An Analysis of Hard Drive Energy Consumption

Anthony Hylick, Ripduman Sohan, Andrew Rice, and Brian JonesUniversity of Cambridge, Computer Laboratory

Firstname.Lastname@cl.cam.ac.uk

October 24, 2008, This version corrects an error in theprevious version of the paper regarding write energy con-sumption (§7.3).

Abstract

The increasing storage capacity and necessary redun-dancy of data centers and other large-scale IT facilitieshas drawn attention to the issue of reducing the power con-sumption of hard drives. This work comprehensively inves-tigates the power consumption of hard drives to determinetypical runtime power profiles. We have instrumented at afine-grained level and present our findings which show that(i) the energy consumed by the electronics of a drive is justas important as the mechanical energy consumption;(ii)the energy required to access data is affected by physicallocation on a drive; and(iii) the size of data transfers hasmeasurable effect on power consumption.

1. Introduction

Reducing power consumption in grid-powered comput-ing contexts has garnered much attention over the past sixyears. Prior to this, power consumption was a concern pri-marily for mobile computing domains. The rising cost ofenergy and increased public awareness surrounding the en-vironment and sustainability has prompted even more at-tention to reducing the power footprint of IT infrastructureas a whole.

Optimizing hard drive power consumption is an impor-tant part of this challenge. The amount of storage neededto fulfil current and growing computing needs seems set tocontinue to increase. Enabling this increased storage ca-pacity is often accomplished through the addition of morehard drives, and these added drives will provide service atthe expense of requiring additional power. The Interna-tional Data Corporation (IDC) reports that storage capac-ity is increasing at a rate of almost 60% per year [1]. Thisincrease in storage capacity is likely to increase the propor-tion of total IT power that is consumed by hard drives.

We focus our attention on obtaining an in-depth under-standing of the power consumption trends of hard drives,instrumenting at a fine-grained level to present results thatprovide insight into the mechanical and electronic powerconsumption of hard drives at runtime. In particular, weshow that(i) the power consumed by the electronic compo-nents of hard drives is just as significant as the consumptionof the mechanical components and should not be consid-ered as a secondary concern when approaching the issue ofdrive power management;(ii) data placement has an effecton the power consumption of data access but does not affectthe power consumption of committing data to disk; and(iii)the size of transfers affects power consumption. Based onobserved results, we make recommendations to aid hard-ware and software designers in optimizing data storage andaccess.

2. Motivation

Previous research has implicitly prioritized the mechan-ical power consumption of hard drives over the electroniccomponents by attempting to reduce power consumptionby spinning disks down. We have attempted to determinepossibilities for power savings while hard drives are in theiractive and idle states (states in which drives are capableof servicing requests without substantial latency penalties).Our interest led us to thoroughly explore the runtime powerconsumption of hard drives; requiring a detailed under-standing of both the mechanical and electronic componentoperation and power consumption of modern drives whilein their different operating modes.The goal is to draw con-clusions regarding drive power consumption in the generalcase for a wide range of commodity disks.

The remainder of the paper is organized as follows:In section 3 we summarize other power measurement ef-forts and highlight existing emphasis on spin-down poli-cies; Section 4 provides a background of drive mechan-ics and operation; Sections 5 & 6 explain the measurementsetup and testing methodology, respectively; Results andanalyses are presented in Section 7; and Section 8 con-cludes.

3. Related Work

Our objective was to discover ways of reducing driveenergy consumption not presented by previous research. Adetailed understanding of the following areas has enabledus in our endeavor:

Detailed Measurement It has been our motivation toaccurately acquire online power consumption measure-ments of hard drive activity for the discovery of opportu-nities to reduce drive power consumption outside of drivespin-down. Our measurement setup provides direct, onlinemeasurements of hard drive power consumption that fewprevious authors have described and employed. For exam-ple, Colarelli and Grunwald limited their measurement ofseeks and state transitions of “large IDE” drives to calibrat-ing their storage array simulator [4]. Similarly, Zedlewskiet. al. took precise measurements of a laptop hard drive’spower consumption for the purpose of creating their disksimulation environment, Dempsey [23]. Weddle et. al.constrained their facility to comparing seek, spin-up/down,and standby measurements to drive datasheets and thosenumbers used in other simulations [21].

With our custom hardware, we aim to uncover hid-den power consumption trends and behaviors that are notexposed through coarser-grained measurement techniques.Our goal is to exploit these hidden trends for reduced en-ergy consumption.

Spin-Down At the forefront of addressing the reductionin power consumption is the need to do so while maintain-ing the “always-on” or “always-available” state of comput-ers and the internet. The same may be applied to reducingthe power consumption of hard drives. However, observethat commodity hard drives do not have any active, lower-power operating modes like modern processors (i.e. volt-age and frequency scaling).

Previous energy-saving research has been based aroundthe standby paradigm; the notion that the only way to savepower is to place drives in standby—an inactive state—foras long as possible. Since accessing a spun-down drivecosts time and energy, devising the most effective and effi-cient spin-down strategy has been a common topic for priorwork.

Wilkes provides an early adaptive prediction approachto determining when to spin the disk down [22]. Hismethod calculates a delay value which is an adaptive waittime before spinning down the drive based upon the user’sprevious activity patterns. Greenawalt models hard drivepower consumption and concludes that spinning downdrives reduces power consumption for carefully selectedtimeout values [8]. Douglis et. al. show that spinning adrive down reduces energy consumption while comparingdifferent spin-down algorithms [6]. Douglis et. al. alsopresent some adaptive spin-down techniques that provide

more desirable and efficient spin-down transitions by re-ducing spin-ups [5]. Their adaptive approach monitors us-age trends and adapts the policy as needed. Papathanasiouand Scott propose a strategy to group drive accesses to-gether in order to maximize standby times and utilizationonce the drive is spun-up [15]. Helmbold et. al. use a ma-chine learning approach to dynamically determine when tospin a drive down, arguing that more intelligent approachesto drive spin-down are needed to be effective in saving en-ergy [11]. All of these approaches focus solely on placingdrives in standby mode as a means to reduce power con-sumption.

More recent work has focused on partitioning largerstorage deployments into groups, allowing subsets to beplaced in standby mode under lighter workload demands.

Colarelli and Grunwald take advantage of the largeamounts of idle drive time that results from the reliabil-ity and redundancy of most high-fidelity storage environ-ments [4]. The pool of available drives is divided into oneset of active drives, which are on and servicing requests,and another set of passive drives, which are spun-downinto standby mode. Pinheiro and Bianchini introduce Pop-ular Data Concentration (PDC) where frequently accesseddata is placed onto a subset of available network storagedrives [16]. The remaining drives may be spun-down toconserve energy. Ganesh et. al. present a method of in-telligently spinning down a large portion of drives in largestorage deployments while managing performance trade-offs [7]. Narayanan et. al. show that enterprise storage en-vironments are amenable to drive spin-down and describean approach to spin-down a large portion of drives during“write-dominated” access periods for reduced energy con-sumption [14]. Pinheiro et. al. take advantage of the re-dundancy of large storage deployments through a methodcalled “Diverted Access” [17]. This technique allows forredundant drives to be spun-down while maintaining highavailability. Weddle et. al. introduce a power-aware RAID(PARAID) method for reducing the power consumption ofcommodity hardware via an adaptive algorithm designedto allow for drives to be powered down [21]. These at-tempts produce the effect of an “always-available” storageenvironment while reducing the number of drives that areactively servicing requests. Still, the underlying premise ofthese approaches is that the reduction of power consump-tion as it relates to hard drives is only achieved by placingthe disks in inactive states.

Other research has attempted to exploit scheduling andintelligent caching to reduce the energy consumption ofstorage elements. Zhu et. al. introduce several power-aware cache management techniques that aim to reduce en-ergy consumption by allowing drives to remain in lowerpower modes longer [25]. Li and Wang describe power-aware scheduling and cache management algorithms for

different RAID configurations [13]. These authors lever-aged power-aware caches and scheduling knowledge to ex-tend drive spin-down periods and most efficiently utilize aspun-up drive.

Son et. al. explore physical layout optimizations acrossdrive arrays for increased performance and increased idleperiods in large systems [20]. The increased idle periodsresulting from the authors’ optimizations allow for drivesto be placed in standby mode for longer periods of time.For applications whose data access patterns are known andunderstood in detail, this work presents a viable option forreduced energy consumption. Again, the energy-savingsproclaimed by this work are only realized after drives areplaced in standby mode.

Throttling Zhu et. al. present Hibernator, an energymanagement system for drive arrays in large storage envi-ronments [24]. This system adaptively changes the speedsat which drives are spinning in order to save energy whilemaintaining performance guarantees. Gurumurthi et. al.explore a method for dynamically controlling the speed atwhich drives spin with DRPM [9]. Focused on large-scalestorage deployments like data centers, this approach re-duces the total energy consumption while preserving highavailability. Modern commodity hard drives do not offerthe multi-speed operating modes needed by this work, lim-iting the usefulness of these approaches.

4. Hard Drive Operation

In analyzing drive power consumption, it is importantto have a detailed understanding of how drives operate andtheir internal make up.

Modern hard drives are composed of (i) magnetic plat-ters, on which data is stored; (ii) a platter spindle motor,responsible for turning the spindle attached to the plat-ters; (iii) read and write heads, which retrieve and commitdata to the platters; (iv) read and write head actuator arms,which suspend the read and write heads above the platters;(v) a voice coil actuator, which positions the actuator armsover the correct location on the platters; (vi) and printedcircuit board electronics, which consists of a buffer cache,a motor driver, and a main controller.

Power is supplied to hard drives via 12-Volt and 5-Voltsupply lines. The 12-Volt supply is responsible for power-ing the motor to spin the hard drive’s platters and supplyingpower to the read/write arm actuator, while the 5-Volt sup-ply is responsible for powering the read and write heads,the buffer cache, the flash memory, and the main controller.Laptop drives do not make use of a 12-Volt supply line.The internal mechanics of laptop drives is the same as 3.5”drives, but they operate on a 5-Volt supply. This wouldmake it more difficult to discern mechanical power con-

sumption from electronic power consumption. Our initialfocus is on server-sized storage infrastructure.

Drive platters spin constantly, only stopping in standbymode. Read/write heads are only powered when readingand writing, and the arm actuator is only powered whenseeking to and residing over locations on a platter. Printedcircuit board electronics are always powered.

According to the ATA/ATAPI-5 specification and theAdvanced Configuration and Power Interface (ACPI), fourpower management states are supported by modern harddrives: active, idle, standby (or spin-down), and sleep (re-quires a system reset to recover from this state) [3].

5. Measurement Environment

We have developed a custom measurement platformthat provides direct, online measurements of individual PCcomponents. We configured our platform to capture two,12-bit inputs—the 5-Volt and 12-Volt supply lines of a con-nected hard drive—at 1100 Hz.

Measurement samples are generated by intercepting the5 and 12-Volt supply lines with .02 Ohm resistors, lead-ing to a small, but detectable, drop in potential across theseresistors. We capture this potential difference using dif-ferential amplifiers and digitize the outputs. A microcon-troller collects the digital outputs and sends them to a sep-arate PC for logging. Our measurement instrumentation ispowered separately and has a negligible effect on the actualpower consumption of the drive under test. Pictures of themeasurement board and measurement setup are provided inFigures 1 and 2, respectively.

6. Test Setup & Methodology

The test set is comprised of ten hard drives of varyingage, capacity, and interface. The test set is chosen specif-ically to model the diversity of drives available in produc-tion environments. A detailed list of drives tested can beseen in Table 1.

Specialized microbenchmarks were created to test eachof the drive components individually:

Arm Actuator The arm actuator, powered by the 12-Volt supply, is tested by specifying reads and writes in dif-ferent locations on the drive and across the entire drive. Thepower consumption of seeking to specified locations andthe power consumed keeping the heads in the correct loca-tion up to request completion is captured. To expose differ-ences in power consumption depending on the originationand destination of seeks, measurements are also taken ofseeking to different parts of the drive from different start-ing locations (i.e. inner to outer, center to outer, inner tocenter, etc.).

Figure 1. Custom Measurement Board Figure 2. Measurement Setup

Table 1. Details of Drives Tested

Make & Model Capacity(GB)

1 IBM Deskstar IC35L060AVER07-0 61.52 Fujitsu MPE3084AE 8.453 Seagate ST380215A 804 Samsung HD501LJ 5005 Seagate ST3250820A 2506 Hitachi Deskstar HDS725050KLA360 5007 Hitachi Deskstar HDP725025GLA380 2508 Western Digital WD5000AACS-00ZUB0 5009 Seagate ST380011A 8010 Seagate ST373455LW 73

Read/Write Heads The read/write heads, powered bythe 5-Volt supply, are isolated by capturing the 5-Volt con-sumption during data transfers. Varying data sizes weretested, and drive accesses were conducted in different phys-ical locations on the drive and across the entire drive.Transfers were large enough as to eliminate the possibleinterference of on-drive caches.

Printed Circuit Board Electronics The printed circuitboard electronics, powered also by the 5-Volt supply, arealways active. Because the printed circuit board electron-ics are always on, capturing measurements when no re-quest was being serviced allowed us to subtract printed cir-cuit board consumption from read/write head consumptionmeasurements.

Platter Spindle Motor The platter spindle motor, pow-ered by the 12-Volt supply, is only deactivated when thedrive is in standby mode. Measuring 12-Volt consumptionin idle mode and in standby mode permitted us to differen-tiate arm actuator power from platter spindle motor power.

7. Results

In this section, we present the results of testing the harddrives listed in Table 1 along with brief analyses. The com-petitive nature of the hard drive industry meant we couldfind no literature on the low-level construction or opera-tion of the drives, and hence we were unable to corroborateour speculations as to the causes of the observed trends.Nonetheless, we are confident that the trends themselvesare reliable and feel it is useful to present our current rea-soning in its speculative form.

These results are most usefully viewed in terms ofthe drive’s operating modes: standby, idle, and read-ing/writing data. The microbenchmarks of the previoussection are combined to explain the power consumption ofthese modes. Lastly, we investigate the mechanical opera-tion of the spindle motor and the actuator arm.

7.1. Standby Mode

Standby mode measurements were conducted by issu-ing the ACPI standby command and recording power con-sumption over a 5-minute interval. Multiple runs were col-lected to expose any variation in the measurements. Fig-ure 3 presents the energy consumption averages for all thedrives tested. Because the mechanical components are offin standby mode, the drive electronics are responsible forthe majority of the total energy consumption for this du-ration of time. The numbers show that even in standbymode (an inactive state where no work is being done), thedrive electronics are still active. For the drives tested, themechanical energy savings of entering standby versus re-maining idle range from 92-99%. The electronic compo-nent energy savings span a wider range between 35% and95%. Throughout our testing, we have confirmed that plac-ing drives in standby mode significantly reduces the me-chanical power consumption of hard drives, but we believe

0

100

200

300

400

500

600

700

1 2 3 4 5 6 7 8 9 10

Sta

ndby

Ene

rgy

(Jou

les)

Drive

Standby - ElectronicsStandby - Mechanics

Figure 3. Summary of 5-Minute Standby EnergyConsumptions

200

400

600

800

1000

1200

1400

1600

1800

2000

1 2 3 4 5 6 7 8 9 10

Idle

Ene

rgy

(Jou

les)

Drive

Idle - ElectronicsIdle - Mechanics

Figure 4. Summary of 5-Minute Idle EnergyConsumptions

that there has not been enough focus on reducing the elec-tronic power consumption for standby and other operatingmodes.

The electronics should strive to minimize power con-sumption while the drive is in standby mode. However,from the numbers that we collected, it seems that savingscould be greater. The power consumptions for Figure 3 areof the order of hundreds of mW, while modern integratedcircuit components in standby/sleep modes consume lessthan 10 mW.

7.2. Idle Mode

Idle mode power consumption is measured by capturingdata when the drive is not servicing I/O requests. Figure 4illustrates that the hard drive electronics are still a signifi-cant power consumer in idle mode as well. In idle mode,the drive is not servicing any requests, but the platters arespinning. The energy consumption values in Figure 4 showthat mechanical components responsible for spinning theplatters do not dominate the electronics as expected. More-over, the electronic energy consumption ranged from 25%to nearly 75% of the total.

7.3. Read/Write Power Consumption

We conducted several tests to thoroughly test and un-derstand the energy cost of data access and storage. Ourfirst test measured the energy consumed while reading andwriting the entire drive divided into 256 MB bins. Thistest was devised to get an average energy consumption forreading and writing a given LBN, determining the relation-ship between power consumption and physical drive loca-tion. Next, we measured the power consumption of read-ing and writing 1 GB of data with block sizes ranging from

512 B to 8 KB. This test was structured to expose differ-ences in power consumption resulting from different blocksize specifications.

Location dependence Figure 5 demonstrates that, gen-erally, read and write energy consumptions were dependenton LBN, increasing as LBN increased. Data density (dataper track length) increases at higher LBNs simply becausethe track length becomes smaller. However, as a result ofthe zoned bit recording employed by modern hard drives,the outer tracks contain more data as they have a larger cir-cumference than the inner tracks. Because hard disk drivesrotate at a constant angular velocity, more data can be readand written—and at faster rates—at the outer tracks (lowerLBNs) than the inner tracks (higher LBNs). The smallerenergy consumption values at the lower LBNs is a resultof more sectors being seen by the Read/Write heads for agiven angle of rotation as compared to the higher LBNs.In addition to the faster linear velocity of larger radii, theinner tracks cannot store as much data as the outer tracks,potentially leading to added seeking in order to service asingle request.

To validate our thinking, We measured read and writebandwidths across all the drives. We observed that read andwrite bandwidths decreased as LBN increased (Figure 6).Bandwidth decreases as a result of transfers taking longerto complete at higher LBNs. This observation is due tothere being less data per track and the fact that the linearvelocity of the Read/Write heads decreases as the radiusdecreases.

All drives tested showed this same behavior of increas-ing read and write energy consumptions. Low LBNs cor-respond to tracks with larger radii near the outer parts ofthe platters, and high LBNs correspond to tracks of smallerradii near the inner edge of the platters. We fit a cubic poly-nomial to the data points of the read and write curves from

20

40

60

80

100

120

140

160

180

200

10480943283847336628852404192314420961048

Ene

rgy

(Jou

les)

LBN (x 10^6)

Energy vs. LBN

ReadWrite

Figure 5. Hard Drive 6, Read and Write EnergyConsumptions from low LBN to high LBN

0

10

20

30

40

50

60

70

10480943283847336628852404192314420961048

Ban

dwid

th (

MB

/s)

LBN (x 10^6)

Bandwidth vs. LBN

ReadWrite

Figure 6. Hard Drive 6, Read and Write Band-widths from low LBN to high LBN

20

40

60

80

100

120

140

160

180

200

10480943283847336628852404192314420961048

Ene

rgy

(Jou

les)

LBN (x 10^6)

Energy vs. LBN

ReadWrite

f(x)

Figure 7. Hard Drive 6, Read & Write Energyvs. Logical Block Number

Figure 5 in Figure 7 with a root mean square (RMS) errorof 6.44 Joules.

We saw a consistent trend in read energy consumptionacross all drives tested. The curve fit to the data points asLBN increases suggests that there is a cubic relationshipbetween read and write energy consumption and LBN, i.e.

Energy∝ LBN3 (1)

Figure 7 and Equation 1 help give some indication asto how a drive’s energy consumption is related to the LBNthat is accessed. Including this type of consumption modeland coefficient values in the technical specifications of harddrives would be a way for drive manufacturers to aid de-velopers in programming with power in mind. This typeof knowledge could be used in reducing drive energy con-

sumption further by laying out data in a way that keepsfrequently accessed data at low LBNs where the energy re-quired to retrieve the data is lowest. For Drive 6, readingthe lowest LBN versus the highest LBN results in a powersavings of approximately 14 Watts.

We took great care in ensuring our tests suppressed bothwrite and read caching. We are confident that neither thesystem cache nor the on-disk caches were a factor in theresults we present. System developers would only need tounderstand the relationships we present with respect to log-ical block number, however, we could roughly transformthis relationship with LBN into a relationship with respectto radius by analyzing our bandwidth and energy/powergraphs and estimating the radius of the tracks and zones.9 of 10 drives were described by this cubic relationship,and RMS errors were less than 8 Joules.

Seeing this read behavior, we then investigated the 5 and12-Volt components that make up this curve. We decom-posed Drive 6’s read energy curve (Figure 5) into the 5-Volt and 12-Volt components and plotted them in Figure8. Both supply lines show a similar shape to the combinedcurve. It is interesting to note that the 5-Volt energy con-sumption is just as significant in the total energy consump-tion as the 12-Volt supply. Here again, we see that both the5-Volt and 12-Volt power consumptions are equally impor-tant, and neither can be ignored when addressing hard drivepower consumption.

Finally, Figure 9 considers the relationship between readenergy (Figure 5) and bandwidth (Figure 6). Our aim wasto uncover any relationship between the performance andenergy consumption of Drive 6.

The quadratic fit in Figure 9, RMS = 3.81 Joules, re-inforces our previous conclusion of favoring data accessesat lower LBNs (outer edges of the platter) to lower energy

0

20

40

60

80

100

120

140

10480943283847336628852404192314420961048

Ene

rgy

(Jou

les)

LBN (x 10^6)

Energy vs. LBN

5-Volt Line12-Volt Line

Figure 8. Hard Drive 6, Read Energy Con-sumption 5 V and 12 V

0

10

20

30

40

50

60

70

80

90

100

110

25 30 35 40 45 50 55 60 65

Ene

rgy

(Jou

les)

Read Bandwidth (MB/s)

Drive 6, Read Energy Consumption vs. Read Bandwidth

Read Energy vs. Read Bandwidthf(x)

Figure 9. Hard Drive 6, Read Energy Con-sumption vs. Read Bandwidth

consumption. In addition, this relationship shows that per-formance is not sacrificed in accessing data at lower LBNsbut enhanced. Contrary to many other power-performanceoptimizations, energy consumption and performance ofhard drives seem to have a symbiotic relationship. It isalso important to note that this quadratic relationship be-tween energy and read bandwidth suggests that the powerconsumption per bit is not constant, unlike commonly be-lieved (e.g., equation 1 in [12]). A 1/x fit to this data yieldedan RMS of 12.08 Joules. The longer a transfer takes, thebandwidth is lower and the energy consumption is greater.What sounds like a simple inverse relationship is actually aquadratic relationship.

These findings are important because they provide newinformation that hardware manufacturers and kernel devel-opers could leverage in attempts to reduce drive energy

0

200

400

600

800

1000

1200

1400

1600

8 KB4 KB2 KB1 KB512 B

Ene

rgy

(Jou

les)

Block Size

Energy vs. Block Size (1 GB Transfer)

ReadWrite

Figure 10. Hard Drive 7, Energy Consumptionas Block Size Increases

consumption without sacrificing performance: (i) read-ing and writing low LBNs for better performance and de-creased energy consumption, (ii) writing frequently readdata at low LBNs and less frequently read data at higherLBNs, and (iii) using our model to understand the rate atwhich read and write energy increases with respect to LBN.

Effect of Transfer Size The data to this point consid-ered block sizes of 256 MB. In general, reading consumesmore energy than writing for large (≥ 2 KB) block sizes,but the opposite is true for smaller (<2 KB) block sizes.Figure 10 shows this happening on Drive 7 as 1 GB is readand written with the block size increasing from 512 B to8 KB. Transfers are made up of both a fixed and variablecost. The fixed cost is due to the pre-processing and setupof the transfer, while the variable cost is a result of the sizeof the actual data transfer. Small-block-size transfers aredominated by the fixed cost, and large-block-size transfersare dominated by the variable cost. After observing similarbehavior on 7 of 10 drives, it seems that the pre-processingand setup for writes takes longer than that of reads, gen-erally, and this becomes apparent as block size decreases.The longer fixed cost for writes at lower block sizes re-sults in the increased energy consumption, however, oncethe variable cost becomes the dominant factor, writes con-sume less energy than reads.

With large-block-size reads consuming more energythan writes, the fact that reads occur at least 4-5 times morethan writes [19] becomes more important when attemptingto reduce the power consumption of data storage. In ad-dition, digital data is growing rapidly as more digital datais created in a year than there is capacity to store it [2].The implication of these observations is that more consid-eration should be given to current data access patterns toobviate unnecessary increases in energy consumption forthe future.

-5

0

5

10

15

20

25

180170160150140130120110100908070605040302010

Pow

er (

Wat

ts)

Time (seconds)

Drive 3, Standby-to-Active Spin-up Cost

Figure 11. Spin-up Cost of Drive 3

7.4. Hard Drive Mechanics

Platter Spindle Motor By investigating how the 12-Volt consumption was affected during our tests, we gainedinsight into the mechanical operation of hard drives. Thepower consumption of the hard drives’ mechanical parts inkeeping the platters spinning was fairly constant and con-sistent across drives. The increased consumption duringspin-up transitions varied across the drives. Costs variedbetween 7.2 Watts and 24 Watts. This transition consumesthe greatest power in hard drives, and prior authors haveexpressed the need to avoid this cost, for the sake of power,performance, and device lifetime, while maximizing powerefficiency and performance. The reduction in energy con-sumption at the expense of increased energy and latencyfrom the subsequent spin-up has been highlighted in previ-ous work [5, 10, 11]. Research has also addressed the po-tential reduction in device lifetime resulting from repeatedspin-down/spin-up transitions [18,21,24]. The mechanicalpower cost of spinning a drive up is large, but short-lived.

Figure 11 shows the spin-up cost of Drive 3. This spin-up transition takes about 2.5 seconds and consumes approx-imately 37 Joules (5-Volt and 12-Volt combined). Spinningup this drive from idle to active mode takes approximately 1second and consumes less than 1 Joule (5-Volt and 12-Voltcombined). When data is not present in memory and mustbe retrieved from disk, latency is naturally orders of mag-nitude larger than typical memory access times. The dif-ference between a spin-up from standby (2.5 seconds) andfrom idle (1 second) is only 1.5 seconds. And, by being instandby mode versus idling, this particular drive saves the37 Joules needed to recover the drive from standby afteronly about 9 seconds in standby.

Our investigation has confirmed previous conclusionsthat placing a drive in standby is a means of reducing driveenergy consumption, barring the drive is spun down for along enough period of time.

4

6

8

10

12

14

16

18

20

22

0 100000 200000 300000 400000 500000

Pow

er (

Wat

ts)

Measurement Sample (Increasing LBN)

Drive 6, Incremental Seek Profile

Figure 12. Seek Profile of Drive 6

Actuator Arm We structured two additional tests toprofile the power consumption cost of seeking to differ-ent locations on the platters (zones) from different start-ing points. The first test grouped logical blocks into 5 GBbins. Our measurements captured the power consumptionof seeks originating at bin 0 to all other bins until the end ofthe drive was reached. Figure 12 shows the results for thistest. This was a 500 GB drive, and the seek power costs ofseeking from bin 0 to all other bins were fairly constant be-tween 18 and 21 Watts. 75% of the seek power costs werein this band, while the remaining quarter were less than18 Watts. Drawing attention to the upward trend at higherLBNs, we saw that the 12-Volt line consumes more powerto keep the actuator arm at higher LBNs. Positioning theactuator arm at higher LBNs (physically located towardsthe inner part of the platters) requires more current to passthrough the voice coil. Figure 13 shows more specificallythat the power cost of seeking from low to high LBNs isabout 20 Watts at max and, just like the spin-up cost, isvery short lived. The energy consumption for this outer toinner seek spike (LBN 0 to LBNMAX) is only about 0.60Joules. Regardless of destination, the power cost of seekingfrom the beginning of the disk is fairly constant.

The second test was setup to see how much power wasconsumed from seeks originating from the middle tracks ofa platter. We measured the power needed to seek from thecenter logical blocks to higher LBNs and to lower LBNsin 1 GB block groupings. From a performance standpoint,data placement offers improvement, but from a power per-spective, placement does not offer any significant reductionin power consumption as far as seeks are concerned (theseek results were similarly constant like those in Figure12). Placement could possibly be explored for sustained orfrequent data access over extended periods of time for en-ergy savings. For example, since less energy is consumedfor reading lower LBNs versus higher ones, frequently readdata that is placed at a low LBN will save energy over ac-cessing the same data placed at a higher LBN. More power

4

6

8

10

12

14

16

18

20

1620144012601080900720540360180

Pow

er (

Wat

ts)

Time (milliseconds)

Drive 6, Seek from Outer to Inner Track

Figure 13. Drive 6, Power Cost of Outer Trackto Inner Track Seek

was saved seeking from the center to lower LBNs than fromthe center to higher LBNs, but the savings were minimal.

8. Conclusion

We set out to uncover hidden power consumption be-haviors of modern hard drives in efforts of discoveringnew possibilities to reduce drive energy consumption. Thestandby and idle mode energy consumption of the driveelectronics could use a great deal of attention. This is anissue that will probably need to be addressed at the microar-chitecture level by hard drive manufacturers. However, thedisparity of read and write energy consumption at differentLBNs and the difference in energy consumptions of same-size transfers specified with different block sizes are bothissues that can be exploited by software developers. Inves-tigating the energy consumption of reading and writing hasled us to the possibility of reducing energy consumption bywriting data that will be accessed frequently at lower LBNsand writing and reading data with the largest possible blocksizes. The energy consumed during seeking is minimal,but restricting disk accesses between low and central LBNscould aggregate into savings over extended usage periods.

Storage is only going to increase into the future, whichmotivates us to reduce the consumption of hard drives inactive as well as passive modes. We have presented severalopportunities to make improvements in energy consump-tion, and the cumulative savings will substantially impactcomputing long-term.

Solid-state drives will have different power consump-tion profiles, but the current price of these drives makesthem an uneconomical option for large storage deploy-ments. Future work will include in-depth analysis ofthe power consumption of solid-state drives and laptophard drives as well as reporting on the implementation of

suggestions made here and quantifying the impact on asystems-level.

Acknowledgements

This work would not have been possible without thehelp and support of Professor Andy Hopper. Robert Harleprovided instrumental guidance and assistance throughoutthis research. Sudhanva Gurumurthi of the University ofVirginia provided helpful insight into our testing. Wes Fel-ter of IBM Research helped to debug the error of our initialwrite-testing script. We also would like to thank our shep-herd, Nikolai Joukov.

References

[1] IDC serves up 10 storage predictions for 2008, 2007.http://www.cio.co.uk/news/index.cfm?articleid=2455.

[2] Study: digital universe and its impact bigger than wethought, 2008. http://www.computerworld.com/action/article.do?command=viewArticleBasic&articleId=9067639.

[3] Advanced Configuration and Power Interface (ACPI) spec-ification, 1999.http://www.acpi.info.

[4] D. Colarelli and D. Grunwald. Massive arrays of idle disksfor storage archives. pages 1–11. In Proceedings of the 2002ACM/IEEE International Conference on Supercomputing,2002.

[5] F. Douglis, P. Krishnan, and B. Bershad. Adaptive diskspin-down policies for mobile computers. pages 121–137. In Proceedings of the 2nd Symposium on Mobile andLocation-Independent Computing, 1995.

[6] F. Douglis, P. Krishnan, and B. Marsh. Thwarting thepower-hungry disk. pages 292–306. USENIX Winter Con-ference, 1994.

[7] L. Ganesh, H. Weatherspoon, M. Balakrishnan, and K. Bir-man. Optimizing power consumption in large scale storagesystems. In Proceedings of the 11th Workshop on Hot Top-ics in Operating Systems (HotOS XI), 2007.

[8] P. Greenawalt. Modeling power management for harddisks. pages 62–66. In Proceedings of the 1994 IEEE Inter-national Symposium on Modeling, Analysis, and Simula-tion of Computer and Telecommunication Systems (MAS-COTS), 1994.

[9] S. Gurumurthi, A. Sivasubramaniam, M. Kandemir, andH. Franke. DRPM: dynamic speed control for power man-agement in server class disks.ACM SIGARCH ComputerArchitecture News, 31(2):169–181, May 2003.

[10] T. Heath, E. Pinheiro, and R. Bianchini. Application-supported device management for energy and performance.In Proceedings of the Workshop on Power-Aware Comput-ing Systems, 2002.

[11] D. P. Helmbold, D. D. E. Long, and B. Sherrod. A dynamicdisk spin-down technique for mobile computing. pages

130–142. In Proceedings of the Second Annual ACM Inter-national Conference on Mobile Computing and Network-ing, 1996.

[12] N. Joukov and J. Sipek. GreenFS: making enterprise com-puters greener by protecting them better. pages 69–80. InProceedings of EuroSys 2008, 2008.

[13] D. Li and J. Wang. EERAID: energy efficient redundant andinexpensive disk array. page 29. In Proceedings of the 11thworkshop on ACM SIGOPS European Workshop, 2004.

[14] D. Narayanan, A. Donnelly, and A. Rowstron. Write off-loading: practical power management for enterprise stor-age. In Proceedings of the 6th USENIX Conference on Fileand Storage Technologies (FAST), 2008.

[15] A. E. Papathanasiou and M. L. Scott. Increasing disk bursti-ness for energy efficiency. Technical Report TR792, Uni-versity of Rochester. pages 1–31, 2002.

[16] E. Pinheiro and R. Bianchini. Energy conservation tech-niques for disk array-based servers. pages 68–78. In Pro-ceedings of the 18th International Conference on Super-computing, 2004.

[17] E. Pinheiro, R. Bianchini, and C. Dubnicki. Exploitingredundancy to conserve energy in storage systems.ACMSIGMETRICS Performance Evaluation Review, 31(1):15–26, Jun 2006.

[18] E. Pinheiro, W.-D. Weber, and L. A. Barrosa. Failure trendsin a large disk drive population. In Proceedings of the5th USENIX Conference on File and Storage Technologies,2007.

[19] D. Roselli, J. R. Lorch, and T. E. Anderson. A comparisonof file system workloads. pages 41–54. In Proceedings ofthe 2000 USENIX Annual Technical Conference, 2000.

[20] S. W. Son, G. Chen, and M. Kandemir. Disk layout opti-mization for reducing energy consumption. pages 274–283.In Proceedings of the 19th Annual International Conferenceon Supercomputing, 2005.

[21] C. Weddle, M. Oldham, J. Qian, A.-I. A. Wang, P. Reiher,and G. Kuenning. PARAID: a gear-shifting power-awareraid. ACM Transactions on Storage, 3(3):33, Oct 2007.

[22] J. Wilkes. Predictive power conservation. Technical Re-port HPL-CSP-92-5, Hewlett-Packard Laboratories. page1, February 1992.

[23] J. Zedlewski, S. Sobti, N. Garg, F. Zheng, A. Krishna-murthy, and R. Wang. Modeling hard-disk power consump-tion. pages 217–230. In Proceedings of the 2nd USENIXConference on File and Storage Technologies, 2003.

[24] Q. Zhu, Z. Chen, L. Tan, Y. Zhou, K. Keeton, and J. Wilkes.Hibernator: helping disk arrays sleep through the winter.ACM SIGOPS Operating Systems Review.

[25] Q. Zhu, F. M. David, C. F. Devaraj, Z. Li, Y. Zhou, andP. Cao. Reducing energy consumption of disk storage us-ing power-aware cache management. pages 118–129. InProceedings of the 10th International Symposium on HighPerformance Computer Architecture (HPCA), 2004.