As data centers have increased in size, the energy consumed in powering and cooling server clusters has become a significant fraction of the total cost of ownership of operating a data center. As a result, there is an increasing trend towards reducing the energy and carbon footprint of data centers as well as in using green sources of energy such as solar and wind power to run data centers. In this project, we will conduct fundamental research on system support for server clusters that are designed to run using variable power, such as that from renewable sources. We will design a new supply-side technique to manage the energy and power footprint of a data center server cluster and to adapt this footprint to supply-side fluctuations. Our enables compute and data nodes in a server cluster to be duty-cycled so that their power and energy footprint can be gracefully adapted to a varying supply curve. We will design study how distributed storage and server virtualization can be adapted to exploit such duty-cycling of servers to reduce energy use. The broader impacts of this work include a new graduate seminar course on Sustainable Computing, summer REU projects to include undergraduate students, especially from underrepresented groups, in this research, and outreach to K-12 teachers and community college faculty via tutorials in sustainability.
Data centers are large server farms that provide compute and storage resources to online applications. Since data centers can house a large number of servers, they consume significant amounts of energy. The design of energy-efficient data centers that have a low energy footprint and lower energy costs has emerged as an important topic for industry. Modern data centers have also begun to use renewable sources such as solar and wind to power their servers. However renewable sources are known to be intermittent in nature and the energy supply from these sources fluctuate over time, resulting in frequent shortfalls to meet the current demand. Our project yielded a new technique called blinking that enables a data center to duty cycle ("blink") its servers during periods of energy shortfalls, allowing it reduce its total energy usage. Blinking is an alternative to turning off servers and is a more attractive option since servers remain active, albeit in duty cycled mode, allowing for applications on these servers to continue to run and data stored on these servers to remain available. We designed a prototype server cluster that implements blinking and designed a blink-aware storage layer for higher level applications. We demonstrated the benefits of the blinking approach by designing blink-aware versions of memory cache and disk-based video cache applications. Finally, we studied the use of blinking in distributed data centers and large-scale distributed systems such as content distribution networks. Overall, our project demonstrated that a blinking-based approach is both feasible and an attractive for data centers and server clusters that use renewable power as one of their energy source. Broader impacts of this project included training of doctoral graduate students, participation in the activies of real green data center built by a local consortium of universities, and incorporation of green computing topics into a graduate course.