A deep learning container cloud for GPU resources

With the development of deep learning, deep learning framework has become an important tool for the deep neural network development. The framework greatly shortens the network construction and computing time, and its powerful computing ability comes from GPU. But It is an important issue that how to effectively allocate and use GPU resources in heterogeneous cluster among many frameworks. In this paper, we propose a Deep Learning Container Cloud (DLC) architecture for GPU resources specifically. With the characteristics of easy deployment and easy migration, the frameworks can be deployed on heterogeneous cluster in the form of container, and the GPU driver and container can be decoupled according to NVIDIA-docker volume. The DLC provides services in the form of the MESOS framework. After obtaining resources through scheduler, a deep learning framework is created quickly to meet the requirements. DLC will loads the specified GPU resource and the corresponding runtime library to achieve the rapid creation of a deep learning environment with specific version. In addition, this paper proposes an allocation algorithm based on GPU topology. DLC constructs the topo-tree by analyzing the GPU topology structure in agent node, and on this basis, assigns the GPU with the P2P function within the node. Our experiment shows that the use of P2P data transmission in containers will significantly increase bandwidth. It is of great significance for promoting the development of deep learning.