Online energy conservation scheduling for geo-distributed data centers with hybrid data-driven and knowledge-driven approach

With the surge in demand for cloud computing services, the problems of high operating cost of data center cluster (DCC) are becoming increasingly challenging, while the strong intermittency of renewable energy source and the randomness of interactive workload make DCC subject to the uncertainty of supply and demand. Therefore, how to maximize the economy of DCC and improve its renewable energy source accommodation rate during online scheduling is particularly important. Based on the spatial-temporal migration mechanism and the complementary distribution of renewable energy source in multiple regions, this paper maximizes the accommodation rate of renewable energy source and reduces the operating cost by flexibly migrating workload in DCC. Firstly, a DCC load balancing model is established and formulated into a deep reinforcement learning framework considering the coupling constraints of energy storage system and batch workload. Then, the advanced deep reinforcement learning TD3 algorithm is adopted for online schedule of DCC. Finally, to improve the convergence characteristics, the ITD3 algorithm is proposed in combination with the imitation learning Dagger algorithm, which achieves the transcendence of expert experience. Results using Google DCC data on the 33-nodes testing system show that the proposed ITD3 online scheduling method achieves near-optimal scheduling beyond expert experience, significantly reduces the operating cost through the spatial-temporal workload migration strategy, and realizes the coordinated scheduling of equipment in DCC.