See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/334030324

Electrical Energy Consumption Model of Internal Components in Data Centers

Conference Paper · September 2019

DOI: 10.1109/ISGTEurope.2019.8905595

CITATIONS READS
11 1,299

4 authors, including:

Kazi Main Uddin Ahmed Jil Sutaria

Luleå University of Technology Luleå University of Technology
16 PUBLICATIONS 161 CITATIONS 19 PUBLICATIONS 86 CITATIONS

SEE PROFILE SEE PROFILE

Math Bollen
Luleå University of Technology
595 PUBLICATIONS 19,263 CITATIONS

SEE PROFILE

All content following this page was uploaded by Kazi Main Uddin Ahmed on 04 July 2019.

The user has requested enhancement of the downloaded file.

 Electrical Energy Consumption Model of Internal
Components in Data Centers
Kazi Main Uddin Ahmed∗ , Jil Sutaria† , Math H. J. Bollen‡ and Sarah K. Rnnberg§
Electric Power Engineering, Lule University of Technology,
Skellefte, Sweden.
Email: ∗ kazi.main.uddin.ahmed@ltu.se, † jil.sutaria@ltu.se, ‡ math.bollen@ltu.se, § sarah.ronnberg@ltu.se

Abstract—In the context of modern information technology cooling infrastructure [5], [6]. However, most of the literature
(IT) industry, cloud computing is gaining popularity for big data focuses on modelling and analyzing subsystems of a data
handling. Therefore, IT service providers like Google, Facebook center, e.g. uninterrupted power supply unit (UPS), power
and Amazon are expanding their technical resources by building
data centers to improve the data processing and data storage distribution unit (PDU), Computer room air handling system
facilities under cloud service pattern. However, data centers (CRAH) etc, as independent entities without considering the
consume a large amount of electrical energy. In recent years, influences they have on the electrical energy consumption of
a lot of research has been done to reduce the electrical energy other parts of the data center [7], [8].
consumption of data centers by high performance computing. This paper presents the models of above-mentioned mod-
However, very few researchers have focused on the electrical
energy consumption by the electrical components inside the ules for a data center with N server racks. Detailed energy
data center. In this paper, a component based electrical energy consumption modelling for each component as well as the
consumption modelling approach is presented to identify the interactions between all components are presented to identify
losses of different components as well as their interactions component losses of the internal grid of data center. Unlike
to the total electrical energy consumption of the data center. existing research, the presented modular simulation model is
The electrical energy consumption models of servers and other
components are presented as a function of server utilization. independent of the main source of electricity and can take
Index Terms—server utilization, component loss modelling, different design structures of data centers into account to
PDU and UPS, data center. provide hourly electrical energy consumption profiles for each
component. The same modelling approach can be utilized in
I. I NTRODUCTION renewable integrated microgrid interfaces for simulation.
With the rapid development of cloud computing and data
storage, data center technologies have become the main con- II. DATA CENTER MODELLING IN LITERATURE
cern part of innovation for IT industries. Thus the performance Numerous academic studies and industrial white papers
improvement is not the only requirement for data centers describe the electrical energy consumption models of indi-
deployment, a lot of attention is being paid for data centers vidual data center subsystems. [9] presented the component
electrical energy consumption [1]. A recent study shows that based energy consumption model of the data center for the
the data centers worldwide consumed 270 TWh of electrcial first time, shown in Fig. 1. A comparison of the energy
energy in 2012 and this consumption had a Compound Annual consumption level of different components is shown in Table I
Growth Rate (CAGR) of 4.4% from 2007 to 2012 [2]. A [5]. Although IT loads and cooling stand for almost 86% of the
typical data center may consumes thousand times higher en- energy consumption in Table I, losses in power conditioning
ergy than households, e.g. Barcelona Supercomputing Center equipment and network equipment (13%) also affect the effi-
(a medium-size data center) pays an annual bill of about e1 ciency [10]. Other studies have focused on understanding the
million for its electrical energy consumption of 1.2MW, which thermal implications of data center design [10], [11]. Energy
is equivalent to 1, 200 households [3]. Against the background consumed by the lighting facilities are reported to be negligible
of a growing energy consumption of data centers, the Nordic in most of the literature [9], [11]. As sophisticated power
region has attracted significant data center investments. For management features proliferate, the dynamic range of data
example, after Google and Facebook entered the region in center power drawn (as a function of utilization) is increasing,
2009 and 2011, the Nordic countries have become a preferred and interactions among power management strategies across
site location by an increasing number of data center investors. subsystems grow more complex. However, few attempts have
A report by Business-Sweden estimates that the Nordics by been made to identify the component-based losses of data
2025 could attract investments for data centers in the order centers, which is presented in this paper.
of e2-4 billion. This is based on the forecast of worldwide The next section will describe a detailed model of the
demand for data center services corresponding the data center IT equipment with server rack fans, power conditioning
investments of the Nordic countries [4]. Considering those equipment and different cooling systems. In Section IV, the
facts, researchers have paid much attention to the problem of simulation results will be discussed and analyzed to identify
enhancing data center efficiency including data scheduling and the amount of consumption of different components. Section
 2
V will conclude the paper with recommendations to reduce Pstatic = a2 Tdie + a3 Tdie (3)
the losses of the data center.
Compared with the CPU energy consumption, other parts of
server like memory units, power supply and hard disk drives
consumes few percentage of server power (1.5%, 4.5% and
8% of CPU energy consumption, respectively) [12]. Thus the
total consumed energy of the CPU (which accounts for most
of the generated CPU heat) is assumed to be equal to the total
energy consumption by IT loads (4).The die temperature, Tdie
of any processor depends on the design fabrication process
Fig. 1: Power consumption in data center model [9] and efficiency of the heatsink on top of that. Based on the
empirical values given in [12], depending on server utilization
Tdie varies between 70°C to 95°C. Thus a linearized model of
TABLE I: Component based energy consumption individual server power can be derived as a function of server
utilization, as (5). Regarding the idle power, Pidle , an IT load
Servers Cooling Power Cond. Eqip. Network Eqip. Lighting
56% 30% 8% 5% 1% scheduling approach (PowerNap) has been proposed in [14],
which is out-of-scope for this paper.

III. M ODELLING OF COMPONENTS IN DATA C ENTER 2

PIT ∝ PCPU = a0 + a1 uCPU + a2 Tdie + a3 Tdie (4)
As mentioned earlier, the main goal of this paper is to
propose a component based energy consumption modelling PITi = Piidle + (Pipeak − Piidle )ui (5)
approach. Finally, the total power consumption profile of any
data center is calculated considering its components and envi- where ui is the utilization of ith server, which varies depending
ronmental data. The following sections will present models for on the work load and resource allocation to this server.
the energy consumption of the IT system and cooling system
of a typical data center. Server utilization, ui is a unitless quantity that can vary
between 0% (no work load) to 100% (maximum work load).
A. IT System Modelling
If the server cluster is assumed to be homogeneous for N
The servers in a typical data center contain processing number of servers, the degree of utilization can be express as
units, memory devices, hard drives, power supply and (6)
cooling fans. Some specialized servers are integrated with
Graphics Processing Unit (GPU) and related supply unit for
1 N
high-end data processing. For simplicity, only the typical U= ∑ ui (6)
server units showing predominant energy consumption have N i=1
been considered for this study. Losses of the server power
supply unit (PSU) are neglected for this study, because 2) Server fan consumption: Apart from heatsink attached
the loss of PSU does not vary with server utilization [12]. cooling fan for CPUs, independent server cooling fans are
For developing the energy consumption model of a typical installed with each individual server. Recent high-density
server unit we consider, 16 blade server rack with 2 CPU, servers have variable airflow control to ensure the reliable
6×4GB memory and 2×75GB hard disk drive per blade [13]. operation of the server cooling system. In [12] the power
drawn by the server fan, Ps f an is shown as a function of
1) CPU heat generation: The CPU heat generation (i.e., fan rotor speed, RPM using empirical values (7). Controlled
power usage) that determines the thermal characteristics of a rotor speed ensures the required airflow into the server, which
server varies with CPU utilization and CPU die temperature relates with the thermal model of server inside environment.
(i.e. junction temperature) [12], expressed in (1). The CPU However, thermal resistance Rtot of the server depends on the
dynamic power, Pdynamic can be expressed in terms of the CPU difference of server room ambient temperature, Tamb and CPU
utilization, uCPU , that varies in proportion to the CPU clock die temperature, Tdie , referred by (8) and (9).
frequency, f and CPU supply voltage, Edd , in (2), where the
Ps f an = a4 + a5 RPM + a6 RPM 2 + a7 RPM 3 (7)
constant term Ce is the effective switching capacitance. Also,
the CPU static power Pstatic is associated with the current
leakage depending on the CPU die temperature Tdie , shown Tdie − Tamb
Rtot = (8)
in (3) [12]. PITi

PCPU = Pdynamic + Pstatic + Pidle (1) a

log R 9−a
!
tot 8
1 a9 a10
2
Pdynamic = Ce Edd f = a1 uCPU (2) Rtot = a8 + =⇒ RPM = 10 (9)
2 (RPM)a10
 3) Server power consumption: In general, it is not a !
practice to measure the actual power consumption of every Loss idle
PUPS = PUPS + ΦUPS ∑ PPDU (14)
server unit (10). Therefore, the impact of task consolidation PDUs
L for the running servers is modelled as in (11), where L = 1
means perfect load balancing between maximum number of where, ΦUPS represents the UPS power loss coefficient. A
running servers and L = 0 represents unbalanced load distri- UPS typically consumes around 9% of its input power at full
bution between running servers. Considering the best case of load. However, it is very unlikely to run the UPS at full load
uniformly distributed work load (U = ui for L = 1) the total constantly, because the PDUs connected with a single UPS
power consumed by the servers, Ps f is the summation of power always tends to change the load level depending on the server
consumed by all running server, shown in (12) load, referred by (13),(14).
B. Cooling System Modelling
Pserver = PITi + Pf an (10)
A datacenter consists of several components for handling
U the heat generated inside the server room, mainly by the
ui = (11)
U + (1 − U)L IT equipment. Normally, a server room consists of several
N racks of IT equipment arranged in rows of hot-aisle and
Ps f = ∑ Pserveri (12) cold-aisle. Rack mounted fans control the air flow and air
i=1 temperature rise inside the rack (discussed in section III-A2).
4) Power conditioning system: Data server equipment is The data center room is serviced by the Computer Room Air
very sensitive to any kind of disturbance in the electrical Conditioners/Air Handlers (CRAC/CRAH), providing precise
supply system. Thus, the internal grid of data center requires temperature and humidity control for critical environments.
considerable electrical infrastructure with specific rating to There is some constructional difference between these two
distribute uninterrupted power, with stable voltage to the types of cooling system. The CRAC unit consists of a
server. PDUs transform the high voltage distributed power condenser for transferring the heat to the outside of the
to different low voltage levels (12V, 5V, 1.2V ) for different data center, while a CRAH unit mainly uses a chiller plant
server components. Typical PDUs have a switch mode power to circulate the chilled water. Therefore, the CRAC unit
supply unit (SMPSU) to deliver DC supply to the racks. In consumes more energy than the CRAH unit [5]. Starting from
the first stage, line AC voltage is rectified and passed to a the Chiller, the following section will describe mathematical
capacitor, followed by a voltage regulator. As the demand of models of CRAH and CRAC.
racks varies, the SMPSU controls the switching frequency to
transfer more or less stored electrical energy from the capacitor 1) Chiller power consumption: The chiller plant removes
[15]. Thus for different power electronic devices with high heat from the warm coolant that returns from the server room;
frequency switching, PDUs incur a constant power loss as hence, this heat is being transferred to external cooling towers
well as a power loss proportional to the square of the server using a compressor. The chiller plants compressor accounts
load [9], as in (13) for the majority of the overall cooling cost in most data
 2 centers. The power drawn by the chiller depends on the amount
Loss idle of extracted heat, the chilled water temperature, the water
PPDU = PPDU + ΦPDU ∑ Pserver (13)
servers flow rate, the outside temperature, and the outside humidity.
Loss According to [8], the chillers power consumption quadratically
where PPDU represents power loss by the PDUs, ΦPDU
idle the idle increases with the amount of heat to be removed and thus with
represents the PDU power loss coefficient, and PPDU
the data center utilization. The size of chiller plant has to be
power drawn by PDU. PDUs typically consume 3% of their
0.7 × Psmax
f in order to provide sufficient cooling; the chiller
input power. As in current practice, all the PDUs remain
plant power consumption model is derived as shown in (15).
connected with the supply system, which increase the idle
losses of PDUs [5]. We consider perfect load balancing to
Pchiller = 0.7 × Psmax αU2 + β U + γ


f (15)
identify the power consumed by the PDUs at data center as
an independent module. where α, β and γ are obtained by performing curve fitting
of several samples from the real data center.
UPSs provide backup support during relatively short voltage 2) CRAH and CRAC power consumption: Based on exper-
interruptions (up to some tens of minutes) voltage dips and imental results, in [17] lists the impacts of several parameters
other severe disturbances in the supply voltage. UPS systems on the power consumption of a CRAC unit, such as the server
are typically placed in series between the utility supply and racks arrangement, ambient temperature, outside temperature,
PDUs. In commercial data centers with large number of humidity, etc. They observed that the power consumption
servers, different UPS topologies are used to allow proper required for heat cycling is linearly proportional to the volume
energy management [16]. However, without considering any of the airflow, f . The required volume of air flow in a server
energy management techniques the power loss of UPS follows room can be obtained by f = fmax × U, where fmax is the
the relation [9] as below maximum standard air flow (14000 CMH for a 7.5kW CRAH
unit). Reference (16) shows the electrical power required for (7) - (9). Though, the RPM generated by the fans is related
CRAH as a function of maximum power consumed by the to Rtot which is inversely proportional to PIT , but fan’s power
servers at the data center and required volume of airflow. consumption does not vary linearly with speed as in (9). The
coefficients chosen for (7) - (9) are given below [21].
Psmax
f
Pheat = 1.33 × 10−5 × f (16) • Fan: a4 = 0, a5 = 0.0003, a6 = −3 × 10−8 , a7 = 7 × 10−12
ηheat • RPM: a8 = 0.1325, a9 = 17440, a10 = 1.56
were ηheat denotes the heat removal efficiency of the system.
Server fan power consumption
9.2

Power (KW)
Electrical power required to transfer the heat Pheat from the
server room is derived in (17), where the idle power of the
idle can be considered as 7% to 10% of Pmax .
CRAH unit, PCRAH sf 9

idle
PCARH = PCRAH + Pheat (17) 24 48 72 96 120 144 168

As mentioned above, the CRAC uses a dedicated condenser IT load consumption

Power (MW)
46.6
as air-conditioning system for the server room, while CRAH
systems condition the air by direct energy exchange from the 46.5
air to a chiller. Considering a CRAC most of the consumed
energy is used to drive the condenser or the air conditioning 46.4
(AC) system. Based on experimental data presented in [18] 24 48 72 96 120 144 168
the energy consumed by CRAC is written as (18). The idle Time (hr)
power consumed by the CRAC unit, PCRAC idle depends on the Fig. 2: Comparison of IT loads and Server fan consumption
maximum load of the data center (10% ∼ 30% of Psmax f ).
The total power consumed by the server farm, obtained
idle from (10), is depicted in Fig. 2. This server farm consumes
PCRAC = PCRAC + ((1 +COP) × Pheat ) (18)
on average 55.54 MW. However, the box plot shown in Fig. 4
where COP is the condenser performance coefficient which gives more information about the fluctuation in the consump-
varies between 3 to 6; the higher values denote higher effi- tion pattern, which is important to control power conditioning
ciency of the condenser. equipment (UPS, PDU). In Fig. 4, the median consumption
value remains almost constant (around 55.55 MW) during the
IV. S IMULATION & R ESULT D ISCUSSION
week, while showing more at the beginning of the week
A. Server loads and related losses (55.6 MW). This type of characteristics is very common in
As a use-case study we assume a data center with a server special community based data center, e.g. data center hospital,
farm consisting of 40,000 server racks, which is a moderate university or government data center. Apart from that, the
size compared to the existing tech-giants [16]. Additional bottom part of each box, which is 25% of the server farm
energy consumption from office buildings, lighting systems power consumption, remains almost the same for each day.
and security systems to facilitate the operation of data center is However, the upper edge of the box, which shows the 75%
not considered in this study. Each of those racks contains, 16- consumption level, varies between 55.65 MW to 55.55 MW
Blade server with typical computational resources described during the week. This analysis shows that there is always a
in Section III and 10 cooling fans to manage the sufficient base-load to support the computational workload, which varies
air flow into the racks (considering Dell PowerEdge M1000e) depending on the CPU utilization. These variations affect the
[19]. Based on the data analyzed in [20] for different server losses of the power conditioning equipment 13) (14).
racks, we choose power consumption for each rack varying Considering the losses of the electrical equipment (UPS and
between 750 W (at idle mode) to 1500 W (at peak load). PDS), which directly relate to the total server farm power
Regarding the utilization of servers, it is also assumed that consumption, as mentioned in (13) (14), hence related with
the server clusters are homogeneous and load is balanced, total server farm utilization. The total losses of UPS and PDU
as mentioned in Section III-A. We used random generated as a function of total utilization are shown in [18], which has
server utilization factors for each server, to calculate the been used to calculate those losses for our case study. It is
power consumption of CPUs and fans of the racks in (1) - (7). assumed that PDUs and UPSs are loaded uniformly where
each PDU is connected with 10 server racks and each UPS
The total consumption pattern of the IT loads and fans over supports 10 PDUs [9].
a week time is shown in Fig. 2 and Fig. 3. Rack mounted fans Fig. 5 is showing the power losses of a single PDU com-
consume around 9 MW while IT loads consume 46.5 MW on pared with a single UPS. A single PDU is more energy
average over a week. Within this time period, fluctuations with efficient than the UPS, because of the built-in power electronic
some peaks are being observed for both cases; peaks in IT converters in the UPS that consume more than the PDU.
loads correlate with peaks of fan consumption as shown in However, total PDU loss is higher than the total UPS loss
 Total Sever Farm Power 2000
55.8
1500

Power Loss (kW)

Total Power loss of PDU
55.7 Total Power loss of UPS
Total Power loss by elec. equip.
1000
Power (MW)

55.6
500

55.5
0
1 6 12 18 24
55.4 Time (one day)
Fig. 6: Total Losses of electrical equipment
55.3
24 48 72 96 120 144 168
Time (hr) every CRAH system is supposed to utilize chillers to circulate
Fig. 3: Total IT Load variation within a Week the chilled water within the data center. Although, the power
consumed by a CRAC unit is higher than CRAH, power
Daily Variation in Server Farm Consumption consumption of a CRAH unit with a chiller cumulatively is
almost the same as CRAC, as shown in Fig. 8. In Fig. 8,
55.7 power consumption of different cooling system components
are compared with the amount of heat generated from the rack
Power (MW)

55.6 level IT load. In [12] the electrical energy consumption level

of the chiller and CRAH unit is analyzed according to data
55.5 centers thermal guidelines of ASHRAE TC 9.9. This correlates
with the results shown in Fig. 8, although different approach
55.4 is used to model the cooling units based on server load and
utilization.
55.3
CRAC Power Consumption (MW)

1 2 3 4 5 6 7 31 6.96

CRAH Power Consumption (MW)

Time (week)
Fig. 4: Boxplot of a week server farm consumption 30.9 6.94

30.8 6.92
as depicted in Fig. 6. Considering the uniform distribution,
the total number of PDU is ten times higher than the number 30.7 6.9
of UPS, which results in the PDUs to be the most energy
consuming power conditioning equipment. 30.6 6.88

B. CRAH and CRAC consumption

30.5 6.86
24 48 72 96 120 144 168
CRAC and CRAH are popular choices for handing the heat Time (hr)
generated in a server farm. As mentioned in Section III-B, Fig. 7: Consumption comparison of CRAH and CRAC
the CRAC unit consumed comparatively more energy than the
CRAH system. A comparison of the consumption, between
the CRAC and CRAH unit, is shown in Fig. 7, showing the V. C ONCLUSION
consumption of CRAC is almost 6 times higher. However,
In this paper, a component based modelling approach for
data centers is presented. The model used for individual
1.2
components consists of a detailed analysis of the power con-
1
sumption as a function of server utilization as well as interac-
Power Loss (kW)

Loss of a unit of UPS

tions between the power consumption of different components.
0.8 Loss of a unit of PDU The presented modular simulation approach can take different
Loss of Power supply equip.
design structures of a data center into account and provide
0.6
hourly power consumption profiles for each component. A
0.4 data center as a whole is a huge load; hence, a reduction
of few percent would save a significant amount of electricity
0.2
1 6 12 18 24 cost. An interesting outcome of this study is the dominant
Time (hr) contribution of PDUs power consumption to the total electrical
Fig. 5: Losses of electrical equipment (single unit) losses, which motivates to reduce the number of PDUs in the
 800 [12] S.-W. Ham, M.-H. Kim, B.-N. Choi, and J.-W. Jeong, “Simplified server
Generated Heat Chiller CRAH CRAC
model to simulate data center cooling energy consumption,” Energy and
Amount of Power (MW)

Buildings, vol. 86, pp. 328–339, 2015.

600 [13] “Dell poweredge m1000e, technical specification,” All rights reserved.,
Tech. Rep., 2012 Dell Inc.
[14] D. Meisner, B. T. Gold, and T. F. Wenisch, “Powernap: eliminating
server idle power,” in ACM sigplan notices, vol. 44, no. 3. ACM,
400 2009, pp. 205–216.
[15] D. Meisner and T. F. Wenisch, “Peak power modeling for data center
servers with switched-mode power supplies,” in 16th ACM/IEEE inter-
200 national symposium on Low power electronics and design. ACM, 2010,
pp. 319–324.
[16] S. K. Uzaman, J. Shuja, T. Maqsood, F. Rehman, S. Mustafa et al.,
“A systems overview of commercial data centers: Initial energy and
0 cost analysis,” International Journal of Information Technology and Web
1 2 3 4 5 6 7 Engineering (IJITWE), vol. 14, no. 1, pp. 42–65, 2019.
Days of Week [17] S. Itoh, Y. Kodama, T. Shimizu, S. Sekiguchi, H. Nakamura, and
Fig. 8: Consumption of different cooling unit compared to N. Mori, “Power consumption and efficiency of cooling in a data
center,” in Grid Computing (GRID), 2010 11th IEEE/ACM International
generated heat Conference on. IEEE, 2010, pp. 305–312.
[18] R. Rahmani, I. Moser, and M. Seyedmahmoudian, “A complete model
for modular simulation of data centre power load,” CoRR, vol.
internal grid of a data center. It is also found that, both of the abs/1804.00703, 2018.
[19] J. Beckett and R. Bradfield, “Power efficiency comparison of enterprise-
existing cooling systems consume 6 ∼ 7 times more than the class blade servers and enclosures,” A Dell Technical White Paper, 2011.
amount of heat generated in the server farm. Hence, there is a [20] V. Debusschere, S. Bacha et al., “Data center energy consumption
scope of improvement in electrical power consumption of the simulator from the servers to their cooling system,” in PowerTech
(POWERTECH), 2013 IEEE Grenoble. IEEE, 2013, pp. 1–6.
cooling system. [21] S. Sampath, “Thermal analysis of high end servers based on development
of detail model and experiments,” The University of Texas at Arlington,
ACKNOWLEDGMENT Arlington. (Master’s Thesis), 2013.

This study was supported by the Swedish Energy Agency

under grant 43090-2, Cloudberry Datacenters.

R EFERENCES
[1] M. Uddin and A. A. Rahman, “Energy efficiency and low carbon
enabler green it framework for data centers considering green metrics,”
Renewable and Sustainable Energy Reviews, vol. 16, no. 6, pp. 4078–
4094, 2012.
[2] W. Van Heddeghem, S. Lambert, B. Lannoo, D. Colle, M. Pickavet, and
P. Demeester, “Trends in worldwide ict electricity consumption from
2007 to 2012,” Computer Communications, vol. 50, pp. 64–76, 2014.
[3] A. Uchechukwu, K. Li, and Y. Shen, “Energy consumption in cloud
computing data centers,” International Journal of Cloud Computing and
services science, vol. 3, no. 3, p. 31Y48, 2014.
[4] J. D. Christensen, J. Therkelsen, I. Georgiev, and H. Sand, Data centre
opportunities in the Nordics: An analysis of the competitive advantages.
Nordic Council of Ministers, 2018.
[5] S. Pelley, D. Meisner, T. F. Wenisch, and J. W. VanGilder, “Understand-
ing and abstracting total data center power,” in Workshop on Energy-
Efficient Design, vol. 11, 2009.
[6] Q. Fang, J. Wang, and Q. Gong, “Qos-driven power management of data
centers via model predictive control,” IEEE Transactions on Automation
Science and Engineering, vol. 13, no. 4, pp. 1557–1566, 2016.
[7] K. Ebrahimi, G. F. Jones, and A. S. Fleischer, “A review of data center
cooling technology, operating conditions and the corresponding low-
grade waste heat recovery opportunities,” Renewable and Sustainable
Energy Reviews, vol. 31, pp. 622–638, 2014.
[8] E. Oró, V. Depoorter, A. Garcia, and J. Salom, “Energy efficiency and
renewable energy integration in data centres. strategies and modelling
review,” Renewable and Sustainable Energy Reviews, vol. 42, pp. 429–
445, 2015.
[9] N. Rasmussen, “Electrical efficiency measurement for data centers,”
White Paper by Schneider Electric-Data Center Science Center, vol.
154, 2007.
[10] X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a
warehouse-sized computer,” in ACM SIGARCH computer architecture
news, vol. 35, no. 2. ACM, 2007, pp. 13–23.
[11] C. D. Patel, R. Sharma, C. E. Bash, and A. Beitelmal, “Thermal
considerations in cooling large scale high compute density data centers,”
in 18th Conference on Thermal and Thermomechanical Phenomena in
Electronic Systems, 2002. ITHERM 2002. IEEE, 2002, pp. 767–776.

View publication stats