It appears BOINC w/ CUDA requires gdm to be running first.

[Steffen Moeller]

Hi,

the CUDA part of the BOINC upstream code was removed for license issues.
I could imagine that this makes a difference of some sort with respect
to initialisation, but having not looked at it I am only guessing here.

I am happy to hear that CUDA works with Debian/Ubuntu BOINC, actually :)

So, the hunch I have is that the commercial nvidia driver is somehow
invoked and provides the underlying CUDA functionality which BOINC then
profits from. You could test this by substituting the GDM with KDM
(which would mean that the desktop manager does not make a difference
but the initialisation of  X does) and the official nvidia driver with
the free one that ships with Debian.

And ... should my hunch be correct, I don't know what this actually
means for our BOINC packages. Comments by the others will tell :)

Best,

Steffen

Skip Guenter wrote:
> Hey ya'll,
>
> First a HUGE THANX for all the good stuff ya'll do!
>
> On to my problem...
>
> I recently installed an 8800GT card into each of two different machines.
> One my normal desktop (Ubuntu v8.10 64b) and another into a headless,
> diskless number cruncher (Dotsch/UX / Ubuntu v8.10).  There was little
> problem with Intrepid machine as I only had to create a link
> in /var/lib/boinc-client to libcudart.so to resolve an error message.
>
> The install on the Dotsch/UX machine was another thing.  I'm trying to
> work with the author of that distro on what I believe the problems are
> when the machine is running off of a served up NFS.  
>
> Anyway, to get around this I installed a HDD back into that machine and
> installed Jaunty on it (v9.04, 64b).  All went well with the same link
> needing to be added into /var/lib/boinc-client.  
>
> On BOTH machines I installed v6.4.5 deb package from 'getdeb' Intrepid
> 64b.
>
> On the 2nd (normally headless) machine I had the gdm NOT starting
> automatically via a rename of the "S30gdm" in rc2.d to "s30gdm". 
>
> I then noticed that BOINC was throwing errors saying "missing CUDA
> device".  I started the gdm back up and then restarted boinc-client.
> This resolved the error.  I then reset the gdm to auto start and
> rebooted the machine to make sure it would be OK on a restart.  It was
> not.  I again had the missing CUDA device message.  
>
> I looked at /etc/rc2.d and it appears to me that boinc-client is started
> at S20 and gdm is started at S30.  IF (that's a BIG IF) I understand
> correctly that means that boinc-client is already running by the time
> xsession/gdm is started.  I think this is a problem because I believe
> the file /etc/X11/xorg.conf contains things needed by the nvidia device
> driver which in turn is needed by BOINC to recognize the CUDA card.
>
> Is there a simple rename that I could do to resolve this locally?
> Something along the lines of:
>
> sudo mv /etc/rc2.d/S20boinc-client /etc/rc2.d/S32boinc-client
>
> I'm afraid there may be something else that would have to match an
> action like the above on another file but I don't know what it would be.
>
> I don't know how to report this as an official problem but still would
> appreciate anybody that could help with a local fix.  
>
> If somebody wants to point me to something on how to report the bug I'll
> try to fill out the forms. 
>
> Thanx in advance, Skip
> República de Tejas, Centro
>
>
> --
> pkg-boinc-devel mailing list
> pkg-boinc-devel at lists.alioth.debian.org
> http://lists.alioth.debian.org/mailman/listinfo/pkg-boinc-devel