[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [HTCondor-users] RANK question; slots unclaimed on a single computer






On Fri, Aug 8, 2014 at 12:02 PM, Todd Tannenbaum <tannenba@xxxxxxxxxxx> wrote:

+ Your start _expression_ on your execute machine could only allow Slot1 to be claimed during the night and/or weekends, when interactive use is unlikely, and/or your preempt _expression_ could kick off the job on Slot1 when keyboard activity is detected.

+ You could permanently reserve one core for interactive use by the owner with
  NUM_CPUS = $(DETECTED_CPUS) - 1

Yes, but then I'd unnecessarily lose 10-15 slots during the day which are not being used for whatever reason.

I've turned off preemption and use SUSPEND exclusively, based on keyboard/console and non-HTC load. However you gave me an idea. By allowing preemption after say 10 minutes of suspension, the job will start on another slot. Since these are not long jobs, starting over is not a great penalty, and better than remaining suspended for an hour or two.

BTW I just tried this and it works well for my query as to assigning rank:

Rank = kflops * SlotID

It's simple, jobs go first to the faster machines with more slots, and will fill up the last slot of a machine last.

Thanks,
Ralph Finch
Calif. DWR