Reinforcement Learning for Smart Data Center

Abstract: Data centers are the key infrastructure backbone powering most IT services worldwide. From text messages, streaming services and voice calls to large organizational services and corporate transactions are all made possible because of data centers. As those services keep growing the scale of the data centers worldwide also keeps growing. Despite the immense usefulness data centers have, there is a serious problem with energy efficiency when it comes to power consumption and electricity demands. Data centers require a large amount of energy, partly because of inefficient scheduling and resource utilization. Associated with this huge energy demand there is always a higher cost. There have been studies to improve the energy efficiency of data centers which some of them has proven to be effective. In this research, the Q-learning algorithm is used to address the problem of inefficient scheduling in a data center with the ultimate goal of reducing the energy and power consumption of a data center. By monitoring the job pool, the reinforcement learning algorithm will make a decision whether to increase or decrease virtual machines. The algorithm will therefore keep a certain task queue length to obtain maximum energy efficiency without harming the task completion time. After the experiments are conducted, results show that the Q-learning algorithm saves up to 33% of energy consumption when compared to traditional scheduling algorithms by keeping the task processing speed at an equal pace. The algorithm effectively saves energy without affecting the performance of the data center.