Enterprise Flash Drives Target the Data Center

With a "green energy" theme taking hold this year and continued pressures on IT managers to shrink storage footprints and reduce energy consumption in data centers, there appears to be an emerging case for flash memory in the enterprise that can coexist with tight 2009 budgets.

The data center "sweet spot" for flash is in its capacity to position itself between higher-performing memory like cache and RAM and slower media like hard drives. By inserting enterprise flash drives (EFDs) in areas where they can deliver low latency and high performance for input/output, enterprises are likely to see a positive difference in I/O processing as they compute their total cost of ownership for data centers.

"Bear in mind that it's not the most optimal solution for enterprises to simply put flash in for hard disk drives, and we are not advocating that," says David Flynn, chief technology officer for Fusion-io Inc. , a provider of solid-state technology. "To make the case for flash in enterprises, you need to insert flash where it can deliver the most benefit and value." That area is in throughput and I/O -- and in the potential power savings that data centers are able to realize.

Power savings must be a factor in any data center's TCO computations. In 2007, Gartner's research vice president Michael Bell projected that 50 percent of data centers would exceed 6 kilowatts of energy per rack within two years -- and that the percentage would increase to 70 percent to 80 percent of all data centers in four years. Bell concluded that the energy consumption and the costs were unsustainable.

"We know that IT decision makers are still very cautious about SSD [solid-state drive] technology," says Greg Goelz, vice president of marketing for Pliant Technology Inc. , a maker of enterprise flash drives established by executives and engineers from Fujitsu Ltd. (Tokyo: 6702; London: FUJ; OTC: FJTSY), Maxtor Corp. , and Seagate Technology Inc. (NYSE: STX) "But they are also starting to see SSD product that is ready for the data center. It's increasingly difficult, if you're a CIO, to continue to work with traditional media and toolsets. In many cases, you've already optimized your media and your energy usage. The challenge now is to find a bolder way to drive efficiency and power savings in the data center. One way to do this is by being able to turn racks into shelves, which EFDs can do."There is no set formula to determine how much an enterprise storage footprint can be reduced by supplementing or by replacing hard disk storage with SSD devices like flash, but Fusion-io reports one customer was able to attain improved MySQL performance after installing two 160-GB Flash ioDrives in a server. The customer then disconnected a 100-disk storage array, shut down three of four Intel processors, and reduced main memory from 64 GB to 24 GB. The number of application transactions per second with the flash was 11,000. Under the original hard-drive system, application throughput had been 9,000 transactions per second. Power consumption for the application fell by 90 percent.

"When you start knocking out hard drives with flash and turning racks into shelves, you will see an 80 percent to 90 percent energy savings," says Pliant's Goelz. "When you combine this with flash performance of 30,000 I/Os per second versus 15,000 RPM hard drive performance of 300 I/Os per second, then it really gets interesting."

Despite the performance data, enterprises still have concerns about flash's durability -- and its predictability within the data center. "The conversation has been about how many write cycles flash memory can perform before it wears out," says Goelz. "We have wear-leveling and benchmark information that will show an enterprise on paper that flash will last for the expected IT lifecycle of five years. But this doesn't totally satisfy the data center people. The typical enterprise range of reads and writes for storage media runs between 60 reads/40 writes and 90 reads/10 writes. It all depends on the storage deployment you have, as well as on the applications you run. Within that range, IT wants predictability in terms of performance."

One challenge to performance predictability is the disparity between read and write performance that characterizes flash. "You might get 30,000 I/Os per second on flash reads, but only 3,000 I/Os per second for writes," says Goelz. "This is a concern for a data center manager focused on performance predictability because he feels he has to use the lowest common denominator for performance, which makes flash a 3,000 I/Os per second device. The goal is to turn that device into a 30,000 I/Os per second device, whether it reads or writes. Then you have the predictability element resolved for the data center manager. That's where the return on investment begins."

A second way to appeal to enterprise decision makers, according to both Pliant and Fusion-io, is by commoditizing flash so that it can interchangeably fit in a variety of data center platforms. This means partnering with leading equipment makers that already talk with data center managers. Pliant works with Fujitsu and IBM Corp. (NYSE: IBM), while Fusion-io partners with IBM, Hewlett-Packard Co. (NYSE: HPQ), and Dell Inc. (Nasdaq: DELL) That approach contrasts with other flash providers that have customized controllers along with flash, presenting it in a highly integrated, highly optimized, and somewhat proprietary architecture."Early on, we decided that a productive relationship with large distribution partners was the key to success," says Flynn. "This is why our goal was not to make memory appliances, but rather an off-the-shelf component that could be used in a variety of machines and applications."

Flynn says the benefit to data center managers is a scalability that, to date, has been elusive. "By componentizing EFDs so they work more like a commodity item that can be dropped in and more readily scaled than proprietary memory devices, we can scale for both performance and capacity," he says. "We have an I/O flash drive that is very simplistic. It does not require a RAID controller like other storage architectures of today that are tree-structured. We have integrated the product with high performance networks by using remote DMA [direct memory access], so it's all about ease of scaling for both capacity and performance. We see third-party integrators doing more of this work in data centers."