Is this sucker nuclear?! —

Apple Fusion Drive—wait, what? How does this work?

New feature sounds less like hybrid drive caching and more like auto-tiering.

We got a slide this morning, but does anyone understand the Apple Fusion Drive yet?
We got a slide this morning, but does anyone understand the Apple Fusion Drive yet?

Apple's new iMac announcement today included an interesting bit of information on an upcoming technology Apple calls "Fusion Drive." According to Phil Schiller this morning, the technology takes a relatively small solid state disk and a relatively large spinning hard disk drive, then "fuses" them together into a single drive.

Speculation in the Ars forums started immediately, with most wondering if "Fusion Drive" works the same as current hybrid disk drives. Those incorporate some amount of NAND flash inside a traditional hard disk drive as an extended cache. Others speculated the Apple technology resembles something like Intel's Smart Response Technology, which uses a dedicated SSD (of up to 64GB in size) as a transparent cache for a larger hard disk drive.

Technical details are scarce, but based on Schiller's descriptions, the answer to both of those questions appears to be "no." Apple's Fusion Drive does not appear to function like an SSD-backed disk cache, but rather seems more like a file-level implementation of a feature that has existed for some time in big enterprise disk arrays: automatic tiering.

Most big disk arrays have different types of storage—some slow spinning disk, some faster spinning disk, and some solid state storage—and some have the ability to monitor what data is being accessed the most and can automatically move that data to a faster tier of disk as needed. These features typically operate at the block level, below the files, and can be done on large or small chunks of data, depending on what's hot and what's not. Auto-tiering also includes the ability to take data that is no longer in demand, or no longer "hot" and demote it down off of fast disk and onto slower stuff. In this way, a file that doesn't get accessed very often might be stored on slow SATA disks, but if a hundred people need to open it repeatedly over a short period of time, it will get pulled up and kept on SSD until it's not needed anymore.

Based on Schiller's explanation, Fusion Drive sounds similar. In a caching solution, like Intel's, files live on the hard disk drive and are temporarily mirrored to the SSD cache as needed. In an enterprise auto-tiering situation, and with Fusion Drive, the data is actually moved from one tier to another, rather than only being temporarily cached there.

Schiller noted that out of the box, a Fusion Drive-equipped Mac will have its core operating system components and preinstalled applications placed on the SSD side of the house, while documents and applications live on the slower spinning disk. As you open files and documents and install applications, the operating system makes note of what you're doing and how often you do it, and the things you use most often are promoted up onto the SSD. This is done transparently, and you see the SSD and hard drives as a single volume.

This is almost certainly done using Apple's Core Storage logical volume manager, introduced in OS X 10.7. As a volume manager, Core Storage has the chops to weld the two separate drives together into a single entity and handle the relocation of files between one tier and the other. I say "files" because it's almost certain that the implementation here is file-level instead of block-level, since a single user doesn't need the granularity that a block-level tiering solution would provide—as users, we tend to care about files and applications. OS X needs only a method of keeping track of how many times and how often a file or application is used and then can automatically have that file slipped onto fast SSD or slow disk when it crosses an upper or lower threshold. The tracking might be implemented through Core Storage attributes or through an external on-disk database like Spotlight and document revisions.

Apple is currently mum, though with the announcement still less than a few hours old, we're sure more information will begin to trickle out shortly. It seems certain, though, that this is not a caching implementation like Intel's but rather actual movement of files between physical disks by the Core Storage volume manager.

Channel Ars Technica