C OV E R F E AT U R E
TO LEASE OR NOT TO LEASE FROM STORAGE CLOUDS Edward Walker, University of Texas at Austin Walter Brisken and Jonathan Romney, National Radio Astronomy Observatory
Storage clouds are online services for leasing disk storage. A new modeling tool, formulated from empirical data spanning many years, lets organizations rationally evaluate the benefit of using storage clouds versus purchasing hard disk drives.
H
ard disk drives provide storage for a broad range of devices, from mobile phones to large IT server farms. In 2008, approximately 590 million hard disk drives were shipped worldwide,1 largely driven by the vast amount of information our digital society is generating. For example, a report from researchers at the University of California, Berkeley, estimates that 92 percent of the five exabytes (1018) of new information created in 2002 was stored on magnetic media, primarily hard disk drives.2 At the same time, the business of selling infrastructure as a service through the Internet is growing. This technology trend, also known as cloud computing, lets individuals and organizations outsource their IT requirements to remote data centers, paying for only what they use. The cloud computing industry’s worth was estimated at more than $16 billion in 2008, and it’s expected to grow to $42 billion by 2012.3 Several online services currently lease storage infrastructure. These storage clouds let anyone with a credit card purchase storage capacity online, paying a monthly fee for the storage they use. Amazon.com’s S3 service (http://aws.amazon.com/s3), for example, lets users store
44
COMPUTER
arbitrary objects of up to 5 Gbytes each in their online storage repositories. The S3 service uses a tiered pricing structure, with storage getting cheaper as more is used. As of late 2008, Amazon.com reported that users stored more than 40 billion objects in the S3 service.4 With the significant growth of society’s storage requirements, and the availability of pay-per-use online storage services, when should a consumer consider using storage clouds? We focus on the problem of resolving this buy-orlease storage decision.
BUY OR LEASE An organization’s decision to buy or lease long-term assets is often a critical dilemma. The literature describes numerous models to assist firms in resolving this acquisition decision.5-7 However, there is a significant lack of work in applying these models to compare the cost of purchasing versus leasing storage from cloud services. Important studies have examined the cost of performing scientific experiments using the Amazon.com cloud services,8,9 but these prior works are difficult to generalize outside the scope of the applications used in the studies. We believe our proposed model for evaluating the benefits of purchasing versus leasing storage from a cloud service is the first reported in the literature, offering a method for deriving a reasonable estimate of the best possible outcomes from the alternative investment choices. Furthermore, our model, which we formulated using empirical disk price data we’ve been collecting weekly from Pricewatch.com over many years, is agnostic to the application using the storage cloud. Our empirical data tracks the lowest prices for serial advanced technology attachment (SATA) disk drives for sale online since the
Published by the IEEE Computer Society
0018-9162/10/$26.00 © 2010 IEEE
Table 1. Parameter terms in the buy-or-lease decision model. technology’s introduction in 2003. This data is significant because SATA is the predominant drive technology used in desktop and nearline systems. Most organizations use financial models as only one data point in deciding whether to buy or lease an asset. Other factors also influence their decision.7 However, our proposed model bridges a critical gap in the literature and, more importantly, the market. Our model helps users understand the premium they’re paying for what seems to be an arbitrarily priced online storage market. Our research agenda aims to provide a quantitative framework for rationally reasoning about the cost of leasing infrastructure from cloud services, and extends our prior work on leasing from compute clouds.10 Importantly, the systematic understanding of the real cost of leasing assets from cloud services lets users make rational decisions with innovative pricing structures. This promotes market transparency and ultimately supports a competitive product market.
PRELIMINARIES A key principle in economic finance is the time value of money. Basically, this principle states that an investor always prefers to receive some fixed amount of money today rather than in the future. Hence, when making buy-or-lease decisions, investors often compare future cash flows in an investment over time, discounted to their present value by some interest rate.5-7 The discounting interest rate used reflects the risk involved in raising the capital needed to invest. In equation form, the simplified standard capitalbudgeting format for calculating a purchased asset’s net present value (NPV) is as follows: PT − CTP S + −E , T (1+ I K )N (1+ I ) T =0 K N
NPVP = ∑
Term
Description
δ
Cost of electric utility ($/kilowatt hour)
Ω
Size of purchased disk drives (Gbytes)
ρ
Proportional difference between human effort in maintaining a purchased versus a leased storage infrastructure
ϒ
Used disk depreciation factor on salvage ([0.0, 1.0])
C
Disk controller unit cost ($)
HT
Annual human operator salary ($)
IF
Risk-free interest rate (percent)
Κ
Current per-Gbyte storage price ($/Gbyte)
LT
Expected annual per-Gbyte lease payment ($/Gbyte/year)
Pc
Disk controller power consumption (kW)
PD
Disk drive power consumption (kW)
VT
Expected storage requirement in year T (Gbytes)
the involved payment structure’s predictability.5 With this NPV formulation for asset purchase and lease, investors can make the buy-or-lease decision using the following criteria: If the incremental NPV (ΔNPV) ≥ 0 ⇒ buy; if ΔNPV < 0 ⇒ lease, where ΔNPV = NPVP – NPVL.
DECISION MODEL We derive a buy-or-lease decision model that calculates the comparative NPVs from storage purchase versus lease. Our model accounts for the expected capital expenditure from SATA hard disk drive purchase, replacement, and end-of-use salvage. It also accounts for the expected operational expenditure for utility consumption and human operator cost. The “Derivation of the Decision Model” sidebar describes how we derived our decision model. We summarize it as follows: If ΔNPV ≥ 0 ⇒ buy; if ΔNPV < 0 ⇒ lease, where N
where PT is the annual profit resulting from the purchased asset in year T; CTP is the asset’s expected annual operating cost at year T; IK is the firm’s cost of capital, defined as the interest rate of its outstanding debt used to finance the purchase;11 N is the asset’s productive life in years; S is the asset’s salvage value after N years; and E is the asset’s purchase (capital) cost. Similarly, the equation for calculating a leased asset’s NPV is as follows: N
NPVL = ∑
T =0
PT − CTL
(1+ I )
T
K
N
−∑
T =0
LT
(1+ I )
T
,
R
where CTL is the leased asset’s expected annual operating cost at year T; LT is the lease payment at year T; and IR is the interest rate for financing the lease payments. In this formulation, the lease’s financing rate is generally regarded as smaller than the cost of capital, IK, because of
ΔNPV = ∑
T =0
CT − ET + LT
(1+ I )
T
+
F
S
(1+ I )
N
−C
F
S = γ ∗ Ω ∗ ⎡⎢VT ⎤⎥ Ω ∗ K ∗e −0.438T
)
(
CT = −ρ ∗ HT − ( 365 ∗24 ∗δ ∗ PC + PD ∗ ⎡⎢VT ⎤⎥ Ω
(
)
ET = 1.03 ∗ ⎡⎢VT ⎤⎥ Ω − ⎡⎢VT −1 ⎤⎥ Ω ∗ Ω ∗ K ∗e
−0.438T
(1)
) .
Table 1 lists the model parameters. We assume ⎡V⎤Ω is an operator returning the minimum number of Ω-sized disk drives needed to store V Gbytes of data. The derived terms S, CT, and ET represent the expected end-of-life disk salvage value, the operating cost in year T, and the capital cost in year T, respectively.
EXAMPLE APPLICATION We used our decision model to evaluate the advantages of purchasing versus leasing storage from a hypothetical storage cloud vendor.
APRIL 2010
45
C OV E R F E AT U R E
DERIVATION OF THE DECISION MODEL
W
e calculate the incremental net present value (ΔNPV) as follows:
ΔNPV = NPVP − NPVL PT − C TP
N
⇒∑
T =0 N
⇒∑
T =0
(1+ I ) (1+ I )
T
K
−E −∑
T =0
K
C TL − C TP
N
S
+
(1+ I ) (1+ I ) T
K
PT − C TL
N
S
+
T
N
+∑
T =0
K
(1+ I )
T
K
LT
(1+ I )
T
N
+∑
T =0
LT
(1+ I )
T
disk controller and the disk units, plus the cost of a human operator to manage the system/data. These cost components contribute to C TP as follows:
(
C TP = (365⋅24)∗δ ∗ PC + PD ∗ ⎡⎢VT ⎤⎥
−E.
R
Capital cost
where δ is the utility cost ($/kilowatt per hour); PC is the controller’s power requirement in kWs; PD is the power requirement in kWs per disk drive; and α is the proportion of the human operator cost, HT, required to maintain the system/data at year T. The operating cost for leased storage, C TL , only includes the cost of a human operator to manage the data. Thus, we calculate C TL as C TL = β ∗H T , where β is the proportion of the human operator cost required to maintain the data on the leased storage at year T. Substituting ρ for (α – β), we can modify ΔNPV to reflect the operating costs as follows: N
Three components make up the cost of purchasing storage E:
ΔNPV = ∑
) ( (1+ I )
− ρ ∗H T − ( 365⋅24 ∗δ ∗ PC + PD ∗ ⎡⎢VT ⎤⎥
t The consumer needs a disk controller to house the purchased disks. t The consumer purchases disk drives in blocks based on increasing storage needs. t The consumer must periodically replace disks due to failure. These components contribute to the future cash flow E as follows: T
− ⎡⎢VT −1 ⎤⎥
Ω
Ω
) ∗Ω + R ) ∗G T
T
(1+ I )
T
+C
K
⇒E=
ET
(1+ I )
T
+C
K
((
)
)
E T = ⎡⎢VT ⎤⎥ − ⎡⎢VT −1 ⎤⎥ ∗Ω + RT ∗GT , Ω Ω where Ω represents the hard disk drive size in Gbytes; VT is the storage requirement in Gbytes at year T; ⎡VT⎤Ω is an operator that returns the minimum number of Ω-sized disk drives needed to store VT ; GT is the predicted cost per Gbyte of disk storage at year T; RT is the disk replacement in Gbytes at year T; and C is the disk controller cost. The important insight in this formulation is that the capital cost isn’t all incurred at the start of the project, unlike in traditional NPV models. Rather, it’s a time-varying formula in which users can grow their storage systems as their requirements evolve. We modify ΔNPV to reflect this growth in capital cost: N
C − C − ET
T =0
(1+ I )
ΔNPV = ∑ +
L T
P T
T
−C
K
N
S
(1+ IK
)
N
+∑
T =0
LT
(1+ IR
)
T
.
Operating cost We estimate the operating cost of purchased storage, C TP , by calculating the electric utility cost associated with running the
46
COMPUTER
− C +
Ω
)− E
T
T
T =0
(( ⎡⎢V ⎤⎥
T
R
Assuming that profit from storage is equal in both the buy and lease cases—that is, using 1 Tbyte of storage from a purchased disk or from a storage cloud results in the same level of productivity—the above simplification results in the removal of the profit term PT .
E=
) +α ∗H ,
Ω
K
N
S
(1+ I )
N
+∑
T =0
K
LT
(1+ I )
T
.
R
Approximating the best outcomes Consumers often find it difficult to estimate their capital cost and the lease financing rates available to them. We therefore suggest an approximate method of evaluating the buy-or-lease decision by deriving the best possible outcome in both the purchase and lease cases. From the well-known economic Law of One Price,1 we can derive the upper bounds for NPVP and NPVL by substituting IK and IR with the risk-free interest rate IF. The “Corollary for Estimating the Upper-bound Present Value” sidebar explains this in more detail. In turn, we can estimate the risk-free interest rate, IF, from the published return on an instrument such as government treasury bills. We can therefore derive an approximation of the decision criteria using these best NPVs, letting us further simplify the currently derived version of ΔNPV to: ΔNPV = N
∑
)
(
− ρ ∗H T − ( 365∗24 ∗δ ∗ PC + PD ∗ ⎡⎢VT ⎤⎥
(1+ I )
T =0
− C +
Ω
)− E
T
+ LT
T
F
S
(1+ I )
N
.
F
Disk price trends We haven’t yet provided a function for the term GT, which we need to calculate ET and to estimate the cost of disk storage at year T. Since April 2003, we’ve been collecting SATA disk price data from Pricewatch.com on a weekly basis. We collected price data for all available disk drive sizes each week—250 Gbytes, 500 Gbytes, and so on—regardless of manufacturer or model. Figure A plots the 10 lowest prices each week for SATA disk drives across all the available drive sizes. Approximating the observed exponential trend line using regression analysis, we obtain the formula G X′ = 1.2984e − 0.0012 X .
2.5 Disk price Exponential (disk price)
2.0
$/GByte
The function G X′ predicts the cost per Gbyte of SATA disk storage X days from 20 April 2003, with G X′ = 1.2984 when X = 0. We can therefore approximate the function GT by assuming that the future disk price trend conforms to the equation K ∗ eC ∗ T, where K represents the lowest storage price per Gbyte available to the consumer at T = 0; and T represents the number of years in the future. We therefore derive GT as GT = K ∗ e–0.0012 ∗ 365 ∗ T ⇒ GT = K ∗ e–0.438 ∗ T.
Disk replacement rates
1.5
y = 1.2984e – 0.0012 X
1.0 0.5
Any realistic cost model for disk storage ownership must estimate the disk replacement cost. A recent large-scale study of disk failures mea0 200 400 600 800 1,000 1,200 1,400 1,600 1,800 2,000 sured the annualized replacement rate (ARR) of Days Start date: 20 April 2003 disk drives in real data centers.2 The study observed ARRs in the range of 0.5 to 13.5 percent, with the most commonly observed ARRs in Figure A. Weekly SATA disk price data collected from Pricewatch.com from the 3 percent range. 20 April 2003 to 19 August 2008. In our model, we approximate RT with this empirical approximation of the disk replacerange [0, 1]. In equation form, this gives us the salvage value S = γ ment rate by using the formula RT = 0.03 * Ω * * Ω * ⎡VT⎤Ω ∗ K ∗ e–0.438 ∗ T, simplifying into ΔNPV as shown by ⎡VT⎤Ω. In this formula, the constant 0.03 represents the observed Equation 1 in the main text. 3 percent disk replacement rate.2 Thus, we can simplify ET to: ET = (( ⎡VT⎤Ω – ⎡VT – 1⎤Ω) ∗ Ω + 0.03 ∗ Ω ∗ ⎡V T⎤Ω) ∗ K ∗ e–0.438 ∗ T ⇒ ET = (1.03 ∗ ⎡V T⎤Ω – ⎡VT – 1⎤Ω) ∗ Ω ∗ K · e–0.438 ∗ T.
Disk salvage value We assume a hard disk drive can be sold in the used market for some salvage value at the end of its life. To predict this salvage value, we leverage the future disk price prediction formula, discounting the predicted price by some depreciation factor, γ, in the
We assume this vendor lets users purchase raw disk storage over the Internet. We’re only interested in the storage cloud’s pricing structure for the end user, and we don’t assume any particular service access technology. Table 2 shows the assumed tiered monthly pricing structure. For illustrative purposes, we don’t consider data uploading or downloading costs.
Single-user computers Single-user computers make up the vast majority of all shipped disk storage capacity.2 For our study, we assume the same user owns and operates the single-user computer. This user’s storage requirement grows at a moderate rate of 100 Gbytes per year. Therefore, regardless of where data is stored, we assume the level of effort in managing the system/data is approximately equal in both cases because of the low storage volume involved—that is, ρ = 0. We assume the user must purchase a new disk controller (C = $1,000), specified to consume 0.5 kW of power. Also, as storage is required, the user will purchase
References 1. O.A. Lamont and R.H. Thaler, “Anomalies: The Law of One Price in Financial Markets,” J. Economic Perspective, vol. 17, no. 4, 2003, pp. 191-202. 2. B. Schroeder and G. Gibson, “Disk Failures in the Real World: What Does an MTTF of 1,000,000 Hours Mean to You?” Proc. 5th Usenix Conf. File and Storage Technologies (FAST 07), Usenix Assoc., 2007, pp. 1-16.
COROLLARY FOR ESTIMATING THE UPPER-BOUND PRESENT VALUE
W
e derive the following corollary from the economic Law of One Price, which states, “In an efficient market, identical goods will have only one price.” Corollary 1. The upper bound of the present values for the purchase and lease price equilibrium in an efficient market is derived from the risk-free interest rate, IF. Informal proof: In an efficient market, an arbitrager has no opportunity to make a risk-free profit. Assuming an efficient market in which financing at the risk-free rate is possible for a subset of agents, the proof is by contradiction. Thus, if the present value of the lease price is higher than that derived from the risk-free rate, IF, an arbitrager can purchase disks from monies borrowed at the riskfree rate, lease at this higher price, and pocket the risk-free profit. Also, if the present value of the purchase price is higher than that derived from the risk-free rate, IF, an arbitrager can purchase a disk from monies borrowed at the risk-free rate, sell the disk at some future instance at this higher price, and pocket the risk-free profit.
APRIL 2010
47
C OV E R F E AT U R E Table 2. Hypothetical storage cloud pricing structure. Default
Storage > 50 Tbytes
Storage > 100 Tbytes
Storage > 500 Tbytes
$0.15/Gbyte/month
$0.14/Gbyte/month
$0.13/Gbyte/month
$0.12/Gbyte/month
Table 3. Buy-or-lease decision model worksheet for a single-user computer. Year (T)
Storage requirement in Gbytes (VT)
Purchased drives
Operating cost (CT)
Capital cost (ET)
Lease payment (LT)
Recurring cost (CPVT)
Salvage value (SPVT)
Incremental net present value (ΔNPV)
0
100
1
($178.70)
$154.50
$98
($236)
$15.00
($1,221)
1
200
0
($178.70)
$2.90
$278
($141)
$9.58
($1,131)
2
300
0
3
400
0
($178.70)
$1.87
$458
$131
$6.12
($863)
($178.70)
$1.21
$638
$575
$3.91
($421)
4
500
0
($178.70)
$0.78
5
600
1
($182.21)
$17.79
$818
$1,188
$2.50
$190
$998
$1,947
$3.19
$950
6
700
0
($182.21)
$0.65
$1,178
$2,884
$2.04
$1,886
Notes: C = $1,000, Ω = 500 Gbytes, Κ = $0.30 ($150 per 500 Gbytes), ρ = 0, γ = 0.1, δ = $0.04, PC = 0.5 kW, PD = 0.01 kW, and IK = 1 percent.
sion model for an expected storage life expectancy of 0 to 6 years. We compute the values for the “purchased drive” column using VT 500 – VT–1 500, and the ΔNPVs using Equation 1. In addition, we calculate the recurring costs (CPVT) and salvage values (SPVT) representing terms in the formula:
ΔNPV (thousands of dollars)
0 –50
–100 –150 –200 –250 –300
ΔNPV = CPVT + SPVT – C 0
2
4
6 Operational years
8
10
12
T
CPVT = ∑
CT – ET + IT
(1 + I )
T
T – 0
Figure 1. The calculated ΔNPV values for a medium-size enterprise deciding whether to buy or lease storage with life expectancy from 0 to 10 years.
SPVT =
K
S
(1 + I )
T
.
K
ΔNPV (thousands of dollars)
150 100 50 0 –50
–100 –150 –200
0
2
4
6 Operational years
8
10
12
Figure 2. The calculated ΔNPVs for a large-size enterprise deciding whether to buy or lease storage with life expectancy from 0 to 10 years.
500-Gbyte disk drives (Ω = 500), with a present cost of $150 per drive (K = $0.30), each specified to consume 0.01 kW of power. Finally, we assume the electric utility cost is $0.04 per kilowatt hour (kWh), and the end-of-life disk salvage depreciation factor is 0.1. Table 3 shows the worksheet for the buy-or-lease deci-
48
COMPUTER
The work sheet shows that for a n expected storage life expectancy of less than four years, leasing is always the preferred option. This is because of the high cost of the disk controller the user purchases at the outset. If the storage system will likely exist for more than four years, however, purchasing storage is the superior choice.
Medium-size enterprises
Medium-size enterprises—those with clusters of tens to hundreds of servers— made up approximately 50 percent of all data centers installed in the US from 2000 to 2007.12 We assume our medium-size enterprise has a storage requirement that grows at 1 Tbyte a year to serve a moderate-size data center of tens or low hundreds of servers. For such an enterprise, the human operator burden in owning and operating a storage cluster will be larger
than that of leasing from the storage cloud vendor. We therefore assume the proportional difference in the operator’s level of effort required to manage the system/data (H T = $70,000 per year) is ρ = 0.5. We assume that the firm must purchase an enterprise-class RAID disk controller (C = $2,000), specified to consume 0.7 kW of power. Also, as storage is needed, the enterprise will purchase 1-Tbyte disk drives (Ω = 1,000), with a present cost of $300 per drive (K = $0.30), each specified to consume 0.01 kW of power. Finally, we assume the electric utility cost is $0.04 per kWh, and the end-of-life disk salvage depreciation factor is 0.1. Figure 1 shows the calculated ΔNPV values for a storage life expectancy of 0 to 10 years. In all the operational lifetimes examined, the model shows that leasing is always preferable to purchasing storage. In this case, the clear recommendation to the medium-size enterprise is to lease storage from the storage cloud vendor.
Large-size enterprises Next, we look at the benefits of purchasing versus leasing storage for a large-size enterprise—for example, a data center with thousands of servers. In this scenario, we assume the large enterprise’s storage requirement grows at 10 Tbytes per year. In this case, the human operator burden in owning and operating a storage cluster is even larger than that of leasing from the storage cloud vendor. We therefore assume the proportional difference in the operator’s level of effort required to manage the system/data (HT = $70,000/year) is ρ = 1.0. We also assume that the firm must purchase an enterprise-class RAID disk controller (C = $2,000), specified to consume 0.7 kW of power. Furthermore, we assume the controller has a peak capacity of 100 Tbytes, and the firm will purchase additional controllers as the storage need arises. For the actual storage, the enterprise will purchase 1-Tbyte disk drives (Ω = 1,000), with a present cost of $300 per drive (K = $0.30), each specified to consume 0.01 kW of power. Finally, as before, we assume the electric utility cost is $0.04 per kWh, and the end-of-life disk salvage depreciation factor is 0.1. Figure 2 shows the calculated ΔNPVs for a storage life expectancy of 0 to 10 years. As the graph shows, leasing storage is advantageous up to a nine-year storage life expectancy. After that, it becomes more advantageous for the enterprise to purchase and maintain a storage cluster. Thus, the final decision to buy or lease storage will depend on the expected use of the storage and data. For example, if the storage is destined for use by a server cluster with a five-year life expectancy, the enterprise should lease storage. However, if the storage is destined for a long-term archival system with an indefinite life expectancy, the enterprise should purchase storage instead.
LATENCY IS NOT ZERO A common flawed assumption in designing distributed systems is the notion that latency is zero. In fact, latency isn’t zero for cloud services. Accessing storage from across the commodity Internet can incur a substantial cost in terms of I/O latency. Our model doesn’t account for this latency. However, future extensions can incorporate this factor by estimating the profit parameter, PT, which we assumed to be equal when deriving our current model. We can use this profit parameter to reward services with faster response times. For example, an enterprise might naïvely consider a service that’s two times faster to be more productive, and hence two times more profitable. For the moment, we leave this substantive extension to a future time.
O
ur primary purpose in this article is to stimulate discussion, debate, and future work in the quantitative modeling of the cloud computing industry. To this end, we propose a model to assist consumers, researchers, and policy makers in estimating the benefit of leasing from storage clouds. Ultimately, an organization’s buy-or-lease decision will depend on their anticipated parameters in the analysis. Our model simply provides a first stepping-stone for rational decision making to prevail in the cloud computing market.
Acknowledgments This article is based on work supported in part by US National Science Foundation grant 0721931.
References 1. K. Chandar, “SSD & HDD Market Tracker,” iSuppli research report, 2009; www.isuppli.com/Abstract/ ABSTRACT - SSD_HDD Market Tracker 2009.pdf. 2. P. Lyman and H.R. Varian, “How Much Information?” Oct. 2003, School of Information Management and Systems, Univ. of Calif., Berkeley; www2.sims.berkeley. edu/research/projects/how-much-info-2003. 3. N. Leavitt, “Is Cloud Computing Really Ready for Prime Time?” Computer, Jan. 2009, pp. 15-20. 4. A. Henry, “Keynote Address: Cloud Storage FUD (Failure, Uncertainty, and Durability),” presented at the 7th Usenix Conf. File and Storage Technologies, 2009; www.usenix.org/media/events/fast09/tech/ videos/henry.mov. 5. R.W. Johnson and W.G. Lewellen, “Analysis of Leaseor-Buy Decision,” J. Finance, vol. 27, no. 4, 1972, pp. 815-823. 6. G.B. Harwood and R.H. Hermanson, “Lease-or-Buy Decisions,” J. Accountancy, vol. 142, no. 3, 1976, pp. 83-87.
APRIL 2010
49
C OV E R F E AT U R E 7. P.F. Anderson and J.D. Martin, “Lease vs. Purchase Decisions: A Survey of Current Practice,” Financial Management, vol. 6, no. 1, 1977, pp. 41-47. 8. M. Palankar et al., “Amazon S3 for Science Grids: A Viable Solution?” Proc. Int’l Workshop Data-Aware Distributed Computing, ACM Press, 2008, pp. 55-64. 9. E. Deelman et al., “The Cost of Doing Science on the Cloud: The Montage Example,” Proc. ACM/IEEE Conf. Supercomputing, IEEE Press, pp. 1-16. 10. E. Walker, “The Real Cost of a CPU Hour,” Computer, Apr. 2009, pp. 35-41. 11. W.G. Lewellen, The Cost of Capital, Wadsworth Publishing, 1970. 12. US Environmental Protection Agency, Energy Star Program, Report to Congress on Server and Data Center Energy Efficiency, Public Law 109431, Aug. 2007; www.energystar.gov/ia/partners/ prod_development/downloads/EPA_Datacenter_ Report_Congress_Final1.pdf. Edward Walker is a researcher with the Texas Advanced Computing Center at the University of Texas at Austin. His research interests include distributed systems, storage systems, and operating systems. Walker received a PhD in
computer science from the University of York, UK. He is a member of IEEE and the ACM. Contact him at ewalker@ computer.org. Walter Brisken is a scientist at the National Radio Astronomy Observatory. His research interests include pulsars, the ionized interstellar medium, and very long baseline interferometry. Brisken received a PhD in physics from Princeton University. He is a member of the American Astronomical Society. Contact him at
[email protected]. Jonathan Romney is a scientist at the National Radio Astronomy Observatory. His research interests include active galactic nuclei and astronomical instrumentation. Romney received a PhD in astronomy from the California Institute of Technology. He is a member of the American Astronomical Society, the International Astronomical Union, and the International Union of Radio Science (URSI). Contact him at
[email protected].
Selected CS articles and columns are available for free at http://ComputingNow.computer.org.
UÊ
Welcomes Your Contribution Computer magazine looks ahead to future technologies
UÊComputer, the flagship publication of the IEEE Computer Society, publishes peer-reviewed technical content that covers all aspects of computer science, computer engineering, technology, and applications. UÊArticles selected for publication in Computer are edited to enhance readability for the nearly 100,000 computing professionals who receive this monthly magazine. UÊReaders depend on Computer to provide current, unbiased, thoroughly researched information on the newest directions in computing technology. To submit a manuscript for peer review, see Computer’s author guidelines:
www.computer.org/computer/author.htm
50
COMPUTER
