5s run time limit on Linux 5s runtime limit on Linux

Hello Folks,

I’ve installed an 8800GT on the headnode of a Linux cluster running RHEL 4, the only node in which the card physiclally fits. After a successful installation - both hardware and software -
the 8800GT AUTOMATICALLY defaults as the video card on boot. I ran some of the examples
and all went fine. I have not been able to change this through the BIOS, and with this in mind, I noted in the toolkit release notes that:

“Individual GPU program launches are limited to a run time of less than 5 seconds on a GPU with a display attached. Exceeding this time limit causes a launch failure reported through the CUDA driver or the CUDA runtime. GPUs without a display attached are not subject to the 5 second run time restriction. For this reason it is recommeded that CUDA is run on a GPU that is NOT attached to an X display.”

I really want to get started developing code, and will undoubtedly have need for more than 5s
of compute time with the card. I only occasionally need to sit at the console to do software installs, etc., at the console.

QUESTION: Can this timeout issue be avoided simply by removing the monitor cable from the back of the card? I would just login remotely and develop/run code, never having to sit at the console. Better yet, since I don’t login in at the console and therefore don’t execute startx, does this again preclude the 5s timeout, i.e., can the cable be left attached as long as X is not started
to avoid the timeout?

If any NVIDIA folks read this, it would be VERY HELPFUL to the community to precisely clarify this issue in future release notes!

             Many Thanks in Advance,     Hans

If you do not start X, you will not have the 5s problem. I believe it is in he FAQ, it is sticky on one of the forums.

Yep, just run in the linux console without X and you can run kernels without the 5s limitation.

Although, I myself have never found the limitation to get in my way. The most compute intensive kernel I’ve written to date still finishes in 50ms: a long ways off from the 5s watchdog timeout.

I’m currently running cuda on a GTX 260 which is also used to drive my display (through X11). When my kernel code gets into an infinite loop, X11 hangs and the only way I can recover is to kill the X11 server and let GDM start a fresh one. Actually, even killing the X11 server requires about 5 minutes of the screen frozen before X11 is restarted…

Is there any way to activate the watchdog timer? (Should I compile with dbg=1?)

I think this is a known bug (I’ve filed a very similar one involving the watchdog timer and X hanging). Can you give me your machine specs and tell me how long it took for your X server to be killed?

Here are the system’s specs:

CPU: Intel Core 2 duo (E6400) @ 2.13 GHz
RAM: 3 Gb (2 x 1 Gb, 2 x 512 Mb)
HP dc7700cmt (with power supplied replaced by a 650W, and original PCIe 16x card removed)
NVIDIA GTX 260 (fitted in a PCIe 16x slot)
Intel Q965 Express chipset

The X11 hangs for around 1 min. before being restarted from the time I kill it.