IRJET- A Statistical Approach Towards Energy Saving in Cloud Computing

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 4687
A statistical approach towards energy saving in cloud
computing
Anushree chausalkar
Assistant Professor Computer Science Engineering. , ITMGOI Gwalior, India
----------------------------------------------------------------------------------------------------------------------------------------
Abstract : Due to the budget and the environmental issues,
achieving energy efficiency gradually receives a lot of
attentions these days. In our previous research, a prediction
technique has been developed to improve the monitoring
statistics. In this research, by adopting the predictive
monitoring information, our new proposal can per-form the
optimization to solve the energy issue of cloud computing.
Actually, the optimization technique, which is convex
optimization, is coupled with the proposed prediction
method to produce a near-optimal set of hosting physical
machines. After that, a corresponding migrating instruction
can be created eventually. Based on this instruction, the
cloud orchestrator can suitably relocate virtual machines to a
designed subset of infrastructure. Subsequently, the idle
physical servers can be turned off in an appropriate manner
to save the power as well as maintain the system
performance. For the purpose of evaluation, an experiment is
conducted based on 29-day period of Google traces. By
utilizing this evaluation, the proposed approach shows the
potential to significantly reduce the power consumption
without affecting the quality of services.
CCS CONCEPTS
• Computing methodologies → Supervised learning by
regression; • Computer systems organization → Cloud
computing; • Mathematics of computing → Stochastic
processes;
KEYWORDS
IaaS, Cloud Computing, Predictive Analysis, Convex
Optimization, Energy Efficiency, VMs, Gaussian process
regression
1 INTRODUCTION
In recent years, a number of data centre recognized cloud
computing as a popular platform to manage most of the
operations. Naturally, cloud computing improves utilization
and scalability of underlying physical infrastructure. As a
substitution for independently allocating the computing
facilities when being requested, cloud computing is able to
deliver the ordered resource as a virtual package
conveniently via internet connection. Besides, it is worth
noting that cloud computing can be used to enhance the
utilization of the infrastructure by virtualizing the service
composition in a higher level. Hence, the capacity of the
physical facilities can be unified to provide better quality of
services. Finally, implementing cloud computing can lessen
the management cost to consequently save the money.
In order to achieve the reduction of power consumption
in cloud computing, it would be a must to understand the
sources that consume the energy and how to efficiently
reduce the corresponding consumption. Obviously, when a
computing system is online, most of the internal
components burn the power to do the assigned jobs.
Because of this reason, any inefficiently running devices,
which are in the idle state, actually waste the power for very
limited value. Critically, this kind of facilities should be
minimized to save the energy. Regularly, the conventional
approach is to decline the number of working physical
machines to an optimal quantity. By using the virtualization,
cloud computing has a chance to implement this approach
through stacking the virtual machines (VMs). In order to do
that, the VMs can be migrated to an optimal designated
physical machines (PMs). Subsequently, the remaining idle
PMs are turned off to fulfil the requirement of mitigating the
power burning. In the recent research, we have developed
an enhanced prediction technique based on Gaussian
process regression to improve the monitoring statistics. In
this research, we would like to propose an optimization
scheme to reduce the power consumption in cloud
computing. The remaining parts of the paper are organized
as follows. In section 2, we included some related works of
energy efficiency in cloud computing area. Section 3 shows
the description of our proposed architecture. In this section,
the summary of our preceding research, including the
prediction, is also briefly intro-duced. Section 4 includes the
proposal of optimization technique to do the energy
optimization. Section 5 presents how we conduct the
performance evaluation to show the usefulness of the
solution. In the end, section 6 attaches the conclusion of
paper and outlines our future work.

2 RELATED WORKS
Energy efficiency in cloud computing is mostly related to VM
consolidation philosophy. It means that the problem of
interest focuses on choosing the suitable placement for VMs
with regards to the utilization of PMs [3]. Basically, we can
model the VMs and PMs as a regular object-bin problem.
Hence, the VM consolidation can be simplified to bin-packing
problem, which is NP-hard [1]. Consequently, the heuristics-
oriented techniques might be the promising solutions.
Whereby, some well-known approaches popularly adopt this
methodology, namely best fit decreasing [1] and first fit
decreasing [5]. By engaging these techniques, the cloud
orchestrator has a tendency to assign VMs to minimize the
number of hosting PMs. Due to this attractive feature, the
mentioned bin-packing model is widely used to generate the
solution to deal with the energy efficiency. However,
heuristics approaches have a critical drawback when
implementing in action. In order to produce good solution,
this family requires the fixed number of objects and bins at
the beginning of time. In other words, the quantity of VMs
and PMs must be recognized in advance. Apparently, this
requirement is unfeasible since it breaks the principles that
make cloud computing, which are the elasticity and the
multi-tenancy. In addition, the rapid changes in
infrastructure’s utilization clearly degrade bin-packing
approaches in term of accuracy. Because of that, this issue
eventually casts bad effects on system performance.
In order to breakthrough the mentioned obstacle, other
approaches utilize the prediction techniques as a
preprocessing step to enhance the input data. By predicting
the infrastructure’s workload, the cloud orchestrator can
produce more reasonable decision to lessen only unexpected
effect of utilization fluctuation. There would be a number of
research take into account this method to their proposals.
The candidates for prediction algorithms are various from
hidden Markov model [8] to polynomial fitting [16].
Unfortunately, these authors do not pay enough attention to
the designed philosophy of versatile resource provision in
cloud computing. Therefore, these techniques, might not
provide good prospect of underlying system to the
orchestrator. Besides, there is another research trying to
utilize the Wiener filter [7] to predict the workload.
However, to the best of our knowledge, Wiener filter
performs properly only with the stationary signal and noise
spectrum. Bringing signal processing technique to the cloud
computing domain without a rigorous analysis might not be a
good idea. Due to this reason, Wiener filter might be
inapplicable for prediction purpose in the domain of interest.
Furthermore, one different kind of approaches that
should be included is the modified specific schedulers in
[6], [14] and [2]. These schedulers are the efforts to solve
other aspects of energy efficiency in network traffic,
resource reconfiguration and communication rates. By
proposing these schedulers, the authors claim that they can
optimize the network throughput as well as balance the
resource utilization, eventually save the energy. However,
these research do not consider the importance of system
performance preservation. Therefore, the referred
schedulers are unable to implemented in service providing
systems.
By investigating the research area, a conclusion can be
made that even though the energy efficiency is a hot topic
in computer engineering these days, not enough research
has comprehensively been successful in equating the
energy savings with an acceptable performance, especially
in a predictive and optimized manner. Be-cause of that, we
would like to propose a solution which engages our
previous prediction method [4] and convex optimization
tech-nique to reduce the energy consumption in cloud
computing. The rest of the proposal is described in the
next sections.
3 PROPOSED ARCHITECTURE 3.1
System description
Assume that the infrastructure of interest is homogeneous
system. That means all the physical computing facilities are
identical. This assumption is only to make the equation
derivatives more convenient. In fact, this configuration does
not degrade the generality because the heterogeneous
system can be transformed to homogeneous system just by
adding some weighted arguments. As stated previously, the
target of the research is to reduce the power consumption in
cloud computing. In order to do that, we follow the VMs
stacking philosophy. In the other words, VMs consolidation
is chosen to compact the size of running PMs. This choice
relies on the fact that an idle PM actually burns an amount
of power up to 60% [9][11][10] of the peak power, which is
used to maintain the same PM in peak performance. It is
worth mentioning that booting up a PM just burns 23.9%
[13] of the same power. Furthermore, reducing the number
of running PMs delivers additional reduction of the extra
power for maintaining the cooling system as well as the
networking devices. Due to these reasons, stopping idle PMs
can help to save more power than leaving them serving no
specific purpose, even an extra expense is needed to re-
activate the offline computing facilities subsequently.
Relying on this reasoning, we design an architecture, namely
the energy efficiency management (E2M) system shown in

Figure 1. The main target of this architecture is to optimally
create VMs consolidation strategy and send it to the
orchestrator periodically. Finally, the idle PMs are
temporarily deactivated to reduce power consumption.
Following is the functionality of each component in the
architecture:
• Ganglia: this component collects most of operation
statistics for both PMs and VMs. The collected
information is actually used as the input for the
prediction step in the next stage. Note that Ganglia is
known to be trusted platform for years for monitoring
purposes. This component is light-weight but powerful
and versatile enough to integrate to any solution.
• Predictor: this component is the data sink for Ganglia’s
statistics. After receiving the aforementioned data, the
enhanced Gaussian process regression is activated to
do the prediction step. The output of this step is the
predictive monitoring statistics. In other words, the
predictor provides the futuristic perspective of the
working status of infrastructure. This kind of
anticipated system utilization is more valuable for the
optimization step than the original data.
• Energy optimizer: the predictive monitoring statistics,
that are retrieved from the predictor, can be engaged as
the precious input for creating near-optimal
consolidating instruction. The strategy, if possible, has
to save as much power as possible without
deteriorating the quality of services. In fact, the
responsibility of this component is to decide the mini-
mum but feasible set of PMs to normally host the
increasing

VM Migrating Instruction Energy
Optimizer
Predictive
Physical Physical Statistics
Server Server
Predictor
Cloud
Monitoring
Ganglia
Statistics
Orchestrator
Physical Physical
Server Server
Energy Efficiency
Management
Physical Physical
Server Server
Figure 1: Architecture of energy efficiency management (E2M) system.
VMs. Finally, the complete package of VM consolidation is
delivered to the cloud orchestrator for implementation.
3.2 Prediction model
As mentioned above, the migrating instruction would be created in the
energy optimizer component. Before the optimization procedure can be
issued, it is mandatory to enhance the monitoring data in advance. The
reason for the need of this enhancement is twofold. Firstly, it is known to be
true that the monitoring statistics is always the delayed information. It
means that the data we has received at the time t actually reflects the
system status at the time t − τ , in which, τ is the monitoring window that
triggers the data collecting process. Any decision making based on this
obsolete data might not be reasonable at the time the reaction is executed.
Considering this fact, there is obviously a requirement for data prediction.
The second reason is that sometimes it is better to apply proactive reaction
rather than reactive model. In that case, there would be higher chance for
the orchestrator to reduce the violation to quality of services in advance.
Regularly, the target of the predictor is to provide the futuristic utilization
of resources to the optimizer. In order to do that, the Bayesian learning and
the Gaussian process regression are chosen to make the regression. The
guidance on how to build this prediction model is provided in detail in our
preceding research [4].
4 ENERGY OPTIMIZATION
Coming to this step, we assume that the energy optimizer receives enough
information from the predictor, it is right time to conduct the optimization
for power consumption. As said previously, a minimum-but-feasible
number of PMs is required as the output of this stage. Note that the output
is subsequently used to con-struct the instruction for VM migration.
Primarily, there are two sub-components in the energy optimizer, namely
power manage-ment and cluster optimizer. The power management
observes the
resource pool and incorporates the energy decision that has been
made from the cluster optimizer. The final decision can be
referred to as the instruction for VM migration. This instruction is
sent to the cloud orchestrator to actuate.
4.1 Performance modeling
Since CPU is one of the most sensitive parameters, this factor should be
chosen to model the performance. Denote the global utilization as Umfi ∈
R+ and the individual utilization as Imfi ∈ R+ regarding the resource fi (For
instance, fc stands for CPU). The number of active PMs, which is denoted by
am at the monitoring window m, are the target to calculate. It is crucial to
mention that consolidating the VMs into a number am of PMs might result
the infrastructure to its peak performance. Hence, this procedure needs to
be controlled. Otherwise, the whole sys-tem might suffer very high latency
[15] and violate the quality of services, which is described in the service
level agreement (SLA) document. Therefore, the utilization of CPU resource
should be formulated as follows:
Im = max{Imfc } = max{
Ufc
m
}. (1)
amC
fc
fc fc
Observing (1), Im is known to be a decreasing function of am . In other
words, decreasing the number of PMs might cast high latency to entire
system. Denote the average latency of task processing in CPUs as lm . This
parameter can be computed by engaging the expectation waiting time E(fc
) of the exhausted CPU:
l I = fc
) = λm 1/µ2 , (2)
m ( m ) E(
2(1 −
Λ 1 µ
m / )
in which, λm is the arrival rate of the tasks, µ is the service rate of
homogeneous CPU. By comparing lm to the threshold l (which is depicted
in the SLA document), the quality of services can be estimated to be
violated or not. If the violation occurs, the penalty

cost Cmp needs to be calculated as follows:
Cmp = wmsp (lm (Im ) − l)+, (3)
in which, wm and sp stand for the weight factor that reflects the
magnitude of violation, and the fine that needs to be paid for the
penalty, respectively. The weight factor wm is also supposed to
extinguish the trend of the average latency increment. In other words,
this parameter flexibly allows a controlled number of under-
performance PMs as a preventive method for reducing system
overhead. Essentially, this weight plays the role of preserving the SLA
execution.
4.2 Energy modeling
As we all know that the power consumption in running clusters can be
broken down to two periods: the period of processing the assigned
tasks and the period of maintaining the idle state. This fact can be
modeled by using following equation:
em
= Pidl e
+ Pr unninд . (4)
Assume that sm stands for the fine of electricity at monitoring
window m, the power expense, denoted by Cme , is represented as
follows:
Cme (am ) = smamem = smam (Pidl e + Pr unninд ). (5)
In (5), the energy, which is used for processing the tasks, is un-
touchable. As a result, the represented parameter Pr unninд should
not be considered in the optimization procedure. So on, (5) is re-duced
to:
Cme (am ) = smam Pidl e . (6)
4.3 Cluster optimizer
The heart of energy optimizer is represented in this section. As a brief
recapitulation, our objective is to reduce the power consump-tion but still
preserve the quality of services. This objective can be achieved via
minimizing the number am of active PMs. Mathe-matically, the variable am
needs to be found optimally. Firstly, we model the problem by engaging
convex optimization as follows:
0 min (wmsp (lm (Imz ) − l)+ + smam Pidl e ). (7)
≤am ≤Pm
As stated before, the function lm (Im ) is a decreasing function of am .
Therefore, the condition of am can be depicted as below:
a δm
. (8)
m ≤ l−1 l
In case a and
m( )
+ = 0,≥ δm lm−1
l lm Iz l ) = lm δm am ) − lm / ( ) ( ( m) − ( ( / )
then any reduction of a to δm lm−1 l
)
can also reduce the burning
m / (
power. Due to this reason, (7) can be re-formulated as shown below:
minδm
(wmsp
(lm
(Im
z ) − l
)+ + smam Pidl e
). (9)
0≤am ≤ l −1 l
m ( )
The Lagrangian function of this problem can be expressed as
below:
L(am , γ ) = wmsp (lm (δm ) − l)++
am
(10)
δm
sm
am
Pidl e
+γ(am
−
lm−1(l)
) +α(0−am
).
Table 1: Summary of Google Traces’ Characteristics
Time span # of PMs #VM requests # of users
29 days 12583 >25M 925
This function can be solved by applying Karush−Kuhn−Tucker (KKT)
conditions to find the near-optimal value am .
5 PERFORMANCE EVALUATION 5.1
Experiment design
The testbed is a cluster of 16 homogeneous servers. For the detail
configuration, an Intel Xeon E7-2870 2.4Ghz and 12GB of RAM are geared
towards the purposed of hosting upto 8 VMs in each serves. With these
equipments, the infrastructure can host up to 128 VMs at maximum to
conduct the experiment. For the dataset, we use Google traces as a
simulation for the workload. Announced by Google, these traces actually
comprise the monitoring data from more than 12,500 machines over a
duration of 29 days. However, only a set of 6732 machines is chosen to
satisfy the assumption of homogeneous system. In this set, we also extract
randomly 2.26 GB from 39 GB of compressed data for the experiment. The
chosen dataset consists of many parts. Each part represents a period of 24-
hour of traces. For the convenience of presentation, we scale the maximum
length of measurement to 60 seconds. This length is also adopted as the
monitoring window. Moreover, the summary of Google traces’
characteristics is described in Table 1.
5.2 Implementation
The experiment is conducted under four schemes for comparison
as follows:
• The default schemes: all of the PMs are activated all the time.
No power savings is acquired at all.
• The greedy first fit decreasing (FFD) scheme [12]: the VMs are
sorted into queue by descending order in term of internal CPU
utilization. This queue is subsequently submit to the first host that
matches the resource requirement. Basically, the bin-packing
approach is used to relocate VMs.
• The proposed approach (E2M) scheme: the proposed method is
implemented to create near-optimal energy consumption and
preserve the quality of services.
• The optimal energy-aware scheme: an optimal solution is pre-
calculated to achieve minimum energy consumption. In this scheme,
the quality of services is not taken into ac-count. In order words, the
quality of services is sacrificed to significantly save the energy.
5.3 Results
The Google traces is actually a set of synthesized data. Therefore, in order
to measure the energy consumption, an external equiva-lent energy
calculation [13] is applied to compute the result. The description of
calculation and related parameters is depicted in the original paper and
summarized in Table 2. As shown in Fig-ure 2 and 3, because of activating
the PMs all the time, the default scheme consumes egregious amount of
power. Meanwhile, in the

Table 2: Energy Estimation Parameters
Parameter Value Unit
Esleep 107 Watt
Eidle 300.81 Watt
Epeak 600 Watt
Eactive→sleep 1.530556 Watt-hour
Esleep→active 1.183333 Watt-hour
Eactive→off 1.544444 Watt-hour
Eoff→active 11.95 Watt-hour
100
90
(%)
80
servers
70
ofphysical
60
50
Percentage
40
30
20 default scheme
FFD
10 Optimal
E2M
0
0 100 200 300 400 500 600
Time (hour)
Figure 2: Percentage of active physical servers in Google
traces experiment.
FFD scheme, even the power utilization is less than the default scheme, a
remarkable amount of power is wasted since many idle PMs are kept alive
when the workload fluctuates. The reason for this issue is that, without the
capability of prediction, the FFD is unable to appropriately perform the bin-
packing algorithm in ma-jority of times. Another reason is the obsolete
status information of underlying computing facilities. Oppositely, the
proposed ap-proach, namely E2M, can save much better energy by
equipping with the prediction on resource utilization and the optimization
on the pool of active PMs. There is also another additional aspect of this
achievement, which is the gap between E2M and the optimal scheme.
Apparently, the optimal scheme has better energy savings regardless the
system performance. Because the quality of service is totally not considered
in this scheme, this optimal solution brings to the infrastructure too much
overhead and tends to frequently violate the SLA.
For more detail on quantitatively measuring the energy savings, our
proposal can obtain the reduction of power consumption up to
100
90
(kWh)
80
Consumption
70
Power
60
50
default scheme
FFD
Optimal
E2M
40
0 20 40 60 80 100 120
Sampling time (minute)
Figure 3: Power consumption evaluation of the proposed
method in Google traces experiment (lower is better).
100
Power Consumption (%)
90 Latency (second)
80
70
60
50
40
30
20
10
0
default scheme FFD E2M Optimal
Name of approach
Figure 4: Power consumption vs average latency in Google
traces experiment.
34.89% compared with the default scheme. The detail evaluation can be
found in Figure 4. This achievement can be taken into account as a
significant improvement. As a side note, the optimal scheme can only
achieve up to 37.08%. It means that the proposed method can be seen as
a near-optimal solution. Also in Figure 4, our method suffers around
54.72% less than the optimal solution in term of average latency of
system scheduling. Therefore, the quality of services can be preserved in
an acceptable level.

6 CONCLUSION
In this research, a near-optimal energy efficient solution is proposed based
on the utilization prediction of infrastructure and the power consumption
optimization. By engaging the mentioned techniques, our proposal can
create suitable VM migration strategy. Based on this migration scheme, the
cloud orchestrator can issue more rea-sonable VMs consolidation and
condense near-optimal the pool of active PMs. As a result, a significant
reduction in energy con-sumption can be achieved while still preserving the
SLA. In future, we plan to integrate the heuristics algorithm to build a
knowledge base that might help to reduce the overhead when performing
the prediction. This integration might boost up the prediction part to even
more quickly create the VM migrating instruction.
(ICEAC), 2011 International Conference on. IEEE, 1–6.
[14] M. Shojafar, N. Cordeschi, and E. Baccarelli. 2016. Energy-efficient Adaptive Resource
Management for Real-time Vehicular Cloud Services. IEEE Transactions on Cloud
Computing PP, 99 (2016), 1–1. https://doi.org/10.1109/TCC.2016.2551747
[15] Qi Zhang, Mohamed Faten Zhani, Shuo Zhang, Quanyan Zhu, Raouf Boutaba, and
Joseph L Hellerstein. 2012. Dynamic energy-aware capacity provisioning for cloud
computing environments. In Proceedings of the 9th international conference on
Autonomic computing. ACM, 145–154.
[16] Yuanyuan Zhang, Wei Sun, and Yasushi Inoguchi. 2006. CPU load predictions on the
computational grid*. In Cluster Computing and the Grid, 2006. CCGRID 06. Sixth IEEE
International Symposium on, Vol. 1. IEEE, 321–326.
REFERENCES
[1] Yasuhiro Ajiro and Atsuhiro Tanaka. 2007. Improving packing algorithms for
server consolidation. In Int. CMG Conference. 399–406.
[2] Enzo Baccarelli, Nicola Cordeschi, Alessandro Mei, Massimo Panella, Mohammad Shojafar, and
Julinda Stefa. 2016. Energy-efficient dynamic traffic offloading and reconfiguration of
networked data centers for big data stream mobile computing: review, challenges, and a case
study. IEEE Network 30, 2 (2016), 54–61.
[3] Anton Beloglazov, Jemal Abawajy, and Rajkumar Buyya. 2012. Energy-aware resource
allocation heuristics for efficient management of data centers for cloud computing.
Future generation computer systems 28, 5 (2012), 755–768.
[4] Dinh-Mao Bui, Huu-Quoc Nguyen, YongIk Yoon, SungIk Jun, Muhammad Bilal Amin, and
Sungyoung Lee. 2015. Gaussian process for predicting CPU utilization and its application to
energy efficiency. Applied Intelligence 43, 4 (2015), 874–891.
[5] Edward G Coffman Jr, Michael R Garey, and David S Johnson. 1996. Approx-
imation algorithms for bin packing: a survey. In Approximation algorithms for
NP-hard problems. PWS Publishing Co., 46–93.
[6] Nicola Cordeschi, Mohammad Shojafar, and Enzo Baccarelli. 2013. Energy-saving
self-configuring networked data centers. Computer Networks 57, 17 (2013),
3479– 3491.
[7] Mehiar Dabbagh, Bechir Hamdaoui, Mohsen Guizani, and Ammar Rayes. 2015.
Energy-efficient resource allocation and provisioning framework for cloud data
centers. IEEE Transactions on Network and Service Management 12, 3 (2015),
377–391.
[8] Christopher Dabrowski and Fern Hunt. 2009. Using markov chain analysis to study
dynamic behaviour in large-scale grid systems. In Proceedings of the Seventh
Australasian Symposium on Grid Computing and e-Research-Volume 99. Australian
Computer Society, Inc., 29–40.
[9] Xiaobo Fan, Wolf-Dietrich Weber, and Luiz Andre Barroso. 2007. Power provi-
sioning for a warehouse-sized computer. In ACM SIGARCH Computer
Architecture News, Vol. 35. ACM, 13–23.
[10] Ajay Gulati, Anne Holler, Minwen Ji, Ganesha Shanmuganathan, Carl Wald-
spurger, and Xiaoyun Zhu. 2012. Vmware distributed resource management:
Design, implementation, and lessons learned. VMware Technical Journal 1, 1
(2012), 45–64.
[11] David Meisner, Brian T Gold, and Thomas F Wenisch. 2009. PowerNap: eliminat-ing server idle
power. In ACM Sigplan Notices, Vol. 44. ACM, 205–216.
[12] Thiago Kenji Okada, Albert De La Fuente Vigliotti, Daniel Macêdo Batista, and Alfredo
Goldman vel Lejbman. 2015. Consolidation of VMs to improve energy efficiency in
cloud computing environments. (2015), 150–158.
[13] Imad Sarji, Cesar Ghali, Ali Chehab, and Ayman Kayssi. 2011. CloudESE: Energy efficiency
model for cloud computing environments. In Energy Aware Computing

IRJET- A Statistical Approach Towards Energy Saving in Cloud Computing

More Related Content

What's hot (18)

Similar to IRJET- A Statistical Approach Towards Energy Saving in Cloud Computing (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- A Statistical Approach Towards Energy Saving in Cloud Computing