A Deep Reinforcement Learning based VM Scheduling Strategy Decreasing Data Center Communication costs

With the rapid development of cloud computing, more and more complicated services are deployed on the cloud. The frequent communication between services places a great burden on the network of the data center and leads to a decline in the quality of service. Researches have shown that the costs of communication between services vary with the placements of virtual machines on which services are running. Therefore, the virtual machine strategy should be optimized by migrating virtual machines with high communication frequency closer. Based on fat-tree network topology, a Deep Q network based online virtual machines scheduling strategy(DQN-VMS) is proposed. The objective is to obtain an optimized virtual machine scheduling strategy that brings down communication costs while guaranteeing the quality of service. The calculation model for communication costs is built for the fat-tree network topology. A reward function that takes into account the communication costs and Service Level Agreements violation(SLAV) rate is designed. The experiment shows that the proposed strategy performs about 26.81% and 40.55% better in comparison with two virtual machine scheduling strategies based on graph partition algorithm and greedy algorithm, which are commonly used in existing data centers.