Abstract
Container technology has shown its advantages, such as lightweight and low latency, over traditional virtual machines for the applications in edge computing. However, the current container design is not for system virtualization and therefore lacks of many important features. In this paper, we present an enhanced version of Docker, the state-of-art container, to support multiple GPU device drivers. We proposed the idea of kernel assisted virtualization for effective container implementation. More specifically, we added additional modual namespace in host OS so that different containers can utilize its own modules. We demonstrated our implementation on containerized CUDA environment. Experiments show that our enhanced container has close to native performance, but with more flexibility for resource sharing.