This article is more than 1 year old
Facebook's request to the flash industry: 'Make the worst flash possible'
Counterintuitive, perhaps – but eminently sensible
Flash Memory Summit Flash-memory designers may currently be focused solely on improving speed and endurance, but Facebook's director of capacity engineering and analysis wants something completely different from them.
"The Facebook ask of the industry is make the worst flash possible," Facebook's Jason Taylor told his keynote audience at the Flash Memory Summit on Tuesday in Santa Clara, California. "Just make it dense and cheap."
What Taylor was talking about was flash to be used in "cold storage" – not only for rarely accessed data such as logs and metrics that Facebook uses to do its analytics, but also user content that's rarely – if ever – accessed.
"Photos, video – essentially, after you first create these, they're almost never updated," he said. "The majority of that data will probably be written once and read never – really, it's sad."
Slow, low-endurance flash would be ideal for such an application. "Write-once, read-never is probably the spec for a lot of this," he said, noting that when you want to keep something for a very long time and essentially never touch it, write speeds could be up to 10 times as slow as current flash, and the user experience would be hardly compromised.
Facebook currently handles its "cold storage" by assembling a rack of hard drives that sit in racks, spun down, and are only spun up, individually, when a request is made. Needless to say, when a Facebook user actually wants to access something that's in cold storage, the latency involved is considerable.
Cold storage that's done using flash, however, doesn't have the same spin-up latency as does hard drive–based cold storage. Taylor called this type of rarely used storage cold flash. "I wouldn't say it was completely cold," he said. "It's more like 'tepid'."
Cold flash wouldn't need to be fast and long-lived, as the type of flash that the industry is continually working on perfecting. "Essentially the silicon could be developed in such a way that we could get really dense flash modules that could then be created for very high-fidelity, long-term read-only memory," Taylor said.
And cheap – don't forget cheap. ®