Green, as we now know, is, indisputedly, the new black. Seems like you can't turn on the television or pick up a newspaper to read about the latest green initiative. Lots of people are talking.

Feng Zhao is doing something about it.

Zhao's Networked Embedded Computing group is showing a TechFest demo called Auto-Shift: Energy-Aware Server Provisioning, which addresses server resource management for Internet services, such as Live Messenger and Hotmail. Data centers for such services require potentially expensive decisions about how many computers to allocate and how those are deployed.

"No. 1," Zhao says, "you have to buy the servers. No. 2, once you buy a server, you have to manage it. And third, you have to have an infrastructure, such as power supply. In this particular study, we looked at the power usage of the servers that are running one of our largest Web services. If you look at the load as it varies over the course of the day, it peaks around noon and slows down around midnight. That clearly shows that not all the servers are needed all the time. Can we shut down some of the servers? Can we actually save energy?"

Initially, Zhao and colleagues thought the challenge might be straightforward. But complications arose. With a Web service such as Messenger, users are serviced continually. When servers are turned off, users lose access, and when they sign back in, they create a spike in the data-center load.

"We have developed a set of techniques that allow us to intelligently forecast when a load is going to spike," Zhao says, "and to provide a period of buffer time so that we gradually shut down machines--and only shut down machines when the load on the machine is below some small threshold. So we don’t have to migrate too many users, which, in turn, would create too many spikes in the load. The techniques that we have been developing are different ways of managing that process of gradually ramping down a server--and at the same time having some knowledge about what’s coming and actually prepare enough resources."

Some techniques, though, save more energy, some slightly less.

“If you want to be aggressive about saving energy, you are going to incur some penalties," Zhao says. "Maybe the response time will be longer. We’re looking at these tradeoffs. If you care more about energy, it will have to be at the price of a little more degraded user experience. Or you can provide a very good user experience, but the energy saving might be a little bit less. That’s a very interesting set of findings."

Auto-Shift is based on Messenger data over a period of 45 days running on machines used to emulate the Networked Embedded Computing's group's algorithm. Zhao says that algorithm can achieve energy savings of as much as 25 percent without significantly affecting the user experience.

"Energy savings are possible," he says, "but they require some careful scheduling, management, and predicting, and we have algorithmic smarts in the system that allows us to do this."

Work remains. The team needs to ensure sure that its algorithm works with other existing components. Other Internet services need investigation to see if the savings work in other scenarios. In the meantime, the server-provisioning work could prove even more beneficial.

"We also have all these sensors in the data centers," Zhao says. "Some of the machines work harder than others. If we can move the workload around, from hotspots to cool spots, the air conditioning doesn’t have to work as hard, because of the efficiency of cooling the hottest spots. If you move that workload and even out the temperature disparities, that means good energy savings. Incorporating environmental-sensor readings such as temperature and humidity, and couple that with smart scheduling and workload migration, and we believe we can even save more resources."

That sounds green, indeed--and economical, too. 

"What it translates to," Zhao concludes, "is that you use less power and that, with these smarts, we can figure out that maybe we don’t need to buy that many machines to start with, because we can do the same work, with very little difference in performance, and actually run it on a smaller set of machines. Reduce energy cost and reduce hardware investment in the first place--that would reduce service cost, reduce staffing, and reduce the space you need to build."