Re: [HTCondor-users] scheduling behavior

[Marcus Ebert <mebert@xxxxxxx>]

Hi Todd,

After running with this setting for 2 days, it seems jobs are now evenly 
distributed between VMs instead of filling up one first before sending 
jobs to another one.
Is there any other option that needs to be set so that 
NEGOTIATOR_DEPTH_FIRST has an effect?

Cheers,
 Marcus


On Wed, 2 May 2018, Marcus Ebert wrote:

> Hi Todd,
>
> On Wed, 2 May 2018, Todd Tannenbaum wrote:
>
> >  On 5/2/2018 3:53 PM, Marcus Ebert wrote:
> > >  Hello,
> > > 
> > >  We have a condor 8.6 installation with about 4000 jobs slots. Most of
> > >  these are on 8-core VMs. When there are free slots available, the
> > >  scheduling of idle jobs to a specific machine seems somehow random.
> > >  What we would like to have is that HTCondor fills up all slots on a
> > >  machine before sending jobs to another (empty) machine.
> > > 
> > >  Is there any config option that would allow to change the scheduling
> > >  behavior in that way (something like a node allocation policy in Maui)?
> > > 
> >
> >  Hi Marcus,
> >
> >  Are you using static slots (the default) or partitionable slots on your
> >  pool?
> >
> >  If you are using static slots, the default configuration should already
> >  prefer filling up a machine before using another machine.
> >
> >  If you are using partitionable slots, and you are running HTCondor v8.6.x
> >  or later, I suggest you append the following into the
> >  condor_config[.local] on your central manager and then do a
> >  condor_reconfig:
> >
> >     # If using partitionable slots, fill depth first.  See
> >     http://tinyurl.com/y75k3k7p
> >     NEGOTIATOR_DEPTH_FIRST = True
> Yes, we are using partitionable slots (1 partitionable slot per VM).
> I'll try that setting and let you know how it works.
> Thanks!
>
> Cheers,
>  Marcus