Re: [Condor-users] Preemption question

[Bruce Beckles <mbb10@xxxxxxxxx>]

On Fri, 30 Jun 2006, Miskell, Craig wrote:

<snip>
> I, too, thought I had disabled pre-emption with:
> PREEMPT                 = FALSE
>
> Lo and behold, some jobs got pre-empted last night (the first time my
> pool has had any real contention for resources).  From my reading of the
> previous thread, this can potentially happen when PREEMPT is false, but
> only if the machine has a RANK expression which results in it preferring
> a different job to the one currently running.  If that understanding is
> correct, then something weird is going on, as all my startds have a Rank
> of 0 (I've done nothing special).

To *completely* disable preemption you also need to set:

PREEMPTION_REQUIREMENTS = FALSE

on the *CENTRAL MANAGER* (strictly speaking, on the machine running the 
condor_negotioator daemon).  If you don't do this, you can still get 
"preemption by users with better priority in the pool".


> Interestingly, in this case the preferred/pre-empting job had a Rank =
> Mips, whereas the pre-empted/evicted job had a Rank of 0.  This may be a
> red-herring, but does seem suspiciously related.  Should the job Rank
> have any effect on preemption (I would have thought it wouldn't even be
> considered until the Negotiator had decided that a given Startd was a
> candidate for satisfying the claim).

If you didn't have PREEMPTION_REQUIREMENTS set to FALSE on your central 
manager, you might want to check to see what the priority of the user 
submitting the preferred/pre-empting job was and compare it with the 
priority of the user who had submitted the pre-empted/evicted job.

But an additional thought - and I'm speculating here - it would make sense 
to me if it turned out to be the case that the negotiator might look at a 
startd that was already running a job if it was trying to match a 
different job that had been submitted by the _same_ user.  I.e. when 
trying to match jobs to startds for user A, the negotiator might say, 
something like "I'll consider idle startds and then startds claimed by 
user A when trying to match this job".

That's pure speculation on my part - I may be completely wrong.

	-- Bruce

--
Bruce Beckles,
e-Science Specialist,
University of Cambridge Computing Service.