[Halld-offline] Hall D batch farm priority
marki at jlab.org
Thu Mar 10 12:05:52 EST 2016
Talked to Sandy. Here is my current understanding.
Our priority has been raised from 27% (= 30/110) to 60% on the farm.
After the run, we will drop back to 50%. This was Rolf Ent's decision
based on a proposal from Graham Heyes based on suggestions from many of us.
There is no more "exclusive" queue on the farm. The "multi-core" queue
that you may or may not have heard being discussed recently does not
exist either. There is only one production queue, "prod64". If you want
to run a multi-threaded job, ask the batch system for an appropriate
number of "CPU" in your request. The scheduler will make an effort to
create space on one of the nodes so that your job can run. That means,
for example, if you request CPU=48, you will run exclusively on one of
the 48-hyperthreaded nodes, once a node is freed up by the batch system.
I still have some uncertainties in my own mind on details of the
scheduler, so questions and corrections from others are welcome. I can
pass on questions that I cannot answer myself.
More information about the Halld-offline