In an InfoWorld (Feb 26, 2007, p. 29) column titled Gauging Net Consumption, David Margulis writes about about a report estimating that data centers account for about 1.2 percent of U.S. Gross Domestic Energy Consumption. He thinks that household computers might be more of a problem, because there are so many more PCs (75 million) in homes, than there are servers (10 million) in datacenters. He suggests that the PCs should come with a yellow energy guide, just like those attached to appliances, estimating the energy use and cost of the PC during a year. I think it’s a great idea, but the devil is in the details.
When we buy appliances, we have a pretty intuitive feeling for their capabilities — how much food they can hold or how much laundry they can wash per load. But PCs are a different story. More gigahertz is better, right? More RAM is better, right? These all drive up the energy use, but appear to supply additional capabilities. We may never use all the PCI slots or fill all the drive bays, but we better get the bigger power supply just in case. This relentless pressure for more and bigger, coupled with FUD (Fear, Uncertainty and Doubt), needs some kind of countervailing force if we are to change behavior. Fortunately, we have another sticker to consider — the EPA fuel economy ratings on vehicles that allow comparison of energy use per some meaningful measure.
A similar measure could be developed for PCs, with a set of usage scenarios to create a few workflow profiles. Student, casual web surfing, and business consumption could then be measured for a computer with its operating system. Then a shopper could have benchmark to compare PCs. Similar work was done by NIST for appliance estimates.
But I think it’s more interesting and meaningful in the data center. Workload benchmarks are already well established for web, application, database and other servers. But these are generally oriented toward determining the ultimate capabilities of a system. Imagine if these workloads were then coupled with a scheduler so typical diurnal and weekly variations could be applied. Energy measurements could incorporate effects ranging from sleeping un-needed execution units to shutting down servers whose work can be handled by its partner in a load sharing arrangement.
These kinds of realistic scenarios should be the bread and butter of virtualization vendors.