2.5 Comparing SR-Array with Striped Mirror

In an SR-Array, all replicas exist on the same disk. Removing this restriction, we can place these replicas at rotationally even positions on different disks in a ``synchronized'' mirror, a mirrored system whose spindles are synchronized. We call this layout strategy a striped mirror, one flavor of the RAID-10 systems known in the disk array industry. (RAID-10 is a broader term that typically does not necessarily imply the requirement of synchronized spindles and the placement of replicas at rotationally even positions.) To make the performance of the striped mirror competitive to a corresponding SR-Array, we must choose replicas intelligently based on rotational positioning information.

Even with these assumptions, a striped mirror is not equivalent to an SR-Array counterpart. Consider a simple example involving only two disks: blocks A and B reside on different disks in an SR-Array; but each of the disks in a corresponding striped mirror has both blocks. Now consider a reference stream of AAB. On an SR-Array, the two accesses to A are satisfied by two rotational replicas on one disk, consuming less than a full rotation in terms of rotational delay; and the access to B is satisfied by a different disk so its access time is independent of the first disk and the first two accesses. In an attempt to emulate the behavior of the SR-Array, we must send the two accesses to A to the two replicas on different disks in a striped mirror; but now the access time of B is affected by the first two accesses because both disks are busy. In general, it is impossible to enforce identical individual access time for a stream of requests to an SR-Array and a striped mirror.

Statistically, the read latency of a striped mirror should be slightly better than the latency on a corresponding SR-Array. This is because the average of the minimum of the sum of seek and rotational delay is smaller than the sum of the average seek and average minimum rotational delay.

In terms of throughput, the simple example above shows that for an arbitrary stream of requests, there does not exist a general schedule on a striped mirror that is equivalent to that on a corresponding SR-Array. We do not pretend to know how to optimally choose replicas on a striped mirror. Section 3 discusses a number of heuristics. The performance of our best effort implementation of a striped mirror has failed to match that of an SR-Array counterpart.

In terms of feasibility, as spindle synchronization is becoming a rarer feature on modern drives, one can only approximate striped mirrors on unsynchronized spindles. In terms of reliability, a striped mirror is obviously better than an SR-Array.

In general, it is possible to combine an SR-Array with a striped mirror to achieve the benefits of both approaches so that some of the replicas are on the same disk and some are on different ones. The result is the most general configuration: a $D_s \times D_r \times D_m$ ``SR-Mirror'', where D_s implies that only 1/D_s of the space is used (to reduce seek time), D_r is the number of replicas on the same disk, and D_m is the number of replicas on different disks. A $D \times 1 \times 1$ system is D-way striping. A $1 \times 1 \times D$ system is a D-way mirror. A $D_s \times D_r \times 1$ system is an SR-Array. A $D_s \times 1 \times 2$ is the most common RAID-10 configuration. We may approximate the performance of an SR-Mirror by replacing D_r in the SR-Array models with $D_r \times D_m$.