Distributed battery control for peak power shaving in datacenters

Datacenters are large cyber-physical systems with continuous performance and power measurements, and real-time control decisions related to workload placement, cooling and power subsystems etc. In our work we focus on the non-ideal UPS system used to shave peak power demands. Our novel distributed battery control design has no performance impact, reduces the peak power needs, and accurately estimates and maximizes the battery lifetime. We demonstrate that models which do not take into account physical characteristics of batteries overestimate their lifetime by 2.4x. In contrast, our design is within 3.3% of the centralized battery control in terms of battery lifetime with 10x reduction in the communication costs, while shaving 23MWhrs/week of energy in a 10MW datacenter, equivalent to adding 8760 more servers at no additional power cost.