[torqueusers] TORQUE GPU support

Andrew Keen keenandr at msu.edu
Fri Sep 6 17:24:11 MDT 2013


On 9/6/2013 6:06 PM, torqueusers-request at supercluster.org wrote:
> Message: 5 Date: Fri, 6 Sep 2013 17:04:48 -0500 (CDT) From: Dave 
> Ulrick <d-ulrick at comcast.net> Subject: Re: [torqueusers] TORQUE GPU 
> support To: Ken Nielson <knielson at adaptivecomputing.com> Cc: Torque 
> Users Mailing List <torqueusers at supercluster.org> Message-ID: 
> <alpine.LFD.2.11.1309061703030.3235 at lion.cso.niu.edu> Content-Type: 
> TEXT/PLAIN; charset=US-ASCII; format=flowed On Fri, 6 Sep 2013, Ken 
> Nielson wrote:
>> >The following explains how to get rid of the nvidia-smi call and get torque
>> >to call the api instead.
> Question: will the call to the NVML library cause the nVidia driver to be
> loaded if it's currently unloaded? If so, wouldn't we have the same driver
> load/unload overhead with the API that we're seeing with nvidia-smi?
>
> Dave
Yes, I believe so. I'd suggest putting something in your epilogue that 
resets the card on job end; if you're scheduling at a node-exclusive 
level you could do a service nvidia restart .


More information about the torqueusers mailing list