Observing the Effects of Multi-Zone DisksBy Rodney Van Meter, Information Sciences Institute, University of Southern California
Summary by Peter Collinson
Data are stored on disks in concentric rings, placing sectors around the disk. These rings are called "cylinders," because a single disk usually comprises several surfaces, each with its own head, and the rings that the heads traverse make a hollow cylinder. Recent disk design has increased storage space by recognizing that the circumference of a circle is related to its radius and more sectors can be added to each track as the circumference increases, assuming that the disk is spinning at a constant rate. The sectors are added in regular increments, and the effect is that each disk consists of areas (or zones) containing different numbers of sectors in each cylinder. The technique is called Zoned Constant Angular Velocity (ZCAV).
However, UNIX filesystems were designed for disks that had a constant number of sectors per track. When using ZCAV disks, the system ignores the actual disk geometry. No one has performed a study of the effect of ZCAV on data throughput or has looked at filesystem throughput when ZCAV disks are used by the standard UNIX filesystem. This paper provides such a study.
The study contains a description of the hardware, an analytical analysis of the system that predicts what should be happening, and a description of an experiment that measured the real performance when a standard FFS filesystem is placed on the disk.
The author came to several conclusions. First, it is difficult to get accurate information about the real layout of ZCAV disks. The vendors don't often provide it in manuals or in the information that the disk is supposed to provide about itself via the SCSI interface. It's often necessary to map the disk using seek to move to a particular linear block and then reading back the (cylinder, head, sector) information.
Second, the raw throughput of the disk at the rim may be as much as 90% faster than the throughput at the center; the drive measured was approximately 75% faster. A linear model that relates to the block number can be used to predict the degradation in performance as the disk nearer the central spindle is used. Block zero is usually at the rim of the disk, so it's clear that you should try to place your most frequently accessed large files in the low numbered blocks of the filesystem.
Third, an experiment on the filesystem showed that the performance dropped roughly 25% (or, rose roughly 33%). This effect matched the linear model acceptably. Larger effects were not seen because the filesystem was on one partition covering the inner 70% of the disk drive. Greater differences can be expected on partitions covering whole drives or between partitions on drives.
These effects should be considered when designing and benchmarking filesystems.
The paper and the slides for the talk are available on https://www.isi.edu/netstation/zcav.
Originally published in ;login: Vol. 22, No.2, April 1997.
Last changed: May 28, 1997 pc