Re: [HTCondor-users] jobs to start on busy machines first

[Rolf Seuster <seuster@xxxxxxx>]

Hi all,

just a little follow up, in case other face similar problems.
There's still a confusion about the sign of the Cpu term, at lest on my 
side ;)

My config that works for me has this:
# condor_config_val -negotiator NEGOTIATOR_PRE_JOB_RANK
(10000000 * MY.Rank) + (1000000 * (RemoteOwner =?= UNDEFINED)) - (- 
100000 * MY.Cpus) - MY.Memory

# condor_config_val -negotiator NEGOTIATOR_POST_JOB_RANK
(RemoteOwner =?= UNDEFINED) * (ifthenElse(isUndefined(KFlops), 1000, 
Kflops) - SlotID - 1.0e10*(Offline=?=True))

With this, one machine gets filled first. I'll check this a little bit 
more and
report back anything I find out.

Thanks, Rolf


On 2018-08-07 12:38 p.m., Todd Tannenbaum wrote:
> On 8/7/2018 2:26 PM, Rolf Seuster wrote:
> > Hi Todd,
> >
> > > So I had a typo in the recipe which may make a difference depending on
> > > how you are submitting the jobs. The clause "(10000 * Cpus)" was
> > > supposed to be "(-10000 * Cpus)"... I changed the wiki recipe.
> >
> > Thanks, looks like that is now filling one machine at a time !
>
> Yay! Glad I could help.
>
> > > Do your execute nodes have config knob CONSUMPTION_POLICY=True (this
> > > is not the default) ?
> >
> > No, it seems our config doesn't touch this variable. Reading up on it 
> > now.
>
> My advice is to keep it that way... i.e. unless you need to do so, do 
> not specify CONSUMPTION_POLICY, just use the default....
>
> regards,
> Todd