A doctoral candidate in mechanical engineering, Song studies data centers — the nervous systems of our digital lives — and ways to cool them more efficiently.
“I like computers, math and engineering, but it’s not just about technology. It’s more about people,” Song says. “It’s the demand of people who are used to this digital life today that drives the need for the growth of data centers.”
As indispensable as they are, data centers can also be energy hogs. The tens of thousands of data centers across the country used roughly 76 billion kilowatt-hours of energy in 2010, or about 2 percent of the nation’s entire electricity consumption, according to the New York Times. Much of that energy goes into cooling, or thermal management. The processors in a data center emit heat, and overheating can lead to slower processing or even system failure.
Song’s research is part of a growing effort looking at how companies can save money and energy by cooling their data centers more efficiently.
Right now, the standard method used to control energy usage in data centers involves using large-scale computational modeling and costly measurements of the temperatures throughout the data center for specific room configurations. It’s a precise method, but processing all the information and making changes to improve cooling performance can take hours or even days.
Song thinks that’s too long, especially when changes in temperature can mean drastic changes in cost. If a company can maintain healthy operating conditions in its data center, it won’t need to spend as much cooling down overheated processors. If companies can raise the temperature of the cool air they send into data centers by just four degrees Celsius (from 18 to 22 degrees), they use on average 30 percent less energy per minute.
Song is working on a greener solution, which combines smarter scientific modeling, simpler measurement requirements and much, much faster monitoring — so that cooling needs can be diagnosed and adjusted within minutes, rather than days.
The trick, Song thinks, is to develop a smart compact model that can not only provide more effective guidance to heat sensors setup, but can also learn from them via real-time feedback and predict the stuff that cannot be measured.
Because of the wide range of data center configurations and sizes, there is no one-size-fits-all model for thermal management, notes Bruce Murray, professor of mechanical engineering and Song’s advisor.
“Song has shown a lot of initiatives to develop a broad spectrum of compact models,” Murray said. “He’s already published four peer-reviewed journal articles.”