[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [Condor-users] questions about job , user priority and machine rank
- Date: Thu, 17 Feb 2005 11:44:53 -0500
- From: "Ian Chesal" <ICHESAL@xxxxxxxxxx>
- Subject: RE: [Condor-users] questions about job , user priority and machine rank
I can't say anything about condor-g because I'm not a condor-g user. But
I can throw my $0.02 in on some of your other questions...
> Other than that, I am a little bit confuse about Ranking and
> priority, it seems that the higher rank (I mean the number),
> the machine should be chosen. For higher priority (I mean the
> number also), the jobs should not be chosen. Should be
> reverse to that of rank, right?
You have RANK correct: higher is better. For priority you have to be
careful. There are two priorities in condor: the user-specified job
priority and the condor-calculated user priority. For the job priority,
set by the user when they submit their cluster, higher is better. Jobs
with higher job priority on a schedd will run before jobs with lower job
priority on the same schedd. The condor-calculated user priority is a
running count of how many machines the user is currently using. The
lower the number the better if they user is trying to get more machines.
The negotiator sorts users during the negotiation cycles from lowest
user priority to highest and the lowest priority user gets more
resources when compared to higher priority users.
> One more questions about retirementtime, I am not so
> understand what it means, it means the amount of time the
> jobs can be run without preempted?
Retirement time and preemption work together. If a job (and/or a
machine) defines a retirement time for running jobs then when another
job wants to preempt a running job on that machine a count down timer is
started and the running job is given the amount of time specificed in
the retirement time to complete what ever it is doing before normal
preemption (checkpointing if it's a standard universe job, then hard
killing and sending the job back to it's queue) occurs. If the running
job can finish in the time alloted by retirement time then the machine
is handed over to the preempting job without any additional fuss.