Hi,
Am 27.04.2011 um 00:04 schrieb William Deegan:
> Is there a way to set the maximum rate at which new jobs will be launched on
> a cluster (and/or a given machine)?
>
> My client's worried about submitting a 100 jobs and having them all start at
> the same time crushing the fileserver.
To throttle the overall job scheduling you will have to define two things:
$ qconf -ssconf
...
job_load_adjustments np_load_avg=1
load_adjustment_decay_time 0:7:30
and in the queue configuration:
$ qconf -sq all.q
...
load_thresholds np_load_avg=0.75
Each started job will put amn artificial load on each machine and avoid that
too many jobs can start at the same time. For the waiting job you will see an
output like:
$ qstat -j 416
...
scheduling info: queue instance "all.q@fubar1" dropped because it is
overloaded: np_load_avg=1.000000 (= 0.010000 + 1 * 0.990000 with nproc=1) >=
0.75
queue instance "all.q@fubar2" dropped because it is
overloaded: np_load_avg=1.020000 (= 0.030000 + 1 * 0.990000 with nproc=1) >=
0.75
until the artificial load decreased again. This virtual alarm state of a queue
won't show up in `qstat -f` though. The exact values of "job_load_adjustments",
"load_adjustment_decay_time" and "load_thresholds" need to be adjusted by
experiment to your needs.
-- Reuti
_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users
