Re: [Condor-users] Specifying Max Job Run Time

[Burt Holzman <burt@xxxxxxxx>]

Terrence Martin wrote:
> I do not suppose there is a recipe out there for restricting how long a 
> job runs while at the same time making sure that jobs are not 
> interrupted prior to that run time and for the long run jobs only to be 
> kicked if there are other jobs waiting for the queue? Say a value of 72 
> hours, that is both the min and max runtime for jobs but if the jobs are 
> the only one on the cluster that they can just keep running.
>
> There are a lot of settings it seems, from PREEMPTION_REQUIREMENTS to 
> MaxJobRetirementTime to PREEMPT_LATENCY. It is just not all that clear 
> to me how to go about getting all these settings to do what I want as 
> far as putting an upper limit on jobs after which they can not be 
> guaranteed to run, while at the same time not kicking off jobs that may 
> be running long for legitimate reasons on an otherwise underused cluster.

Terrence,

It sounds like you just need jobs to preempt quickly but with a long 
retirement time.  We have this at the CMS Tier 1 at FNAL. 
MaxJobRetirementTime is set to 48 hours.  PREEMPTION_REQUIREMENTS is 
essentially set to

(CurrentTime - EnteredCurrentState) > (10*60)

In this configuration, jobs get preempted after 10 minutes of running if 
jobs are waiting in the queue; but the preemption does not evict the 
running jobs until it hits 48 hours.

- B