Apologies for the duplicate mailing. I sent this earlier, but I think
it was recorded as a response to a thread, not a new one.|
We're running a fairly small condor cluster (16 dual core cpus, so 32nodes) on machines that are also general purpose compute servers.
I think we're running into problems where condor is marking a CPU as busy when users are running other (non-condor) processes on the machines.
I think the confusion is because some users are bypassing condor and running their jobs directly on the machines. This causes CpuIsBusy =
TRUE (e.g., condor_status -l s11) and prevents these machines from getting matched to jobs. Meanwhile, condor_q reports these machines
misleadingly as idle.
Is there a way around this? I've read the Preemption and scheduling sections of the manual, and they all appear to deal with how to handle scheduling WITHIN condor. Is there a way to make condor's threshold for flagging a CPU as busy significantly higher? How about increasing the priority that ALL condor jobs run as?
-- Chris Mason