BOINC Not Seeing GPU's With PCLinuxOS_64

[Terror Australis]

I have just installed PCLOS_64 RC5. When this is finally released it looks like it will be a very good distro.

However, BOINC does not recognise my GPU's. I get the message "No usable GPU's found".

I have the V302.17 Nvidia drivers installed, CUDA installed (both from the PLOS repo), there are no bugs in my xorg.conf file and Nvidia settings works properly and the GPU temps appear in gkrellm so I'm pretty sure all is ok there.
 
I have all the correct BOINC files installed and configured as they should be (I'm getting pretty good at this).  The permissions are correct ( I set them to "all" just to make sure). There are no hardware problems as the PCLOS installation is on hardware that had a working Mageia installation on it up until very recently and is untouched

When I start the program from a console there are no error messages.

I'm pretty sure this is a Distro problem, not a BOINC problem.

Does anyone have any suggestions as to what may be broken and what I can do to fix it ?

T.A.

[sunu]

1) What gpus are you trying to use?
2) What cuda version did you install? What "cuda" exactly did you install? You only need two files libcudart and libcufft. And they must be version 3.2
3) When you say "When I start the program from a console there are no error messages" which program did you start? boinc? seti?
4) Which boinc version did you use?

[Terror Australis]

Hi Sunu
1) GTS250's
2) 302-17 The only one available from the repo (that is the PCLOS version number) and these are the CUDA/Open CL libraries for the V302.17 Driver
3) BOINC (of course).
4) 6.10.58

A note about Q2 I have found that you need to install the CUDA libraries separately  for CUDA to work on both Mandriva and Mageia systems. As PCLOS is also a Mandriva fork the same possibly applies. On all systems I've built, there is no CUDA unless these libraries are present even with libcudart and libcufft in the SAH project directory (which they are).

You may have a point about the CUDA version, when I installed the libraries it also upgraded the Nvidia driver version from 295 to 302.17. But the cards were not recognised before the update either. libcudart.so.3 and libcufft.so.3 are in the SAH folder and are pointed to in the app_info file.

To make sure I copied libcudart and libcufft from the SAH directory to the /usr/local/cuda/lib and lib64 directories and added them to my $LD_LIBRARY_PATH but the cards are still not recognised.

T.A.

[sunu]

Get the relevant cuda libraries from http://developer.nvidia.com/cuda-toolkit-32-downloads. There are 3 files for each libcudart and libcufft, 2 symbolic links and 1 "real" file.  In app_info.xml, libcudart.so.3 and libcufft.so.3 need to be referenced. To avoid any collision, it would be better to remove any cuda related packages from the repository. Do an ldd on the setiathome executable to see if it finds the libraries it needs.

I also never install any nvidia drivers from the repository. I get the drivers straight from nvidia and install them manually. You can also try that.

In your first post you said that you installed boinc and set the permissions correctly. How exactly did you install boinc? You should never have to tinker with any boinc related permissions. Get boinc from berkeley and install it manually in your home directory. Leave whatever permissions alone.

[Terror Australis]

Some testing with a raw Mageia install shows that for CUDA to work under a Mandriva related system either the CUDA toolkit libraries or the Toolkit development files have to be installed. I'm not sure which ones as I loaded them both at the same time.

With Mandriva/Mageia you can't install the drivers using the Nvidia .run packages as they conflict with something inside the system and crash it. However there is a script available that converts the .run file into 5 RPM's which can be installed using urpmi. Obviously this effects PCLOS as it is Mandriva derived as well.

There is a possible version clash here. The CUDA libraries version from Mageia is labelled 3.2.16, as I posted above the PCLOS version is 302.17. Whether the numbering sequence is related or not I don't know.

FYI the sequence of the fresh install was as follows.
1) Clean Mageia install
2) Copied BOINC directory over from the PCLOS Install
3) Installed V295.59 drivers using the above mentioned script
4) Copied the xorg.conf file over from PCLOS.
5) Started BOINC - GPU's not found
6) Installed the V295.59 CUDA/ opencl libraries
7) Started BOINC - GPU's not found
Installed the CUDA toolkit and toolkit devel. libraries
9) Started BOINC, cards recognised and crunching.

This is the same method I used with the PCLOS install. The only difference is the version numbers of the Driver and Tool kit libraries. When I get some time I will roll the PCLOS drivers back to 295 and copy the toolkit files over from the Mageia install and see what happens. Likewise I'll do the same back the other way.

If you can think of any reason why the PCLOS install did not work I'd be interested to hear your thoughts.

Thanks for your help, it did provide some inspiration.

T.A.

[sunu]

With Mandriva/Mageia you can't install the drivers using the Nvidia .run packages as they conflict with something inside the system and crash it. However there is a script available that converts the .run file into 5 RPM's which can be installed using urpmi. Obviously this effects PCLOS as it is Mandriva derived as well.

Where did you read that? Searching google for "pclinuxos nvidia driver how to" didn't show anything remotely close to an nvidia driver installation guide for pclinuxos in the first 2-3 pages I looked at.            

Anyway working from a livecd (why isn't there any 64bit livecd mentioned at all at the official site and wiki?            ? I dug it out from the ftp mirrors) I installed the nvidia driver 302.17 with no problem (just a garbled screen during installation), seti works fine and boinc 6.10.58 recognized a 470 nicely. Before I installed the nvidia driver I --purge removed:
dkms-nvidia173
dkms-nvidia96xx
dkms-nvidia-current *
lib64dri-drivers-experimental
nvtv
x11-driver-video-nouveau
x11-driver-video-nvidia173
x11-driver-video-nvidia96xx
x11-driver-video-nvidia-current *
Probably only those 2 with an asterisk needed to be removed but I got purge happy.

There is a possible version clash here. The CUDA libraries version from Mageia is labelled 3.2.16, as I posted above the PCLOS version is 302.17. Whether the numbering sequence is related or not I don't know.

Don't be confused by those two numbers. They are completely different. 302.17 is a driver version, 3.2.16 is a cuda version.

If you can think of any reason why the PCLOS install did not work I'd be interested to hear your thoughts.

-Maybe the script you mention doesn't do a good job. You don't need it anyway.
-You also don't need any cuda related packages installed from the repositories. Only the files I mentioned in my other post.
-Did you run ldconfig after you copied the cuda libraries (with the corresponding changes to ld.so.conf.d) to rebuild the cache?
-Constant fiddling with boinc/seti related permissions that you shouldn't and needn't touch.

[Terror Australis]

With Mandriva/Mageia you can't install the drivers using the Nvidia .run packages as they conflict with something inside the system and crash it. However there is a script available that converts the .run file into 5 RPM's which can be installed using urpmi. Obviously this effects PCLOS as it is Mandriva derived as well.

Where did you read that? Searching google for "pclinuxos nvidia driver how to" didn't show anything remotely close to an nvidia driver installation guide for pclinuxos in the first 2-3 pages I looked at.            

It's mentioned many times on the forums for the MDV/MGA distros. The first post in this thread tells the story. It may not be necessary for PCLOS. I didn't have try it on that install, the latest version drivers were on the Repo.

Anyway working from a livecd (why isn't there any 64bit livecd mentioned at all at the official site and wiki?            ? I dug it out from the ftp mirrors) I installed the nvidia driver 302.17 with no problem (just a garbled screen during installation), seti works fine and boinc 6.10.58 recognized a 470 nicely.

It isn't officially released yet, the version I'm using (and probably the one you have too) is RC5. I was not working from a live CD it was installed to HDD

Before I installed the nvidia driver I --purge removed:......

The install I used had V295.59 drivers on it but not the  CUDA/OpenCL libraries. When I installed the libraries from the Repo it also updated the driver to 302.17 to suit the libraries

There is a possible version clash here. The CUDA libraries version from Mageia is labelled 3.2.16, as I posted above the PCLOS version is 302.17. Whether the numbering sequence is related or not I don't know.

Don't be confused by those two numbers. They are completely different. 302.17 is a driver version, 3.2.16 is a cuda version.

My mistake there (it was 4AM when I typed that post) the 302.17 libs I was referring too are the CUDA/OpenCL libraries for that version driver which I agree are not the libs for CUDA itself.    [Last sentence edited for clarity]

If you can think of any reason why the PCLOS install did not work I'd be interested to hear your thoughts.

-Maybe the script you mention doesn't do a good job. You don't need it anyway.

With Mandriva/Mageia you definitely need the script. Using the Nvidia .run file without it definitely makes a hash of things.l

-You also don't need any cuda related packages installed from the repositories. Only the files I mentioned in my other post.

I've done roughly a dozen installs of Mandriva/Mageia and have never gotten CUDA to work without them. Maybe there are other ways to do it but installing the actual CUDA packages is the quickest and easiest for a second rate hacker (as the term is used in the golfing world) like me    Come to think of it, IIRC  it's mentioned somewhere in an old post on the MDV forum that they are required but I couldn't tell you where.

T.A.

[sunu]

Lets see, mageia 2 livecd (may 2012 says the site):

-nvidia driver installed. Again removed some nvidia related packages already present before install. Mageia's livecd wasn't compile ready (like pclinuxos's was) so installed a couple kernel and X11 devel packages
-boinc sees the 470
-seti runs nicely


The first post from the thread you mention says:

If you do not use RPM packages to install the driver and instead run the
NVIDIA installer directly, it will overwrite some system files with the
following consequences:
- X server will fail to start if you update kernel or x11-server packages

This isn't a problem just an added step. Every time you install a new kernel or x server you rerun the driver installation

- 3D acceleration will stop working if you update mesa packages

Didn't check but I doubt it. A driver reinstall should be able to solve it anyway. What effects the livecd had (transparencies, jumping icons, some other stuff) continued to work with nvidia's driver. This is irrelevant for seti anyway, you don't want 3d effects to steal computing cycles.

- switching to non-proprietary driver using XFdrake will not work

What's the point of this?


What I gather from all these is what mageia, pclinuxos, scripts, etc mean about "nvidia driver" is not the full driver you get from nvidia but just a part from it. If you need the full functionality that nvidia's driver offers, you need to install multiple packages in these systems.