Characterizing Power Usage in Zero Reserved Power Data Centers to Enable Planned Maintenance

The prevalent convention in data centers (DCs) has been to provide reserved power to handle any failures of power device. To reduce the environmental impacts and to increase resource utilization, recently, DC providers have started shifting towards zero reserved power (ZRP), where a DC can safely consume power up to 100% of the capacity of the power sources. As DCs have only recently started shifting to ZRP, little is known about the trends in the power usage of ZRP DCs. In this paper, we fill this gap and present, for the first time, a thorough characterization of the power usage trends in operational ZRP DCs of a large data center provider (LDCP). We analyze various spatial, temporal, organizational, and statistical factors that impact the power usage in LDCP's ZRP DCs. Based on our observations, we further develop an approach that accurately predicts future time windows during which the power usage can be expected to stay below a given threshold. Such prediction is required to enable planned maintenance of power sources in ZRP DCs because as ZRP DCs do not have reserved power, when one power source is disconnected for maintenance, the power usage of the running workloads should not exceed the capacity of the remaining power sources. Otherwise, the ZRP DC could go into a blackout. Our evaluations on data from ZRP DCs show that the proposed prediction approach successfully identifies 91% of all the 1-hour time-windows that are safe for maintenance up to 14 days in the future.