Hi Eygene,
> > If you want to drain the site for downtime, then you should stop the
> > queues in plenty of time beforehand, i.e. don`t start jobs. It is not
> > sufficient to say you are closed, lock the door. Torque qstop(or qdisable)
> > will stop new job starts and submissions and this has some effect on the
> > info, but it doesn`t rely on the info to block submissions.
>
> Yes, and the usual sequence is as the following (at least it is what
> I use).
>
> a) Shut down globus-gatekeeper. As I understand (and used to verify),
> this will prevent new jobs to come via Globus pathway, but already
> running jobs won't be harmed.
No! Running jobs generally _will_ be harmed! An RB/WMS/Condor-G will not
be able to clean up finished jobs, so the jobs will be considered failed.
Furthermore, after 1 hour an RB/WMS/Condor-G will no longer be able to
monitor the active jobs, since it cannot relaunch grid_monitors on the CE.
|