[torqueusers] Problem with ppn and routing : Possible way to get the routing you want continued.
jamagallon at ono.com
Thu Dec 2 17:07:19 MST 2010
Sorry if I do some childish questions, but I have not been able to get
the answers or misunderstood some things I read, perhap you could include this
in a kind of newbie FAQ, if not already there...
On Thu, 2 Dec 2010 14:33:01 -0700 (MST), Ken Nielson <knielson at adaptivecomputing.com> wrote:
> The TORQUE "resource manager" knows nothing of ncpus. When a job is submitted and the ncpus keyword is used the string is passed through to the scheduler. In the case of PBS this would be pbs_sched. If you run torque without a scheduler and call qrun for a job you will get a single node and a single processor to run the job.
I was not aware that torque was pretty useless without a scheduler, ie,
jobs just sit in the queue in Q state if pbs_sched is not running. I had
understood that without a scheduler there was a FIFO policy, and that a
scheduler just let you do more complicated things. But using pbs_sched
is mandatory for a real system. IT is THE FIFO.
> >> ncpus and nodes are competing ways to specify a resource request. Don't mix
> >> them and everything works better.
> >> ncpus pre-dates clusters and is used to specify the number of cpus on 1 node.
> >> nodes was grafted into OpenPBS later in life to deal with clusters.
> > The TORQUE resource manager has no concept of ncpus. It is interpreted
> > by the scheduler. In the case of Moab it indicates the number of procs
> > to be used per task.
> I don't know much about how ncpus are handled, but torque must have
> some kind of concept, as it predates nodes, and I don't think anyone
> went through and stripped out all kinds of code related to ncpus. I
> do know that your $PBS_NODEFILE only gets one host for requests using
> ncpus unless you are using something like Maui or Moab.
If you read man pbs_resources, 'nodes' is described with the full spec,
including 'ppn', needed for job submission. The if you read docs about
queue limits here
it looks like resource limits are just the same, but it not really, ppn is
ignored, for example. Then that page mentions 'nodect', not referenced
in man pbs_resources. So I'm really messed up...
Lets suppose for further examples I have a cluster with 16 quad-core boxen.
So if I understand things, in queues I can only limit the number of nodes,
not ppn. So I can set up a queue for long jobs, allowing a max of 4 nodes
for example. That queue will allow jobs like nodes=1:ppn=4 or nodes=4:ppn=4.
It just will be able to run only one 4:4 job, or two 2:4, or two 4:2,
or sixteen 1:1, etc...
But I can not limit the number of ppn's to force jobs to ppn=2 and be able
to run several long jobs in parallel.
Thinking about it, perhaps it is enough to limit the number of 'cores' a
job can use. Suppose I limit it to 4 (from 16), so max valid jobs would be
like 1:4, 2:2, 4:1. Is there a way to do that ? Limit on total number of
cores used ? ncpus is for a single node (not total) and outdated, procs is
the same as nodes, nodect is also the same, so nothing seems to fit for
J.A. Magallon <jamagallon()ono!com> \ Software is like sex:
\ It's better when it's free
More information about the torqueusers