I have cuda installed on RH 4.4 with 8800 Ultra. I am accessing the machine remotely, and all the samples fail with the following diagnostics:
./clock
NVIDIA: could not open the device file /dev/nvidiactl (No such file or directory
).
There is no device supporting CUDA.
The only way to make it work that I found is to physically log on to the x server on an console as root, and run one of the examples. After that remote sessions start to work as well. Is there any way to set it up so I won’t need physical access?
Even when accessing remotely via SSH or something else, X must have run at least once since boot time. It would suffice to run “startx” as a normal user, wait a few seconds, then kill it.
at my university the problem you describe consisted even if a user logged in locally into X. So additionally we had to make sure that the users that work with CUDA are members of the “video” group as only those have proper access writes to /dev/nvidiactl.
This has been described a lot of times already on this board. Please search the board.
In short - depends on how you access the remote machine (VNC or ssh). Direct VNC connection is fine out-of-the-box. With ssh, make sure ssh does not forward the X protocol and you are actually using the remote server. If your program does not have an X window, there is no need for a remote X server at all. In that case, just make sure the nvidia kernel module is loaded (ie. put modprobe nvidia in /etc/init.d/boot.local or similar). No need for silly starting & killing X.