With 5,000 employees and 280 branch offices, Associated Bank is a prime example of a midsize business with big data demands.
After acquiring another bank in 2006, its storage area network (SAN) grew from 17TB to 300TB in less than a year. Since then, because more applications have come online and because of the need to comply with federal regulations requiring more data retention, the SAN has grown to 900TB, or 52 times its original size.
Dan Marbes, a systems engineer at the Green Bay, Wis.-based bank, decided to try solid-state drives (SSD) to increase the performance on I/O-hungry applications, while reducing his spindle footprint.
He bought three SSDs to serve as top-tier storage for business intelligence (BI) applications on his SAN. The flash storage outperformed 60 15,000-rpm Fibre Channel disk drives when it came to small-block reads.
However, when Marbes used the SSDs for large-block random reads and any writes, "the 60 15K spindles crushed the SSDs," he said, demonstrating that flash deployments should be strategically targeted at specific applications.
SSDs use nonvolatile NAND flash memory chips, which are cheaper than DRAM chips but are still as much as 18 times more expensive than 15,000-rpm Fibre Channel or serial SCSI (SAS) drives, according to Gregory Wong, an analyst at market research firm Forward Insights. But prices continue to fall, mostly in response to consumer adoption of solid-state drive and NAND flash card technology, which ultimately affects the cost of data-center-class products.
According to the market research firm IDC, the global outlook for client-side SSDs -- those used in consumer products such as laptops and tablets -- is expected to grow roughly tenfold in five years, from 11 million units in 2011 to 100 million units in 2015. The use of NAND in 256GB SSDs is forecast to more than double from 19% of all NAND used in SSDs in this year to 42% in 2015. Further, demand for 512GB SSDs is expected to grow from a 0.3% portion in 2011 to 8% in 2015, also underscoring the growing interest in higher density SSDs.
Wong expects mass adoption of SSDs by consumers won't occur until the price has reached about $1 per gigabyte, some time in the second half of 2012. The more consumer NAND flash that is purchased, the lower the overall price of flash products, even in the data center.
Currently, the price for NAND flash in an SSD form factor is about $9 per gigabyte for high-end, single-level cell (SLC) flash and $3 per gigabyte or multilevel cell (MLC) flash. A new class of MLC, called enterprise MLC or eMLC, can withstand up to 30 times more writes than consumer-grade MLC flash technology can, but it also costs about 20% more, Wong said.
In comparison, a Fibre Channel or SAS drive costs 50 to 60 cents per gigabyte.
Many array manufacturers, such as EMC and Hitachi Data Systems, offer an SSD option, and are able to fill array drive slots with SSDs in 2.5-in or 3.5-in hard drive form factors. Some vendors also offer automated tiering software that migrates the most frequently accessed data to the SSDs, while keeping less critical data on lower-cost hard drives.
When it comes to PCIe NAND flash cards, like those sold by Fusion-io, Texas Memory Systems, Micron or Virident Systems, which can be used in all-flash arrays or in application servers themselves, prices can go through the roof, but so does performance thanks to the higher speed interconnect and the proximity of the flash storage to the server processors.
Even so, not everyone "gets" why flash-based PCIe cards are so expensive. Chris Rima, supervisor of infrastructure systems for the Information Services Department at Tucson Electric Power, a subsidiary of UniSource Energy Corp., paid only half of the retail price for flash cards to accelerate the performance in his company's NetApp 3170 NAS arrays. That means he paid $30,000 for each card.
Tony Edlebrock, senior systems administrator at Tucson Electric, said NetApp's flash cards, called Performance Acceleration Modules (PAM), boosted the performance of his PeopleSoft and Oracle customer care and billing reporting system to the point that it cut the nightly batch process in half, from eight and 12 hours to four to six hours.
"The nice thing about flash cache is you just put them in and they begin to improve performance everywhere. Protocol latency went down and throughput shot up," Rima said. "I just don't get why they're so expensive."
According to NetApp, using its PAM flash cards in its NAS arrays makes it possible to reduce the number of disk drives by as much as 75% while achieving the same or better performance.
Flash strategies, use cases
However, because of the exorbitantly high cost of some PCIe-based flash cards, their purchase should be based on multiple business use cases, not a single task, Rima said.
Tucson Electric Power uses its six flash cards on the front end all of its PeopleSoft and Oracle databases. The cards also act as front-end cache for two-thirds of Unisource's 500 VMware virtual machines (VM), GIS mapping systems as well as databases that manage power outages. The cards hold from 40TB and 60TB for every two NAS servers, which are configured as a cluster.
In 2012, Tucson Electric might purchase all-SSD storage arrays from Nimbus Systems, similar to storage systems that eBay rolled out earlier this year, Rima said.
When eBay had problems meeting the I/O storage demands of business units starved for more virtual machine (VM) deployments, the online auction site's quality assurance division had an idea: swap out hard-disk-drive-based arrays with SSD storage.