IT execs explain their moves to solid-state storage
- 06 August, 2012 11:39
Fred Abounader, a performance systems engineer at chip maker AMD, recently deployed a 3TB all-NAND flash storage array into a virtual server test environment.
The result, he said, was astounding.
In AMD's virtualization benchmarking test environment, the SSD array helped reduce latency by a factor of 50 and yielded a 40% improvement in performance compared to hard disk arrays.
Although the system was operating in a test environment, the data was real, compiled from emails, databases, and Web 2.0 applications -- it even included data migrations between VMs. "It was like a real life data center," Abounader said.
The goal of the test was to see if AMD could overcommit its servers and still get the performance it needed on the back end without being bottlenecked. The experiment was an overwhelming success, Abounader said.
With just one 6TB flash array from WhipTail, the system was able to achieve 86,000 I/Os per second (IOPS) for virtual machines that typically saw only 6,000 IOPS using a SAN with 400 15,000-rpm hard drives connected by a Fibre Channel network.
Now AMD is looking to scale out its SSD environment even further.
"It's been installed for a few months now, and we've run it pretty hard," Abounader said. "There have been zero issues."
Picking up momentum
Abounader is part of a small but fast-growing universe of enterprise users who have deployed flash storage. In a recent TheInfoPro survey of 255 IT managers and storage admins, 37% of the respondent said that they plan to deploy SSD technology, up from just 7% in 2011.
Adoption of more-expensive all-flash arrays has been slower -- only 7% of the respondents said they that are currently using them, while 86% said the technology isn't in their IT plans at this time. About 4% of the respondents indicated that they plan to purchase all-flash arrays in six to 18 months, and 2% said they expect to do so after 18 months.
Industry experts say that those who aren't using NAND flash are missing out on the huge performance advantages that the technology can offer. And they contend that purchasing NAND flash doesn't have to break the bank.
"Every big organization has got to be looking at having a full range of solid-state devices if they don't want to be left behind," said Mark Peters, an analyst at ESG. "Solid-state storage is not one thing. It has the potential to be even broader than spinning disk products."
"We see that flash is starting to change the business world," said Kobi Rozengarten, a managing partner at Jerusalem Venture Partners (JVP). "Magnetic media formerly used for storage is starting to move toward solid-state flash, mostly because of performance and partially because of lower power and smaller form factor."
JVP, a leading Israeli venture capital firm that invests in media technology companies, manages more than $900 million in eight investment funds. EMC in May paid some $430 million for one of its technology investments, XtremIO, an Israeli maker of all-flash primary storage arrays.
Rozengarten said the next step is for flash to be used as a primary storage medium, not just in virtual tape libraries to speed backups. But that move that is happening far more slowly than many experts had anticipated.
"Flash is basically an unreliable device if it's just operated as flash. That's one reason for the slow uptake," he said, citing the need for better flash management software.
Virident's FlashMAX flash module is used at Vail Systems.
"The other reason is that large enterprise companies such as EMC, Dell, HP and others are totally committed to their own storage product structure today. This structure is optimized around HDD, and it doesn't really work well when it comes to flash," Rozengarten added.
Most major storage vendors have been offering solid-state drives or PCIe flash cards in storage arrays for at least a couple of years. Rozengarten said those products are very expensive, and do not take full advantage of the performance potential of flash technology.
Rozengarten is quick to admit that NAND flash technology will never beat the per gigabyte price of HDDs. However, when targeted at specific applications, such as virtual desktop infrastructures and online relational databases, the costs to achieve the same performance with flash compared with HDDs can be vastly lower, he added.
More and faster customer response times
David Fruin, vice president of engineering at Vail Systems, said HDDs didn't cut it when it came to his SQL database's response times to customer inquiries.
Vail Systems, a telephony service provider for banks, insurance companies and others in the Fortune 500, runs an interactive voice response system for customer care and conference calls. The system can guide callers through processes such as automated credit card activation, Fruin said.
All phone calls become electronic records, storing in log information such as how long the call took, when it was made, how the caller used the system and if anything unusual occurred during the call. The company's Microsoft SQL Server database holds all log information and keeps billing records.
Vail processes more than 48 million billing records a day, and while each call log is relatively small, they add up to about 4.5GB of data in the SQL database per day.
Customer automated queries had required significant disk access at Vail because the server's memory could not handle random access across the whole database. The databases are mirrored for redundancy, but while redundant writes occurred, they effectively blocked read operations.
Prior to turning to flash storage, Vail had been using standard Dell servers with 12 10,000 rpm SAS drives set up in a RAID 10 configuration. The spindles didn't provide the response times the company needed. For example, response time for customers calling to check billing-records averaged five seconds, Fruin said.
"With SQL, if you're doing a query and you're looking for something that's not indexed, the SQL engine has to go to disk and inspect it and see what query you're looking for. For us, even one day's worth of data could be 100,000 records," Fruin said. "With no index, there are many seeks, and seeks are what kills performance in database queries."
Fruin said the company had to eliminate the disk-drive bottleneck without changing the architecture of the billing system or the customer-care interface. "Re-engineering the billing system on a No-SQL solution wouldn't achieve these goals," he said.
Thus, Fruin selected flash storage, first by adding 2.5-in Intel SSDs to the Dell servers, and then PCIe modules from Virident Systems. The SSDs improved performance, but only about four times. The two 1TB Virident flash modules offered up a 10x performance improvement over hard drives.
Virident's FlashMAX PCIe modules come as either high cost, high performance single-level cell (SLC) NAND flash or lower-performance, lower cost multi-level cell (MLC) NAND flash devices. They range in capacity from 550 gigabytes to 2.2 terabytes per module. The MLC cards, which Fruin chose, offer read performance of up to 1.3GBs or 325,000 IOPS using 4k blocks.
Fruin said the flash allowed 10 times as many caller queries to be handled in the same amount of time.
"We were looking for four times improvement based on projected call volume increase over time and we got 10X, so we were surprised," he said.
After Virident's FlashMAX flash module was installed, response times to billing inquiries dropped from an average of five seconds to virtually real time -- as low as 500 milliseconds.
Fruin considers Vail's flash storage technology a competitive advantage, as it improved response times and allowed the business to service more clients every day.
Database load capacity also doubled, allowing Vail to store more billing records for clients to use as trending data.
The flash technology also addressed what had been a significant source of downtime with the hard drive configuration.
Fruin is candid about the cost. A single flash module cost Vail $13,000, he said, but added that alternative solutions were no cheaper.
"They are a lot of money [but] the alternative was to throw a lot of RAM into the boxes, which is even more expensive," he said.
It's about how, not how much
Ryan Chien, a storage and memory analyst with market researcher IHS iSuppli, said using flash to accelerate the speed of databases could be very inexpensive if users are careful with how they deploy it. A little flash, he said, can go a long way.
"Probably the most cost effective way to [accelerate Oracle databases] is to put the write tables and hash tables on flash memory," Chien said. "That still gives you a 2X to 3X performance boost without a huge investment involving moving whole database over to flash. That's really the best way to do it in 2012. Most companies don't need 100TB capacity arrays yet."
Associated Bank and three SSDs
Dan Marbes, a systems engineer at Associated Bank in Green Bay, Wis. recently deployed three SSDs in one of its three Dell Compellent storage arrays.
The bank installed a SAN about seven years ago with a capacity of 17TB. When it acquired another bank in 2006, SAN capacity grew to 300TB. Since then, the SAN's capacity has tripled again.
Associated Bank now has more than 5,000 employees and facilities in 280 locations.
Within the last year, the bank brought its public Internet servers in house, which increased I/O workloads on backend storage and archive servers. The bank also recently rolled out a business intelligence application and upgraded its website.
The bank's three Dell-Compellent storage arrays are classified by the applications that are use them: database clusters and business intelligence; VMware servers; and archive data.
Marbes deployed three SSDs to support the database clusters and business intelligence array. The performance boost was immediate, he said. In terms of sequential, small-block data reads, the three SSDs outperformed what 60 15,000 rpm Fibre Channel hard drives formerly handled.
However, when it came to large-block and random data read operations, the 60 Fibre Channel drives "crushed the SSDs."
"Obviously, with SSDs, unless you're going to make a massive investment and put a whole shelf of SSDs in, you're going to be hamstrung by capacity," Marbes said.
SAP and an all-flash array
Matt Wattles, enterprise infrastructure architect for Mitsubishi Power Systems Americas, said his organization installed an all-flash array from Nimbus Storage last December to address an I/O bottlenecks it the company's SAP SQL Server environment.
While less than a half a terabyte in size, nightly SAP SQL Server database loads on the company's Hewlett-Packard EVA storage array were taking four to five hours to complete. That and the company's other nighttime backups were beginning to move into daytime hours.
Wattles said he needed ERP load times to be reduced to between one and two hours in order to fit his daily window.
The slow backups were also affecting SAP response times. "Developers were not happy with the response times," he said.
Wattles initially tried putting a second set of processors in all of his SAP SQL servers and bumped up the RAM to its maximum capacity, but the problem persisted.
He finally narrowed the performance problem down to the terabyte-sized SAP database and loading data from primary storage, which at the time was an HP EVA storage array and fibre-attached, technology adapted (FATA) and SAS drives.
Wattles initially looked into simply adding SSDs to his HP EVA storage array, but he said that option was even more expensive than purchasing an all-flash array, he said.
Mitsubishi finally settled on an all-flash S Class array from Nimbus Systems.
The 2TB array listed for $50,000 and had a street price of around $40,000, he said. The Nimbus array could be used solely for the SAP environment, offering the highest performance possible without a multi-tenant setup, an added benefit.
The superior performance of the flash storage on the Nimbus array not only eliminated the database bottleneck, it also slashed data backup times from four hours on the HP EVA array to 15 minutes using the new flash array.
Additionally, the all-flash, Fibre Channel-attached array cut backup time from nearly two hours to 20 minutes, an 83% reduction. Now able to fit within his nightly windows, SAP response times for developers also bounded back.
"It made them very happy," he said. "I like the performance so far. There hasn't been a single glitch. I think got the array in one morning and had it up and running that afternoon. That's how easy it was."
Lucas Mearian covers storage, disaster recovery and business continuity, financial services infrastructure and health care IT for Computerworld. Follow Lucas on Twitter at @lucasmearian, or subscribe to Lucas's RSS feed . His e-mail address is email@example.com.
Read more about data storage in Computerworld's Data Storage Topic Center.