Slashdot Mirror
Intel Shows Data Centers Can Get By (Mostly) With Little AC
Ted Samson IW writes "InfoWorld reports on an experiment in air economization, aka 'free cooling,' conducted by Intel. For 10 months, the chipmaker had 500 production servers, working at 90 percent utilization, cooled almost exclusively by outside air at a facility in New Mexico. Only when the temperature exceeded 90 degrees Fahrenheit did they crank on some artificial air conditioning. Intel did very little to address air-born contaminants and dust, and nothing at all to deal with fluctuating humidity. The result: a slightly higher failure rate — around 0.6 percent more — among the air-cooled servers compared to those in the company's main datacenter — and a potential savings of $2.87 million per year in a 10MW datacenter using free cooling over traditional cooling."
Makes sense to me. The most efficent places to store data centers is in the northern US or Canada where you have sub-zero temperatures from November - March and ranging between 0-15 in April/May and Sept/Oct and the rest of the year 20-30+ (Celcius of course) With these lower temperatures they could run a data center entirely off outside air from September - May each year. Put a heppa filter in between to scrub out dirt and dust and vola, o'natural cooling solutions
I asked the president of an engineering firm that I work for about this. He ships racks of boxes, each holding DSP boards on backplanes, each backplane has it's own PSU.
When I asked him why he doesn't just have one or two -big- power supplies in the unit, he said that he tried that, but the cost of the non-standard PSU was higher than all the ATX PSUs put together, and then some, and replacing the units when they eventually fail would be tricky, as opposed to just stocking more ATX PSUs.
I agree that it's a good idea, but until there's enough volume of large multi-output PSUs shipping, the cost of manufacture makes the product unworkable (unless you think big-picture and want to spend more up front for power savings over the whole unit's life).
Generally, the people who use the hardware aren't the ones building it, and buyers usually go for the lowest bid.
"Sometimes, I think Trent just needs a cup of hot chocolate and a blankie." -Tori Amos on Nine Inch Nails
So this study is not actually useful for people who need to build data centers in more humid places then new mexico
Humidity only really matters for two reason - If too low, you get a lot of static buildup, and if too high, you get condensation.
Condensation only tends to happen on objects cooler than ambient, which doesn't really apply to running servers. Static matters a lot more, but you can raise humidity a lot cheaper than you can lower it, so, not as much of an issue there.
And as a bonus, more humid air can carry away more heat than the same volume of less humid air.
One benefit to going DC is that you can wire your battery modules directly into the DC distribution grid for the CPUs (with appropriate charge and cutover circuits), and forgo the inefficiencies in converting AC to DC at the UPS, and then back out again, only to convert the AC back to DC at the CPU.
Having multiple of a commonly used voltage used in renewable energy also helps if, for example, you want to feed your datacenter directly from say wind or solar, in addition to a set of AC to DC converter.
Sun is also running a comparable experiment with Belgacom and allows you to log in to a live interface to view stats on in- and outlet temperatures and more at http://wikis.sun.com/display/freeaircooling/Free+Air+Cooling+Proof+of+Concept For more details and analysis see http://www.datacenterknowledge.com/archives/2008/09/18/intel-servers-do-fine-with-outside-air/ or http://securityandthe.net/2008/09/18/intel-sees-the-future-of-datacenters-and-it-does-not-include-airconditioning/
DC Knowledge also has a nice video of this experiment at http://www.datacenterknowledge.com/archives/2008/09/18/video-intels-air-side-economization-test/
Having very large PSUs is a pain in the ass. Failures tend to be catastrophic and dangerous. They're more expensive to build and maintain. (think basic economy of scale problems) They also may not be any more efficient than distributed conversion. You also tend to distribute much lower voltages with DC than you do with AC. (240vac vs 48vdc) This gives very high amperages which requires much thicker wiring. Copper is EXPENSIVE right now, this makes it a big factor in the cap-ex of building a new DC.
This is why a lot of work is going into improving the efficiency of commodity power supplies. Groups like 80plus.org are doing great things.
Also some other links:
http://www.treehugger.com/files/2007/07/secret_efficien.php
http://services.google.com/blog_resources/PSU_white_paper.pdf
We're talking less than a minute needed. In the end they couldn't use the several large and expensive flywheels because they could not provide power long enough.
If you're powering your whole data center 'for a while' with these... you must have very few servers (like a handful).
Experience teaches only the teachable. -AH