With each passing day, more than 350 million digital photos find their way onto Facebook, joining the more than 240 billion that are already there. Jay Parikh is the man who makes sure all those photos can remain online from now until the end of time, and that’s going pretty well. But he wants a little help. He wants the world to build him a new kind of flash memory ideally suited to storing all those old pics you rarely look at anymore.
The way things are going, he’ll probably get it. As Facebook’s vice president of infrastructure engineering, Parikh oversees the development of both the software and the hardware needed to operate the world’s most popular social network, and in recent months, he and his colleagues have developed a certain knack for convincing the world’s hardware companies to build stuff that suits Facebook’s particular needs — and the needs of other massive online operations. Just yesterday, Intel, AMD, and two other chip makers put their weight behind Facebook’s effort to create a new type of server that lets you add and remove processors whenever you like.
In the spring of 2011, Facebook founded something called the Open Compute Project, hoping to remake the hardware industry through communal collaboration, and on Wednesday, even as the company was showing off the latest fruits of the project, Parikh was calling for more.
Today, Facebook stores all your photos on good old-fashioned mechanical hard disks. This works well enough, but now that the company is juggling hundreds of billions of photos, many of them dating back three or four years, it makes less sense by the day. Those older photos aren’t accessed that often, and if they remain on hard drives, they continue to sap relatively large amounts of power.
“We can’t say: ‘Hey, you want to look at that picture from Halloween five years ago? We’ll go grab the tape and put it on the drive and send it to you in the week.’”
— Jay Parikh
Traditionally, companies have moved their older digital data to, say, magnetic tape. With Facebook, there’s an added rub. Even those older photos must be stored in a way that lets you instantly retrieve them. “We can’t say: ‘Hey, you want to look at that picture from Halloween five years ago? We’ll go grab the tape and put it on the drive and send it to you in the week,’” Parikh explained on Wednesday, during a speech at the Open Compute Project’s latest summit meeting. “That just doesn’t work.”
Parikh and company could move those older photos to the sort of flash memory devices that are now invading other parts of the computer data center — devices that consumes far less power than hard disks — but these carry relatively high price tags. Today’s super-speedy flash devices are a great to store information that’s accessed all the time, but they’re too expensive for the kind of “cold storage” Facebook needs for old photos.
So Parikh wants someone to build him some flash storage that’s cheaper. Basically, today’s flash is built to maximize the number of reads and writes you can make — the medium tends to degrade as you read and write — but he wants a version that specifically designed for data that’s accessed only occasionally.
“We want something with a low write-endurance,” he told us on Wednesday. “We want to get to something with a much, much lower unit cost per gigabyte.” This will also reduce the company’s power consumption, and according to Parikh, it will give his engineers more flexibility when it comes to writing software. “You play all sorts of neat software tricks in terms of how you optimize it.” When you’re reading and writing data on spinning hard drive platters — and you want to keep speeds up — there are tight restrictions on how you can do so.
On stage, during his speech, Parikh likened storage devices to cars. Right now, he said, it was as if his only options were a minivan and a high-end sports car. What he really wants is a third option — something more like a Prius. “We want is a continuum of options,” he said. “Today, the spectrum is very coarse. There’s tape on one end. There’s flash on the other end. And spinning disks in the middle. And that’s it.”
That may seem like a tall order, but Facebook is now so large — and so influential — it can make such requests in earnest. According to Facebook hardware design and supply chain guru Frank Frankovksy, Facebook engineers have long worked hand-in-hand with the Utah-based flash storage outfit Fusion-io, playing directly into the company’s latest designs, including the new 3.2 terabyte flash introduced on Wednesday at the Open Compute Summit, and Fusion-io chief David Flynn is only too happy to advertise this relationship.
Fusion-io’s high-speed flash cards are already driving Facebook’s database software — where so much of your text-based today is stored — but someday, Facebook will also move flash into the cold storage facility it’s building outside its data center up in Prineville, Oregon. It needs to get its hands on an entirely new type of flash before that can happen. But that shouldn’t be a problem.