2007
> So there is no job start or job finish trigger. Scheduler tries to schedule
> in each scheduling cycle. Am I right?
First, what kind of scheduler are you using, pbs_sched or maui/moab?
Don't forget that anything can change at any time! I frequently down
nodes for maintainance. I frequently re-add nodes or change their settings.
Also consider that when a lot of jobs finish, every job finish would
mean that you would have to contact the scheduler and update it. This
works for few objects (nodes, jobs, ...) but not for big setups.
The scheduler should not re-schedule at any job exit or job entrance,
but only update the already made scheduling decisions when there is
really a need (e.g. a new high-priority job). Usually it is quite rare
that a new job is to be run NOW. Also there is usually a scheduling
window, e.g. only a certain amount of objects is considered at once, to
limit complexity and runtime.
Torque was meant to be scalable, and in my setting it has proven that
quite well (using it since the 1.0 days :)
> Then how do we decide the scheduling interval?
Am I right that you are trying to optimize your settings/configuration?
Considering that I have only 14 Nodes, but sometimes up to 50.000 jobs
in the queue I find a scheduling intervall for maui of 45 seconds to be OK.
Also take into account that when you have a lot of short jobs (I
sometimes do, sometimes even only 20sec runtime per job) a shorter poll
intervall helps improving the utilization of your compute nodes.
More information about the torqueusers