GoKawiilDiscovering Hard Disk Physical Geometry through Microbenchmarking
Modern hard drives store an incredible amount of data in a small space, and are still the default choice for high-capacity (though not highest-performance) storage. While hard drives have been around since the 1950s, the current 3.5″ form factor (actually 4″ wide) appeared in the early 1980s. Since then, the capacity of a 3.5″ drive has increased by about 106 times (from 10 MB to about 10 TB), sequential throughput by about 103 times, and access times by about 101 times. Although the basic concept of spinning magnetic disks accessed by a movable stack of disk heads has not changed, hard drives have become much more complex to enable the increased density and performance. Early drives had tens of thousands of sectors arranged in hundreds of concentric tracks (sparse enough for a stepper motor to accurately position the disk head), while current drives have billions of sectors (with thousands of defects), packed into hundreds of thousands of tracks spaced tens of nanometers apart.
Beyond just the high-level performance (throughput and seek time) measurements, which drive characteristics can be characterized using microbenchmarks? I had initially set out to detect the number of platters in a disk without opening up a disk, but in modern disks this simple-sounding task requires measuring several other properties before inferring the count of recording surfaces. Characterizing disk drive geometry has been done in the past [1, 2], and the algorithms I used aren’t very different. However, older algorithms often make assumptions that are no longer true on modern drives. For example, the Skippy [2] algorithm (a fast algorithm to measure the number of surfaces, cylinder switch times, and head switch times) no longer works on modern drives because the algorithm assumes one particular ordering of tracks onto multiple platters that is no longer used on modern disks (that several head switches occur before a seek to the next cylinder).
Hard disk drives store data on a stack of one or more rotating magnetic disks. Data is written in concentric tracks. A stack of read/write heads move (radially) across the disks to position the head over the desired track. There are typically two heads per platter (one for each side), and the entire stack of heads move together as a single unit. Reading data occurs by moving the disk head to the desired track (a seek), waiting until the beginning of the desired data passes under the disk head, and then continuing to read sequentially until either all of the requested data is read, or the end of the track, when the head needs to be moved to the next track. A hard drive’s “geometry” describes how data is arranged into platters, tracks, and sectors. Historically, this was described using three numbers: Cylinders (number of concentric rings from outside to inside), Heads (number of recording surfaces, or the number of tracks per cylinder), and Sectors per track, leading to the well-known acronym CHS. The capacity of a hard drive in sectors is simply C×H×S. Today, C and S are variable and only H is still constant. The number of tracks is not necessarily the same on each recording surface, and the number of sectors per track varies across the disk (more sectors in the longer outer tracks than the inner tracks).
This article describes several microbenchmarks that try to extract the physical geometry of hard disk drives, and a few other related measurements. These measurements include rotation period, the physical location (angle and radius) of each sector, track boundaries, skew, seek time, and some observations of defective sectors. I use these microbenchmarks to characterize a variety of hard drives from 45 MB (1989) to 5 TB (2015). There is no attempt to characterize other important performance aspects such as caching. The remainder of this article begins with a background on hard drive geometry. It then describes the collection of microbenchmarks, starting with a basic read access time measurement and building towards increasingly complex algorithms. The second part of the article presents microbenchmark measurements for each of the 17 drives that were tested.
Summary Background: Hard drives consist of spinning disks, and a stack of heads. Data is arranged on recording surfaces (2 sides per platter), tracks, and sectors. “Cylinders” no longer exist in drives newer than around 2000.
What can be measured: RPM, angular position of every sector, and seek times, by timing specific sequences of sector reads. These basic measurement methods can then be used to find track boundaries, how tracks are arranged on a surface, and the number of surfaces.
Access time: Out of the drives I tested, it takes 1.3 to 3.6 revolutions to do a full-stroke seek. Heads accelerate slowly: Very few tracks are accessible in the first revolution. Short stroking offers limited reduction in seek times because even short seeks take a relatively long time.
Seek time is non-trivial to measure. A seek time plot can be used to observe acoustic management (AAM). AAM slows down long-distance seeks to reduce noise, but not short-distance seeks.
Track boundaries can be found by searching the disk for track skew. Newer disks use different densities (track size) on each surface. Track density and bit density can be estimated by knowing track count and size. In the newest drive I have, average track pitch is 80 nm and an average bit is 17 nm in length.
... continue reading