Master of Science in Computer Science Engineering

October 2018

Energy-aware adaptation in Cloud

datacenters

Srivasthav Mahadevamangalam

Faculty of Computing, Blekinge Institute of Technology, 371 79 Karlskrona, Sweden

 This thesis is submitted to the Faculty of Computing at Blekinge Institute of Technology in
partial fulfilment of the requirements for the degree of Master of Science in Computer
Science. The thesis is equivalent to 20 weeks of full time studies.

Contact Information:
Author(s):
Srivasthav Mahadevamangalam
E-mail: [email protected]

University advisor:
Dr. Emiliano Casalicchio
Department Computer Science and Engineering

Faculty of Computing Internet : www.bth.se

Blekinge Institute of Technology Phone : +46 455 38 50 00
SE-371 79 Karlskrona, Sweden Fax : +46 455 38 50 57
ii
 ABSTRACT

Context: Cloud computing is providing services and resources to customers based on pay-per-
use. As the services increasing, Cloud computing using a vast number of data centers like
thousands of data centers which consumes high energy. The power consumption for cooling the
data centers is very high. So, recent research going on to implement the best model to reduce the
energy consumption by the data centers. This process of minimizing the energy consumption
can be done using dynamic Virtual Machine Consolidation (VM Consolidation) in which there
will be a migration of VMs from one host to another host so that energy can be saved. 70% of
energy consumption will be reduced/ saved when the host idle mode is switched to sleep mode,
and this is done by migration of VM from one host to another host. There are many energy
adaptive heuristics algorithms for the VM Consolidation. Host overload detection, host
underload detection and VM selection using VM placement are the heuristics algorithms of VM
Consolidation which results in less consumption of the energy in the data centers while meeting
Quality of Service (QoS). In this thesis, we proposed new heuristic algorithms to reduce energy
consumption.

Objectives: The objective of this research is to provide an energy efficient model to reduce
energy consumption. And proposing a new heuristics algorithms of VM Consolidation
technique in such a way that it consumes less energy. Presenting the advantages and
disadvantages of the proposed heuristics algorithms is also consdered as objectives of our
experiment.

Methods: Literature review was performed to gain knowledge about the working and
performances of existing algorithms using VM Consolidation technique. Later, we have
proposed a new host overload detection, host underload detection, VM selection, and VM
placement heuristic algorithms. In our work, we got 32 combinations from the host overload
detection and VM selection, and two VM placement heuristic algorithms. We proposed
dynamic host underload detection algoirthm which is used for all the 32 combinations. The
other research method chosen is experimentation, to analyze the performances of both proposed
and existing algorithms using workload traces of PlanetLab. This simulation is done using
CloudSim.

Results: To compare and get the results, the following parameters had been considered: Energy
consumption, No. of migrations, Performance Degradation due to VM Migrations (PDM),
Service Level Agreement violation Time per Active Host (SLATAH), SLA Violation (SLAV),
i.e. from a combination of the PDM, SLATAH, Energy consumption and SLA Violation (ESV).
We have conducted T-test and Cohen’s d effect size to measure the significant difference and
effect size between algorithms respectively. For analyzing the performance, the results obtained
from proposed algorithms and existing algorithm were compared. From the 32 combinationsof
the host overload detection and VM Selection heuristic algorithms, MADmedian_MaxR (Mean
Absolute Deviation around median (MADmedian) and Maximum Requested RAM (MaxR))
using Modified Worst Fit Decreasing (MWFD) VM Placement algorithm, and
MADmean_MaxR (Mean Absolute Deviation around mean (MADmean), and Maximum
Requested RAM (MaxR)) using Modified Second Worst Fit Decreasing (MSWFD) VM
placement algorithm respectively gives the best results which consume less energy and with
minimum SLA Violation.

Conclusion: By analyzing the comparisons, it is concluded that proposed algorithms perform

better than the existing algorithm. As our aim is to propose the better energy- efficient model
using the VM Consolidation techniques to minimize the power consumption while meeting the
SLAs. Hence, we proposed the energy- efficient algorithms for VM Consolidation technique

Faculty of Computing, Blekinge Institute of Technology, 371 79 Karlskrona, Sweden

and compared with the existing algorithm and proved that our proposed algorithm performs
better than the other algorithm. We proposed 32 combinations of heuristics algorithms (host
overload detection and VM selection) with two adaptive heuristic VM placement algorithms.
We have proposed dynamic host underload detection algorithm, and it is used for all 32
combinations. When the proposed algorithms is compared with the existing algorithm, we got
22 combinations of host overload detection and VM Selection heuristic algorithms with MWFD
(Modified Worst Fit Decreasing) VM placement and 20 combinations of host overload
detection and VM Selection heuristic algorithms with MSWFD (Modified Second Worst Fit
Decreasing) VM placement algorithm which shows the better performance than existing
algorithm. Thus, our proposed heurisitc algorithms gives better results with minimum energy
consumption with less SLA violation.

Keywords: Cloud Computing, Virtualization, CloudSim, Energy-efficient, and VM

Consolidation.

3
 ACKNOWLEDGEMENT
I would like to show my gratitude to persons who had helped me through my education in
Sweden. The first person I want to thank is my supervisor Dr. Emiliano Casalicchio for his
continuous support. He always encouraged me in every task, I do not have a background in
Cloud Computing, but my supervisor always provided guidelines and clarified even my small
doubts. Thanks, will be a small thing to say for his patience, motivation and great knowledge. I
will always be grateful to him.

I want to thank my parents (Muralidhar and Dakshayani) and my brother (Prasad) for their love
and support. A special thanks to all my friends and my seniors for their support and
encouragement. I want to thank my friend Swetha Penmetsa for her great support throughout the
Maters. I want to thank everyone who supported me in Sweden.

4
 CONTENTS
ABSTRACT ........................................................................................................................................................... 2

ACKNOWLEDGEMENT .................................................................................................................................... 4

CONTENTS........................................................................................................................................................... 5

LIST OF ABBREVIATIONS............................................................................................................................... 7

LIST OF TABLES ................................................................................................................................................ 8

LIST OF FIGURES .............................................................................................................................................. 9

1 INTRODUCTION ...................................................................................................................................... 10
1.1 AIM AND OBJECTIVES ........................................................................................................................... 10
1.2 RESEARCH QUESTIONS ......................................................................................................................... 11
2 BACKGROUND AND RELATED WORK ............................................................................................. 12
2.1 CLOUD COMPUTING ............................................................................................................................. 12
2.1.1 Service models ................................................................................................................................. 12
2.1.2 Deployment models ......................................................................................................................... 12
2.2 VIRTUALIZATION.................................................................................................................................. 13
2.3 VM CONSOLIDATION PROBLEM ............................................................................................................ 13
2.4 VM CONSOLIDATION TECHNIQUE ......................................................................................................... 13
2.5 CLOUDSIM ........................................................................................................................................... 14
2.6 RELATED WORK ................................................................................................................................... 15
2.7 WORK DISTRIBUTION: .......................................................................................................................... 16
3 METHOD.................................................................................................................................................... 19
3.1 LITERATURE REVIEW ............................................................................................................................ 20
3.2 EXPERIMENT ........................................................................................................................................ 20
3.2.1 The system model ............................................................................................................................ 20
3.2.2 Power model.................................................................................................................................... 21
3.2.3 Cost of VM live migration ............................................................................................................... 22
3.2.4 SLA violation metrics ...................................................................................................................... 22
3.2.5 ESV (Energy consumption and SLA Violation) ............................................................................... 23
3.2.6 Adaptive heuristics for dynamic VM consolidation ......................................................................... 23
3.2.7 Base-Line Algorithm for comparison (Existing algorithm): ........................................................... 24
3.2.8 Proposed new Adaptive heuristics for dynamic VM consolidation ................................................. 24
3.2.8.1 Host overloaded detection ...................................................................................................................... 24
3.2.8.2 VM selection policies ............................................................................................................................. 26
3.2.8.3 Host Underloading Detection algorithm ................................................................................................. 27
3.2.8.4 VM placement ........................................................................................................................................ 27
3.2.9 Dependent and Independent variables ............................................................................................ 29
3.2.10 Experimental Setup ..................................................................................................................... 30
3.2.11 Workload Dataset ....................................................................................................................... 31
3.2.12 Hypothesis Testing ...................................................................................................................... 31
3.2.13 Cohen's d: ................................................................................................................................... 32
4 RESULTS AND ANALYSIS ..................................................................................................................... 33
4.1 MODIFIED WORST FIT DECREASING ALGORITHM (MWFD) ................................................................ 33
4.1.1 Results of algorithms under 10 different workloads ........................................................................ 33
4.1.2 Analysis: .......................................................................................................................................... 36
4.1.3 Step 1: VM selection algorithm ....................................................................................................... 36
4.1.4 Step 2: Host Overload detection algorithm ..................................................................................... 37
4.2 MODIFIED SECOND WORST FIT DECREASING ALGORITHM (MSWFD) ................................................ 39
4.2.1 Results of algorithms under 10 different workloads ........................................................................ 39

5
 4.2.1 Analysis: .......................................................................................................................................... 42
4.2.2 Step 1: VM Selection Algorithm ...................................................................................................... 42
4.2.3 Step 2: Host Overload detection algorithm ..................................................................................... 43
4.3 COMPARISON OF RESULTS FOR MWFD, MSWFD AND PABFD .......................................................... 45
4.3.1 MWFD is compared with the existing algorithm PABFD ............................................................... 45
4.3.2 MSWFD is compared with the existing algorithm PABFD ............................................................. 47
5 DISCUSSION ............................................................................................................................................. 51
5.1 RESEARCH QUESTIONS ......................................................................................................................... 51
5.2 THREATS VALIDITY .............................................................................................................................. 52
6 CONCLUSION AND FUTURE WORK .................................................................................................. 53
6.1 CONCLUSION ........................................................................................................................................ 53
6.2 FUTURE WORK ..................................................................................................................................... 53
REFERENCES .................................................................................................................................................... 54

6
 LIST OF ABBREVIATIONS

QoS Quality of Service

SLA Service Level Agreements
VM Virtual Machine
VMC Virtual Machine Consolidation
SaaS Software as a service
PaaS Platform as a service
IaaS Infrastructure as a service
SLAV SLA Violation
SLATAH SLA Time per Active Host
PDM Performance degradation due to Migration
ESV Energy consumption and Service level agreement Violation
IQR Interquartile Range
MMT Minimum Migration Time
MADmean Mean Absolute Deviation around the mean
MADmedian Mean Absolute Deviation around the median
MaxR Maximum requested RAM
MinR Minimum requested RAM
MaxU Maximum Utilization
MinU Minimum Utilization
PABFD Power Aware Best Fit Decreasing
MWFD Modified Worst Fit Decreasing
MSWFD Modified Second Worst Fit Decreasing

7
 LIST OF TABLES
Table 1: VM Consolidation technique and number of proposed algorithms ........................................ 17
Table 2: Combinations of proposed algorithms (when algorithms are divided)................................... 17
Table 3 Combinations of proposed algorithms (When algorithms are shared) .................................... 17
Table 4: Power consumptions ............................................................................................................... 28
Table 5: 32 combinations (host overload detection and VM selection algorithms) .............................. 30
Table 6: Host characteristics ................................................................................................................ 30
Table 7: VM characteristics .................................................................................................................. 30
Table 8 VM selection algorithms using six paired t-test ...................................................................... 37
Table 9: Host Overload detection algorithms with MaxR using 28 paired t-test.................................. 39
Table 10: VM selection algorithms using six paired t-test .................................................................... 43
Table 11: Host Overload detection algorithms with MaxR using 28 paired t-test................................ 45
Table 12: Paired t-test of PABFD & MWFD ........................................................................................ 47
Table 13: 10 combination algorithms rejected the alternative hypothesis (null hypothesis accepted). 47
Table 14: 22 combination algorithms accepted the alternative hypothesis (null hypothesis rejected). 47
Table 15: Paired t-test of PABFD & MSWFD ...................................................................................... 49
Table 16: 12 combination algorithms rejected the alternative hypothesis (null hypothesis accepted). 49
Table 17: 20 combination algorithms accepted the alternative hypothesis (null hypothesis rejected) . 50

8
 LIST OF FIGURES
Figure 1: Cloud computing Services [11] ............................................................................................... 12
Figure 2: The CloudSim architecture [19] .............................................................................................. 14
Figure 3 Research methods and what they achieve .............................................................................. 19
Figure 4: Cloud System Architecture [1] [8] ............................................................................................ 21
Figure 5: Energy consumption of MWFD ............................................................................................. 33
Figure 6: No. of VM migrations of MWFD ........................................................................................... 34
Figure 7: Performance Degradation due to VM Migration (PDM) of MFWD .................................... 34
Figure 8: SLA violation Time per Active Host (SLATAH) of MFWD ................................................... 35
Figure 9: SLA Violation of MWFD ....................................................................................................... 35
Figure 10 : ESV of MWFD .................................................................................................................... 36
Figure 11: Energy consumption of MSWFD ......................................................................................... 39
Figure 12: No. of VM migrations of MSWFD ....................................................................................... 40
Figure 13: Performance Degradation due to Migration (PDM) of MSWFD ....................................... 40
Figure 14: SLA violation Active per Active Host (SLATAH) of MSWFD ............................................. 41
Figure 15: SLA Violation of MSWFD ................................................................................................... 41
Figure 16 ESV of MSWFD .................................................................................................................... 42

9
1 INTRODUCTION
In recent years, the Cloud computing model takes advantage of Virtualization of computing
resources providing the resources to the customers based on the need of the customer on a pay-
as-you-go basis [1]. The Cloud computing provisioning the computing resources to user
applications and services [2] [3]. Instead of experiencing in paying high costs in advance to
purchase IT infrastructure and proceeding with upgrades and maintenance of software and
hardware, these computational needs of the organization can be outsourced to the Cloud [1].
The growth in the Cloud computing results in the establishment of the services and the services
are provided by running the applications on large data centers which contains thousands of
hosts. Based on the activity, typical server systems consume power at different levels [4]. Here
the problem is, even when the typical server systems are idle they consume nearly half of the
actual peak power [4]. Total Cost of Ownership (TCO) is increased when the resources are
maintained and managed.

The other critical problem is that an enormous amount of electrical energy is needed and
consumed by the datacenters to run computing and need cooling operations [5] [2]. More
cooling of the servers is required when it is inefficient is load distribution on servers which
results in the emission of carbon-dioxide [2]. In paper [6], the author said that from 2005 to
2010, 56% increase in the energy consumption by the data centers virtualization and committed
1.1% to 1.5% of global electricity usage [7] [56]. The energy consumption of datacenter equals
to power/energy consumed by 25,000 households [8] [9].

To overcome the problem of energy consumption by data centers, the resources should be used
in efficient manner such that power consumption can be reduced without affecting the
performance [8]. The idle power consumption can be eliminated by switching idle nodes to
sleep mode or hibernate mode, this reduces the energy consumption [1]. There should be
reliable Quality of Service (QoS) between the Cloud providers and customers, it is essential for
Cloud computing environment, QoS is defined via SLAs (Service Level Agreements). Cloud
providers have to maintain and minimize power consumption while meeting SLAs. The energy
management of Cloud data center is essential, so our aim of this thesis is to reduce the energy
consumption of Cloud datacenters along with not compromising on SLAs contracted. To
achieve this, we need to model some energy-efficient resources management techniques. The
technique we are going to use in this thesis work to overcome this issue is VM consolidation
technique (VMC) which is explained in Section 2.4, this VMC is used by many researchers for
any Cloud system to minimization the energy consumption.

The focus of this work is to propose an energy efficient algorithm for VM Consolidation. These
algorithms were evaluated by simulating using CloudSim toolkit. Proposed algorithms which is
explained in Section 3.2.8 effectively reduce the energy consumption while meeting the SLA’s.
The main contribution of this thesis work is described as follows:
x Proposed host underload detection, host overload detection, VM selection algorithm and
VM placement algorithm to minimize the energy consumption by using VM
consolidation technique.
x Analysis and simulation-based evaluation of the proposed algorithms.

1.1 Aim and objectives

Aim:
The aim is to perform the energy-efficient adaptation model using CloudSim, i.e. proposing an
energy- efficient model to reduce the energy consumption of the system.

10
Objectives:
The following are the research objectives that will help us to reach our aim:
1. To learn about the VM consolidation technique and virtualization concept.
2. To analyze the existing energy-efficient heuristic algorithms to get knowledge about VM
consolidation of datacenters, i.e. the output of RQ1.
3. To analyze the CloudSim simulation and execute the energy-efficient model.
4. To analyze the existing performance of the energy-efficient model.
5. To propose an energy efficient heuristic algorithms those consumes less energy and
maintain SLA using CloudSim simulation model.
6. To obtain the performance of the algorithms, i.e. the output of RQ3.
7. To compare the results of the energy-efficient heuristic algorithms and identify best
algorithm among them, i.e. the output of RQ2.

1.2 Research Questions

Research Question1: What are the algorithms suitable for the energy-efficient model of the
Cloud-based system?
Motivation: By doing a literature review on the existing algorithms, we can implement a new
algorithm of the energy-efficient model with better results than the already proposed algorithms
(existing algorithms).

Research Question2: What are the advantages and disadvantages of the proposed algorithm of
the energy-efficient model?
Motivation: By statements of the advantages and disadvantages of proposed algorithms, we can
compare the results of the proposed algorithms with existing algorithms and can state the best
energy-efficient algorithm and can differentiate with others.

Research Question3: What are the performance of the algorithms, i.e. the output of the energy-
efficient model?
Motivation: By performance of the algorithms we can understand and state the heuristic
algorithms with better results and we can compare our proposed heuristic algorithms with the
existing algorithms to determine the energy efficient model.

Thesis Work Distribution:

This thesis is divided based on the VM Consolidation technique. As the topic is described
detailly in the next Chapter 2 Background and Related Work, we mentioned the work
distribution in the Section 2.7.
The contents of this thesis document are organized as follows. In Chapter 2, the background
knowledge, the topics related to the chosen research area and the techniques used for this
research area were provided and mentioned. Chapter 3 describes the method chosen and used
for this study. The results and analysis of the study have been presented in Chapter 4. In
Chapter 5, the discussion has been provided in which answers to the research questions have
been discussed. Chapter 6 gives the conclusion and future work.

11
 2 BACKGROUND AND RELATED WORK
In this Chapter, we discuss about the background and techniques used to overcome the problem of
energy efficient by Cloud data centers and the simulation model in detail. From the related work
which is presented in this Chapter explains about the various energy efficient models. This Chapter
brings more insights about the models and techniques of the energy efficient model.

2.1 Cloud Computing

Definition: Cloud Computing is a model using the network access of a computing resources like
networks, servers, and storage that can maintained and released with less management effort. It serves
as a local server or more like a personal computer [3]. Cloud consists of three service models and four
deployment models.
2.1.1 Service models
The following are the Service models [10]:

x Software as a service (SaaS):

SaaS model allows consumers to use the software applications as a service, which is
accessed by the end users through a web browser [10].

x Platform as a service (PaaS):

PaaS is a Cloud computing offering that provides the consumers with a consumer-created
Cloud infrastructure [10]. Here the consumer can develop, manage and deliver applications
created using programming languages, which are aided by the provider [10].

x Infrastructure as a service (IaaS):

IaaS is a Cloud computing offering in which a vendor provides user access to computing
resources such as servers, storage, and networking [10]. Organizations use their own
platforms and applications within a service provider’s infrastructure [10].

Figure 1: Cloud computing Services [11]

2.1.2 Deployment models
The Cloud consists of following deployment models [10]:

x Private Cloud [10]:

The single organization holding multiple consumers (i.e., business units) maintains the
Cloud infrastructure for the exclusive use. The organization owns manages and operates
the Cloud infrastructure, the third party may also maintain the Cloud infrastructure.

12
x Community Cloud [10]:
An organization which has a specific community or association of consumer’s that have
shared concerns (e.g., Policies, requirements regarding security, operations) maintains a
Cloud infrastructure for the exclusive use [3]. One or more organization which is in the
association can manage, own and operate the Cloud infrastructure.

x Public Cloud [10]:

The public Cloud allows systems and services to be accessible to general public [10]. The
academics, business or government organization can manage, operate and own the Cloud
infrastructure [10].

x Hybrid Cloud [10]:

Hybrid Cloud is the composition of two or more distinct Cloud infrastructure (public,
private and community) [10]. It is bound by standard or proprietary technology [10].

2.2 Virtualization
The utilization of the computing resources by the users is done through virtualization techniques
[12] [13, 14, 15, 25]. Virtualization is the main concept in Cloud computing. In computing,
through virtualization virtual things can be created like storage devices, virtual computer
hardware and network resources of a computer, aids in better resource utilization and increasing
the Return on Investment (ROI) [1]. “A virtualization technology is a promising approach to
consolidating multiple virtual machines (VM) onto a minimum number of servers” [2]. There
are two concepts in virtualization called VM Consolidation and VM live migration. VM
consolidation is a technique for shifting the VMs to a single host. VM live migration is
transferring the VMs between one host to another host [16].

2.3 VM consolidation problem

Data Centers uses VM Consolidation to reduce energy and power consumption and to increase
resource utilization [17]. Server consolidation is important when the user uses unpredictable and
complex workload, it needs to be revisited time to time. When the demands of the user change,
VMs can be migrated or it can be resized. Migration is done between the physical servers if
needed.

Two type of consolidation process can be performed

x Static consolidation: Configure the VMs capacities, by using peak demand of load available
(a static value). Due to this hosts may lead to idleness. VMs stay in the same host at any
time, i.e. no VM migrations are possible.
x Dynamic consolidation: Periodically change configurations of VMs capacities according to
the current demand in workload (dynamic value). Doesn’t lead idleness of hosts, as it
changes configurations of VMs capacities dynamically using current demand in workload. It
Enables VM live migrations between nodes in the datacenter. Due to this reason dynamic
consolidation perform better than static.

2.4 VM consolidation technique

From [1] [9], VM Consolidation can be distributed with four categories/scenarios. The
following are the categories used by VM consolidation for migration of the VMs:

1. Host underload detection: In this underload detections, hosts were checked for underload,
if there is underload host then the VMs present in that host are migrated to another host
by changing it to low-power mode.

13
 2. Host Overload Detection: In this detection, if the hosts overload then the VMs present in
the host is migrated to another active or reactivated host to mitigate the SLA violation
(QoS requirements).
3. VM Selection: Selection of the VMs to transfer from the one overload host to active host.
4. VM placement: VM placement is to place the selected VM on the active host.
Algorithms need to be designed where the migration of VMs is done so that there is less energy
consumption and it can mitigate the SLA violation and can transfer VMs so that only some
hosts are inactive.
2.5 CloudSim
Motivation: In this thesis among the various tools CloudSim [18] is considered, as it is an
established tool to achieve simulation of Cloud environments, modeling, and interpretation of
resource algorithms. The main benefit of using the tool CloudSim in terms of performance
includes applicability, time effectiveness and flexibility. Also, it helps in meeting the necessities
of our study, by showing the activities that resemble large-scale Cloud data centers. Any
organization that supports services related to the Cloud should test the services in advance
before deploying to the Cloud, this is to lower the risk in the time of deploying. To assess the
results and to model the mechanisms, usually, simulations are used by researchers.

Figure 2: The CloudSim architecture [19]

CloudSim Architecture and Modeling [18]:

CloudSim core simulation engine: It is the CloudSim first version released. This supports
functionalities alike host, datacenter, datacenter broker and VMs etc., In order to support
leading operations, this layer was taken out from the present version of CloudSim.

14
CloudSim simulation: For the creation of resources like the host, VMs, datacenter this layer
gives support. Also, it is used for interface management, storage, bandwidth, memory to
simulate any Cloud environments and manage the execution of applications, this layer monitors
dynamically to evaluate the provisioning algorithms

User Code: This layer is present in the top of CloudSim architecture, this shows vital entities
necessary for requirements application such as the type of users, datacenter broker policies,
tasks. These entities can be extended to generate model Cloud scenarios, configurations and
conduct tests based on implementing techniques for Clouds.

Modeling the Cloud using CloudSim: In this thesis, we work on the middle layer, i.e. CloudSim
simulation. Here, by extending datacenter, hosts and VMs entities we can simulate Clouds
related to IaaS using CloudSim. Cloudsim allows both multi-core and single functionalities for
the hosts. All these entities will be assigned with required characteristics. Based on VM policy
specified by CSP, datacenter allocates hosts to several VMs [18].

2.6 Related Work

Many researchers worked on the concept of VM consolidation, resulting in required amount of
research work under this topic. Few of them are as follows:

A Beloglazov and R Buyya et al. [1] [9] proposed heuristic algorithms for all four categories
present in the VM consolidation technique. The algorithms proposed are host overload detection
algorithms: Median Absolute Deviation (MAD) and Interquartile Range (IQR) and Regression.
Algorithms proposed for VM selection are Random Choice (RS), Minimum Migration Time
(MMT) and Maximum Correlation (MC). And finally, for VM placement they have proposed
Power Aware Best Fit Decreasing (PABFD) based on the concept of bin-packing problem. All
these algorithms are designed as adaptive heuristics, this is the only research where all four
categories of VM consolidation were analyzed compared to any other research done under the
concept of VMC.

In [20] [57], authors have proposed a heuristics method and the method is for bin packing
problem. The method is for placement of the VMs on the physical servers. They have not
considered the QoS (Quality of Service) and did not mitigate the SLA Violations. In paper [21],
authors have developed a Bayesian network-based estimation model (BNEM), the algorithm is
used for the VM migration. The validation is done by performing the algorithms in the
CloudSim.

In paper [22], the authors mainly focused on the energy consumption which is impacted by the
RAM, since many of the experiments on the energy efficient is mainly focused on the energy
consumption by the CPU and the other hardware components were ignored. The authors in this
paper, focused both CPU and RAM energy consumption. They have proposed two approaches
based on both CPU and RAM usage. The approaches were CPU and RAM Energy-Aware
(CREW) and Underload Detection and CPU and RAM were taken into consideration. Traces of
the Google Cluster has taken as workload and CloudSim is used for the implementation and
evaluation.

In [22] [23] paper, to deliver an energy- aware resource allocation system with the combination
of the local and global managers, the authors have proposed an model. The physical server
contains the local manager and the local managers engages the policies of the system-level and
sends to the control power consumption. Repeatedly, the global manager receives the
information from the local manager and modify the VM placement. The VM placement is done
using live migration and information to the global manger is related to the resources usage. The
global consolidation strategy was not provided by the authors.

15
In [24], authors have proposed a heuristic approach based on the genetic algorithm. It is for the
Virtual Machine packing optimization. They did not do concerning the SLA Violation and
Migration costs. Instead of Virtual Machines they proposed that uses Virtual Clusters.

In [9] VM Consolidation with migration control is instituted. In this, VMs which are steady are
not migrated to other hosts, when the VMs are not steady they are migrated from one host to
another host for better performance. Heuristics approach has been proposed for the migration of
the VMs. It is static and have minor error but acceptable performance.

In [25], the authors focus on the QoS requirements and satisfaction of the Cloud users and states
the issues in the performance and methods for the resource allocation. This was done by survey,
the author conducted survey for addressing and summarizing the issues in performances and
methods of the resource allocation of the Cloud computing. The author describes the final goal
of the resource allocation is that Cloud provides should have maximum profits and consumer
should have the minimum cost.

In [26], for the minimizing the cost infrastructure and SLA violation, authors proposed a
resource allocation algorithm, it is for providing the satisfaction for the service customer. Three
algorithms were proposed. The first algorithm is to get the profit maximum by minimizing the
number of SLA, the second one is to increase the profit by minimizing the cost by the Virtual
Machine reuse (most access is needed), the third algorithm is to get profit maximum by
minimizing the cost by reusing the VMS when the less access is needed. It is for the customer
and SaaS.

In [25] [27], the authors described about the issues in the resource allocation. The issues were
based on the knowledge of the workload and performance of the services. Author had
implemented scheduling heuristics and resource allocation algorithms. The performance
feedback is provided to the source which was taken from the targeted machine performance.
The resource allocation method proposed is improved by considering the feedback (feedback
mechanism).

In [25] [28], the authors focused on the two main objectives overload avoidance and dynamic
resource allocation through virtualization techniques in the green computing. They have
proposed a load prediction algorithm which is used to predict the resource utilization. Prediction
is done without seeing into virtual machines. For designing this algorithm, the authors used
trace driven simulation.

Research Gap: The research work available till now regarding energy efficient model using
VM consolidation, researcher did not attempt to implement the heuristic algorithm in all four
categories of VM consolidation technique, they have proposed algorithms in any one of the
categories of VM Consolidation. In this research work, we have proposed new algorithms for
four categories, i.e. new algorithms for host overload detection, host underload detection, VM
Selection and VM placement heuristic algorithms has been proposed and compared with the
existing one and proposed one to get best energy efficient algorithm.

2.7 Work Distribution:

This thesis is done under the supervision of Dr. Emiliano Casalicchio and shared with Swetha
Penumetsa (swpe16). In this research work, we use the VM Consolidation techniques (described
in the Section 2.4) has 4 categories: host underload, host overload detection, VM Selection and
VM placement algorithms. VM migration from the hosts are done using these techniques. In
this work, we have proposed new heuristics for all four categories of VM Consolidation
technique which are host underload, host overload detection, VM Selection and VM placement

16
heuristic. The VM Consolidation algorithm is a set of all four categories of VM Consolidation
techniques.

VM Consolidation algorithm = host underload detection + host overload detection +

VM selection + VM placement algorithms

The working of a final VM consolidation algorithm changes when we change heuristic

algorithms belonging any of the four categories within the VM consolidation.

In this thesis, work is divided among the proposed heuristic algorithms of four categories of VM
Consolidation technique. We have shared the three categories of VM Consolidation technique ,
i.e. host underload detection, host overload detection and VM selection with different VM
placement heuristic algorithms. The three categories were shared because we could cover all the
possible combination algorithms (proposed host overload detection, host underload detection
and VM selection algorithms) with different VM placement algorithms. If we did not share
them and use different algorithms in each category, then we may have only limited possible
combination of algorithms that we would work with. And we would have missed the other
possible combination of algorithms that might not be explained in both the thesis work, these
missed algorithm combinations might contain the best heuristic algorithm in it.

For example:

VM Consolidation technique and number of proposed algorithms:

Host underload Host overload detection VM selection VM placement

1 (U) 2 (X, Y) 2 (A, B) 2 (M, N)
Table 1: VM Consolidation technique and number of proposed algorithms

If we divide the algorithms and implement them, person 1 and person 2 will get only one
combination of host underload, host overload detection and VM selection with VM placement
algorithm as follows:

Person 1 Person 2
st nd
1 combination 2 combination
UXAM UYBN
Table 2: Combinations of proposed algorithms (when algorithms are divided)

If we share the algorithms and implement them, total we get 4 combinations of host underload,
host overload detection and VM selection with VM placement algorithm and person 1 will get 2
combinations of host underload detection, host overload detection and VM selection algorithm
with one VM placement algorithm and person 2 will get 2 other combinations of host underload
detection, host overload detection and VM selection algorithm with one VM placement
algorithm as follows:

Person 1 Person 2
st nd rd
1 combination 2 combination 3 combination 4th combination
UXAM UYBM UXAN UYBN
Table 3 Combinations of proposed algorithms (When algorithms are shared)

From comparing the above combinations, we missed UYBM and UXAN combination when the
algorithms are divided. There may be best heuristic algorithm among those missed

17
combinations. So, to purpose and implement all possible combinations we have shared the 3
categories of the VM Consolidation technique.

Since we have different VM placement algorithms and proposed different energy efficient
model, the final VM consolidation algorithms presented in both the thesis were totally different,
which means each result and working presented was unique and doesn’t match at any case.

In many research papers, authors have proposed the energy efficient model by considering one
of the VM Consolidation techniques, i.e., they have proposed new heuristic algorithm in any
one part of the VM Consolidation technique, while the heuristic algorithms for other three
categories were used from any of the existing literature they found.

18
3 METHOD
In this method, we explain the research Methodology, to which our study complies to. We have
conducted the experiments and done Literature Review to obtain the details of the existing
models to do experimentation for the optimize solution. Research Design is the process of
method selection for a particular research problem, working on its strengths, while avoiding and
mitigating its weaknesses [29]. The following process explains our research design:

LITERATURE REVIEW:
For answering the Research Question, we need to know about the model and existing algorithms
which works on that model. We defined to present a best energy efficient algorithm using
CloudSim, to obtain the parameters and to know about the independent and dependent variables
and to learn about the background of the existing algorithms of the energy efficient model, we
have done the Literature review.

EXPERIMENTATION:
“A controlled experiment is an investigation of a testable hypothesis where one or more
independent variables are manipulated to measure their effect on one or more dependent
variables” [29]. To answer the remaining questions, we have chosen experimentation method.
Through this experimentation we are going to implement our heuristics algorithms based on the
host overload detection, VM Selection and VM placement. These parameters are extracted from
the Literature Review.

In this methodology, we have not chosen Survey and Case Study. In [30], author introduces the
case study as “an empirical inquiry that investigates a contemporary phenomenon within its
real-life context, especially when the boundaries between phenomenon and context are not
clear.” Case Study is the process of understanding the existence of the tasks. Since, we are
implementing a new algorithm; we have not chosen Case Study. Survey is to interview the
people and having a questionnaire for collecting the data [29]. This is not a relevant study for
our research. So, survey method is ignored.

Figure 3 Research methods and what they achieve

19
3.1 Literature review
For this study literature review is used as the research method. Cronin et al stated that “A
literature review provides the reader with the comprehensive background for understanding
current knowledge and highlighting the significance of new research” [31]. The main objective
to conduct the literature review for this study is to gather the relevant information and
background from the existing literature on the research topic.

The literature review is carried out in 5 phases. They are search string selection, collecting
literature, evaluating literature, writing the review and building bibliography [32]. Snowball
sampling method is used in collecting the literature review to mitigate the chances of missing
the relevant literature for the research [33]. The forward and backward sampling is done using
the references and citation from each article.

STEPS IN CONDUCTING LITERATURE REVIEW

1. Search string selection:

The search string is used to retrieve the relevant literature. The search string of our study is
created using the keywords “Cloud Computing", "Green Computing" "Virtualization",
"CloudSim", "energy-efficient", "VM Consolidation".

2. Collecting literature:
In this step, the relevant literature is collected form the databases Google Scholar, IEEE Xplore,
Scopus using the selected search strings. Then the documents are selected based on inclusion
and exclusion criteria. The criteria are based on:
x The articles available in full text.
x The articles written in English.
x The articles published after the year 2007.
x The articles discussing about the Cloud Computing, CloudSim, Virtualization, SLA
Violation, Energy consumption were selected.
x The articles that are not repeated.

3. Evaluating the literature:

The collected literature after the inclusion and exclusion criteria are analyzed and evaluated.
The relevant literature is included after excluding the irrelevant article by through reading of
titles, abstract and full text. After filtering the resulted documents are used to write the literature
review.

4. Writing the review: After the selection and evaluation of literature, the collected
information in the literature is used for documentation.

5. Bibliography: The references of the literature that are used in this review is cited in this step.

3.2 Experiment
3.2.1 The system model
The user submits their request for the provisioning of the N heterogeneous VMs through the
data centers. The VMs provision is characterized by the CPU performance which is defined by
the MIPS, amount of RAM and network bandwidth. The provisioning of the VMs is done
between the hosts, i.e. allocation of the VMs.

In our model, multi core CPU architecture is provided to the physical servers. Here modeling of
multi core CPU is done with n core in which each core has m MIPS as a single core CPU with

20
having of nm MIPS total capacity. The only restriction is that VM CPU capacity must be less
than or equal to the single core capacity. If the VM CPU capacity is greater than single core
capacity, then VM should be executed parallelly in multiple cores. This VM parallelization is
complicated research problem.

Figure 4: Cloud System Architecture [1] [8]

The system architecture is classified into consisting of local and global managers [1] [8]. The
local managers are place on the node. The CPU utilization is monitored by the mangers
regularly, the VMs resizing were done accordingly based on their needs and the provisioning of
the VMS has been decided continuously [1] [8].

The global manger is placed on the master node and collects the information from the local
manager and it maintains and regulates the (optimization of the VM placement) [1] VM
placement optimization and resource utilization to VMM (execute the functions accordingly)

3.2.2 Power model

Power consumption by nodes in the datacenter is governed by the CPU, memory, disk storage
[9], among these the CPU utilization takes the major part of data centers energy consumption.
Recent studies [34, 35, 36, 37] [9] have shown that CPU utilization of data center that linearly
proportional to its power consumption. The techniques used in the experiment are based on
power consumption linear model.

The recent studies state that the server which is idle consumes 70% of the energy/ power when
compared with the server which is running at full CPU speed. By this, we can justify that the
switching from the idle mode to sleep mode can reduce the power consumption by 70%.

The system configuration of the power consumption model is defined in equation 1 as follows
[9] [38]:

Where,
o is the power consumed when the server fully works.
o is the fraction of the power consumption by the idle server (i.e., 70%)
o is the CPU utilization
o We have taken as 250W, which is a common value for the modern servers [9].

Due to the workload variability, there will be changes in the CPU utilization over time to time
[9] [39]. CPU utilization is function of time and defined as u(t). Therefore, time period ( )

21
the total consumption of energy by a physical node (E) can be stated as an integral of the power
consumption function over as defined in equation 2 [9] [39].

3.2.3 Cost of VM live migration

Live migration of VM is nothing but a VM transfer between the physical nodes. The
transferring of VM is done without interruption and with short downtime [1]. Voorslys et al.
have performed same experiment to study on the live migration negative impact on the
performance of the applications running in VM while migration [40] [1]. They have found that
application behavior is the cause for the performance degradation and short downtime [it is
estimated approx 10% of the CPU utilization].

In our simulation, during the migration we prototype the same amount of CPU capacity to
allocation of VM on the end node [1]. This results in SLA violation, so there is a need to
minimize the SLA violation by decreasing the number of VM migrations. The length of VM
migration is dependent on the amount of memory used by a VM and network bandwidth. The
migration time and performance degradation which is accomplished by the VM j is defined in
equation 3 and 4 [1].

Where,
o is completion time of migration
o is memory used by VM j.
o is available network bandwidth.
And,
o is total performance degradation by VM j.
o is starting time of migration
o is CPU utilization of VM j.

3.2.4 SLA violation metrics

In Cloud computing environment, the QoS requirement must be achieved and it is considered as
the important task. The QoS requirements are standardized in the form of SLAs.

The occurrence of SLA violation happens when VM running cannot get its provisioning
resources (such as MIPS, memory) from the datacenter [41]. When SLA violation occurs the
Cloud service provider should pay penalty to Cloud service users.

In an IaaS environment, to measure the level of SLA violation we have considered the metrics
from [1]
x SLA violation Time per Active Host (SLATAH):
It is the time percentage, during which active hosts accomplished 100% CPU utilization.

22
 Where,
o is number of hosts active
o is total time of the host i accomplished the 100% CPU utilization leading to SLA
violation [1].
o is total time of the host i being in the active state.

x Performance Degradation due to Migration (PDM):

Comprehensive performance degradation by VMs while migration

Where,
o is number of VMs
o is the estimate of the performance degradation of VM, j which is caused due to
migration (In [1] the estimate is considered as 10% CPU utilization in term of MIPS
during all VM j migrations).

x SLA violation (SLAV):

Both SLATAH and PDM distinguish the level of SLA violation, [1] combined metric SLA
violation (SLAV) expressed as

3.2.5 ESV (Energy consumption and SLA Violation)

As energy consumption (measured by power model defined in Section 3.2.2) and SLA Violation
(defined in Section 3.2.4) are the main metrics to be measured and these two-metrics results
opposite to each other, i.e. algorithm with less energy consumption might result with high SLA
Violation and vice versa. So, one more combined metrics has been defined in [1] named as ESV
which is a combination of metrics Energy Consumption and SLA Violation.

3.2.6 Adaptive heuristics for dynamic VM consolidation

The problem of dynamic VM consolidation is divided into four categories:
1. Host overloading detection
2. Host underloading detection
3. VM selection from the overload host
4. Placement of VMs which is selected for migration

The VM placement optimization algorithms are shown in [1]. First phase, the overload detection
algorithm checks whether host is overloaded in the host list. When the host is overloaded, VM
selection is applied to select the VMs that need to be mitigated from the hosts and VM
placement algorithm is applied to select the host to place the VMs for the migration [1]. The
second phase of the algorithm is to find the underloaded hosts in the host list and VM placement
algorithm is invoked for the VMs migration. The algorithm provides information about the new
VM placement of the specified VM selected which are to be migrated from the overloaded and
underloaded hosts [1]. The complexity is 2N, where N is the number of hosts.

23
3.2.7 Base-Line Algorithm for comparison (Existing algorithm):
In [1], they have proposed few heuristic algorithms for host overloading detection, host
underload, VM selection and VM placement (VM Consolidation). They have represented the
following heuristic algorithms as the best heuristics, since they take less energy consumption
and less SLA Violation.

The following are the heuristics proposed in [1]

VM placement:
In [1] the author proposed Power Aware Best Fit Decreasing (PABFD), which is described in
Section (3.2.7)

Host overloading detection:

Interquartile Range (IQR) - An adaptive utilization threshold: It is calculated as a difference
between the third and first quartile, it is a measure of statistical dispersion.
IQR=Q3-Q1

VM selection policy:
The Minimum Migration Time policy (MMT): MMT migrates a VM which needs a minimum
time for completion of the migration when compared with another VM in the host.

In this thesis, we considered these above described heuristic algorithms as a base line for
comparing with our proposed algorithm. And we refer them as PABFD (IQR_MMT) which
means VM consolidation algorithm with Power Aware Best Fit Decreasing VM placement
heuristic algorithm (Interquartile Range host overloading detection heuristic algorithm
_Minimum Migration Time VM selection heuristic algorithm).

3.2.8 Proposed new Adaptive heuristics for dynamic VM consolidation

3.2.8.1 Host overloaded detection

An adaptive dynamic utilization threshold:
Depending on the workload patterns presented by the application the system must automatically
adjust its behavior accordingly [1]. For the adjustment of the utilization threshold automatically
we proposed a technique based on the statically analysis. For the analysis the data is collected
from the lifetime of the VM, it is the historical data of the VM. The following are the
descriptive statistics we have used for the adaptive dynamic utilization threshold.

24
1. Mean: Mean is nothing but a average of the taken values. It is calculated as adding the
input numbers ( ) and divide it by total numbers n

2. Median: Median is considered as the middle value in the dataset. For the dataset in
statistics, it is used as measure of the property.

3. Mode: Mode is the number that is repeated in the dataset.

4. Range: Range is the value we get from the difference from the highest and lowest value of
the dataset.

5. Variance: Variance is a measurement of the data value in the dataset from its mean. First,
Variance is calculated as mean in the dataset and gets the difference with mean value and
data value in the dataset. It makes square of the difference values and divides it with the
number values of the dataset.

Where, is the mean

6. Standard Deviation: It is nothing but a square of the variance.

7. MADmean: MADmean refers to Mean Absolute Deviation around the mean. The dataset
consists of elements, the MADmean is stated as mean of the absolute
deviation from the data’s mean.

8. MADmedian: MADmedian refers to Mean Absolute Deviation around the median. The
dataset consists of elements, the MADmedain is stated as mean absolute
deviation from the data’s median.

Upper utilization threshold (Tu) [1] is expressed as

Tu = 1-s. (Host overload detection algorithm)
Where,
x s€R+ is a parameter defines the system VM consolidation.
o Lower s, the lower energy consumption the higher the SLA violations caused by the
consolidation.
o By getting lower s, there will be lower energy consumption but higher the level of
SLA violation due to consolidation.

25
3.2.8.2 VM selection policies
The VM selection policies is implemented after detecting the overload host and selection
policies are applied such that VM is selected in a way the host is not in overloaded condition.
This process of selecting VM perform the overload host is done repeatedly until the host is not
in overloaded condition.

The following are the VM selection policies:

1. Maximum Requested RAM: The VM migration of the host is done which request
maximum amount of RAM between the VMs.

2. Minimum Requested RAM: The VM migration of the host is done which request
minimum amount of RAM between the VMs.

3. Maximum Utilization: The VM is migrated with the VM has the maximum CPU
utilization between the VMs which are present in the host.
CPUu(a) ≥ CPUu(b)
Where, CPUu(a) is VM a CPU utilization present in the host.

4. Minimum Utilization: The VM is migrated with the VM has the minimum CPU utilization
between the VMs which are present in the host.
CPUu(a) ≤ CPUu(b)

26
 Where, CPUu(a) is VM a CPU utilization present in the host.

3.2.8.3 Host Underloading Detection algorithm

For underloading detection, we have used a simple approach. We have used some of the
overloading detection algorithms to find the overloaded host from the host list in the datacenter
to migrate VMs from the overloaded host. This is done using some algorithms to detect
overloaded hosts.

While in underloading detection, we check the CPU utilization and consider the minimum CPU
utilization host as the underloaded host and will find a suitable host to migrate the VM from this
underloaded host and transfer the underloaded host is changed to sleep mode. If there is no
chance of migration of all VM s in the detected underloaded host, the host is maintained in
active mode until further migration of VMs are done.

3.2.8.4 VM placement
The VM placement can be viewed as a bin packing problem with flexible bin sizes and prices,
where physical node is represented by bins, available node CPU capacities as bin sizes and the
prices correlate to the power consumption by the nodes [1].

Paper [1] proposed Power Aware Best Fit Decreasing in which the VMs are sorted in the
decreasing order of the CPU utilization and allocate VM to a host based on the host that
provides least increase of the power consumption to allocate the VM. This process is repeatedly
done until all the VMs are allocated in their suitable hosts.

27
An example is explained in the following
Scenario: In the first, the VMs are sorted in the decreasingly order based CPU utilization for a
host. The VM migration to the host in the host list (VMs occupy the host with the minimum
power).
For example:

Case power consumption of hosts

Before After allocation of VM to host
allocation of power Increase of power Remaining power of
VM to host consumption consumption each host
1. Host1 = 20kw VM to host1 = 22kw 22kw-20kw = 2kw Host1 = 18kw
2. Host2 = 30kw VM to host2 = 33kw 33kw-30kw = 3kw Host2 = 17kw
3. Host3 = 40kw VM to host3 = 41kw 41kw-40kw = 1kw Host3 =19kw
Table 4: Power consumptions
In the above table we have three different cases, where each host power consumptions were
presented both before and after allocation of VM to hosts. Where in column 2, before allocation
of VM to host is the initial power available in the hosts. In column 3, 4 and 5 are power
consumption of the hosts after allocation of VM to it. Here, column 3 describes the power
consumption of each host when a VM is allocated in it, column 4 describes amount of increase
in power consumption of the hosts after allocation of VM to it, and column 5 describes
remaining power available in the hosts after allocation of VM to it.

By considering the PABFD algorithm, it takes the case3 due to increase in power consumption
is minimum and VMs are allocated to the Host3 with remaining power of 19kw.

From the above consideration of the algorithm of VM placement in [1] we have proposed an
two new VM placement based on concept of bin packing problem. The proposed VM placement
algorithms are: Modified Worst Fit Decreasing (MWFD) and Modified Second Worst Fit
Decreasing (MSWFD).

1. Modified Worst Fit Decreasing (MWFD)

In this VM placement algorithm, the VMs are sorted in the decreasing order based on
CPU utilization of a host and VM is allocated to a host which takes maximum increase in
power consumption. This is different and opposite from the PABFD algorithm.

From the above example, if we consider MWFD, case2 will be taken due to increase in
power consumption is maximum and VMs are allocated to the Host2 with remaining
power of 17kw.

28
 2. Modified Second Worst Fit Decreasing (MSWFD):
The second proposed model is Modified Second Worst Fit Decreasing (MSWFD). In this
VMs are selected that need to be allocated in the host which have the second maximum
increase in power consumption.

From the above example, if we consider MSWFD, case1 will be taken due to increase in
power consumption is second maximum and VMs are allocated to the Host1 with
remaining power of 18kw.

3.2.9 Dependent and Independent variables

The following are the dependent and independent variables of our research:

Independent Variable:
The variables that can be revised, changed and controlled in an experiment are called
independent variables. If the independent variables are changed it will have an effect on the
dependent variables [42]. The choices of these variables are done in reverse order or
simultaneously which includes the measurement scaled, the range and definite levels at which
test will be made [42].

In our experiment, the independent variables are all proposed algorithms (described in Section
3.2.8), existing algorithm (described in Section 3.2.7) and workloads (described in Section
3.2.11), since study mainly focuses on the performance of the algorithms. The proposed
algorithm should result in minimum power consumption while meeting SLAs and this depends
on the metric ESV (described in Section 3.2.5) i.e. the proposed algorithm is optimal when the
ESV is minimum than the existing algorithm and it is not optimal when the ESV is maximum
than the existing algorithm. If any changes done to the proposed and existing algorithms, it has
reflection on the metrics used and the experiment results. Hence in our experiment, those
algorithms are classified as independent variables. The performance of the algorithm is based
on the metrics used, i.e. ESV. The independent variable has 2 levels and the levels are
maximum ESV and Minimum ESV.

Dependent Variable:
Those variables which has effect when there are changes in the independent variable are called
dependent variables [42]. In our experiment, the dependent variables are the performance
metrics (described in the Sections 3.2.2, 3.2.3, 3.2.4 and 3.2.5). The performance metrics

29
considered, i.e., Energy Consumption, (PDM) Performance Degradation due to migration, SLA
violation Time per Active Host (SLATAH), SLAV and ESV.

In Sections 3.2.7 and 3.2.8 we have presented the existing and proposed heuristics. For the
experiment results, we have considered the combination of the heuristic algorithms, i.e. eight
descriptive statistics host overloading detection algorithms based on dynamic threshold and four
VM selection. Total combinations of 32 heuristics are extracted in the results and compared
with the existing heuristics. We have proposed a single dynamic underload detection algorithm
(described in the Section 3.2.8.3) and it is used for all 32 combinations. These 32 combinations
with two proposed VM placement algorithm, i.e. MWFD (Modified Worst Fit Decreasing) and
MSWFD were analyzed in the following Chapter 4.

Table 5: 32 combinations (host overload detection and VM selection algorithms)

3.2.10 Experimental Setup

For our thesis work, we used the CloudSim for simulation, modeling and evaluation of our
proposed heuristic algorithms and existing algorithms. According to description given in system
model, we initially simulated a datacenter with 50 physical machines/hosts in it. We divided the
hosts into two types, i.e. half hosts of type 1 and half hosts of type 2, and these are multi-core
hosts as described in Section 3.2.1. We used 4 types of VM which corresponds to the instance
type of Amazon EC2. VMs used are single core architecture as described in Section 3.2.1.
Characteristics of hosts and VMs are described in table 6 and 7.

Host Cores MIPS RAM

(Type) (GB)
1 2 1860 4
2 2 2660 4
Table 6: Host characteristics

Host Cores MIPS RAM

(Type)
1 1 2500 870
2 1 2000 1740
3 1 1000 1740
4 1 500 613
Table 7: VM characteristics

30
3.2.11 Workload Dataset
The workload used for this thesis work is Planetlab workload which is provided as a part of
CoMon project [1] [43]. This workload consists of real VMs CPU utilization collected from
different number of servers that are present all-over the world, the data is collected day-wise
manner for a time interval of 5 minutes. The main reason for choosing this Planetlab workload
is to make our simulation applicable for other real Cloud environments, as it contains the real
dataset of VMs workload traces.

We have used 10 different workloads from this real dataset on our proposed algorithms and also
for existing algorithms, to get accurate results. The chosen workloads are: Workload 1 with
10VMs; Workload 2 with 20VMs; Workload 3 with 30VMs; Workload 4 with 40VMs;
Workload 5 with 50VMs; Workload 6 with 60VMs; Workload 7 with 70VMs; Workload 8 with
80VMs; Workload 9 with 90VMs and Workload 10 with 100VMs; and number of hosts is
common for all workloads, representing three different scenarios (small, medium and heavy
loads) i.e. 50 hosts as stated in experimental Setup (3.2.9).

Motivation: From the researched articles, all experiments were done using single workload and
analysis of the results were done. To have the better conclusion multiple workloads (10
workloads) has been considered. The range for the workload is considered by different
researchers is in between the 1.2 VMs to 1.5 VMs for 1 host. In my thesis work, 1 host takes
minimum of 0.2 VMs to maximum 2 VMs, the maximum limit depends on our 2-core
architecture and memory bounds. So, the maximum limit for our experiment is 2 VMs, this limit
can be varied depending the multi-core architecture we consider.

3.2.12 Hypothesis Testing

Hypothesis is a working statement which is supposed to be consistent with given data and is
neither proved false nor verified to be true. [44] [45] [46] [47] [48]

Alternative Hypothesis [44] [45] [46] [47] [48]:

Alternative hypothesis is a type of hypothesis that says that there is significant relationship
between the variables in the hypothesis. It is the hypothesis that the researcher wants to prove. It
is a statement created by researchers when they speculate upon the outcome of an experiment. It
is usually represented by H1. It is a statement which a researcher is testing. In the alternative
hypothesis, the researcher tries to prove that the pattern of data which is observed in the
experiment is not due to chance.

Null Hypothesis [44] [45] [46] [47] [48]:

Null Hypothesis is a type of hypothesis that says that there is no statistical significance between
the variables in the hypothesis and the researcher is trying to disprove it. It is generally denoted

31
by H0 and states exact opposite of what researcher predicts or expects. When we do find that a
relationship exists then null hypothesis is rejected and accept the alternative hypothesis.

T-test [49]:
T-test often called Student's T-test is used to compare two different set of values. This test
normally compares the mean of two samples and it uses and standard deviations of two samples
to make a comparison. There are many types of t test:
x One sample t-test [49]: To compare the mean of a population with a theoretical value.
x Independent t-test/unpaired two sample t-test: to compare the mean of two unrelated groups
of samples.
x Paired t-test [49]: to compare the mean of two related samples. In this thesis, the paired t-
test is calculates using the SPSS tool.

3.2.13 Cohen's d:
“Cohen’s d is an Effect size used to specify the systematic difference between two means” [50]
[51] [52]. Cohen’s d works best for large samples. As an effect size, Cohen’s d is typically used
to represent the magnitude of differences between two or more groups on a given variable, with
larger values representing a greater differentiation between the groups on that variable.

Cohen’s d Effect /size = (M1-M2)/SDpooled

M1: Mean of First group

M2: Mean of Second group
SDpooled: pooled standard deviation of both groups

Effect Size [50] [51] [52]: Effect size is a statistical concept that measures the strength of the
relationship between two variables on a numerical scale. Effect size is an important tool in
reporting and interpreting effectiveness.

Cohen’s d effect size values:

Greater than 0.2-represent small effect
Greater than 0.5-represent medium effect
Greater than 0.8-represent large effect

32
4 RESULTS AND ANALYSIS
Using the workloads described in the above Section 3.2.11, 32 combinations of the heuristic
algorithms which are stated in the Table 5 were simulated. The 32 combinations are
combination of the 8 host overload algorithms and 4 VM Selection algorithms using the VM
placement algorithms (MWFD, MSWFD).
4.1 Modified Worst Fit Decreasing Algorithm (MWFD)
4.1.1 Results of algorithms under 10 different workloads
In this algorithm, the results were obtained using the combinations of heuristics which were
explained in the experiment method, there are total of 32 heuristics combinations (explained in
the Table 5) i.e. from host overload detection and VM Selection using MWFD VM placement
algorithm. Here 10 different workloads were considered which was explained in the above
Section 3.2.11.

The following figures represent the performance metrics (explained in the Sections 3.2.2, 3.2.3,
3.2.4, 3.2.5) of the experiment and they are represented using box plots (Box plots distinguish
samples using the 25th, 50th and 75th percentiles, they are represented as lower quartile (Q1),
median (m or Q2) and upper quartile (Q3) [53]).

Figure 5: Energy consumption of MWFD

The above Figure 5 represents the energy consumption of 32 heuristic algorithm combinations
(explained in the Table 5) using MWFD VM placement algorithm under 10 different workloads.

33
 Figure 6: No. of VM migrations of MWFD

The above Figure 6 represents the number of VM migrations took place for each of 32 heuristic
algorithm combinations (explained in the Table 5) using MWFD VM placement algorithm
under 10 different workloads.

Figure 7: Performance Degradation due to VM Migration (PDM) of MFWD

The above Figure 7 represents the performance degradation due to migration of VM for each of
32 heuristic algorithm combinations (explained in the Table 5) using MWFD VM placement
algorithm under 10 different workloads.

34
 Figure 8: SLA violation Time per Active Host (SLATAH) of MFWD

The above Figure 8 represents the SLA violation occurred per active host for each of 32
heuristic algorithm combinations (explained in the Table 5) using MWFD VM placement
algorithm under 10 different workloads.

Figure 9: SLA Violation of MWFD

The above Figure 9 represents the overall SLA violation of 32 heuristic algorithm combinations
(explained in the Table 5) using MWFD VM placement algorithm under 10 different workloads.

35
 Figure 10 : ESV of MWFD

The above Figure 10 represents the ESV, i.e. overall energy consumption and SLA violation for
each of 32 heuristic algorithm combinations (explained in the Table 5) using MWFD VM
placement algorithm under 10 different workloads.

4.1.2 Analysis:
We have considered ESV metric (described in the Section 3.2.5) as a performance metric to
decide the best algorithm of the host overload detection and VM selection. ESV metric is
selected as it is combination of all metrics, i.e. Energy consumption and SLA Violation, as SLA
Violation is obtained from the combination of PDM (Performance Degradation due to VM
Migration) and SLATAH (SLA violation Time per Active Host).

According to Ryan-Joiner’s normality test, the values of the ESV (overall energy consumption
and SLA violation) metric we got from the 32 combinations of the heuristic algorithm obey the
normal distribution with p-value > 0.1.

The analysis is done in two different steps to select the best algorithm from the VM selection
and combination of the host overload detection and VM selection algorithm. The steps are
defined in the following manner:
Step 1: A six paired t-test is conducted to select the best VM selection algorithm (which takes
the less energy consumption and have less SLA violation).
Step 2: A 28 paired t-test is conducted to select the best algorithm from the combination of host
overload detection and VM selection.

4.1.3 Step 1: VM selection algorithm

Here, six paired t-test is conducted to select the best VM selection algorithm. The paired t-test is
conducted by comparing the means of ESV metrics of the proposed algorithms of VM selection.
The VM selection algorithm which has the minimum ESV mean value is considered as the best
VM selection algorithm, i.e. VM selection algorithm which takes the less energy consumption
and have less SLA violation. This is represented in the Table 8.

In this analysis, we have considered the null hypothesis and alternative hypothesis to detect
whether the algorithms have statistical significance difference between the mean values. It is
important to have the significance difference between the ESV means of the VM selection

36
algorithm to decide the best algorithm. The following are the null hypothesis and alternative
hypothesis of this analysis:

Null Hypothesis: Here we consider the null hypothesis as the differences in mean between the
Algorithm 1 and Algorithm 2 is equal to zero (μ = 0).
Alternative Hypothesis: The difference in mean between the algorithms is not equal to zero,
i.e. μ≠0.

Algorithm (ESV) Mean 95 % confidence Interval (CI) P-value

1 2 Difference Lower Upper
MinR MaxR 0.04267 0.01632 0.06902 < 0.05
(0.043857) (0.001186)
MaxU MaxR 0.04477 0.01663 0.07290 < 0.05
(0.045955) (0.001186)
MinU MaxR 0.03436 0.01190 0.05682 < 0.05
(0.035542) (0.001186)
MaxU MinR 0.00210 0.00076 0.00344 < 0.05
(0.045955) (0.043857)
MinR MinU 0.00832 0.00063 0.01726 < 0.05
(0.043857) (0.035542)
MaxU MinU 0.01041 0.00767 0.02849 < 0.05
(0.045955) (0.035542)
Table 8 VM selection algorithms using six paired t-test

From the above table, the Null hypothesis is rejected since the mean difference between the
algorithms are not equal to zero and p<0.05. Therefore, the alternative hypothesis is considered.
From these results, we got MaxR perform better than the other algorithms with less value of
ESV which has p<0.05.

4.1.4 Step 2: Host Overload detection algorithm

To select the best algorithm from the combination of the host overload detection and VM
selection algorithm, i.e. combination algorithm which takes less energy consumption and less
SLA Violation, we have examined the eight combinations of the host overload algorithms with
the best VM selection algorithm (MaxR). In this, 28 paired t-test is conducted to select the best
algorithm from the combination of host overload detection and VM selection. This is
represented in the following Table 9. The test is conducted by the comparing the means of ESV
metric of the algorithms.

In this analysis, we have considered the null hypothesis and alternative hypothesis to detect
whether the algorithms have statistical significance difference between the mean values. It is
important to have the significance difference between the ESV means of the VM selection
algorithm to decide the best algorithm. The following are the null hypothesis and alternative
hypothesis of this analysis:

Null Hypothesis: Here we consider the null hypothesis as the differences in mean between the
Algorithm 1 and Algorithm 2 is equal to zero (μ = 0).
Alternative Hypothesis: The difference in mean between the algorithms is not equal to zero,
i.e. μ≠0.

Algorithm (ESV) Mean 95 % confidence Interval P-value

Difference (CI)
1 2 Lower Upper
Median_MaxR Mean_MaxR 0.00007 0.00000 0.00025 < 0.05

37
(0.002973) (0.002905)
Var_MaxR Mean_MaxR 0.00254 0.00019 0.00489 < 0.05
(0.005448) (0.002905)
Mean_MaxR Std_MaxR 0.00198 0.00052 0.00343 < 0.05
(0.002905) (0.000927)
Mean_MaxR Mode_MaxR 0.00185 0.00038 0.00331 < 0.05
(0.002905) (0.001058)
Mean_MaxR Range_MaxR 0.00172 0.00007 0.00337 < 0.05
(0.002905) (0.001186)
Mean_MaxR MADmean_Ma 0.00197 0.00048 0.00346 < 0.05
(0.002905) xR (0.000935)
Mean_MaxR MADmedian_M 0.00221 0.00066 0.00376 < 0.05
(0.002905) axR (0.000693)
Var_MaxR Median_MaxR 0.00248 0.00015 0.00480 < 0.05
(0.005448) (0.002973)
Median_MaxR Std_MaxR 0.00205 0.00045 0.00364 < 0.05
(0.002973) (0.000927)
Median_MaxR Mode_MaxR 0.00192 0.00020 0.00363 < 0.05
(0.002973) (0.001058)
Median_MaxR Range_MaxR 0.00179 0.00007 0.00350 < 0.05
(0.002973) (0.001186)
Median_MaxR MADmean_Ma 0.00204 0.00048 0.00359 < 0.05
(0.002973) xR (0.000935)
Median_MaxR MADmedian_M 0.00228 0.00054 0.00402 < 0.05
(0.002973) axR (0.000693)
Var_MaxR Std_MaxR 0.00452 0.00164 0.00740 < 0.05
(0.005448) (0.000927)
Var_MaxR Mode_MaxR 0.00439 0.00133 0.00745 < 0.05
(0.005448) (0.001058)
Var_MaxR Range_MaxR 0.00426 0.00166 0.00686 < 0.05
(0.005448) (0.001186)
Var_MaxR MADmean_Ma 0.00451 0.00167 0.00735 < 0.05
(0.005448) xR (0.000935)
Var_MaxR MADmedian_M 0.00476 0.00168 0.00783 < 0.05
(0.005448) axR (0.000693)
Mode_MaxR Std_MaxR 0.00013 0.00064 0.00038 < 0.05
(0.001058) (0.000927)
Range_MaxR Std_MaxR 0.00026 0.00005 0.00047 < 0.05
(0.001186) (0.000927)
MADmean_Ma Std_MaxR 0.00001 0.00000 0.00009 < 0.05
xR (0.000935) (0.000927)
Std_MaxR MADmedian_M 0.00023 0.00003 0.00043 < 0.05
(0.000927) axR (0.000693)
Range_MaxR Mode_MaxR 0.00013 0.00009 0.00017 < 0.05
(0.001186) (0.001058)
Mode_MaxR MADmean_Ma 0.00012 0.00004 0.00020 < 0.05
(0.001058) xR (0.000935)
Mode_MaxR MADmedian_M 0.00036 0.00000 0.00073 < 0.05
(0.001058) axR (0.000693)
Range_MaxR MADmean_Ma 0.00025 0.00002 0.00048 < 0.05
(0.001186) xR (0.000935)
Range_MaxR MADmedian_M 0.00049 0.00017 0.00081 < 0.05
(0.001186) axR (0.000693)

38
 MADmean_Ma MADmedian_M 0.00024 0.00019 0.00028 < 0.05
xR (0.000935) axR (0.000693)
Table 9: Host Overload detection algorithms with MaxR using 28 paired t-test

From the above table, null hypothesis is rejected since the mean difference is not equal to Zero
and p<0.05. It satisfies the both p-value <0.05 and μ≠ 0. Therefore, the alternative hypothesis is
considered. From these results, we got MADmedian_MaxR perform better than the other
combinations of the host overload algorithm with VM Selection (MaxR) algorithms with less
value of ESV which has p<0.05.

Based on the results, MADmedian_MaxR is the best resulted algorithm that takes the least
energy consumption and less SLAV when compared to other combinations of the host overload
algorithm with VM Selection (MaxR) using MWFD VM placement algorithm.

4.2 Modified Second Worst Fit Decreasing Algorithm (MSWFD)

4.2.1 Results of algorithms under 10 different workloads
In this algorithm, the results were obtained using the combinations of heuristics which were
explained in the experiment method, there are total of 32 heuristics combinations (explained in
the Table 5) i.e. from host overload detection and VM Selection using MSWFD VM placement
algorithm. Here 10 different workloads were considered which was explained in the above
Section 3.2.11.

The following Figures represent the performance metrics (explained in the Sections 3.2.2, 3.2.3,
3.2.4, 3.2.5) of the experiment and they are represented using box plots (Box plots distinguish
samples using the 25th, 50th and 75th percentiles, they are represented as lower quartile (Q1),
median (m or Q2) and upper quartile (Q3) [53]).

Figure 11: Energy consumption of MSWFD

The above Figure 11 represents the energy consumption of 32 heuristic algorithm combinations
(explained in the Table 5) using MSWFD VM placement algorithm under 10 different
workloads.

39
 Figure 12: No. of VM migrations of MSWFD

The above Figure 12 represents the number of VM migrations took place for each of 32
heuristic algorithm combinations (explained in the Table 5) using MSWFD VM placement
algorithm under 10 different workloads.

Figure 13: Performance Degradation due to Migration (PDM) of MSWFD

The above Figure 13 represents the performance Degradation due to Migration of VM for each
of 32 heuristic algorithm combinations (explained in the Table 5) using MSWFD VM
placement algorithm under 10 different workloads.

40
 Figure 14: SLA violation Active per Active Host (SLATAH) of MSWFD

The above Figure 14 represents the SLA violation occurred per active host for each of 32
heuristic algorithm combinations (explained in the Table 5) using MSWFD VM placement
algorithm under 10 different workloads.

Figure 15: SLA Violation of MSWFD

The above Figure 15 represents the overall SLA violation of 32 heuristic algorithm
combinations (explained in the Table 5) using MWFD VM placement algorithm under 10
different workloads.

41
 Figure 16 ESV of MSWFD

The above Figure 16 represents the ESV, i.e. overall energy consumption and SLA violation for
each of 32 heuristic algorithm combinations (explained in the Table 5) using MWFD VM
placement algorithm under 10 different workloads.

4.2.1 Analysis:
We have considered ESV metric (described in the Section 3.2.5) as a performance metric to
decide the best algorithm of the host overload detection and VM selection. ESV metric is
selected as it is combination of all metrics, i.e. Energy consumption and SLA Violation, as SLA
Violation is obtained from the combination of PDM (Performance Degradation due to VM
Migration) and SLATAH (SLA violation Time per Active Host).

According to Ryan-Joiner’s normality test, the values of the ESV metric we got from the 32
combinations of the heuristic algorithm obey the normal distribution with p-value > 0.1.

The analysis is done in two different steps to select the best algorithm from the VM selection
and combination of the host overload detection and VM selection algorithm. The steps are
defined in the following manner:

Step 1: A six paired t-test is conducted to select the best VM selection algorithm (which takes
the less energy consumption and have less SLA violation).
Step 2: A 28 paired t-test is conducted to select the best algorithm from the combination of host
overload detection and VM selection.

4.2.2 Step 1: VM Selection Algorithm

Here, six paired t-test is conducted to select the best VM selection algorithm. The paired t-test is
conducted by comparing the means of ESV metrics of the proposed algorithms of VM selection.
The VM selection algorithm which has the minimum ESV mean value is considered as the best
VM selection algorithm, i.e. VM selection algorithm which takes the less energy consumption
and have less SLA violation. This is represented in the Table 10.

42
In this analysis, we have considered the null hypothesis and alternative hypothesis to detect
whether the algorithms have statistical significance difference between the mean values. It is
important to have the significance difference between the ESV means of the VM selection
algorithm to decide the best algorithm. The following are the null hypothesis and alternative
hypothesis of this analysis:

Null Hypothesis: Here we consider the null hypothesis as the differences in mean between the
Algorithm 1 and Algorithm 2 is equal to zero (μ = 0).
Alternative Hypothesis: The difference in mean between the algorithms is not equal to zero,
i.e. μ≠0.

Algorithm (ESV) Mean 95 % confidence Interval (CI) P-value

1 2 Difference Lower Upper
MinR MaxR 0.05717 0.01791 0.09643 < 0.05
(0.05961) (0.002438)
MaxU MaxR 0.06259 0.02246 0.10271 < 0.05
(0.065027) (0.002438)
MinU MaxR 0.05162 0.01501 0.08823 < 0.05
(0.054059) (0.002438)
MaxU MinR 0.00542 0.00376 0.00708 < 0.05
(0.065027) (0.05961)
MinR MinU 0.00555 0.00293 0.00817 < 0.05
(0.05961) (0.054059)
MaxU MinU 0.01097 0.00645 0.01549 < 0.05
(0.065027) (0.054059)
Table 10: VM selection algorithms using six paired t-test

From the above table, the null hypothesis is rejected since all the mean difference is not equal to
zero and the p-value < 0.05. Therefore, the alternative hypothesis is considered. From these
results, we got MaxR perform better than the other algorithms with less value of ESV which has
p<0.05.

4.2.3 Step 2: Host Overload detection algorithm

To select the best algorithm from the combination of the host overload detection and VM
selection algorithm, i.e. combination algorithm which takes less energy consumption and less
SLA Violation, we have examined the eight combinations of the host overload algorithms with
the best VM selection algorithm (MaxR). In this, 28 paired t-test is conducted to select the best
algorithm from the combination of host overload detection and VM selection. This is
represented in the following Table 11. The test is conducted by the comparing the means of
ESV metric of the algorithms.

In this analysis, we have considered the null hypothesis and alternative hypothesis to detect
whether the algorithms have statistical significance difference between the mean values. It is
important to have the significance difference between the ESV means of the VM selection
algorithm to decide the best algorithm. The following are the null hypothesis and alternative
hypothesis of this analysis:

Null Hypothesis: Here we consider the null hypothesis as the differences in mean between the
Algorithm 1 and Algorithm 2 is equal to zero (μ = 0).
Alternative Hypothesis: The difference in mean between the algorithms is not equal to zero,
i.e. μ≠0.

43
 Algorithm (ESV) Mean 95 % confidence Interval P-
Difference (CI) value
1 2 Lower Upper
Mean_MaxR(0.00 Median_MaxR(0. 0.00001 0.00000 0.00002 < 0.05
1619) 001611)
Var_MaxR Mean_MaxR(0.00 0.00341 0.00084 0.00597 < 0.05
(0.005024) 1619)
Mean_MaxR(0.00 Std_MaxR 0.00011 0.00001 0.00020 < 0.05
1619) (0.001504)
Mode_MaxR Mean_MaxR(0.00 0.00027 0.00015 0.00039 < 0.05
(0.001886) 1619)
Range_MaxR Mean_MaxR(0.00 0.00082 0.00065 0.00112 < 0.05
(0.002438) 1619)
Mean_MaxR(0.00 MADMean_Max 0.00018 0.00004 0.00032 < 0.05
1619) R (0.001434)
Mean_MaxR(0.00 MADMedian_Ma 0.00004 0.00001 0.00005 < 0.05
1619) xR (0.001577)
Var_MaxR Median_MaxR 0.00341 0.00082 0.00601 < 0.05
(0.005024) (0.001611)
Median_MaxR Std_MaxR 0.00011 0.00006 0.00016 < 0.05
(0.001611) (0.001504)
Mode_MaxR Median_MaxR 0.00027 0.00013 0.00041 < 0.05
(0.001886) (0.001611)
Range_MaxR Median_MaxR 0.00083 0.00017 0.00148 < 0.05
(0.002438) (0.001611)
Median_MaxR MADMean_Max 0.00018 0.00003 0.00032 < 0.05
(0.001611) R (0.001434)
Median_MaxR MADMedian_Ma 0.00003 0.00000 0.00001 < 0.05
(0.001611) xR (0.001577)
Var_MaxR Std_MaxR 0.00352 0.00112 0.00592 < 0.05
(0.005024) (0.001504)
Var_MaxR Mode_MaxR 0.00314 0.00110 0.00517 < 0.05
(0.005024) (0.001886)
Var_MaxR Range_MaxR 0.00259 0.00058 0.00459 < 0.05
(0.005024) (0.002438)
Var_MaxR MADMean_Max 0.00359 0.00098 0.00620 < 0.05
(0.005024) R (0.001434)
Var_MaxR MADMedian_Ma 0.00345 0.00060 0.00629 < 0.05
(0.005024) xR (0.001577)
Mode_MaxR Std_MaxR 0.00038 0.00007 0.00068 < 0.05
(0.001886) (0.001504)
Range_MaxR Std_MaxR 0.00093 0.00033 0.00152 < 0.05
(0.002438) (0.001504)
Std_MaxR MADMean_Max 0.00007 0.00001 0.00013 < 0.05
(0.001504) R (0.001434)
MADmedian_Ma Std_MaxR 0.00007 0.00002 0.00014 < 0.05
xR (0.001577) (0.001504)
Range_MaxR Mode_MaxR 0.00055 0.00047 0.00063 < 0.05
(0.002438) (0.001886)
Mode_MaxR MADMean_Max 0.00045 0.00028 0.00061 < 0.05
(0.001886) R (0.001434)
Mode_MaxR MADMedian_Ma 0.00031 0.00015 0.00046 < 0.05
(0.001886) xR (0.001577)

44
 Range_MaxR MADMean_Max 0.00100 0.00026 0.00173 < 0.05
(0.002438) R (0.001434)
Range_MaxR MADMedian_Ma 0.00086 0.00047 0.00124 < 0.05
(0.002438) xR (0.001577)
MADMedian_Ma MADMean_Max 0.00014 0.00008 0.00019 < 0.05
xR (0.001577) R (0.001434)
Table 11: Host Overload detection algorithms with MaxR using 28 paired t-test

From the above table, the null hypothesis is rejected since the mean difference is not equal to
Zero and p<0.05. Therefore, the alternative hypothesis is considered. It satisfies the both p-value
<0.05 and μ≠ 0. From these results, we got MADmean_MaxR perform better than other
combinations of the host overload algorithm with VM Selection (MaxR) algorithms with less
value of ESV which has p<0.05.

Based on the results, MADmean_MaxR is the best resulted algorithm that takes the least energy
consumption and less SLAV when compared to other combinations of the host overload
algorithm with VM Selection (MaxR) using MWFD VM placement algorithm.

4.3 Comparison of results for MWFD, MSWFD and PABFD

In this section, the comparison of the proposed heuristic algorithms (Modified Worst Fit Decreasing
and Modified Second Worst Fit Decreasing) with the existing algorithm Power Aware Best Fit
Decreasing algorithm (described in the section 3.2.7) is done and based on the results the
combinations of the host overload detection and VM selection algorithms with VM placement
algorithms which shows the better results has been selected. The results are described in the
following sections 4.3.1 and 4.3.2.

4.3.1 MWFD is compared with the existing algorithm PABFD

The alternative hypothesis is considered as the ESV (overall energy consumption and SLA
Violation) metric of the proposed algorithm should be less than the ESV metric of the existing
algorithm, so that the best proposed heuristic algorithms of host overload detection and VM selection
using MWFD (Modified Worst Fit Decreasing) with minimum energy consumption and less SLA
violation can be obtained.
Null hypothesis: The ESV metric of the existing algorithm (PABFD) is less than the proposed
algorithm (MWFD) i.e. μ≤0.
Alternative Hypothesis: The ESV metric of the existing algorithm (PABFD) is greater than the
proposed algorithm i.e. μ > 0.

Algorithm (ESV) Mean 95 % confidence t P-value Cohens d

Difference Interval (CI) effect size
Old Proposed Lower Upper
IQR_MMT Mean_MaxR 0.03829 0.01577 0.06081 3.847 < 0.001
(0.041195) (0.002905) 1.2165
IQR_MMT Mean_MinR 0.01296 -0.00585 0.03177 1.558 > 0.05
(0.041195) (0.028239) 0.492734
IQR_MMT Mean_MaxU 0.01020 -0.00760 0.02800 1.296 > 0.05
(0.041195) (0.030995) 0.409864
IQR_MMT Mean_MinU 0.02192 0.00271 0.04112 2.582 < 0.001
(0.041195) (0.019279) 0.816373
IQR_MMT Median_MaxR 0.03822 0.01601 0.06043 3.893 < 0.001
(0.041195) (0.002973) 1.231201
IQR_MMT Median_MinR 0.01023 -0.00647 0.02693 1.385 > 0.05
(0.041195) (0.030966) 0.438088

45
IQR_MMT Median_Max 0.01015 -0.00545 0.02575 1.472 >0.05
(0.041195) U (0.031043) 0.465483
IQR_MMT Median_MinU 0.01630 -0.00339 0.03599 1.873 < 0.001
(0.041195) (0.024892) 0.592357
IQR_MMT Var_MaxR 0.03575 0.01449 0.05701 3.804 < 0.001
(0.041195) (0.005448) 1.202893
IQR_MMT Var_MinR 0.02563 0.01067 0.04059 3.876 < 0.001
(0.041195) (0.015564) 1.225583
IQR_MMT Var_MaxU 0.02442 0.00856 0.04028 3.483 < 0.001
(0.041195) (0.016773) 1.101401
IQR_MMT Var_MinU 0.02315 0.01061 0.03568 4.178 < 0.001
(0.041195) (0.018049) 1.321248
IQR_MMT Std_MaxR 0.04027 0.01760 0.06293 4.019 < 0.001
(0.041195) (0.000927) 1.27102
IQR_MMT Std_MinR 0.02763 0.01196 0.04331 3.987 < 0.001
(0.041195) (0.015564) 1.260835
IQR_MMT Std_MaxU 0.02809 0.01157 0.04460 3.847 < 0.001
(0.041195) (0.013109) 1.216527
IQR_MMT Std_MinU 0.02911 0.01264 0.04559 3.997 < 0.001
(0.041195) (0.012083) 1.263943
IQR_MMT Mode_MaxR 0.04014 0.01738 0.06290 3.989 < 0.001
(0.041195) (0.001058) 1.261411
IQR_MMT Mode_MinR -0.06961 -0.10295 - -4.722 < 0.001
(0.041195) (0.110803) 0.03626 -1.49335
IQR_MMT Mode_MaxU -0.05798 -0.08755 - -4.435 < 0.001
(0.041195) (0.099176) 0.02841 -1.4025
IQR_MMT Mode_MinU -0.05925 -0.09034 - -4.310 < 0.001
(0.041195) (0.100443) 0.02815 -1.363
IQR_MMT Range_MaxR 0.04001 0.01762 0.06240 4.042 < 0.001
(0.041195) (0.001186) 1.278331
IQR_MMT Range_MinR -0.00266 -0.01359 0.00826 -0.551 > 0.05
(0.041195) (0.043857) -0.17431
IQR_MMT Range_MaxU -0.00476 -0.02164 0.01212 -0.638 > 0.05
(0.041195) (0.045955) -0.20168
IQR_MMT Range_MinU 0.00565 -0.00366 0.01496 1.374 > 0.05
(0.041195) (0.035542) 0.434373
IQR_MMT MADmean_M 0.04026 0.01756 0.06296 4.013 < 0.001
(0.041195) axR
(0.000935) 1.268993
IQR_MMT MADmean_M 0.03123 0.01417 0.04829 4.142 < 0.001
(0.041195) inR
(0.009961) 1.309717
IQR_MMT MADmean_M 0.02899 0.01235 0.04563 3.942 < 0.001
(0.041195) axU (0.0122) 1.246473
IQR_MMT MADmean_M 0.02892 0.01285 0.04499 4.072 < 0.001
(0.041195) inU
(0.012272) 1.287653
IQR_MMT MADmedain_ 0.04050 0.01769 0.06331 4.016 < 0.001
(0.041195) MaxR
(0.000693) 1.270059
IQR_MMT MADmedain_ 0.02971 0.01278 0.04665 3.969 < 0.001
(0.041195) MinR
(0.011483) 1.255045
IQR_MMT MADmedain_ 0.02952 0.01239 0.04665 3.898 < 0.001 1.232811

46
(0.041195) MaxU
(0.011675)
IQR_MMT MADmedain_ 0.02781 0.01111 0.04451 3.767 < 0.001
(0.041195) MinU
(0.013384) 1.191247
Table 12: Paired t-test of PABFD & MWFD

From the literature review, we have considered the existing algorithm PABFD (IQR_MMT). In
this section, we have conducted the 32 paired T-test to compare the ESV means of the 32
combinations of the proposed algorithm using MWFD and existing algorithm PABFD. This is
done to obtain the best algorithms from the combination of the host overload detection and VM
Selection using the VM Placement (MWFD).

We have calculated the effect size since effect size is to measure the strength and interpretation
of the effectiveness between the variables. It represents the value as 0.2 for the small effect, 0.5
for the medium effect and 0.8 for the large effect.

From the results, 22 combinations (host overload and VM Selection) of the proposed algorithm
using MWFD rejects the Null hypothesis. It shows the p-value <0.001, t- value and mean
difference are positive values. There are 10 combinations which did not reject the NULL
hypothesis. They have not rejected the Null hypothesis because of the p-value is not less than
0.05, t- value and mean values are negative. The 10 combinations of proposed algorithms which
did not reject the null hypothesis are mentioned in the following Table 13:
Mean_MaxR Mean_MaxU Median_MinR Median_MaxU Mode_MinR
Mode_MaxU Mode_MinU Range_MinR Range_MaxU Range_MinU
Table 13: 10 combination algorithms rejected the alternative hypothesis (null hypothesis
accepted)
Best resulted 22 combinations (host overload and VM Selection) of the proposed algorithms
using MWFD are mentioned in the following Table 14:

Mean_MaxR Mean_MinU Median_MaxR Median_MinU

Var_MaxR Var_MinR Var_maxU Var_MinU
Std_MaxR Std_MinR Std_MaxU Std_MinU
Mode_MaxR Range_MaxR MADmean_MaxR MADmean_MinR
MADmean_MaxU MADmean_MinU MADmedian_MaxR MADmedian_MinR
MADmedian_MaxU MADmedian_MinU
Table 14: 22 combination algorithms accepted the alternative hypothesis (null hypothesis
rejected)

By this analysis, the remaining 22 combinations of host overload detection and VM Selection
using MWFD perform better than the existing algorithm, i.e. the energy consumption is
minimum and minimum SLA violations. These algorithms have p-value < 0.001 and the effect
size is greater than 0.8, it has large effect.

4.3.2 MSWFD is compared with the existing algorithm PABFD

The alternative hypothesis is considered as the ESV (overall energy consumption and SLA
Violation) metric of the proposed algorithm should be less than the ESV metric of the existing
algorithm, so that the best proposed heuristic algorithms of host overload detection and VM selection
using MSWFD (Modified Second Worst Fit Decreasing) with minimum energy consumption and
less SLA violation can be obtained.

47
Null hypothesis: The ESV metric of the existing algorithm (PABFD) is less than the proposed
algorithm (MWFD) i.e. μ≤0.
Alternative Hypothesis: The ESV metric of the existing algorithm (PABFD) is greater than the
proposed algorithm i.e. μ > 0.

Algorithm (ESV) Mean 95 % confidence t P-value Cohens d

Difference Interval (CI) effect size
Old Proposed Lower Upper
IQR_MMT Mean_MaxR 0.03958 0.01717 0.06198 3.995 < 0.001
(0.041195) (0.001619) 1.263487
IQR_MMT Mean_MinR 0.00074 -0.01984 0.02132 0.082 > 0.05
(0.041195) (0.040453) 0.025802
IQR_MMT Mean_MaxU -0.00359 -0.02378 0.01661 -0.402 > 0.05
(0.041195) (0.044781) -0.12701
IQR_MMT Mean_MinU 0.01288 -0.00742 0.03319 1.435 > 0.05
(0.041195) (0.028311) 0.453837
IQR_MMT Median_MaxR 0.03958 0.01688 0.06229 3.943 < 0.001
(0.041195) (0.001611) 1.24702
IQR_MMT Median_MinR -0.00291 -0.02510 0.01928 -0.297 > 0.05
(0.041195) (0.044106) -0.09386
IQR_MMT Median_Max -0.00477 -0.02554 0.01600 -0.519 >0.05
(0.041195) U (0.045964) -0.16425
IQR_MMT Median_MinU 0.00883 -0.01351 0.03117 0.894 > 0.05
(0.041195) (0.032364) 0.282803
IQR_MMT Var_MaxR 0.03617 0.01506 0.05728 3.876 < 0.001
(0.041195) (0.005024) 1.225801
IQR_MMT Var_MinR 0.02522 0.01034 0.04010 3.835 < 0.001
(0.041195) (0.015972) 1.212632
IQR_MMT Var_MaxU 0.02489 0.00924 0.04053 3.598 < 0.001
(0.041195) (0.01631) 1.137927
IQR_MMT Var_MinU 0.02339 0.00941 0.03736 3.786 < 0.001
(0.041195) (0.017808) 1.197141
IQR_MMT Std_MaxR 0.03969 0.01715 0.06223 3.984 < 0.001
(0.041195) (0.001504) 1.259764
IQR_MMT Std_MinR 0.02771 0.01135 0.04407 3.831 < 0.001
(0.041195) (0.013489) 1.21145
IQR_MMT Std_MaxU 0.02777 0.01057 0.04498 3.652 < 0.001
(0.041195) (0.01342) 1.154726
IQR_MMT Std_MinU 0.02731 0.01123 0.04340 3.841 < 0.001
(0.041195) (0.01388) 1.214541
IQR_MMT Mode_MaxR 0.03931 0.01690 0.06172 3.968 < 0.001
(0.041195) (0.001886) 1.25485
IQR_MMT Mode_MinR -0.09509 -0.13854 - -4.951 < 0.001
(0.041195) (0.136281) 0.05164 -1.56551
IQR_MMT Mode_MaxU -0.09127 -0.13401 - -4.831 < 0.001
(0.041195) (0.132468) 0.04853 -1.52769
IQR_MMT Mode_MinU -0.07814 -0.11783 - -4.453 < 0.001
(0.041195) (0.119332) 0.03844 -1.40806
IQR_MMT Range_MaxR 0.03876 0.01622 0.06129 3.890 < 0.001
(0.041195) (0.002438) 1.230186
IQR_MMT Range_MinR -0.01842 -0.04003 0.00320 -1.927 < 0.001
(0.041195) (0.05961) -0.60942
IQR_MMT Range_MaxU -0.02383 -0.04722 - -2.306 < 0.001
(0.041195) (0.065027) 0.00045 -0.72908

48
IQR_MMT Range_MinU -0.01286 -0.03218 0.00645 -1.506 > 0.05
(0.041195) (0.054059) -0.47636
IQR_MMT MADmean_M 0.03976 0.01721 0.06232 3.988 < 0.001
(0.041195) axR
(0.001434) 1.26107
IQR_MMT MADmean_M 0.03012 0.01311 0.04713 4.006 < 0.001
(0.041195) inR
(0.011073) 1.266809
IQR_MMT MADmean_M 0.02787 0.01176 0.04398 3.913 < 0.001
(0.041195) axU
(0.013324) 1.23746
IQR_MMT MADmean_M 0.02780 0.01171 0.04388 3.910 < 0.001
(0.041195) inU
(0.013398) 1.236305
IQR_MMT MADmedain_ 0.03962 0.01676 0.06248 3.920 < 0.001
(0.041195) MaxR
(0.001577) 1.23966
IQR_MMT MADmedain_ 0.02964 0.01254 0.04674 3.921 < 0.001
(0.041195) MinR
(0.011554) 1.239916
IQR_MMT MADmedain_ 0.02996 0.01261 0.04732 3.906 < 0.001
(0.041195) MaxU
(0.011231) 1.235264
IQR_MMT MADmedain_ 0.02666 0.01014 0.04319 3.650 < 0.001
(0.041195) MinU
(0.014534) 1.15423
Table 15: Paired t-test of PABFD & MSWFD

From the literature review, we have considered the existing algorithm PABFD (IQR_MMT)
(described in the Section 3.2.7). In this section, we have conducted the 32 paired T-test to
compare the ESV means of the 32 combinations of the proposed algorithm using MSWFD and
existing algorithm PABFD. This is done to obtain the best algorithms from the combination of
the host overload detection and VM Selection using the VM Placement (MWFD). We have
calculated the effect size since effect size is to measure the strength and interpretation of the
effectiveness between the variables. It represents the value as 0.2 for the small effect, 0.5 for the
medium effect and 0.8 for the large effect.

From the results, 20 combinations (host overload detection and VM Selection) of the proposed
heuristic algorithms using MSWFD rejects the Null hypothesis. It shows the p-value <0.001, t-
value and mean differences as positive values. There are 12 combinations which did not reject
the NULL hypothesis. They have not rejected the null hypothesis because one of the reason is
either p-value is not less than 0.05 or t- value and mean values are negative. The 12
combinations proposed algorithms which did not reject the null hypothesis are mentioned in the
following Table 16:

Mean_MinR Mean_MaxU Median_MinU Median_MinR Median_MaxU

Median_MinU Mode_MinR Mode_MaxU Mode_MinU Range_MinR
Range_MinR Range_MaxU Range_MinU
Table 16: 12 combination algorithms rejected the alternative hypothesis (null
hypothesis accepted)

49
Best resulted 20 combinations (host overload and VM Selection) of the proposed algorithm
using MSWFD are mentioned in the following Table 17:

Mean_MaxR Median_MaxR Var_MaxR Var_MinR

Var_MaxU Var_MinU Std_MaxR Std_MinR
Std_MaxU Std_MinU Mode_MaxR Range_MaxR
MADmean_MaxR MADmean_MinR MADmean_MaxU MADmean_MinU
MADmean_MaxR MADmean_MinR MADmedian_MaxU MADmedian_MinU
Table 17: 20 combination algorithms accepted the alternative hypothesis (null hypothesis
rejected)
By this analysis, the remaining 20 combinations of host overload detection and VM Selection
using MSWFD perform better than the existing algorithm, i.e. the energy consumption is
minimum and minimum SLA violations. These algorithms have p-value < 0.001 and the effect
size is greater than 0.8, it has large effect.

50
5 DISCUSSION
5.1 Research Questions
RQ1) What are the algorithms suitable for energy-efficient model of Cloud-based system?
ANS) The research on energy-efficient model has been increased. Literature review has been
done to learn more about the algorithms of the existing energy efficient model. From the many
articles, the structure and insights of the CloudSim simulation model has been studied to
conduct the experiments on the energy efficient algorithms. The background of the energy
efficient model had been researched to implement new heuristics for the algorithms which
consumes less energy and show minimum SLA Violations which is related to the QoS (Quality
of Service). The answer to this research question, i.e. the algorithms and their explanation are
clearly explained in the Sections 3.2.6 and 3.2.8.

RQ2) What are the advantages and disadvantages of the proposed algorithm of the energy-
efficient model?
ANS) We have conducted experiment on the CloudSim simulation model to implement our
energy efficient algorithms. The results were obtained and analysis of the results (explained in
the Chapter 5) were done to observe the performance of our proposed algorithms of the energy
efficient model. And, the comparison between the proposed algorithms and existing algorithms
were done (explained in the Section 4.3). From the results and analysis, the advantage of our
proposed energy efficient model is that maximum number of heuristic algorithms of host
overload and VM selection gives the best performance under the VM placement algorithms.
When compared with the existing algorithm from the 32 combinations of proposed algorithm
(host overload detection and VM Selection), we got 22 best combinations (mentioned in the
Table 14) of heuristics algorithms (host overload detection and VM selection) using MWFD
(Modified Worst Fit Decreasing) VM placement algorithm (described in the section 3.2.8.4) and
we got 20 best combinations (mentioned in the Table 17) of proposed heuristic algorithms (host
overload detection and VM selection) using the MSWFD (Modified Second Worst Fit
Decreasing) VM placement algorithm. The proposed heuristics algorithms take the minimum
power consumption with less SLA Violation. This is considered as advantage of this work. The
disadvantage of the model is that 10 combinations (mentioned in the Table 13) of proposed of
heuristics algorithms (host overload detection and VM selection) using MWFD (Modified
Worst Fit Decreasing) VM placement algorithm and 12 combinations (mentioned in the Table
16) of proposed heuristic algorithms (host overload detection and VM selection) using the
MSWFD (Modified Second Worst Fit Decreasing) VM placement algorithm takes high energy
consumption and have high SLA Violations and have the poor performance when compared
with the existing algorithm.

RQ3) What are the performance of the algorithms, i.e. output of the energy-efficient model?
ANS) To answer this research Question, we have conducted the experiments on the CloudSim
and through the literature Review learning the existing models, we have proposed new
algorithms by combining the constraints of the Host Overload detection (Mean, Median, Mode,
Range, MADmedian, MADmean) and VM selection (MaxR, MinR, MaxU, MinU), these
algorithms are explained in the experiment Sections 3.2.8.1 and 3.2.8.2. Using these techniques,
we have proposed new heuristics algorithms (host overload detection and VM selection) using
the VM placement algorithms (Modified Worst Fit Decreasing and Modified Second Worst Fit
Decreasing). These proposed heuristics have shown the better results in energy consumption
and SLA Violation. These experiments are conducting using the Workload which is taken from
PlanetLab (explained in the experiment Section 3.2.11) and the paired t-test is conducted to
obtain the best result from the proposed algorithms combination of host overload detection and
VM Selection using MWFD and MSWFD VM placement algorithm. We got
MADMedian_MaxR as a best proposed algorithm (combination of host overload detection and
VM Selection) using MWFD and MADMean_MaxR as a best proposed algorithm (combination

51
of host overload detection and VM Selection) using MWFD since they take the less value of
ESV (Energy consumption and SLA Violation) i.e. they consume very less energy and has less
SLA Violation. When the comparison is done with the existing algorithm (described in the
Section 3.2.7), from the 32 combinations, 22 combinations of host overload detection and VM
Selection using MWFD VM placement algorithm and 20 combinations of host overload
detection and VM Selection using MSWFD VM placement algorithm shows the better results
than the existing IQR_MMT host overload detection and VM Selection using PABFD (Power
Aware Best Fit Decreasing) VM placement algorithm.

5.2 Threats Validity

Internal Validity [29] [54] [55]: Internal validity refers to the extent to which independent
variables have effect on the result of an experiment. Higher the internal validity, higher the
confidence on results tested with independent variables. Confounding variables (extra
independent variables having effect on dependent variables) must be avoided for higher internal
validity as it increases variance in program/experiment. In this workload is considered as threat
as the results are depends on the workload and dependent variables can have the effect, So the
different workloads had been taken to overcome problems while experimenting.

External Validity [29] [54] [55]: External validity refers to the extent at which findings of one
sample are of relevance to other sample. It means the extent to which the results of findings can
be generalized to the world at large. Higher the external validity, higher the generalization, i.e.
research/experiment is applicable to other researches/experiments. External validity can be
improved by using random sampling to select participants. Since, results of findings can be
generalized, and experiment is applicable to other research experiments. The PlanetLab
workload which is the CPU Utilization collected from the hundreds of the VMs. The results can
be generalized and can be used for the future we have used PlanetLab Workload.

52
6 CONCLUSION AND FUTURE WORK
6.1 Conclusion
At present, Cloud computing provisioning the resources to the customers and maintaining the
large data centers is crucial concept that needs to handle. There are thousands of computing
nodes in the datacenter, it consumes very high amount of energy. For the cooling the data
centers, Air Conditioners are applied, and it will consume lot of electrical energy. The main
problem of the Cloud computing is maintaining the data centers and it is not efficient in saving
the energy consumption. As solution to this problem, we have proposed an energy efficient
model using CloudSim. We have done simulation on the workload traces of the PlanetLab VMs.
We have proposed dynamic host underload detection algorithm, such that host can be
switched/changed to sleep mode from idle mode. This will save the energy consumption by the
data centers up to 70% (since data centers consumes 70% of energy even in idle mode). We
have proposed the new heuristics algorithms (32 combinations) of the host overload detection
and VM selection using VM placement algorithms. Paired t-test has been conducted to identify
the best proposed heuristic algorithms of host overload detection and VM selection using the
VM placement (Modified Worst Fit Decreasing and Modified Second Worst Fit Decreasing)
algorithm. From the 32 combinations, the heuristic algorithm MADmedian- MaxR (host
overload detection and VM Selection) using MWFD (Modified Worst Fit Decreasing) VM
Placement algorithm, and MADmean- MaxR (host overload detection and VM Selection) using
MSWFD (Modified Second Worst Fit Decreasing) VM placement algorithm gives the best
results which consume less energy while meeting SLA Violations. Paired t-test has been
conducted to extract the best combinations of proposed heuristic algorithms of host overload
detection and VM selection using VM placement algorithm when compared with the existing
algorithm IQR_MMT of host overload detection and VM selection using PABFD VM
placement algorithm. From the 32 combinations, 22 combinations of host overload detection
and VM Selection using MWFD VM placement algorithm and 20 combinations of host
overload detection and VM Selection using MSWFD VM placement algorithm shows the better
results than the existing IQR_MMT host overload detection and VM Selection using PABFD
VM placement algorithm and Cohen’s – d test has been conducted to calculate the effect size
and we acquired large effect size for the combinations.

6.2 Future Work

The work can be continued on different workloads and complex workloads. Different
algorithms can be implemented on the different workloads to implement an energy efficient
algorithm to reduce the energy cost. At present, OpenStack is trending in Cloud computing, this
energy efficient model can be implemented within OpenStack.

53
REFERENCES
[1] Beloglazov, Anton, and Rajkumar Buyya. "Optimal online deterministic algorithms and adaptive
heuristics for energy and performance efficient dynamic consolidation of virtual machines in cloud
data centers." Concurrency and Computation: Practice and Experience 24.13 (2012): 1397-1420.
[2] Choudhary, Anita, et al. "Energy-efficient resource allocation approaches with optimum virtual
machine migrations in cloud environment." Parallel, Distributed and Grid Computing (PDGC),
2016 Fourth International Conference on. IEEE, 2016.
[3] Mell, Peter, and Tim Grance. "The NIST definition of cloud computing." National institute of
standards and technology53.6 (2009): 50.
[4] Fan, Xiaobo, Wolf-Dietrich Weber, and Luiz Andre Barroso. "Power provisioning for a warehouse-sized
computer." ACM SIGARCH computer architecture news. Vol. 35. No. 2. ACM, 2007.
[5] Luo, Jianying, Lei Rao, and Xue Liu. "eco-idc: Trade delay for energy cost with service delay
guarantee for internet data centers." Cluster Computing (CLUSTER), 2012 IEEE International
Conference on. IEEE, 2012.
[6] Koomey, Jonathan. "Growth in data center electricity use 2005 to 2010." A report by Analytical
Press, completed at the request of The New York Times 9 (2011).
[7] Shidik, Guruh Fajar, Nani Sri Sulistyowati, and Manggiardi BW Tirta. "Evaluation of cluster K-
Means as VM selection in dynamic VM consolidation." Communications (APCC), 2016 22nd Asia-
Pacific Conference on. IEEE, 2016.
[8] Monil, Mohammad Alaul Haque, Romasa Qasim, and Rashedur M. Rahman. "Energy-aware VM
consolidation approach using combination of heuristics and migration control." Digital Information
Management (ICDIM), 2014 Ninth International Conference on. IEEE, 2014.
[9] Beloglazov, Anton, Jemal Abawajy, and Rajkumar Buyya. "Energy-aware resource allocation
heuristics for efficient management of data centers for cloud computing." Future generation
computer systems 28.5 (2012): 755-768.
[10] Mell, Peter, and Tim Grance. "The NIST definition of cloud computing." (2011).
[11] Jung, Sung-Min, Nam-Uk Kim, and Tai-Myoung Chung. "Applying scheduling algorithms with
QoS in the cloud computing." 2013 International Conference on Information Science and
Applications (ICISA). IEEE, 2013.
[12] Ranjbari, Milad, and Javad Akbari Torkestani. "A learning automata-based algorithm for energy and
SLA efficient consolidation of virtual machines in cloud data centers." Journal of Parallel and
Distributed Computing 113 (2018): 55-62.
[13] Jennings, Brendan, and Rolf Stadler. "Resource management in clouds: Survey and research
challenges." Journal of Network and Systems Management 23.3 (2015): 567-619.
[14] Manvi, Sunilkumar S., and Gopal Krishna Shyam. "Resource management for Infrastructure as a
Service (IaaS) in cloud computing: A survey." Journal of Network and Computer Applications 41
(2014): 424-440.
[15] Li. Yunfa, Li. Wanqing and C. Jiang. "A survey of virtual machine system: Current technology and
future trends." In Electronic Commerce and Security (ISECS), 2010 Third International Symposium
on,IEEE, pp. 332-336, 2010.
[16] T. Veni and S. Bhanu. "A survey on dynamic energy management at virtualization level in cloud
data centers." Computer Science & Information Technology, pp. 107-117, 2013.
[17] Ferreto, Tiago C., et al. "Server consolidation with migration control for virtualized data
centers." Future Generation Computer Systems 27.8 (2011): 1027-1034.
[18] Calheiros, Rodrigo N., et al. "CloudSim: a toolkit for modeling and simulation of cloud computing
environments and evaluation of resource provisioning algorithms." Software: Practice and
experience 41.1 (2011): 23-50.
[19] Long, Wang, Lan Yuqing, and Xia Qingxin. "Using cloudsim to model and simulate cloud
computing environment." In Computational Intelligence and Security (CIS), 2013 9th International
Conference on, IEEE,pp.323-328,2013.
[20] Verma, Akshat, Puneet Ahuja, and Anindya Neogi. "pMapper: power and migration cost aware
application placement in virtualized systems." Proceedings of the 9th ACM/IFIP/USENIX
International Conference on Middleware. Springer-Verlag New York, Inc., 2008.
[21] Li, Zhihua, et al. "Bayesian network-based virtual machines consolidation method." Future Generation
Computer Systems69 (2017): 75-87.
[22] Castro, Pedro HP, et al. "A joint CPU-RAM energy efficient and SLA-compliant approach for cloud data
centers." Computer Networks 94 (2016): 1-13.

54
[23] Nathuji, Ripal, and Karsten Schwan. "VirtualPower: coordinated power management in virtualized
enterprise systems." ACM SIGOPS Operating Systems Review. Vol. 41. No. 6. ACM, 2007.
[24] Nakada, Hidemoto, et al. "Toward virtual machine packing optimization based on genetic
algorithm." International Work-Conference on Artificial Neural Networks. Springer, Berlin,
Heidelberg, 2009.
[25] Parikh, Swapnil M. "A survey on cloud computing resource allocation techniques." Engineering
(NUiCONE), 2013 Nirma University International Conference on. IEEE, 2013.
[26] Wu, Linlin, Saurabh Kumar Garg, and Rajkumar Buyya. "SLA-based resource allocation for
software as a service provider (SaaS) in cloud computing environments." Proceedings of the 2011
11th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing. IEEE Computer
Society, 2011.
[27] Jiang, Congfeng, et al. "Resource allocation in contending virtualized environments through VM
performance modeling and feedback." Chinagrid Conference (ChinaGrid), 2011 Sixth Annual. IEEE,
2011.
[28] Xiao, Zhen, Weijia Song, and Qi Chen. "Dynamic resource allocation using virtual machines for cloud
computing environment." IEEE Trans. Parallel Distrib. Syst. 24.6 (2013): 1107-1117.
[29] Easterbrook, Steve, et al. "Selecting empirical methods for software engineering research." Guide to
advanced empirical software engineering. Springer, London, 2008. 285-311.
[30] Aberdeen, Trudie. "Yin, RK (2009). Case study research: Design and methods . Thousand Oaks,
CA: Sage." The Canadian Journal of Action Research 14.1 (2013): 69-71.
[31] Cronin, Patricia, Frances Ryan, and Michael Coughlan. "Undertaking a literature review: a step-by-
step approach." British journal of nursing 17.1 (2008): 38-43.
[32] Levy, Yair, and Timothy J. Ellis. "A systems approach to conduct an effective literature review in
support of information systems research." Informing Science 9 (2006).
[33] Wohlin, Claes. "Guidelines for snowballing in systematic literature studies and a replication in
software engineering." Proceedings of the 18th international conference on evaluation and
assessment in software engineering. ACM, 2014.
[34] Gandhi, Anshul, et al. "Optimal power allocation in server farms." ACM SIGMETRICS Performance
Evaluation Review. Vol. 37. No. 1. ACM, 2009.
[35] Kusic, Dara, et al. "Power and performance management of virtualized computing environments via
lookahead control." Cluster computing 12.1 (2009): 1-15.
[36] Raghavendra, Ramya, et al. "No power struggles: Coordinated multi-level power management for
the data center." ACM SIGARCH Computer Architecture News. Vol. 36. No. 1. ACM, 2008.
[37] Verma, Akshat, Puneet Ahuja, and Anindya Neogi. "pMapper: power and migration cost aware
application placement in virtualized systems." Proceedings of the 9th ACM/IFIP/USENIX
International Conference on Middleware. Springer-Verlag New York, Inc., 2008.
[38] Casalicchio, Emiliano, Lars Lundberg, and Sogand Shirinbab. "Energy-aware adaptation in
managed Cassandra datacenters." 2016 International Conference on Cloud and Autonomic
Computing (ICCAC). IEEE, 2016.
[39] Long, Wang, Lan Yuqing, and Xia Qingxin. "Using cloudsim to model and simulate cloud
computing environment." Computational Intelligence and Security (CIS), 2013 9th International
Conference on. IEEE, 2013.
[40] Voorsluys, William, et al. "Cost of virtual machine live migration in clouds: A performance
evaluation." IEEE International Conference on Cloud Computing. Springer, Berlin, Heidelberg,
2009.
[41] Yadav, Rahul, et al. "MuMs: Energy-Aware VM Selection Scheme for Cloud Data
Center." Database and Expert Systems Applications (DEXA), 2017 28th International Workshop on.
IEEE, 2017.
[42] Wohlin, Claes, et al. Experimentation in software engineering. Springer Science & Business Media,
2012.
[43] Park, KyoungSoo, and Vivek S. Pai. "CoMon: a mostly-scalable monitoring system for
PlanetLab." ACM SIGOPS Operating Systems Review 40.1 (2006): 65-74.
[44] “Wolfram MathWorld” [Online]. Available: http://mathworld.wolfram.com/
[45] “Null Hypothesis” [Online]. Available: https://math.tutorvista.com/statistics/null-hypothesis.html
[46] “Alternative Hypothesis” [Online]. Available: https://math.tutorvista.com/statistics/alternative-
hypothesis.html
[47] “Hypothesis Testing” [Online]. Available: https://math.tutorvista.com/statistics/hypothesis testing.html
[48] “NULL AND ALTERNATIVE HYPOTHESIS” [Online].
Available:https://researchbasics.education.uconn.edu/null-and-alternative-hypotheses/
[49] “t test formula” [Online]. Available: http://www.sthda.com/english/wiki/t-test-formula

55
[50] “Statistics How To” [Online]. Available: https://www.statisticshowto.datasciencecentral.com/
[51] “Scientifically Sound” [Online]. Available: https://scientificallysound.org/
[52] “Statistics Solutions” [Online]. Available: http://www.statisticssolutions.com/
[53] Krzywinski, Martin, and Naomi Altman. "Points of significance: visualizing samples with box plots."
(2014): 119.
[54] Runeson, Per, and Martin Höst. "Guidelines for conducting and reporting case study research in
software engineering." Empirical software engineering 14.2 (2009): 131.
[55] Yin, Robert K. "Case study research: Design and methods (applied social research
methods)." London and Singapore: Sage (2009).
[56] Sekhar, Jyothi, Getzi Jeba, and S. Durga. "A survey on energy efficient server consolidation
through vm live migration." International Journal of Advances in Engineering & Technology 5.1
(2012): 515.
[57] Raj, Rakhi K., Getzi Jeba, and P. Leelipushpam. "Live Virtual Machine Migration Techniques-A
Survery." International Journal of Engineering Research & Technology (IJERT) 1.7 (2012): 1-7.

56