GPU Method and System to manage and schedule GPU memory resource in Container-based virtualized environment

摘要

Provided are a GPU memory resource management and scheduling method and system in a container-based virtualization environment. The GPU memory resource management and scheduling method in a container-based virtualization environment comprises: a step of redirecting a command to an original docker with an option for using an NVIDIA GPU when the command is transmitted to a user defined nvidia-docker; a step of ConVGPU controlling and communicating with the NVIDIA GPU to insert a CUDA wrapper API module into a container; a step of the CUDA wrapper API module capturing some APIs to transmit the same to a GPU memory scheduler during a container running time; and a step of the GPU memory scheduler scheduling the use of the container and a GPU memory.