New kernel came out for Ubuntu 11.04 and I installed it today.
Unfortunately my system no longer boots with NVidia driver 270.41.06. So I updated to "nvidia-current" in the ppa.
The system seems to work fine except for BPRcuda32nv270 which gives me the error:
[09:22:42][2797][INFO ] Starting data processing... Error: API mismatch: the NVIDIA kernel module has version 280.13, but this NVIDIA driver component has version 275.09.07. Please make sure that the kernel module and all NVIDIA driver components have the same version. [09:22:42][2797][ERROR] Couldn't initialize CUDA driver API (error: 100)!
example:
242778750
Anybody else seen this? Any suggestions?
Joe
Copyright © 2024 Einstein@Home. All rights reserved.
Linux NVidia 280.13
)
I'm running 280.13 and the new driver is working okay for me. It looks like your libcuda was possibly not updated. You can check for those files and links here:
/usr/lib/libcuda.so.280.13
/usr/lib/libcuda.so -> libcuda.so.280.13
/usr/lib/libcuda.so.1 -> libcuda.so.280.13
/usr/lib64/libcuda.so.280.13
/usr/lib64/libcuda.so -> libcuda.so.280.13
/usr/lib64/libcuda.so.1 -> libcuda.so.280.13
Then if you have everything set, run ldconfig.
You can also manually extract the nvidia driver with the -x flag copy those libraries in place manually.
Thanks Jeroen, I think you
)
Thanks Jeroen,
I think you nailed it.
Same for lib64.
when you say:
Are you suggesting I download from nVidia instead of the ppa? I'll try that and see what happens. I'm getting pretty adept at backing out nVidia changes, unfortunately.
Joe
It's working now. Thanks
)
It's working now. Thanks Jeroen!
The problems seems to be an incompatibility or rather a different installation paradigm between the nVidia supplied drivers and the Ubuntu supplied drivers.
You cannot mix them so don't because it's not easy to get things squared away.
It seems the nVidia installer puts the dynamic libraries in /usr/lib and /usr/lib32 while the Ubuntu installer puts them in /usr/lib/nvidia-current and /usr/lib32/nvidia-current. Since ld.conf has /usr/lib first and everything is referred to by symbolic links without version number version mismatches are inevitable.
What I ended up doing, which is probably far from optimal and I recount it to discourage mixing the installers not as a recommended procedure to fix them if you have.
1. Boot in recovery mode
2. uninstall the drivers from nVidia with sudo sh NVIDIA-Linux-x86_64-275.09.07.run --uninstall
3. Run the Synaptic package manager and uninstall all the nNvidia packages
4. Reboot in recovery mode
5. Install nvidia-settings and nvidia-current-dev
6. Reboot in normal mode
I believe I had to install the drivers from nVidia in the past to get a CUDA version that worked with E@H. What I should have done is uninstall it before allowing Ubuntu to upgrade. One of these days I will learn something the easy way, it's bound to happen.
Joe
RE: Thanks Jeroen, I think
)
Either way should be fine. You could try and reinstall using the package management tool. Or you could download the driver from nvidia.com and extract the libcuda library that way.
If you wanted to try and install the library manually, this is how I normally do it:
./NVIDIA-Linux-x86_64-280.13.run -x
cp NVIDIA-Linux-x86_64-280.13/libcuda.so.280.13 /usr/lib64
ln -sf /usr/lib64/libcuda.so.280.13 /usr/lib64/libcuda.so.1
ln -sf /usr/lib64/libcuda.so.280.13 /usr/lib64/libcuda.so
cp NVIDIA-Linux-x86_64-280.13/32/libcuda.so.280.13 /usr/lib
ln -sf /usr/lib/libcuda.so.280.13 /usr/lib/libcuda.so.1
ln -sf /usr/lib/libcuda.so.280.13 /usr/lib/libcuda.so
ldconfig
What I saw as I was
)
What I saw as I was investigating this is that the 280 versions were loaded but the older 27x were not removed.
I think as long as Ubuntu is providing new enough drivers to run E@H, I'm inclined to let them maintain them.
Now that I have a (minimal) understanding of the problem it's much less scary.
Joe
Glad to hear you have the 280
)
Glad to hear you have the 280 driver installed and up and running.