Since NVIDIA has released its GPU kernel modules. Sharing GPU between different containers has a easier way than before. The project is still base on GaiaGPU: Sharing GPUs in Container Clouds
mkdir -p build
cd build && cmake ..
make
- set LD_PRELOAD environemnt
export LD_PRELOAD=libcuda_hook.so
1.1 for gpu memory limitation:
export CUDA_MEM_LIMIT=<device index>=<memory limitation>
1.2 for sm utilization limitation:
export CUDA_CORE_LIMIT=<device index>=<core limitation>
If you have sm utilization limit enabled, you must start a server_monitor to control the utilization ./server_monitor <device idx> <cgroup id> <core limit>