Privacy-Preserving Energy Sharing Among Cloud Service Providers via Collaborative Job Scheduling
Yimeng Sun, Zhaohao Ding, Yuejun Yan, Zhaoyang Wang, Payman Dehghanian, Wei‐Jen Lee
Abstract
With the growing digitalization of the economy and society, the scale of energy consumption in cloud computing is continuously expanding. Leveraging the flexible scheduling characteristics of computing jobs, data centers operated by different cloud service providers can reduce their energy costs by spatiotemporally shifting jobs to periods and locations with lower energy prices. However, privacy concerns on critical operation information hinder such collaboration among different cloud service providers. In this paper, we propose a privacy-preserving federated reinforcement learning scheme for collaborative job scheduling to enable energy sharing among cloud service providers. First, we establish the collaborative energy management model via job transfer and computing resource allocation as a decentralized partially observable Markov decision process. Then, we develop a personalized federated reinforcement learning approach under a decentralized training with decentralized execution framework, where decisions are made adaptive to the heterogeneous environments of different cloud service providers while protecting their operation privacy. Finally, the real-world traces from Alibaba are used to illustrate and verify the effectiveness of the proposed scheme.