[GE users] Fwd: Subordinate queue help
d.gruber at sun.com
Thu Apr 29 20:34:50 BST 2010
In SGE 6.2u5 there is a feature called "slotwise preemption"
(which was discussed also as slotwise suspend on subordination).
It does exactly what you need: You define a slot limit
in the high priority queue. This is the maximum amount
of jobs which can run in the low priority queue when
no job is running in the high priority queue. If there is one
job in the high queue then in the low priority queue one
job less than the limit can run. And so on. If the amount
of running jobs in the high queue and the low queue together
is exceeded, low priority jobs are suspended accordingly.
The high queue is not affected by the limit, there more
than 'limit' jobs can run. It is for non-parallel jobs only.
The scheduler is not aware if a job is going to be suspended
before the job is started. But to overcome this limit you
can make use of a host consumable, which counts the amount
of running jobs in both queues on the host for example (you
can give the high priority jobs more "costs" for the consumable)
If you set your scheduler queue sort method to 'load' and
select the consumable (or -consumable) as load formula, the
least loaded host according to high priority jobs is taken.
But this is for sequential (non-parallel) jobs only.
See also man page queue_conf(5):
Am Donnerstag, den 29.04.2010, 10:40 -0700 schrieb jlforrest:
> On 4/29/2010 7:30 AM, futurity wrote:
> > The other reason for choosing to use 12 "low" queues is that the
> > associated 12 "high" queues can suspend just a single job rather than
> > all the jobs on the same machine.
> I've just started looking into subordinate queues and I've
> learned about the behavior described above. That is,
> all the jobs in a subordinate queue get suspended when
> there is just one job in the higher queue. This seems
> unduly harsh. It would make more sense if there
> were a one to one suspension. By this I mean that if
> there's one job in the higher queue, then just one
> job in the subordinate queue should be suspended.
> Two jobs results in two suspensions, and so on.
> Searching this list shows that various other people have commented
> on this in the past and have submitted feature requests. Can
> anybody from Sun/Oracle comment on the status of these
To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
More information about the gridengine-users