Power-efficient Server-class Performance from Arrays of Laptop Disks - PDF

Please download to get full document.

View again

of 16
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Information Report
Category:

Business

Published:

Views: 55 | Pages: 16

Extension: PDF | Download: 0

Share
Related documents
Description
Power-efficient Server-class Performance from Arrays of Laptop Disks Athanasios E. Papathanasiou and Michael L. Scott The University of Rochester Computer Science Department Rochester, New York Technical
Transcript
Power-efficient Server-class Performance from Arrays of Laptop Disks Athanasios E. Papathanasiou and Michael L. Scott The University of Rochester Computer Science Department Rochester, New York Technical Report 837 May 4 Abstract The disk array of a server-class system can account for a significant portion of the server s total power budget. Similar observations for mobile (e.g. laptop) systems have led to the development of power management policies that spin down the hard disk when it is idle, but these policies do not transfer well to server-class disks. On the other hand, state-of-the-art laptop disks have response times and bandwidths within a factor of 2.5 of their server class cousins, and consume less than one sixth the energy. These ratios suggest the possibility of replacing a server-class disk array with a larger array of mirrored laptop disks. By spinning up a subset of the disks proportional to the current workload, we can exploit the latency tolerance and parallelism of typical server workloads to achieve significant energy savings, with equal or better peak bandwidth. Potential savings range from 5% to 8% of the total disk energy consumption. This work was supported in part by NSF grants numbers EIA-8124, CCR , and CCR-4344, by DARPA/AFRL contract number F2961--K-182, and by Sun Microsystems Laboratories. 1 Introduction Data centers capable of providing Internet, application, database and network services are an increasingly important component of the world s computing infrastructure. In 1995 there were, servers in the world. As of June 1, that number had reached six million [Maximum Throughput Inc, 2]. Most existing research on data center design has aimed to improve performance, reliability and availability. Recently, however, researchers have begun to recognize the importance of energy efficiency [Chase and Doyle, 1; Bohrer et al., 2]. Scientists at the Lawrence Berkeley National Laboratory have estimated that the annual energy used by servers, minicomputers, and mainframes exceeded 19 terawatt-hours (TWh) in 1999 [Kawamoto et al., ] (equivalent to the total output of four large commercial power plants). Moreover, the increase in demand for data center services suggests that by 5 new data centers will require approximately TWh or $4B at $1 per MWh unless they become more energy efficient [Chase and Doyle, 1]. Increased data center power consumption translates directly into higher total cost of ownership, attributable to operating power, cooling, and decreased reliability. A recent white paper suggests that disk drives in a data center can account for 27% of total electric consumption [Maximum Throughput Inc, 2]. In some configurations the fraction can be significantly higher. A Dell PowerEdge 665 [DELL, 3], for example, comes equipped with 4 Intel Xeon 2. GHz processors and KRPM hard drives. The processors are rated at 58W each, while an operational SEAGATE ST KRPM 18GB server-class disk drive consumes 15W [Seagate, 3]. In such a configuration the hard disks consume 15 times more power than the processors. Previous work on the energy efficiency of storage systems has focused on the hard drives used in mobile environments. Such drives support multiple non-operational low power modes that can provide significant energy savings during relatively long periods of inactivity (on the order of tens of seconds). The key idea in this work is to place the hard drive into a non-operating low power mode during relatively long periods of inactivity [Douglis et al., 1995; Douglis et al., 1994; Helmbold et al., 1996; Li et al., 1994]. Unfortunately, such policies do not transfer in an obvious way to the server environment. Server-class disks require tens of seconds to move from an inactive to an active state, and consume tens of Watts while spinning up. In order to save energy by spinning down, they must be idle for significantly longer than their laptop-class counterparts. At the same time, server workloads are characterized by disk access patterns that have significantly shorter request inter-arrival times than are typical in mobile workloads. Finally, server-class disks are not designed for frequent stop-start cycles; heavy use of standby mode can be expected to significantly lower their mean time to failure. To improve the energy efficiency of server-class storage systems Gurumurthi et al. have suggested the use of DRPM [Gurumurthi et al., 3], an approach that dynamically modulates disk speed depending on current workload, decreasing the power required to keep the platters spinning when the load is light. Given the load fluctuations typical of web and transaction processing [Bohrer et al., 2], such an approach can save significant amounts of energy during periods of low activity while maintaining adequate response times during peaks. Unfortunately, DRPM disks are not yet commercially available. As an alternative, we suggest replacing each server-class disk in an array with a modest number (e.g. 3) of mirrored, energy-efficient laptop-class disks. This alternative can provide comparable or even improved throughput during workload peaks by exploiting read parallelism, while consuming significantly less energy when the load is light, by spinning down disks that represent excess capacity. The rest of the paper is structured as follows. Section 2 describes the characteristics of laptop-class and server-class disks and makes the case for using the former in server disk arrays. Section 3 describes an array of laptop disks designed to minimize energy consumption. Section 4 shows the energy efficiency potential of the proposed design. Finally, section 5 describes related work and section 6 presents our conclusions. 1 Disk DK23DA-F3 TravelStar 7K6 UltraStar 15K73 Cheetah 15K.3 Price $127 $21 $195 $199 Capacity 3GB 6GB 73.9GB 73GB Form Factor 2.5inch 2.5inch 3.5inch 3.5inch Cache Size 48KB 8192KB 8192KB 8192KB Rotational Speed RP M 7RP M 1537RP M 15RP M Max. Transfer Rate 35M B/s 64M B/sec 129M B/s 112M B/s Avg. Seek Time 13ms 1ms 3.9ms 3.8ms Max. Seek Time 25ms 16ms ms Operating Shock 18G/2ms G/2ms 15G/11ms 6G/2ms Recording Density 6KBP I 624KBP I N/A 533KBP I Track Density 55.KT P I 88.2KT P I N/A 64KT P I Avg. Active Power 2.1W 2.5W 15.6W 14.9W Performance Idle 1.6W 2W 12.W 12.2W Active Idle N/A 1.3W N/A N/A Low Power Idle.6W.85W N/A N/A Standby.25W.25W N/A N/A Sleep.1W.1W N/A N/A Standby to Ready 3sec 3sec 25sec sec Table 1: Characteristics of four mobile and server-class hard drives. Hitachi DK23DA-F3 [Hitachi, 1] is a 1 mobile hard drive, Hitachi TravelStar 7K6 [Hitachi, 3a] is the newest mobile hard drive implemented by the Hitachi Global Storage Technologies and the first mobile hard drive with a rotational speed of 7 RPM. Hitachi UltraStar 15K73 [Hitachi, 3b] and Seagate Cheetah 15K.3 [Seagate, 3] represent modern server-class drives. The prices presented were the lowest prices retrieved from on April 22, 4. The rest of the parameters have been taken from the respective hard drive specifications. The Transfer rate represents the maximum media-to-buffer transfer rate. For the two mobile drives the row Standby to Ready represents the time to transition from the standby low power mode (heads are parked and platters are spun down) to ready (performance idle mode), while for the server-class disks the time from power on to ready. 2 Disk Arrays: Mobile or Server-class Disks? The constant need of data-centric services for higher throughput and shorter response times is driving server-class hard drives toward higher rotational speeds and shorter seek latencies. State-of-the-art server class disks have rotational speeds of 15 thousand RPM and seek times shorter than 4ms. As shown in Table 1, such performance characteristics come at the cost of significantly increased power budgets. Even when idle, a 15 KRPM hard drive consumes more than 12 Watts. In comparison to their server-class counterparts, modern hard disks for mobile systems have worse performance characteristics. However, the need of the mobile system market for longer battery lifetimes makes energy efficiency a very important factor in the development of mobile hard drives. Even when active, modern mobile hard drives typically consume less than 3 Watts. They also typically support several nonoperational low power modes (Table 1): Active Idle, Low Power Idle, Standby and Sleep 1. Non-operational 1 In the Active Idle state the head servoing mechanism and portion of the electronics are off. In the Low Power Idle state the disk is still spinning, but the electronics may be partially unpowered, and the heads may be parked or unloaded. In the Standby state the disk is spun down. The Sleep state powers off all remaining electronics; a soft reset is required to return to higher states. 2 low power modes can save significant energy during periods of disk inactivity. A comparison of the hard disks parameters presented in Table 1 shows that a modern 7 RPM mobile hard disk, such as the Hitachi TravelStar 7K6, has the same cache size as, and similar capacity to, a 15 KRPM server-class disk. The server-class disks provide up to 2.5 times better performance, but consume 6 times as much power, even when the low power modes of the mobile disk are not being used. Such differences suggest that by replacing each high performance disk in a server environment with a mirrored (RAID Level 1) array [Patterson et al., 1988] of several (Table 1 suggests 3) mobile hard disks, one can achieve similar or higher I/O throughput at significantly lower power. Individual request response times will be higher, but for most large secondary storage systems aggregate I/O throughput is (within limits) more important [Chen et al., 1994]. The lower power consumption of mobile hard drives has additional advantages. First, it can lead to improved reliability. Increased temperatures caused by heat dissipation in densely packed storage systems can lead to component malfunctions: for every degree that the operating temperature of a hard disk exceeds the recommended level the failure rate increases two to three percent [Maximum Throughput Inc, 2]. Over time, a hard disk operating at just five degrees above its recommended temperature is 1 to 15 percent more likely to fail. The resistance of mobile disks to operating shocks (Table 1) also increases their reliability. Second, reduced power consumption can lead to a smaller cost for cooling equipment. Third, the smaller form factor, combined with lower heat dissipation, means that several (certainly three) mobile disks can be squeezed into the same physical space as their server-class counterpart, while generating much less acoustic noise. We believe that the attractiveness of mobile hard drives as an alternative to server-class disks will increase over time. The notebook market is growing faster than the desktop and server markets. Rising demand, together with the trend toward higher performance mobile processors, operating systems, interfaces, and buses, is fueling the development of ever faster mobile hard drives. Recent technological advancements, such as adaptive formating [Laroia and Condon, 3], antiferromagnetically-coupled (AFC) media [IBM, 1], fluid dynamic motors [Blount, 1], and fempto sliders [Best et al., 3], have led to faster, more reliable, higher capacity mobile hard drives at (almost) the same power budget. The 4 TravelStar 7K6 hard drive, when compared to the 1 Hitachi DK23DA hard drive (Table 1), has twice the capacity, 1.7 times the rotational speed, and improved seek times, while its idle power consumption has increased by only 25%. The same technologies that have lead to the development of the 7K6 promise additional future improvements. For example the antiferromagnetically-coupled media [Laroia and Condon, 3] and the fempto slider [Best et al., 3] suggest future areal densities of 1Gbit/inch 2, compared to roughly 7Gbit/inch 2 for current commercial products. The above trends make arrays of mobile disks an attractive alternative to high performance server class disks. The principal disadvantage of such an array is its price: with mobile and server-class disks of comparable capacity costing roughly the same amount (Table 1), replacing each server-class disk with several mobile disks will increase the initial investment several fold. This investment may be partially offset by reduced electric bills. In addition, the redundancy of mirroring should eliminate the need for parity disks, allowing an array of n + 1 server-class disks to be replaced by 3n mobile disks, rather than 3n + 3, where n is the parity width. (Striping may still be desirable for high bandwidth sequential reads.) Finally, economies of scale suggest that mobile disk prices may improve faster than those of server-class disks. 3 Design Issues in an Array of Laptop Disks The idea of using arrays of laptop disks in place of server-class disks has as a goal to minimize power consumption while maintaining acceptable aggregate I/O throughput. Based on the discussion in section 2, 3 replacing each server-class disk in a large scale storage system with a mirrored array of three 7 RPM mobile hard disks will provide similar or even better aggregate throughput at half the power budget, even if the disks remain active constantly. Additional power savings can be achieved by taking advantage of the varying workload intensity of web server systems and transaction processing systems [Bohrer et al., 2]. During low intensity periods, only the portion of the hard disks that is necessary in order to sustain acceptable throughput need actually be active. Depending on the exact characteristics of the workload, the additional disks may enter one of several low power modes. The choice of mode depends on the intensity of the workload and more specifically on the rate at which the secondary disks accept requests. Traditional mirrored disk array systems aim to maximize aggregate throughput without regard to power consumption. Hence, common policies used to select the disk to service a request attempt to balance the load evenly across all mirrored disks. Examples of such policies include random selection, round-robin selection, or selection of the disk with the shortest request queue. Such load balancing schemes are inappropriate for power efficiency: the disk array controller may keep all disks active even during light workloads by submitting requests to all disks even during low intensity workloads. A more power-friendly approach would be to use a policy that starts using secondary disks only when individual response times exceed a certain threshold. Such a policy has the advantage of increasing the request inter-arrival time to secondary disks, allowing them to drop into low power modes when the load is low. At the same time, by tracking the response times of individual requests and spinning up additional disks when those times exceed some acceptable threshold, we can guarantee not to damage aggregate throughput. A simple way to limit worst case response time is to track the number of pending requests in the queue of each mirrored disk. New requests can be scheduled to one of the currently active disks until the number of requests in the queue exceeds a certain some queuing threshold, at which point one of the secondary disks can be activated in order to service additional requests. If all disks in the system have exceeded their queuing threshold, a traditional load balancing scheme will become appropriate. Unfortunately, while reads can be spread across disks, writes must be performed on all copies. This may lead to increased response times in write intensive workloads, since the aggregate write throughput is limited to that of a single disk. Power consumption may also increase with a decrease in the length of secondary disk idle intervals, which can lead to inefficient use of low power modes. Fortunately, Internet content delivery is characterized mostly by read activity. It may also be possible to reduce the power impact of writes (though not their performance impact) by updating only those disks that are currently active. Idle disks may be synchronized periodically using data from the primary disks or from a disk array write cache. We plan to explore such options in our future work. 4 Energy Efficiency Potential of Laptop Disk Arrays In this section we evaluate the idea of using laptop disk arrays in server-class storage systems. We conducted our experiments using the Disksim simulator [Ganger et al., 1999; Ganger, 1995]. We augmented the simulator with code to model energy consumption. Part of the code is based on the Dempsey simulator [Zedlewski et al., 3]. The disk configuration parameters used in the simulations are based on the Hitachi TravelStar 7K6 and Hitachi UltraStar 15K73 disks (Table 1). In the simulations of the TravelStar 7K6 disk we assume a simplified disk power model with the characteristics shown in Table We do not model the active idle power mode, since we do not have information about its transition time parameters. Our results are hence conservative, since the use of the active low power mode can provide additional savings even during short idle intervals. In addition, we do not model the Sleep mode. 4 State Active Performance Idle Low Power Idle Standby Power Low Power Idle to Active Energy Low Power Idle to Active Time Active to Low Power Idle Energy Active to Low Power Idle Time Standby to Active Energy Standby to Active Time Active to Standby Energy Active to Standby Time Mobile Disk 2.5 W 2. W.85 W.25 W 1.45 J.53 s 2.15 J.85 s 9. J 3. s 2.94 J 1.25 s Table 2: Abstract disk model parameters used calculate energy consumption. Values are based on the TravelStar 7K6 and UltraStar 15K73 disk. Workload Type Bandwidth (Rqs/sec) Fraction of Max. Srv Bandwidth Exp-8 % Exp-1 % Exp-1 1..63% Par-1 % Par-5 % Max. Srv Bandwidth % Max. Mbl Bandwidth % Table 3: Aggregate bandwidth in requests per second imposed by each workload tested and its fraction of the maximum bandwidth supported by the server-class disk (Srv). The maximum aggregate bandwidth of the laptop-disk array (Mbl) is also shown. In our evaluation, we compare the performance and energy consumption of a server-class disk with that of mirrored disk array system consisting of three mobile hard drives. Since we want to evaluate the proposed idea under workloads with various degrees of intensity we have conducted an open system simulation using synthetic workloads. In contrast to a closed system simulation, in an open system simulation requests are considered to be independent of each other. In all the experiments, one million requests are issued. We use two types of distributions in order to model the inter-arrival time among requests generated by the synthetic workloads: an exponential distribution that leads to a smoother workload and a Pareto distribution that leads to a more bursty workload. We experiment with a wide range of mean inter-arrival times. In the remainder of the paper, we will follow the naming convention Distribution-Mean in order to refer to a specific workload. For example Exp-1 represents a workload with request inter-arrival times that follow an exponential distribution with a mean of 1 ms, while Par-5 represents a workload with request interarrival times that follow a Pareto distribution with a mean of 5 ms. Table 3 presents the aggregate bandwidth in requests per second imposed by each workload to the storage system. Similar workload configuration parameters were used in the evaluation of the DRPM approach [Gurumurthi et al., 3]. Across all experiments we present results for five systems. Srv represents a storage system consisting of a single server-class disk. The remaining four systems are variations of the laptop disk based array. We 5 1 Percentage of Requests Response Time (milli-seconds) Srv Mbl-E1 Mbl-E3 Figure 1: Request Response Time Distribution for a read-only Exp-8 workl
Recommended
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks