Benchmarks Can Answer Solid-State Drive Question

Sears and other retailers use a company called richrelevance to help personalize product recommendations for their online customers. Richrelevance tracks customer preferences and Website habits, analyzes the data, and develops profiles on consumers so e-commerce sites can better target product and service recommendations for customers' individual interests during visits to retail Websites.

"The recommendations that we send to Website consumers are built by us several times a day, and are based on a mathematical model," says Elya Kurktchi, the company's senior director of IT and operations. To process this data, richrelevance runs five data centers that use a mix of Hewlett-Packard Co. (NYSE: HPQ) and Dell Inc. (Nasdaq: DELL) computers to host Web servers and applications, with load balancers to ensure that the work is evenly distributed for maximum processing efficiency.

"Each data center deploys a highly distributed network architecture, which is why load balancing and the ability to concurrently process transactions on many servers and disks is critical," Kurktchi says. "Nevertheless, we were discovering that on our consumer-facing runtime servers that supported our retailers' Websites, the personalized 'buy' recommendations that ultimately were presented to consumers also had to travel to disk when there was a need to access older data that was not stored in cache. Whenever we had to access a hard drive during the course of running a mathematical model for a recommendation that we would send to a consumer, we encountered latency of up to 100 milliseconds. This affected our response time to consumers."

That isn't good for online retailers, for whom response time is crucial. Richrelevance thought solid-state drive (SSD) technology might be the key to reducing latency and improving response times, so it decided to benchmark SSDs from several vendors against the performance of its existing hard drives.

"SSDs had the capability to respond to the mathematical model formulation and to deliver quick response to consumers," Kurktchi says. "We recognized that the reads of our hard drives had become a bottleneck, and that new generation SSDs could come very close in performance to what we were caching in RAM [random access memory] on our servers."To quantify the performance, richrelevance set up benchmark tests that ran SSDs against hard drive performance benchmarks. "To establish a baseline, we collected historical behavior of what one server using hard drive technology does in a given day," Kurktchi says. "We replayed this baseline in several different test scenarios."

The first test scenario replayed benchmark data with a Unix "unzip" to unpack a typical mathematical model that was developed during customer behavior on a Website. Several other scenarios were run using Unix and also Java in both a fast and a throttled mode. "We performed numerous tests of sequential read performance, and also random and concurrent reads," Kurktchi says.

In a mixed read environment, a Ridata 32-Gbyte drive delivered 1,211 IOPS, a 64-GB SanDisk Corp. (Nasdaq: SNDK) SSD drive produced 2,163 IOPS, and an Intel Corp. (Nasdaq: INTC) X25-M SSD drive delivered 9,272 IOPS. For a random write, the Ridata SSD performed 8 writes, the SanDisk SSD did 12 writes, and the Intel X25M SSD performed 7,700 writes.

Richrelevance had expected to see significant performance improvements with SSD, but Kurktchi says the results were "eye opening... I expected to see 100 milliseconds of hard drive latency reduced to 80 milliseconds -- but not the 16 milliseconds that SSD delivered."

A second area of concern for Kurktchi and staff was longevity of the SSD products, and their ability to match up with IT asset acquisition and retirement cycles. "We realized quickly that not all SSDs are created equal," she says. "We were initially attracted to several alternatives because of their more attractive price points. However, because of the limited numbers of writes that the media could do, we did not see these assets lasting in our operation for more than a year."In durability testing, the Intel X25M SSD again stood out. X25M utilized SSD "wear leveling" algorithms that spread the writes evenly over SSD NAND memory, effectively extending the life time of SSD drives to five years -- a perfect fit with the asset acquisition, retirement and depreciation cycle.

"We were excited, because we didn't want products wearing out within the first year," Kurktchi says. "We also felt that the controller was slow in one of the alternatives we were considering. ... When we looked at the Intel X25M 80-GB server with SSD, we saw an advanced wear leveling algorithm that would keep the SSD in service for around five years, along with an advanced controller design. This was what we had been looking for."

Richrelevance proceeded in a swap-out of hard drives with Intel X25M SSDs. "It was a very straightforward process," Kurktchi says. "We procured the SSD drives, shipped them to the data centers, installed them, and produced new file systems. We performed a series of I/O tests to bake them in and then implemented them."

A successful implementation of SSD technology saved richrelevance the pain of upgrading server cache memory and also improved performance four to five times for mathematical modeling and interactive personalization with retail consumers while they shopped.

"The ability to respond quickly on our consumer-facing servers is absolutely critical to us and to our online retail clients, since in e-commerce, personalized recommendations must come instantly," Kurktchi says. "We will continue to look for other applications that can take advantage of SSD."Kurktchi foresees a not-too-distant future when servers will come equipped with banks of SSD drives that will allow the servers in richrelevance's data centers superior recovery, since SSDs function like RAM, retaining data even when the power is turned off.

"In today's SSD market, there are a lot of options and you really have to do your homework," Kurktchi says. "The real barrier continues to be controller design. An intelligent controller with superior processing power and RAM that can translate all the random writes and place them into their required streams is one key that could unlock this."