A group supported by the federal Environmental Protection Agency (EPA) has finalized a protocol measuring energy...
efficiency in 1U and 2U rack servers.
The standard differs little from the first draft put out this summer, according to Jonathan Koomey, a professor at Stanford University and staff scientist at the Lawrence Berkeley National Laboratory, who helped create the standard.
The basis is this: You want to buy servers and are worried about how much power they're going to suck up in your data center. So you look at different vendors' performance using this industrywide metric and use that to help decide.
The metric uses a power meter to measure frequency, voltage, power factor and total harmonic distortion compared to CPU utilization. What comes out the other end is a curve with power output on the y-axis and percentage of workload on the x-axis. Measurements for workload are in 10% increments from 0% to 100%, allowing data center managers to examine the curve depending on how busy they expect the server to be.
"We'll be talking to the manufacturers and customers, trying to get them to actually use the protocol," Koomey said. "We want manufacturers to start generating data in the format we have."
Major vendors are on board. Among the 16 people listed as authors of the metric are representatives from IBM, Hewlett-Packard Co. (HP), Sun Microsystems Inc., Dell Inc., Intel Corp. and Advanced Micro Devices Inc. (AMD). Other authors are from the Berkeley Lab, Stanford, Rumsey Engineers, the California Data Center Design Group and the Uptime Institute.
The group wants manufacturers to make the results available to users. There is also talk about having results from all vendors available on a central Web site, such as the EPA's, but the group hasn't decided on that yet. Koomey said having all information in one spot would be beneficial but could lead to a lag between vendors getting results and administrators of the Web site posting them online.
The hope is that end users will decide what servers they buy based in part on the machine's energy efficiency.
How long it takes for power and cooling issues in the data center to affect purchasing decisions is the question. Ronald Witt, an operating system (OS) administrator at Mortgage Guaranty Insurance Corp. in Milwaukee, said the metric won't have an immediate effect on what servers are bought, but down the road it likely will.
"It's something new in the 'Intel' world and may take time to gain traction in smaller data centers," he said. "In very large installs I think the metric would be used without question."
Lance Kekel, data center operations manager at a jewelry company in the Midwest, had similar thoughts; he said it would initially have "very little, to no bearing" on purchasing habits. But as power consumption becomes an all-consuming issue in his data center, things could change.
"Some of my budget requests will likely bring a little more visibility to this, but I think it will probably take a year or two for things to catch on," he said.
For some data center managers out there, server buying decisions are based more on applications than power consumption. Frank Walsh, data center manager at Lahey Clinic Medical Center in Burlington, Mass., said much of what is in his data center is driven by projects. A doctor may have received a grant or wants to do research on a particular topic. That research depends on a particular application, and the company that sells that application only does it on certain hardware.
Walsh does have some say over some of the server's attributes; for example, it has to have dual power supplies and dual processors. Beyond that, it's rare that he's choosing servers from a particular vendor.
"The project decides what the application is going to be and the application many times determines the hardware platform that it's going to run on," Walsh said. "I'm kind of in a position where they really get thrown on my dock, and we roll them in the door."
The server energy efficiency metric will be available on an EPA Energy Star program Web site and will focus on 1U and 2U rack servers.
Meanwhile, the Standard Performance Evaluation Corp. (SPEC), a nonprofit that measures server performance benchmarks, has formed a committee to develop a standard of its own and hopes to have one by the end of March. Industry leaders hope the two groups can come together to merge their two metrics into one and expand the energy efficiency standard beyond transaction-based performance benchmarks.
After that, the group will see if the metric gains traction. If vendors start publishing their results from the metric and end users decide what servers to buy based in part on the metric, it could expand. Vendors have already talked about developing a similar metric for blade servers, as well as rack servers larger than 4U.
That's where virtualization could make measurements tricky.
"This is kind of like the pilot," said Christian Belady, a technologist at HP who has represented the company at meetings about the standard. "The whole concern is that it's even more complicated with the introduction of virtualization. You can use virtualization to increase utilization or you can use power management. This really addresses the power management side."
"With virtualization, there's just a lot of complexity on how you would deal with the performance side of the equation," Koomey added.
Nonetheless, the group does want to move beyond 1U and 2U rack servers. Koomey said some manufacturers have intimated that blade servers -- which have been gaining popularity -- might be next on the radar screen. Again, though, there are more factors to consider. What counts as one server? Do you measure energy and performance when the chassis is full, when it's half full or when there's just one blade in it? And what's to be said for chassis that hold a varying number of blades depending on what vendor it comes from?
"We haven't thought that through and hashed it out with the manufacturers," Koomey said. "That's going to be one of the earlier things we work on."
There is also interest in the industry to look at the power consumption of storage equipment, but that will likely be tackled further down the road.