To help researchers receive data in a more ad hoc manner, flash storage may be just the thing. "We have no religious attachment to flash, but we can construct flash-based storage at a reasonable cost and around 25ms latency, and we are doing so."
SLAC has developed its own SSD-based system that is in the final debugging stages, Mount explains. "The first version of this will provide about 2TB of storage, but we can easily grow this to 5 or 10TB just by buying flash chips," though he reckons the scalability will require "more serious expenditure." At the 2TB level, it will serve as a test and development system only.
Eventually, the goal is to use SSD technology as a cache for all particle accelerator research, which will allow scientists to access data at any time from any data store. "SSDs help the entire system run more efficiently by ensuring the I/O capability is in balance with the rest of the application system," adds IDC's Janukowicz. "The characteristics of flash-based SSDs make them a well-suited alternative for high-IOPS applications that are read intensive. SSDs have no rotational latency and have high random-read performance. Thus, with SSDs the time to access the data is consistent and very small regardless of where on the device the data is held."
Considering SSD at the Pacific Northwest National Laboratory
At the Pacific Northwest National Laboratory (PNNL) in Washington, solid-state technology could help alleviate a supercomputer bottleneck. At the lab, researchers run tests that sustain a write speed of 80Gbit/sec. and a read speed of 136Gbit/sec. Yet, one or two slow hard disk drives running at one quarter the speed of other disks causes performance to degrade quickly.
"Solid-state devices such as flash drives can use a RAID striping technique to achieve high streaming bandwidth -- just like [hard] disk drives -- while also maintaining very low latency for random access," says Robert Farber, a senior researcher at PNNL. "This is a very exciting combination."
The lab has not moved to solid-state technology yet. But Farber says the real debate is whether low-latency access for "seek-limited applications" -- in other words, many requests for small amounts of data -- can alleviate the pressure of computing bandwidth. It is not solely a price-per-gigabyte debate. "It remains to be seen how much of a price premium consumers will tolerate before robustness, power, storage capacity and physical space differences cause a mass departure from magnetic media," Farber says.
At the PNNL, the latency goal for its last supercomputer was 25Mbit/sec., per gigaflop of peak rate floating-point performance. This is mostly to be able to handle the data-intensive nature of the NWChem scientific software calculations running. The lab's new environmental molecular sciences facility contains a new supercomputer with a theoretical peak floating point performance of 163 teraflops. And, like at the Stanford lab, disk speed is a critical part of the equation, so solid-state is the forerunner in solving the bottleneck.
One breakthrough Farber expects in the not-too-distant future: Operating systems will change their memory hierarchy to directly access SSD, turning the technology into a hard drive replacement for mass storage.
Complementary, not replacement tech for most users
One question that remains: When will SSD really impact the corporate world? Some say SSD in the data center is just on the horizon, since laptops such as the Dell XPS M1330 uses a Samsung 64GB SSD. Alienware also offers a 64GB option in some of its desktop computers. And SSD is applicable across the commercial landscape; while researchers need the speed to study proteins, retailers may need or want faster POS transactions.