[torqueusers] [torquedev] GPU Hardware

Gareth.Williams at csiro.au Gareth.Williams at csiro.au
Thu Nov 4 17:47:37 MDT 2010



> -----Original Message-----
> From: Gareth.Williams at csiro.au [mailto:Gareth.Williams at csiro.au]
> Sent: Thursday, 4 November 2010 9:29 PM
> To: torqueusers at supercluster.org
> Subject: Re: [torqueusers] [torquedev] GPU Hardware
> 
> > -----Original Message-----
> > From: Christopher Samuel [mailto:samuel at unimelb.edu.au]
> > Sent: Wednesday, 3 November 2010 9:53 AM
> > To: torqueusers at supercluster.org; Torque Dev
> > Subject: Re: [torqueusers] [torquedev] GPU Hardware
> >
> > -----BEGIN PGP SIGNED MESSAGE-----
> > Hash: SHA1
> >
> > On 22/10/10 03:00, "Mgr. Šimon Tóth" wrote:
> >
> > > Several cards in one machine. Users should be able to
> > > select the amount of cards (or even specific type of
> > > card), Torque needs to make sure that each job will
> > > get its own requested cards (dedicated).
> >
> > I guess the simplest way to do this would be for Torque
> > to set all the GPU cards to root only access on startup
> > (if no jobs are running) and then set file permissions
> > appropriately per job.
> >
> > The main issue there will be if a user starts two jobs
> > on the same box then there will be the possibility of
> > clashes over which GPUs it can access.
> -snip-
> 
> We are using a group for access to the gpu devices (using
> pro/epilogue).
> It's not great and we're looking forward to doing something better...
> 
> (dual socket quad core, dual fermi-gpu nodes)
> 
> - Gareth

We can post more detail next week, but it is worth noting 
that we found that changing permissions on the device and also 
having different permissions on two devices caused us problems.  
That is why we settled on a group permission approach - but we 
Would prefer something better.

- Gareth


More information about the torqueusers mailing list