Fusion-io Targets Flash Storage at Hyperscale Datacenters

Fusion-io has announced the Fusion ioScale, an attempt to displace disk storage in hyperscale datacenters.

Each ioScale device offers up to 3.2 terabytes of flash storage, with performance tuned to the needs of hyperscale environments. Prices will begin at $3.89 per gigabyte—assuming hundreds of gigabytes of storage are purchased.

Gary Orenstein, senior vice president of products at Fusion-io, couldn’t recall the exact performance in IOPS (Input/Output Operations Per Second) of the company’s new storage option. But he insisted that’s the wrong question to ask.

Hyperscale data centers, built by either top-tier providers or cloud providers, usually mean one thing: low-power, optimized CPUs running specialized workloads. In many cases, hyperscale is understood to mean an ARM CPU, although Intel has targeted the space with its “Centeron” Atom processor, which uses the X86 instruction set. Meanwhile, flash-based solutions from Fusion-io and other firms offer high speeds at a relatively high price (although recent competition in the SSD space may hold prices down).

“The way to look at it is that a hyperscale customer is not looking to combine a cheap CPU with an expensive storage product,” Orenstein said. “Instead, they are considering the… most transactions per second for minimal space and the lowest power.”

Fusion-io’s ioScale is based on the Fusion ioFX workstation acceleration products, but can incorporate commodity flash to save cost. As an example of a customer who could take advantage of the ioScale, Orenstein suggested an American Idol-style broadcaster with a need to process “votes” made up of zillions of SMS votes. Each vote would have to be tallied, Orenstein noted, but the total capacity used would be tiny: “In that case, it’s about IOPS, not the total storage.”

The ioScale consists of single half length PCIe slot, with a single controller, enabling a small form factor server to scale to 12.8 TB or more. Servers supporting UEFI can boot from Fusion ioScale, and the device includes support for atomic multiblock native writes, eliminating the need for a double write to flash memory to maintain correctness—with a corresponding 30 to 40 percent increase in performance, Orenstein said.

Fusion ioScale will be showcased at the Open Compute Summit this week in Santa Clara, where the company suggested it will appear within a Facebook server. Other announcements expected this week at the Summit include AMD’s first Open Compute-compliant server board, as well as a logo and certification program instituted by OCP members.


Image: Teerawut/Shutterstock.com