Risk-Aware Energy Management of Data Center via Job Scheduling

With the rapid acceleration of digitalization, the energy consumption of data centers has increased consistently. Given latency-tolerant computing jobs, data centers can reduce operation costs by delaying job execution during peak electricity price periods. However, fluctuations in electricity prices and the unpredictability of job arrivals expose data centers to the risk of high operation costs, which presents a significant challenge for job scheduling. To address these challenges, we propose a reinforcement learning algorithm for risk-aware real-time job scheduling that minimizes energy costs in data centers. First, we formulate the job scheduling problem as a Markov decision process in an uncertain environment. Subsequently, we adopt Conditional Value at Risk to quantify the risk of high operation costs and incorporate it as a constraint in the reinforcement learning model to enable risk-aware scheduling. Finally, we propose a policy gradient algorithm combined with Monte Carlo sampling to develop a solution whose convergence is theoretically proven. The simulation results based on real-world data from Alibaba demonstrate the effectiveness of the proposed approach.