Evolutionary Storage

Evolutionary Storage is proposed to counteract the disk devolution process. Degradation in performance over time is a serious and common concern in magnetic disks. Sometimes, people benchmark their disks when new, and then many months later, and are surprised to find that the disk is getting slower. In fact, the disk most likely has not changed at all, but the second benchmark may have been run on tracks closer to the center of the disk. Modern hard disks employ zoned bit recording (ZBR): tracks are grouped into zones. The raw data transfer rate of the disk when reading the outside cylinders is much higher than when reading the inside ones. This is because the tracks on the outside are much longer than the ones on the inside, typically doubling the circumference or more, therefore containing more data, but the angular velocity of the platters is constant regardless of which track is being read.

An evolutionary storage system uses object frequency to cluster frequent data. It is a conceptually simple technique for dramatically improving disk performance. Two different three-level architectures may be used: inner-outer-cache (fast band disk implementation as a comparison) and disk-flash-cache (Hybrid Disk implementation). Using a real system disk trace, the experimental results showed that the performance speedup due to the disk re-organization is 2.4 for the fast band disk and 4.0 for the hybrid disk, respectively. The Hybrid Disk outplays the fast band disk at little additional cost on a flash memory.

This technique can be implemented in a device driver (or controller). No changes to the file system are required. This is a “black box” design, which automatically infers object frequency in a disk drive. Unlike this, “White Box” or “Gray Box” design relies on storage front-ends, such as a file system, to directly or indirectly pass semantic information to obtain object frequency in a storage system.

The strategy is to move data to their optimal locations during idle periods, so as not to conflict with system use. Much of the benefit of disk re-organizing could be obtained from reorganizing only 10% of the data. For example, it takes less than 1 hour to move 50% of fast band data out of the fast band and 1 hour to move the same amount of data to fill the fast band. It is still practical for a storage system, because frequency distribution is relatively stable, it is unnecessary to keep swapping the data between the fast band/flash memory and slow band. Instead, it might be acceptable to spend one hour every week on running Object Mover to update the layout.