Energy Efficiency techniques in cloud computing

Altaf Ur Rahman, Fiaz Gul Khan, Waqas Jadoon

Abstract— Cloud computing gaining popularity at enormous Every organization and individual use CC in one or other
rate since from its emergence. CC changed the way that computing form using online services. Providing and managing such a
services are provided. On demand platform (PaaS), infrastructure flexible platform is a challenging task. Provider built data
as a service (Iaas) and software (SaaS) as a service through center that consist of hundreds of thousands of servers
internet. Consumer use third party services instead of building his
continuously running to provide uninterrupted services.
own infrastructure which need up-front investment and expertise.
Cloud computing becoming popular for unlimited computing Data center are energy hungry station for continuous
power, availability, nice pricing, on demand services and quality of running it requires a nonstop supply of power for running and
service. For availability and computing power the service provider cooling. Data center is a resource rich platform and the
expands their resource capacity to handle user requirements. This scheduling of resources is one of the challenging task. Cloud
expansion in resources capacity lead to high energy demand. Two refer to a data centre where all the user requirements like
big issues for cloud computing is energy demand and hardware and software are provided in the form of
security/privacy requirements. In this survey we will give a review preconfigured resources and remotely hosted applications [2].
on the latest techniques for energy efficiency in cloud computing. Cloud service provider objective is to maximize their profit by
The main focus is on software base energy efficiency techniques in
efficiently utilizing the data center resources. Consumer of a
which we will explain the workload consolidation and resource
management in detail. service aim to be best served in term of cost and quality.
Providing QOS provider have to increase the number and
capacity of the resources which in turn increase the
Index Terms—cloud computing, data center, energy efficiency
techniques. maintenance expenditure and decrease the profit that can be
made possible by the use of virtualization. According to [3]
energy consumption in data center can be classified in to two
I. INTRODUCTION classes computing resources and physical resources.
Computing resources consume about 50% while physical
Cloud computing is a platform which enable individual
resources 40% the remaining 10% by power supply and other
users and conglomerates to use infrastructure, platform and
miscellaneous things.
software as a service through internet instead of buying
Energy efficiency is one of the biggest issue faced by data
managing and developing their own. By using cloud
center. Two percent of the total power produced in USA is
computing (CC), companies reduce their operational cost
consumed by data center. Energy demand double since 2000
while increasing their operational efficiency. With the
[4]. Amazon estimate that energy related cost is forty-two
emergence of Internet of Things (IOT) the CC usage will
percent of its budget. Data center are not environment friendly
grow, in 2003 connected devices were 6.3 billion. The figure
and contribute two percent of co2 emission. Energy
of active devices per person was 0.8. This figure will grow
consumption can be reduced by efficient utilization of data
from 6.3 to 7.6 in 2020 and connected device per person will
center resources. In a survey conducted in 2010 it is find out
be 6.58 [1]. IOT device will use cloud platform because of
that idle server cause of producing 11M ton of co2 emission
low processing power and limited storage.
on yearly basis [5]. Energy efficient data center are not only
economical but also environment friendly.
The rest of the paper consist of five section. Section I
introduction give a brief overview of cloud computing and
Altaf Ur Rahman student of MS program in department of computer science energy efficiency challenges in cloud computing. Section II is
at Comsats Abbottabad Pakistan. (e-mail: [email protected]). about the energy efficiency trends in data center. Section III is
Dr. Fiaz Gul Khan is in Department of Computer Science at COMSATS
about the software based techniques for improving energy
Institute of Information Technology, University Road Tobe Camp, efficiency in data center. The techniques presented by
Abbottabad. (e-mail: [email protected]) Belogazo in 2011 is more focused on power management. We
will focus on software side because so far both the Hardware
Dr. Waqas Jadoon is in Department of Computer Science at COMSATS
Institute of Information Technology, University Road Tobe Camp,
and software side is considered and the techniques on
Abbottabad. (e-mail: [email protected]) virtualization level is not studied in detail. Section four is
about comparative analysis of different software base
techniques on the basis of their strength and limitations.
 II. ENERGY EFFICIENCY TRENDS AT DATA CENTER LEVEL
Data center is a collection of connected server used by an
organization for remote processing and storage. Data center
(DC) give flexible platform to customer by hiding platform
dependency. User has no need of any special hardware all he
need is thin client. Cloud service provider trying to make DC
flexible but harder for the service provider and data center
manager. To efficiently perform user request DC have a
hundred and thousands of servers which should be managed
intelligently. Servers in DC consume huge amount of energy
which is expensive and cause co2 that is 2% of global co2
emission [6]. Building energy efficient data center is not only Figure 1: Energy optimization techniques at data center level
benefited for cloud service provider but also environment A. DVFS
friendly.
Initially researcher was focusing on the power distribution The process of intentionally limiting the performance of the
unit and power management module. They were involved in processor at the time when it consuming too much power
improving the flow of current and avoiding multiple while the task can be accomplished processing at lower
conversion of alternating and direct current. This research frequency. Voltage and frequency have a direct relationship. It
suggests the use of energy star rating IT and non IT equipment has been observed that server consume about 70% power
in data center and green DC. Cooling requirements of DC when it is idle [3] saving this energy will have great impact on
resources is one of the prime energy demanding entity. The energy reduction. DVFS can handle this problem effectively.
suitable location of data center can help in minimizing energy DVFS is basically intended and design for energy efficiency
consumption in this regard. Location selection is based on the in embedded system. But this can be implemented and
temperature and low-cost availability of hydroelectric power consider one of the most effective energy efficiency
[7]. Using these techniques, no major gain is obtained. techniques for computation intensive server. However, this is
Researcher shifted from focusing on designing green DC and considering not good for input output intensive processes [4].
trying to find the energy efficient way in the traditional DC.
DC operator work on implementing hardware base techniques Methods using DVFS techniques mostly consider
for energy optimization. homogeneous physical machine while less interest is shown
Using the hardware base optimization techniques data by researcher toward heterogeneity. In [4] Jacob Leverich
center operator did not get the required gain in energy. suggest a method to power on some of the resources and
According to [8] it was poor program and software design. perform work on them while the rest should be in off state to
This initiated the trend of writing parallel program that could save energy. The difference between consolidation and DVFS
run parallel. Latter on it was found that resource management is that consolidation is applied as a whole regardless DVFS
play important role in energy efficiency. This initiated the which is local [9]. But Wilies lang propos that running all the
trend of developing algorithm for energy efficient resources will perform the work in less time and more energy
management of DC resources. will be saved. Energy efficiency achieved through DVFS
depend on the type of SLA either strict SLA lead to low
III. ENERGY EFFICIENCY TECHNIQUES AT DATA CENTER LEVEL energy saving 1.11% while more effective in relax SLA
At data center level energy efficiency actions can be 6.69% [3].
divided into hardware base, infrastructure base, software base
and location based. In [8] author discus the various power DVFS can be effectively used in CPU intensive task while it
management techniques both at hardware and software level is not suitable in memory and Input output intensive task [10].
while in [7] author focus more on software level techniques A history table is maintained when a request is received it is
for energy efficiency. Further software based techniques for compared against the table to find out whether it is CPU or
energy optimization are classified into five group’s resource I/O intensive. After this calcification the frequency is adjusted.
management, Dynamic voltage and frequency scaling, parallel B. RESOURCE MANAGEMENT
program and workload consolidation. In this survey we will
Resource management is the process of selecting computer
extend the work of [7] by adding the latest techniques in
resources such as computing, storage, network intelligently
resource management and workload consolidation.
and allocating against the single or set of request received to
meet performance objective of the user [11]. Resource
management play starring role in energy efficiency.
Identifying the type of request assigning the best possible
resources in term of performance and price by looking at SLA.
Management is basically the scheduling of resource. When a
request is received the scheduler identify an optimal resource
allocation by looking and analysing the current state of the
system [12]. Scheduling mostly involve with Virtual Machine combine them by bringing it to the least number of pm. let
management. Virtual machine management deals with VM suppose we have two pm its CPU utilization is (80%, 60%)
migration. VM migration performed for three things and (80%, 95%) if a request is made for creating a new VM of
performance, load balancing and energy efficiency. But as our above 40% processing it cannot be served. But we have 85%
focus is energy efficiency so we will look at the energy processing power combinly on two pm if we combine it we
efficiency aspect. can serve a request of creation of new VM above 80%.

In [15] author proposed a scheme based on virtual machine

consolidation, reducing the number of active servers by
consolidating virtual machines dynamically on a minimum
number of servers, and take advantage of the lacking energy
proportionality of commonly used hardware. VM
consolidation is a methodology to maximize resource usage
while keeping low the energy need in a virtualized
environment. Most of the paper does not consider network
energy consumption and cooling energy consumption of a
Figure 2: Resource Management Hierarchy
data center. Consider both the network data structure and
cooling energy when consolidating VM. Fever racks and
In [13] energy aware VM migration techniques is proposed
routers are utilized without compromising SLA to improve
based on firefly algorithm. This algorithm work using three
energy efficiency.
principle (I) fireflies attracted toward each other without
considering his sex because fireflies are unisex (II) less bright C. WORKLOAD CONSOLIDATION
will travel toward brighter firefly’s attractiveness decrease as
Data center are physically distributed over a geographical area.
the distance increase (III) the brightness of the firefly is
When a request is received it is put on the nearest
controlled by the setting of the objective functions to be
geographically located data center. This help to reduce
improved. Firefly algorithm perform energy aware virtual
network delay which in turn reduce turnaround time. Inside
machine migration by migrating VM from the most loaded
data center it is tried to fulfil the request using currently active
active node to the least loaded active node. The decision of
PM to save energy. A large percentage of energy is consumed
migration consists of four steps most loaded active node
as high as 50% more than when server performing at the peak
selection, VM selection from the loaded node, destination
power when server in data center are idle [13]. This high
node where to be migrated and distance update. For selecting
percentage is because 70% of the time server are idle mean
loaded node first computational energy is calculated and a list
under loaded [3]. For saving this energy an effective way of
is generated then a node computation time is calculated and
choosing the right node for the request processing is needed.
values is put in table. In step third attractiveness is measured
A strategy is required to put the under loaded node into low
on the base of CE value the least CE is the first in the list. In
energy state or turnoff. When to turnoff and when to awake
step four distance is calculated and a list is obtained a node is
these are some of the decision that should be devised in this
selected based on CE value which is closest to the calculated
strategy. Similarly, if a new request comes and no active
distance. Compute load on individual VM and create a list in
server can perform then a new server must be turn on or
descending order. Migrate the top VM in VM list to the top
wakeup to fulfil this request. This flexibility is possible by
node on AI list and update the distance to reinitiate the
virtualization technology to put more than one request on a
process.
single server to improve resource utilization and decrease the
To improve energy efficiency is to minimize energy intake
resource demand [16] [17] [18].
of resources such as memory, CPU, network and storage.
Putting more than one VM on a single pm to optimize
Suggested strategies use the approaches of utilizing limited
energy consumption is named VM consolidation. The
resources effectively. Turning off or bringing to sleep mode
consolidation of VM on fever number of pm is online bin
idle resources for saving energy. Network traffic is kept
packing problem. Where pm is a bin and VM is an object and
minimum to optimize network energy. To keep network
the process is real time so decision has to be quick and
traffic minimum VM which have high traffic are shifted on
efficient both in performance and resource utilization
same PM. Keeping network traffic minimum less number of
efficiency [19]. VM are of different sizes according to the
network element will be in active state this will lead to save
resource demand and bin have different capacity in term of the
considerable amount of energy [14].
resources it has. At the time of assigning VM to pm it is tried
VM on pm are not utilized to the fullest of his capacity in
to use the active server instead of turning on or waking up a
[10] author propose a technique called residual resource
new server. How it will select the pm from multiple on pm to
fragmentation. There are different resources but we are more
put the request on them. One techniques are to put the request
interested in processing and memory resources. When a
on the pm which have minimum free space to accommodate
resource usage is limited or blocked by some other resource is
the request. This technique is used to acquire maximum
called resource fragmentation. In this paper author basically
resource utilization by leaving the large free space on other
combine the fragmented resource in each single pm and
pm to be used for another request. However, putting a server
 into sleep mode and waking up considered an overhead but software optimization which is further classified as code
this overhead is tiny [20]. Aggressive consolidation may lead optimization and runtime optimization. Parallel program deal
to performance degradation as well as energy inefficient. with code optimization. For obtaining energy efficiency
Because when a pm performing at the maximum capacity its programmer have a good knowledge of underlying hardware.
performance efficiency and speedup is degraded and its
cooling requirement rise up. IV. COMPARATIVE ANALYSIS OF DIFFERENT ENERGY
EFFICIENCY TECHNIQUES
D. PARALELL PROGRAMING All techniques have some limitation and strength. There is no
Multi core systems improve performance by the cost of technique which have good enough up in all aspects. Mostly
increasing computational energy. Handling fastest processing techniques are focusing on the area which is more responsible
unit in effective way is one of the most complex task in for energy usage like server/storage and cooling which is
computer system. A large number of techniques are available responsible for 50% and 34% power usage of the data center
for improving the energy consumption of these respectively.
multicore/parallel system. All these techniques improve
energy consumption [21] by compromising throughput to
optimize performance per watt. Parallel program came under

Table 01: Comparison of Energy Efficiency Techniques

Technique General comments/Strength Limitation

 Work well for computational intensive tasks.  Didn’t perform well for i/o intensive
task.
 No overhead of maintaining data center level  dependent on SLA not applicable for
load information applicable to a single server. hard quality of service application/task.
Strict SLA lead to low energy saving.
DVFS  effectively save energy while server is idle and
server are 70% of the time idle.  did not consider the priority of the tasks
consider only the workload on CPU.
 turning on a switch off or bringing a sleep node to an  quality of service compromised
active state consume more energy as compared
using DVFS.  no need of data replication because all
the servers are active using the least
possible energy

 Resources are not utilized all the time 70% of the  accurate prediction for resource,
time it is idle. resource capacity and resource time for
which an application required a
particular resource is not known in
 using efficient management of resources more than advance.
50% power can be saved.
Resource  resources management is possible only
Management  resource management is the only techniques that on accurate resource demand prediction
have the potential of more power saving as
compared to other techniques.  performance may be degrading depend
on SLA
 profit and utilization maximization of infrastructure  chances of SLA violation is higher
is a service.  data replication for data availability
 quality of service is maintained
  increase network traffic
 Avoid overloading
 delay involve due to network
 balancing the load for energy efficiency and
 turnaround time increase
performance.
 only consider the workload doesn’t
Workload consider the priority of the tasks
consolidation  turning idle node to sleep or turn off state
 overhead of maintain all the records of
server’s workload
 increase service provider profit by maximum  take time to bring sleep or turn off
utilization of active resources server into active state which increase
chances of SLA violation.
 using consolidation, we have to
replicate data for availability

 Depend on the percentage of code in a

 Time for processing decrease
program that can be parallelized.
 Throughput increase
 less energy saving is achieved as
Parallel  Energy efficiency achieved by decreasing the
compared to resource management and
program processing time workload consolidation.
 component idle time is minimized  programmer is responsible for code
 quality of service is maintained parallelization
 not dependent on SLA

Table 02: energy efficiency effectiveness of different paper on the basis of different parameter

Netwo R- Energy efficiency Exp Resp Fault

Ref Perf rk Load utili Setup time toleranc
consi Perf balancing e

[11] yes No yes yes 10% Cloudsi Decreas yes

m e

[12] yes yes yes yes NA NA Stable NA

1.11 kwh less thean

 [13] yes yes yes yes des-2-n and 12.78 Stable Yes
from ses-1 CloudSi
m

[15] yes yes yes yes 25% NA Decreas NA

e

20.8% static
[16] yes No yes yes 22% dynamic VM Custom NA Yes
load Built
Simulato
r
s
[17] yes No No yes 20% Cloudsi Stable NA
m

[5] A. Paya and D. C. Marinescu, “Energy-aware Load

V. CONCLUSION Balancing and Application Scaling for the Cloud
In this paper we discuss four major energy efficiency Ecosystem,” vol. 7161, no. c, 2015.
techniques DVFS, resource management, work load [6] S. F. Piraghaj, A. V. Dastjerdi, R. N. Calheiros, and R.
consolidation and parallel programing. These four are Buyya, “Efficient Virtual Machine Sizing for Hosting
software base energy efficiency techniques we tried to include Containers as a Service (SERVICES 2015),” 2015 IEEE
the best and latest techniques for energy efficiency in these World Congr. Serv., pp. 31–38, 2015.
four categories. In future we will extend this work by deeply [7] T. Kaur and I. Chana, “Energy efficiency techniques in
studding and analysing the resource management and cloud computing: A survey and taxonomy,” ACM
workload consolidation techniques in detail. Most of the Comput. Surv., vol. 48, no. 2, 2015.
current research based on these techniques. The margin for [8] A. Beloglazov, R. Buyya, Y. C. Lee, and A. Zomaya, A
energy efficiency is high in these techniques as compared to Taxonomy and Survey of Energy-Efficient Data Centers
other techniques and Cloud Computing Systems, vol. 82. 2011.
[9] P. Arroba, J. M. Moya, J. L. Ayala, and R. Buyya,
“DVFS-Aware Consolidation for Energy-Efficient
Clouds,” 2015 Int. Conf. Parallel Archit. Compil., pp.
REFERENCES 494–495, 2015.
[10] A. P. Florence and V. Shanthi, “Energy aware load
[1] V. Albino, U. Berardi, and R. M. Dangelico, “Smart
balancing for computational cloud,” 2014 IEEE Int.
Cities: Definitions, Dimensions, Performance, and
Conf. Comput. Intell. Comput. Res. IEEE ICCIC 2014,
Initiatives,” J. Urban Technol., vol. 22, no. 1, pp. 3–21,
pp. 40–46, 2015.
2015.
[11] B. Jennings and R. Stadler, “Resource Management in
[2] A. Goyal and S. Dadizadeh, “A Survey on Cloud
Clouds: Survey and Research Challenges,” J. Netw. Syst.
Computing,” no. December, pp. 1–14, 2009.
Manag., vol. 23, no. 3, pp. 567–619, 2014.
[3] H. Rong, H. Zhang, S. Xiao, C. Li, and C. Hu,
[12] N. Whittington, L. Liu, B. Yuan, and M. Trovati,
“Optimizing energy consumption for data centers,”
“Investigation of Energy Efficiency on Cloud
Renew. Sustain. Energy Rev., vol. 58, pp. 674–691,
Computing,” 2015 IEEE Int. Conf. Comput. Inf. Technol.
2016.
Ubiquitous Comput. Commun. Dependable, Auton.
[4] Y. Ding, X. Qin, L. Liu, and T. Wang, “Energy efficient
Secur. Comput. Pervasive Intell. Comput., pp. 2080–
scheduling of virtual machines in cloud with deadline
2087, 2015.
constraint,” Futur. Gener. Comput. Syst., vol. 50, pp.
[13] N. J. Kansal and I. Chana, “Energy-aware Virtual
62–74, 2015.
Machine Migration for Cloud Computing - A Firefly
Optimization Approach,” J. Grid Comput., 2016.
[14] J. Dong, H. Wang, and S. Cheng, “Energy-performance
tradeoffs in IaaS cloud with virtual machine scheduling,”
China Commun., vol. 12, no. 2, pp. 155–166, 2015.
[15] S. Esfandiarpoor, A. Pahlavan, and M. Goudarzi,
“Structure-aware online virtual machine consolidation
for datacenter energy improvement in cloud computing
q,” Comput. Electr. Eng., 2014.
[16] W. Vogels, “Beyond server consolidation,” Queue, vol.
6, no. 1, p. 20, 2008.
[17] G. Prekas, M. Primorac, A. Belay, C. Kozyrakis, and E.
Bugnion, “Energy proportionality and workload
consolidation for latency-critical applications,” Proc.
Sixth ACM Symp. Cloud Comput. - SoCC ’15, pp. 342–
355, 2015.
[18] S. Soni and V. Tiwari, “Energy Efficient Live Virtual
Machine Provisioning at Cloud Data Centers - A
Comparative Study,” vol. 125, no. 13, pp. 37–42, 2015.
[19] G. Li, Y. Jiang, W. Yang, C. Huang, and W. Tian, “Self-
Adaptive Consolidation of Virtual Machines For Energy-
Efficiency in the Cloud,” 2016.
[20] M. Zotkiewicz, M. Guzek, D. Kliazovich, and P. Bouvry,
“Minimum Dependencies Energy-Efficient Scheduling
in Data Centers,” IEEE Trans. Parallel Distrib. Syst., vol.
9219, no. c, pp. 1–1, 2016.
[21] B. Goska, J. Postman, M. Erez, and P. Chiang,
“Hardware / software codesign for energy-efficient
parallel computing.”

Altaf ur Rahman was born in Dir lower

district, in 1990. He received his B.S.
computer science degree from SBBU
sheringal and M.S. degree in progress from
Comsats institute of sciences and technology
Abbottabad Pakistan.

Dr. Fiaz Gul was born on 22-11-1982, in a

beautiful valley Abbottabad of KPK. My
graduation and MS is from COMSATS
Institute of Information Technology
Abbottabad in the field of Computer Science.
For specialization master and Doctorate, he
won the HEC scholarship under the project
UESTP for Politecnico di Torino Italy.
Currently he is serving as an Assistant
Professor in Computer Science Department at
COMSATS Abbottabad.

Waqas Jadoon received the Ph.D. degree in

Computer Science from Sichuan University
China in 2014. Currently, he is an Assistant
Professor at COMSATS University, Pakistan.
His research interests focus on Pattern
Recognition, Image Processing, theory and
applications of Machine Intelligence.