GreenDRL: Managing Green Data Centers Using Deep Reinforcement Learning
Wednesday, February 24, 2021, 02:00pm - 04:00pm
Speaker: Kuo Zhang
Location : Remote via Zoom
Thu D. Nguyen (Advisor)
Event Type: Qualifying Exam
Abstract: Managing data centers to maximize efficiency and sustainability is a complex and challenging problem. In this work, we explore the use of deep reinforcement learning (RL) to manage “green” data centers, bringing a robust approach to optimizing management systems for specific workload and datacenter characteristics. Specifically, we consider data centers that are partially powered by on-site generation of renewable energy and partially cooled by “free-cooling.” We design and evaluate GreenDRL, a hybrid system that combines a deep RL agent and a simple dispatcher to schedule a workload while jointly managing server power consumption and the cooling system to minimize cost. Our design addresses several important challenges, including scalability, robustness ,and effective learning in an environment comprising an enormous state/action space and multiple stochastic processes. Evaluation results show that GreenDRL is able to learn important principles such as delaying deferrable jobs to leverage variable generation of renewable (solar) energy , and avoiding the use of power-intensive cooling settings even at the expense of leaving some renewable energy unused. Simulation of a realistic (small) green data center running a workload containing a fraction of jobs deferrable by up to 12 hours shows that GreenDRL can reduce grid electricity by 26–57% compared to a baseline system, depending on outside temperatures and availability of renewable energy. Overall, our work shows that deep RL is a promising technique for building efficient management systems for green data centers.