Re: [HTCondor-users] Fw:Can condor_exec get the slot number when running?

[Todd Tannenbaum <tannenba@xxxxxxxxxxx>]

On 5/16/2017 6:57 AM, çå wrote:
> Hi
>    I installed two GPUs on one machine,and want to bind each GPU to a
> slot,for example,
> There are GPU0 and GPU1, and i modified condor config, slot1@yffs and
> slot2@yffs can be found on the same machine.
>
>
> I want all the jobs running by condor (condor_exec) get the slot number
> of which it is running on,
> So that I can manual bind the GPU with the slot.
>
> Is there any way to bind each GPU to a slot?
> Or just get the slot number when condor_exec running?
>
> Thanks very much

Append the following line to the condor_config on your execute machines:

  use feature : GPUs

By default, this will evenly distribute GPU devices across your static 
slots, and also set up the environment variables for CUDA (and i think 
OpenCL) to bind programs running on that slot to the device.  In other 
words, I think adding this one line will do exactly what you want.

More details and other options can be found in the HTCondor HOWTO 
recipes; specifically see
  https://htcondor-wiki.cs.wisc.edu/index.cgi/wiki?p=HowToManageGpus

regards,
Todd