SSDs Ready For The Enterprise

Because the storage subsystem is the performance bottleneck for most commercial applications, system designers have sought to speed the flow of data from disk to main memory using a variety of schemes. The latest move relies on solid-state disk technology. Over the past year, flash-memory-based solid-state devices have emerged that deliver data in a few microseconds; are significantly less expensive than RAM on a per-gigabyte basis; and, like disks, store data reliably when powered off.

Enterprise-class flash devices are still pricey--averaging more than $100 per gigabyte. But for those with deep pockets and a real need for speed, these systems can deliver up to 45,000 read I/O operations per second, or 16,000 write IOPS, compared with the 170 IOPS typical of a 15-000 RPM drive. A single mirrored pair of SSDs can outperform 100 spinning disks that would cost several times as much after drive enclosures, array features, software that's licensed by capacity, and other so-called slot costs are figured in. Flash devices also can help you go green because that pair of high-end SSDs will use much less power, and generate less heat that the data center cooling system must remove, compared with a group of 15,000-RPM drives delivering the same IOPS.

Vendors from EMC to Xiotech are debuting innovative SSD technologies suitable for enterprise data centers and branch offices. But there are a few downsides. Long-term reliability is unproven, the bottom-line cost on a per-gigabyte basis could send your CFO into cardiac arrest, and advances in data classification and tiering are still needed to gain maximum benefit. IT groups need to be aware of the trade-offs.

How Flash Works
While flash is semiconductor memory, like the RAM that makes up a computer's main memory, it doesn't allow direct read and write access to each byte, the way RAM does. Just as a disk drive is divided into sectors, the NAND flash chips typically used in SSDs are organized into pages, typically of 4 KB each. These pages are in turn collected into blocks of 256 KB to 1 MB.

Data can be read from the flash memory on a page-by-page basis and written to empty pages. However, to overwrite data in a previously used page, the entire block containing that page must be erased--a relatively slow process. If other pages in that block contain valuable data, that information must either be relocated to pages in another block or loaded into RAM cache in the SSD and written back once the block has been erased. This results in flash-memory devices being three to 10 times slower to write data than to read it. In addition, the high voltages needed to erase blocks cause wear on the drives' microscopic transistors and connections, eventually wearing them out.

DIG DEEPER

Put Data On A Diet

Find out how deduplication can trim bloated file systems.

Download this
InformationWeek Report

See all our InformationWeek Reports

Flash memory comes in two basic varieties, differentiated mainly by the ability to handle multiple erase cycles. As the name implies, single-level cell, or SLC, flash stores 1 bit in each memory cell, while multilevel cell memory stores 2 or 3 bits of data in each cell.

Storing 2 or 3 bits per cell increases density, and therefore reduces cost, but it also slows access and reduces the longevity of the device. Therefore, most server and array flash systems use SLC technology.

Even among SSDs using the same flash technologies--or even the same flash chips--performance can vary significantly, however. A main variable is that individual vendors design the controllers that put a storage interface on their flash chips.

Vendors Flock To Flash
EMC started the trend of SSD adoption by adding STEC's Zeus IOPS SSD to its Symmetrix, Clariion, and Celerra product lines about 15 months ago. Following EMC's lead, most array vendors, including IBM, HDS, and Hewlett-Packard, are replacing standard Fibre Channel drives in their arrays with SSDs.

The benefit for storage administrators is that they can build logical unit number, or LUN, RAID drives from these flash drives and move their most I/O-intensive data to these new, blazingly fast LUNs.

Of course, identifying this I/O-intensive data and moving it to flash isn't always a simple matter. You'll need someone with detailed knowledge of the organization's data and a database administrator who can move Oracle tables, or high-use portions of tables, to the new flash LUN. Users of applications like Exchange that treat the entire database as a single file will have to transfer the whole enchilada to gain benefits.

The next step in accommodating SSDs is for vendors to adopt automatic storage tiering in their devices. Then, the storage system will track what data needs faster access and automatically move it to faster SSD LUNs, lightening the load on admins. There is movement here: Symantec's VxFS file system, a component of its Storage Foundation storage management package, can move frequently accessed files to faster disk, while Compellent's Storage Center has supported automated tiering for block data for over a year. EMC has announced its Fully Automated Storage Tiering technology, which will move files on Celerra devices later this year and manage block data on Clariion and Symmetrix boxes in 2010.

Sun's (now Oracle's) Open Storage products currently employ high-performance SLC SSDs to hold the frequently accessed file system logs for its ZFS file system and can use hundreds of gigabytes of lower-cost multilevel cell memory, which is slow to write but fast to read, as a read cache. This combination allows Open Storage NAS systems with flash and SATA drives to perform like rival devices with costlier 10,000- or 15,000-RPM drives.

Bucking the trend toward using a small number of very fast but expensive SSDs, Pillar Data Systems and Dell/EqualLogic have chosen to implement whole enclosures of lower-cost, but still SLC-based, SATA SSDs from Intel and Samsung. While these lower-cost drives deliver just one-sixth the write speed of STEC's, they're a small fraction of the cost--$10 per gigabyte for Intel X25-E vs. $110 per gigabyte for STEC. They can deliver plenty of performance at a price that branch-office and midsize customers can afford.

Rather than bundling flash chips into modules that emulate disk drives in size and interface, another group of vendors, led by Fusion-io and RAM SSD pioneer Texas Memory Systems, are putting flash memory on PCI Express cards. These products can achieve astounding performance in both throughput and latency by bypassing drive interface electronics, RAID controllers, and SAN interconnects.

Our Take

SOLID-STATE DRIVES

SSDs are performance monsters--a single mirrored pair can outperform 100 spinning disks.They provide green benefits such as reduced power and cooling costs.Major vendors are bringing more SSDs to market, driving down prices.That said, they're still expensive. SSDs will take a bite out of your storage budget.Long-term viability for SSDs is still unproven.

While PCIe flash drives are fast, they're also direct-attached storage that's owned by a single server. This is at odds with the move toward ubiquitous server virtualization, which relies on a shared storage back end--either SAN or NFS--something current PCIe flash systems can't support.

Time To Flash?
We're just in the first generation of data center flash implementations, and there's no consensus on how best to take advantage of this new technology. Even so, flash SSDs are too powerful a tool for enterprise data center engineers to ignore, and the devices based on these drives will only get more compelling as prices fall and storage system designers integrate their advantages more tightly into next-generation products.

For now, IT groups that can easily identify, and relocate, 5% or more of their stored data that require significantly higher I/O rates should be looking seriously at adding a flash-based Tier 0 to their storage infrastructures. Companies that can't yet easily identify their hot spots should launch a data classification project to do so, while closely following the development of automated tiering.

Howard Marks is chief scientist at DeepStorage.net, a testing lab and analyst firm.