PRINEVILLE, Oregon — Tucked deep in a Facebook data center, amid hundreds of humming server racks, nearly 2,000 smartphones are running some some version of a Facebook app.
Some are scrolling through the News Feed, Facebook's notoriously
algorithmic and somehow always-controversial stream of updates from its
users. Other devices are launching Messenger,
the spin-off messaging app Facebook made a requirement for users about
two years ago. Some of the phones are rebooting. At least one is
scrolling through Lady Gaga's official Facebook page.
The phones are part of Facebook's Mobile Device Lab, a new system the
company is using to test the performance of Facebook, Messenger,
Instagram and the company's other apps. The goal? To ensure updates and
features Facebook is testing at its Melo Park headquarters won't break
the app when they get pushed out to the company's 1.6 billion users —
including ones using four-year-old devices half a world away.
In total, Prineville has about 60 racks, each containing 32
smartphones, mounted on boards behind layers of insulation. Most of the
phones themselves are a at least a year old, and some are significantly
older. Some represent cutting-edge smartphone tech from half a decade
ago: the iPhone 4S, Samsung Galaxy Nexus and Nexus 5 are all there. In
each rack, above the tangle of cables and phones, sits a camera,
recording every onscreen movement in case a developer needs to review
specific hiccup.
"When
a developer makes a change to one of the mobile applications, we take
that change, we build the app with the change, and then we install it on
one of the devices that are here and we run the app while collecting
metrics," explains Facebook production engineer Antoine Reversat.
If the update negatively impacts a device's memory usage, battery
life or the performance of the app, like slowing down News Feed
scrolling, then it's sent back to the developer for a fix before the
update can be pushed live.
The whole point is to go fast, we have to get better quickly
The setup is an important one for Facebook, which is aggressively
pursuing users in developing countries — many of whom are using older
devices and operating systems — as it looks to get its next billion
users online. At stake is much more than a few app crashes or bug
reports. The elaborate testing setup helps Facebook push
performance-enhancing updates faster, which could be the difference
between whether or not a frustrated user deletes the Facebook app from
their phone after noticing it's a battery hog.
"The whole point is to go fast; we have to get better quickly," says
Ken Patchett, Facebook's Director of Western Data Center Operations
Right now the Mobile Device Lab is only looking at how updates affect
each app's performance while on Wi-Fi networks. But this eventually
could grow to include how they perform while on reduced network speeds —
an issue that is particularly important to the company as it looks to grow its presence un developing markets.
"That's sort of the next step," Reversat says.
Besides
device testing, Facebook's sprawling Prineville facility is also home
to one of the social network's biggest artificial intelligence projects:
the hardware Facebook has designed specifically for machine learning,
also known as Big Sur.
At first glance, Big Sur doesn't look much different than what's
inside other racks in the data center. But inside are eight powerful
graphical processing units (GPUs) that allow Facebook to train its
artificial intelligence at previously unprecedented speeds.
Big Sur is used to train neural networks — a type of AI that learns
from data much the same way as the human brain does. Previously, the
development of a single neural network could take weeks or even months,
depending on the hardware setup. With Big Sur, Facebook has cut that
time down to less than a day.
If
you're a Facebook user, chances are you are already experiencing the
benefits of this. Big Sur helps power the social network's real-time translations and
photo-recognition abilities — where Facebook actually describes what's
in a photo instead of relying solely on tagging and user-written
captions, which are often notoriously unspecific. Visually impaired
users experience this through the app's ability to recognize photos and read what's in them.
In the future, Big Sur could be used for even more ambitious AI
projects. The field is increasingly moving toward something unsupervised
learning — a type of AI that's able to learn from data on its own —
which could have even bigger implications for how Facebook uses
artificial intelligence in its products.
