My quad-core tower suddenly feels wimpy What does it take to power the world's most popular search engine? Lots of CPU cycles. Which is just what Google's new data centers provide. No one is talking, thanks to Google's tight NDA policy, but with satellite imagery and some deft estimation we can figure it out.
Powering a warehouse-sized computer? OK, how big is a warehouse-sized computer? Google wrote a paper about it (see Google’s warehouse-size power problem), but they were vague on details like the number of processors.
Take Google's new data center on the banks of the Columbia river in The Dalles, Oregon. The area had been hurting since the aluminum smelters shut down after power went over $30 per megawatt. Pricing is complex, but it looks to be about $45/mw today.
If that sounds cheaper than what you pay, it is. A lot cheaper. That's why Google, Microsoft and Yahoo are building data centers in the hydropower-rich Columbia river basin. The world's best windsurfing is just a bonus.
Sizing the Google data centers I can't republish copyrighted photos, but using aerial and satellite photos, Google Maps (of course!) and some mapping software, I determined that each of the two identical buildings contain about 100,000 square feet of Class A data center floor space, not including the covered loading docks. Here's a picture of the cooling end of one data center.
Counting racks Each 19" rack is about 24" wide and 30" deep (I saw a picture of one at the Seattle Conference on Scalability) or 5 square feet. Fire codes require a 3 foot hallway in front of each rack so the total footprint is 11 square feet - just over 1 square meter in the metric system. That means each data center has room for 100,000/11 or 9,090 racks.
There are also requirements for gigE fabric racks, power conversion and conditioning equipment, cooling and air-handling equipment. Let's say that 10% of the racks are used for non-server equipment. That leaves 8,180 racks.
According to published reports, each rack contains 80 cores in 40 dual-core Intel chips. 8,180 x 80 = 654,400 processor cores. And don't forget there are 2 of those buildings on the Oregon site. Over 1.3 MILLION cores in that one facility alone.
The Storage Bits take As I noted in Google's three rules Google embraces cheap computers. They optimize their workloads for parallel computing, so the number of processors is more important than their speed.
And they certainly have the numbers!
Comments welcome, of course.