The demands of the computing world have shifted in the last few years, leading hardware vendors to scramble to meet the needs of Web 2.0 companies and cloud computing providers. That shift was already well under way when the recession hit, but it wasn’t until then that the effects of it really started to take hold. As the enterprise buyers of high-end servers saw their IT budgets slashed, buyers of web servers such as Amazon and Facebook suddenly rose in importance.
But those Web 2.0 buyers’ needs are not being met, and they won’t be for quite some time, as the hardware industry tries to catch up with the innovations the web and cloud crowd have pioneered, according to Margaret Lewis, a product marketing director at AMD. She told me companies like Facebook and Amazon are “on the bleeding edge so you can’t make the optimizations in the hardware fast enough to suit them.”
Which is why in June Facebook VP of Technical Operations Jonathan Heiliger got on stage at our Structure 09 event and slammed chipmakers and server vendors for being behind the times. “The biggest thing (that) surprised us is…less-than-anticipated performance gains from new micro-architectures,” he said. “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. And we’re, literally in real time right now, trying to figure out why that is.”
Lewis doesn’t disagree, and says AMD is making changes. In 2007 year the company announced plans for something it calls lightweight profiling that will allow Java-based programming languages to run faster as well as boost the performance of software in multicore environments, but that’s not expected in actual chips until 2011 or later. AMD’s also working to build run-time environments and compilers that are optimized for the types of languages that many web app vendors are running, rather than more traditional C languages.
And while you can’t redesign and release a chip overnight, which is why lightweight profiling will be a long time coming, Lewis admits that some things can be changed quickly. For example, web servers, such that many of these customers are running, are part of a network of servers running an application rather than a single application that runs on one server. However, AMD doesn’t benchmark its chips in networked servers, which means such tests can provide misleading data for customers like Facebook.
“We always count and look at servers based on how they compute by themselves — in benchmarking tests we don’t even connect them to the networks — so now we have to look at the whole system,” Lewis says.
Still, these changes can only take performance so far in the near term, a fact that Lewis regrets. She’s aware that Facebook’s issues are pressing, in part because the company is innovating in public rather than in a research lab. Instead of hacked-off scientists experiencing hardware limits and writing research papers to present at obscure trade shows, Facebook’s hardware snafus are felt by its 250 million users whose complaints have a widespread effect on the brand.
“This whole cloud computing thing is driving the technology and innovation from the consumer market and they have very different expectations around innovation than the research and enterprise guys,” Lewis says. “Facebook is doing innovation in a production environment, and that’s a new level of stress.”
But thanks to Heiliger’s public comments, he’s passed that stress along to suppliers who are scrambling to keep up with the pace of change.