Server Storage Systems Use Numerous Disks to Achieve High Performance, Thereby Consuming a Significant Amount of Power. Intradisk Parallelism Can Significantly Reduce Such Systems’ Power Consumption by Letting Disk Drives Exploit Parallelism in The

......Many applications today, used by millions of people around the clock, store and process massive data sets. These data-centric computing applications include transaction processing, search engines, and e-mail services as well as newer types of applications, such as social networking and photo and video sharing. The amount of data these applications must handle is growing tremendously, and experts estimate that the amount of data generated and stored worldwide will reach nearly 1,000 exabytes in a few years. Data centers will need to store and centrally manage this data, and applications will use it to process and deliver content to users. In addition to storage capacity, storage systems within data centers must provide these data-intensive applications with high I/O performance. The conventional approach to building high-performance storage systems is to aggregate several disk drives and form disk arrays. However, because disk access times are relatively large, high-performance storage systems use a large number of disks, and performance (rather than capacity) is the primary driver for selecting the number of disks. Moreover, many high-performance storage systems also tend to underutilize disk capacity to leverage the higher data rates on the outer zones of the platters and reduce the impact of disk-arm positioning delays. This approach to storage system design has resulted in a significant increase in data centers’ storage power consumption. Moreover, researchers expect disk drive performance to improve relatively slowly due to limitations in the magnetic recording technology and thermal constraints associated with making the platters spin faster. Although modern disk drives offer some parallelism in the form of prefetching and tagged command queuing, which helps aggregate multiple I/O requests within the drive to improve disk arm scheduling, achieving good parallel I/O performance requires multiple .......