MultiGreen: Cost-Minimizing Multi-source Datacenter Power Supply with Online Control

Abstract

Faced by soaring power cost, large footprint of carbon emis- sion and unpredictable power outage, more and more mod- ern Cloud Service Providers (CSPs) begin to mitigate these challenges by equipping their Datacenter Power Supply Sys- tem (DPSS) with multiple sources: (1) smart grid with time- varying electricity prices, (2) uninterrupted power supply (UPS) of finite capacity, and (3) intermittent green or re- newable energy. It remains a significant challenge how to operate among multiple power supply sources in a comple- mentary manner, to deliver reliable energy to datacenter users over time, while minimizing a CSP’s operational cost over the long run. This paper proposes an efficient, online control algorithm for DPSS, called MultiGreen. MultiGreen is based on an innovative two-timescale Lyapunov optimiza- tion technique. Without requiring a priori knowledge of system statistics, MultiGreen allows CSPs to make online decisions on purchasing grid energy at two time scales (in the long-term market and in the real-time market), leveraging renewable energy, and opportunistically charging and dis- charging UPS, in order to fully leverage the available green energy and low electricity prices at times for minimum op- erational cost. Our detailed analysis and trace-driven sim- ulations based on one-month real-world data have demon- strated the optimality (in terms of the tradeoff between min- imization of DPSS operational cost and satisfaction of data- center availability) and stability (performance guarantee in cases of fluctuating energy demand and supply) of Multi- Green.