Fusion-io founders David Flynn (middle) and Rick White (right), together with chief scientist Steve Wozniak. Image: DEMO

The Bad Idea That Spawned Fusion-io

If you know anything about Fusion-io, you know that its chief scientist is Steve Wozniak, the man who founded Apple alongside high-school friend Steve Jobs. Woz, as he’s affectionately known, serves as a kind of walking PR campaign for Fusion-io’s new breed of memory, but the idea wasn’t his. It sprang from Rick White and David Flynn.

Before they founded Fusion-io, White and Flynn worked together at a company called Realm Systems. Realm sold a computer that fit into your pocket. It was designed to take your online identity anywhere — you would plug it into another computer and identify yourself with a fingerprint scan or the like, and it would instantly log into all your software applications — but the thing never really took off, in part because the company’s funding dried up.

‘Having worked with supercomputers, I knew that the biggest problem with these big systems was the lack of memory.’

— David Flynn

As they brain-stormed ideas for a new company, White suggested they repurpose the processors inside those tiny Realm machines. If they wired hundreds of them together, he told Flynn, they could create a new kind of low-power supercomputer capable of running massive applications inside the data center. Flynn, an engineer who had once built supercomputers for the likes of the Lawrence Livermore National Labs, told White that was a really bad idea — “He reminded me that processors on those tiny mobile computers couldn’t do floating point,” White remembers, referring to the complex calculations typically handled by server chips — but it sparked another.

Those tiny Realm computers also included flash. They used flash in lieu of hard drives, and as a kind of supplement for the DRAM that made up their memory subsystems. It wouldn’t make sense to wire all those tiny processors together, Flynn realized, but you could remake big servers so that they looked more like the Realm machines. You could remake big servers with flash.

Many companies were already offering solid-state drives, or SSDs. These are flash devices designed to mimic hard drives. They look like hard drives, and they plug into the same connectors. But Flynn aimed for something different. He wanted to build a flash device that could also mimic a server’s memory subsystem. “Having worked with supercomputers,” Flynn says, “I knew that the biggest problem with these big systems was the lack of memory.”

Today’s server processors, you see, are designed to use only so much DRAM. Theoretically, this limit is 4 terabytes, aka 4,000 gigabytes, but most servers are equipped with far less. Flynn’s idea was to build what amounted to a second memory subsystem — one that could house many more terabytes of data — and the result was the Fusion-io flash card. It didn’t plug into a hard drive controller. It plugged straight into the PCI Express bus, or PCIe bus, that shuttles information across a server. In other words, this card sat closer to the processor. It bypassed all the extra hardware and protocols needed to operate a hard disk.

In hindsight, this seems like an obvious thing to do. But it was very much a departure from the norm. It’s not just that companies were accustomed to using just hard drives and DRAM inside the server. They stored most of their data outside the server, on vast arrays of hard disks — big-beefy systems called storage area networks, or SANs. Packing its card with such enormous amounts of flash, Fusion-io sought to replace these massive appliances as well. It called the card a “SAN in your hand.”

Facebook and the New Memory

Fusion-io uncloaked its cards in the fall 2007, and its first big customer was Facebook. David Flynn was friends Paul Saab, a Facebook software developer and a well-known contributor to the FreeBSD open source operating system, and the relationship between the two companies progressed from there. Nowadays, Fusion-io will even design new cards in tandem with Facebook engineers.

Apple came later. The natural assumption is that this relationship was driven by Steve Wozniak — who joined Fusion-io as chief scientist in late 2008 — but according to Fusion-io, this wasn’t the case. In typical fashion, Apple didn’t respond to a request to discuss its use of flash cards — and Fusion-io demurred — but Facebook is another matter. Unlike Apple, the social networking giant is willing to open the curtain on the hardware and software used inside its data centers, and its new-age systems give you a pretty good idea of how these flash cards can reshape the server.

‘It creates this new platform architecture a couple of years down the road — where the application can now be more judicious about how to balance the performance and the power.’

— Kushagra Vaid

Initially, Facebook used the cards in the machines that run its database software. This is where companies typically use flash SSDs as well. Basically, the cards replaced hard drives. Housing its database software on flash rather than hard disks, Facebook can significantly speed the delivery of status updates and comments and other social networking chatter.

“When a request gets to the database tier, we want to be able to serve it up really, really quickly,” Facebook’s Frank Frankovsky once told us. “It’s a significant improvement on that overall round-trip time to the user.” What’s more, these flash devices can reduce hassle. As Frankovsky says, a hard disk breaks down more than any other device in the data center. Today, Facebook even boots its database servers from these flash cards.

But Facebook is now moving this sort of flash card into other machines, and it’s replacing memory as well as hard drives. Traditionally, Facebook operated another tier of servers that ran something called Memcached. These machines would, well, cache data in memory. The idea was to provide even quicker access to the stuff that users looked at the most. With this data, you wouldn’t even have to visit the database tier. This worked well enough, but Facebook has now cranked the dial on these servers as well, creating a new software platform that caches data on flash cards rather than main memory.

With this new platform — known as McDipper — the company can cache much larger amounts of data. “Flash provides up to 20 times the capacity per server and still supports tens of thousands of operations per second,” the company said in a recent blog post. “It was the obvious answer to this problem.”

This sort of setup is still significantly slower than main memory, says Microsoft’s Kushagra Vaid. But it’s fast enough, and it consumes less power. When you reach the size of a Microsoft or a Facebook or an Apple, keeping power down is just as important as anything else.

The point is that flash gives you more options — and in the long run, Vaid says, these options will only expand. “It creates this new platform architecture a couple of years down the road,” he explains, “where the application can now be more judicious about how to balance the performance and the power — by choosing the appropriate balance between system memory and non-volatile memory.”

Uh, We’re a Software Company

But, yes, Facebook built McDipper from scratch. And that too shows you how much these flash cards are changing the data center. Today’s software isn’t designed to run on flash cards. You’re forced to write new software.

‘The storage devices that are coming out now are almost nothing like the storage devices that people have been using for the past few decades. They break a whole bunch of rules and work in very different ways and do a whole lot of things that just weren’t possible before.’

— Chris Mason

That’s why Fusion-io has engineers working inside the data centers run its largest customers, including Apple and Facebook. And it’s why the company employs several software developers known for their work on the Linux kernel, the software at the heart of the open source operating system that runs most machines inside the big internet data centers. This includes Chris Mason, who oversees a Linux file system called Btrfs (pronounced “butter fs”) and Jens Axboe, who built several tools related to data input and output and other tasks inside the Linux kernel.

“The storage devices that are coming out now are almost nothing like the storage devices that people have been using for the past few decades. They break a whole bunch of rules and work in very different ways and do a whole lot of things that just weren’t possible before,” says Mason. “But you have to change the core software to enable this.”

The aim, explains Mason and CEO David Flynn, is to tune Linux and other software to get the most out of these cards, but also to make it easier for companies to build their own software. They company now offers a software development kit, or SDK, for building tools atop the cards. In the end, says Flynn, Fusion-io is a much a software company as a hardware company.

This is borne out by the company’s recent contribution to the Open Compute Project, a Facebook-led effort to create new hardware designs for the data center — and share those designs with the rest of the world. This winter, Fusion-io open sourced the design of its flash card, including everything but the controller that drives the card. The aim is to feed the creation of still more flash devices that can run its new breed of software.

The hope, Flynn says, is that computer makers like Quanta and Wistron will build flash straight into the server motherboard, pushing it even closer to the processor. After all, that’s where the data center is headed.

Robert McMillan contributed to this story.