Like Google and Facebook, Twitter Designs Its Own Computer Servers
Twitter juggles and analyzes data using sweeping software systems based on tools originally developed at Google. It meticulously spreads computing tasks across thousands of machines using a clone of a Google system called Borg. It just built an artificial intelligence engine much like the Google Brain. It even customizes hardware a la Google.
According to Alex Roetter, Twitter’s head of engineering, the company now employs a team of hardware designers that customize computer servers for its data centers. These designers don’t build individual computer chips, but they do piece together systems from existing parts. Then Twitter contracts with outside manufacturers—likely companies in Asia—to build them. “We’re not designing individual circuits,” Roetter says, “but we’ve putting machines together.” This is similar to what Google has done for years.
Just What’s Needed
All this is partly because so many ex-Google engineers now work at Twitter. But it’s also because Google has so effectively solved the problems that faces Internet companies as they expand from a small startup to a global operation. Google provides a blueprint for the modern internet as a whole.
Yes, that blueprint includes new hardware as well as new software. What Google and Twitter have learned is that, for their enormous web services, traditional servers from the likes of Dell, HP, and IBM are too expensive and too power-hungry and saddled with too much extraneous hardware. So they’re streamlining this hardware, building just what’s needed for the tasks at hand. “We look at efficiency. We look at the workload of our jobs, what we’re trying to do,” Roetter says in describing how the Twitter designs its machines.
For instance, Twitter recently fashioned a new type of machine for its AI system, which is based on “neural networks”—networks of machines that mimic the web of neurons in the human brain. These systems includes scads of graphics processing units, or GPUs, which are particularly adept at running neural nets.
The New Breed
Google and Twitter aren’t the only ones. Following Google’s lead, many companies are now moving away from the beefy, expensive machines of the past. According to Jason Waxman, who oversees the cloud infrastructure group at Intel, the world’s largest chip maker, about 15 percent of data center machines are now this new breed of streamlined system—“some sort of optimized, scale-based server.” Among companies that run Google-style cloud computing services, he says, that number jumps to 60 percent.
This is not to say that all these companies are designing their own machines. After building its own streamlined servers, Facebook open sourced its designs, sharing them with the world at large. This forced a shift in the worldwide hardware market. Dell and HP and other hardware vendors now sell machines based on Facebook’s designs. What this means that businesses can use Google-like servers without designing their own.
Some have questioned how many companies really need this new breed. But Twitter’s servers are yet another sign that the market is shifting. This makes life more difficult for the HPs and the Dells. Even if they can sell machines for use inside modern web services, their margins take a hit. But they don’t have much choice but to get into the market. The Internet demands what it demands.