You are on page 1of 88

Energy Efficient Multi-Connectivity for

Ultra-Dense Networks

Valentin Poirot

Computer Science and Engineering, master's level (120 credits)


2017

Luleå University of Technology


Department of Computer Science, Electrical and Space Engineering
Luleå University of Technology
Department of Computer Science, Electrical and Space Engineering
Erasmus Mundus Master’s Programme in Pervasive Computing & Communications for
sustainable Development PERCCOM

Valentin Poirot

ENERGY EFFICIENT MULTI-CONNECTIVITY FOR ULTRA-DENSE


NETWORKS

2017

Supervisors: Mårten Ericson (Ericsson Research)


Mats Nordberg (Ericsson Research)
Associate Professor Karl Andersson (Luleå University of Technology)

Examiners: Professor Eric Rondeau (University of Lorraine)


Professor Jari Porras (Lappeenranta University of Technology)
Associate Professor Karl Andersson (Luleå University of Technology)
This thesis is prepared as part of an European Erasmus Mundus programme
PERCCOM - Pervasive Computing & COMmunications for sustainable development.

This thesis has been accepted by partner institutions of the consortium (cf. UDL-DAJ, no 1524,
2012 PERCCOM agreement).

Successful defense of this thesis is obligatory for graduation with the following national diplo-
mas:

• Master in Complex Systems Engineering (University of Lorraine)

• Master of Science in Technology (Lappeenranta University of Technology)

• Master of Science (120 credits) - Major; Computer Science and Engineering, Speciali-
sation; Pervasive Computing and Communications for Sustainable Development (Luleå
University of Technology)
ABSTRACT

Luleå University of Technology


Department of Computer Science, Electrical and Space Engineering
PERCCOM Master Programme

Valentin Poirot

Energy Efficient Multi-Connectivity for Ultra-Dense Networks

Master’s Thesis

2017

87 pages, 23 figures, 9 tables

Examiners: Professor Eric Rondeau (University of Lorraine)


Professor Jari Porras (Lappeenranta University of Technology)
Associate Professor Karl Andersson (Luleå University of Technology)

Keywords: energy efficiency; multi connectivity; ultra dense network; 5G; secondary cell as-
sociation; reliability

In 5G systems, two radio air interfaces, evolved LTE and New Radio (NR), will coexist. By
using millimeter waves, NR will provide high throughputs, but the higher frequencies will also
lead to increased losses and a worse coverage. Multi-connectivity is therefore envisioned as
a way to tackle these effects by connecting to multiple base stations simultaneously, allowing
users to benefit from both air interfaces’ advantages. In this thesis, we investigate how multi-
connectivity can be used efficiently in ultra-dense networks, a new paradigm in which the num-
ber of access nodes exceed the number of users within the network. A framework for secondary
cell association is presented and an energy efficiency’s condition is proposed. Upper and lower
bounds of the network’s energy efficiency are analytically expressed. Algorithms for secondary
cell selection are designed and evaluated through simulations. Multi-connectivity showed an
improvement of up to 50% in reliability and and an increase of up to 20% in energy efficiency.
ACKNOWLEDGEMENT

Luleå, May 31, 2017

I would like to express all my gratitude to the PERCCOM consortium [1] which allowed me to
participate in this unique adventure for the past two years. It was a life-changing event that was
only possible by the help of all the people involved in the project.

Special thanks to Eric Rondeau, Jari Porras, Karl Andersson and Jean-Phillipe Georges for
everything they did for PERCCOM: all the activities organized, the classes given and their time.
I would also give a special mention to Karl for his assistance during the thesis.

I would also like to thanks Mats Nordberg and Mårten Ericson for allowing me to pursue this
thesis with them and for their continuous support throughout the work. Thanks to all the other
people working here at Ericsson, for their feedback and the coffee breaks.

And, of course, a special mention to all my classmates with whom I shared wonderful moments
during those past two years: Victor, Tamara, Chandara, Joseph, Emil, Rafiul, Manish, Felipe,
Atefe, Carlos, Mustaqim, Henrique, Olga, Giang, Nhi and Aigerim. To them, I want to say
Bonzon!

Valentin Poirot
5

CONTENTS
1 Introduction 10
1.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Objectives and Research Questions . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Thesis Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2 Background and Related Work 15


2.1 5th Generation of Mobile Communications . . . . . . . . . . . . . . . . . . . . 15
2.1.1 The European Vision . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.2 Technical Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Ultra-Dense Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 HetNet and UDN Characteristics . . . . . . . . . . . . . . . . . . . . . 19
2.2.2 Ultra-Dense Networks in the Literature . . . . . . . . . . . . . . . . . 20
2.3 Multi-Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 MC Scenarios and Procedures . . . . . . . . . . . . . . . . . . . . . . 22
2.3.2 Multi-Connectivity in the Literature . . . . . . . . . . . . . . . . . . . 24
2.4 Energy Efficiency in Mobile Networks . . . . . . . . . . . . . . . . . . . . . . 25
2.4.1 Power Consumption Models . . . . . . . . . . . . . . . . . . . . . . . 25
2.4.2 Sleep Mode Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.3 Sleep Mode Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.4 Other Studies on Energy Efficiency . . . . . . . . . . . . . . . . . . . 29
2.5 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Models and Multi-Connectivity Algorithms 31


3.1 Models Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1 Handover Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.2 Radio Link Failure Model . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.3 Sleep Mode Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.4 Network Related Models . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.5 User and Traffic Models . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Multi-Connectivity Implementation . . . . . . . . . . . . . . . . . . . . . . . 39
3.2.1 Basic Algorithm Decomposition . . . . . . . . . . . . . . . . . . . . . 39
3.2.2 Classification of Metrics Usage . . . . . . . . . . . . . . . . . . . . . 41
3.3 Presentation of our proposals . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3.1 Max Bitrate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.2 Max SINR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3.3 Max Bitrate-EE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6

3.3.4 Max Clustered-Bitrate . . . . . . . . . . . . . . . . . . . . . . . . . . 46


3.3.5 Analytic Hierarchy Process . . . . . . . . . . . . . . . . . . . . . . . . 47

4 Analytical Study of Energy Efficiency 49


4.1 Energy Efficiency’s Expression . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.1 Single Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.2 Multi-Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.3 Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.1.4 Low Performance Bound . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.1.5 Upper Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1.6 Ensuring Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . . 54
4.2 Power Condition for Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . 55
4.2.1 Definition of a Power Condition . . . . . . . . . . . . . . . . . . . . . 55
4.2.2 Power Condition based on the offered capacity . . . . . . . . . . . . . 57
4.2.3 Power Consumption Comparison of a Clustered-based and non Clustered-
based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5 Simulation Results 61
5.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Reliability Improvement with Multi-Connectivity . . . . . . . . . . . . . . . . 62
5.3 Multi-Connectivity Algorithms’ Comparison . . . . . . . . . . . . . . . . . . 66
5.3.1 Power Consumption and Sleeping Cells . . . . . . . . . . . . . . . . . 67
5.3.2 Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3.3 Theoretical Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3.4 Probability of Multi-Connectivity . . . . . . . . . . . . . . . . . . . . 72
5.3.5 User Throughput . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.4 Sustainability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

6 Summary and Discussions 77


6.1 Multi-Connectivity Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 Multi-Connectivity Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

7 Conclusion and Future Work 79


7

List of Figures
1 Envisioned Traffic Growth, from [2] . . . . . . . . . . . . . . . . . . . . . . . 10
2 5G Radio Access Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Heterogeneous Network (a) and Ultra-Dense Network (b) . . . . . . . . . . . . 19
4 Multi-Connectivity Architecture . . . . . . . . . . . . . . . . . . . . . . . . . 22
5 Inter-frequency MC (a) and intra-frequency MC (b) . . . . . . . . . . . . . . . 23
6 Power Consumption in Cellular Networks . . . . . . . . . . . . . . . . . . . . 25
7 Power Model for LTE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
8 Sleep Mode Levels, adapted from [57] . . . . . . . . . . . . . . . . . . . . . . 28
9 Handover Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
10 Sleep Mode Model for 5G . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
11 Deployment Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
12 Association Procedure Flowchart . . . . . . . . . . . . . . . . . . . . . . . . . 40
13 Disassociation Procedure Flowchart . . . . . . . . . . . . . . . . . . . . . . . 41
14 Radio Link Failure Rate depending on the user velocity . . . . . . . . . . . . . 63
15 Linear Regression of the Simulated RLF Rate . . . . . . . . . . . . . . . . . . 64
16 Standard Deviation of the RLF Rate . . . . . . . . . . . . . . . . . . . . . . . 65
17 Simulation results distribution depending on the velocity and number of con-
nections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
18 System Power Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
19 Percentage of Sleeping Cells . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
20 System Energy Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
21 Results Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
22 Simulated Results and Theoretical Lower Bound for Energy Efficiency . . . . . 72
23 10th Percentile User Throughput . . . . . . . . . . . . . . . . . . . . . . . . . 74
8

List of Tables
1 Handover Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2 Radio Link Failure Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3 Network Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Users Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5 Traffic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6 Multi-Connectivity Schemes Parameters . . . . . . . . . . . . . . . . . . . . . 42
7 Scenario Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
8 Deployment Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
9 Probability of Multi-Connectivity . . . . . . . . . . . . . . . . . . . . . . . . 73
9

ABBREVIATIONS AND SYMBOLS


3GPP Third Generation Partnership Project
5G Fifth Generation of Mobile Communications
5GPPP 5G Public Private Partnership
BS Base Station
CoMP Coordinated MultiPoint
DC Dual Connectivity
DL Downlink
EE Energy Efficiency
ETSI European Telecommunications Standards Institute
FTP File Transfer Protocol
HetNets Heterogeneous Networks
ITU International Telecommunications Union
LTE Long Term Evolution
LTE-A Long Term Evolution Advanced
MC Multi-Connectivity
METIS Mobile and wireless Enablers for the Twenty-twenty Information Society
MIMO Multiple Input Multiple Output
mmWave Millimeter Waves
NR (5G) New Radio
RAT Radio Access Technology
RLF Radio Link Failure
SINR Signal over Interference plus Noise Ratio
UDN Ultra-Dense Networks
UE User Equipment
UL Uplink
10

1 Introduction

Throughout the world, mobile networks are being more and more used. It is expected that there
will be a total of 8,900 million mobile subscriptions in 2022, smartphones accounting for 6,800
million of the total [2]. Moreover, the traffic growth is also following an exponential growth. In
2022, it is expected that the total mobile traffic will increase by a factor of 10, with video traffic
making most of it.

Figure 1. Envisioned Traffic Growth, from [2]

Tackling this rise of traffic has become the priority of many actors in telecommunications. Mul-
tiples projects and collaborations are carried out around the globe to investigate, discuss and
prepare the next generation of mobile communications, 5G.

Moreover, 5G will have to fully support new use cases. Machine-type communications, coming
from the Internet of Things, driverless cars or the industry will require very low latency and
high reliability. Accurate positioning will be an important service for localized and personalised
applications, but also for emergency responses. High reliability should also be ensured for high
velocity vehicles, such as trains. Finally, 5G should provide acceptable quality of experience
even in highly dense areas, like stadiums or shopping malls [3].

Current developments of 5G follow two tracks: an evolution of LTE, which will improve its
capacity while ensuring backward compatibility, and a new radio access technology, called New
radio (NR), which will work at higher frequencies to provide improved throughputs. [4]
11

With higher frequencies also come low coverage, leading to a need for more serving points. It
is expected that the network densification will be so important that the number of access nodes
will exceed the number of active users within the network. This paradigm is called ultra-dense
network, and will allow optimal throughput for the user, regardless of its location. Further-
more, multi-connectivity will allow users to connect to multiple access points at the same time.
This will allow them to benefit from the coverage of LTE and the performance of New Radio
simultaneously, at the cost of more active base stations for the network.

1.1 Problem Definition

Energy consumption is becoming more and more a problem for network operators. In 2011,
base stations alone were representing 4.5 GW of power and 20 Mt of CO2 per year [5]. Since
that, the figures has been constantly increasing. In terms of costs, this equates to up to 18% of
operational expenses in Europe, and up to 32% in India for that same year [6]. Energy efficiency
has therefore been targeted at the international level as one of the key capabilities of 5G [3].

Ultra-dense networks, where tremendous number of base stations are deployed in a never seen
manner, will probably cause a pike of energy consumption in specific areas. Energy-efficient
hardware and efficient sleep procedures will be critical in urban areas. However, at the same
time, it is envisioned that multi-connectivity could be used to provide better quality of expe-
rience (QoE). Users could then connect to multiple access nodes, thus aggregating additional
bandwidth for high data rate and improving reliability. That feature could undermine energy re-
duction techniques put into place. A trade-off between QoE and power consumption is therefore
critical.

The impacts of 5G in terms of sustainability are not yet known. On one hand, most predictions
tend to believe the energy consumption of the network will increase worldwide, while a lot of
research has been done recently towards more energy efficient solutions. Through the study of
energy efficiency, the impacts of these new technologies should also be assessed with regard to
sustainability.

1.2 Objectives and Research Questions

This thesis focuses on studying multi-connectivity. First, we investigate its effects on the net-
work performance. We do it both theoretically and by simulation. The second part of our work
12

consists in designing new algorithms for multi-connectivity association. Specifically, we give a


particular attention to the system’s energy efficiency, and try to improve it.

We will answer the following research questions:

1. How does multi-connectivity affect the mobile network performance?

(a) What is the impact of multi-connectivity on the power consumption?


(b) How multi-connectivity affects the energy efficiency of the system?
(c) How can multi-connectivity improve reliability for the user?

2. How can we design Multi-Connectivity schemes to optimize the system performance?

(a) What common characteristics define multi-connectivity schemes?


(b) How can we compare different schemes?
(c) How can we estimate energy efficiency when choosing a secondary cell?

3. How do these results translate in terms of sustainability?

We will answer these questions in four steps. After a survey of the state of the art, we will de-
sign algorithms for secondary cell association aimed at focusing the energy efficiency of multi-
connectivity. We will then perform a numerical analysis of energy efficiency and evaluate our
proposals through simulations. Finally, we will look at the power consumption and the energy
efficiency of networks implementing our solutions to establish the impact of multi-connectivity
in terms of sustainability.

As 5G networks are only deployed in laboratories at the time of this study, actual measurements
are not possible and we therefore decided to use a simulation tool in the evaluation of the work.
Models used by our simulator are explained later in this thesis and are based on others’ work
from empirical measurements made on real life equipment.

As one focus of this work is to design energy efficient solutions, we adopt a Green IT approach
of sustainability. The energy consumption and energy efficiency of the system are our most
important metrics related to the problem, and we deduce from them direct impacts on carbon
emissions and costs. Other indirect impacts, or rebound effects, are discussed when possible.
Therefore, we consider only two of the three main aspects, or pillars, of sustainability, related
to economical and ecological challenges.
13

1.3 Scope

In this thesis, we focus on the usage of multi-connectivity within ultra-dense networks. This
means that this work takes place in dense, urban areas. We use a combination of macro cells
and small cells, which are explained further in this thesis. Moreover, a FTP traffic model is used
to represent Internet-related user traffic within the network.

Furthermore, we propose several algorithms that are evaluated in this work. These algorithms
treat the cell selection for secondary cell association and disassociation. We do not propose here
new procedures (i.e. communications between BSs and UEs) for cell association and resource
scheduling, as these parts will be defined during the standardization process.

1.4 Thesis Structure

The thesis is structured as follows:

Chapter 2 - Background and Related Work


All important aspects necessary for the good understanding of the thesis are presented here.
After a brief introduction of the history of mobile communications, a brief tour of 5G and its
most promising technologies is given. The key terms of the title are then defined, and a survey
of the literature is carried out.

Chapter 3 - Models and Multi-Connectivity Algorithms


This chapter first introduces our simulator. The different models used and implemented
throughout our work are explained. The second part presents our framework for multi-connectivity.
Multiple algorithms are introduced and their functionalities are detailed.

Chapter 4 - Analytical Study of Energy Efficiency


The analytical work is shown in this chapter. A general expression of multi-connectivity is
given, and its evaluation for energy efficiency is carried out. An approach to power savings is
also expressed in this chapter, and a study of several algorithms is performed.
14

Chapter 5 - Results
The work is evaluated in chapter 5. After presenting the methodology, results are presented
for each scenario.

Chapter 6 - Summary and Discussions


A summary of the main findings is done in chapter 6. A discussion of the results is also done.

Chapter 7 - Conclusion and Future Work


Finally, we conclude this thesis in chapter 7. Our contributions are highlighted and potential
tracks are given for future studies.
15

2 Background and Related Work

This chapter encompasses fundamentals notions that are needed for the good understanding of
this thesis as well as the state of the art of the field. 5G is presented in 2.1 through its technical
solutions. Sections 2.2 and 2.3 define ultra-dense networks and multi-connectivity respectively.
A review of the literature is given for these two concepts. Previous works related to energy
efficiency are then listed in section 2.4. Finally, an overview of the different metrics used in the
literature is given in section 2.5.

2.1 5th Generation of Mobile Communications

During the past ten years, multiple projects spawned worldwide to study and develop the next
generation of mobile communications. In America, Europe or Asia, there is a push for na-
tional or continental efforts to lead its development. At the international level, the International
Telecommunications Union (ITU) started planning the IMT-2020 specification in 2014, for a
release in 2020.

2.1.1 The European Vision

The European Union decided to invest a lot in the development of the next generation of mobile
communications. The METIS project (Mobile and wireless Enablers for the Twenty-twenty
(2020) Information Society) is a FP7 funded collaboration between industry and academia
started in 2012 and finished in 2015. Its main objective was to lay the foundation for the future of
mobile communications, by bringing a consensus between multiple European actors in prevision
of the international standardization. With 29 partners including manufacturers, telecommuni-
cations operators, automotive industries and academia, they successfully managed to propose a
unified vision of what 5G should look like by stating challenges and requirements that must be
met by the new technology in order to be considered a success.

In February 2014, The European Commission and the European ICT industry jointly launched
The 5G Public Private Partnership (5G-PPP) under the Horizon 2020 funding framework. Its
objective is to investigate and produce solutions, architectures and standards for 5G.

A few defining attributes were stated by 5G-PPP [7]:


16

• Amazingly fast

• Great service in a crowd

• Best experience follows you

• Super real-time and reliable connections

• Ubiquitous things communicating

5G is thus not only envisioned as an improvement of LTE, with higher capacity and increased
quality of service, but it should also cover new use cases that were not present before. Pervasive
computing, IoT, connected vehicles and the industry 4.0 are a few of these new applications that
must be served in 5G.

In order to evaluate potential technologies, these scenarios were also transposed into technical
requirements. 5G must not only deliver good results during in first deployment, around 2020,
but must also be able to evolve as the traffic grows in the following years.

The METIS 5G requirements were :

• “1000 times higher mobile data volume per area,

• 10 to 100 times higher number of connected devices,

• 10 to 100 times higher user data rate,

• 10 times longer battery life for low power massive machine communication (MMC), and
finally, and

• 5 times reduced end-to-end latency” [8]

These requirements might not need to be satisfied all at once. Indeed, with the appearance
of the scenarios listed before, only a few of them will be required at once. These led to the
division of 5G into three main services: (i) extreme mobile broadband, (ii) massive machine-
type communications and (iii) ultra-reliable machine-type communications.

These services, as well as the attributes defined by 5G-PPP were refined in twelve test-cases,
each of them aiming at representing one possible deployment and utilisation of 5G. The entire
list can be found at [9]. We can, as an example, cite Shopping mall, Emergency Communications
or Massive deployments of sensors and actuators.
17

At the international level, the International Telecommunications Union (ITU) published a draft
for the technical requirements for IMT-2020, the technological family for 5G [10].

2.1.2 Technical Solutions

As explained in the introductory chapter, 5G will be composed of two radio access technologies
(RATs): an evolution of LTE, and New Radio (NR) [11]. We list here few of the most promising
tracks to achieve the technical requirements fixed by the ITU. Figure 2 represents 5G RATs.

Figure 2. 5G Radio Access Technologies

The main concept of NR is to use higher frequency bands. Most of the studies consider the
millimeter waves as the best candidate. mm-waves refer to the frequencies in the 3-300 GHz
[12], and is composed of the Super High Frequency (SHF) and Extremely High Frequency
(EHF) bands, with wavelength between 1 to 100 mm. Numerous frequencies were also proposed
within this range, and theoretical and experimental studies have been carried out to determine
which is the most suitable. Amongst them, we can cite the 15Ghz band [13], the 28 and 38 GHz
bands [14, 15] and the 60 GHz band [16, 17]. A late consensus seems to indicate the 28 GHz
as the most suitable candidate in the beginning. However, the ITU might select multiple bands
during the IMT-2020 standardisation procedure.

It can also be worth noting that unlicensed spectrum might be available in 5G, as it was with
LTE-Unlicensed [18]. It might however only be complementary, as opposed to the new fre-
quency band which will play an important role.

The second most promising technology is the usage of massive MIMO (multiple input multiple
18

output) [19, 20, 21]. In conventional point-to-point MIMO, base stations and user equipment are
equipped of more than one receiver and transmitter. MIMO can be used for precoding, where
all the transmitters emit the same signal in order to increase the received signal power at the
receptor. MIMO is also used for spatial multiplexing, where each transmitter emits a part of
the message at the same time in order to increase the bitrate. However, MIMO technology was
mainly used for one user at a time [22]. Massive MIMO is a scaled-up MIMO, with hundreds of
antennas elements in one base station, with the capacity to serve hundreds of users at the same
time.

Massive MIMO offers numerous advantages. The usage of cheap low-power transceivers makes
it energy-efficient, secure, robust and offer a great spectral efficiency. Different deployment sce-
narios are possible: distributed, where the antennas are spread across some area with a common
radio cloud for processing, cylindar, to improve covered area, linear and rectangular [23].
Massive MIMO is expected to increase the capacity by 10 or more and the energy efficiency
by 1000 or more. High mobility is however a challenge for this system as localisation-aware
spatial beamforming and channel estimation might require a lot of signalling.

Software Defined Networking (SDN), Network Function Virtualisation (NFV) and Network
Slicing are also envisioned [24, 25]. Virtual evolved packet core (vEPC) can be used to lower
the costs for operators and improve the network efficiency.

Finally, device-to-device (D2D) communications tackle the traditional view on architecture,


where users and providers are two different entities. New nodes might serve as relay between
a user equipment and a base station, or two equipment might want to communicate together
without using the serving base station [26, 27]. With connected vehicles, both vehicle-to-vehicle
and vehicle-to-infrastructure communications will be important, either to serve users within the
vehicle or for new services such as collision avoidance for unmanned cars. Internet of Things
can also benefit from D2D communications, especially for reliability, low latency and energy
efficiency [28, 29].

2.2 Ultra-Dense Networks

As stated before, ultra-dense networks (UDN) are seen as one of the most important way of
tackling the new requirements. We can give a first and simple definition of UDN as:

A deployment in which the number of access nodes exceeds the number of users.
19

While this paradigm is seen as a cornerstone of 5G, densified networks were already present in
the past. We can see Heterogeneous Networks (HetNets), a concept designed for LTE, as a first
step in this direction. However, HetNets were only used locally to increase the performance,
whereas UDN is seen as a cornerstone of future deployments.

Figure 3. Heterogeneous Network (a) and Ultra-Dense Network (b)

Another possible definition of UDN, which can be extracted of the quantitative work of Ding
et al. [30], is related to the density of cells. In their work, 103 cells/km2 appears as the lower
bound for an ultra-dense deployment. This definition is linked to the first one in [31]. However,
their density of users might not reflect the one in 2020 and onward. It is solely included in this
review to provide another view on how UDN can be defined.

2.2.1 HetNet and UDN Characteristics

HetNets are based on the principle of using cells of different size or coverage within the same
network [32]. Typically, a HetNet is composed of two layers: a macro-layer, with traditional
base stations, and a micro-layer, where small cells, either micro, femto or pico-cells, are used in
strategic locations. Such scenario is often referred as “multi-tier deployment”.

Small cells have two main objectives: provide coverage and/or performance. At the edge of
macro-cells, where downlink throughput is usually low and uplink suffers even more, users
can connect directly to that cell and maintain a good quality of service. Another possibility is
using Coordinated Multi-Point (CoMP), where multiple base stations cooperate by transmitting
or listening simultaneously and in the same subframe to decrease the probability of erroneous
transmissions [33].

But small cells can also be deployed within macro-coverage to serve small areas with high traf-
fic. Solutions like Carrier Aggregation or Dual Connectivity can be set up to boost performance.
20

From this concept, the main idea was to deploy even more small cells, as it has been shown that
densifying the deployment does increase the system throughput [34].

Access nodes will not be exclusively deployed by network operators in 5G. It is expected that
indoor deployment can be done by users within their homes. Moreover, in device to device
communications, some devices might play the role of user, relay or access node depending of
the situation. These possibilities make management quite complex in some cases.

From these definitions and observations, we can list some crucial characteristics of UDN. (i)
Firstly, one user can be within the vicinity of multiple cells. That was already true with multi-
tier deployments, but users can be within coverage of more than two access nodes in UDN.
(ii) Secondly, Interferences will be highly critical. With low inter-site distance, transmissions
might interfere between each other, leading to a decrease of the signal quality. (iii) Some cells
might have no users within their coverage. This characteristic is crucial, as sleep mode can be
used to save energy. (iv) In high frequency bands, Line Of Sight deployment will be of great
importance. The propagation losses will be so important that user losing a clear view towards
the access node might lose its connection.

2.2.2 Ultra-Dense Networks in the Literature

Numerous papers have investigated UDN. We review here some of them by highlighting the
different approaches used. A recent survey on ultra-dense networks can be found at [31].

The basic foundation of UDN is laid down in [35]. A list of objectives and challenges are
presented, from scaling laws to coordination problems. Some basic impacts of the density
of base stations are evaluated, such as the guaranteed rate. The authors conclude by stating
that more performance modelling and realistic scenarios are needed to push UDN to their full
potential. Theoretical work has been carried out to model UDN. Two main mathematical tools
were used: stochastic geometry, and game theory.

Stochastic geometry is often used to represent wireless networks, with users and base stations
usually represented by Poisson processes, as it offers the possibility to evaluate probabilistic
deployments. Indicators such as outage or coverage probability or the bitrate can be expressed
as a lot of mathematical background is available. Stochastic geometry is used in [36] to express
spectral and energy efficiency in relation to the density of base stations. The authors show
that densifying up to a certain value is not efficient. Also, this optimal density depends of
the transmitted power per base station. An optimal density can be found. Similarly, energy
21

efficiency is at its minimum (as EE is defined as the power used to transmit one bit here) for the
same optimal density.

Game Theory, the second mathematical tool, is used to study behaviour and decision making in
cooperation or conflict between rational agents. In [37], Sun et al. formulates a non-cooperative
game for cell selection in association. Centralized and decentralized cluster control are tested,
and CoMP is used as a way to achieve better performance. A new SINR measurement is also
proposed, and their results show a higher connectivity towards small cells and higher through-
put for low throughput users. [38] uses mean-field game theory, a sub-field of that mathematical
tools, where the number of players is significantly higher. In their work, a two-level interfer-
ence management framework is proposed. Especially, co-layer interferences are tackled with
MFG, and the tested algorithms show an increase in energy efficiency compared to traditional
frequency reuse solutions.

Another study shows that densification has its limits. In [39], a deployment with constant area
is densified and tested with different propagation model. Results showed that a theoretical limit
to densification exists, and it is related to the baseline (constant) power consumption of access
nodes.

Simulations are also largely used in the literature. Yunas et al. [40] also studied spectrum and
energy efficiency. Their work focused on three deployment strategies: dense indoor femtocells,
densified macrocells and dynamic distributed antenna system, thus making the comparison with
[36] impossible.

2.3 Multi-Connectivity

The concept of multi-connectivity is not new with 5G. Indeed, some similar mechanisms were
defined and standardized for LTE. Dual Connectivity (DC) is defined as an “operation where a
given UE consumes radio resources provided by at least two different network point connected
with non-ideal backhaul” [41]. Although the official definition allows more than two connec-
tions, the scenarios covered contain only one macro cell (master eNodeB) and one small cell
(secondary eNodeB).

As for multi-connectivity (MC), Its standardization is only a draft yet and no final definition
has been given [42]. However, most authors agree and use a similar concept, derived from the
definition of DC. In this work, we define multi-connectivity as:
22

The ability for a user equipment to connect to multiple access nodes at the same time.

We will develop more this definition further in this section.

In order to understand how DC and, by extension MC, works, we will briefly explain its archi-
tecture and related procedures. A deeper presentation of DC can be found at [43]. We will call
master cell the main base station, also called master eNodeB, MeNB in the literature, and sec-
ondary cell a base station used as an additional connection. It is also called secondary eNodeB,
SeNB in the literature. Unlike the DC standard, we do not consider the master cell as a macro
cell only and the secondary cell as a small cell only, but accepts scenarios where the master cell
can be a small cell and a secondary cell can be a macro cell.

2.3.1 MC Scenarios and Procedures

In order to reduce signalling, the control plane is present at the master cell only. The UE RRC
(Radio Resource Control) layer is terminated at the master cell, which is itself directly connected
and communicating with the core network (Evolved Packet Core, EPC for LTE; Next Gen Core,
NGC for NR) through the interfaces S1 or NG. The serving cells are communicating together
via the X2 or Xn interfaces. The UE is then connected through the user plane to every serving
cells. Figure 4 shows the control and user planes in MC.

Figure 4. Multi-Connectivity Architecture

For the user plane, multiple scenarios are possible. First, since DC was designed for LTE, both
23

the master and the secondary cells were using the same frequency band. This scenario is called
intra-frequency DC. Within this deployment, the split of the data transmission could be either
at the bearer level, i.e. each cell use a different radio bearer, or at the packet level, i.e. the
master and secondary cells share the same bearer, but packets are either send by the master
or a secondary cell. In MC, since multiple frequency bands are available, another scenario is
possible. We can have inter-frequency MC, where the master cell and secondary cell does not
use the same frequency band. The architecture is, in this case, similar to the bearer level split.

We can further improve our definition of MC:

Multi-Connectivity is the ability for a user equipment to connect to multiple access nodes at the
same time. These connections can happen either within the same frequency band, which we
call intra-frequency MC, or in different frequency bands, which we call inter-frequency MC.

Figure 5. Inter-frequency MC (a) and intra-frequency MC (b)

In DC, the decision to connect a user to a secondary cell is made by the master cell, based on
the UE measurements. Although the user might have more information at its level, especially
concerning the radio signals, its battery usage, etc., having the decision at the network level
allows a better control of resources, especially when it comes to energy efficiency.

Concerning the uplink (UL), different scenarios exist. UL split can be proven difficult as it
increases the complexity of handling reports and prioritization (BSR, LCP, PHR, etc.). How-
ever, another technique, named Coordinated Multi-Point (CoMP), can be used to increase UL
performance. More information on CoMP can be found at [44].
24

2.3.2 Multi-Connectivity in the Literature

In [45], MC in mmWave is investigated. Specifically, a procedure is defined for directional beam


tracking, where both the user and small cells are composed of highly directional antenna array.
The work focused on how tracking can be done in order to maintain the multiple connections.
Since antenna arrays are another promising technology for 5G, this approach can be of great
interest. However, it might induce an increase in signalling, which is the opposite of what 5G is
trying to achieve.

Another work on MC tackles the problem of mobility [46]. Indeed, higher frequency bands
are typically associated with lower reliability, as propagation losses are higher and the channel
quality can quickly evolve. This work examines how MC can improve the reliability of the
connection by looking at the radio link failure (RLF) rate for walking users at 3km/h and users
at 60 km/h (in cars along a linear road). A few assumptions are made regarding the architecture,
in particular a cloud radio is considered, in order to avoid handover latencies. Furthermore, a
strong assumption is made by considering that the control plane is handled by “all of the co-
ordinated cells.” A MC scheme is also proposed, and multiple parameters are tested to see which
provide the highest reliability. Their results showed that the RLF can be completely resolved
and the 5%ile-throughput can be improved by 43% with specific settings. However, it was also
shown that too extreme settings will decrease the user throughput.

Da Silva et al. proposes different approaches to define MC in order to provide a tight integration
of LTE and 5G [47]. Different scenarios of common layers are presented to allow such integra-
tion, and procedures such as fast switching, user plane aggregation or control plane diversity
are presented along with their advantages. However, these solutions are not tested or put into
practice in any way. These can be seen as possible improvements and propositions towards a
standardized solution.

In [48], spectrum aggregation in DC is studied for decoupled UL/DL in 5G. Previous studies
showed that the best cell for downlink (DL) might not always be the best for uplink (UL), as
the latter depends on the user equipment transmission capabilities. Moreover, spectrum aggre-
gation was also shown at being not energy efficient in certain cases for UL. The work used the
stochastic geometry framework to prove analytically how UL/DL decoupling can be used with
DC in order to improve the user throughput. The results showed that interesting improvements
can be made for DL, but it was rather limited for UL because of the power constraints.
25

2.4 Energy Efficiency in Mobile Networks

Studies were carried out to decompose the power consumption at the network level [49] and
at the component level in base stations [50]. It was discovered that base stations account for
around 57% of the entire operative cost. Mobile switching and the network core account to-
gether for around 35%, and datacentres and retail represents the remainder. Figure 6 shows the
consumption decomposition. The figure is adapted from the values of [49] and [50].

Figure 6. Power Consumption in Cellular Networks

Therefore, most works are aimed at improving the energy efficiency of base stations. More
energy efficient and adaptive hardware were designed to reduce their footprint. However, since
only 10% of the BS consumption is caused by signal processing, against 50 to 80% in the
power amplifier, most studies concentrate on techniques such as sleep mode to reduce the BS
consumption to some lower level.

2.4.1 Power Consumption Models

Models for power consumptions are necessary in order to evaluate energy efficiency. Multiple
models have been proposed in the literature, covering different aspects such as the differences in
consumption between macro-cell and femto-cell, or between abstract or more realistic models.
Ismail et al. summarize some power models in their survey [51]. The most common and most
widely used model for macro-cell consumption is defined in [52].

A few papers also focus on femtocell consumption. Femtocell is the main enabler for HetNets
and UDN, and its consumption must be correctly modelled for their studies. Deruyck et al.
[53] proposed a power model independent of the load. This choice can be explained by the low
26

radiated power of such base stations, but the power is likely going to evolve following the load,
as for more traditional BS. Riggio et al. studied real femtocell consumption in order to abstract
a model out of their measurements [54]. This model is this time load dependent, although it is
only modelling 3G femtocells.

The EARTH project (Energy Aware Radio neTwork tecHnologies) is a FP7 financed European
project aimed at investigating energy efficiency of mobile communication networks. In its E3
Framework (Energy Efficient Evaluation Framework), Auer et al. a power model for base sta-
tions in LTE as they are used nowadays [52]. Figure 7 shows the power decomposition for the
different components of a base stations of different size.

Figure 7. Power Model for LTE

Mathematically, the power consumption is defined as:

n P + P ∗ ∆ ,0 6 P 6 P
0 tx P tx max
PBS,LTE = NT ∗ NS ∗ NC ∗ (1)
Psleep , sleep mode

Where NT is the number of transceivers (antennas), NS is the number of sectors, NC the number
of carriers, P0 the static power consumption caused by the cooling and signal processing, Ptx the
radiated power, ∆P the power consumption dependent of the radiated power, due to feeder losses
and the power amplifier, Pmax the maximum radiated power and Psleep the power consumption
when the base station is in sleep mode.
27

Regarding 5G, most studies envision or assume that 5G base stations will follow a model similar
to the one used for LTE. However, the development of more energy efficient hardware will most
likely reduce the overall consumption. Moreover, METIS, for its 5G concept, envisioned a lean
system control plane [55], sometimes also called ultra-lean design, as a way to reduce signalling.
Such an improvement will bring more possibility for more often and deeper sleep-mode usage.

In order to incorporate lean design to the power model of 5G, we define the power consumption
as:
( P0 + Ptx ∗ ∆P , 0 6 Ptx 6 Pmax
PBS,5G = NT ∗ NS ∗ NC ∗ Psleep , micro-sleep mode (2)
δNR ∗ Psleep , deep-sleep mode
where we reuse the notation in (1). δNR corresponds here, in a similar manner to [56], to the im-
provement made by the ultra-lean design in sleep consumption. Thus, we denote the difference
in LTE and 5G models by a second, deeper, sleep mode level, which corresponds to a fraction
of the usual sleep mode consumption. The reason of this expression is detailed below.

2.4.2 Sleep Mode Model

In order to save energy, equipment vendors decided to develop more energy-efficient techniques
for their products, especially for scenarios with low load. Sleep mode and cell breathing are two
of these solutions.

Cell breathing consists of increasing or reducing the radiated power, thus allowing an increase
or decrease in coverage. Even if increasing the transmitted power does also increase the energy
consumption, it can be orchestrated in such a way that neighbours cells consumption can be
decreased, either by reducing their coverage or even switching them off.

Sleep mode corresponds to a state for the base station when its functions are reduced to its strict
minimum, allowing a reduction in energy consumption at the cost of not serving eventual users.

More complex sleep mode is also envisioned for new base stations. Unlike in LTE where one
sleep mode is defined, 5G could see the apparition of multiple sleep mode levels, each with a
different power consumption [57]. A base station could thus progress through different level,
reducing at each step its consumption, and for longer periods of time, as depicted in figure 8.
28

Figure 8. Sleep Mode Levels, adapted from [57]

2.4.3 Sleep Mode Techniques

Sleep mode usage has already been intensively investigated to improve energy efficiency in
networks. Using stochastic geometry, Soh et al. showed analytically its importance in hetero-
geneous networks [58]. More extensive surveys of the different sleep mode techniques can be
found in [59] for switch-off techniques and in [60] for cell zooming algorithms.

Coordinated Multi-Point (CoMP) is used in [61] in addition to the sleep mechanism. Neighbours
of the sleeping cell coordinates together to provide coverage. It is shown that the SINR actively
increased when using CoMP. However, the important signalling to coordinate neighbours can
quickly lead to inefficient energy consumption. A threshold for CoMP is thus necessary and
proposed. Channel estimation is also presented as a factor for CoMP configuration.

Dudnikova et al. proposed a multi-criteria decision algorithm to select cells that should be
sleeping and redistribute spectral resources in [62]. Analytic Hierarchy Process and Grey Rela-
tional Analysis are used to rank the different cells. Coverage rate indicator, cell load factor and
the number of interfering neighbours are used as criterion. Results showed improved energy
efficiency between 5 and 11% against other algorithms from the literature. In [63], the same
authors also include Fuzzy Logic to their proposal.
29

A low complexity algorithm is proposed in [64]. By ordering cells that could be switched off
depending on the possible savings and the load steering, the authors managed to obtain results
similar to the optimal, NP hard, optimization solution. An energy saving ratio of 0.14 is found,
compared to 0.06 and 0.1 from two other algorithms of the literature (the higher means more
savings).

In [65], cells are ordered based on its distance to its users. The average distance of the traffic
load is computed and shared with neighbouring cells. The cell with the highest distance is
switch off if it does not degrade quality of service. Results showed an improvement of 28% in
energy consumption.

Cell breathing also received a lot of attention. In [66], antenna tilt is modified after the cell
switch-off selection in order to modify the coverage. Results show a slight improvement of
quality of service and energy efficiency after tilting optimization. However, we believe that this
solution cannot be used in deployment where switch-on and switch-off are frequent. Moreover,
the tilt-optimisation algorithm is based on knowledge on the antenna characteristics as well as
other cells relative positions. This won’t be possible in user-deployed networks.

Finally, a centralized cell zooming algorithm and its procedure is described in [67]. Cell zoom-
ing can either be done by augmenting the transmitted power, CoMP or relaying in this case. This
algorithm was tested against a distributed algorithm based on the same principle, and the cen-
tralised one behaved better in terms of energy consumed, with savings of the order of 20-50%
compared to the baseline without algorithm.

2.4.4 Other Studies on Energy Efficiency

Discontinuous Transmission (DTX) impact on energy efficiency is studied in [68]. At the net-
work level, cell DTX consists of putting to a lower power mode the base station during two
broadcast or signalling when no transmission is needed. These periods are usually 2 or 3 OFDM
symbols long. A 15% decrease in energy per bit is obtained with cell DTX. Combined with
adaptive hardware, which can attain lower level of consumption in low power mode, a decrease
of 57% can be achieved.

Multi-RAT deployments and their relation to energy efficiency are investigated in [69]. In co-
sited deployments, it is shown that shared power amplifiers can save up to 42% of energy. Load
balancing is also studied, where some traffic is balanced from one rat to another, here from LTE
to HSDPA as LTE can use cell DTX, and results showed an improvement of 29%. However,
30

shared power amplifiers could be used only for big sites, since femtocells probably won’t be
co-sited with other technologies.

In [56], Tombaz et al. investigated the new air interface in terms of energy performance. LTE
and 5G NR, in that paper called 5G-NX, have been tested at 2.6 and 15 GHz. A real life scenario
inspired from Asian cities is used, along with beamforming for 5G. It is shown that there is
by average a 65% decrease in power consumption for 5G NR at 15 GHz compared to LTE at
2.6 GHz. Moreover, even using LTE and NR together offers a reduction of power consumption
compared to LTE alone (35% decrease) at high load. Furthermore, DTX is shown to offer a
66% percent decrease of power consumption at low load and 13% decrease at high load, which
is similar to the previous studies.

2.5 Metrics

Metrics are needed in order to quantify power or energy efficiency. The most common one,
simply call energy efficiency, corresponds to the quantity of information transmitted per unit of
energy, usually Joule. It is expressed in [bit/J]. It can also be expressed as the bitrate offered
per power consumed, in [bps/W]. Some work also quantifies energy efficiency in [J/bit], as
the quantity of power for 1 bit transmitted. It is sometimes called Energy Consumption Ratio
(ECR). Their usage mainly depends of the objective of the study: maximisation or minimisation.
At the system level, the energy efficiency is simply defined as the sum of data transmitted for a
unit of time over the entire consumption of the system during that same period of time.

Extensions of these metrics can sometimes be found. The energy efficiency per area, expressed
in [bit/J/km2 ], can be more precise for urban scenario, especially for ultra-dense network, where
energy consumption can be non-homogeneous. In a similar fashion, the spectral efficiency
is defined in [bit/J/Hz]. Some other means of quantifying efficiency can be used in specific
scenarios, such as UDN. The Sleep Ratio is defined as the number of access nodes in sleep
mode over the total number of nodes. Similarly, the active ratio can be expressed as 1-sleep
ratio.

Many other metrics were also defined. Power Usage Efficiency (PUE) came from datacentre
evaluation and can be used to express the efficient usage of energy at the base station level.
Daily consumed energy can assess the evolution of energy consumption throughout the day,
whereas averaged daily consumed energy can differentiate two deployments.
31

3 Models and Multi-Connectivity Algorithms

This chapter details the implementation done throughout this work. Section 3.1 lists all the
models used in our tools. In 3.2, we present how multi-connectivity is implemented. We present
our vision on the algorithm steps, and how we can define schemes. Section 3.3 goes further and
a detailed description of our proposals are given.

3.1 Models Implementation

Our study relies on the usage of a simulator tool for the evaluation. We decided to use a solution
developed and used internally at Ericsson. Our tool is a Matlab-based discrete time event simu-
lator. Since this is a property of Ericsson, no more information can be disclosed. We however
present below the parameters or models we modified for this thesis.

3.1.1 Handover Model

The Handover (HO) refers to the procedure of transferring user resources from one cell to an-
other. In 3G, the upkeep of the connection is assured by so-called soft-handovers [70]. In
4G, only hard handovers are available, being mainly due to the usage of OFDM (Orthogonal
Frequency-Division Multiplexing), an orthogonal modulation. However, the seamless transition
must also be assured in LTE.
3 types of handovers are defined for LTE:

• Intra-LTE Handover

• Inter-LTE Handover

• inter-RAT

The first type corresponds to the scenario where both eNodeB are within the same LTE network.
In this case, the X2 interface is used when available, otherwise the S1 interface is used through
the evolved packet core (EPC).
The second type happens when the user also needs a handover the Mobility Management Entity
(MME) and/or Serving Gateway (S-GW).
Finally, LTE implements the possibility for the user to switch from one Radio Access Technol-
ogy to another, i.e. from LTE to 5G, or UMTS to LTE.
32

Studies have been carried out to investigate handovers in LTE-A and 5G. Ferng and Huang
[71] proposed a new handover scheme for heterogeneous networks, while Vasudeva et al. [72]
investigated HO failures by analysis and Lee et al. [73] showed the impact of TTT on HO
performance.

Some radio quality measurements are needed in order to instantiate a handover procedure. Dif-
ferent metrics have been specified by 3GPP [74] for this purpose. We list here a few of them:

• Reference Signal Received Power (RSRP)

• Received Signal Strength Indicator (RSSI)

• Reference Signal Received Quality (RSRQ)

• Signal over Interference plus Noise Ratio (SINR)

• etc...

In our work, we decided to use the Signal over Interference plus Noise Ratio (SINR) as our
metric to detect and initiate the HO procedure. This choice is due to our interest in allowing both
intra-LTE and inter-RAT handovers. Propagation losses, transmitted power and interferences
must be taken into account when selecting the cell.

The SINR is mathematically defined by:

Gc,u ∗ Ptx,c
γc = P (3)
ci 6=c Gci ,u ∗ Ptx,ci + N

where u is the user, c,ci are base stations,Gc,u the channel gain between the base station c and
the user u, and Ptx,c the transmitted power of base station c.
Extended to multiple cells transmitting the same information (in CoMP):
P
c∈Cu Gc,u ∗ Ptx,c
γCu = P (4)
ci 6∈Cu Gci ,u ∗ Ptx,ci + N

where Cu is the set of base stations transmitting the useful data.

In the handover procedure, the event A3 describes the time where the connected cell signal falls
below the signal of another cell. In order to avoid a ping-pong effect of handovers, a hysteresis is
defined. The ping-pong effect relates to the case where a user switches repeatedly between two
cells in a very short period of time. If the SINR γc stays below γt +hysteresis for a certain period
of time called Time To Trigger (TTT), the handover procedure is initiated. After a texec time and
33

Figure 9. Handover Procedure

if the connection does not have a failure, the handover is successful. Figure 9 represents a HO
procedure and table 1 contains all the parameters related to HO.

Mathematically, we have:

γc (t) < γt (t) + hysteresis; for t0 − T T T < t < t0 . (5)

where γc (t), γt (t) the SINR value of the connected BS and the target BS respectively, and T T T
the Time To Trigger.

Table 1. Handover Parameters

Parameters Values
Hysteresis [dB] -2
Time To Trigger [ms] 50
Execution Time [ms] 40

3.1.2 Radio Link Failure Model

Radio Link Failure (RLF) denotes the status where a wireless connection is considered dropped.
Its definition hasn’t changed much from 3G to 4G [75, 76], and is unlikely to evolve in a drastic
34

way for 5G.

A link is considered in RLF if the received SINR is below a threshold Qout for a certain period
of time noted TRLF . If the signal goes higher than the threshold during that time, the counter is
reset and the link is not in failure. If, after that time, the signal is still below the threshold, the
link is considered in RLF and the connection is dropped. Therefore, the user has to initiate a
new connection procedure, which can take some time. Table 2 contains the values used in our
model.

Mathematically, we write:

γc (t) < Qout ; for t0 − TRLF < t < t0 . (6)

where Qout is the threshold and TRLF is the time after which the connection is dropped.

Table 2. Radio Link Failure Parameters

Parameters Values
Qout [dB] -5
TRLF [ms] 50
Reconnection time [ms] 1300

3.1.3 Sleep Mode Model

As presented in section 2.4.2, will be improved in 5G. For our implementation, we follow the
concept of sleep levels from [57] that we already discussed. Figure 10 shows our implementa-
tion. Unlike Debaillie’s paper, we define only two levels: micro-sleep and deep-sleep.

As defined in equation (2), the deep-sleep mode consumption is set as a fraction of the micro-
sleep consumption by a factor δNR . For our simulations, we choose δNR = 0.29, as used in
[56].

We define the sleep and wake-up procedures as the following:

• After 10ms in active mode and without any transmission, the base station goes to micro-
sleep mode;

• The base stations wakes up directly if it needs to transmit something. Without any trans-
mission, the base station stays in micro-sleep for a maximum duration of 10ms. After that
35

Figure 10. Sleep Mode Model for 5G

time, it goes into deep-sleep.

• The base station stays in deep-sleep until it needs to transmit something. If data arrived,
it directly goes to active mode.

The consumption values we use are slightly higher than real consumption. It allows us to ac-
count within the sleep consumption small wake-up phase. No wake-up transition phase is mod-
elled as its duration would be inferior to the simulation time step.

3.1.4 Network Related Models

In this work, LTE or the LTE evolution is considered using the 2 GHz carrier. As stated in the
related work, 5G NR will likely use frequencies around 30 GHz. Therefore, we use the 28 GHz
carrier for NR, which seems to be the global consensus. In order to model the higher capacity
of the new air interface, we use a bandwidth of 100 MHz, unlike LTE which is modelled with a
carrier bandwidth of 20 MHz.

Since LTE is used as the macro-layer, we use tri-sector antennas, as often deployed in reality.
We also set the height to 30 meters, such as the cells are a few meters above the rooftop level,
or at the top of a tower. As 5G NR is used as small cells, we use a simple omni-directional
model for the antenna. The antenna gain is set to 0 dBi, such as there is no gain compared to
36

Table 3. Network Parameters

Parameters LTE 5G NR
Frequency [GHz] 2 28
Carrier Bandwidth [MHz] 20 100
Antenna Type Tri-sector Omni
Antenna Gain [dBi] 15 0
Antenna Height [m] 30 10
NS (Number of Sectors) 3 1
NC (Number of Carriers) 4 4
P0 [W] 130 56
∆P 4.7 2.6
Pmax [W] 20 6.3
Psleep [W] 75 39
δNR 0.29
Number of Sites 3 61
Inter Site Distance 400 100

the isotropic antenna. The height is set to 10 meters, which can be seen as an antenna on a
façade. Sometimes, antenna are also considered at the top of a lamp post. This is the usual
height defined in the METIS guidelines.

For the consumption model, the EARTH framework defined a set of values for different size
of cells [52]. LTE values are based on the macro-cell power decomposition, whereas 5G NR
are taken from the micro-cell values. Table 3 lists all the different network parameters used
in our simulations. Both radio access technologies will use four carriers in this thesis. Carrier
Aggregation is allowed.

For the network deployment, we use a hexagonal grid positioning system. We therefore always
have a deterministic network. RAT are considered non co-sited as they have different inter-site
distances. Some cells might however have the same location; this is not a problem. Figure
11 represents the deployment used. Furthermore, we use a wrap around parameter, such as
the network is repeated at its edges. Cells at the edge receives interferences from the wrapped
network and user going out of the area are wrapped around.

The COST Hata model is used for radio propagations. As the original covers frequencies up to
2 GHz, we extend it via linear interpolation. Fading is modelled as shadowing, and follows a
Gaussian distribution with a mean of 0 dB and σ 2 = 5 dB.
37

Figure 11. Deployment Strategy

3.1.5 User and Traffic Models

A user is represented by its equipment, called user equipment, or UE. User mobility is defined
as a linear movement, i.e. the UE follows a straight line. A bouncing circle is used to ensure that
all UEs remain within the simulated area. When the user reaches the circle, it simply “bounces”
back towards the simulated region with a random angle.

The number of users remains fixed throughout a simulation. However, UEs can be in different
states. When connected to a base station, it is considered in active state. After some inactivity
period, the user can go into idle mode, which relax the signalling transmission between the
serving BS and the UE. As explained before, a UE can also be in RLF if the received signal is
too weak.

Users also move with a fixed speed. Unless stated otherwise, UEs are considered pedestrians
with a velocity of 3 km/h. Table 4 summarizes the different parameters. Once created, we keep
the user alive until the end of the simulation. After transmitting and a counter, the user goes to
idle mode.

Daily, weekly and even monthly consumption are well known and can be found in [77, 78].
Some model for future networks are also described for the METIS test-cases in [79]. Based on
38

Table 4. Users Parameters

Parameters Values
Number of UEs Fixed
Movement Model Linear, Bouncing Circle
Velocity Fixed
Traffic Type Bursty, FTP

these different solutions, we propose our own model which aims at modelling data traffic at a
smaller time scale, typically a few seconds or minutes.

As recommended by the simulation guidelines of METIS [79], a bursty user-driven model is


chosen. Specifically, the File Transfer Protocol (FTP) is selected for modelling data transfer.
Although, in their guidelines, the 3GPP FTP Model 2 [80] is used with a data size of 20 MB,
in order to model web browsing traffic. In our model, FTP is used in conjunction with different
packet size to realistically emulate the diversity of flows.

Packets are generated randomly with three different size (headers included): 1 MB, 10 MB or
100 MB. Packet size distribution follows a Poisson law, i.e. there are more chances of generated
a small packet rather than a big one. Two successive packets have independent distributions.
The inter-arrival time is non-periodic and also modelled following an exponential distribution.
The mean time between two packets is set to 3 seconds. Again, the inter-arrival time distribution
is independent of the packet size and between two successive packets.

Parameters of the traffic model are summarized in table 5.

Table 5. Traffic Model

Parameters Values
Traffic Type FTP, Bursty
Size Distribution Exponential
File Size 1 MB, 10 MB, 100 MB
Size Probability 5/8, 2/8, 1/8
Inter-arrival Time Distribution Exponential
Mean Inter-arrival Time 3 seconds
39

3.2 Multi-Connectivity Implementation

One of the objectives of this thesis is to develop energy efficient multi-connectivity schemes.
In this section, we describe the working principles of our proposals. For this, we present first a
breakdown of the mechanics of the MC procedure. We will then elaborate on each scheme.

3.2.1 Basic Algorithm Decomposition

In order to provide the best service and based on the specification of dual connectivity, we can
enumerate few principles that should be followed for the implementation:

• User association logic is based on one or a set of metrics.

• Channel quality, BS information and user information are the inputs available. They are
obtained through the same assessments and reports used in the handover procedure.

• As for handovers, ping-pong effects (consecutive connection, disconnection or switching)


should be avoided.

• The association procedure should be initiated at the network side. The user might have
the ability to request a specific connection, but the cell is the final decision maker.

• The base station serving the user (the master cell) is the one running the scheme. It can
receive measurements from target cells but should be the decision maker.

Each scheme might use or combine metrics differently. Different metrics are also available to
estimate the channel quality, as presented before. The X2 interface is used to share information
between two base stations.

Counters and hysteresis are commonly used to tackle the ping pong problem. In our work, we
solve this problem by isolating the connection and the disconnection procedures into two mech-
anisms. Counters are also used to avoid noisy metrics such as high variation in the channel
quality that could lead to early disconnection. Hysteresis can also be used within the mecha-
nisms.

Figures 12 and 13 are flowcharts representing respectively the connection and disconnection
procedures. It can be noted that both have a similar structure.
40

Following the reception of information, our inputs, a first selection is made. If a target is se-
lected, a Time To Trigger (TTT) counter, associated with a specific user, is started. When the
counter reaches the threshold and if that target is still available, the actual connection (or dis-
connection) procedure is initiated.

Figure 12. Association Procedure Flowchart

The scheme can either be implemented for serving one user only, and therefore the algorithm
will loop for each user, or it can be implemented to serve all the connected users at the same
time. Of course, taking into accounts all the UEs is better when trying to find a global optimal
solution.
41

Figure 13. Disassociation Procedure Flowchart

3.2.2 Classification of Metrics Usage

Numerous metrics can be taken as input. We decide to combine them into the following cate-
gories:

• Robustness: it typically regroups the channel quality metrics, such as RSRP, RSSI and
SINR. They can be used to provide reliable connection.

• Performance: the bitrate can be used as a metric for performance. Since the exact bitrate
cannot be estimated before establishing the connection, a simple estimation such as the
Shannon law can be useful. It allows connection to cell with high bandwidth (and possibly
high bitrate) but with signal strength not as good.

• Energy: Power required for transmitting the data for the user or the entire energy used to
42

keep the base station active can be used as metric as a way to estimate energy efficiency.
Power models need to be implemented within the scheme in order to provide the best
estimations.

• Cell Utilization: this class contains all the information relative to the target cell. The
cell state (active, sleep), its load or the number of users associated with it can be used for
traffic steering and load balancing.

By looking at the metrics used as input, it is possible to classify the purpose of a scheme. It
is also possible to combine metrics coming from different categories, in order to create an all-
purpose solution. However, in that case, a trade-off would be needed. Indeed, optimizing the
energy usage will have a negative impact on the performance or the robustness, and the most
reliable connection might not offer the best performance.

Differentiating schemes also let us compare them more easily. We can define a scheme as a
baseline if it takes input from one category only and if it uses simple logic. More strictly, we
can specify a baseline algorithm if it uses one single metric and if its logic consists of optimizing
(by maximizing or minimizing) that input.

3.3 Presentation of our proposals

We present here five of the algorithms developed during our work. Table 6 show the relation
between these schemes and the classification introduced earlier.

It can already be noted that the first two algorithms follow the definition of a baseline given
before. The naming convention tries to include the optimization principle and the metric used
when possible.

Table 6. Multi-Connectivity Schemes Parameters

Schemes Performance Robustness Energy Utilization


Max Bitrate x
Max SINR x
Max Bitrate-EE x x
Max Clustered-Bitrate x x x
Analytic Hierarchy Process x x x x
43

For each of these algorithms, we explain how the secondary cell target selection and cell asso-
ciation and disassociation are done.

3.3.1 Max Bitrate

The Max Bitrate scheme can be seen as a simple algorithm aimed at maximizing the perfor-
mance of a user. The bitrate is used as the sole metric. The bitrate here is computed with the
Shannon Law, such as R(ci ) = BWi ∗ log2 (1 + SINR). Here, we use the bandwidth of the entire
carrier as BWi .

Mathematically, the target selection of a secondary cell can be written as:

ctarget = argmax R(ci ) = {ci | ci ∈ Cθadd ⊆ C} (7)

Where ctarget is the chosen target, ci the cell numbered i, R(ci ) the bitrate offered to the user if it
was connected at the cell ci , C the set of all cells and Cθadd a subset of C. Its definition is given
later. argmax is the function returning the index of the maximum value. Here, argmax R(ci )
returns ci when R(ci ) is the maximum.

We choose the target cell from a subset and not the entire set of cells. Indeed, we added a
condition where the bitrate of the target cell should be higher than a certain threshold. This is
also done in the disassociation procedure and thus create a hysteresis to avoid the ping pong
effect.

Moreover, without this condition, there is no proof that the target cell does actually provide
some improvement, as we only know it is the best available. We then define the threshold as
dependent to the bitrate provided by the master cell. In this manner, we ensure that the target
cell will provide a real improvement.

We define the subset as:

Cθadd = {cmaster ∈ C ∧ ∀ci ∈ C : R(ci ) > R(cmaster ) ∗ θadd } (8)

thus, the set of targets is composed of cells offering a bitrate higher than at least θadd times higher
than the master bitrate. Since the master cell is chosen based on its SINR, it is possible to obtain
targets when θadd > 1. However, it is quite rare. In our implementation, we choose θadd = 0.8.
This means we choose a secondary cell that will provide at least 80% as much throughput as the
44

master cell. Although it means we can obtain a secondary cell with a worse bitrate, as we can
request resources from both base station, we should obtain a higehr user throughput.

Regarding the disassociation procedure, we follow a similar principle. We disconnect all the
secondary cells with a bitrate below a certain threshold. That limit is also defined based on the
master cell bitrate. Thus, we obtain our hysteresis. The subset of target to disconnect are defined
as:
Cθrmv = {cmaster ∈ C ∧ ∀ci ∈ C : R(ci ) 6 R(cmaster ) ∗ θrmv } (9)

where θrmv < θadd .

We select the value θrmv = 0.5 as threshold.

Finally, the connection and disconnection procedure only initiate the necessary mechanisms for
the transactions and allocation for the user.

3.3.2 Max SINR

In a similar manner, we define Max SINR as a baseline for schemes trying to improve robustness
or reliability. Schemes such as Max RSRP or Max RSSI could also be used to serve the same
purpose.

The logic behind the algorithm is strictly the same. A target with the best SINR is selected from
the subset of cells. Here, the subset is defined as cells with a SINR higher than a threshold,
which depends on the SINR of the master cell. We have:

ctarget = argmax γ(ci ) = {ci | ci ∈ Cθadd ⊆ C} (10)

Cθadd = {cmaster ∈ C ∧ ∀ci ∈ C : γ(ci ) > γ(cmaster ) ∗ θadd } (11)

With γ(ci ) the SINR of the cell numbered i. In this algorithm, we set θadd = 0.25 ( = −6dB).
The disassociation procedure also follows the same concept:

Cθrmv = {cmaster ∈ C ∧ ∀ci ∈ C : γR(ci ) 6 γ(cmaster ) ∗ θrmv } (12)

where θrmv < θadd , and we set θrmv = 0.125 ( = −9dB).

Such scheme can be useful when reliability is critical. Moving nodes, especially at high speed,
might lose connection during the handover establishment. Multi-connectivity can offer a sec-
45

ondary link that can still be used during the handover of the master cell. A scheme with a similar
principle is presented in [46], although the architecture of the algorithms differs. In their work,
Tesema et al. investigates how MC can tackle RLF. Moreover, multiple settings, which corre-
sponds to our thresholds, are tested. We refer the reader to that paper if it wishes to understand
the effect of the threshold over the RLF rate.

3.3.3 Max Bitrate-EE

One observation that can be made while using one of the above scheme is that energy efficiency
is not quantified. As more base stations stay in active mode to provide the increased performance
or reliability, the overall energy consumption of the network is also increased. An estimation of
the impact of multi-connectivity must be made during the selection procedure. Thus, we could
reject associations that would be too costly in terms of energy. We therefore introduce another
condition, that we refer as the power condition.

Mathematically, we write our condition as:

CpowerCond = {cmaster ∈ C ∧ ∀ci ∈ Cθrmv : PC|cmaster (t) > PC|ci (t)} (13)

Where PC|cmaster (t) > PC|ci (t) simply means that the power consumed by the network knowing
that the transmission is done using the master cell must be higher that the power consumed if the
transmission was done with that secondary cell. It can also be noted that this condition is applied
to the set satisfying the precedent condition, where the bitrate is already above the threshold.

The validity of this inequality is presented in our theoretical study, later in this thesis. We can
however already point out that this condition requires either the knowledge of the size of the
transmission, or an estimation. A second important point is that this inequality is tested for
one user only. It doesn’t evaluate the case where two users could use the same secondary cell,
therefore lowering even more the energy cost.

Two solutions can be implemented to tackle this problem: treating all the users with the same
target together for the condition, or using a cluster approach. The latter is presented in the next
scheme.

After choosing the targets of each user connected, the condition is tested for each secondary
cell, considered all the possible users that can connect to it. If the condition is true, all these
users can continue the procedure. If not, that target is removed. The interest of this approach is
also studied in the theoretical section. It is the one we implemented.
46

The disassociation procedure follows the same principle as described in the precedent schemes.

3.3.4 Max Clustered-Bitrate

Whereas the precedent scheme does take into account the impact of MC on the energy con-
sumption, it tries to optimize the performance of only one user at a time without considering all
the associations as a whole. This principle is more of a local optimization than a global one.
Moreover, local optimization of some subset of the problem is equivalent to the global optimum
only if each subset is independent, which is not the case here when users might share the same
secondary cell.

For this scheme, called Max Clustered-Bitrate, we introduce another metric, categorized in the
Cell Utilization class: the number of links. The goal here is not only maximize the performance,
but also trying to group users in such a way that they will connect to the same secondary cell.

Mathematically, we create a 2D matrix called A of size M × N , where M is the number of


users and N the number of cells. It has to be noted that A is not a square matrix. In order to
limit the size for the operation, a clustering of the network can be done. Typically, all the small
cells within the coverage of a macro-cell can form a cluster. Cells can also be part of multiple
clusters, such as the one being at the edge of a macro-coverage. This clustering approach is the
reason why this scheme as this name.

The matrix A is then filled following this principle:



1 if Ri (cj ) > Ri (cmaster ) ∗ θadd
Aij = (14)
0 otherwise

Where Ri (.) is the bitrate for the user i.

Therefore, we have Aij ∈ {0; 1}. By choosing the column with the highest sum, you can
maximize the number of users connected to the same target, or:
X
ctarget = argmax Aij = {cj | cj ∈ Ccluster } (15)
i

However, experiments showed that maximizing the number of associations only does not pro-
vide a great performance. We improve that scheme by introducing a second matrix, B, again of
47

size M × N , defined as:


Ri (cj )
Bij = Aij ∗ (16)
Ri (cmaster ) ∗ θadd
We thus have Bij ∈ {0} ∪ [1; +∞[, although it is unlikely that Bij is higher than the order of
10.

Now, the matrix is weighted by the bitrate, more specifically the improvement offered compared
to the master cell. We can simply modify (15) and use B instead:
X
ctarget = argmax Bij = {cj | cj ∈ Ccluster } (17)
i

Finally, we also apply the power condition explained for the precedent algorithm.

For this scheme, we reuse the same θadd as presented before. We also reuse the same principle
for the disassociation procedure.

3.3.5 Analytic Hierarchy Process

The Analytic Hierarchy Process (AHP) is a technique for multi-criteria decision developed by
Thomas L. Saaty [81, 82].

In this framework, a goal is divided in sub-problems, or criterion. Each of them is given a


weight, with the sum of the weights equal to 1. Each tested solution are then evaluated regarding
each criteria and given a score between 0 and 1. The final score of each alternative is then
computed with a simple weighted mean of each individual score.
X
S(x) = wi ∗ Si (x)
i

Where S(x) is the final score of the solution x, wi the weight of the criteria i and Si (x) the
individual score of x for the criteria i.

AHP can be a quite powerful tool because it enables decision over multiple criterion in a simple
and fast manner. Moreover, the weights can be easily modified, allowing the same implementa-
tion with different parameters to be deployed in various scenarios to tackle drastically opposite
goals. We can, for example, think of a use case with fast moving nodes where the reliability
is critical using AHP, as well as a second test-case with stationary nodes where performance is
48

more important.

In our implementation, we opted for three sub-problems: performance, characterized by the


bitrate, robustness, with the SINR, and energy efficiency, which is defined using the power
condition presented before. More metrics could be used, such as the cell load or other channel
quality measures, but we consider that these three already offers a wide range of possibilities.
Mathematically, we have a matrix S of size M × N such as:

Sij = wperformance ∗ Sbitrate (i, cj ) + wrobustness ∗ SSINR (i, cj ) + wEE ∗ SEE (i, cj ) (18)

Ri (cj )
where Sbitrate (i, cj ) = ,
max Ri (ck )
ck ∈C

γi (cj )
SSINR (i, cj ) = ,
max γi (ck )
ck ∈C

1 if PC|cmaster (i) > PC|cj (i)
and SEE (i, cj ) =
0 otherwise

By dividing each individual score by its maximal value, we normalize its value and ensure it is
in [0;1]. After we obtained the score for each cell and for each user, we remove the possible
connections with a bad score by applying a threshold. We then select the best scores and start
the procedure.

ctarget (i) = argmax Sij = {cj | cj ∈ CSi } (19)

CSi = {∀cj ∈ C : S(i, cj ) > θSmin }

In our implementation, we fixed the minimal score θSmin = 0.5. Since our main objective is to
increase the energy efficiency, we also defined our weights such as wperformance = 0.5, wEE = 0, 5
and wrobustness = 0 since robustness would only decrease performance in our scenario.

Regarding the disassociation procedure, we simply disconnect all users with a score below a
certain threshold. Here the threshold is set to θSrmv = 0.3.
49

4 Analytical Study of Energy Efficiency

This chapter contains our mathematical analysis of multi-connectivity. In 4.1, we give a general
expression of energy efficiency and try to find its bounds. Section 4.2 presents the power condi-
tion mentioned earlier and shows how it can be used to determinate algorithm performance.

4.1 Energy Efficiency’s Expression

In telecommunications, the ITU defines energy efficiency as “the relationship between the spe-
cific functional unit for a piece of equipment and the energy consumption of that equipment”
[83]. At the network level, this definition corresponds to the quantity of data transmitted at the
physical layer by an equipment per unit of energy. It is expressed in [bit/J]. We use this defini-
tion to express energy efficiency at the network level, in scenarios with only one connection per
user and scenarios where the multi-connectivity feature is activated.

4.1.1 Single Connectivity

At the radio access network level, we define energy efficiency as the ratio between the quantity
of transmitted data by the network and the operating cost of equipment in terms of power. For
the latter, we discard intermediary components, and only consider the consumption of base
stations, as it represents an important share of the operational cost of a cellular network [49].
For the former, we consider the system throughput as the sum of the transmission occurring
through the user plane between a user and its serving cell. Control plane transmission are thus
discarded as they not represent useful data from the application point of view. We also make
this choice as the impact of multi-connectivity in terms of additional transmission is not known
yet. We can therefore write the energy efficiency of a single connectivity (SC) system as:

RSC
ηSC = (20)
PSC
where RSC is the system throughput (for a single connectivity scenario) and PSC the power
consumed by the system, such as:

X
RSC = Rmaster (u)
u∈U
50

where u is a user, U the set of active users in the network, Rmaster (u) the downlink bitrate for the
user u by its master cell, and

X X
PSC = Psleep (c) + (P0 (c) − Psleep (c))
c∈C c∈Cused
X
+ Ptx (c, u)
u∈Uc ,c∈Cused

where c is a cell, C the set of all cells within the network, Cused a subset of C containing all the
active cells, Uc the subset of users connected to the cell c, and Ptx (c, u) = Ptx ∗ ∆P from now
on. These expressions are equivalent to say that the system throughput is the sum of each user
downlink throughput and the power consumed is equal to the power consumed by the entire
network in sleep mode, plus the cost of turning on the active BSs and the cost of transmitting
data to the user, as defined in Eq. (1).

4.1.2 Multi-Connectivity

This expression covers all scenarios where only one link is allowed per user. For a scenario with
multi-connectivity, the expression must be extended. We therefore write:

RMC
ηMC = (21)
PMC

where
X X X
RMC = Rmaster (u) + Rc (u)
u∈U u∈UMC c∈Csec,u

where UMC is the set of active users using multi-connectivity, Csec,u is the set of secondary cells
the user u is connected to, and Rc (u) the downlink bitrate from c to u. The expression for PMC
is similar to the one given for PSC , except that the set Cused contains more elements, such as
PMC 6= PSC . More precisely, it contains the set of master cells that was already present in single
connectivity as well as the set of secondary cells. Furthermore, we have RMC ≥ RSC such as
MC SC
Rmaster (u) 6= Rmaster (u) ∀u ∈ U . We can further expand as:
X
RMC = RSC − δR (u) ∗ Rmaster (u)
u∈U
X X (22)
+ Rc (u)
u∈UMC c∈Csec,u
51

where δR (u) ∈ [0, 1] represents the offloading factor between the master cell and secondary
cells and equals 1 when RMC = RSC , such as traffic consisting in small payloads. For high
payload, δR (u) will be equal to 0 and represents a scenario where MC is useful.

This general expression, however, does not provide much help to determine whether MC is
more or less energy efficient than a classical scenario. Therefore, we propose here to estimate
the lower and upper bounds of this expression, and compare them to the energy efficiency of the
same scenario with single connectivity.

4.1.3 Lower Bound

Let us consider a scenario where the usage of multi-connectivity induces an offloading of data
from the primary cell towards the second. Moreover, let us also consider that this traffic steering
does not provide any improvement towards the capacity, such as:

RMC = RSC (23)

Regarding the power consumption, as in Eq. (22), the offloading induces a decrease in the
transmitted power of the master cells, and an increase in the secondary cells, written as:
X
PMC = PSC + (P0 (c) − Psleep (c))
c∈Csec ,c6∈Cmaster
X X
+ Ptx (c, u) − δP (u) ∗ Ptx (c, u)
u∈Uc ,c∈Csec u∈Uc ,c∈Cmaster

where Csec is the subset of cells containing all the BSs used for secondary links, Cmaster the
subset of C containing all the BSs used as master cell, Uc the subset of users connected to the
cell c and δP (u) the offloading factor associated to the user u. We can simplify the expression
as:
PMC = PSC + ∆MC − ∆offload (24)
P P
where ∆offload = u∈Uc ,c∈Cmaster δP (u)∗Ptx (c, u) and ∆MC = c∈Csec ,c6∈Cmaster (P0 (c) − Psleep (c))+
P
u∈Uc ,c∈Csec Ptx (c, u).

From Eq. (23) and Eq. (24), we can express the energy efficiency as:

RSC
ηMC = (25)
PSC + ∆MC − ∆offload
52

where the efficiency of multi-connectivity is therefore confined to its ability to less consuming
transmitters. In case of non-ideal backhaul, its impact should also be assessed within ∆MC .

4.1.4 Low Performance Bound

A second and less general bound can be defined where we consider this time an increase in
capacity with multi-connectivity. We define later in this paper a disconnection threshold, below
which a secondary link is not maintained. It is expressed as a ratio of the bitrate offered by
the master cell, such as threshold = θrmv ∗ Rmaster (see section 3.3.1). Therefore, we consider a
worst-case scenario using the following assumptions:

Assumption 1: Each secondary cell serves at most one user.


This assumption ensures us that the power consumption is maximized. Moreover, it is a reason-
able one with few users, such as in an ultra-dense network.

Assumption 2: We consider that the multi-connectivity feature induces an increase in the sys-
tem throughput such as RMC > RSC .
This assumption is sound and reasonable if the traffic does not consist of small packets only.

X
RMC = Rmaster (u)
u∈U
X X
+ θrmv ∗ Rmaster (u)
u∈UMC c∈Csec,u

≈ RSC ∗ (1 + |UMC | ∗ mean(|Csec,Umc |) ∗ θrmv )


≈ RSC ∗ (1 + ∆C,lwr )

Such as:
∆C,lwr = |UMC | ∗ mean(|Csec,Umc |) ∗ θrmv (26)

where |UMC | = Pr(u ∈ UMC ) ∗ |U | is the number of users with multiple links, and can also
be defined as the factor of the total number of users and the probability than one user can have
multiple links, mean(|Csec,Umc |) the average cardinality of Csec,u , ∀u ∈ UMC , which corresponds
to the mean number of additional connections for a user in MC. For the worst case scenario, this
parameter is equal to 1.
53

In a similar manner, PMC can be simplified as:


X
PMC = PSC + (P0 (c) − Psleep (c))
c∈Csec ,c6∈Cmaster
X
+ Ptx (c, u)
u∈Uc ,c∈Csec
X
= PSC + [P0 (c) − Psleep (c) + Ptx (c)]
c∈Csec ,c6∈Cmaster

= PSC ∗ (1 + ∆P,lwr )

where c ∈ Csec , c 6∈ Cmaster represents a cell element used as a secondary cell that was is not
used as a master cell for another user, and ∆P,max the evolution of the consumed power related
to MC such as: P
c∈Csec ,c6∈Cmaster [P0 (c) − Psleep (c) + Ptx (c)]
∆P,lwr = (27)
PSC

We can thus describe energy efficiency of a multi-connectivity scenario as:

1 + ∆R,lwr
ηSC ∗ ≤ ηMC (28)
1 + ∆P,lwr

where ∆R,lwr and ∆P,lwr are defined in Eq. (26) and Eq. (27) respectively.

4.1.5 Upper Bound

In a similar manner, one can wonder what is the highest improvement possible using multi-
connectivity. Such upper would be dependant on the architecture deployment and the user
distribution, but is nonetheless interesting to fully understand the interest of MC. We again
assume two characteristics:

Assumption 1: Each secondary cell serves at most one user.

Assumption 2: Each user maintain the maximal amount of secondary link and each link pro-
vides the maximum achievable bitrate, following assumption 1.
Such as we have the bitrate is maximized for every user.
54

We therefore obtain:
X X X
RMC = Rmaster (u) + Rmax (c, u)
u∈U u∈UMC c∈Csec,u

≈ RSC + |U | ∗ max(|Csec,Umc |) ∗ mean(Rmax (c, u))


≈ RSC ∗ (1 + ∆R,lwr )

where Rmax (c, u) is the maximum achievable bitrate between u and c, max(|Csec,Umc |) the maxi-
mum number of secondary link that a user can handle, mean(Rmax (c, u)) the average maximum
achievable bitrate taken for all users and

|U | ∗ max(|Csec,Umc |) ∗ mean(Rmax (c, u))


∆R,upr = (29)
RSC

The power consumption follows a similar evolution as for the lower bound, such as:
X
PMC = PSC + [P0 (c) − Psleep (c) + Ptx (c)]
c∈Csec ,c6∈Cmaster

= PSC ∗ (1 + ∆P,upr )
P
c∈Csec ,c6∈Cmaster [P0 (c) − Psleep (c) + Ptx (c)]
∆P,upr = (30)
PSC
wehere the power transmitted is maximized as we are using all available resources.

Therefore, from Eq. (29) and Eq. (30), we can express the energy efficiency of MC as a bounded
value such as:
1 + ∆R,upr
ηMC ≤ ηSC ∗ (31)
1 + ∆P,upr

4.1.6 Ensuring Energy Efficiency

Following Eq. (28) and Eq. (31), the energy efficiency of multi-connectivity can be written as:

1 + ∆R,lwr 1 + ∆R,upr
ηSC ∗ ≤ ηMC ≤ ηSC ∗ (32)
1 + ∆P,lwr 1 + ∆P,upr

It can be of great interest to compare the lower bound against the evolution of ηSC , such as we
can decide under which conditions Mc is more energy efficient than single connectivity. To do
55

so, we search parameters such as:


1 + ∆R,lwr
≥1 (33)
1 + ∆P,lwr
⇔ ∆R,lwr ≥ ∆P,lwr
P
[P0 (c) − Psleep (c) + Ptx (c)]
c∈Csec ,c6∈Cmaster
⇒ θrmv ≥ (34)
PSC ∗ |UMC | ∗ mean(|Csec,Umc |)

By choosing the right value for the disconnection threshold, it is therefore possible to ensure
that MC is more energy-efficient than single connectivity.

4.2 Power Condition for Energy Efficiency

In this section, we want to express a condition on the power consumption to ensure that transmit-
ting on a secondary cell is more energy efficient than transmitting the same information directly
from the master cell. For this, we compare the energy consumption in both scenarios, i.e. the
information is transmitting using the master connection or the information is transmitting using
a secondary connection.

4.2.1 Definition of a Power Condition

As we want an energy-efficient offloading, we start by considering a transmission more efficient


with the secondary connection.

Our condition can be written as:


PC|M (t) > PC|S (t) (35)

With PC|M (t) and PC|S (t) the power consumed at time t when the transmission is done on the
master or secondary connection respectively.

We can decompose it as the consumption of each element:

⇔ PC,master|M (t) + PC,sec|M (t) > PC,master|S (t) + PC,sec|S (t)

With PC,master (t) and PC,sec (t) the power consumed by the master BS and the secondary BS
respectively.

We can also express the power consumption by the difference of consumption at time t and t-1,
56

since the consumption at t-1 did not depend on the link selection.

⇔PC,master|M (t) − PC,master (t − 1) + PC,sec|M (t) − PC,sec (t − 1)


> PC,master|S (t) − PC,master (t − 1) + PC,sec|S (t) − PC,sec (t − 1)

Assumption: Let’s consider that the traffic on the non-selected BS is constant from t to t-1.

Although this might not be true in reality, we want to assess the impact of the user u only. The
event of a BS going to active mode for another transmission cannot be planned when choosing
our place of transmission. This assumption means that if a BS is sleeping at t-1, it will stay
asleep at the time t if we don’t choose it.

Following this assumption, we can simplify our inequality:

⇔ PC,master|M (t) − PC,master (t − 1) > PC,sec|S (t) − PC,sec (t − 1)

Using the power models defined in (1) and (2), we obtain the formal expression:

P0,master + PTX,master (t − 1) + PTX,u|M (t) ∗ ∆P,master − PC,master (t − 1)

> P0,sec + PTX,sec (t − 1) + PTX,u|S (t) ∗ ∆P,sec − PC,sec (t − 1)

With P0 the constant consumption in active mode, PTX the radiated power, PTX,u|M or S the radi-
ated power related to the transmission for user u, ∆P the slope relative to the dynamic power
consumption and PC (t) the power consumed at the time t.

We define ∆PC (t) as the evolution of the power consumption, i.e. ∆PC (t) = PC (t) − PC (t − 1).
We use this definition only to simplify the expression, as it does not change the content. We can
now formulate our condition as:

∆PC,master (t) + PTX,u|M (t) ∗ ∆P,master > ∆PC,sec (t) + PTX,u|S (t) ∗ ∆P,sec (36)

We can derive some simpler expressions out of (36) if we consider simpler scenarios. For
example, by considering a scenario where the master BS does not go to sleep and with the
precedent assumption of constant traffic, we obtain:

PTX,u|M (t) ∗ ∆P,master > ∆PC,sec (t) + PTX,u|S (t) ∗ ∆P,sec


57

Another simplification comes from the assumption that the secondary BS was in sleep mode at
time t-1. We could then obtain:

PTX,master|UE (t) ∗ ∆P,master > P0,sec − Psleep,sec + PTX,master|UE (t) ∗ ∆P,sec

We clearly see that in this case, the cost of transmitting on the master BS must be higher than
the cost of putting the secondary BS in active mode and transmitting with it.

4.2.2 Power Condition based on the offered capacity

One interesting development of (36) could be to express it as an inequality involving the ca-
pacities of both the master and the secondary BS. Its interest relies in the fact that selection of
secondary cells is also based on the bitrate achievable if connected to the targeted BS. Starting
from (36), we can write:

(∆PC,master (t) − ∆PC,sec (t)) + PTX,u|M (t) ∗ ∆P,master > PTX,u|S (t) ∗ ∆P,sec

As defined in the power models (1) and (2), PTX represents the radiated power, and is linear with
the load. We can also write it as PTX = load ∗ Pmax where load ∈ [0, 1] and Pmax the maximum
radiated power. Moreover, we can also express the load relative to the user u as the number of
resource blocks needed to transmit data to this user, which can be written as Cpayload u
RB ∗nbRB
where
payloadu is the number of bits for the transmission for u, CRB the capacity (i.e. number of bits
that can be coded) in one resource block and nbRB the total number of resource blocks of the
base station, a constant.

We can thus write:

payloadu (t) payloadu (t)


∆PC,sys (t) + (t) ∗ Pmax,M ∗ ∆P,M > ∗ Pmax,S ∗ ∆P,S
CRB|M (t) ∗ nbRB CRB|S (t) ∗ nbRB

with ∆PC,sys (t) = ∆PC,master (t) − ∆PC,sec (t) the evolution of the power consumption for both
BSs and CRB|M (t) the capacity of one RB on the master BS. Let us consider the case where the
payload is inferior to the maximum offered throughput, as the latter expression does not depend
on the throughput and is equivalent to Eq. (36).

We can now try to isolate CRB|S (t), by writing:


 
1 1 nbRB Pmax,M ∗ ∆P,M
6 ∗ ∆PC,sys (t) ∗ +
CRB|S (t) Pmax,S ∗ ∆P,S payloadu (t) CRB|M (t)
58

Which leads to:

Pmax,S ∗ ∆P,S ∗ payloadu (t)


CRB|S (t) > CRB|M (t) ∗ (37)
Pmax,M ∗ ∆P,M ∗ payloadu (t) + nbRB ∗ CRB|M (t) ∗ ∆PC,sys (t)

This inequality shows that the power condition is correlated with the offered capacity of each
BS as well as the impact of the transmission over the system, here ∆PC,sys (t) = ∆PC,master (t) −
∆PC,sec (t), which for example means that putting a sleeping secondary BS in active mode is
power efficient only if the offered capacity of such BS is much greater than the capacity offered
by the master BS.

4.2.3 Power Consumption Comparison of a Clustered-based and non Clustered-based


Algorithms

In this section, we try to show that using a cluster approach is more power-efficient than trying
to have power efficient connection for each target separately. By doing this, we are demonstrat-
ing that the EE-Weighted-Clustered algorithm is more power efficient than the EE-Best-Bitrate
algorithm.

Here, we define power efficient as an algorithm that consumes less energy, in opposition to en-
ergy efficient, which aims at increasing the bit-per-Joule metric.

First, the inequality in (36) can be extended to multiple users. Considering multiple users trying
to connect to the same secondary BS, our condition can now be written as:
X
(P0,master|u − PC,master|u (t − 1)
u∈UM C

+ PTX,master|u (t − 1) + PTX,master,u|u (t) ∗ ∆P,master )
(38)
> P0,sec − PC,sec (t − 1)
!
X
+ PTX,sec (t − 1) + PTX,master|u (t) ∗ ∆P,sec
u∈UM C

With u a user, UM C the set of users that want to transmit in MC and Px,master/secondary|u the power
consumed by the master/secondary BS where u is connected.

From (36) and (38), we can write that the problem of reducing the energy consumption for the
59

Max Bitrate-EE algorithm and the Max Clustered-Bitrate. algorithm can be written as:

X 
PC (t) = min PC,u|master (t), PC,u|sec (t)
MCB u∈UM C

with PC|MBREE (t) the total energy consumption of the system at the time t using the Max Bitrate-
EE algorithm,u a user,UM C the set of users that want to transmit in MC,PC,u|master (t) and PC,u|sec (t)
the power consumed knowing that the data is transmitted to u by the master BS or by the sec-
ondary BS respectively, at time t, where the expression is conceived on the same principle as
(36), and:


PC (t) = min PC,u|master (t), PC,u|sec (t)
MCBR

where the expression is based on the same principle as (38), for the Max Clustered-Bitrate al-
gorithm.

We now want to study the evolution of the power consumption when a transmission for the
user u can be transmitted either on the master BS or a secondary BS. By using once again the
notation ∆PC (t) = PC (t) − PC (t − 1), we can have the expression:

∆PC,master (t) + ∆PC,sec (t) + PTX,u|M (t) ∗ ∆P,master ,
X 
∆PC (t) = min (39)
MBREE u∈UM C
∆PC,master (t) + ∆PC,sec (t) + PTX,u|S (t) ∗ ∆P,sec

And:
 P  
u∈UM C ∆P C,master (t) + ∆P C,sec (t) + P TX,u|M (t) ∗ ∆P,master ,
∆PC (t) = min P  (40)
MCBR ∆PC,sec (t) + u∈UM C ∆PC,master (t) + PTX,u|S (t) ∗ ∆P,sec

Considering a scenario where using the master cell is always more power efficient (which is
equivalent to a scenario without multi-connectivity), we have the same consumption. However,
we can define a scenario where transmitting using the secondary BS is always more power
efficient than using the master BS. With this assumption, we now have:

Su 6=Sv Mu 6=Mv
X X X
∆PC (t) = ∆PC,sec (t) + ∆PC,master (t) + PTX,u|S (t) ∗ ∆P,sec (41)
MBREE u,v∈UM C u,v∈UM C u∈UM C

with Su the secondary BS of user u, Mu the master BS of user u, Su 6= Sv meaning we sum only
60

once BSs used by multiple users, and:

Mu 6=Mv
X X
∆PC (t) = ∆PC,sec (t) + ∆PC,master (t) + PTX,u|S (t) ∗ ∆P,sec (42)
MCBR u,v∈UM C u∈UM C

We can clearly see that, for a common set UM C :

∆PC,MCBR (t) 6 ∆PC,MBREE (t)

since we try to group transmissions on the same secondary BS. This does not, however, prove
that Max Clustered-Bitrate is more energy efficient than Max Bitrate-EE. Consuming less power
is not the only impact on EE as shown before, and trying to group users on the same secondary
BS does actually impact the offered throughput negatively.
61

5 Simulation Results

This chapter deals with the evaluation of the simulation results. In section 5.1, we present the
methodology we followed during our simulations. Section 5.2 details our results on the user
reliability. Finally, in 5.3, we evaluate our algorithms.

5.1 Methodology

As real experiments are hard to perform given our study, we decide to follow the simulation
based research. The usual methodology for this can be decompose as: i. Model Design, ii.
Model Verification, iii. Simulation Design, iv. Simulation, v. Evaluation.

Chapter 3 presented how we designed the most important models our simulation relies on. The
verification of these models have been made internally. Since most of our models are similar to
the ones used in the literature, results comparison was made to validate them.

Here, the simulation design is the most important aspect. We must have a good design of
experiments to obtain statistically good results. To do so, we will run multiple randomized
runs. The goal is to reduce the effect of packet and user distributions. For reliability, we use
the average failure rate over 100 runs. For the algorithm comparison, we use 20 runs of each
scheme, since the simulation duration evolves exponentially with the number of users.

The second important aspect of our design of experiments is its reproducibility. Our simulator
allows reproducible results via the usage of its pseudo-random number generator and a seed. As
such, each algorithm is run once with the same seed. The exact same traffic and the same users
are therefore applied to each scheme. We thus ensure that randomization does not impact our
results.

When simulating dynamical systems, a special interest must be given on the stability of the sys-
tem. Here, we define the warm-up time of the system as the first third of the results. They are
discarded from our evaluation, since connection establishments and cells sleep/wake-up proce-
dures have a real and important impact on the system. Our simulations are running 15 seconds
(in simulated time) for reliability and 20 seconds for algorithm comparison. Our sampling fre-
quency is 100 Hz (data points are collected every 10 ms). The temporal average is taken from
the stable period.

For reliability, we define the single connectivity scenario as our baseline. For algorithm com-
62

parison, we also consider single connectivity as our baseline. We might also consider our first
algorithm as a second baseline when comparing scheme performance.

5.2 Reliability Improvement with Multi-Connectivity

Our first experiment tries to answer how multi-connectivity can improve reliability. To achieve
this, we investigate the Radio Link Failure rate for different user speed. The RLF rate is defined
as the percentage of users in RLF at any instant t.

For this study, we consider a deployment where only small cells are present. Furthermore, these
cells are all 5G NR base stations. We also consider an inter-site distance of 200 meters. The
main reason is to insure validity of the results, as a macro-coverage will decrease the risks of
failures. We compare two experiments: one where users can only connect to one cell at a time,
refereed here as single connectivity, and a second where users can connect to two base stations
at the same time, refereed as multi-connectivity, and following the Max SINR scheme. 50 users
are present and active throughout the simulations.

Table 7. Scenario Parameters

Deployment Values
RAT 5G NR
Number of sites 61
Inter Site Distance [m] 200
Experiments Schemes
Single Connectivity No scheme
Multi-Connectivity Max SINR

One way to model the RLF rate can be a function depending on the base stations density, the
user speed, the value for the Time To Trigger parameter, as well as the noise, such as PRLF (u) =
f (λBS , vu , T T T, N ). Vasudeva et al. showed that the RLF increases with the UE velocity [72].

Our hypotheses are as follows:

1. the Radio Link Failure rate decreases with Multi-Connectivity,

2. the RLF rate decreases with the number of secondary connections, but the interest of MC
decreases as the number of connections increases,
63

such as the gain from single connectivity to dual connectivity is higher than the gain from dual
connectivity to three connections. The first hypothesis can be seen as obvious, since more con-
nections can be seen as a security against connection drops in a first, naive approach. However,
we should be reminded that the MC scheme takes the main SINR as input, such as the probabil-
ity of MC is not independent from the main signal. Figure 14 shows the relation between RLF
rate and user velocity.

8
Percentage of Radio Link Failure (%)

6 Multi Connectivity (n=3)


Dual Connectivity (n=2)
5 Single Connectivity

0
0 20 40 60 80 100 120 140
User Velocity (km/h)

Figure 14. Radio Link Failure Rate depending on the user velocity

As expected, multi-connectivity does improve reliability, especially at high speed. We can note
a decrease of 37.4% with dual-connectivity compare to single connectivity, and 52.8% for MC.
The differences are however much less significant at medium or low speed. At 50 km/h, DC
decreases the failure rate by 26.8%, compared to a decrease of 32.3% with MC. Below 20 km/h,
there is no significant improvement.

One first conclusion that can be drawn from this figure is that the radio link failure rate evolves
exponentially with the velocity, such as PRLF = f (v, n)eh(v,n) where v is the velocity, n the
number of connections and h an unknown function. We decide to apply linear regression to
log(PRLF ) to obtain an estimate of f . Specifically, we use the least squares method.

We obtain fˆ(v, n) = e(0.0245−0.0028∗n)∗v ∗ eh(n) where h(n) = 0.0992 ∗ n2 − 0.1414 ∗ n − 0.3689.


64

The standard deviation of the simulation results (averaged) and the estimated function are 0.03
for MC, 0.046 for DC and 0.173 for SC. The latter can be decreased to 0.077 by applying non
linear regression, at the cost of a more complex fˆ. The standard deviation against the entire set
of results (2700 values) is similar to the deviation from the mean and given below. Figure 15
shows the linear regression expression against the averaged RLF for each speed.

8 Linear Regression − Single Connectivity


Percentage of Radio Link Failure (%)

Linear Regression − Dual Connectivity (n=2)


7 Linear Regression − Multi Connectivity (n=3)
Mean − Single−Connectivity
6 Mean − Dual Connectivity (n=2)
Mean − Multi−Connectivity (n=3)
5

0
0 20 40 60 80 100 120 140
User Velocity (km/h)

Figure 15. Linear Regression of the Simulated RLF Rate

As said in the methodology section, the RLF rate was obtained by averaging the results of 100
runs. As the process involved random variables, we must analyse the standard deviation over the
results. Figure 16 shows the standard deviation evolution depending on the user speed. Again,
it can be seen that the deviation is evolving with the velocity and the number of connections. To
better understands these results, we look in figure 17 at two cases, at 15 and 125 km/h.

We can see that the higher the velocity, the higher the standard deviation. However, increasing
the number of connections decreases both the variance (from a few users at 20% failures in SC
against 12% maximum in MC) and the mean (around 3% for MC against 6% for SC).

Following our hypotheses, we can conclude that:

1. The Radio Link Failure rate increases exponentially with the user velocity,
65

5
Standard Deviation of the RLF Rate (% of RLF)
4.5
Multi−Connectivity (n=3)
4 Dual Connectivity (n=2)
Single Connectivity
3.5

2.5

1.5

0.5

0
0 20 40 60 80 100 120 140
User Velocity (km/h)

Figure 16. Standard Deviation of the RLF Rate

Single−Connectivity, 15 km/h Dual−Connectivity, 15 km/h Multi−Connectivity (n=3), 15 km/h


60 60 60

50 50 50
Occurences (%)

Occurences (%)

Occurences (%)

40 40 40

30 30 30

20 20 20

10 10 10

0 0 0
0 2 4 6 8 10 0 2 4 6 8 0 2 4 6 8
RLF Rate (%) RLF Rate (%) RLF Rate (%)

Single−Connectivity, 125 km/h Dual−Connectivity, 125 km/h Multi−Connectivity (n=3), 125 km/h
20 20 15

15 15
Occurences (%)

Occurences (%)

Occurences (%)

10

10 10

5
5 5

0 0 0
0 10 20 30 0 5 10 15 20 0 5 10 15
RLF Rate (%) RLF Rate (%) RLF Rate (%)

Figure 17. Simulation results distribution depending on the velocity and number of connections

2. Increasing the number of connections decreases the RLF rate, and

3. The decrease between multi-connectivity and dual connectivity is lower than the decrease
between dual connectivity and single connectivity.
66

Table 8. Deployment Parameters

Parameters LTE 5G
Number of sites 3 61
Inter Site Distance [m] 400 100

5.3 Multi-Connectivity Algorithms’ Comparison

In this scenario, we want to evaluate our multi-connectivity schemes against each other and
against single connectivity. For this comparison, we will look at several metrics: i. the power
consumption, ii. the percentage of sleeping cells, iii. the energy efficiency, iv. the probability of
multi-connectivity and v. the 10 percentile user throughput.

For this scenario, we use a two-tier deployment. LTE is used as the macro-layer, with three sites
(therefore nine cells, since LTE uses tri-sector antennas here) and 61 5G NR sites. Since we
placed ourselves in a dense urban scenario, we choose an inter-site distance of 400 meters for
LTE and 100 meters for NR.

We will test single connectivity and four MC schemes: Max Bitrate, Max Bitrate-EE, Max
Clustered-Bitrate and AHP. We fix the maximum number of secondary connections to one (sim-
ilar to dual connectivity).

Our hypotheses are as follow:

1. Multi-connectivity consumes more than single connectivity,

2. Max Clustered-Bitrate consumes less than Max Bitrate-EE,

3. Multi-connectivity is more energy-efficient than single connectivity,

4. Max Bitrate-EE is more energy-efficient than Max Bitrate, and

5. The user throughput is increased with Multi-connectivity.


67

4
x 10

2
Power Consumption (W)

1.5

Single Connectivity
1
Max Bitrate
Max Bitrate−EE
Max Clustered−Bitrate
0.5 AHP

0
0 10 20 30 40 50 60 70 80 90 100
Number of Users

Figure 18. System Power Consumption

5.3.1 Power Consumption and Sleeping Cells

Figure 18 presents the power consumption depending on the number of users present. We can
see that the consumption increases with the number of UEs non-linearly. The limit when the
number of users surpass the number of BSs and approaches infinity is the consumption of all the
cells active and with full load, such as limnbu →+∞ Psystem (nbu ) = nbcells ∗ Pload=100% ≈ 41 kW.

One first observation is that most of the MC schemes lead to an increase of consumption. Single
Connectivity represents 1.419 kW for 35 users. Max Bitrate consumes 1.475 kW, which repre-
sents an increase of 4%. Max Bitrate-EE consumes 1.454 kW, or an increase of 2.5% compared
to SC. Finally, AHP represents 1.449 kW, which is equivalent to a 2.11% increase. These results
are too similar to differentiate them and rank the algorithms. This is due to the fact that all these
algorithms consider small cells as targets, with low consumption.

A second observation can be made from the Max Clustered-Bitrate. Here, the consumption is
actually lower compared to single connectivity. This is due to the fact that the algorithm try to
minimize the number of additional cells to wake up, and group the additional transmissions on
low consumption BSs. Since the power condition checks whether it is more efficient to use a
secondary connection or to keep the primary one, we can achieve this reduction of 2.9% with a
consumption equals to 1.378 kW.
68

Percentage of Sleeping Cells Comparison for different MC Schemes

Percentage of Sleeping Cells (%) 90 Single Connectivity


Max Bitrate
80
Max Bitrate−EE
70 Max Clustered−Bitrate
AHP
60

50

40

30

20

10

0
0 10 20 30 40 50 60 70 80 90 100
Number of Users

Figure 19. Percentage of Sleeping Cells

The growth in power consumption is due to the fact that multiple cells are used concurrently to
improve the user throughput. This also means the transmission is finished earlier, and the cells
could go to sleep faster if the load is not too high.

Similar findings can be obtained when looking at the number of sleeping cells, as shown in figure
19. Here, we can see that both single connectivity and Max Clustered-Bitrate have more users
in sleep mode, compared to the others. Max Bitrate, Max Bitrate-EE and AHP have between
6 to 8% less cells in sleep mode. (Max Bitrate: 7.97%, Max Bitrate-EE: 7.01%, AHP: 6.04%
compared to SC).

From this, we can already conclude that:

1. Multi-connectivity often consumes more than single connectivity (Hypothesis 1), but

2. Some schemes might consume less than single connectivity (Hypothesis 2), depending on
the sleep strategy, and

3. The increase of power consumption is rather limited.


69

7
x 10
10

8
Energy Efficiency (b/J)

6 Single Connectivity
Max Bitrate
5
Max Bitrate−EE
4 Max Clustered−Bitrate
AHP
3

0
0 10 20 30 40 50 60 70 80 90 100
Number of Users

Figure 20. System Energy Efficiency

5.3.2 Energy Efficiency

The energy efficiency is defined here by the system throughput divided by the entire power
consumption of the system, and is expressed in bits per Joule ([b/J]). It also corresponds to the
quantity of data transferred per unit of energy. Figure 20 shows the energy efficiency of our
deployment depending on the number of UEs.

First, we can see that the system is not energy efficient for a very low number of users. This is
due to the fact that we have a huge overhead of power consumed by the sleeping cells for a low
traffic in the area. With more users, the traffic grows almost linearly with the number of users,
leading to a more efficient deployment. We can also note that the EE stabilizes above 40 users
until 100.

In a similar manner, we can see that Max-Bitrate, Max Bitrate-EE and AHP provides an ap-
proximately similar efficiency, respectively 62.41 Mb/J, 62.97 Mb/J and 63.2 Mb/J compared
to the single connectivity scenario (51.67 Mb/J), for 10 users. This represents an increase of
20.8% for Max-Bitrate compared to SC, 21.9% for Max Bitrate-EE and 22.3% for AHP. For
Max Bitrate-EE and AHP, the usage of the power condition can explain the result. For Max Bi-
trate, we assume its performance is mainly due to the maximization of the bitrate, which evolve
faster than the power consumption, leading to this good performance.
70

As the number of users (and, by extension the load) increases, we see a slight decrease in gains
from MC. At 40 users, Max Bitrate offers a increase of 14.1% compared to single connectivity,
Max Bitrate-EE 15.8% and 15.9% for AHP.

When the number of users becomes too high, and we leave the ultra-dense network paradigm,
the performances tend to be less significant. With 75 users in our network, the three algorithms
obtain an increase between 7 and 9%.

Concerning the Max Clustered-Bitrate scheme, we can observe that it behaves badly with very
few users, giving the same performance as single connectivity. For 45 users, Max Clustered-
Bitrate represents an increase in EE of only 4.2% compared to SC (while the others are above
14%). Its performance slightly increases with the number of users, 5.2% with 75 UEs, and tends
to the same gain as the others at 100 users. However, this is mostly due to the decrease of gain
of the other algorithms, more than the increase in performance form Max Clustered-Bitrate.

Regarding the standard deviation, we obtain results varying around 10-15% around the average
value (the deviation being slightly higher with few users). Figure 21 shows the EE distribution
for 20 runs with 60 users. We can clearly see that results with MC have an higher mean and
slightly lower deviation, but they sometimes overlap. This means that for specific patterns and
specific times, single connectivity obtains better results. In reality, this might be explained by
the connection establishment and channel quality estimation, during which secondary cells are
active, but not used for transmissions.

We can therefore expand our findings with:

4. Multi-connectivity is more energy-efficient than single connectivity (Hypothesis 3), but

5. The gain in energy-efficiency decreases as the number of users increases,

6. Max Clustered-Bitrate results are rather limited (Hypothesis 4), and

7. AHP offers the best results, closely followed by Max Bitrate-EE and Max Bitrate (Hy-
pothesis 4).
71

Single Connectivity Max Bitrate Max Bitrate−EE


7 7 7
6 6 6
Occurrences 5 5 5

Occurrences

Occurrences
4 4 4
3 3 3
2 2 2
1 1 1
0 0 0
0 5 10 15 0 5 10 15 0 5 10 15
Energy Efficiency (b/J) 7 Energy Efficiency (b/J) 7 Energy Efficiency (b/J) 7
x 10 x 10 x 10
Max Clustered−Bitrate AHP
7 7
6 6
5 5
Occurrences

Occurrences
4 4
3 3
2 2
1 1
0 0
0 5 10 15 0 5 10 15
Energy Efficiency (b/J) 7 Energy Efficiency (b/J) 7
x 10 x 10

Figure 21. Results Distribution

5.3.3 Theoretical Lower Bound

We want to compare our values with the theoretical lower bound expressed in chapter 4. Based
on the results for single connectivity and the number of users, we compute the lower bound.
Figure 22 shows the simulated results in dashed lines and the theoretical bound.

As we used a disconnection threshold of 0.5, we obtain a bound above the single connectivity
efficiency, we ensure us that multi-connectivity will always be more energy efficient. We can
observe that our simulations obtained better results than the bound, which was expected. How-
ever, as the number of users increases, the simulated results seems to tend to the theoretical
bound. By average, Max Bitrate and Max Bitrate-EE are above the theoretical bound by 11.5%
and 12.5% respectively. The theoretical bound also ensure an increase of energy efficiency of
6% by average over single connectivity.

The AHP scheme, however, can not be compared with the theoretical bound. In AHP, we cannot
certify that a target with a bitrate below the threshold will not be selected. It will also depend
on the score of the other metrics. Nonetheless, the previous results showed that AHP does still
increase energy efficiency.

It has to be noted that the formula used for the lower bound does not scale to an infinite number
of users. Indeed, the assumptions do not hold true when there is more than one user on a
72

7
x 10
12

10
Energy Efficiency (B/J)

6
Low Performance Bound
Simulated Single Connectivity
4 Simulated Max Bitrate
Simulated Max Bitrate−EE
2

0
0 10 20 30 40 50 60 70 80 90 100
Number of Users

Figure 22. Simulated Results and Theoretical Lower Bound for Energy Efficiency

secondary cell, which can also be seen as a higher density of users than access nodes.

From this low performance bound, we could influence the gains in terms of energy efficiency by
choosing the right threshold. With a higher disconnection threshold, the low performance bound
would increase. However, the probability of multi-connectivity is also important and used in the
formula. As the threshold increases, the probability decreases. Our simulations did however not
prove if there is a linear relation between the two parameters.

5.3.4 Probability of Multi-Connectivity

As the selection process changes between each scheme, the different requirements might change
how often a user can be connected to multiple base stations. We also decide to observe the
probability of MC as a metric to compare schemes. For this, we define the Max Bitrate scheme
as the baseline, as its selection procedure is the simplest. Table 9 lists our results.

We can first note that, as expected, the Max Bitrate scheme has the highest probability, with
19.01% of users in MC. AHP is close second with 18.07%, whereas Max Bitrate-EE has
17.23%. Both Max Bitrate-EE and AHP use the same energy efficiency condition, but Max
Bitrate-EE applies selection sequentially, removing the non efficient targets from the ones sat-
isfying the bitrate condition, whereas AHP computes the score separately and sum to get the
73

Table 9. Probability of Multi-Connectivity

Scheme Probability [%] Standard Deviation


Max Bitrate 21.53 2.46
Max Bitrate-EE 18.71 2.27
Max Clustered-Bitrate 13.16 4.93
AHP 19.11 2.22

final set. AHP therefore contains targets that are not present with the other scheme, which ex-
plains the slightly better results. In a similar fashion, Max Bitrate doesn’t apply the efficiency
condition, which is why he has the highest number of UEs in MC. The deviation is mostly due
to simulations with few users, which are either more sensible to randomness or obtain slightly
higher probability.

Once again, we can see that Max Clustered-Bitrate offers worse performance compared to the
other schemes, with a probability of being in MC of only 7.37%. However, it has to be noted
that, not unlike the energy-efficiency, the probability is dependent on the number of users and
increases as the number of UEs increases. Therefore, we obtain around 10% with a low number
of users against 17% with a high number. These findings confirm that Max Clustered-Bitrate
performs really badly compared to the other schemes with few users.

While looking at the performance of Multi-Connectivity, we can also observe the distribution of
connections. The first finding is that NR is almost always used for the secondary connections.
This is due to the choice of bitrate as a metric, and the higher bandwidth of 5G compared to LTE.
The second observation is that most of the multi-connectivity happens when the master cell is
a LTE base station (between 80-90% of cases). Once again, since LTE has a lower bandwidth,
it is easier to find a target with better performance. However, NR as both master and secondary
connections also happens, and we believe other deployment scenarios might see an increase of
that cases. Since two 5G NR cells consume the same power in our scenario, it is unlikely that
using the second one is more energy efficient.

Finally, we want to explain the impact of the θadd and θrmv when designing the algorithms. We
saw after simulations that the higher the hysteresis, the higher the probability for Max Bitrate.
by stetting an hysteresis at 0.2, we saw the probability of MC decreasing by 2-3%. We also saw
a decrease in energy efficiency with low disconnection thresholds.

Other algorithms are not as much impacted (when reducing the disconnection threshold), since
most users will not find a more energy efficient BS to connect to. This is particularly teh case
when UEs are connected to NR as a master cell.
74

Max Clustered-Bitrate did also behave very badly with high disconnection threshold, falling to
a probability around 5-9%, while the other suffered a decrease of 1-2% only.

5.3.5 User Throughput

A percentile is defined as the value below which a given percentage of observations fail. Here,
it means that we look at the user throughput below which 10% of the users are. Figure 23 shows
the 10%-ile user throughput as a function of the number of users.

50

45
10th%−ile User Throughput (Mb/s)

Single Connectivity
40 Max Bitrate
Max Bitrate−EE
35 Max Clustered−Bitrate
AHP
30

25

20

15

10

0
0 10 20 30 40 50 60 70 80 90 100
Number of Users

Figure 23. 10th Percentile User Throughput

As a general observation, we can start by stating that the 10%-ile decreases as the number of
users increases. This is simply explained by an increase in traffic demand for an equivalent
number of resources. More interesting to our study, we see that every algorithm performs better
than single-connectivity. Once again, it is simply because we offer more resources (from two
or more cells) to the user. With a single connection, the 10%-ile user throughput is at 5.966
Mb/s with ten users. Max Clustered-Bitrate obtains the worst result of all the MC schemes, with
6.646 Mb/s (an increase of 11.4% only). Max Bitrate offers an increase of 77.2% compared
to SC, with 10.57 Mb/s, while AHP offers 10.88 Mb/s (82.4% increase) and Max Bitrate-EE
has the highest result here with 11.02 Mb/s, which corresponds to an increase of 84.7%. The
last three algorithms obtained pretty similar results anyway and we cannot differentiate which
one is the best. With more users, here 75, Max Clustered-Bitrate obtained better results with
75

an increase of 43% compared to single connectivity, while the other algorithms obtained again
similar results (between 73 to 80% increase).

Unsurprisingly, we can conclude that the user throughput is increased with multi-connectivity
(Hypothesis 5).

5.4 Sustainability

As we stated in the introductory chapter, one of the big problems of cellular networks are their
energy consumption. Depending on the sources, their share in the carbon emissions of ICT is
between 16% and 24% [84, 85]. Moreover, with 5G on the way, and scenarios like ultra-dense
networks, its consumption will most likely increase even further, as the never-stopping growth of
data traffic requires these new deployments. To promote sustainability, one can try to challenge
this power consumption growth and propose more energy-efficient solutions.

These results can also be interpreted for sustainability. Although we saw an increase in con-
sumed power, multi-connectivity can be used with specific parameters to actually decrease the
energy used by the network. Furthermore, other algorithms lead to a small increase, less than
4%, which is lower that what we feared. More energy-efficient hardware could easily tackle this
small growth.

Moreover, our results for energy efficiency are very interesting. We showed that multi connec-
tivity allows a network operator to provide a better service, here the offered capacity, for the
same amount of energy. A more energy efficient solution is however not always more sustain-
able. The Jevons paradox states that a more efficient technology might cause a growth in the
consumption rate, causing such technology to not be sustainable. In our case, it is a known fact
that the consumption rate is increasing, regardless of how efficient our solution is. On the other
hand, the rate is also affected by other factors. For wireless systems, the spectral efficiency is
one such bounding parameter and, nowadays, we are tending to the maximal theoretical effi-
ciency. This means, amongst other things, that the gains from a more efficient system will not
be overturn by an increase of deployed base stations far above what is expected in this study.

Therefore, we can consider our solution as green, or greener than actual system, as envisioned
within a Green IT aspect. Sustainability, on the other hand, is a much broader field. It is often
decoupled into three aspects, called the Pillars of Sustainability: Economical, Ecological, and
Social (also known as people, planet and profit). Our work can easily be linked with economical
aspect, as energy consumption, but also network deployment, as a cost. We can also link energy
76

usage towards ecological aspects, especially considering energy production, although its analy-
sis depends on the region considered. The third pillar is somewhat harder to connect with this
study, as users’ behaviour does affect the traffic, but the system might not affect back the users
if the quality of experience is satisfied. Furthermore, long term effects, also known as rebound
effects, are even harder to grasp regarding this topic. We did not find any possible negative long
term effects, other than the one covered before.
77

6 Summary and Discussions

Our work tried two answer two big questions: how multi-connectivity affect the system perfor-
mance and how can we design MC schemes to improve the network services.

6.1 Multi-Connectivity Benefits

Multi-connectivity has been investigated regarding its impact on reliability, on power consump-
tion, and its energy-efficiency.

We showed first that indeed, multi-connectivity can reduce the failure rate. Having multiple
connections allow continuous transmission even if one link is in failure. Depending on the
procedures, it could also reduce the time for handover. However, we saw that the benefits of
adding more connections decreases as the number of simultaneous links increases. This means
than there are little benefits to have ten connections instead of only three or four. In fact, we
believe three or four should be the maximum number of connections for a best trade-off between
reliability and signalling complexity. However, we can also note that users at low speeds will
have little benefits to connect to multiple cells as their failure rate is quite low already, and will
not obtain the optimal performance due to the usage of SINR as a metric. They would, however,
still benefit from a substantial increase in user throughput, event with that metric.

Then, we showed the impact of multi-connectivity over the power consumption. The expectation
was an increase of power consumption with multi-connectivity. This is indeed true, but the
increase is around 4%, whereas we were expecting more. We believe it is mostly due to the MC
distribution: almost all secondary connections were small cells, with smaller consumptions.
Concerning the results of Max Clustered-Bitrate, our observations are the same. The small
probability of MC, coupled with the energy efficiency condition allowed these results. We can
thus deduce that with a good sleep strategy, multi-connectivity can reduce the network’s power
consumption, and will only increase it by a few margins without a good sleep strategy.

Finally, we showed that multi-connectivity was more energy-efficient than single connectivity
in our scenario. Since the system throughput increases faster than the system’s power consump-
tion, multi-connectivity is a good feature for 5G. With a large amount of users, however, the
gains become less significant. This is largely due to the scarcity of resources as more users use
resources.
78

6.2 Multi-Connectivity Schemes

In this work, we designed a general mechanism on how multi-connectivity can be executed on


base stations. We then proposed five different algorithms based on this procedure; one specifi-
cally aimed at improving robustness, the four others for energy-efficiency and performance. We
tested our solutions against each other, and evaluated them using numerous metrics. We found
that the energy condition theorized did improve the energy efficiency. To be more precise, the
condition allowed a slight mitigation of the increase in power consumption and did not im-
pact the improvement on the system throughput. The condition also affected the probability of
maintaining multiple connections by a few percent, but it is still an excellent metric for energy
efficiency.

We also found that the Analytic Hierarchy Process (AHP) slightly outperformed the other
schemes in terms on energy efficiency, while offering a good probability of MC. Its princi-
ple also allows its utilization in a variety of different scenarios, by simply changing the weights,
whereas the rest of the algorithms cannot be easily altered to different use. We consider this
scheme as the most promising for future improvements.

Max Clustered-Bitrate, however, performed poorly with various metrics. This algorithm was
designed to specifically tackle the power consumption, as grouping more users together could
reduce the number of active cells. However, although its goal was fulfilled concerning the power
consumption, the system throughput was quite low compared to the other solutions, which led
to non satisfactory results for energy efficiency. We decided to keep it in our final results for
comparison purposes. This algorithm is however the result of multiple iterations, and had the
best results amongst the proposed solutions using a clustering technique. Nonetheless, this
algorithm might yield better results in other scenarios, for example dense networks with a large
amount of users.
79

7 Conclusion and Future Work

In this thesis, we investigated multi-connectivity’s effects on the network performance within


an ultra-dense deployment. Multi-connectivity has been demonstrated as an important feature
for meeting the 5G requirements within ultra-dense networks.

To study MC performance, we presented a framework for secondary cell association. A condi-


tion on energy efficiency for cell association was also theorized. Simulations were carried out
to evaluate multi-connectivity schemes.

Four algorithms were proposed and evaluated in terms of power consumption, system’s energy
efficiency and user throughput. Our solutions showed an increase in energy efficiency of up to
22% with ten users and up to 15% with forty compared to single connectivity. It was also shown
that multi-connectivity decreased the radio link failure rate of users at high speeds, by up to
37% with two connections and up to 52% with three connections. A discussion on how green
this solution can be considered was done, and what we can deduce for sustainability.

This work can be used as a foundation for more research on multi-connectivity, as few pa-
pers investigated its interest. Our results could be extended by looking at multi-connectivity in
stochastic networks or real case scenarios, as we used a hexagonal grid deployment here. Cov-
erage islands and holes in the micro-layer could impact locally the effect of multi-connectivity.

A second possible research direction would consist of designing adaptive algorithms. Indeed,
taking our AHP scheme as an example, one could design a solution in such a way that the weight
of each metric would evolve depending on the user history, such as its precedents handovers,
failures or traffic evolution.
80

REFERENCES
[1] A. Klimova, E. Rondeau, K. Andersson, J. Porras, A. Rybin, and A. Zaslavsky, “An in-
ternational master’s program in green ICT as a contribution to sustainable development,”
Journal of Cleaner Production, vol. 135, pp. 223–239, 2016.

[2] Ericsson AB, Ericsson Mobility Report, nov 2016.

[3] ITU-R M.2083-0, IMT Vision - Framework and overall objectives of the future develop-
ment of IMT for 2020 and beyond, sep 2015.

[4] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong, and J. C.


Zhang, “What will 5G be?,” IEEE Journal on Selected Areas in Communications, vol. 32,
pp. 1065–1082, jun 2014.

[5] H. Bogucka and A. Conti, “Degrees of freedom for energy savings in practical adaptive
wireless systems,” IEEE Communications Magazine, vol. 49, pp. 38–45, jun 2011.

[6] I. Humar, X. Ge, L. Xiang, M. Jo, M. Chen, and J. Zhang, “Rethinking energy efficiency
models of cellular networks with embodied energy,” IEEE Network, vol. 25, pp. 40–49,
mar 2011.

[7] M. Maternia, S. Eddine El Ayoubi, M. Fallgren, P. Spapis, Y. Qi, D. Martín-Sacristán,


Óscar Carrasco, M. Fresia, M. Payaró, M. Schubert, J. S. Bedo, and V. Kulkarni, “5G PPP
use cases and performance evaluation models,” tech. rep., 5G-PPP, 2016.

[8] A. Osseiran, F. Boccardi, V. Braun, K. Kusume, P. Marsch, M. Maternia, O. Queseth,


M. Schellmann, H. Schotten, H. Taoka, H. Tullberg, M. A. Uusitalo, B. Timus, and M. Fall-
gren, “Scenarios for 5G mobile and wireless communications: the vision of the METIS
project,” in IEEE Communications Magazine, IEEE, may 2014.

[9] METIS, Deliverable D1.5: Updated scenarios, requirements and 5G for 5G mobile and
wireless system with recommendations for future investigations, 2015.

[10] ITU-R, Working Party 5D, [Draft] Minimum requirements related to technical perfor-
mance for IMT-2020 radio interface(s), feb 2017.

[11] 3GPP TS 23.799, Technical Specification Group Services and System Aspects; Study on
Architecture of Next Generation System (Release 14), 2016.

[12] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broadband systems,” IEEE
Communications Magazine, vol. 49, pp. 101–107, jun 2011.
81

[13] P. Okvist, H. Asplund, A. Simonsson, B. Halvarsson, J. Medbo, and N. Seifi, “15 GHz
propagation properties assessed with 5G radio access prototype,” in 2015 IEEE 26th An-
nual International Symposium on Personal, Indoor, and Mobile Radio Communications
(PIMRC), IEEE, aug 2015.

[14] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz,


M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5G cellular:
It will work!,” IEEE Access, vol. 1, pp. 335–349, 2013.

[15] W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and F. Aryan-
far, “Millimeter-wave beamforming as an enabling technology for 5G cellular communi-
cations: theoretical feasibility and prototype results,” IEEE Communications Magazine,
vol. 52, pp. 106–113, feb 2014.

[16] R. Daniels and R. H. Jr., “60 GHz wireless communications: emerging requirements and
design recommendations,” IEEE Vehicular Technology Magazine, vol. 2, pp. 41–50, sep
2007.

[17] T. Yilmaz, E. Fadel, and O. B. Akan, “Employing 60 GHz ISM band for 5G wireless
communications,” in 2014 IEEE International Black Sea Conference on Communications
and Networking (BlackSeaCom), IEEE, may 2014.

[18] R. Zhang, M. Wang, L. X. Cai, Z. Zheng, X. Shen, and L.-L. Xie, “LTE-unlicensed: the
future of spectrum aggregation for cellular networks,” IEEE Wireless Communications,
vol. 22, pp. 150–159, jun 2015.

[19] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, “An overview of massive
MIMO: Benefits and challenges,” IEEE Journal of Selected Topics in Signal Processing,
vol. 8, pp. 742–758, oct 2014.

[20] F. W. Vook, A. Ghosh, and T. A. Thomas, “MIMO and beamforming solutions for 5G
technology,” in 2014 IEEE MTT-S International Microwave Symposium (IMS2014), IEEE,
jun 2014.

[21] S. Han, C. lin I, Z. Xu, and C. Rowell, “Large-scale antenna systems with hybrid ana-
log and digital beamforming for millimeter wave 5G,” IEEE Communications Magazine,
vol. 53, pp. 186–194, jan 2015.

[22] M. Sawahashi, Y. Kishiyama, A. Morimoto, D. Nishikawa, and M. Tanno, “Coordi-


nated multipoint transmission/reception techniques for LTE-advanced [coordinated and
distributed MIMO,” IEEE Wireless Communications, vol. 17, pp. 26–34, jun 2010.
82

[23] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO for next generation
wireless systems,” IEEE Communications Magazine, vol. 52, pp. 186–195, feb 2014.

[24] C. Liang, F. R. Yu, and X. Zhang, “Information-centric network function virtualization


over 5g mobile wireless networks,” IEEE Network, vol. 29, pp. 68–74, may 2015.

[25] C. Bouras, A. Kollia, and A. Papazois, “SDN & NFV in 5G: Advancements and chal-
lenges,” in 2017 20th Conference on Innovations in Clouds, Internet and Networks (ICIN),
IEEE, 2017.

[26] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu, “Device-to-device communication in 5G


cellular networks: challenges, solutions, and future directions,” IEEE Communications
Magazine, vol. 52, pp. 86–92, may 2014.

[27] J. Qiao, X. Shen, J. Mark, Q. Shen, Y. He, and L. Lei, “Enabling device-to-device com-
munications in millimeter-wave 5G cellular networks,” IEEE Communications Magazine,
vol. 53, pp. 209–215, jan 2015.

[28] L. Militano, G. Araniti, M. Condoluci, I. Farris, and A. Iera, “Device-to-device communi-


cations for 5G internet of things,” EAI Endorsed Transactions on Internet of Things, vol. 1,
p. 150598, oct 2015.

[29] Y. Wu, J. Chen, L. P. Qian, J. Huang, and X. S. Shen, “Energy-aware cooperative traffic
offloading via device-to-device cooperations: An analytical approach,” IEEE Transactions
on Mobile Computing, vol. 16, pp. 97–114, jan 2017.

[30] M. Ding, D. Lopez-Perez, G. Mao, P. Wang, and Z. Lin, “Will the area spectral efficiency
monotonically grow as small cells go dense?,” in 2015 IEEE Global Communications
Conference (GLOBECOM), IEEE, dec 2015.

[31] M. Kamel, W. Hamouda, and A. Youssef, “Ultra-dense networks: A survey,” IEEE Com-
munications Surveys & Tutorials, vol. 18, no. 4, pp. 2522–2545, 2016.

[32] 3GPP TS 36.839, Technical Specification Group Radio Access Network; Evolved Univer-
sal Terrestrial Radio Access (E-UTRA); Mobility enhancements in heterogeneous networks
(Release 11), 2012.

[33] 3GPP TR 36.819, Technical Specification Group Radio Access Network; Coordinated
multi-point operation for LTE physical layer aspects (Release 11), 2011.

[34] M. Thurfjell, M. Ericsson, and P. de Bruin, “Network densification impact on system ca-
pacity,” in 2015 IEEE 81st Vehicular Technology Conference (VTC Spring), IEEE, may
2015.
83

[35] A. Gotsis, S. Stefanatos, and A. Alexiou, “UltraDense networks: The new wireless frontier
for enabling 5G access,” IEEE Vehicular Technology Magazine, vol. 11, pp. 71–78, jun
2016.

[36] Q. Li, G. Wu, and R. Q. Hu, “Analytical study on network spectrum efficiency of ultra
dense networks,” in 2013 IEEE 24th Annual International Symposium on Personal, Indoor,
and Mobile Radio Communications (PIMRC), IEEE, sep 2013.

[37] Y. Sun, Y. Chang, M. Hu, and B. Wang, “A cluster-based hybrid access strategy using
non-cooperative game theory for ultra-dense HetNet,” in 2015 IEEE 17th International
Conference on High Performance Computing and Communications, 2015 IEEE 7th Inter-
national Symposium on Cyberspace Safety and Security, and 2015 IEEE 12th International
Conference on Embedded Software and Systems, IEEE, aug 2015.

[38] A. Y. Al-Zahrani, F. R. Yu, and M. Huang, “A joint cross-layer and colayer interference
management scheme in hyperdense heterogeneous networks using mean-field game the-
ory,” IEEE Transactions on Vehicular Technology, vol. 65, pp. 1522–1535, mar 2016.

[39] M. Ericson, “Total network base station energy cost vs. deployment,” in 2011 IEEE 73rd
Vehicular Technology Conference (VTC Spring), IEEE, may 2011.

[40] S. Yunas, M. Valkama, and J. Niemelä, “Spectral and energy efficiency of ultra-dense net-
works under different deployment strategies,” IEEE Communications Magazine, vol. 53,
pp. 90–100, jan 2015.

[41] 3GPP TS 36.842, Technical Specification Group Radio Access Network; Study on Small
Cell enhancements for E-UTRA and E-UTRAN; Higher layer aspects (Release 12), 2013.

[42] 3GPP TR 23.799, [Draft] Technical Specification Group Radio Access Network; Evolved
Universal Terrestrial Radio Access (E-UTRA) and NR; Multi-Connectivity; Stage 2 (Re-
lease 15), 2017.

[43] Wireless World Research Forum, LTE Small Cell Enhancement by Dual Connectivity, nov
2014.

[44] R. Irmer, H. Droste, P. Marsch, M. Grieger, G. Fettweis, S. Brueck, H.-P. Mayer, L. Thiele,
and V. Jungnickel, “Coordinated multipoint: Concepts, performance, and field trial re-
sults,” IEEE Communications Magazine, vol. 49, pp. 102–111, feb 2011.

[45] M. Giordani, M. Mezzavilla, S. Rangan, and M. Zorzi, “Multi-connectivity in 5G mmWave


cellular networks,” in 2016 Mediterranean Ad Hoc Networking Workshop (Med-Hoc-Net),
IEEE, jun 2016.
84

[46] F. B. Tesema, A. Awada, I. Viering, M. Simsek, and G. P. Fettweis, “Mobility modeling and
performance evaluation of multi-connectivity in 5G intra-frequency networks,” in 2015
IEEE Globecom Workshops (GC Wkshps), IEEE, dec 2015.

[47] I. D. Silva, G. Mildh, J. Rune, P. Wallentin, J. Vikberg, P. Schliwa-Bertling, and R. Fan,


“Tight integration of new 5G air interface and LTE to fulfill 5G requirements,” in 2015
IEEE 81st Vehicular Technology Conference (VTC Spring), IEEE, may 2015.

[48] M. A. Lema, E. Pardo, O. Galinina, S. Andreev, and M. Dohler, “Flexible dual-


connectivity spectrum aggregation for decoupled uplink and downlink access in 5G het-
erogeneous systems,” in IEEE Journal on Selected Areas in Communications, IEEE, nov
2016.

[49] M. Gruber, O. Blume, D. Ferling, D. Zeller, M. A. Imran, and E. C. Strinati, “EARTH - en-
ergy aware radio and network technologies,” in 2009 IEEE 20th International Symposium
on Personal, Indoor and Mobile Radio Communications, IEEE, sep 2009.

[50] C. Han, T. Harrold, S. Armour, I. Krikidis, S. Videv, P. Grant, H. Haas, J. Thompson, I. Ku,
C.-X. Wang, T. Le, M. Nakhai, J. Zhang, and L. Hanzo, “Green radio: radio techniques
to enable energy-efficient wireless networks,” IEEE Communications Magazine, vol. 49,
pp. 46–54, jun 2011.

[51] M. Ismail, W. Zhuang, E. Serpedin, and K. Qaraqe, “A survey on green mobile networking:
From the perspectives of network operators and mobile users,” in IEEE Communications
Surveys & Tutorials, IEEE, 2015.

[52] G. Auer, V. Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson, M. Imran, D. Sabella,


M. Gonzalez, O. Blume, and A. Fehske, “How much energy is needed to run a wireless
network?,” in IEEE Wireless Communications, IEEE, oct 2011.

[53] M. Deruyck, D. D. Vulder, W. Joseph, and L. Martens, “Modelling the power consumption
in femtocell networks,” in 2012 IEEE Wireless Communications and Networking Confer-
ence Workshops (WCNCW), IEEE, apr 2012.

[54] R. Riggio and D. J. Leith, “A measurement-based model of energy consumption in femto-


cells,” in 2012 IFIP Wireless Days, IEEE, nov 2012.

[55] H. Tullberg, P. Popovski, Z. Li, M. A. Uusitalo, A. Hoglund, O. Bulakci, M. Fallgren, and


J. F. Monserrat, “The METIS 5G system concept: Meeting the 5G requirements,” in IEEE
Communications Magazine, IEEE, dec 2016.

[56] S. Tombaz, P. Frenger, F. Athley, E. Semaan, C. Tidestav, and A. Furuskar, “Energy perfor-
mance of 5G-NX wireless access utilizing massive beamforming and an ultra-lean system
85

design,” in 2015 IEEE Global Communications Conference (GLOBECOM), Institute of


Electrical and Electronics Engineers (IEEE), dec 2014.

[57] B. Debaillie, C. Desset, and F. Louagie, “A flexible and future-proof power model for
cellular base stations,” in 2015 IEEE 81st Vehicular Technology Conference (VTC Spring),
IEEE, may 2015.

[58] Y. S. Soh, T. Q. S. Quek, M. Kountouris, and H. Shin, “Energy efficient heterogeneous


cellular networks,” in IEEE Journal on Selected Areas in Communications, IEEE, may
2013.

[59] J. Wu, Y. Zhang, M. Zukerman, and E. K.-N. Yung, “Energy-efficient base-stations sleep-
mode techniques in green cellular networks: A survey,” in IEEE Communications Surveys
& Tutorials, IEEE, 2015.

[60] K. C. Tun and K. Kunavut, “An overview of cell zooming algorithms and power saving
capabilities in wireless networks,” in KMUTNB International Journal of Applied Science
and Technology, King Mongkut University of Technology North Bangkok, sep 2014.

[61] G. Cili, H. Yanikomeroglu, and F. R. Yu, “Cell switch off technique combined with co-
ordinated multi-point (CoMP) transmission for energy efficiency in beyond-LTE cellular
networks,” in 2012 IEEE International Conference on Communications (ICC), IEEE, jun
2012.

[62] A. Dudnikova, P. Dini, L. Giupponi, and D. Panno, “Multi-criteria decision for small
cell switch off in ultra-dense LTE networks,” in 2015 13th International Conference on
Telecommunications (ConTEL), IEEE, jul 2015.

[63] A. Dudnikova, P. Dini, L. Giupponi, and D. Panno, “Fuzzy multiple criteria switch off
method for dense heterogeneous networks,” in 2015 IEEE 20th International Workshop on
Computer Aided Modelling and Design of Communication Links and Networks (CAMAD),
IEEE, sep 2015.

[64] T. En, P. Zhiwen, Y. Xiaohu, and W. Yu, “A low complexity energy efficiency optimization
algorithm based on optimal switching-off eNB selection in LTE networks,” in 2014 Sixth
International Conference on Wireless Communications and Signal Processing (WCSP),
IEEE, oct 2014.

[65] A. Bousia, A. Antonopoulos, L. Alonso, and C. Verikoukis, “Green distance-aware base


station sleeping algorithm in LTE-advanced,” in 2012 IEEE International Conference on
Communications (ICC), IEEE, jun 2012.
86

[66] G. Micallef, P. Mogensen, and H.-O. Scheck, “Cell size breathing and possibilities to
introduce cell sleep mode,” in 2010 European Wireless Conference (EW), IEEE, 2010.

[67] Z. Niu, Y. Wu, J. Gong, and Z. Yang, “Cell zooming for cost-efficient green cellular net-
works,” IEEE Communications Magazine, vol. 48, pp. 74–79, nov 2010.

[68] M. Fallgren, M. Olsson, and P. Skillermark, “Energy saving techniques for LTE: Integra-
tion and system level results,” in 2013 IEEE 24th Annual International Symposium on
Personal, Indoor, and Mobile Radio Communications (PIMRC), IEEE, sep 2013.

[69] P. Frenger and M. Ericson, “Assessment of alternatives for reducing energy consumption in
multi-RAT scenarios,” in 2014 IEEE 79th Vehicular Technology Conference (VTC Spring),
IEEE, may 2014.

[70] 3GPP TS 23.009, Digital cellular telecommunications system (Phase 2+); Universal Mo-
bile Telecommunications System (UMTS); Handover procedures (Release 10), 2011.

[71] H.-W. Ferng and Y.-Y. Huang, “Handover scheme with enode-b pre-selection and param-
eter self-optimization for LTE-a heterogeneous networks,” in 2016 International Confer-
ence on Machine Learning and Cybernetics (ICMLC), IEEE, jul 2016.

[72] K. Vasudeva, M. Simsek, D. Perez, and I. Guvenc, “Analysis of handover failures in hetero-
geneous networks with fading,” in IEEE Transactions on Vehicular Technology, pp. 1–1,
IEEE, 2016.

[73] Y. Lee, B. Shin, J. Lim, and D. Hong, “Effects of time-to-trigger parameter on handover
performance in SON-based LTE systems,” in 2010 16th Asia-Pacific Conference on Com-
munications (APCC), IEEE, oct 2010.

[74] 3GPP TS 36.214, LTE; Evolved Universal Terrestrial Radio Access (E-UTRA); Physical
layer - Measurements (Release 9), 2010.

[75] 3GPP TS 36.331, LTE; Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Re-
source Control (RRC); Protocol specification (Release 9), 2012.

[76] 3GPP TR 36.839, Technical Specification Group Radio Access Network; Evolved Univer-
sal Terrestrial Radio Access (E-UTRA); Mobility enhancements in heterogeneous networks
(Release 11), 2012.

[77] EARTH, Energy efficiency analysis of the reference systems, areas of improvements and
target breakdown, jan 2012.

[78] UMTS Forum, Mobile traffic forecasts 2010-2020, jan 2011.


87

[79] METIS, Deliverable D6.1 - Simulation Guidelines, oct 2013.

[80] 3GPP TR 36.814, Technical Specification Group Radio Access Network; Evolved Univer-
sal Terrestrial Radio Access (E-UTRA),Further advancements for E-UTRA physical layer
aspects (Release 9), mar 2010.

[81] T. L. Saaty, “Decision making with the analytic hierarchy process,” International journal
of services sciences, vol. 1, no. 1, pp. 83–98, 2008.

[82] T. L. Saaty, “How to make a decision: The analytic hierarchy process,” European Journal
of Operational Research, vol. 48, pp. 9–26, sep 1990.

[83] Recommendation ITU-T L.1310, “Series l: Construction, installation and protection of


cables and other elements of outside plant; energy efficiency metrics and measuerement
methods for telecommunication equipment,” tech. rep., 2014.

[84] W. Vereecken, W. Heddeghem, M. Deruyck, B. Puype, B. Lannoo, W. Joseph, D. Colle,


L. Martens, and P. Demeester, “Power consumption in telecommunication networks:
overview and reduction strategies,” IEEE Communications Magazine, vol. 49, no. 6,
pp. 62–69, 2011.

[85] APC and Hivos, Global Information Socitey Watch, 2010.

You might also like