How smart batteries create efficient data centers
How smart batteries create efficient data centers
The electric grid experiences the same daily peak demand issues as our freeways, seeing pronounced surges during the critical parts of the day. This makes providing power more expensive because extra power plants have to be built to meet this peak demand.
Servers that provide the world's Internet content are also their busiest during this peak power period. And with the Internet consuming two percent of the world's energy and predicted to surpass the airline industry by 2020, the problem is only getting worse.
Furthermore, companies that have large IT deployments are being challenged by rapid expansion and rising energy costs.
For companies like Akamai that host their IT infrastructure in third-party collocation data centers, energy is typically priced based on the total supplied power in kilowatts (KW) charged at a fixed rate in $/KW, for example 50KW at $200/KW per month, similar to a fixed number of minutes for a mobile phone plan. While the supplied power is fixed, the power drawn by the servers vary with server activity which peaks and lulls daily. Energy costs could be reduced if the server peak-power demand, and hence the supplied power, can be lowered.
One of the innovative ways Akamai is looking at mitigating this challenge and reducing network operational costs is to reduce the grid-based power supplied to its servers during peak demand by supplementing with batteries and recharging the batteries at night when most energy consumers are asleep and power production is cheap and plentiful.
A recently published research paper by Ramesh Sitaraman, Akamai Fellow and UMass professor, written in collaboration with researchers at Penn State and BBN, evaluates using smart batteries in an Internet-scale distributed network such as the Akamai Network. The smart batteries that are placed inside a server or within a server rack detect when a server's power draw crosses a threshold level, as it becomes busier, and then supply battery power until the server power draw drops below the threshold. The batteries would recharge at night or whenever energy and server demand are low.
The effect would be that the peak power draw from the electricity grid of a cluster of servers would flatten and lessen, with peak demand shifting from the electricity grid to the batteries.
A trial of aggregate server power demand over a month in an Akamai New York data center showed smart batteries dropped the required power supply from 432 KW to 381 KW.
Despite the control intelligence of these batteries, the chemistry is conventional lead-acid, similar to a deep-cycle marine batteries. Lead-acid batteries are fully recyclable.
The capital cost of the battery is a function of the energy storage capacity of the battery in kilowatt-hours (kWh) which determines how long the battery can charge the server. This study uses reasonable estimates ranging from $100-300/kWh and a lifespan ranging from three-to-five years.
Batteries don’t have to supply power for very long to achieve a compelling peak power reduction benefit. Even a battery that can power a server for only five minutes, comparable to those batteries used in uninterruptable power supply (UPS) systems today, can provide a 7 percent power savings, while a larger 40-minute battery supply can save up to 14 percent.
Furthermore, most of the power savings are achievable with a small cycle rate of one full discharge/recharge cycle every three days which is conducive to a five-year battery life. Power savings improve demonstrably as servers get better at power management.
In the near term, cost savings are derived from a reduction in the power supplied to each server cluster -- power circuit size. Using smart batteries to reduce peak power loads would reduce the required power circuit size and, hence, the cost of each circuit. An alternate but equivalent way of viewing the benefit is that batteries allow more servers to be powered using the same power supply than before. (A key assumption is that the batteries could be fit into an existing server rack without taking up additional data center floor space and costs.)
Longer term as the smart grid matures and electric utilities implement time-of-day energy pricing, electricity prices will be more expensive during the day when demand is high and cheaper at night. Batteries make even more sense with time-of-day pricing since energy can be stored during the cheap off-peak hours and discharged to servers during the expensive peak hours. In addition, large-scale adoption of this power demand-shifting technique would increase demand for wind energy which typically has higher production capacity at night, and help slow the need to build power plants to meet peak energy demand.
Dr. Sitaraman emphasized "Our research helps establish batteries as a key part of the architecture of Internet-scale distributed networks. In addition to providing a distributed UPS function, batteries can also provide a cost reduction by decreasing the required power supply. Further, as servers become more energy-efficient and as batteries become cheaper and better, the case for batteries will only become more compelling."