On Tue, 8 Apr 2003, Steve Traylen wrote:
> I think this would work well. People might be surprised if they submit
> a sleep 30 job which then fails because the queue is only 20 seconds long.
> Maybe it makes sense to stop accepting jobs for the last hour or so.
It would need a bit of slack anyway, because people are likely to know
the cpu time for their job better than the wallclock time, so you could
set a cpu time limit an hour or so less than the cutoff time.
Also, whatever happened to the TIMEL subroutine in CERNLIB, does anyone
know if it works with PBS or if any software still uses it? That used to
give a way for jobs to die gracefully before they got terminated. (The H1
event reader would return EOF once you got to the limit - very annoying
with VMS interactive sessions where you could hit the time limit even
though there wasn't one!)
Stephen
|