Am 31.07.2012 um 13:57 schrieb Semi:
> I would like to define general CPU load limit.
> For example: if node have 24 cpu, normal uptime for it will < 24, if 8 < 8.
This is not feasible in Linux with recent kernels, as also processes in state D
count as running and might pass load_thresholds while having idle CPUs.
Unless you have interactive tasks in addition (which you could route through
SGE though), it's best to define slots=24 in the exechost definition, as it
will be across all queues residing on an exechost.
> Do not run any job if uptime is more. Does it possible do in general
> configuration, not per queue?
> All of my nodes have various CPU quantity.
You can either define it in a loop using:
a) `for i in `seq -w 120`; do qconf -mattr exechost complex_values slots=24
node$i; done`
or
b) an RQS:
limit name slots hosts {*} to slots=$nproc
> From other hand I have subordinate queues and I want that any job from
> subordinate queue will have their priority.
(I'm not sure whether I understand what you mean by "priority".)
With an RQS you can (and will have to) set a limit across the subordinated
queues only, otherwise the superordinated ones will never start as all slots
are already used up:
limit name sub queues !super.q hosts {*} to slots=$num_proc
limit name sup queues super.q hosts {*} to slots=$num_proc
(Depending on the version of SGE this might not work, due to a bug in RQS)
==
But: if you have only two queues, then you can limit the number of slots in the
queue definition per hostgroup, and you don't need any setup in the exechost or
RQS at all. With slotwise premption, you can also achieve that the number of
active slots will never pass the overall CPU count:
$ qconf -sc sub.q
...
slots 1,[@large=24],[@small=8]
$ qconf -sc sup.q
...
slots 1,[@large=24],[@small=8]
...
subordinate_list
NONE,[@largeslots=slots=24(sub.q:0:sr)],[@smallslots=slots=8(sub.q:0:sr)]
-- Reuti
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
