CUDA processing problems
9 Feb 2010 20:53:07 UTC
Topic 194766
(moderation:
I have a crashing ABPS task. It lists its CPU time as 00:00:01, progress as 100%, and status as either waiting to run or running, both with (1.00CPUs, 1 CUDA).
My system is 32-bit Kubuntu 9.10 with BOINC manager 6.4.5 (from repos). CUDA is enabled and my 9600GT is found by BOINC at startup. CUDA works successfully with the SDK tests and my own programs.
The stderr.txt for the crashing task is here: http://www.pastebin.ca/1791822
It seems to be having problems with the CUDA libraries, but I have no idea beyond that what's going on. Does anyone have any suggestions?
Edit: I should say that this is the first ABPS task I've run since enabling CUDA.
CUDA processing problems
)
gpu driver version?
I run 64-bit Ubuntu 9.10. I had to install driver 190.42 from the nVidia website (its not in the repositories) to get cuda to work (running a 9800GT card).
Note, I only do Einstein@Home. I think it's Einsten that requires a minimum CUDA compatibility level of 2.3 ?? (hence ABP2cude23), which for linux is 190.42 and above. BOINC may be happy and recognise the device but Einstein may not be.
http://ubuntuforums.org/showthread.php?t=990978 is a good guide to doing the install. If you do, you will have to repeat this each time the kernel is updated.
I'm running 190.18, which
)
I'm running 190.18, which were the latest CUDA-supporting drivers it offered me when I set this up. I note .53 is now available, so I'm downloading that. I am running the latest stable version of CUDA (2.3).
Unrelated to Einstein, but you might be interested in this guide to getting nvidia drivers to automatically update with new kernel versions: http://ubuntuforums.org/showthread.php?t=835573
RE: Unrelated to Einstein,
)
Cheers for this! I have update manager insisting kernel 2.6.31-19 is a really good idea. I have been ignoring it 'cause of this very issue.
To your problem, have a read of this post: http://einsteinathome.org/node/194500&nowrap=true#99973. "For all of you who report CUDA-related problems ... Please make sure that you use at least BOINC 6.6 because 6.4 still has several known CUDA-related issues."
Follow the thread, some of the early posts match your issue.
I'm running BOINC 6.10.17, again not in the repo.
RE: To your problem, have a
)
Ah, yes, that does look much like my issue. Problem persists with the new drivers, so looks like I'm doing a manual BOINC update today then. Thanks!
Ah, success! BOINC updated
)
Ah, success! BOINC updated successfully and painlessly to latest version by copying the relevant files over the package manager installed files and CUDA processing on that task is now proceeding. Thanks for the help!
It's actually surprising that
)
It's actually surprising that with a 6.4.5 client you get CUDA Tasks at all. I am pretty sure we limited the CUDA Application to 6.10 Clients. If you got CUDA tasks anyway, it means that this feature is actually broken in the current scheduler.
BM
BM
RE: It's actually
)
It must be; 6.4.5 is definitely reporting its version correctly and I was sent several CUDA tasks.
I doubt there's anything that can be done about this, but if Einstein is the only CUDA-supporting project you have enabled, BOINC will try and keep the GPU busy as well as the CPUs, and thus will constantly be running Einstein modules using one CPU core plus the GPU, ignoring project resource balances. I upgraded to 64 bit so I could add GPUGRID and PrimeGrid's CUDA WUs to fix that.
For future reference if anybody stumbles upon this, the easiest way to install CUDA-compliant nvidia drivers on Ubuntu is by adding the nvidia VDPAU team's PPA from here: https://launchpad.net/~nvidia-vdpau/+archive/ppa This has updated nvidia driver packages, making installing and building new modules for new kernel versions totally painless.