In this paper, we set out to answer a simple question: how do we systematically increase the performance of a disk array by adding more disks?
This question is motivated by two phenomena. The first is the presence of a wide variety of performance-enhancing techniques in the disk array literature. These include striping[17], mirroring[3], and replication of data within a track to improve rotational delay[18]. All of these techniques share the common theme of improving performance by scaling the number of disks. Their performance impacts, however, are different. Given a fixed budget of disks, an array designer faces the choice of what combination of these techniques to use.
The second phenomenon is the increasing cost and performance gap between disk and memory. This increase is fueled by the explosive growth of disk areal density, which is at an annual rate of about 60% [8]. On the other hand, the areal density of memory has been improving at a rate of only 40% per year [8]. The result is a cost gap of roughly two orders of magnitude today.
As disk latency has been improving at about only 10% per year [8], disks are becoming increasingly unbalanced in terms of the relationship between capacity and latency. Although cost per byte and capacity per drive remain the predominant concerns of a large sector of the market, a substantial performance-sensitive (and, in particular, latency-sensitive) market exists. Database vendors today have already recognized the importance of building a balanced secondary storage system. For example, in order to achieve high performance on TPC-C [26], vendors configure systems based on the number of disk heads instead of capacity. To achieve D times the bandwidth, the heads form a D-way mirror, a D-way stripe, or a RAID-10 configuration [4,11,25], which combines mirroring and striping so that each unit of the striped data is also mirrored. What is not well understood is how to configure the heads to get the most out of them.
The key contributions of this paper are:
More specifically, we present a disk array configuration, called an SR-Array, that flexibly combines striping with rotational replication to reduce both seek and rotational delay. The power of this configuration lies in that it can be flexibly adjusted in a balanced manner that takes a variety of parameters into consideration. We present a series of analytical models that show how to configure the array by considering both disk and workload characteristics.
To evaluate the effectiveness of this approach, we have designed and implemented a prototype disk array that incorporates the SR-Array configurations. In the process, we have developed a method for predicting the disk head location. It works on a wide range of off-the-shelf hard drives without special hardware support. This mechanism is not only a crucial ingredient in the success of the SR-Array configurations, it also enables the implementation of rotational position sensitive scheduling algorithms, such as Shortest Access Time First (SATF) [14,23], across the disk array. Because these algorithms involve inter-disk replicas, without the head-tracking mechanism, it would have been difficult to choose replicas intelligently even if the drives themselves perform sophisticated internal scheduling.
Our experimental results demonstrate that the SR-Array provides an effective way of trading capacity for improved performance. For example, under one file system workload, a properly configured six-disk SR-Array delivers 1.23 to 1.42 times lower latency than that achieved on highly optimized striping and mirroring systems. The same SR-Array achieves 1.3 to 2.6 times better sustainable throughput while maintaining a 15 ms response time on this workload.
The remainder of the paper is organized as follows. Section 2 presents the SR-Array analytical models that guide configuration of disk arrays. Section 3 describes the integrated simulator and prototype disk array that implement the SR-Array configuration models. Section 4 details the experimental results. Section 5 describes some of the related work. Section 6 concludes.