Energy Stat China

Applied Energy 265 (2020) 114806
Contents lists available at ScienceDirect
Applied Energy
journal homepage: www.elsevier.com/locate/apenergy
A review of power consumption models of servers in data centers T

Chaoqiang Jin, Xuelian Bai , Chao Yang, Wangxin Mao, Xin Xu
⁎
Key Laboratory of the Three Gorges Reservoir Region's Eco-Environment, Ministry of Education, Chongqing University, Chongqing 400045, China
National Centre for International Research of Low-carbon and Green Buildings, Chongqing University, Chongqing 400045, China
HIGHLIGHTS
• Analyze the server power trends from generation to generation.

• Classify and compares existing power consumption models of servers.
• Summarize the application scenarios of the power consumption models.
• Discuss the model parameters for model construction.
• Identify the future research directions for improving data centers’ energy efficiency.
ARTICLE INFO ABSTRACT
Keywords: This study provides an overview of power consumption models of servers in data centers. The server is the basic
Server unit of both power and heat flow paths; therefore, its power consumption model can be used for both energy
Power consumption model management and thermal management. Investigations of server power trends were carried out according to the
Data center data from the Standard Performance Evaluation Corporation (SPEC). It is found that a heavier workload can be
Energy performance
handled without consuming more energy, and the difference between the peak power and idle power of the
servers is not consistent from generation to generation. Furthermore, the existing power consumption models are
categorized as additive models, baseline power + active power (BA) models, and other models based on cal-
culation formula and other factors. Specifically, there are four forms of BA models: linear regression models,
power function models, non-linear models and polynomial models. Besides, these models have been compared in
terms of accuracy. It can be found that the polynomial model and the linear regression model perform better in
terms of accuracy. Additionally, the model applications are summarized. Considering server architecture up-
grades and technological innovation, the establishment of the new model and its application scenarios are
discussed. Moreover, in-depth and accurate power consumption models must be extensively researched and
applied to effectively improve data centers, including information technology (IT) equipment and cooling
equipment, in terms of overall energy performance.
1. Introduction over-year increase of 30%, the overall market size of cloud computing
reached 94.61 billion yuan with a growth rate of 32.4%, and the overall
As the digital transformation of industry continues to accelerate, the energy consumption of data centers reached 120–130 billion kWh,
new generation of IT integration innovation (i.e., mobile Internet, which accounts for about 2% of the total Chinese electricity con-
Internet of things, cloud computing, big data, artificial intelligence, sumption [4]. In 2014, data centers in the U.S. consumed an estimated
etc.) is increasingly active [1], which drives the growth of the economy 70 billion kWh, representing approximately1.8% of the total U.S.
and improves labor productivity [2]. According to the Synergy Re- electricity consumption [5].
search Group, there were nearly 400 large-scale data centers in the Data centers have four major components: power equipment,
world in 2017. Of these data centers, 44% were in the United States, cooling equipment, IT equipment, and miscellaneous components [6];
20% were in China, Japan, and the United Kingdom, and 3–5% were in IT equipment and cooling equipment are the two main components,
Australia, Germany, Canada, Singapore, India, and Brazil [3]. In China, accounting for approximately 90% of the total energy consumption of
the size of the big data industry in 2017 was 470 billion yuan, a year- the data center [7]. As shown in Fig. 1, servers consume energy and
⁎
Corresponding author.
E-mail address: xuelianbai@cqu.edu.cn (X. Bai).
https://doi.org/10.1016/j.apenergy.2020.114806
Received 27 November 2019; Received in revised form 3 March 2020; Accepted 5 March 2020
0306-2619/ © 2020 Elsevier Ltd. All rights reserved.
C. Jin, et al. Applied Energy 265 (2020) 114806
Nomenclature OS Operating system

PCM Phase-change memory
Abbreviations PMC Performance monitoring counters
PRO The ratio of peak power to rated power
BA Baseline power + Active power RCI Rack cooling index
CFD Computational fluid dynamic RHI Return heat index
CRAC Computer room air conditioner RTI Return temperature index
CRAH Computer room air handler SHI Supply heat index
DDR Double data rate SLAs Service-level agreements
DVFS Dynamic voltage and frequency scaling SPEC Standard Performance Evaluation Corporation
I/O Input/output SSD Solid-state drive
ICT Information and communications technology TDP Thermal design power
IPO The ratio of idle power to peak power UPS Uninterrupted power supply
IRO The ratio of idle power to rated power VM Virtual machine
IT Information technology VRM Voltage regulator modules
Fig. 1. The power flow and heat flow in general data centers.
dissipate heat to the thermal environment, which determines the maximized. Furthermore, such a model plays an important role in the
cooling load of the cooling system and thereby influences the electricity cooling system design and power trend forecasting of data centers.
consumption of the cooling equipment. In turn, changing the operating The purpose of this paper is to provide an overview of the recent
parameters of the cooling system affects the inlet air temperature and power consumption models of servers and related studies. Previous
air flow of servers, which impact the energy consumption of the servers. reviews of data center thermal environments and energy efficiency are
Hence, the energy consumption of servers is the basic unit of power summarized. Subsequently, the development trend of server power
flow and heat flow in data centers, and IT equipment and cooling consumption, which is one of the factors to be considered in the es-
equipment are coupled due to the thermal environment. Moreover, tablishment of power consumption model, is analyzed. Then, the recent
with the continuous improvement of equipment efficiency, the energy power consumption models of servers are classified, and their merits
savings of IT and cooling equipment will be maximized, and the energy and demerits are compared. Furthermore, the application of these
conversation potential will be reflected in power and thermal man- models is discussed. Finally, further studies of power consumption
agement. To perform these two management functions in data centers, models are analyzed and discussed.
an accurate power consumption model of servers is necessary. Such a
model helps in the identification of optimization opportunities and in
the prediction of the consequences of the decisions and results in more 2. Related work
effective management/control; consequently, energy savings can be
Although the modeling, prediction, and application of the energy
2
Table 1
Summary of review articles about thermal environment, energy efficiency and power models for data centers.
Reference Major contents Major contributions and conclusions
Lu et al. [8] (1) Airflow distribution in data centers including ventilation Geometry of plenum and inner obstructions are major factors for airflow
configurations, underfloor plenum, and row and rack-based distribution in underfloor data centers. Numerical simulations have shown
solutions great agreements with the experiment data. Different combinations of air
(2) Airflow management and optimization in cold aisle containment distribution systems and airflow manage methods need to be explored to
and hot aisle containment systems improve the thermal environment
Alkharabsheh et al. [9] (1) Numerical modeling of raised-floor plenum, perforated tiles, rack, Numerical modeling with emphasis on CFD, experimental measurements, and
CRAC/CRAH, room level, and dynamic models recent cooling technologies (containment systems, economizer cooling,
(2) Experimental measurements hybrid-cooling, and device level liquid cooling) are extensively reviewed.
(3) Recent technologies including containment systems, economizer There are many challenges facing thermal management in data centers, such a
cooling, hybrid-cooled systems, and device level liquid-cooling workload variation, environment changes, and scaling issues (data centers
solutions vary in size, complexity, and business objective).
Chu and Wang [10] (1) Methodologies in exploration the performance of data centers CFD plays vital role in data center design and simulation. However, the
(2) Long-distance cooling including raised-floor and overhead air simulation scale may change from sub-micro to hundreds meters. For long-
supply distance cooling systems, the outside ambient air can be directly brought into
(3) Short-distance cooling and airflow management of rack-level data centers through filters which temperature is lower than that in data
cooling center. For short-distance cooling systems, the heat pipe or thermosyphon
system may have better energy efficiency and cooling capacity.
Rambo and Joshi [11] (1) Data center modeling objectives Future trends in reduced or compact modeling of data center airflow and heat
(2) Numerical modeling transfer are presented to serve as an overview of integrating rack-level
(3) Model validation compact models into full-scale facility level numerical computations. Compact
(4) Rack-level compact modeling models can be used to efficiently model data centers through varying model
(5) Data center dynamics fidelity across length scales. Dynamic effects can be included to develop next-
generation control schemes to maximize data center energy efficiency.
Wang and Khan [12] (1) Green data center practice and benchmarking Performance metrics are important for designing, building and evaluating a
(2) Green performance metrics data center. It also encourages the use of new technologies for next generation
(3) Greenhouse gas emission data centers. The key driving features to develop above metrics include:
(4) Humidity economic features and environment concern. Metrics for balancing purchasing
(5) Thermal metrics cost and operation cost and for balancing compute performance and green
(6) Power/energy metrics performance need to be addressed.
(7) Multiple indicators
(8) TCO: total cost of ownership
Jin et al. [13] (1) Factors influencing the thermal environment from room level to As for the room level, it is recommended that the CRACs locate at the ends of
server level each rack row could be the optimal choice, and the height of plenum and
(2) Numerical study including simulation software and validation ceiling is 0.79–0.91 m and 2.52 m, respectively. Besides, it is suggested that
(3) Airflow performance metrics and thermal optimization the open area of the perforated tiles is 25%, and the airflow outlet angle is 60°.
Comparing existing simulation tools, 6SigmaDCX is recommended because of
its specialty, easy to use. If engineers want to assess the magnitude of
recirculation and mixing of hot and cold air, SHI or RHI combining with β
index can be the best choice. If the engineers want to address the cooling
efficiency, RCI and RTI are recommended.
Ge et al. [14] (1) General power-saving strategies Dynamic provisioning-based strategy is the most effective power-saving
(2) Taxonomy of power-saving research in data center networks and approach. Response time plays an important role for data center network and
content delivery networks content delivery network to pursue better power-saving effects while keeping
(3) Power-saving schemes in data center network and content the power-performance tradeoff well balanced.
delivery network
Mittal [15] (1) The techniques for managing power consumption of embedded Power management techniques which ensure high-performance and low-costs
system are expected to become a crucial part of future enterprise architectures.
(2) The need of power management
(3) A classification of the techniques on similarities and differences
Mittal [16] (1) DVFS based techniques The leakage energy consumption is increasing and hence the dynamic range of
(2) Server consolidation and power state transitioning based energy consumption that DVFS can utilize has reduced.
techniques
(3) Workload scheduling based techniques
(4) Thermal-aware power management techniques
Orgerie et al. [17] (1) The techniques for improving the energy efficiency of computing Energy-efficiency approaches and technologies are often combined and
and networking resources applied in a coordinated way in large-scale distributed systems. Although
(2) Studies and models for estimating the energy consumption of many research directions have been studied to save energy, several key
these resources problems remain open.
(3) A classification of existing solutions and research work
Beloglazov et al. [18] (1) Causes and problems of high power/energy consumption Currently IT infrastructures contribute about 2% of total CO2 footprints. And
(2) A taxonomy of energy-efficient design of computing systems the power management problem becomes more complicated when considered
covering the hardware, operating system, virtualization and data from the data center level.
center levels
Bostoen et al. [19] (1) Research on power-aware enterprise storage systems A majority of power-reduction techniques is based on dynamic power
(2) Exiting power-reduction techniques management. For every energy-conservation technique, the fundamental
trade-offs between power, capacity, performance, and dependability are
uncovered.
Wang et al. [20] (1) Energy efficient computer servers in hardware-level, system- For building more energy efficiency systems, we must construct a model about
level, and application-level the relationship between energy and performance in a given hardware
(2) Energy efficient cluster systems in virtualization, cluster-level environment. The model must consider such factors as resource utilization and
optimization, heterogeneous energy strategy, and green cloud system performance requirements.
computing
Hammadi and Mhamdi (1) Conventional tree-based data center network architecture
[21] (2) The data center architecture evolution
(continued on next page)
3
Table 1 (continued)
Reference Major contents Major contributions and conclusions
(3) Data center network architectures A good architecture design should be tailored along with different algorithms
(4) The energy-saving approaches and technologies to address not only power consumption but also network performance along
with cost.
Darahmeh and Wang (1) Airside economizers in direct free cooling, indirect cooling and The free cooling technologies applicable for data centers were summarized,
[22] multi-stage evaporative cooling systems including airside economizers, waterside economizers and heat pipe
(2) Waterside economizers in integrated dry cooler-chiller system application. The indirect airside coolers such as air-to-air heat exchanger
and cooling tower system system and heat wheel shows very high efficiency, yet the thermosyphon
(3) Cold energy storage systems and integrated system of mechanical system reveals even more promising features.
refrigeration and thermosyphon
Ni and Bai [23] (1) Air conditioning summary including the indoor thermal Analyzing energy consumptions for air-conditioning system from 100 data
guidelines, cooling methods and air distribution in data centers centers. Some currently available energy efficiency strategies like economizer,
(2) Energy performance in mechanical cooling equipment, cooling airflow optimization, energy management and simulations tools are reviewed
distribution equipment and heat rejection equipment. and summarized. The collected data from articles and reports showed that the
(3) Energy efficiency strategies by applying economizer cycles, range of air-conditioning system energy usage was 21% for the most efficient
airflow optimization and energy management. system and 61% for the least efficient system.
Shuja et al. [24] (1) Computing systems including server architectures, power Energy-efficiency techniques conflict with the business objectives and call for
distribution, and cooling reduced redundancy of cloud resources that may lead to violation of user
(2) Storage systems including storage architectures, technologies, SLAs. Energy-efficient hardware devices such as SSDs and optical
and DRAM architectures interconnects do not fit into current cloud paradigm due to their high cost.
(3) Communication systems including data center network Furthermore, energy-efficient hardware devices such as ARM processors and
architecture, optical data centers, and data center routing PCM drives do not provide comparable performance to prevalent hardware
technology.
Bhattacherjee et al. [25] (1) Existing resource management techniques including resource Prediction models can be used to predict the future behaviors of the physical
scheduling, load balancing, and migration machines and the virtual machines by recording and analyzing their past
(2) Energy efficient resource management behaviors.
Atiewi et al. [26] (1) Virtual machine scheduling Most algorithms perform scheduling based on one or two parameters. A better
(2) Scheduling model in cloud data center scheduling algorithm can be developed from existing methods by adding more
(3) Scheduling parameters metrics, which can result in good performance and outputs that can be
(4) Existing energy-efficient task scheduling algorithms deployed in a cloud environment in the future.
Vasques et al. [7] (1) Energy efficiency strategies of CPU, memory, disk, network Studies to further develop strategies in small and medium data centers are
interface, virtualization framework, uninterrupted power supply, needed, including more case studies, simulation, planning of polices, and
and cooling overall evaluation to provide a clearer view of how the aforementioned actors
(2) Demand response in data centers can as a matter of fact benefit from each other, thereby promoting effective
(3) Small and medium data center perspective transformations in the energy market.
Reda and Nowroz [27] (1) Techniques in power modeling and characterization for three Power consumption is contributed by two components: dynamic power and
computing substrates static power. Power estimation is highly challenging due to many modeling
(2) The basic principles that govern power consumption in digital complexities and unknown factors during design time. Popular methods to
circuits adjust performance and power consumption include dynamic frequency and
(3) Utilize these principles to describe high-level power modeling voltage scaling, nap modes, and sleep modes.
techniques
Zhang et al. [28] (1) Cooling infrastructure of data centers includes air cooling, liquid Based on the available techniques of data center operation and previous work
cooling, and immersion cooling on data center management (i.e., modeling, optimization, algorithms and
(2) Techniques of ICT and cooling management from chip level testbeds), the learning-based approach can be a promising framework for the
management to room level management joint ICT and cooling management in the data center, from data profiling,
(3) Models of ICT and cooling management including ICT workload, learning, optimization to execution. Equipment safely and hardware reliability
ICT power, heat transfer, ambient temperature, and cooling are important for the performance and operational cost in the data center. In
power addition, the hybrid solution of cooling technologies and the integration of
(4) Optimization of ICT and cooling management including ICT renewable can also be the future for enhancing the energy efficiency.
control, cooling control optimization
(5) Algorithms of workload scheduling including heuristic, static
cooling, and dynamic cooling scheduling algorithms
Mobius et al. [29] (1) Estimation models’ essential steps: model inputs and training All of the models use either OS-provided metrics, PMC, or a mixture of these.
model with benchmarks Most of the models are trained and tested with standard benchmarks. A
(2) CPU models feedback system or a belief revision mechanism is more realistic to implement
(3) Virtual machine models as is done in probabilistic parameter estimation.
(4) Server models
Dayarathna et al. [30] (1) Data center energy consumption More than 200 power models were examined in this survey. While there has
(2) Digital circuit level energy consumption modeling been a large number of studies conducted on the energy consumption
(3) Server energy models modeling at lower levels of the data center hierarchy, much less work has been
(4) Processor power models done at the higher levels. The accuracy, generality, and practicality of the
(5) Memory and storage power models majority of the power consumption models remain open.
(6) Data centers level energy consumption modeling
(7) Software energy models
(8) Energy consumption modeling using machine learning
consumption of data centers have been researched, few reviews have centers via ventilation configurations, underfloor plenum, and row- and
been conducted in this area. Most published reviews of data centers can rack-based solutions [8]. Alkharabsheh et al. [9] and Chu and Wang
be classified into two categories: thermal environment and energy ef- [10] summarized the existing thermal management techniques by
ficiency. Table 1 summarizes the existing reviews of thermal environ- considering the solutions and the air supplies, respectively. Rambo and
ment, energy efficiency and power models for data centers. The reviews Joshi overviewed data center modeling by considering model valida-
of the thermal environment mainly focus on air flow distribution, tion, rack-level compact modeling and dynamic modeling [11]. The
thermal managements, numerical simulation, airflow performance, and survey conducted by Wang and Khan focused on the green performance
thermal optimization. Lu et al. surveyed the airflow distribution in data metrics, thermal metrics, power/energy metrics of the data centers
4
[12]. Jin et al. surveyed the factors influencing the thermal environ- energy savings and emission reductions of data centers. The major
ment, from room level to server level, as well as simulation software, contributions of this review are as follows.
airflow performance metrics and thermal optimization [13]. These re-
views cover thermal environment design, optimization, control, and • The server power trends are analyzed by considering IT develop-
evaluation, but they do not consider the power consumption model at ment.
the component, server, or data center level. • The existing power consumption models of servers are reviewed by
The reviews of energy efficiency mainly focus on energy-efficient considering the calculation formula and other factors, and these
techniques cover component-level, server-lever, and room-level. For models are compared in terms of advantages and disadvantages.
example, Ge et al. [14] conducted a survey of the power-saving tech- • The power consumption models of servers applied for energy pre-
niques used in data centers and content delivery networks. Mittal diction and management are reviewed to realize energy savings for
[15,16] surveyed the techniques for managing the power consumption server and cooling systems.
of embedded systems and data centers, including dynamic voltage and • The model parameters for model construction and the application
frequency scaling (DVFS)-based techniques, power-aware scheduling scenarios of the power consumption models of servers are discussed.
techniques, thermal-aware power management techniques, the use of • The future research directions for improving the energy efficiency
power modes, energy savings in specific components, and the use of and reducing the emissions of data centers are identified.
unconventional cores. Orgerie et al. [17] reviewed the techniques and
solutions that aim to improve the energy efficiency of computing and The review method is briefly introduced as follows. Studies that are
wired network resources and discussed techniques that operate at the directly or indirectly related to the power consumption models of ser-
infrastructure level by coordinating the energy-saving capabilities of vers in data centers, have been selected. These studies include journal
individual hardware and software components. Beloglazov et al. [18] papers, conference papers, white papers, handbooks and reports.
created a taxonomy of energy-efficient design at the hardware level, Through preliminary searches, we compared databases and chose two
operating system level, virtualization level and data center level. Bos- that generated especially relevant results: Science Direct and IEEE.
toen et al. [19] summarized the power-reduction techniques for data- Combined searches have been conducted based on two or more of the
center storage systems. Wang et al. [20] overviewed the energy-saving following keywords: data center, server, power consumption, model,
techniques developed at the hardware, system software, and applica- and cloud computing. Closely related studies were also checked.
tion levels, with an emphasis on energy-efficient computer servers and
cluster systems. Similar work was also conducted by Hammadi and 3. Classification of servers and power consumption models
Mhamdi for data center networks [21]. Furthermore, Refs. [22] and
[23] discussed the applicability of free cooling, which is an energy- 3.1. Classification of servers and power trends
efficient technology for cooling systems. Additionally, energy efficiency
strategies, workload balance and demand response are the main focus 3.1.1. Classification of servers
for cluster servers and cloud data centers. Shuja et al. [24] discussed the Server is a broad term describing a specific piece of IT equipment
conflict between maximizing service quality and minimizing energy that provides computing capability and runs software applications in an
consumption. Bhattacherjee et al. [25] conducted a review on energy environment networked with other IT equipment, including other ser-
resource management strategies, including resource scheduling, load vers. Most servers contain the following major hardware building
balancing, and cloud migration. Vasques et al. [7] reviewed the energy blocks: processors, memory, chipset, input/output (I/O) devices, sto-
efficiency strategies used for the CPU, memory, disk, network interface rage, peripherals, voltage regulators, power supplies, and cooling sys-
card, virtualization framework, uninterrupted power supply (UPS), and tems [31]. Servers can typically be divided into rack type, blade type,
cooling. The demand response in data centers was also discussed. User and tower type according to their configurations. A rack server, which
task scheduling is another factor that plays a significant role in im- is a server that is designed in according to uniform standards in ap-
proving services. Atiewi et al. [26] presented a review of various en- pearance, can be used with cabinets. A blade server has a blade unit
ergy-efficient task scheduling methods for cloud data centers. However, that can be inserted in a standard-height rack chassis; such servers are
these reviews do not focus on power modeling. high-density and low-cost servers. All blades can be connected to pro-
Recently, some surveys of the energy consumption models of data vide a high-speed network environment, share resources and serve the
centers have been conducted. Reda and Nowroz [27] described the same user group. The appearance and structure of a tower server are
main research directions in pre-silicon power modeling and post-silicon similar to those of a common computer, but its main board is expand-
power characterization, focusing on chip/processor level power able. Rack servers and blade servers are widely employed in data
models. Zhang et al. [28] presented the techniques for enabling in- centers. According to the application scenarios, the servers can also be
formation and communications technology (ICT) and cooling manage- classified as database servers, file servers, mail servers, print servers,
ment and the coupling models between them in data centers. Although web servers, game servers, and application servers [32]. Furthermore,
this review combines the power models of IT equipment and cooling servers can be used for scientific and analytic purposes, business pro-
systems, it does not survey power consumption models of servers. cessing, visualization and audio, communications/telecommunications,
Mobius et al. [29] provided a comprehensive survey of the proposed and storage, and in cloud/Internet portal data centers, according to the
power estimation models by considering single-core and multicore workload, and the definitions and subsystem utilization are given in
processors, virtual machines, and the entire server. Moreover, Day- Table 2 [33]. It can be seen that the servers have different usage per-
arathna et al. [30] conducted an in-depth study of the existing literature centages for different components according to their work character-
on power modeling from the hardware level to the data center level, istics, which will result in differences in the power consumption of
covering more than 200 models. Although these two reviews cover the servers and should be considered in the establishment of a power
server power consumption models, none of them only focuses on the consumption model.
power consumption models of servers considering their power trends,
categories, and applications. As mentioned in introduction, the energy 3.1.2. Server power trends
consumption of servers is the basic unit of power flow and heat flow in The SPEC is a non-profit corporation formed to establish, maintain
data centers, and the power consumption models of servers are the and endorse standardized benchmarks and tools for evaluating the
basic consideration for the energy consumption estimation of data performance and energy efficiency of the newest generation of com-
centers or that of the whole region. Therefore, the discussion of the puting systems [34]. Thus, the performance and power consumption
power consumption models of servers are of great significance to the data of servers with different configurations can be obtained from
5
different manufacturers. As of March 2019, there were 641 samples in
Medium-High
Medium-Low
SPEC; the number of samples of each year is given in Fig. 2. It can be
Medium
seen that most (407) samples were from between 2008 and 2012, and
Storage
High
the number of samples has also increased since 2018. Furthermore, 1U
Low
Low
Low
and 2U servers account for 53.8% of the samples, and the number of
servers for each height is shown in Table 3. Servers with 1, 2, 4 and 8
Medium-High
Medium-High
Medium-Low
CPUs make up 89.2% of the samples, and servers with 2 CPUs make up
Medium
Medium
Medium
62.6%. To describe the changes in server power, blade servers and other
Low
I/O
samples without nameplate power are not included, and the total
number of valid samples is 508. In addition, to present the trend over
Medium-High
Medium-High
Medium-Low
Medium-Low
time, all the samples are sorted by time, and the power at 100% is
regarded as the peak power.
Memory
Medium
Medium
As shown in Fig. 3, the rated power of the servers hardly increased

Low
over time, which could be attributed to the full-load power far from the
rated power. Fig. 4 depicts the relationships among rated power, peak
Medium-High
Medium-Low
power and idle power. It can be seen that the ratio of peak power to
Medium
Medium
rated power (PRO) increases slowly. The mean PRO increases from
High
High
CPU
Low
approximately 40% to 50%. Conversely, the ratios of idle power to

rated power (IRO) and idle power to peak power (IPO) decrease. The
mean IRO and IPO decrease from 20% to 5% and from 55% to 15%,
Scientific-based application including biological sciences, geosciences, weather forecasting, engineering, simulation, design, defense,
respectively. Moreover, the workload operation per second (ops) pre-

sents cliff growth for 1, 2, 4, 6 and 8 CPUs from 2014 (see Fig. 5), and
the corresponding mean value is six times the previous value. Then, it
can be concluded that servers are becoming efficient, as a heavier
workload can be handled without consuming more energy. These
changes indicate that the difference between the peak power and idle
Discrete data warehousing, data analysis, big data analytics, and run-time deep machine learning applications
Data center visualization applications including video processing, remote visualization, and audio processing
Enterprise wide line of business applications that manage transactional, operational, and customer databases
power of the servers is not constant from generation to generation.

Hence, these trends should be reflected in the power consumption
models.
Wikis, portals, social media, video-sharing websites, search engines, and online auction websites
Wired and wireless networking applications: application, control, packet, and signal processing
3.2. Power consumption models of the server

Dedicated storage infrastructure and services including backup, tiering, and deduplication
According to recent studies, power consumption models are estab-

lished for the following purposes: (1) to estimate the potential of power
security, and training of deep machine learning applications (versus run-time)
management; (2) to predict the power consumption of servers or data

centers; and (3) to balance the needs of the data center (i.e., reducing
energy consumption and providing high reliability). The power con-
sumption models cover physical machines and virtual machines (VMs),
and they can be classified as additive models, BA models, and other
models by considering the calculation formula and other factors. A total
of 47 models and their factors and benchmark/ testing tools are listed in
Table 4.
The workload types and common subsystem utilization per workload type [33]
3.2.1. Additive models

Fig. 6 depicts the power consumption composition of a server. It can
be found that CPU has the largest contributor to power consumption,
followed by peripheral slots (including network, I/O devices), power
supply unit, memory, motherboard, disk/storage and fan [7]. Hence,
the most direct way to model power consumption is to add the power
contributions of all components. The expression is as follows:
Definition/Examples
Ptotal = PCPU + Pmemory + Pdisk + +PI / O (1)

where PCPU, Pmemory, Pdisk, and PI/O represent the CPU, memory, disk
and I/O power consumption, respectively. Moreover, each component
has its own power consumption model. For example, the power con-
sumption of the CPU is a function of voltage, frequency and utilization,
the power consumption of memory is a function of its size, frequency,
Cloud/Internet Portal Data Center
and usage, the power consumption of the hard disk is a function of read
rate and write rate, and the power consumption of the fan is a function
of width, depth, and rotation speed. However, the main energy-con-
Communications/Telco
Visualization & Audio
suming components can be slightly different owning to the configura-

Business Processing
tions of servers, and the proportion of power consumption of each

Workload Type
component is not constant, as shown in Fig. 7 [35]. Some research

Scientific
Analytics
[36–38] selected the CPU, memory, hard disk, network, mainboard,

Storage
Table 2
and fan as the main components of the power consumption model,

while other studies considered the CPU and memory to be the main
6
Fig. 2. The number of samples of each year.
Table 3
The number of each height of the servers for 641 samples.
Height 1U 2U 3U 4U 5U 7U 10U 12U Blade Others
Number 133 212 2 28 2 5 4 2 59 194
Fig. 5. The workload of servers for 1, 2, 4, 6 and 8 CPUs.
3.2.2. BA models
In data centers, the servers are not always active, as servers can be
idle. Then, the power consumption of the counterparts can be divided
Fig. 3. The rated power of the servers. into two parts. (1) Baseline power (Pbase), which is the power con-
sumption when the machine is idle, includes the power consumption of
the fans, CPU, memory, I/O and other motherboard components in
their idle state and is often considered as a fixed value. (2) Active power
(Pactive), which is the power consumption due to the workload, depends
on the workload of the machine and the way it utilizes CPU, memory
and I/O components. Hence, the power model can be expressed as the
sum of baseline power and active power [43–45]. The formula is as
follows:
PBA = Pbase + Pactive + Delta (2)
where, Delta is the correction term, which can be either a fixed value or
an expression. In addition, the formula can be transformed. Pbase can be
regarded as a constant term, and Pactive can be expressed as a function
(e.g., linear-function, power function, high-degree polynomial). Hence,
the BA model can be classified as linear regression models, power
function models, non-linear model and polynomial model. In particular,
Fig. 4. The relationship among rated power, peak power, and idle power.
from the perspective of heat generation and heat dissipation, the power
consumption of the server includes the IT components and cooling
components. Thus, the server power model can be transformed as their
components, with the other components having a very narrow range or
sum [46,47].
correlating well with CPU activity. Thus, the energy consumed by the
server is a function of CPU, memory, and some other devices [39–42].
3.2.2.1. Linear regression model. Linear regression models include
7
Table 4
The power models and their factors and benchmark/testing tool.
C. Jin, et al.
Category Authors Power model Factors Benchmark/Testing tool Error
Additive model Basmadjian l CPU voltage and Lookbusy benchmark 2% for best case; < 10%
PServer = i = 1 PMainboard ; (for server type )
et al. [36] l m n
frequency, memory for worst case
PServer = i = 1 PMainboard + j = 1 PFan + k = 1 PPSU ; (for tower or rack type ) frequency, hard disk
read and write rate,
and fan rotate speed
Perumal and Enode = ECPU + Ememory + Edisk + ENIC + Emainboard CPU frequency, Eucalyptus
Subbiah [37] voltages and
percentage of active
gates
Song et al. Etotal = ECPU + Ememory + Edisk + ENIC + Enetwork devices Average component MPI shift benchmark < 5%, max; < 3.5%,
[38] power average
Roy et al. [39] E = ECPU + EMemory
Tudor and Teo E = ECPU + EMemory + EI / O Clock frequency, NPB 3.3, memslap 0.44, and PARSEC ~9%
[40] service time of one 2.1 benchmarks
memory request, I/O
service time
Ge et al. [41] Pnode = Pcpu + Pmemory + Pother L1 cache access, L2 NAS EP and FT benchmarks
cache access, main
memory access,
retired operations,
and retired floating-
point operations
Arroba et al. Ptot (m, k ) = Pcpu (m , k ) + Pmem (m, k ) + Pothers (m , k ) Lookbusy, modified RandMem, Web
[42] Search application from the
CloudSuite benchmark suite, SPEC
8
CPU2006 mcf and SPEC CPU2006
perlbench
BA model Dhiman et al. P = Pbase + Pactive CPU utilization Four benchmarks from the SPEC2000 < 10%
[43] suite, namely: mcf, gcc, mesa and gap
Xiao et al. M vm The utilization of each 401.bzip2, 429.mcf, TPC-W, < 5.2%
Pserver = Pstatic + i = 1 Pi = Pstatic + j J (kj Uj )
[44] kind of physical Cachebench, and IOZone benchmark
components
Chen et al. P = Pfix + Pvar Throughput IOzone benchmark, Fibonacci
[45] sequences, JPetstore, and JMeter
Ham et al. 5 2 CPU utilization and
Pserver = PIT + Psfan; PIT = 1.566 × 10 + 42.29ucpu + 0.379Tdie + 0.03002Tdie ;
[46] die temperature, and
Psfan = 0.0012RPM 12 × 10 8RPM2 + 28 × 10 2RPM3
fan revolution speed
Garraghan m n CPU utilization, die SPECPower ssj2008 benchmark 3%
Pserver = PIT + 1 mPpump + 1 nPfan
et al. [47] temperature and fan
speed
Simple Fan et al. [49] P(u) = Pidle + (Pbusy Pidle ) u CPU utilization
regression
model
Kavanagh and EU _Utilx CPU utilization Stress
EU_Px = Host _Idle + (Host _P Host _Idle ) × EU _Count EU _Util
Djemame [50] 1 y
Islam and ConnS [t 1, t 2] Server utilization
PS [t1, t2] = Pidles + (PMaxs Pidles )
Pierson [51] ConnMaxs
Jin et al. [53] Pn = Pn0 + (Pmax Pn0) × rn/rn ; Goodput π-based CPU/RAM intensive
application, HTTP request benchmark
Pv = Pv0 + (Pmax Pv0) × rv /rv
Gupta et al. Throughput and Standard M/M/1 queuing model
P( ) = × (PCPU + POther ) + 1 Idle
[54] µ ( µ )×P speedup
Lefurgy et al. P (k ) = At(k ) + B; The performance state SPEC CPU2000 and the P4MAX
[55] P (k + 1) = P(k ) + Ad (k ) of the processors workload

Table 4 (continued)

C. Jin, et al.
Beloglazov P(u) = k Pmax + (1 k ) Pmax u CPU utilization Non Power-Aware and Single
et al. [56] Threshold policy
Rezaei- Psrv = (Ppeak Pidle )(Tsrv Tin) + Pidle Server temperature
Mayahi et al. and inlet air
[57] temperature
Wang et al. Pestimated = Pidle + UCPU % (P100% Pidle ) + Delta (T ); CPU utilization, inlet Server Efficiency Rating Tool 3.2%–4.6% (Mean
[58] Delta(T ) = a0 + a1 T + a2 T2 air temperature absolute percentage error)
Multiple Economou Pblade = 14.45 + 0.236 × ucpu (4.47E 8) × umem + 0.00281 × udisk + (3.1E 8) × unet ; Utilization of CPU, SPECcpu2000 integer and floating- < 5%
regression et al. [59] Pitanium = 635.62 + 0.1108 × ucpu + (4.05E 7) × umem + 0.00405 × udisk + 0 × unet memory, disk, and point benchmarks, SPECjbb2000,
model network SPECweb2005, the streams benchmark,
and matrix multiplication.
Alan et al. Pt = Ccpu, n × ucpu, t + Cmemory × umemory, t + Cdisk × udisk ,t + Cnic × unic, t CPU utilization, Scp, rsync, ftp, bbcp, and gridftp < 6% for fine-grained
[60] memory access rate, model
hard disk I/O request
rate, and network I/O
request rate
Li et al. [61] n n n
Pserver = k = 1 UCPU (k ) + k = 1 UMEM (k ) + k = 1 UIO (k ) + n e + Ebaseline CPU utilization, Hadoop including PI, Sudoku, < 4%
memory usage, and IO Random, Writer, Sort, Word, and
throughput Count
Lent [62] N 1 C 1 D 1 C 1 Utilization of core, HTTPerf < 2.5 W
P=I+ i=0 N N (i ) + j =0 C C (j ) + k = 0 D D (k ) + m ( j = 0 C (j ) )+ M
disk or port subsystem
C 1
( j = 0 C (j ) )
Kansal et al. Esys = cpu ucpu + mem umen + io udisk + CPU utilization, the SPEC CPU 2006 suite and IOmeter < 5%
[63] number of LLC misses, benchmarks
and the number of
9
bytes read and written
Costa and I J Performance variables SPEC Power
P = P0 + i = 1 i Yi + j = 1 j Sj
Hlavacs [64]
Witkowski y= 0 + 1 f1 (x1) + 2 f2 (x2 ) + + n fn (xn ) Performance and CPU 1%–4%
et al. [65] temperature
Bohra and P{CPU , cache} = a1 + a2 pCPU + a3 pcache ; Performance counter NAS-NPB 2.4, Iozone, Bonnie++, 7% of average mean
Chaudhary P{DRAM , disk } = a4 + a5 pDRAM + a6 pdisk ; values for CPU, cache BYTEmark, Cachebench, Dense matrix error;6% of median error
[66] P(total) = P{CPU ,cache} + P{DRAM , disk } and DRAM multiplication, Gcc
Lewis et al. Esystem = 0 (Eproc + Emem) + 1 Eem + 2 Eboard + 3 Ehdd Ambient temperature, Eight benchmarks from the SPEC < 4%
[67] CPU die temperature, CPU2006 benchmarksuite: bzip2,
system bus, and cactusadm, gromac, lbm, leslie3d, mcf,
misses in the L2 cache omnetpp, and perlbench
Power function Chen et al. P = Pfixed + Pf × f 3 CPU frequency Real HTTP traces
model [68]
Elnozahy et al. P = c 0 + c1 × f 3 CPU frequency Olympics98 and Finance
[69]
Tian et al. i Server utilization and
Pi = i ki µi + Pi
[70] service rate
Yao et al. [71] bi (t ) Server utilization MapReduce jobs
P= + Pidle
A
Wu et al. [72] Pvm (up) = up ; CPU utilization SysBench, Nbench, 4.26% for VMs;0.88% for
Fhourstone + Dhrystone, Ubench host server
n
Pvm (uv , n) = u
() N v
Lien and Bai P = Pbase + (Pmax Pbase ) U CPU utilization Windows Media Load Simulator < 3%, hardware-revised
[73] method; < 6%, software-
revised method;
~11%, filled-manually
method
Table 4 (continued)

C. Jin, et al.
Non-linear Fan et al. [49] P(u) = Pidle + (Pbusy Pidle ) (2u ur ) CPU utilization
model
Polynomial Zhang et al. Ptotal = a + b × R cpu ; CPU utilization SPECPower_ssj2008 benchmark < 4%;
model [74] 2
Ptotal = a + b × R cpu + c × Rcpu ;
2 3
Ptotal = a + b × R cpu + c × Rcpu + d × R cpu
Lin et al. [75] P N Server utilization SPECpower_ssj2008 2.794% for power
E=N× = ;
s EE function model; 2.974%
s 2
c0 + c1 user + c 2 user
= = for quadratic model;
EE (user ) P 2
d0 + d1 user + d2 user 4.506% for linear model;
1.615% for polynomial
model; 5.918% for linear
model of fixed coefficient;
8.222% for power
function model of fixed
coefficient
Other model Horvath and Pi (fi , Ui) = ai3 fi Ui + ai2 fi + ai1 Ui + ai0 Clock frequency and 1%- average error; <
Skadron [76] CPU utilization 4%-worst case
Xu et al. [77] n n CPU utilization, clock SPEC2006 and PARSEC (mcf, bzip2 and 6.7%
powertotal = i=1 (powercpu,i ) + i=1 (powerdisk ,i ) + poweridle
frequency, and disk canneal are chosen to represent user
utilization tasks, and freqmine are chosen to
represent the tasks run by cloud
computing providers.)
Bai et al. [78] i )
Pi = (a1 + ui a2 )(b1 + b 2 Tchip Server utilization and
chip temperature
Enokido and NFEt if NCt ( ) 1
10
Takizawa [79] Et ( ) = ;
minEt otherwise
maxEt if NCt ( ) Mt
Et ( ) = t NCt ( ) + NFEt if 1 NCt ( ) Mt
minEt otherwise
Li et al. [80] Psys Pidle CPU utilization SAP ERP performance model
Pnorm = ; h (U ) = c1 U c2 + c3 U c4 + c5;
Pbusy Pidle
Pnorm (U ) = 1 h (U ) 1
Lim et al. [81] E = Edynamic + Eidle + Esleep = P(s, 1.0) mr /s The average number RUBiS and TPC-W benchmarks < 10%
of instructions, the
predicted number of
requests, CPU
utilization, CPU speed
Kataoka et al. nct 1 Nanosleep, gettimeodday, and taskset
Et ( ) = minEt + t( ) [bCt + i=0 ti ( ) {(cEt + ti ( ) tEt )}]
[82]
Enokido and PCt (trt ( )) = t trt ( ) + minEt + {(act ( ) cEt ) + (nvt ( ) vEt ) + (mvt ( ) mEt )}
Takizawa [83]
individual utilization. In contrast, Kansal et al. regarded CPU utiliza-

tion, the number of LLC misses, and the number of bytes read and
written as the training model parameters [63]. Costa and Hlavacs [64]
described server power consumption based on performance counters
(i.e., perf_cycles, perf_cache.references, perf_cache.misses). Further-
more, Witkowski et al. [65] added the CPU temperature in the model.
VMeter is a power modeling technique based on the online monitoring
of system resources and can predict the instantaneous power with an
average mean and median accuracy of 93% and 94%, respectively [66].
Unlike these models, Lewis et al. [67] combined the ambient tem-
perature, CPU die temperature, system bus, misses in the L2 cache to-
gether to regress the power model, and the accuracy is within 4%.
These models tend to predict power consumption well as long as the
Fig. 6. The power consumption composition of a server [7] characteristics of workload do not change.
simple regression models and multiple regression models. Bellosa [48] 3.2.2.2. Power function model. For some models, CPU is the main
was the first to propose power models based on the correlation between component of the energy consumption model. It operates at a
power consumption and performance counters that captured activity frequency f and its dynamic power consumption is proportional to
across various functional units of the CPU. Fan et al. [49] proposed V2 × f, where V is the operating voltage; the frequency will be reduced
simple linear regression power models based on just CPU utilization to when the voltage is lowered. The power consumption of all other
account for the system power as a whole. The model was validated by components (except the CPU) is independent of the frequency, then, the
the experimental results. A similar model for VMs was also provided power consumption of the server can be expressed as a cubic
[50], and its formula has different independent variables for different relationship with the operating frequency [68,69]. Furthermore, in
application scenarios. For instance, regarding content distribution cloud computing, the server utilization and service rate were regarded
networks, the utilization ratio of the server during the time intervals as the main components of the model by considering the tradeoff
is the variable of the model [51]. Similarly, throughput has a strong between energy cost and performance [70]. However, some studies
correlation with power [52]. For server virtualization, the ratio of the used server utilization as the variable for characterizing the power
goodput to the maximum supportable goodput is regarded as the model. The exponential value is equal to 3 [71]. Additionally, there is a
independent variable [53]. Gupta et al. [54]assumed that the server power function relationship between the power consumption of servers
power consumption scales linearly with the ratio of the throughput to and CPU utilization [72,73].
the processing speed. For feedback control loop, the model is a function
of the performance state of the processors in the control period [55].
3.2.2.3. Non-linear model and polynomial model. In addition to simple
Furthermore, if the idle server consumes a constant ratio of the power
linear regression, Fan et al. [49] proposed a non-linear model to fit the
consumed by the server running at the full CPU speed, then the idle
server power consumption. This model is more accurate than the linear
power can be replaced by the maximum power multiplied by the ratio
model when estimating the total power usage of an individual machine.
[56]. Rezaei-Mayahi et al. [57] adopted the simple linear regression
However, as mentioned above, the simple linear regression models and
model by considering the difference of CPU temperature and inlet air
power function models are a poor fit for all but CPU-intensive servers
temperature to investigate the negative impact of ambient temperature
and resort to multiple-input quadratics to provide an average absolute
on power consumption. The authors assumed that the CPU temperature
error of 9%. Considering that, Zhang et al. [74] used high-degree
is linearly related to server utilization, and the relationship between
polynomial models to fit the server power consumption and found that
inlet temperature and power increment, which is reflected in the delta
the cubic polynomial can be the best choice. Similarly, Lin et al. [75]
in function (2), was analyzed [58].
provided the relationship between power consumption and the second-
Although simple regression models based on CPU utilization are
order polynomial of server utilization.
able to provide reasonable prediction accuracy for CPU-intensive
workloads, they tend to be considerably inaccurate for predicting
power consumption caused by I/O- and memory-intensive applications 3.2.3. Other models
[43]. Hence, to make the model more universal, multivariate linear Additive and BA models are the widely used power consumption
regression is used to build models between power consumption and the models of servers, and other models have also been developed. Horvath
metrics of multiple components. Economou et al. [59] and Alan et al. and Skadron [76] used the CPU utilization, frequency and product to
[60] provided a power model as a function of the utilization of CPU, describe the power consumption of individual active machines. Simi-
memory, disk, and network, and Li et al. [61] proposed a classified- larly, Xu et al. [77] built a model with CPU utilization, the product of
piecewise ternary linear regression model to achieve more accurate CPU utilization and frequency, and disk utilization. Bai et al. [78] used
predictive power. Lent [62] assumed that subsystems such as core, disk the product of the server utilization and chip temperature to represent
and port show linear power consumption with respect to their the power consumption. Enokido and Takizawa [79] created simple
power consumption for a server, where the electric power consumption
Fig. 7. Power breakdown across server components for different rack server configurations [35]
11
is maximally consumed if at least one application process is performed, polynomial model fits the power consumption of the servers best.
and otherwise, the electric power consumption is minimal. In addition
to the CPU and other devices, cooling devices such as CPU fans con-
4. Model application
sume the electronic power. Ham et al. [46] considered the power
consumption of servers to consist of IT and cooling power consumption.
With the development of information technology, the power density
IT power consumption is a function of CPU utilization and die tem-
and energy consumption of IT equipment are increasing. Meanwhile,
perature, and the fan power is related to the fan revolution speed.
the demands for cooling and the energy consumption of the cooling
Unlike the previously described approaches, Li et al. [80] developed a
system are also increasing. Therefore, thermal management and energy
power consumption model through the normalization of system power
management in data centers face enormous challenges. By using the
instead of through direct modeling. Specifically, power capping beha-
server as the basic unit of energy flow and heat flow, the power model
vior will occur when CPU utilization is higher than 80%. This behavior
can be applied for energy prediction and management to reduce the
can be fitted well by a power function by using CPU power over the
amount of energy required by the server and cooling system.
measured time [81]. The relationships among the energy consumed for
processing requests, the idle period, and the sleep period were estab-
lished. Moreover, power consumption models considering multi-core 4.1. Energy management of servers and server clusters
CPU [82] and the number of active VMs [83] were proposed, and the
artificial neural network method was popularly used in power con- According to [85], saving 1 Watt of power at the CPU can easily turn
sumption models for cloud data centers [84]. into 1.5 W of savings due to power delivery efficiency losses inside the
server and up to 3 W in the data center. Hence, considering the data
center energy costs, reducing the energy consumption of servers is a
3.3. Comparison of power consumption models of servers major challenge. According to the review of the literature, from the last
decade, the overall energy consumption of the IT equipment can be
Fig. 8 depicts the statistical result of the main variables in the ex- reduced by reducing the energy consumption of a single device or
isting power models. It can be seen that CPU utilization is the most distributing the workload to the server clusters. To reduce the energy
widely used variable in the power consumption models, followed by used by servers, a higher inlet air temperature will result in higher
CPU frequency, performance counters, and temperature. Different leakage power and fan power, and it is necessary to comprehensively
model components achieve different prediction results, and the accu- evaluate this relationship. Some researchers have provided control
racy will affect the prediction of the energy conservation and the choice policies based on the server power model by considering the optimal
of control strategy. Some studies have validated the accuracy of the thermal environment and obtained the energy savings. Wang et al. [86]
models, and the errors are presented in Table 4. According to the sta- proposed a model-based approach that uses the power and temperature
tistics, the errors of additive models and BA models are < 9% and 11%, models, to create a multiple-input multiple-output fan controller and
respectively. For additive models, the errors of most studies were < 5%, then handled the power optimization problem of blade servers. Com-
and the worst error rate was 10% in the worst case. For BA models, the pared with the feedback controller, this method can reduce fan power
errors of simple linear regression models, multivariate linear regression usage by approximately 20% and guarantee most temperatures be-
models, power function models, and polynomial models were gen- tween 64 °C and 66 °C. However, it cannot consider the influence of the
erally < 4.6%, 7%, 6%, and 4%, respectively. Specifically, the accuracy temperature on the leakage power. Shin et al. [87] comprehensively
of a linear regression model will be improved 13.6% by considering the considered the fan and leakage power. The results showed that tracking
inlet air temperature at 45 °C [58]. To analyze the accuracy of the the energy-optimal steady-state temperature can reduce the overall
models under the same conditions, Zhang et al. [74] compared the energy by up to 17.6%. Moreover, Huang et al. [88] adjusted the server
linear, quadratic, and cubic models and found that the cubic model can thermal setpoint and allowed the system to heat up when doing so saves
fit the measured data well. Lin et al. [75] analyzed the errors among six more fan power than it costs in terms of leakage power. This thermal-
power models and concluded that the polynomial model has the lowest aware power optimization will reduce the total power by 5.2% by in-
error (1.615%), followed by the power function model (2.794%) and creasing the inlet air temperature by 3 °C, and the thermal setpoint
the quadratic model (2.974%). However, if the coefficient is fixed, then converged to 78 °C. Additionally, Zapater et al. [89,90] analyzed the
the errors of the linear model and power function model can increase to relationships among computational power, temperature, leakage, and
5.918% and 8.222%, respectively. Hence, it can be concluded that the cooling power and proposed a cooling management policy that
Fig. 8. The statistical result of the main variables in power models.
12
minimizes the server energy consumption by setting the optimal fan there have been some attempts for the power/energy management of
speed during operation. The energy savings of the whole cluster in the server or server clusters, they are restricted to certain conditions.
terms of leakage plus fan power will increase with the increase in With the continuous improvement of energy-saving requirements, re-
ambient temperature. The maximum savings are 10.3% at 32 °C. fined control/management technology based on the power consump-
Additionally, the server power model can be applied during server tion model has become an important direction for future research.
operation to guide server power mode (i.e., on/off), workload alloca-
tion and virtualization. Servers generally consume more than 50% of 4.2. Cooling load and energy efficiency of cooling system
the peak power when they are idle [43]. However, the word “idle” can
mean several things in practice: (1) the server must almost instantly The matching of cooling supply and demand is key to saving energy
react to a request for new work; (2) the server could pause for a brief of cooling system. According to the power consumption models of
period; (3) the server has plenty of time to wake up [91]. Although servers, the cooling capacity can be calculated accurately to ensure that
much energy can be saved by turning idle servers off, turn a server on IT equipment works in the required environment and avoid wasting the
requires some energy. The compromise is to put the idle server in sleep cooling capacity during design and operation. In the design of the
mode, which can save approximately 140 W of energy [92]. Virtuali- cooling system, the cooling capacity depends on the cooling require-
zation technology allows servers, storage and networking devices to be ments of data center rooms, and the total cooling load can be calculated
shared, which improves resource utilization. In virtualization, appli- by considering various parameters, such as solar radiation through
cations can be easily migrated from one physical server to another [37]. transparent surfaces, heat conduction through the building envelope,
Then, the managers can distribute the workload to the specified servers, heat generated in the space by lights, occupants, and IT equipment, and
and the others can be powered off or put in sleep mode by estimating heat transfer through air infiltration [97]. Among them, the heat dis-
the task needs. In this process, a suitable VM power model can help data sipated by the IT equipment occupies the main part of the cooling load.
center operators save power and price the VM energy consumption of The designers often take the nameplate power of IT equipment as its
cloud computing platforms [93]. Noguchi and Nishi compared the ca- heat dissipation [98]. However, this estimation is seldom true in
pacity of the servers and the summation capacity of the incoming re- practice. As depicted in Fig. 3 and Fig. 4, the mean full-utilization
quests to determine whether to start the servers [94]. Chen et al. [95] power accounts for only approximately 50% of the nameplate power
proposed server provisioning and load dispatching algorithms, which that has hardly changed recent years. Additionally, the servers do not
can reduce energy consumption by 30% without sacrificing user ex- operate at peak power all the time, and the average rack was only one-
periences. Based on a unified energy model and mechanisms for energy- third filled. Hence, the thermal load will be significantly overestimated
aware resource accounting and allocation, a novel framework has been in the data center, and the cooling equipment are designed to be
proposed for energy optimization [96]. Additionally, a joulemeter [63] oversized according to the above design method [99]. At the same time,
is used to monitor actual power usage and to help determine how the the power system has too much redundancy. In [100], the authors used
migration of each VM will affect power level. This method can be used the simple linear regression model [49] to model the power consump-
to realize significant power provisioning cost savings in virtualized data tion of the data center, and they estimated the server’s maximum and
centers. Reference [45] further identified the relationship between idle power consumption by linear regression using 491 data points from
energy consumption and runtime tasks under different cloud system various manufactures submitted to the SPECpower2008 database be-
configures and the correlation with cloud system performance. Elno- tween 2007 and 2017. Giang et al. [101] and Pelley et al. [102] em-
zahy et al. [69] discussed the relationship between power consumption ployed the linear model to investigate the total energy consumption of
and CPU frequency and compared the energy efficiencies of five po- data centers. However, the power draw is not precisely linear during
licies: independent voltage scaling, coordinated voltage scaling, vary-on operation, and if possible, a more accurate representation of server
vary-off, combined policy, and coordinated policy for server clusters. It utilization should be employed.
can be found that the policy that saves the most energy is the co- In addition, the models can be applied for cooling control or airflow
ordinated policy, which achieves 33% and 50% of the cluster energy. management to realize energy savings. Zapater et al. [103] proposed
Furthermore, Dhiman et al. presented a system for online power pre- the cooling control policy that set per-rack inlet temperature and air-
diction using the Gaussian mixture model, and the average prediction flow to minimize data center power based on the BA model and com-
error can be less than 10% [43]. Moreover, Lefurgy et al. proposed a bined cooling and workload management. The results showed that this
feedback controller that uses precise, system-level power measurement policy saves energy during winter (1.2%) and more energy in summer
to periodically select the state with the highest performance while (14.4%). Noguchi and Nishi [94] presented an active controller shutter
keeping the system within a fixed power constraint [55]. Although to prevent cold air from leaking through idle or shut-down servers when
Fig. 9. Server power and airflow rate increase versus ambient temperature for Class A2 [104]
13
the server temperature is less than the threshold value (40 °C). Ham Table 5
et al. [46] applied the server model to simulate the hourly cooling Base idle state level power allowance [105]
energy consumption by considering thermal characteristics. It can be Category Maximum possible Managed Base idle state power
found that cooling energy consumption will be increased when the number of installed server allowance, PBASE (Watts)
supply air temperature is higher than 19 °C, owning to the increase in processors (#P)
fan energy consumption.
A 1 No 47.0
B 1 Yes 57.0
5. Discussion and further work C 2 No 92.0
D 2 Yes 142.0
Resilient 2 Yes 205.0
5.1. Model
Note: A managed server must meet all of the following criteria: 1) is designed to
The physical architecture and configuration of the server differ very be configured with redundant power supplies; and 2) contains an installed
much among manufactures, and become more complicated with the dedicated management controller. The Resilient category applies only to two
innovation of technologies. The difference between peak power and socket systems that meet the definition of Resilient Servers as set forth in
idle power of the servers is not constant from generation to generation, Appendix B of [105].
as discussed in Section 3.1.2. At the same time, different workload types
presented in Table 2 make the CPU, memory, I/O, and storage at dif- utilization levels, which will cause the energy consumption of servers to
ferent usage levels, which result in the power models less portable. be segmented. Then, the future models should be piecewise functions.
Therefore, whether existing power consumption models accommodate According to the statistics of the power consumption models of the
these changes must be determined. Besides, the power increase of a servers, the simple linear regression is the most popular model, and the
server is a result of fan power, component power, and the power con- peak power and idle power are important. However, a method for easily
version of each. The component power increase is a result of an increase acquiring these data is needed. For the peak power model, the power
in the leakage current for some silicon devices. For Class A2, if the consumption is hypothesized to be strongly correlated with thermal
server inlet air temperature is raised from 15 °C to 35 °C, the server design power (TDP), as the variation trend of TDP is the same as that of
power would increase in the range from 7% to 20%, as shown in Fig. 9 the peak power of 1U and 2U servers, as shown in Fig. 10. Energy Star
[104]. At the same time, the required air volume is 1.2–2.5 times that at divided the idle state power into two parts: base idle power and addi-
15 °C, which can increase the power consumption of the server fan. This tional power, and the idle power limits are applicable to one and two
power increase may be ignored for server power consumption but not socket systems only, which not include the blade and multi-node sys-
for the data centers that house tens of thousands of servers. Few studies tems. The base idle power can be divided into five categories (A-D and
[57,58] consider the inlet air temperature variation due to the fluctu- Resilient) by considering the number of processors, redundant power
ating environment in the power consumption model of the server. In supplies, and dedicated management controller. The additional idle
other words, most existing power consumption models are ineffective power mainly considers additional components, such as power supplies,
when the inlet air temperature of the server is changing. Hence, future hard drives, memory, buffered DDR channel and I/O devices [105]. The
models should reflect the coupling relationship between the power relative values are shown in Table 5 and Table 6.
consumption and the thermal environment. Additionally, the power
management of a server is a factor that influences model establishment. 5.2. Application
For instance, demand-based power management of Dell servers per-
forms best in terms of energy-savings at mid-to-low processor A. Current problems for the applications of power consumption models
Fig. 10. The comparison among peak power, rated power and TDP for 1U and 2U servers.
14
Table 6
Additional idle power allowance [105]
System characteristic Applies to: Additional idle power allowance
Additional power supplies Power supplies installed explicitly for power redundancy 20 W per power supply
Hard drives (including solid state drives) Per installed hard drive 8.0 W per hard drive
Additional memory Installed memory greater than 4 GB 0.75 W per GB
Additional buffered DDR channel Installed buffered DDR channels greater than 8 channels (Resilient Servers only) 4.0 W per buffered DDR channel
Additional I/O devices Installed devices greater than two ports of ≥1 Gbit, onboard Ethernet < 1 Gbit: No Allowance
= 1 Gbit:2.0 W/Active Port
> 1 Gbit and < 10 Gbit: 4.0 W/Active Port
≥10 Gbit: 8.0 W/Active Port
Regarding the application of power consumption models, accuracy simplification is attributed to the lack of comprehensive and accurate
and simplicity are the two main requirements, but they are contra- power consumption models of servers, which still limits the diversity of
dictory and restricted. If the model needs to be accurate, then the power the experimental conditions. Therefore, the model-based server simu-
consumption of each component of the server needs to be considered. In lator needs to be developed.
contrast, if the formula is simple and only considers the power con- In addition, with the continuous improvement of equipment effi-
sumption of the main components of the server, then it may not be so ciency, the energy savings of equipment will be maximized, and the
accurate. For example, additive models can explain the components of energy conservation potential will be reflected in operational control.
power consumption according to the configuration of the servers. Regarding thermal management and energy conservation, computer
However, this kind of power model needs to calculate the power con- room air conditioner (CRAC) power leads to cooler server inlet ambient
sumption of each component of the servers, which is complicate, time- air temperature, which makes the server cooling fans work less hard
consuming, and hardly possible for heterogeneous systems. and thus reduces server fan power. Similarly, lower CRAC power leads
Furthermore, the simple linear regression is a popular model for pre- to warmer ambient air temperature, which causes server fans to work
dicting energy consumption. However, it is usually suitable for the harder and consume more IT power [88]. Hence, coupling the server
CPU-dominated servers, especially under moderate utilization and fan and CRAC fan is key to the energy efficiency. The power prediction
constant power consumption. Once the systems become less CPU- models provide a powerful criterion for controlling the operation of the
dominated due to the demand of faster and larger memory and storage fans.
systems, the simple regression model introduces large prediction errors
to these systems [106]. The multi-variable model facilitates monitoring 5.3. Further work
and the collection of parameters at the server level, but it may not
suitable for the system or room level depending on server type and Based on the aforementioned references, the power consumption
workload. The regression coefficient should be adjusted when the models of the servers play a crucial role in the power/energy man-
number of servers changes. Otherwise, the accuracy of the model will agement and thermal management of the data center. However, some
be poor. Therefore, when using existing models for predicting or eval- models are proposed and applied to estimate the potential of power
uating the power/energy consumption of servers or data centers, the management to reduce energy consumption, and further research on
applicability of them needs to be investigated, and the accuracy needs the models and their applications is necessary. The adaptability of the
to be verified. The application scenarios, advantages and disadvantages existing models requires further verification, and the piecewise power
of these models have presented in Table 7. consumption model of servers needs to be investigated according to the
server power trends, server configuration, energy-saving technologies,
B. Power consumption model applications and inlet air temperature. Model-based server simulators that present
the operational characteristics of servers have potential for the ex-
As mentioned in Section 4.2, the existing cooling load calculation in perimental study of the data center, which is important for the devel-
the data center still follows the traditional calculation method. Ac- opment of highly efficient equipment and the implementation of en-
cording to the working characteristics and server configurations, the ergy-saving strategies for data centers. For data centers, thermal safety,
servers hardly consume the rated power; hence, if the value of name- which is the premise of energy savings, is extremely important. Hence,
plate power is regarded as that of heat dissipation, then the equipment regarding the energy savings, energy-aware and thermal-aware man-
will be designed to be oversized, resulting in low equipment efficiency agement based on the power consumption models of servers should be
and significant power and cooling system redundancies, leading to the considered at the same time. Thus, the combination of thermal man-
increase in the initial equipment investment. Due to this redundancy, agement and energy management and the relationship between CRAC
the energy consumed by the cooling system can be offset by the energy fans and server fans should be studied in the future.
savings of the servers. Therefore, the power consumption models of
servers used to estimate the cold load is of great significance to the 6. Conclusion
cooling design and energy saving of the data centers.
Furthermore, considering the safety of servers and data, many ex- The power consumption model of servers plays a critical role in the
periments have been carried out in the laboratory instead of in actual thermal management and energy management of data centers. This
data centers. Hence, the server simulators should determine the heat paper reviewed server classification and power consumption trends,
dissipation characteristics based on the power consumption model of then summarized and classified the existing power consumption models
the servers. Wang et al. [107] and Nelson [108] have used heating of servers. A comparison of these models has been carried out.
elements instead of actual servers to investigate the reliability of the Additionally, the application scenarios of these models were in-
power system and the thermal environment of data centers, which are vestigated, and the outlook for model development to increase energy
similar to therml manikins that an in-line electric heat source replaces savings was discussed. Based on the presented overview, the following
the real person, and the thermal conductivity of the material is basically conclusions can be drawn:
the same as that of human skin [109]. However, the heat dissipated by
these heating elements is always equal to the nameplate power. This (1) The rated power of the servers hardly increased over time, and the
15
peak power of servers accounts for approximately 40% to 50% of
Once the characteristic of components changed, the

dominated systems, and limit to partial utilization
the nameplate power. The idle power decreased from 55% to 15%
Various monitoring parameters and not suit for of the rated power due to energy-saving technologies. Furthermore,
heterogeneous systems or whole data center.
Perform large prediction error for less CPU-
regression coefficient need to be adjusted.

the processing performance of the server per Watt has been sig-
nificantly increased. These changes and the workloads of servers
need to be considered for the application of existing models and the
development of new models.
(2) The existing power consumption models can be classified as ad-
regions and server types.
ditive models, BA models, and other models according to their

calculation formula and other factors. Specifically, the BA models
can be further divided into linear regression models, power function
Disadvantages
models, non-linear models and polynomial models. CPU utilization

is the most common variable in the models, followed by CPU fre-
quency and performance counters. However, these models are for a
particular server under a fixed environment, and most of them are
suitable for single-core systems. Considering simplicity, con-
Few parameters need to be monitor, and suit for predicting energy
consumption of a server or small data centers, and the prediction
Only monitor one parameter and they are suit for prediction for
venience and accuracy, whether at the server, system or room level,

error is between simple linear regression and additive model.
linear regression and polynomial models can be the best choice, and
their errors are < 7% and 4%, respectively. Moreover, to establish a
model, energy-efficient technologies and the inlet air temperature
Precisely predict the power consumption of a server.
cooling load or energy consumption of data centers.
must be considered, and the piecewise function should be used.

(3) The current power consumption models are established to estimate
the potential of energy management, to predict the power con-
sumption of servers or data centers, and to balance energy con-
sumption and reliability. These models have been applied in some
schemes of the IT field and the refrigeration field, but more efforts
are needed. While the accurate power consumption models of ser-
vers result in accurate cooling load calculation and avoid excessive
redundancy, energy- and thermal-aware managements based on the
model results in the best overall energy-savings. Additionally, the
model-based server simulators improve the experimental study, as
Advantages
they closely represent the actual situation.
Declaration of Competing Interest

Artificial neural network method was popularly used in
The authors declare that they have no known competing financial

interests or personal relationships that could have appeared to influ-
power consumption models for cloud data centers
Power consumption estimation of CPU-dominated
servers, medium utilization systems, cooling load
ence the work reported in this paper.

calculation, and cloud computing management
Server power consumption and server power

Server power consumption and optimization
Acknowledgement
The study is supported by the National Natural Science Foundation

The application scenarios, advantages and disadvantages of the power models.
of China (51778080).
References
Application scenarios
[1] CAICT. China Internet Industry Development Trend and Prosperity Index Report.
management
2018.
[2] CAICT. White Paper on AI Development - Industrial Application. 2018.
[3] www.199it.com/archives/670459.html.
[4] Cooling DC. China data center annual research report on cooling technology de-
velopment 2018. China Architecture & Building Press; 2019.
[5] Arman Shehabi SS, Dale Sartor, Richard Brown, Magnus Herrlin United Stated
Simple linear regression model, power
function model, non-linear model and
Data Center Energy Usage Report. Ernest Orlando Lawrence Berkeley National
Laboratory; 2016.
Multiple linear regression model
[6] Dai J, Ohadi MM, Das D, Pecht MG. Optimum cooling of data centers. Springer;
2014.
[7] Vasques TL, Moura P, Almeida A. A review on energy efficiency and demand re-
sponse with focus on small and medium data centers. Energ Effic.
2019;12:1399–428.
polynomial model
[8] Lu HJ, Zhang ZB, Yang L. A review on airflow distribution and management in data
center. Energ Build 2018;179:264–77.
Power consumption model
[9] Alkharabsheh S, Fernandes J, Gebrehiwot B, Agonafer D, Ghose K, Ortega A, et al.

A brief overview of recent developments in thermal management in data centers. J
Electron Packag 2015;137.
[10] Chu WX, Wang CC. A review on airflow management in data centers. Appl Energ
Additive models
2019;240:84–119.
Other models
[11] Rambo J, Joshi Y. Modeling of data center airflow and heat transfer: State of the
BA models
art and future trends. Distributed Parallel Databases 2007;21:193–225.

Table 7
[12] Wang LZ, Khan SU. Review of performance metrics for green data centers: a tax-
onomy study. J Supercomput. 2013;63:639–56.
16
[13] Jin C, Bai X, Yang C. Effects of airflow on the thermal environment and energy [49] Fan XB, Weber WD, Barroso LA. Power Provisioning for a Warehouse-sized
efficiency in raised-floor data centers: A review. Sci Total Environ Computer. Isca'07: 34th Annual International Symposium on Computer
2019;695:133801. Architecture, Conference Proceedings. 2007. p. 13–23.
[14] Ge C, Sun ZL, Wang N. A survey of power-saving techniques on data centers and [50] Kavanagh R, Djemame K. Rapid and accurate energy models through calibration
content delivery networks. IEEE Commun Surv Tut 2013;15:1334–54. with IPMI and RAPL. Concurrency and Computation: Practice and Experience.
[15] Mittal S. A survey of techniques for improving energy efficiency in embedded 2019. p. 1–21.
computing systems. Int J Comput Aid Eng Technol 2014;6:440. [51] Islam Su, Pierson J-M. Evaluating energy consumption in CDN servers. ICT as Key
[16] Power Mittal S. Management techniques for data centers: a survey. Eprint Arxiv Technology against Global Warming. 2012.
2014. [52] Ghosh S, Chandrasekaran S, Chapman B. Statistical modeling of power/energy of
[17] Orgerie AC, De Assuncao MD, Lefevre L. A survey on techniques for improving the scientific kernels on a multi-GPU system. 2013 International Green Computing
energy efficiency of large-scale distributed systems. Acm Comput Surv. 2014;46. Conference (Igcc). 2013.
[18] Beloglazov A, Buyya R, Lee YC, Zomaya A. A taxonomy and survey of energy- [53] Jin YC, Wen YG, Chen QH, Zhu ZQ. An empirical investigation of the impact of
efficient data centers and cloud computing systems. Adv Comput. server virtualization on energy efficiency for green data center. Comput J
2011;82:47–111. 2013;56:977–90.
[19] Bostoen T, Mullender S, Berbers Y. Power-reduction techniques for data-center [54] Gupta V, Nathuji R, Schwan K. An analysis of power reduction in datacenters using
storage systems. Acm Comput Surv. 2013;45. heterogeneous chip multiprocessors. ACM SIGMETRICS Performance Evaluation
[20] Wang J, Feng L, Xue W. A review of energy efficiency technology in computer Review. 2011. p. 87.
servers and cluster systems. 3rd International Conference on Computer Research [55] Lefurgy C, Wang X, Ware M. Server-level power control. international conference
and Development (ICCRD). 2011. on autonomic computing. IEEE; 2007.
[21] Hammadi A, Mhamdi L. A survey on architectures and energy efficiency in data [56] Beloglazov A, Abawajy J, Buyya R. Energy-aware resource allocation heuristics for
center networks. Comput Commun. 2014;40:1–21. efficient management of data centers for Cloud computing. Future Gener Comp Sy.
[22] Daraghmeh HM, Wang CC. A review of current status of free cooling in data- 2012;28:755–68.
centers. Appl Therm Eng 2017;114:1224–39. [57] Rezaei-Mayahi M, Rezazad M, Sarbazi-Azad H. Temperature-aware power con-
[23] Ni JC, Bai XL. A review of air conditioning energy performance in data centers. sumption modeling in Hyperscale cloud data centers. Future Generat Comput Syst
Renew Sust Energ Rev 2017;67:625–40. Int J Esci 2019;94:130–9.
[24] Shuja J, Bilal K, Madani SA, Othman M, Ranjan R, Balaji P, et al. Survey of [58] Wang Y, Nörtershäuser D, Masson SL, Menaud J-M. An Empirical Study of Power
techniques and architectures for designing energy-efficient data centers. Ieee Syst J Characterization Approaches for Servers. ENERGY 2019 - The Ninth International
2016;10:507–19. Conference on Smart Grids, Green Communications and IT Energy-aware
[25] Bhattacherjee S, Khatua S, Roy S. A review on energy efficient resource manage- Technologies. Athens, Greece2019. p. 1-6.
ment strategies for cloud. Adv Intell Syst. 2017;568:3–15. [59] Economou D, Rivoire S, Kozyrakis C, Ranganathan P. Full-System Power Analysis
[26] Atiewi S, Yussof S, Ezanee M, Almiani M. A Review Energy-Efficient Task and Modeling for Server Environments. Workshop on Modeling Benchmarking and
Scheduling Algorithms in Cloud Computing. 2016 Ieee Long Island Systems, Simulation. 2006.
Applications and Technology Conference (Lisat). 2016. [60] Alan I, Arslan E, Kosar T. Energy-aware data transfer tuning. Ieee Acm Int Symp.
[27] Reda S, Nowroz AN. Power modeling and characterization of computing devices: a 2014:626–34.
survey. Found Trends Electronic Des Automat 2012;6:121–216. [61] Li YF, Wang Y, Yin B, Guan L. An online power metering model for cloud en-
[28] Zhang WW, Wen YG, Wong YW, Toh KC, Chen CH. Towards joint optimization vironment. 11th IEEE International Symposium on Network Computing and
over ICT and cooling systems in data centre: a survey. IEEE Commun Surv Tut. Applications (Nca). 2012. 2012. p. 175–80.
2016;18:1596–616. [62] Lent R. A model for network server performance and power consumption. Sustain
[29] Mobius C, Dargie W, Schill A. Power consumption estimation models for pro- Comput Informat Syst 2013;3:80–93.
cessors, virtual machines, and servers. IEEE T Parall Distr. 2014;25:1600–14. [63] Kansal A, Zhao F, Liu J, Kothari N, Bhattacharya AA. Virtual machine power
[30] Dayarathna M, Wen YG, Fan R. Data center energy consumption modeling: a metering and provisioning. Proc of the 1st ACM symposium on Cloud Computing
survey. Ieee Commun Surv Tut. 2016;18:732–94. (SOCC'10). 2010.
[31] ASHRAE. IT Equipment Design Impact on Data Center Solutions: W. Stephen [64] Costa GD, Hlavacs H. Methodology of measurement for energy consumption of
Comstock; 2016. applications. IEEE/ACM International Conference on Grid Computing: IEEE. 2010.
[32] https://en.wiktionary.org/wiki/server. [65] Witkowski M, Oleksiak A, Piontek T, Weglarz J. Practical power consumption
[33] ASHRAE. IT Equipment Power Trends: W. Stephen Comstock; 2018. estimation for real life HPC applications. Future Generat Comput Syst Int J Esci
[34] https://www.spec.org/. 2013;29:208–17.
[35] Malladi KT, Nothaft FA, Periyathambi K, Lee BC, Kozyrakis C, Horowitz M. [66] Bohra AEH. Chaudhary V. VMeter: Power modelling for virtualized clouds. IEEE
Towards Energy-Proportional Datacenter Memory with Mobile DRAM. 2012 39th International Symposium on Parallel & Distributed Processing: IEEE. 2010. p. 1–8.
Annual International Symposium on Computer Architecture (Isca). 2012:37-48. [67] Lewis A, Ghosh S, Tzeng N-F. Run-time energy consumption estimation based on
[36] Basmadjian R, Ali N, Niedermeier F, Meer Hd, Giuliani G. A methodology to workload in server systems. Workshop on Power Aware Computing & Systems.
predict the power consumption of servers in data centres. ACM Sigcomm 2008.
International Conference on Energy-efficient Computing & Networking. 2011. [68] Chen Y, Das A, Qin W, Sivasubramaniam A, Wang Q, Gautam N. Managing server
[37] Perumal V, Subbiah S. Power-conservative server consolidation based resource energy and operational costs in hosting centers. ACM SIGMETRICS Performance
management in cloud. Int J Netw Manag. 2014;24:415–32. Evaluation Review. 2005;33:303.
[38] Song SL, Barker K, Kerbyson D. Unified performance and power modeling of sci- [69] Elnozahy EN, Kistler M, Rajamony R. Energy-efficient server clusters. Power-
entific workloads. Proceedings of the 1st International Workshop on Energy Aware Computer Syst 2003;2325:179–96.
Efficient Supercomputing. 2013. [70] Tian Y, Lin C, Li KQ. Managing performance and power consumption tradeoff for
[39] Roy S, Rudra A, Verma A. An energy complexity model for algorithms. Proceedings multiple heterogeneous servers in cloud computing. Cluster Comput
of the 4th conference on Innovations in Theoretical Computer Science. 2013. 2014;17:943–55.
[40] Tudor BM, Teo YM. On understanding the energy consumption of ARM-based [71] Yao Y, Huang LB, Sharma A, Golubchik L, Data Neely M. Centers power reduction:
multicore servers. ACM Sigmetrics/international Conference on Measurement and a two time scale approach for delay tolerant workloads. IEEE Infocom Ser.
Modeling of Computer Systems. 2013. 2012:1431–9.
[41] Ge R, Feng XZ, Cameron KW. Modeling and evaluating energy-performance effi- [72] Wu W, Lin W, Peng Z. An intelligent power consumption model for virtual ma-
ciency of parallel processing on multicore based power aware systems. Int Parall chines under CPU-intensive workload in cloud environment. Soft Comput 2016.
Distrib, P. 2009. p. 1960. [73] Lien CH, Bai YW, Lin MB. Estimation by software for the power consumption of
[42] Arroba Patricia, Risco-Martín José L, Zapater Marina, Moya José M, Ayala José L, streaming-media servers. IEEE T Instrum Meas 2007;56:1859–70.
Olcoz Katzalin. Server power modeling for run-time energy optimization of cloud [74] Zhang X, Lu JJ, Qin X, Zhao XN. A high-level energy consumption model for
computing facilities. Energy Procedia 2014;62:401–10. https://doi.org/10.1016/j. heterogeneous data centers. Simul Model Pract Th 2013;39:41–55.
egypro.2014.12.402. [75] Lin WW, Wang WQ, Wu WT, Pang XW, Liu B, Zhang Y. A heuristic task scheduling
[43] Dhiman G, Mihic K, Rosing T. A system for online power prediction in virtualized algorithm based on server power efficiency model in cloud environments. Sustain
environments using gaussian mixture models. Des Aut Con 2010:807–12. Comput Informat Syst 2018;20:56–65.
[44] Xiao P, Hu ZG, Liu DB, Yan GF, Qu XL. Virtual machine power measuring tech- [76] Horvath T, Skadron K. Multi-mode Energy Management for Multi-tier Server
nique with bounded error in cloud environments. J Netw Comput Appl Clusters. Pact' 08: Proceedings of the Seventeenth International Conference on
2013;36:818–28. Parallel Architectures and Compilation Techniques. 2008. p. 270–9.
[45] Chen F, Grundy J, Yang Y, Schneider J-G, He Q. Experimental analysis of task- [77] Xu X, Teramoto K, Morales A, Huang HH. DUAL: reliability-aware power man-
based energy consumption in cloud computing systems. Proceedings of the ACM/ agement in data centers. proceedings of the 2013. 13th Ieee/Acm International
SPEC international conference on International conference on performance en- Symposium on Cluster, Cloud and Grid Computing (Ccgrid 2013). 2013. p.
gineering. 2013. 530–45.
[46] Ham SW, Kim MH, Choi BN, Jeong JW. Simplified server model to simulate data [78] Bai Y, Gu LJ, Qi X. Comparative study of energy performance between chip and
center cooling energy consumption. Energ Build 2015;86:328–39. inlet temperature-aware workload allocation in air-cooled data center. Energies
[47] Garraghan P, Al-Anii Y, Summers J, Thompson H, Kapur N, Djemame K. A Unified 2018;11.
model for holistic power usage in cloud datacenter servers. Int Conf Util Cloud [79] Enokido T, Takizawa M. An extended power consumption model for distributed
2016:11–9. applications. Int Con Adv Info Net. 2012:912–9.
[48] Bellosa F. The Benefits of Event–Driven Energy Accounting in Power-Sensitive [80] Li H, Casale G, Ellahi T. SLA-driven planning and optimization of enterprise ap-
Systems. 2000. plications. Proceedings of the first joint WOSP/SIPEW International Conference on
17
Performance Engineering. San Jose, California, USA2010. [94] Noguchi T, Nishi H. Active controlled shutter for effective cooling of servers in
[81] Lim SH, Sharma B, Tak BC, Das CR. A dynamic energy management scheme for data center. IEEE Ind Elec. 2015:1668–73.
multi-tier data centers. Int Sym Perform Anal. 2011:257–66. [95] Chen G, He W, Liu J, Nath S, Rigas L, Xiao L, et al. Energy-aware server provi-
[82] Kataoka H, Duolikun D, Enokido T, Power Takizawa M. Consumption and com- sioning and load dispatching for connectionintensive internet services. 5th USENIX
putation models of a server with a multi-core CPU and experiments. IEEE Symposium on Networked Systems Design & Implementation. San Francisco, CA.
International Conference on Advanced Information Networking and Applications 2008.
Workshops. 2015. [96] Stoess J, Lang C, Bellosa R. Energy management for hypervisor-based virtual
[83] Enokido T, Takizawa M. Power consumption model of a server to perform com- machines. Usenix Association Proceedings of the 2007 Usenix Annual Technical
munication type application processes on virtual machines. 2015 10th Conference. 2007. p. 1–14.
International Conference on Broadband and Wireless Computing, Communication [97] Hatamipour MS, Mahiyar H, Taheri M. Evaluation of existing cooling systems for
and Applications. 2015. reducing cooling power consumption. Energ Build 2007;39:105–12.
[84] Lin W, Wu G, Wang X, Li K. An artificial neural network approach to power [98] Rasmussen N. Calculating Total Cooling Requirements for Data Centers. 2017.
consumption model construction for servers in cloud data centers. IEEE Trans [99] Mitchell-Jackson J, Koomey JG, Nordman B, Blazek A. Data center power re-
Sustain Comput 2019. quirements: measurements from Silicon Valley. Energy 2003;28:837–50.
[85] Gough G, Steiner I, Saunders WA. Energy efficient servers-blueprints for data [100] Cheung H, Wang SW, Zhuang CQ, Gu JF. A simplified power consumption model
center optimization. Apress; 2015. of information technology (IT) equipment in data centers for energy system real-
[86] Wang ZK, Bash C, Tolia N, Marwah M, Zhu XY, Ranganathan P. Optimal Fan Speed time dynamic simulation. Appl Energ. 2018;222:329–42.
Control for Thermal Management of Servers. Ipack 2009: Proceedings of the Asme [101] Tran VG, Debusschere V, Bacha S. Data center energy consumption simulator from
Interpack Conference 2009, Vol 2. 2010:709-19. the servers to their cooling system. 2013 Ieee Grenoble Powertech (Powertech).
[87] Shin D, Kim J, Chang N, Choi J, Chung SW, Chung E-Y. Energy-Optimal Dynamic 2013.
Thermal Management for Green Computing. 2009 International Conference on [102] Pelley S, Meisner D, Wenisch TF, VanGilder JW. Understanding and Abstracting
Computer-Aided Design (ICCAD'09). San Jose, CA, USA: ACM; 2009. Total Data Center Power. Workshop on Energy-Efficient Design. 2009.
[88] Huang W, Allen-Ware M, Carter JB, Elnozahy E, Hamann H, Keller T, et al. TAPO: [103] Zapater M, Turk A, Moya JM, Ayala JL, Coskun AK. Dynamic Workload and
thermal-aware power optimization techniques for servers and data centers. Green Cooling Management in High-Efficiency Data Centers. 2015 Sixth International
Computing Conference and Workshops (IGCC): IEEE. 2011. Green Computing Conference and Sustainable Computing Conference (Igsc). 2015.
[89] Zapater M, Ayala JL, Moya JM, Vaidyanathan K, Gross K, Coskun AK. Leakage and [104] ASHRAE. Thermal guidelines for data processing environment. fourth ed2015.
temperature aware server control for improving energy efficiency in data centers. [105] Star E. ENERGY STAR® Program Requirements for Computer Servers-Partner
Des Aut Test Europe. 2013:266–9. Commitments V2.1. 2016.
[90] Zapater M, Tuncer O, Ayala JL, Moya JM, Vaidyanathan K, Gross K, et al. Leakage- [106] Chatzipapas A, Pediaditakis D, Rotsos C, Mancuso V, Crowcroft J, Moore AW.
aware cooling management for improving server energy efficiency. Ieee T Parall Challenge: resolving data center power bill disputes the energy-performance
Distr. 2015;26:2764–77. trade-offs of consolidation. 2015.
[91] ASHRAE. Server Efficiency-metrics for computer server and storage: W. Stephen [107] Wang CH, Tsui YY, Wang CC. On cold-aisle containment of a container datacenter.
Comstock; 2015. Appl Therm Eng 2017;112:133–42.
[92] Gandhi A, Gupta V, Harchol-Balter M, Kozuch MA. Optimality analysis of energy- [108] Nelson G. Development of an experimentally-validated compact model of a server
performance trade-off for server farm management. Perform Evaluation. rack. Georgia Institute of Technology; 2007.
2010;67:1155–71. [109] Tanabe S, Arens EA, Bauman FS, Zhang H, Nladsen TL. Evaluating thermal en-
[93] Wen CJ, Xiang L, Yang Y, Ni F, Mu YF. System power model and virtual machine vironments by using a thermal manikin with controlled skin surface temperature.
power metering for cloud computing pricing. Third International Conference on Ashrae Trans 1994;100:39–48.
Intelligent System Design and Engineering Applications (Isdea). 2013. p. 1379–82.
18

Energy Stat China

Uploaded by

Copyright:

Available Formats

Energy Stat China

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Energy Stat China

Uploaded by

Copyright:

Available Formats

Applied Energy 265 (2020) 114806

Contents lists available at ScienceDirect

A review of power consumption models of servers in data centers T

• Analyze the server power trends from generation to generation.

ARTICLE INFO ABSTRACT

Nomenclature OS Operating system

Reference Major contents Major contributions and conclusions

different manufacturers. As of March 2019, there were 641 samples in

As shown in Fig. 3, the rated power of the servers hardly increased

approximately 40% to 50%. Conversely, the ratios of idle power to

respectively. Moreover, the workload operation per second (ops) pre-

power of the servers is not constant from generation to generation.

3.2. Power consumption models of the server

According to recent studies, power consumption models are estab-

management; (2) to predict the power consumption of servers or data

3.2.1. Additive models

Ptotal = PCPU + Pmemory + Pdisk + +PI / O (1)

suming components can be slightly different owning to the configura-

tions of servers, and the proportion of power consumption of each

component is not constant, as shown in Fig. 7 [35]. Some research

[36–38] selected the CPU, memory, hard disk, network, mainboard,

and fan as the main components of the power consumption model,

Fig. 2. The number of samples of each year.

Number 133 212 2 28 2 5 4 2 59 194

Fig. 5. The workload of servers for 1, 2, 4, 6 and 8 CPUs.

PBA = Pbase + Pactive + Delta (2)

Category Authors Power model Factors Benchmark/Testing tool Error

(continued on next page)

Category Authors Power model Factors Benchmark/Testing tool Error

Category Authors Power model Factors Benchmark/Testing tool Error

individual utilization. In contrast, Kansal et al. regarded CPU utiliza-

Fig. 8. The statistical result of the main variables in power models.

peak power of servers accounts for approximately 40% to 50% of

Once the characteristic of components changed, the

regression coefficient need to be adjusted.

ditive models, BA models, and other models according to their

models, non-linear models and polynomial models. CPU utilization

venience and accuracy, whether at the server, system or room level,

cooling load or energy consumption of data centers.

must be considered, and the piecewise function should be used.

they closely represent the actual situation.

Declaration of Competing Interest

The authors declare that they have no known competing financial

ence the work reported in this paper.

Server power consumption and server power

The study is supported by the National Natural Science Foundation

[9] Alkharabsheh S, Fernandes J, Gebrehiwot B, Agonafer D, Ghose K, Ortega A, et al.

art and future trends. Distributed Parallel Databases 2007;21:193–225.

You might also like