While testing my kernel on extremely big data sets I’ve faced the following error: cudaErrorLaunchTimeout. This happens after about 14 seconds after the launch. If to reduce the data set so the kernel would cope with it in about 12-13 seconds then all goes fine.
My video card IS NOT connected to the monitor, no watch dogs and other similar things should happen.
Also, I’ve noticed that this error happens with CUDA 2.0 but not with CUDA 1.1. In cuda 1.1 there was another unpleasant thing: infinite kernel could hang the system up, looks like in cuda 2.0 there is one additional monitor-independent watch dog that terminates the kernel after about 15 seconds. It is good for faulty kernels but it is a disaster for kernels that normally work for a long time.
Programming guide does not inform about this new feature, also, no info on how to switch this extra watch dog off.
Does anybody know how to prevent forced kernel termination ? Nvidia guys, the question goes primarily to you …
Thanks in advance!