[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [HTCondor-users] Translating GPU device assignments?
- Date: Thu, 06 Jul 2017 21:22:30 +0000
- From: Michael Pelletier <Michael.V.Pelletier@xxxxxxxxxxxx>
- Subject: Re: [HTCondor-users] Translating GPU device assignments?
> -----Original Message-----
> From: HTCondor-users [mailto:htcondor-users-bounces@xxxxxxxxxxx] On Behalf
> Of John M Knoeller
> Sent: Thursday, July 06, 2017 5:04 PM
> GPU_DEVICE_ORDINAL is the equivalent of CUDA_VISIBLE_DEVICES for OpenCL,
> It would be incorrect for us to renumber it.
> it sounds like you are saying that the job shouldn't look at
> CUDA_VISIBLE_DEVICES at all, it should just look at the number of GPUs it
> has been assigned and then start from 0.
At least in the case of how Caffe is doing it with CUDA devices. I'm not aware of whether or not this is the normal behavior, though. Does it need a "CUDA_DEVICE_ORDINAL" if it is normal?
As far as I can tell the CUDA_VISIBLE_DEVICES is interpreted by the CUDA library, not by Caffe, and from Caffe's perspective it just sees a sequential list of ID numbers corresponding to the visible devices.