Hello
We have some issues running multiple instances of our CNN-pipeline in separated Docker containers. The Linux randomly enters a "Zombie" state (no reaction to anything, even to the "Magic SysRq keys" and so much network traffic that every device on the next switch is down). (see https://software.intel.com/en-us/forums/opencl/topic/804936).
As the problems started when updating the OpenCL driver and only occurred (until now) when multiple Docker containers with OpenCL were running, there might be a multiprocessing issue with the OpenCL driver when used within separated Docker containers (some tests running native (not in Docker containers) caused no problems, but this might just be a coincidence as the tests where only done for a few hours).
So our question is how multiprocessing is handled by the driver and if we need to add more than just the /dev/dri device to the Docker containers to avoid race conditions between the different docker containers.
Greetings,
Thomas