[torqueusers] In search of a long forgotten feature...
jbernstein at penguincomputing.com
Tue Feb 24 12:02:02 MST 2009
Garrick Staples wrote:
> On Tue, Feb 24, 2009 at 10:54:12AM -0800, Joshua Bernstein alleged:
>> Garrick Staples wrote:
>>> On Mon, Feb 23, 2009 at 05:34:02PM -0500, Miller, Timothy E. alleged:
>>>> I seem to recall that there was a way in PBS that, when a job was
>>>> submitted to a particular queue (say i386), PBS would automatically
>>>> add an attribute to that job as specified in the queue's configuration
>>>> (say '32bit'). One could then add a node attribute '32bit' to all
>>>> your 32bit nodes and 'viola!' ensure 32bit jobs ran on 32bit labeled
>>> Others have answered this question, but in this specific case node
>>> aren't the best way to go.
>>> pbs_mom can make an "arch" visible and users can select an arch.
>>> Set arch in mom_config to `uname -m`, and users can add arch=i686 or
>>> arch=x86_64 to their job resource request.
>> Though arch is a special case attribute. Consider something like "IB" for
>> Infiniband, or more recently I've been using "Cuda".
>> I was under the impression that when a job enters a queue, the queue does
>> not modify the job, and thus as it claims in the documentation there is no
>> way to tie a set of nodes to a queue, unless the users themselves asks for
>> the attribute associated with the queue. Am I confused?
>> If the required_property option patch was designed to modify the job upon
>> entering the queue, I would really love to see it incorporated into the
>> upstream TORQUE.
> resources_default properties supply defaults for jobs. resources_default.nodes
> and resources_default.walltime are probably the most common and useful.
Right, but these don't allow you to tie nodes to certain arbitrary properties.
> resources_default.neednodes is used by the scheduler to specify required node
> properties. This is the easy way to "tie" nodes to a queue.
resource_default.neednodes doesn't seem to be honored by pbs_sched, it only
works with Maui or Moab. If I'm mistaken, I'd sure like to know what I've been
doing wrong, and why the documentation suggests I'm correct.
More information about the torqueusers