A number of previous storage systems were designed to take into consideration the tradeoff between capacity and performance. Hou and Patt [10] performed a simulation study of the tradeoff between mirroring and RAID-5.
The HP AutoRAID incorporated both mirroring and RAID-5 into a two-level hierarchy [28]. The mirrored upper level provided faster small writes at the expense of consuming more storage, while the RAID-5 lower level was more frugal in its use of disk space. Its primary focus was solving the small write problem of RAID-5.
We have taken the tradeoff between capacity and performance a step further by 1) improving latency and throughput of all I/O operations, 2) being able to benefit from more than twice the excess capacity, and 3) providing a means of systematically configuring the extra disk heads.
The HP Ivy project [15] was a simulation study of how a high degree of replication could improve read performance. Our study differs from Ivy in several ways. First, Ivy only explored reducing seek distance and left rotational delay unresolved. Second, Ivy only examined mirroring. The third difference is a feature of Ivy that we intend to incorporate into our system in the future: Ivy dynamically chose the candidate and the degree of replication by observing access patterns. We are currently researching a wide range of access patterns (including those at the file system level) that can be used to dynamically tune the array configuration.
Matloff [17] derived a model of linear improvement of seek distance as one increased the number of disks devoted to striping. Bitton and Gray derived a model of seek distance reduction [3] and studied seek scheduling [2] for a D-way mirror. Neither study considered the impact of rotational delay.
Dishon and Liu [6] considered latency reduction on either synchronized or unsynchronized D-way mirrors. A synchronized mirror can reduce foreground propagation latency because the multiple copies can be written at nearly the same time if we insist that the replicas are placed at rotationally identical positions. This advantage comes at the cost of poor read latency because it allows no rotational delay reduction for reads.
Polyzois [20] proposed careful scheduling of delayed writes to different disks in a mirror to maximize throughput, a technique that can potentially benefit delayed writes in our systems when the replicas are on different disks.
The ``distorted mirror'' [19] provided an alternative way of improving the performance of writes in a mirror. It performed writes initially to rotationally optimal but variable locations and propagated them to fixed locations later. This technique can be integrated with our delayed write strategy as well.
Lumb et al. [16] exploited ``free bandwidth'' that is available when the disk head is in between servicing normal requests in a busy system. The free bandwidth was used for background I/O activity. Propagating replicas in our system is a good use of this free bandwidth.
Ng examined intra-track replication as a means of reducing rotational delay [18]. We extend this approach to improve large I/O bandwidth by performing rotational replication across different tracks.
The importance of reducing rotational delay has long been recognized. Seltzer and Jacobson independently examined a number of disk scheduling algorithms that take rotational position into consideration [14,23]. Our work considers the impact of reducing rotational delay in array configurations in a manner that balances the conflicting goal of reducing seek and rotational delay at the same time.
At the time of this writing, the Trail system [12] independently developed a disk head tracking mechanism that is similar to ours. Trail used this information to perform fast log writes to carefully chosen rotational positions. A similar write strategy was in use in the earlier Mime system [5], but Mime relied on hardware support for its rotational positioning information. Aboutabl et al. developed a similar disk timing measurement strategy, which was used to model the response time of individual I/O requests [1].
A number of drive manufacturers have incorporated SATF-like scheduling algorithms in their firmware. An early example was the HP C2490A [9]. Our host-based software solution enables the employment of such scheduling on drives that do not support it internally. Furthermore, it allows experimentation with strategies such as rotational replica selection, strategies that would have been difficult to realize even on drives that support intelligent scheduling internally. On the other hand, if the drive does support intelligent internal scheduling, an interesting question that this study has not addressed is how we can adapt our algorithm for such drives without relying on complex predictions.
One of our goals of studying the impact of altering array configurations is to understand how to configure a storage system given certain cost/performance specifications. The ``attribute-managed storage'' project [7] at HP shares this goal, although its focus is at the disk array level as opposed to individual drive level.