Adaptive Algorithms
For random accesses, a new adaptive algorithm uses the history of previous requests to assign read operations in a way
that minimizes the movement of the disk arm. These optimizations have shown superlinear performance gains on random
read operations. Superlinear means that performance gains are better than linear, with two drives giving greater than two
times the performance of one drive. In the results shown in Figure 2, the performance gain is about 2.3 times that of a single
drive. The reason for this outstanding result is that there are twice as many actuators and each travels less distance than the
average distance when only one drive is used.
In most RAID 1 configurations, each I/O is directed to one of the disks and there is no performance improvement if small
fixed-length stripes are read alternately from the two drives. For instance, if disk 0 reads the even 32K stripes and disk 1
reads the odd 32K stripes, both disks transfer half the time and spend the other half of the time waiting for the head to pass
over data being read by the other drive. This phenomenon is shown on the left side of Figure 2 with small stripe sizes. As the
stripe size is increased, it eventually passes the point where the amount of data being skipped is equal to one track.
At this point the data rate increases sharply because there is almost no time wasted for the head to pass over data being
transferred by the other drive. At the first peak, the data rate is not quite equal to reading the drive sequentially because one
extra disk skew is required when skipping the track read by the other drive. Later peaks have higher bandwidth because the
extra skew is spread across more tracks of transferred data. The position of the peaks and the performance at each peak
vary depending on the bit density, RPM of the drive and the particular zone being measured.
Figure 2. Performance Sensitivity to Stripe Size
The storage switch takes advantage of this phenomenon by setting a stripe size at one of these peaks and simultaneously
accessing alternating stripes from the two drives. In this way, long sequential reads run at nearly twice the rate of a single
drive. The peaks shift to the left at each zone crossing when moving from the outer diameter of the disk toward the inner
tracks. For optimal performance, the zone table is consulted at each zone crossing in order to set the stripe size to the
optimal value for that zone. The combination of the sequential and random access methods gives improved performance
over a wide range of applications
