Solid-state drive (SSD) technology has been in use for decades, but it was hardly a factor in enterprise storage...
circles until approximately 18 months ago, when the price of flash memory started to drop dramatically.
At the present cost of approximately $4 per gigabyte, solid-state drives often make sense only for those applications with the highest performance needs. Although nearly all server and storage vendors have added solid-states drives to their lineups, fewer than 70,000 SSD units shipped last year and fewer than 300,000 will likely ship this year, according to Joseph Unsworth, research director at Stamford, Conn.-based Gartner Inc.
"There's still a lot of challenges out there," Unsworth said. "You need to have the right SSD, properly managed, supported by the right company. Enterprise servers and storage have been architected around hard drives for the last 20, 30, 40 years. All of a sudden you're going to put in a solid-state drive that can access data 100 times faster than even the fastest hard drive? Believe me, these systems aren't architected to take advantage of that and exploit that."
Solid-state disk vs. Fibre Channel
Still, the allure becomes apparent even when looking at the test results of one early adopter. Munder Capital Management in Birmingham, Mich., reduced data access time for its Microsoft Corp. Exchange Server by 89% on average during a 25-day test period of solid-state drives vs. 15K rpm Fibre Channel drives in its Compellent Technologies Inc. Storage Center system. The response time for SQL Server plummeted by an average of 86%.
Munder tested 4K, 8K, 16K, 32K and 64K I/O block sizes with 25 different workloads over a 24 hour period, varying between reading and writing, sequential and random. It claimed the performance of one SSD was equivalent to the performance of 10 to 16 Fibre Channel hard disk drives (HDDs), depending on the workload.
Such a performance boost is especially appealing for mission-critical trading applications that need to process large amounts of data and crunch numbers at lightning-quick speed. Munder also plans to use SSDs for Microsoft SharePoint Server and Exchange Server.
"Right now, that's the best bang for the buck," said Michael Dufek, Munder Capital Management's director of information systems. Until the cost of SSD drops, he added, "It's not geared for everything."
With the Compellent system, the cost of a single 15K 300 GB Fibre Channel HDD is $2,100; a single 146 GB SSD is $24,000, although savings in power consumption and software licensing fees help to offset a portion of the cost differential.
The SSD that Munder Capital Management tested in its Compellent array is a type of flash known as single-level cell (SLC) -- the technology currently favored by most vendors that produce storage systems aimed at enterprise-level needs.
Different types of SSD technology for enterprise data storage
While users such as Munder can take advantage of SSDs today, it's still early days for the technology in enterprise storage. The technology is still evolving, with different types of solid-state drives available -- each with its advantages and drawbacks. But the first point of debate has nothing to do with technology. It's the acronym itself.
Purists argue that SSD stands for solid-state drive -- not solid-state disk -- because there's no spinning disk in an SSD. Still, it's not uncommon to hear the term solid-state disk, perhaps because SSD is often evaluated against HDDs and, depending on the system, uses the same Fibre Channel, SAS or SATA storage interfaces as traditional hard disk drives.
The first type of SSD technology to gain any noteworthy foothold in the enterprise was Dynamic Random Access Memory (DRAM) SSD. Texas Memory Systems Inc. is a leading vendor of DRAM SSD. But the technology generally has been cost-prohibitive for all but those companies whose businesses depend on application performance for success.
The main distinguishing characteristic of DRAM SSD is its volatility. If the power supply is removed, data is lost. DRAM makes use of capacitors, but because the capacitors leak charge over time, they must be refreshed continually to preserve the data. To address those issues, manufacturers typically build-in an alternate power source, such as a back-up internal battery, and the ability to move data to an HDD or a flash-based memory system in the event of a power failure.
Dynamic Random Access Memory SSD's main selling point is performance, and it remains the fastest technology for accessing data and performs reads and writes at approximately the same speed. On the downside, DRAM SSD is expensive and has greater power needs, creating an opening for the cheaper flash-based, energy-efficient SSD technology to gain momentum in the enterprise.
NAND flash memory
Unlike DRAM, flash memory is non-volatile, so the storage cells retain data whether the power is on or off. Flash SSDs have floating-gate transistors that can store charge for an extended period of time, even if there's no power supply connected. Oxide insulation surrounding the floating gate traps the electrons there.
Another contrast between flash DRAM SSDs is data access. DRAM is random. Flash is serial and relies on a controller to bring the data out of the chip and correctly present it to the processor.
There are two types of flash memory: NAND and NOR. NAND flash technology strings together floating-gate transistors to achieve greater density. NOR flash has no shared components, is more expensive to produce and is found mainly in consumer and embedded devices, primarily as a mechanism to boot them. NAND flash, by contrast, is used for data storage and is increasingly making inroads into the enterprise.
What may increasingly become an important consideration for enterprise users is whether NAND flash is single-level cell or multi-level cell (MLC). In SLC NAND flash, each memory cell stores an analog representation of the data and two levels of charge. The less-costly MLC flash can store twice -- as two or more bits per cell and multiple levels of charge -- but it doesn't perform as well and is less reliable than SLC.
As a result, single-level cell is the dominant flash solid-state drive technology in enterprise storage systems today. Gartner's Unsworth estimated that SLC factors into more than 80% of enterprise-grade flash SSDs. But he predicted the breakdown could shift to 60% SLC and 40% MLC by 2011, if vendors sufficiently improve the controller technology and storage management software. MLC is better suited to read-intensive, not write-intensive applications, Unsworth added.
One of the distinguishing characteristics of NAND flash is its process for writing data to a chip. All of the bits in a flash block must be erased before any writing can take place. The term "flash" derives from the violent charge to erase the data. The block size is typically 128 KB to 256 KB, and will likely expand in the future.
NAND flash wear-out
Unfortunately, the erase/program process eventually takes a toll, breaking down the oxide layer that traps the electrons. The gradual deterioration can distort the manufacturer-set threshold value at which a charge is determined to be a zero or a one.
The deterioration is less a problem in SLC flash than it is in MLC flash. In SLC flash, there's only one manufacturer-set threshold value, so the likelihood of a problem is lower. The number of electrons controls the switch-on voltage of the floating gate, and the voltage will either be above the threshold point or below it.
With MLC flash, the manufacturer can set multiple threshold values. As the oxide layer deteriorates, those values can shift across the pre-set threshold points and become difficult to discern, leading to errors.
Both single-level cell and multi-level cell flash rely on error-correction algorithms to ensure the data remains intact, but eventually, NAND flash SSDs wear out. The wear-out figures typically used by the industry are 100,000 program/erase or endurance cycles for single-level cell flash and 10,000 cycles for multi-level cell flash, but those figures vary widely by manufacturer.
Michael Cornwell, lead technologist for flash memory at Sun Microsystems Inc., claimed the MLC endurance cycle figure has been worsening to a figure closer to 3,000 program/erase cycles. He said as flash dies get smaller, fewer electrons fit on the floating gate. That trend, coupled with the natural tendency of electrons to escape, will lead to the use of more sophisticated data correction.
"The flash is going to become a lot more unreliable, but the good news is that error-correction techniques in disk drives will be applied in SSDs, and that will make it more reliable for storing data," Cornwell said. He said most flash solid-state drives now use between 8 bits and 16 bits of correction per sector. A hard disk drive, by comparison, uses 500 bits of correction for the same amount of data.
As vendors try to reduce the cost of solid-state drives, they're working to improve MLC to make it more suitable for certain types of enterprise applications. Lower-cost MLC currently accounts for more than 90% of the overall worldwide NAND output, but only a small percentage of the enterprise space.
"There's a race among SSD makers to find a way to make it work reliably in an enterprise SSD," said Jim Handy, a Los Gatos, Calif.-based analyst who focuses on memory chips and SSDs for the market research firm Objective Analysis.
Issues such as wear-out and error correction don't factor into the equation with the more expensive DRAM solid-state drive, according to Handy. He said DRAM SSDs can be written to an infinite number of times and feature high-integrity memory.
Manufacturers are also exploring a variety of different mechanisms to improve solid-state drives, such as phase change and resistive memory, in addition to the classic approaches of shrinking the technology and packing more bits into cells.
"There are a dozen different implementations that may extend the life of flash technology as we know it," said Patrick Wilkison, vice president of marketing and business development at STEC Inc., which supplies SSDs to major storage vendors such as Dell Inc., EMC Corp., Hewlett-Packard (HP) Co., Hitachi Data Systems Corp., IBM, NetApp Inc. and Sun Microsystems Inc.
In the meantime, users in need of high-performance storage need to weigh the benefits against the potential pitfalls. Munder Capital's Dufek said he doesn't worry about the solid-state drive wear-out factor because his data is backed up, his Compellent storage system monitors for errors and, if an SSD fails, the system can automatically move data from the SSD tier to a spare HDD. Dufek said that under the terms of his agreement with Compellent, Munder Capital would have to wait no more than four hours for a replacement SSD to arrive.
"With flash, the chances of a chip death are very small," said Robert Ober, an LSI Corp. fellow for business development and strategy, who's driving the company's solid-state drive efforts. Plus, SSD failures don't happen without warning, similar to the way a tire sensor can alert the owner when the tread is wearing out, he added.
Solid-state drive trends and future directions
The term "tier 0" has come into vogue to describe solid-state drive storage, as interest in the technology heats up. Framingham, Mass.-based IDC predicts worldwide revenue for enterprise solid-state drive storage (SSD storage) will increase at a compound annual growth rate of 78% from 2009 to 2012. IDC predicts that SSD sales will hit $382 million in 2010, up 132% over 2009.
But several things must happen for SSDs to take off in enterprise storage systems. First, the price must fall. Then storage system vendors must optimize their arrays and management software to work better with solid state.
Jeff Boles, senior analyst and director, validation services at Hopkinton, Mass.-based Taneja Group, said most of the arrays on the market currently support only a limited number of solid-state drive devices because the controller can rapidly become a bottleneck.
"You put a few SSD devices behind an array controller, and you can bottleneck because the array controller can't keep up with the performance of the SSD devices," he said. "To work around all the media issues and the array controller performance issues, the array vendors are pretty sensitive on qualifying specific media."
They also need to make their storage management software more intelligent to exploit the benefits of solid-state drives and enable end users to see true returns on their investments, according to Gartner's Unsworth.
"This is still a very, very nascent market that's just starting to get its footing," he said. "It does have a meaningful, impactful, disruptive role in enterprise storage, but companies need to be careful."