High-Performance Computing Environment with Cooperation Between Supercomputer and Cloud

Due to the recent popularization of machine learning, such a deep reinforcement learning as AlphaGO has advanced to analyze large-scale data and is attracting great attention. In deep reinforcement learning, users evaluate many functions in large-scale computer environments, including supercomputer and cloud systems. Cloud services can provide computer resources based on the scale of the computer environment desired by users. On the other hand, in conventional large-scale computer environment that only consists of CPUs or GPUs, the processing time greatly increases according to the scale of the calculation processing. In this paper, we propose a high-performance computing environment for deep reinforcement learning that links supercomputer and cloud systems. Our proposed system can construct a high-performance computing environment based on the scale of the computing process by the cooperation of the supercomputing and cloud systems with short physical distance and short network distance. In our evaluation of deep reinforcement learning using our proposed system, we confirmed that computer resources can be effectively used by allocating suitable processing for the supercomputer and the cloud according to the usage situations of the CPU, the GPU, and the memory.