Energy-Cost-Driven Scheduling Scheme for Data Centers Incorporating Job Dependency Constraints

The digital transformation drives data centers (DCs) to expand rapidly in both scale and quantity. Consequently, the power consumption and energy procurement cost of DCs continue to increase. Generally speaking, the power consumption of a DC is mainly determined by its job execution pattern, which is affected by complex factors such as the sequential dependency constraints and quality of service (QoS) requirements among heterogeneous jobs. However, most of the existing DC energy management works either neglect or simplify those factors. Therefore, this paper proposes a large-scale job scheduling scheme driven by both energy procurement costs and QoS requirements of DCs. It incorporates a graph neural network to capture the intricate characteristics of job dependency represented by directed acyclic graphs. Meanwhile, it utilizes a policy-based reinforcement learning algorithm to generate scheduling decisions. A scalable end-to-end learning approach is employed to solve this problem. The numerical results indicate the proposed scheme can achieve energy procurement cost reduction while meeting the complex job scheduling constraints under varied job quantities and arrival patterns.