smartphone or tablet device. (The 160-gigabyte iPod classic is probably the last major information appliance with a hard drive.) Instead, today’s mobile gadgets and thumb drives store data in NAND Flash, based on chips with billions of non-volatile “floating gate” transistors, each of which can hold a bit of data for years on end, even without power.
But only in the last few years has Flash memory started to make its way into the enterprise computing market. Manufacturers have long been wary of Flash’s tendency to wear out with use; the more read-write operations a single memory cell is subjected to, the more likely it will fail, or start to give erroneous readings. “In 2007, Dell tried to put consumer Flash drives into laptops, and a number of them burned out,” giving a lot of industry insiders heartburn, Basile recounts. To get past the problems, Flash manufacturers eventually came up with a method called wear leveling, which spreads write access equally across a Flash chip and keeps the average cell going longer. They also introduced virtual mapping systems to allow specific bits of data to be relocated on the chip as cells fail.
By 2009, the technology was working well enough to earn an endorsement from Joe Tucci, the CEO of Hopkinton, MA-based storage giant EMC. “Flash will dominate for the foreseeable future and this will totally change the game in arrays,” Tucci said in an EMC World keynote talk that year. It was a remarkable prediction, coming from the head of the company that basically invented the modern high-capacity disk array and whose bedrock storage product, Symmetrix VMAX, comes with as many as 2,400 hard drives storing up to 2 petabytes of data. But in fact, EMC had already begun to build flash drives into a related product, Symmetrix DMX-4.
Then, late last year, came the “seminal event in our industry,” at least in Basile’s eyes. Oracle CEO Larry Ellison announced in December 2010 that his company would build a “Supercluster” enterprise server customized to run Oracle’s database software, using SPARC processor technology acquired in Oracle’s 2009 purchase of Sun Microsystems. The interesting thing about the Supercluster, from Basile’s point of view, is that it will include terabytes of Flash memory as a kind of queuing area for data traveling between its processors and its disk arrays. Basically, the Flash layer is needed because disk input/output speeds aren’t fast enough to keep the processors busy.
The announcement made Ellison into the “chief spokesman for the memory industry,” Basile says. “There is no more powerful voice in saying how software and hardware should transform the enterprise.”
Oracle is also building Flash into its Exadata storage servers. The big difference between Exadata and Violin’s hardware: you won’t find any spinning disk in the Violin machine. “Exadata has 5 terabytes of Flash and 100 terabytes of disk, but Violin makes the full array in Flash,” says Basile. As a result, he says, “we are able to run Oracle faster and cheaper than Exadata.”
In terms of the cost per gigabyte stored, enterprise-quality Flash memory is no cheaper than disk technology: Violin Memory’s storage arrays cost about the same as a disk-based arrays of equivalent capacity. But speed is the real payoff—Violin’s arrays are “radically cheaper” measured in input-output operations per second, Basile says. “One of our customers is going to
