SAN FRANCISCO – Intel and AMD just got a nasty smackdown from the person who run’s Facebook’s datacenters, saying the chips don’t deliver on their promises.
Jonathan Heiliger, vice president of technical operations for Facebook, was being interview by GigaOm Network founder Om Malik here at the GigaOm Structure 09 conference. Malik asked him about unexpected problems in managing the fast-growing company’s datacenters.
Reps from Intel (NASDQ: INTC) and AMD (NYSE: AMD), both running panels and present at the show, must have clenched their teeth when they heard this:
“The biggest thing … was less-than-anticipated performance gains from new microarchitectures, so new CPUs from guys like Intel and AMD. The performance gains they’re touting in the press, we’re not seeing in our applications,” Heiliger told the audience.
He didn’t let the tier one server vendors off any easier. “I’m not sure if I’m embarrassed or pleased for OEM vendors in the audience, but you guys don’t get it. To build servers for a company like Facebook and Amazon, and other people who are operating fairly homogeneous applications, they have to be cheap and super power efficient,” said Heiliger.
He added he’s not sure why hardware vendors fail at the job, but thinks customers need to step up and apply pressure. “”Perhaps in the coming months we’ll see more collaboration with people running small clusters and large clusters,” he said.
AMD spokesman Phil Hughes said the company would never recommend a customer make a purchase based solely on benchmarks.
“One of the things to keep in mind is that performance claims are based on industry benchmarks,” Hughes told InternetNews.com. “We always encourage customers to run their own applications and test performance.”
Hughes also said AMD’s recently launched Shanghai EE is a 40-watt processor that is “really targeted at that cloud environment.”
Intel could not be reached for comment by press time.
The Facebook Usernames Challenge
Prior to the criticism, Heiliger talked about what it was like for Facebook to launch its Facebook Usernames project. That involved an awful lot of creativity on the firm’s part to avoid a crush that usually brings down lesser servers.
From a tech standpoint, Facebook created a segmented memecached pool to handle queries to ease the load on servers. The company segmented its load between east and west coast servers with some standby hardware in case of failure.
The company then isolated the username function from the rest of the site so that maximum server payload could be given to handling the crush of requests. It turned off all other non-critical functionality that normally runs on all Facebook pages.
“When you come to Facebook, you’re not just looking at your data. You’re looking at your data and your friends’ data and their friends’ data,” said Heiliger. By isolating the username servers, we were able to get through the early stampede.
