Re: [Condor-users] Condor with CUDA on windows

[Lans Carstensen <lans.carstensen@xxxxxxxxxxxxxx>]

+1.  To me, that means working towards adding arbitrary per-system
resource counters for partitionable slots.  Supporting GPU's on a
partitionable slots configuration today is not practical - i.e.
advertising a partitionable resource of say 12 CPU's, 48G memory, and
2 GPU's, and enabling passing device-related configuration to the
partitionable slots requesting a GPU.

This model could then be extended to any other specialized compute.
It could also be used to solve node-locked licensing.  You could also
stretch it into some of the other resource counting methodologies
discussed elsewhere, e.g. you could advertise host network bandwidth
as a partitionable resource.

On Mon, Jan 16, 2012 at 6:45 AM, Tim St Clair <tstclair@xxxxxxxxxx> wrote:
> Maybe it's worth investigating rCUDA.?.?
>
> In general, we really to start treating GPU as a 1st class citizen, maybe target for next series?
>
> Cheers,
> Tim
>
> ----- Original Message -----
>> From: "Ziliang Guo" <ziliang@xxxxxxxxxxx>
>> To: "Todd Tannenbaum" <tannenba@xxxxxxxxxxx>
>> Cc: "Condor-Users Mail List" <condor-users@xxxxxxxxxxx>
>> Sent: Saturday, January 14, 2012 12:53:15 PM
>> Subject: Re: [Condor-users] Condor with CUDA on windows
>>
>> None of my preliminary attempts worked.  This included running the
>> service under my user account and trying that trick with allow
>> interaction with user desktop.
>>
>> I dug some more and the only "solution" out there is one Nvidia
>> provided for their Tesla card drivers:
>>
>> http://forums.nvidia.com/index.php?s=2cd29b11a0be8fa2e10b06aacd57beda&showtopic=93450&st=20
>>
>> People apparently have had success tricking the drivers to work with
>> non-Tesla cards, but at this point it seems a crapshoot.  The only
>> "solution" unless Microsoft or Nvidia and AMD work out a way to
>> provide access to GPUs in session 0 is for Condor to start up the
>> process inside a non-session 0 session, at which point we need to
>> test
>> to make sure the starter can communicate correctly with the startd or
>> the starter can actually keep track of the running process, etc, etc.
>> We MIGHT be able to tweak the use visible desktop code to do this.
>>
>> Z
>>
>> On Fri, Jan 13, 2012 at 4:31 PM, Todd Tannenbaum
>> <tannenba@xxxxxxxxxxx> wrote:
>> > On 1/13/2012 4:04 PM, Ziliang Guo wrote:
>> >>
>> >> Another user sent a similar question and I responded here:
>> >> https://lists.cs.wisc.edu/archive/condor-users/2012-January/msg00038.shtml
>> >>
>> >> Simplistically, due to the changed security model in Vista and
>> >> higher,
>> >> access to the video card is severely restricted in the session 0
>> >> that
>> >> Condor and jobs run in.  Currently the only way to run a CUDA or
>> >> any
>> >> job that requires a GPU on Windows using Condor is to run Condor
>> >> as a
>> >> personal Condor.  This however requires that a user be logged into
>> >> the
>> >> machine to start up and keep Condor running.  You also lose the
>> >> sandboxing from the user that Condor does when running as a
>> >> service.
>> >>
>> >
>> > So even running Condor as a regular user under the service control
>> > manager
>> > is not enough to allow access to the video card on Vista and
>> > higher?  Is
>> > there perhaps some Windows registry setting on some such to relax
>> > this
>> > restriction?
>> >
>> > Sorry I only have questions instead of answers...
>> >
>> > Todd
>>
>>
>>
>> --
>> Condor Project Windows Developer
>> _______________________________________________
>> Condor-users mailing list
>> To unsubscribe, send a message to condor-users-request@xxxxxxxxxxx
>> with a
>> subject: Unsubscribe
>> You can also unsubscribe by visiting
>> https://lists.cs.wisc.edu/mailman/listinfo/condor-users
>>
>> The archives can be found at:
>> https://lists.cs.wisc.edu/archive/condor-users/
>>
> _______________________________________________
> Condor-users mailing list
> To unsubscribe, send a message to condor-users-request@xxxxxxxxxxx with a
> subject: Unsubscribe
> You can also unsubscribe by visiting
> https://lists.cs.wisc.edu/mailman/listinfo/condor-users
>
> The archives can be found at:
> https://lists.cs.wisc.edu/archive/condor-users/