NVidia

From Nuclear Physics Group Documentation Pages
Revision as of 16:53, 5 August 2010 by Aduston (talk | contribs)
Jump to navigationJump to search

To make sure that the CUDA (graphics card computation system) is working (currently only available on Taro), check the following:

  • When a new kernel is install, the NVidia kernel driver also needs to be updated. This is done by re-running the NVidia installer in /root/CUDA_INSTALL. Run the one with the highest version number.
  • The permissions on the /dev/nvidiactl and /dev/nvidia* devices needs to allow users to read them. To set these automatically create a file named nvidia in /etc/modprobe.d/ (if it isn't there already) any put this in it: (The NVreg_DeviceFileGID number should be the GID for the video group).
options nvidia NVreg_DeviceFileUID=0 NVreg_DeviceFileGID=98 NVreg_DeviceFileMode=0666
    • When you reinstall the driver (after a kernel update, for example), you'll need to make sure these permissions are correct. Rebooting should reset them automatically if the modprobe file exists, but it isn't necessary.
  • Make sure the video group exists.
  • Make sure users that want to use the CUDA system are in the video group.
  • You do not need to run X to use the computational system (in the newer versions).

Checking the device

To run the graphical NVidia settings program run:

 Xorg &
 nvidia-settings --display=$DISPLAY  -crtl-display=:0

Where of course you know to set $DISPLAY to your current output device (i.e. your screen, which is a tunnel for ssh, something like localhost:14.0) That first command runs X in the background, which is needed for this program to access the card.

Another code is nvclock, this is 3rd party and beta and would allow overclocking. We don't use this for that reason.