3 Replies Latest reply on Mar 7, 2012 2:42 PM by yurtesen

    Watchdog timer under Linux?

    rahulgarg

      I was wondering if there are any watchdog timers for OpenCL kernels under Linux? I have actually managed to run some large kernels (40 seconds) on AMD cards just fine. The system did become unresponsive while the kernel ran, but once it finished everything was fine.

       

      However, I wonder if problems can be expected for larger kernels? Is there any official recommendation/guideline about the issue? Under Nvidia implementation, OpenCL kernels cannot run for more than 10 seconds if the GPU is driving a display, which is very unfortunate.

        • Re: Watchdog timer under Linux?
          yurtesen

          No, there is no such limitation as far as I know. But it would actually be good to have the timeout, because otherwise you can lock yourself out of the console. But I didnt go much beyond 2 minutes either.

           

          It is a 'feature' of AMD GPUs which people complain about a lot... If your program crashes, you will have to reboot the whole machine and loose all unsaved work. I guess that explains why

          1 of 1 people found this helpful
            • Re: Watchdog timer under Linux?
              rahulgarg

              Thanks for the info. Well, I would prefer not having a watchdog as I want to run long kernels. 10 seconds is way too short for some of the computations I want to do. I do dislike the hard crashes that can result but for my purposes better than being limited to 10 seconds.

                • Re: Watchdog timer under Linux?
                  yurtesen

                  Well, you can run kernels on nvidia platform which take longer than 10 seconds even when X is up and running, and it requires no change to the programs whatsoever.  The only thing you loose is crashing your machine and having to reboot. So much convenience will only spoil you