Hi Stephen,
> A couple of things which I think UK people have mentioned: I seem to
> recall someone asking something about running low-priority background jobs
> on WNs (one of the London sites?), but I don't remember the details and I
> couldn't find it in the archive.
You might be thinking of UCL. On our HEP cluster we have a "bulk" queue
that is allowed to fill the entire cluster, but with jobs running "nicely"
so that jobs on other queues will push the "bulk" jobs into the background
when they start running.
> There's also a question of how to deal with hyperthreading (e.g. in cpu
> counts), does anyone know exactly how pbs, or indeed other batch systems,
> deal with that, i.e. does it just see twice as many CPUs?
I don't know about other batch systems (though I sometimes wish I did
given the problems we have with PBS) but as far as I know PBS sees as many
CPUs as you tell it to see in the "nodes" config file. With dual
hyperthreading CPUs in our WNs, the obvious approach would be to declare
each node as having 4 CPUs. However, because we want to be able to run 4
high-priority jobs on a node while still having 4 bulk jobs running slowly
in the background, we actually declare 8 CPUs on each node.
Hope this helps!
Ben
--
Dr Ben Waugh Tel. +44 (0)20 7679 3783
Dept of Physics and Astronomy Internal: 33783
University College London
London WC1E 6BT
|