Facebook users upload 350 million pictures to its data centers each day, and the company is looking for a new way to store them. Many of the pictures are quickly forgotten and not viewed again for months or even years, but Facebook can't throw them away -- it must store them so users can view their Facebook activity as a kind of lifetime repository, with pictures available for historical reference.
On a tour last February of the massive Facebook data centers in Prineville, Ore., operations manager Joshua Crass pointed out to this reporter a third building under construction at the site. The plain concrete structure would serve as a cold storage facility, he said. It would be packed with disks, many of which would be idle, to maintain old pictures. Users who decided to access infrequently viewed photos might wait longer than usual for disks to be activated and images to be retrieved, Crass explained, but that was the price of enabling Facebook to save pictures long-term and remain an energy-efficient, low-cost operation.
On Tuesday, Jason Taylor, Facebook's director of capacity engineering, called on the memory industry to produce a new type of flash storage that he could use in a future cold storage facility. Taylor, who's responsible for producing picture storage on schedule, termed it "cold flash," meaning it would be used for the same secondary purposes that cold disks were occasionally fired up to fulfill.
[ Want to learn more about Facebook's state-of-the-art data center? See Facebook's Data Center: Where Likes Live. ]
The memory industry has been struggling to produce long-lived devices that sustain frequent data writes, lose less performance through sustained use than early SSDs and hold more data. What Taylor sought was pretty much the opposite: SSDs that are not necessarily engineered for many writes and that are allowed to deteriorate in performance over time. That would be perfect for Facebook's little-used picture storage -- provided the devices could be mass-produced cheaply and in large quantities. The low energy consumption of SSDs is also a good fit with Facebook's preferred plan of operations.
Taylor said that long writes -- like 10 times longer than usual -- do not matter, according to InformationWeek's sister publication EETimes.com. "Make the worst flash possible -- just make it dense and cheap," Taylor said. Part of the rationale behind that approach seeks to lower the $1.24 billion Facebook spent last year building and provisioning data centers.
Those statements were still producing buzz on Thursday as speakers and panels convened on the final day of the Flash Summit. At a panel titled "What Will It Take To Live In A Disk-Free World?" a member of the audience asked, "Do we really want cheap, unreliable SSDs?"
On the contrary, several vendors said they were striving to make their products as reliable and well managed as the best disk drives currently used in tier 1 storage. That's the only way to let flash drives, with their advantages of speed and low power consumption, take over from earlier generations of storage. Simply advocating more cheap parts sounds like mass production of today's throwaways or poor-quality specimens coming off the manufacturing process in a bad lot.
Don Basile, CEO of Violin Memory, said Thursday that a rich opportunity awaited those manufacturers who could produce longer-lived, highly reliable flash. Companies and organizations are spending $15 billion a year on storage, and solid state drives should assume an increasing share.
Violin, which Basile described as "the largest supplier of flash memory to the enterprise" citing a ranking by Gartner, announced on Monday its new 6264 Memory Array, which offers twice the density of its predecessor product at one-third the cost. The array contains 64 TBs of storage in a tray that takes up 3u of space in a server rack. The array is based on memory chips manufactured with 19 nanometer circuits. "Your data center in 2020 will be connected to a vast quantity of real-time data," Basile predicted, and businesses will need to be processing that data 24-7-365. To remain competitive, he said, businesses will need to be evaluating the latest information available from its customers -- and its customers' customers -- on a constant basis.
Many businesses will be communicating via mobile devices, which must depend on the larger memories of server systems to be productive. "If you don't [implement larger systems], Amazon, [Chinese search engine] Alibaba, somebody will," Basile said. Flash memory, with its speed-of-light operation, lends itself to use in real-time data analysis better than spinning disk.
Likewise, Scott Dietzen, CEO of Pure Storage, urged production of better solid state products that could assume the mantle of Tier 1 and 2 storage from disks. "[Flash] allows us to achieve the dream of software-defined storage," he said, where intelligence and policies captured in software produce the right type of storage in the right amount where it's needed. But before such flexibility can be realized, flash products need to be driven by better software systems that capitalize on their advantages, rather than simply trying to duplicate the patterns of today's disk arrays.
Such software would automatically impose de-duplication as data is created and stored rather than applying the process periodically after storage. By applying de-duplication and data compression, flash can store data with fewer writes, which is what tends to deteriorate the reliability of SSDs over time. And that makes the flash system more reliable with better price/performance, Dietzen pointed out.
"The storage market should go flash in the next five years," he predicted. "The winners will be the providers of the best flash storage software."
This year's Flash Memory Summit postulated two countervailing trends. The first is that much enterprise storage work will fall to SSDs as they gain in reliability, sophistication and price/performance. The second is illustrated by Facebook's hypothesis that the best device in the cloud is one that performs well at an unprecedented low price. The device may fail, but the cloud's software writers will have anticipated that and will have another cheap device ready to take its place.