Cross-sector management for workload scheduling and thermal environment in data centers

Data centers serve as crucial infrastructure for the development of the digital economy, and maintaining their thermal environment is essential for ensuring that computing equipment can execute workload properly. Thermal environment management and workload scheduling are mutually influential; however, current research often treats them separately, overlooking their bidirectional coupling relationship and the potential for cross-sector collaboration. This work proposes a cross-sector management approach that integrates heterogeneous workload scheduling and thermal environment management into a unified Markov decision process. By introducing action masking to filter out invalid actions and leveraging proper orthogonal decomposition to accelerate thermal environment simulation, real-time decision is achieved through deep reinforcement learning. This method reduces cooling energy consumption while maintaining workload scheduling performance and thermal environment stability. Results of case study show that compared to common methods of Tetris, Balance, and only workload optimization, the proposed approach ensures workload scheduling performance with only a slight increase in ADR, enhances the alignment of server heat generation and cooling supply, and reduces cooling energy consumption by 28.9%. It also minimizes fluctuations in server outlet temperatures while keeping them below the maximum threshold, demonstrating the effectiveness of cross-sector optimization. Overall, the method and results of this study provide a novel perspective for the operation optimization of data centers.