CUDA failures with 260 series drivers

[Aaron Haviland]

This describes the problem as seen in
http://www.primegrid.com/forum_thread.php?id=2839
and in
http://forums.nvidia.com/index.php?showtopic=186645

This is not an Ubuntu bug, but an upstream one. I don't believe there is anything we can do to fix it. Am filing this bug to provide information and a workaround for people searching for this problem.

This problem has been observed with 260.19.06, 260.19.12, 260.19.21, and also the beta 260.19.26

Application source is available here:
https://github.com/Ken-g6/PSieve-CUDA

The PrimeGrid application intentionally overloads the CUDA processor with more threads than it needs (much like people use "make -j(CPUs+1)"). This seems to expose a bug in the 260 series drivers that did not exist previously.

This results in the application bailing out with an unpredictable frequency (on my card, sometimes it would die instantly, others it would die after a couple minutes. People with different CUDA-capable cards may find different timings.)

The error reported in the application's stderr.txt is:
| Computation Error: no candidates found for p=1795516820654281
where p=xxxx whatever number was being tested at that point.

Workarounds that are known to work at this point:
1) Downgrade to 256 driver series. (e.g. https://launchpad.net/ubuntu/maverick/amd64/nvidia-current/256.53-0ubuntu3)
2) Use BOINC's app_info.xml mechanism to specify that the program uses fewer threads, with the flag "-m 2"

[bugbot (bugbot)]

Hey Aaron,

Hi, have you had a chance to test if this bug is still present in natty?

If it does (and if you're the original reporter), please boot into natty
and run the command:

  apport-collect <bug-number>

which will update the bug with fresh logs and tag the bug as affecting
natty. (It is best to run this right after reproducing the problem.)

[bugbot (bugbot)]

We're closing this bug since it is has been some time with no response from the original reporter. However, if the issue still exists please feel free to reopen with the requested information. Also, if you could, please test against the latest development version of Ubuntu, since this confirms the bug is one we may be able to pass upstream for help.