Behind the cat photos and Oscar opinions of one billion humans is Facebook’s burgeoning network of data centers. As complex as managing people’s personal data and sending them ads is, the company tries to buy from a relatively small number of suppliers, whom it keeps on a short leash.
“In computing at this scale, the data center is a factory floor,” says Frank Frankovsky, vice president of hardware design and supply chain at Facebook. “We try to keep things simple.”
Equipping the data factory works like this: Mr. Frankovsky and his team of about 25 people design five types of computers to run most of Facebook. These are Web page servers, database computers, data storage systems, news feed servers and something called memcache servers, which speed overall performance. They then figure out what it costs to build each kind of machine, and buy from a supplier, known in the industry as an O.E.M., or original equipment manufacturer, if that maker can equal or better Facebook’s performance, at a price within 5 percent of what Facebook figures is its own cost.
“When we first said we’d build our own machines three years ago, the O.E.M.’s didn’t believe it,” Mr. Frankovsky says. In the last release of Intel’s mass-market semiconductor for servers, he says, “we were internally ready three months before everyone else.”
The speed with which Facebook made the new computers, he said, was a result more of the incumbent manufacturers’ older design, manufacturing and sales processes than anything novel that Facebook did.
The Facebook purchasing model has been disruptive to many traditional suppliers, like Dell and Hewlett-Packard. Initially. at least, it seems to have helped ultra-cheap Asian manufacturers like Quanta Computer. Quanta, which makes the laptops sold by Apple, H.P. and others, began selling Facebook and others shrink-wrapped racks of servers instead of individual boxes. It now now supplies 80 percent of Facebook’s Web servers, Mr. Frankovsky says.
H.P. has responded by taking a more active role in the Open Compute Project, an open source project started by Facebook to build big data centers that are more energy efficient. A Web server from H.P., code-named “Coyote,” is now being tested at Facebook. Its “Moonshot” server, built for big data centers with the type of chips usually in cellphones, should also be a candidate when it comes out later this year.
While H.P. is selling substantially fewer computers to Facebook than it used to, Mr. Frankovsky credits it for reacting to a trend. “This industry is shifting from supplying lots of small and medium-sized businesses with lots of different computer requirements, towards clouds,” he says. By selling new products to Facebook with little profit, he says, “they get low margins, but they understand what they could design to sell in the long term.”
It works out nicely for Facebook that way, too.
As painful and testing as this customer-led price demand may be, at least some suppliers say it is a healthy system. “The burden falls on the tech guys to keep their product from becoming a commodity,” says David Yen, senior vice president of the data center technology group at Cisco Systems. “Facebook is pushing new standards that force companies to innovate.”
In addition to Mr. Frankovsky, the Open Compute Project’s board includes Andy Bechtolshiem, a founder of the innovative server maker Arista Networks; Jason Waxman, the head of high-density computing at Intel; and representatives from Rackspace Hosting and Goldman Sachs, which consumes a lot of computing power.
What is missing from the group, it seems, is much of a presence from Amazon Web Services or anything at all from Google. Google, which builds its own machines inside perhaps the biggest data centers of all, has played alone, to a point of not patenting its inventions for fear of disclosing its methods and giving up secrets around power consumption and data transfer speed.