学术文献库

An electricity bill of a data center (DC) is determined not only by how much energy the DC consumes, but especially by how the consumed energy is spread over time during the billing cycle. More specifically, these electricity costs are essentially made up of two major charges: Energy Charge, a cost based on the amount of consumed energy (in kWh), and Peak Charge, a cost based on the maximum power (in kW) requested during the billing cycle. The latter charge component is forced to encourage DCs to balance and regulate their power demands over the billing cycle, allowing the utility company to manage congestion without increasing supply. This billing model has thus called for the development of peak power shaving approaches that reduce costs by smoothing peak power demands over the billing cycle to minimize the Peak Charge component. In this paper, we investigate peak power shaving approaches, and begin by using Google data traces to quantify and provide a real sense of how much electricity cost reduction can peak power demand shaving achieve on a Google DC cluster. We then discuss why peak power shaving is well-suited for reducing electricity costs of DCs, and describe two commonly used peak shaving approaches, namely energy storage and workload modulation. We finally identify and describe key research problems that remain unsolved and require further investigation.