Exploiting Flash for Energy Efficient Disk Arrays

A growing concern, energy consumption in data centers has been the focus of numerous white papers, research studies, and news reports [1, 11, 4, 3]. According to a report to U.S. congress [11], the total energy consumption by servers and data centers in U.S. was about 61 billion kWh in 2006, and is projected to nearly double by 2011 [11]. Among the components in data centers, it has been shown that storage experienced the fastest annual growth (20% between 2000 and 2006) in energy consumption [11]. A key goal in energy efficient system design is to achieve energy proportionality [2], i.e., energy consumption being proportional to the system utilization. However, hard disk drives (HDDs), the dominant technology for data storage today, contain moving components, making it difficult to achieve this goal. For example, an enterprise class Seagate Cheetah 15K.4 HDD consumes about 15W under load and 12W when idle for spinning the disk platters [9]. While a disk can be spun down to standby mode for saving energy, it takes on the order of 10 seconds to spin up a disk, potentially incurring significant slowdowns in application response times. Previous Approach: Exploit Redundancy and NVRAM. One promising solution is to exploit the inherent redundancy in storage systems for conserving energy [5, 12, 7]. Today, most storage systems employ redundancy (e.g., RAID) to achieve high reliability, high availability, and high performance requirements for many important applications. For example, the TPC-E benchmark, which models transaction processing in a financial brokerage house, requires redundancy for the data and logs [10]. As seen by published TPC-E reports on the TPC web site, this requirement is typically achieved by the use of RAID 10 disk arrays. For saving energy, the idea is to keep only a single copy of the data active and spin down disks containing redundant copies of the data under low load. We call the disks containing the active copy of data, the active disks, and the disks that are spun down, the standby disks. In order to guarantee the same level of reliability for write operations under low load (e.g., writing to two non-volatile devices), previous studies [5, 12, 7] propose to use NVRAM (i.e., battery-backed RAM) as non-volatile write buffers. When the system is under low utilization, reads are sent to the active disks, while writes are sent to both the active disks and to the NVRAM buffers. When the system sees high load or when the NVRAM buffers are full, the standby disks are spun up and the buffered writes are applied to bring the disks up to date. Limitations of Previous Approach. There are two main limitations of the previous approach. First, battery-backed RAM is expensive. Its size is often limited to a few hundred MB for a RAID array. Typically, a server-class disk can support about 100MB/s read/write bandwidth. Suppose that under low load, a disk sees 1MB/s write traffic. Then, a 500MB NVRAM buffer will be filled up for the write traffic of a single disk in less than 9 minutes. When the buffer is filled, the standby disks must be spun up to apply the buffered writes. However, a disk supports only a limited number of spin-up/down operations because they introduce wear to the motor and the heads in a disk. In particular, server and desktop disks are often rated at 50,000 spin-up/down cycles (a.k.a. startstop cycles) [8]. Given a five-year lifetime, this puts a limitation of an average 1.1 spin-up/down per hour. Therefore, the above example with a single standby disk will significantly shorten the