Hi all,
we have a problem with ops jobs running on our site when it is
completely full with real jobs.
The site has 38 job slots (19 WN with 2 cpus). When it is full
and a SAM test comes it must wait for one real job to end. Since
the number of jobs slots is limited the waiting time can be too long.
Currently I have these solutions:
1) reservations
2) limitation of the number of allowed 'real' jobs (via QOS in maui) -
currently we use this one
3) preemption
I don't like either of them because it means that at least one cpu is
free for most of the time (also it seems now that solution 2 would make
our CE an attractor for jobs because there would be 1 free cpu published).
From my point of view the ideal solution would be if jobs from one
queue (ops) are allowed to run even on a busy node.
Is it feasible? I had no luck going through maui/torque documentation.
--
Tomas Kouba
Institute of Physics, Academy of sciences of the Czech Republic
|