Technische Universität München

Fakultät für Elektrotechnik und Informationstechnik

Optimal Selection and Adaptation of a

Flexible Functional Split in 5G Radio
Access Networks

Alberto Martínez Alba, M. Sc.

Vollständiger Abdruck der von der Fakultät für Elektrotechnik und Informationstech-
nik der Technischen Universität München zur Erlangung des akademischen Grades
eines

Doktors der Ingenieurwissenschaften (Dr.-Ing.)

genehmigten Dissertation.

Vorsitzender: Prof. Dr. Andreas Herkersdorf

Prüfer der Dissertation: 1. Prof. Dr.-Ing. Wolfgang Kellerer
2. Prof. Dr. Navid Nikaein

Die Dissertation wurde am 28.09.2021 bei der Technischen Universität München ein-
gereicht und durch die Fakultät für Elektrotechnik und Informationstechnik am 28.02.2022
angenommen.
Optimal Selection and Adaptation of a
Flexible Functional Split in 5G Radio
Access Networks

Alberto Martínez Alba, M. Sc.

September 27, 2021

Abstract

The fifth generation of radio access networks aims at ambitious performance objec-
tives, such as offering low-latency ultra-reliable communication, supporting massive
machine-type deployments, and providing high data rates to conventional users. In
order to accomplish these objectives, the network operator must procure an efficient
management of the available network resources. A promising strategy to improve
this efficiency is to increase the density of the deployed cells in the network, which
enables the use of high frequency bands and reduces power consumption. However,
increasing cell density may lead to aggravated inter-cell interference, thus hindering
the achievement of high data rates and compromising network efficiency. In order
to reduce the impact of inter-cell interference, it is proposed to split the functions of
5G base stations into two units: a centralized unit and a distributed unit. This func-
tional split facilitates interference coordination among base stations and, in addition,
reduces the operating and deployment costs of the network. Nonetheless, there is
not a single optimal manner to perform this split. Depending on the location of the
users and their instantaneous activities, the functional split that provides the best per-
formance or the lowest cost may change. As a result, it would be clearly beneficial
to adapt the optimal split to the instantaneous network situation. In this thesis, we
investigate the problem of selecting and dynamically adapting the optimal functional
split of a 5G radio access network. In order to do this properly, we present a dedicated
cost model for flexible communication networks, which takes into account the cost of
operating and changing the network state. Then, we formulate the functional split se-
lection as non-linear mixed-integer optimization problems for various objectives, and
evaluate multiple candidate strategies to solve them in a timely manner. In compari-
son with static approaches, we observe that a dynamically optimized functional split
may lead to substantial increases in the user performance and operating cost reduc-
tions. Besides the theoretical contribution, we also describe a novel implementation of
a radio access network that is able to change its functional split during runtime, with
the intention of demonstrating the feasibility of a dynamic functional split adaptation.
Finally, we propose and compare several adaptation strategies for the network opera-
tor to decide when and how to change the functional split. We conclude that a lively
adapted functional split is indeed possible and may result in considerably improved
user data rates and lower operating cost with respect to conventional approaches.

iii
Kurzfassung

Die fünfte Generation von Funkzugangsnetzen verfolgt ehrgeizige Leistungsziele, wie

zum Beispiel die Bereitstellung einer extrem zuverlässigen Kommunikation mit niedri-
gen Latenzzeiten, die Unterstützung massiver maschinenartiger Anwendungen und
die Bereitstellung hoher Datenraten für konventionelle Nutzer. Um diese Ziele zu er-
reichen, muss der Netzbetreiber für eine effiziente Verwaltung der verfügbaren Net-
zressourcen sorgen. Eine vielversprechende Strategie zur Verbesserung dieser Ef-
fizienz besteht darin, die Dichte der im Netz eingesetzten Zellen zu erhöhen, was die
Nutzung von Hochfrequenzbändern ermöglicht und den Stromverbrauch reduziert.
Die Erhöhung der Zelldichte kann jedoch zu einer verstärkten Interferenz zwischen
den Zellen führen, was das Erreichen hoher Datenraten behindert und die Netzef-
fizienz beeinträchtigt. Um die Auswirkungen von Interferenzen zwischen den Zellen
zu verringern, wird vorgeschlagen, die Funktionen von 5G-Basisstationen in zwei
Einheiten aufzuteilen: eine zentrale Einheit und eine dezentrale Einheit. Diese funk-
tionale Aufteilung erleichtert die Interferenzkoordination zwischen den Basisstatio-
nen und senkt darüber hinaus die Betriebs- und Errichtungskosten des Netzes. Es
gibt jedoch nicht die eine optimale Art und Weise, diese Aufteilung vorzunehmen. Je
nach Standort der Nutzer und ihren momentanen Aktivitäten kann sich die Funktion-
saufteilung, die die beste Leistung oder die niedrigsten Kosten bietet, ändern. Daher
wäre es eindeutig von Vorteil, die optimale Aufteilung an die momentane Netzsitu-
ation anzupassen. In dieser Arbeit untersuchen wir das Problem der Auswahl und
der in Echtzeit laufenden Anpassung der optimalen Funktionsaufteilung eines 5G-
Funkzugangsnetzes. Zu diesem Zweck stellen wir ein spezielles Kostenmodell für
flexible Kommunikationsnetze vor, das die Kosten für den Betrieb und die änderung
des Netzzustands berücksichtigt. Anschließend formulieren wir das Problem der
Funktionsaufteilung als nichtlineares gemischt-ganzzahliges Optimierungsproblem
für verschiedene Ziele und bewerten mehrere Kandidatenstrategien, um sie zeitnah
zu lösen. Im Vergleich zu statischen Ansätzen stellen wir fest, dass eine dynamisch
optimierte Funktionsaufteilung zu einer erheblichen Steigerung der Nutzerleistung
und einer Senkung der Betriebskosten führen kann. Neben diesem theoretischen
Beitrag demonstrieren wir die Machbarkeit einer dynamischen Anpassung der Funk-
tionsaufteilung anhand einer neuartigen Implementierung eines Funkzugangsnetzes,
welches seine Funktionsaufteilung während der Laufzeit kann. Schließlich schlagen
wir verschiedene Anpassungsstrategien für den Netzbetreiber vor und vergleichen
diese, um zu entscheiden, wann und wie der Funktionssplit geändert werden soll.
Wir kommen zu dem Schluss, dass ein dynamisch angepasster Funktionssplit in der
Tat möglich ist und im Vergleich zu konventionellen Ansätzen zu erheblich besseren
Nutzerdatenraten und niedrigeren Betriebskosten führen kann.

v
Contents

1. Introduction 1
1.1. Research challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2. Main contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3. Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2. The Dynamic Functional Split Selection Problem 11

2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1. Motivation, scope, and challenges . . . . . . . . . . . . . . . . . . 11
2.1.2. Key contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2. Centralized 5G RAN architectures . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1. 5G use cases and requirements . . . . . . . . . . . . . . . . . . . . 13
2.2.2. Interference mitigation in 5G networks . . . . . . . . . . . . . . . 14
2.2.3. The Cloud-RAN architecture . . . . . . . . . . . . . . . . . . . . . 16
2.2.4. Functional Split in 5G networks . . . . . . . . . . . . . . . . . . . 19
2.3. Functional split selection problem . . . . . . . . . . . . . . . . . . . . . . 22
2.3.1. System modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.2. Static FSSP formulation . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4. Solving the FSSP dynamically . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.1. Flexibility in softwarized communication networks . . . . . . . . 27
2.4.2. Flexible functional split platform . . . . . . . . . . . . . . . . . . . 30
2.4.3. Dynamic FSSP formulation . . . . . . . . . . . . . . . . . . . . . . 31
2.5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3. Modeling the Cost of Flexible Communication Networks 35

3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.1. Motivation, scope, and challenges . . . . . . . . . . . . . . . . . . 35
3.1.2. Key contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2. Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.1. Static cost models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.2. Dynamic cost models . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.3. Flexibility models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3. System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.1. Network states and demands . . . . . . . . . . . . . . . . . . . . . 39
3.3.2. The adaptation process . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.3. Flexibility measure . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.4. Cost model of a flexible network . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.1. General definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.2. Readiness cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

vii
Contents

3.4.3. Proaction cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.4.4. Reaction cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5. Application example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5.1. Network description . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5.2. Selection of a profitable population size . . . . . . . . . . . . . . . 56
3.5.3. Optimal parallelization level . . . . . . . . . . . . . . . . . . . . . 57
3.6. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4. Optimal Functional Split Selection in 5G Radio Access Networks 61

4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.1.1. Motivation, scope, and challenges . . . . . . . . . . . . . . . . . . 61
4.1.2. Key contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2. Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.3. System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3.1. Network description . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3.2. Functional splits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3.3. Interference mitigation . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.4. Fronthaul network . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.4. Performance-maximizing FSSP . . . . . . . . . . . . . . . . . . . . . . . . 67
4.4.1. Proportionally-fair formulation . . . . . . . . . . . . . . . . . . . . 68
4.4.2. Fractional approximations . . . . . . . . . . . . . . . . . . . . . . . 73
4.4.3. Quadratic formulation . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4.4. Heuristic approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5. Operating-cost-minimizing FSSP . . . . . . . . . . . . . . . . . . . . . . . 83
4.6. Readiness-cost-minimizing FSSP . . . . . . . . . . . . . . . . . . . . . . . 85
4.7. Experimental evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.7.1. Simulation setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.7.2. Convergence time . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.7.3. Spectral efficiency and cost evaluation . . . . . . . . . . . . . . . . 95
4.8. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

5. Implementation of an Adaptive Functional Split 109

5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.1.1. Motivation, scope, and challenges . . . . . . . . . . . . . . . . . . 109
5.1.2. Key contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.2. Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.1. Academic research . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.2. 4G/5G Software Platforms . . . . . . . . . . . . . . . . . . . . . . 112
5.3. Considered functional splits . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.3.1. PDCP-RLC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.3.2. MAC-PHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.4. Adaptive functional split . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.4.1. Objectives and challenges . . . . . . . . . . . . . . . . . . . . . . . 116
5.4.2. Migration strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.4.3. State transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

viii
 Contents

5.4.4. Migration platform . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

5.4.5. Proposed migration procedure . . . . . . . . . . . . . . . . . . . . 121
5.5. Implementation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.5.1. Hardware platform description . . . . . . . . . . . . . . . . . . . . 123
5.5.2. Measurement results . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.6. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

6. Dynamic Functional Split Adaptation in Real Time 129

6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.1.1. Motivation, scope, and challenges . . . . . . . . . . . . . . . . . . 129
6.1.2. Key contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.2. Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.3. Optimal state monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.4. Time-dependent cost modeling . . . . . . . . . . . . . . . . . . . . . . . . 134
6.4.1. Demand modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.4.2. Time dependency modeling . . . . . . . . . . . . . . . . . . . . . . 135
6.4.3. Mean readiness and action costs . . . . . . . . . . . . . . . . . . . 136
6.5. Adaptation strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.5.1. Uniform-static strategy . . . . . . . . . . . . . . . . . . . . . . . . 143
6.5.2. Mean-static strategy . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.5.3. Impatient strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.5.4. Random deferral strategy . . . . . . . . . . . . . . . . . . . . . . . 145
6.5.5. Dynamic programming strategy . . . . . . . . . . . . . . . . . . . 146
6.5.6. Greedy strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.5.7. Adaptive threshold strategy . . . . . . . . . . . . . . . . . . . . . . 148
6.6. Reference action cost frontiers . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.6.1. Impatient frontier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.6.2. Dynamic frontier . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.6.3. Static frontier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
6.7. Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.7.1. Simulator description . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.7.2. Derivation of K∆ (δ) and FT (t) . . . . . . . . . . . . . . . . . . . . 156
6.7.3. Selection of monitoring period Zmon . . . . . . . . . . . . . . . . . 158
P
6.7.4. Proaction cost C estimation . . . . . . . . . . . . . . . . . . . . . 160
6.7.5. Comparison of adaptation strategies . . . . . . . . . . . . . . . . . 161
6.7.6. Reference action-cost frontiers . . . . . . . . . . . . . . . . . . . . 162
6.8. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

7. Conclusion and Outlook 167

7.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.2. Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

A. Proofs and derivations 171

A.1. Proof of Property 3.4.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
A.2. Proof of Property 3.4.10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

ix
Contents

A.3. Estimation of FT (t) from sampled durations . . . . . . . . . . . . . . . . 174

A.4. Estimation of RDF from time-based cost degradation . . . . . . . . . . . 175

Bibliography 177

List of Figures 197

List of Tables 203

x
1. Introduction

Mobile network operators rely on constant innovation to satisfy the increasingly strin-
gent demands of their ever-growing number of customers. The ubiquity and conve-
nience of mobile communications attract both end users and application providers,
between which operators are forced to unceasingly improve their management and
infrastructure in order to keep up with the expectations. Indeed, the mobile operators
that provided mobile connectivity to 8.8 billion mobile devices in 2018 are foreseen to
face a demand of 13.1 billion devices by 2023 [Cis20], of which 1.4 billion devices will
be 5G-capable. Moreover, the average download speed of these 5G devices is expected
to grow from 76 Mb/s in 2019 to 575 Mb/s by 2023 [Cis20], owing to the emergence of
applications requiring high data rates, such as those featuring high-definition video.
In fact, while 59% of mobile traffic in 2017 is already video traffic, by 2022 video traffic
will be almost 80% of all mobile traffic [Cis19]. These trends are spurring the redesign
of 5G radio access networks (RAN) so as to improve their provided service.
Not only mobile operators are compelled to provide increasingly better service to
more users, but they also need to do this at relatively lower prices. In fact, the world-
wide median price of providing a data service of 1.5 GB per month dropped an aver-
age of 10.4% every year from 2013 to 2020, resulting in an accumulated price reduction
of around 50% in seven years [Uni21]. This trend is even steeper for developing and
least-developed countries, exhibiting yearly price reductions of 13.3% and 18.5%, re-
spectively, which translates in ca. 60% and 75% accumulated price reductions [Uni21].
As direct consequence, the revenue associated with providing these mobile services
tends to decrease as well. For example, European mobile operators reported a total
revenue of 177 billion Euros in 2006, whereas in 2016 this revenue was only 138 billion
Euros, a 22% difference, even though the total number of connected devices increased
by 40% [GAO17].
Having to provide better user service at lower prices would be enough motivation
for mobile operators to pursue high resource efficiency. Furthermore, nowadays cus-
tomers explicitly demand this feature to operators, either because of environmental
awareness [Eur17; Eur21] or even because of health concerns [Ips20]. Thus, efficiency
is not only strictly required to profitably operate a communication network, but it can
be also used as a selling feature to attract potential users. Consequently, mobile op-
erators continuously seek strategies to improve the performance-to-cost ratio of their
current infrastructure before considering to invest in additional resources, and prefer
efficient resource utilization over simpler but more costly operation.
A mandatory step in order to increase resource efficiency of mobile radio access net-
works is to increase network density, that is, to deploy more base stations per area

1
1. Introduction

unit [KS17]. In dense RANs, users are always close to its serving base station, which
enables the use of high-frequency bands and increases the received signal power with
comparatively low transmission power. In contrast, sparse RANs tend to produce
more shadow areas even though the transmission powers of the base stations are
higher. However, increased RAN density also leads to increased inter-cell interfer-
ence, which may jeopardize the overall resource efficiency.
Techniques for dealing with interference exist since the the inception of mobile net-
works and have evolved over time, ranging from simple frequency allocation [Eng+98]
to advanced joint transmission and reception [Sey+16]. Simple techniques can be
implemented easily, but their performance is limited and are not sufficient to oper-
ate dense deployments. Advanced techniques promise, in theory, powerful interfer-
ence avoidance or cancellation, but their actual effectiveness may be severely affected
by implementation details [Jun+10]. Namely, the more advanced the interference
technique, the more and better coordination is needed among involved base stations
[Döt+13]. This translates into strict latency and capacity constraints between coordi-
nating base stations, which legacy 3G and 4G networks struggle to satisfy.
In conventional 4G networks, each base station individually receives and processes
user data packets, and takes scheduling, encoding, and modulation decisions inde-
pendently from other cells [3GP21i]. This RAN architecture is depicted in Fig. 1.1a. Al-
though there are logical interfaces designed to provide direct communication between
base stations and promote inter-cell coordination [3GP20a], their limitations, owing
to the fact that cells may be located far away from each other, often restrict their abil-
ity to coordinate to static or semi-static interference mitigation techniques [Kos+12].
However, with the emergence of network softwarization techniques, new network
architectures are proposed so as to provide better inter-cell coordination. Namely, if
mobile processing functions, which are usually deployed on dedicated hardware, are
replaced by software functions, these can be easily relocated into commodity data cen-
ters. This leads to a centralized RAN architecture, in which the processing of all base
stations is moved to a single central location [Che+14a], as depicted in Fig. 1.1b. As
a result, this architecture enables fast communication between all cells, thus allowing
the use of advanced interference mitigation techniques, which are required to operate
dense networks.
Function centralization also entails lower deployment and operating costs, since the
computing resources of off-the-shelf equipment can be pooled to leverage multiplex-
ing gain [Che+14a]. Nonetheless, centralized RAN architectures have also a major
disadvantage. Moving all processing functions away from the remote sites, which
contain the antennas and radio equipment, requires the presence of high-capacity,
low-latency links between them and the data center [Döt+13]. Current mobile opera-
tors, however, cannot meet these requirements without re-deploying the whole RAN,
which incurs in high additional costs [Che+16a].
The solution proposed by 3GPP to benefit from the advantages of function centraliza-
tion without compromising the capacity of currently-deployed RANs is to opt for a
partially centralized architecture [3GP21i], as shown in Fig. 1.1c. With partial centraliza-

2
 Core network Centralized unit Distributed unit

Remote unit RAN functions

(a) Distributed RAN. (b) Fully centralized RAN.

(c) Partially centralized RAN. (d) Dynamically centralized RAN.

Figure 1.1.: Architecture of distributed, fully centralized, partially centralized, and dy-
namically centralized radio access networks.

tion, only a subset of the processing functions is moved to the data center, so that the
links between the data center and the remote sites are not overloaded. This may still
lead to cost reductions and enhanced interference mitigation techniques, depending
on which and how many functions are centralized. In turn, a partially centralized ar-
chitecture introduces a new problem: selecting which functions should be centralized.
Since function centralization affects both the operating cost and the user performance,
via interference mitigation, the operator has to use these two performance indicators
to decide on the optimal centralization level. In addition, as mobile networks are rather
dynamic due to the mobility of the users and their time-dependent patterns, the opti-
mal centralization level may vary over time. Hence, being able to adapt the central-
ization level to the instantaneous network conditions, as depicted in Fig. 1.1d, may
result in higher profits with respect to a static configuration [MAK19a]. A dynamic
adaptation of the centralization level is, nonetheless, a completely novel feature for
a mobile RAN. As such, a careful consideration of all the time-varying cost compo-

3
1. Introduction

nents, including those related to adapting the centralization level and those reflecting
user performance, is strictly required. Furthermore, it is also necessary to show that
the current state of technology can support this novel adaptation, hopefully without
incurring in service disruptions.
In the light of this, the objectives of this doctoral thesis can be summarized as follows.
First, we intend to improve the understanding of the influence of the centralization
level on the performance and operating cost of a partially centralized 5G RAN. This
also includes modeling cost of operating with an inadequate centralization level, and
the cost of updating it when so decided. The second goal is to find a timely and
cost-efficient algorithm to find the optimal centralization level, so that the network
operator can be aware of its evolution and anticipate the benefits of updating it. Third,
we want to show that the set of centralized functions can be dynamically changed
in an actual 5G network without severely affecting user performance or increasing
operating cost, thus proving that a dynamic reconfiguration of a 5G RAN is technically
feasible. Finally, we aim to apply our knowledge about the optimal centralization
level and the changing capabilities of a 5G RAN to derive an optimal rule for deciding
when to trigger a reconfiguration.

1.1. Research challenges

The modeling, design, and optimization of 5G radio access networks featuring flexible
centralization levels entail several research challenges. This section summarizes these
challenges, which are discussed in detail in Chapters 3 to 6.
Optimal selection of the RAN centralization level
Being constrained by the capacity of the network connecting the data center with the
remote units, the first challenge is to find a feasible subset of centralized RAN func-
tions that maximizes a chosen performance indicator. This can be formulated as an
optimization problem, in which the objective function reflects our performance indi-
cator (such as operating cost or user throughput) and the constraints model the ca-
pacity of the network. If each centralized function had an associated constant utility
and the network constraints could be converted into an equivalent upper limit to the
number of centralized functions, we could formulate this problem as an instance of
the well-known knapsack problem [MT90; MAK19a]. Nonetheless, in real networks
the objective function may be non-linear and the network constraints may impose
more complicated restrictions. Therefore, the actual optimization problem has to be
formulated from a more complicated network model.
We need to model how the centralization level affects the required capacity on every
link and make sure that any obtained solution is within their actual capacity. If the ob-
jective function depended only on routing costs, we could achieve this by formulating
an instance of the multicommodity flow problem [AMO13]. However, the objective
function needs to reflect multiple additional factors. On the one hand, the operator

4
 1.1. Research challenges

may be interested in minimizing the operating cost associated with each selection
of centralized functions. On the other hand, satisfying user demands is also impor-
tant for the operator, so that we could also focus on centralizing those functions that
maximize user utility. For this, we need a model of the mutual interaction between
coordinating base stations and its conversion into interference mitigation, then trans-
form interference mitigation into user throughput, and finally relate this throughput
to actual utility or quality of experience.
After the optimization problem is formulated, we have to figure out an adequate ap-
proach to solve it. We can safely anticipate that this optimization problem is hard to
solve, based for example on the fact that a very simplified version of it would result in
an instance of the knapsack problem, which is already NP-Hard [MT90]. At the same
time, since an actual dense RAN may include hundreds of base stations, it is clear that
problem instances may be rather large. Besides, if we intend to dynamically adapt the
centralization level, we need to be able to calculate optimal or near-optimal solutions
in a timely manner, which may be a difficult feat for large, complex optimization prob-
lems. Consequently, we may need to derive approximating approaches or heuristics
to find good-quality solutions that converge faster than exact approaches.
Cost-efficient management of a flexible 5G RAN
Once the mobile network operator has a timely and effective algorithm to find the op-
timal centralization level for a given situation, an important design decision has to be
taken. Namely, the operator must decide between a static configuration, in which the
set of centralized functions is not modified during operation, and a dynamic opera-
tion, in which the centralization level is adapted to the instantaneous situation. Owing
to the dynamic nature of mobile networks, the disadvantages of the former are clear:
the instantaneous traffic and user distribution may differ substantially from the one
at which the static configuration performs optimally. Conversely, a RAN featuring
a dynamically-adapting centralization level could track changes in the environment
and thus maximize performance and minimize cost over time. Nonetheless, we need
to take into account the cost implications of dynamically reconfiguring the centraliza-
tion level in order to conclude that dynamic adaptation is indeed preferable.
A model of the cost of a dynamically-adapting network has to reflect various compo-
nents. First, we need to find out how costly it is to operate on suboptimal, obsolete
centralization levels with respect to the optimal one. If the cost difference is small,
there may be no incentive to perform a change. For this, characterizing the degra-
dation of optimal solutions over time is required. Second, we have to estimate how
costly it is to perform a change in the centralization level. There may be multiple
factors involved in this estimation: additional resource consumption, compensations
for service disruptions, increased end-to-end latency, etc. Finally, we have to ensure
that the network can indeed cope with environmental changes fast enough. That is,
that the delay between a change in the optimal centralization level and its associated
response is not too long, otherwise dynamic adaptation may be pointless.

5
1. Introduction

Realizing a reconfigurable 5G RAN

The ability to dynamically change the centralization level, that is, to modify the set
of centralized functions during runtime, is completely novel for mobile radio access
networks. Although there are works from the research community investigating this
feature [Die+21; DHA20; HR18b], current 5G specifications only describe a static, par-
tially centralized architecture that is identical for all base stations [3GP21d], and at
most they consider the possibility of supporting multiple static centralization levels
on different base stations in the future [3GP17]. Therefore, in order to show the effec-
tiveness of a dynamically-adapting RAN, we must ensure that this is indeed techni-
cally feasible with the current state of technology.
Realizing a software platform that supports standard 5G operation while being able
to centralize or distribute functions is a challenging task. Numerous constraints have
to be taken into account not to disrupt the user connections, such as ensuring that no
packets are lost, end-to-end latency is not severely increased, signaling is still routed
and processed correctly, etc. Albeit there are software platforms supporting live mi-
gration of virtual network functions [Rup+18], they still may not be ideal at meeting
these constraints. As a result, a dedicated migration platform may be required.
Dynamic adaptation of a RAN centralization level
Let us assume that the previous three challenges have been successfully overcome.
We have derived a fast and good-quality optimization approach to find the optimal or
near-optimal centralization levels for a given traffic and distribution of users. In addi-
tion, we have accurately modeled the cost of operating at each possible centralization
level, and the cost of changing it. Finally, we have a migration platform that enables
such changes in a timely and cost-efficient manner. Even in that case, we still need
to have a rule to help us decide when to change the centralization level. If the cost of
moving RAN functions is larger than the average cost difference between the optimal
and recently obsolete centralization levels, it would make sense not to change the cen-
tralization level continuously, but to wait until the change is worthwhile. However,
waiting too long may be counterproductive, since it would also result in increased op-
erating cost. There must be, hence, an optimal waiting time between both extremes,
which is not trivial to find.
Therefore, in order to profitably operate a flexible 5G RAN, we need to devise a way to
find an optimal or near-optimal rule to trigger adaptations of the centralization level.
The main difficulty of this task stems from the fact that variations in user traffic and
mobility are hard to estimate accurately. This may result in inadequate adaptation
rules that perform suboptimally.

1.2. Main contributions

In this section, we provide a summary of the main contributions of this doctoral thesis,
in relation to the aforementioned research challenges. We can classify these contribu-

6
 1.2. Main contributions

tion into four major groups: (i) proposal of an optimization framework to find the best
centralization level, (ii) derivation of a cost model for flexible networks, (iii) imple-
mentation of a proof-of-concept adaptive 5G RAN, and (iv) proposal and evaluation
of dynamically adaptive strategies.
The first contribution comprises the network modeling, problem formulation, and
approach derivation and assessment required to find the optimal centralization level
for any instantaneous network situation. Namely, we model the interference coor-
dination capabilities of all base stations according to their centralization level, how
this coordination translates into user throughput, and the final transformation into
proportionally fair utility to better reflect the quality of experience of the user. This
model is then combined with the network constraints to formulate an optimization
problem, which may be used to maximize user throughput in a proportionally fair
manner, minimize operating cost, or maximize combined operating revenue, which
depends on both performance and operating cost. Furthermore, we relax the orig-
inal, intractable problem formulation into increasingly simpler and faster approxi-
mations that nonetheless yield good quality solutions, with the intention of finding
a suitable algorithm to support dynamic optimization. All these formulations are
fully independent on the network parameters, and thus they can be employed for
an arbitrary number of base stations, users, and centralization options, link capaci-
ties, interference-mitigation capabilities, etc. Finally, the convergence time and per-
formance of all considered approaches are thoroughly evaluated under a wide range
of network conditions and compared against static approaches.
The second contribution addresses the derivation of a model that reflects all cost com-
ponents playing a role in the operation of a flexible, dynamically-adapting network.
Namely, we define and propose expressions to estimate the cost of operating the net-
work in optimal and obsolete states, as well as the cost of deciding and realizing state
changes. We present new functions characterizing the cost degradation of a commu-
nication network and base upon a probability-theory framework to derive a complete
estimation of the average cost. This estimation is then combined with that of the cost
of deciding and realizing state changes, which are analyzed separately, in order to
yield a final expression for the total operating cost. This can be used to configure de-
sign parameters and decide whether a dynamic network is profitable. In addition to
using this cost model for our own use case, we provide example applications of how
to employ it on other types of communication networks.
The third contribution is more practical, since it concerns the implementation of a
proof-of-concept 5G RAN that features the ability of changing its centralization level
without service interruptions. Even though we show via theoretical analysis and sim-
ulations that a dynamic centralization level often leads to better performance and
lower cost than static solutions, it would be unclear whether these benefits are achiev-
able in practice without an actual implementation, owing to the novelty of the pro-
posed architecture and adaptation approaches. That is the reason why we present the
design of a simple 5G RAN that can change between two centralization levels during
runtime. We identify the challenges associated with centralizing and distributing run-

7
1. Introduction

ning functions, such as rerouting their data and control traffic, transferring the state
of transmission and reception buffers, switching on and off processing elements, etc.
Then, we propose a migration strategy that provides a seamless transitions between
a lower and a higher centralization level, with minimal packet losses and negligible
migration latency. In addition, we discuss a strategy to completely remove packet
losses at the expense of slightly higher latency, and vice-versa, so that these can be
configured to the preference of the operator.
The fourth and last contribution consists in the derivation and evaluation of adapta-
tion strategies, which allow us to identify the best moments to trigger a change in the
centralization level. For this, we use the previously derived cost model to characterize
the cost of keeping the current the centralization level and compare it against the cost
of changing it and switching to the optimal configuration. We realize that the deriva-
tion of the optimal rule is not possible unless full information is known about the
future evolution of the system. Nonetheless, we propose near-optimal, self-adjusting
strategies that approximate the performance of the optimal rule. Finally, we evalu-
ate all considered strategies under a wide range of mobility dynamics and network
conditions.

1.3. Outline

A detailed diagram showing the outline of the thesis and summarizing the content of
each chapter is depicted in Fig. 1.2. In essence, the remainder of this thesis is organized
as follows.
Chapter 2 introduces the problem of optimally selecting the functional split which de-
fines the centralization level of each base station in a 5G RAN. We first describe the
architectural options that are considered in 5G and discuss their requirements, their
associated interference-mitigation capabilities, and their overall advantages and dis-
advantages. Then, we introduce the general network model that we use throughout
the thesis and use it to formulate static and dynamic optimization problems.
Chapter 3 presents the model that characterizes the cost components of an adaptive
communication network. We propose new terminology to denote the distinct adap-
tation phases, along with the features of user demands and network states, based on
a probability-theory framework. We analyze the properties of the discussed compo-
nents and derive expressions to combine in order to assess the overall network prof-
itability under dynamic conditions. Finally, we apply the cost model into a generic
simulation example so as to demonstrate how it can be used.
Chapter 4 addresses the formulation of the optimization problem to select the instan-
taneously best centralization level according to user performance, operating cost, and
combined revenue (also known as readiness cost). We tackle mixed-integer non-linear
problems by proposing approximations and alternative reformulations to increase
their tractability. In addition, we discuss two simple heuristic approaches. Then, we

8
 1.3. Outline

Chapter 1
Introduction

Chapter 2
The Dynamic Functional Split Selection Problem
Main References

[MAJK21] [MAK21] [He+19]

Chapter 3
Modeling the Cost of Flexible Communication Networks

Investigated Fields
and Contributions Methodologies Main References

Adaptation Process Analysis Probability Theory [MA+21]

Cost Model Definition Simulations [Bab+20]

Chapter 4
Optimal Functional Split Selection in 5G Radio Access Networks

Investigated Fields
and Contributions Methodologies Main References

Performance Maximization Mixed-Integer Programming [MAJK21]

Operating Cost Minimization Heuristics [MAK21]

Readiness Cost Minimization Simulations [MAPK21]

Chapter 5
Implementation of an Adaptive Functional Split

Investigated Fields
and Contributions Methodologies Main References

Migration Proposal Protocol Stack Analysis [MAGVK19a]

4G/5G Software Development Measurements [MAGVK19b]

Chapter 6
Modeling the Cost of Flexible Communication Networks

Investigated Fields
and Contributions Methodologies Main References

Cost Model Application Dynamic Programming [MAK21]

Dynamic Problem Formulation Heuristics [MAJK20]

Strategy Definition Simulations [MAK19a]

Chapter 7
Conclusion and Outlook

Figure 1.2.: Diagram of the outline of this thesis, including a summary of the contri-
butions, methodologies, and main references of each chapter.

9
1. Introduction

evaluate the convergence times and performance of all formulations by means of a

dedicated simulator, which is designed to represent a wide range of network condi-
tions regarding user distribution and network topology.
Chapter 5 discusses the implementation of our proof-of-concept adaptive 5G RAN,
which features a dynamically reconfigurable centralization level. This implementa-
tion is realized by modifying an underlying 4G/5G software platform. We explain
the challenges of achieving this along with our proposed strategies to migrate run-
ning RAN functions with minimal end-user impact. Finally, we present a selection of
measurement results to back our conclusions.
Chapter 6 explains the application of the cost model presented in Chapter 3 to a dy-
namic 5G RAN, and extends the problem formulations presented in Chapter 4 to in-
clude dynamic optimization. We derive multiple adaptation strategies in order to re-
alize an adaptive optimization, for which we use dynamic programming and heuristic
methods. We lay out the problems of dynamically adapting, propose several adapta-
tion strategies, and evaluate them in extensive simulations.
Finally, Chapter 7 concludes this thesis by summarizing our findings and contribu-
tions and discusses possible directions for future work.

10
2. The Dynamic Functional Split Selection
Problem

2.1. Introduction

2.1.1. Motivation, scope, and challenges

The evolution from 4G to 5G entails improving a large number of features in legacy

radio access networks (RAN). The motivation behind every 5G improvement can be
related to two distinct objectives. On the one hand, operators intend to enhance user
experience by reducing latency, increasing data rates, or supporting a high number of
devices. This is performed in order to attract new users and facilitate the appearance
of novel network applications. On the other hand, increasing network efficiency is
also of utmost importance, since resources are limited and relative user income tends
to decline over time. As a result, mobile network operators always seek to benefit
from the latest advances in technology.
A good example of such advances is the emergence of network function virtualiza-
tion (NFV) and software-defined networking (SDN), which have not gone unnoticed
in the field of mobile networks. Indeed, replacing hardware equipment with software
functions promises important cost reductions and improved performance, whereas
being able to dynamically control the network operation from a centralized point also
offers increased flexibility and scalability. One important application of the paradigms
of NFV and SDN into the 5G RAN architecture is the definition of the functional split,
with which the processing of each base station in the network is divided into central-
ized and distributed portions. Nonetheless, owing to the different advantages and
disadvantages of each functional split option, the operator has to carefully decide
which functions can be centralized and how to do this properly given the underlying
network dynamics.
In this chapter, we provide a detailed background and define the problem of dynami-
cally selecting the optimal functional split in a 5G RAN, which is the central matter of
this thesis. In order to do this, we address a series of challenges. We first have to be
aware of the objectives, opportunities, and limitations of 5G radio access networks, so
as to approach this problem from a realistic perspective. Namely, we need to consider
all the possible options for performing functional splits and model their impact on
the performance and cost of a 5G RAN. This also implies taking into account the role
of all network elements, such as those connecting centralized and distributed units.
Moreover, we have to find an appropriate formulation of the functional split selection

11
2. The Dynamic Functional Split Selection Problem

problem with the objective of finding the best realizable option that optimizes cost
and/or performance. Finally, since our ultimate intention is to operate the network in
real time, we also need to include the dynamic aspects of the system into the problem.

2.1.2. Key contributions

The main content and contributions of this chapter are based on those presented in
[MAJK21], [MAK21], and [He+19], which can be summarized in relation to the afore-
mentioned challenges as follows:
1. We provide a detailed description of the requirements, architectures, and pro-
posed functional splits of 5G radio access networks. We put the emphasis on the
impact of interference on the user experience and on how this interference can
be mitigated.
2. We propose a complete network model to reflect all relevant parameters that
may affect the performance and cost of each functional split option.
3. Based on this model, we formulate the functional split selection problem (FSSP)
in a generic manner, only for static scenarios at first, and finally for dynamic
scenarios after all the variable components of the network have been presented.
The remaining of this chapter is organized as follows. In Sec. 2.2, we discuss the al-
ternatives for implementing a centralized 5G RAN, based on the 5G requirements.
Sec. 2.3 presents our modeling approach and formulates the static functional split se-
lection problem. In Sec. 2.4 we describe the flexibility features that are required in a
5G RAN implementing a dynamically adapting functional split and formulate the dy-
namic functional split selection problem. Finally, Sec. 2.5 summarizes and concludes
the chapter.

2.2. Centralized 5G RAN architectures

Whereas 4G radio access networks usually feature distributed architectures, central-
ized or partially-centralized architectures have been considered for the 5G RAN since
its inception, with the objective of improving user performance and reduce manage-
ment costs. In this section, we introduce the proposed centralized architectures for
5G radio access networks. We then describe the multiple techniques for interference
mitigation that can be leveraged by these architectures. Finally, we present the archi-
tectural options that are considered for current 5G and next-generation networks.

12
 2.2. Centralized 5G RAN architectures

2.2.1. 5G use cases and requirements

5G mobile networks address three main use cases [NGM15]: massive machine-type
communications (mMTC), ultra-reliable low-latency communications (URLLC), and
enhanced mobile broadband (eMBB). The first use case, mMTC, considers networks
connecting hundreds or even thousands of machine-type devices, that is, automated
devices that operate without direct human intervention. In URLLC, the focus is on
providing very reliable and low latency end-to-end connectivity, so as to be able to
operate life-critical remote devices through a mobile network. Finally, eMBB, the third
use case, is a direct evolution of previous use cases, since its objective is to provide
higher data rates to human-type devices in urban or rural scenarios.
The mMTC and URLLC use cases envision a transformation of the traditional utiliza-
tion of mobile networks as they move away from human-type communication, which
has been the main driver for the deployment of these networks in the third and fourth
generations. Nonetheless, high-throughput human-type communication is still a ma-
jor concern for mobile network operators. The reason for this is twofold. On the one
hand, owing to the novelty of mMTC and URLLC use cases, their underlying market
is not yet consolidated and the potential profit of supporting these use cases is still
unclear. Conversely, human-type markets are better known and thus their associated
profits are easier to predict. On the other hand, users require ever-increasing data
rates as new applications appear, such as those based on high-quality video stream-
ing [Upd21]. As a result, supporting the eMBB use case is crucial for mobile operators.
In this work, we address the problem of dynamically modifying the RAN architecture
so as to enhance user data rates, and hence we focus on the eMBB use case.
There are three basic strategies to increase data rates in a wireless communication
system. The first one is to increase spectral efficiency, that is, to improve modulation,
coding, and transmission schemes so that the amount of information transmitted over
a fixed channel bandwidth is optimized. In 5G, this is accomplished by using adap-
tive modulations, ranging from Binary Phase-Shift Keying (BPSK) to 256 Quadrature
Amplitude Modulation (256-QAM) [3GP21e], high-efficiency polar and low-density
parity-check (LDPC) codes [3GP21c; Bae+19], and advanced MIMO and beamforming
techniques [Ali+17]. There are, nonetheless, physical limits to the maximum spectral
efficiencies that can be achieved, such as that provided by the Shannon–Hartley the-
orem [Sha49]. As a result, we cannot rely on this strategy alone in order to arbitrarily
increase user data rates.
The second strategy is to allocate additional radio-frequency spectrum, as the chan-
nel bandwidth is directly proportional to the achieved data rates. In 5G, this trans-
lates into the utilization of new frequency bands within the conventional, sub-6 GHz
frequency range (also known as Frequency Range 1, or FR1) [3GP21g], as well as
within millimeter-wave bands above 6 GHz (also known as Frequency Range 2, or
FR2) [3GP21h]. However, although simple from a theoretical point of view, adding
new spectrum entails two major disadvantages. On the one hand, spectrum is a
scarce resource that has to be shared among multiple wireless services, such as satel-

13
2. The Dynamic Functional Split Selection Problem

lite communications, TV broadcasting, government, emergency, military and profes-

sional communications, etc. Consequently, the price of buying new spectrum can be
very high for mobile operators. On the other hand, high-frequency communications
such as those performed in FR2 often exhibit bad propagation characteristics, since
they are heavily attenuated by buildings and other obstacles between the transmis-
sion and the receiver [Mac+13]. As a resut, high-frequency communications are often
constrained to short-range and/or line-of-sight radio links.
The third strategy consists in increasing cell density, that is, deploying more 5G cells
per area unit. This strategy comes out as a consequence of the limitations of the
previous two strategies. On the one hand, for fixed bandwidth and spectral effi-
ciency, the only remaining possibility to increase data rates is to enhance the signal-
to-interference-and-noise ratio (SINR), which can be accomplished by increasing re-
ceived signal power. Although theoretically possible, increasing transmitted signal
power of sparse deployments is usually not the best approach, since it may lead to
shadowy coverage [Mac+13] and even raise health concerns [Lin16]. Hence, the radio
equipment needs to be brought closer to the users so that SINR can be improved with-
out higher transmission power. On the other hand, the bad propagation characteris-
tics of high-frequency signals also require abundance of radio transmission points, so
as to ensure that there are as few obstacles as possible between users and base stations.
However, increasing cell density is not an ideal solution either. Indeed, when trying
to enhance the SINR by bringing cells closer together, we are at risk of worsen the situ-
ation because of the additional inter-cell interference. This problem can be addressed
with the use of interference mitigation techniques. In the next section, we provide a
summarized description of these techniques.

2.2.2. Interference mitigation in 5G networks

We classify the interference mitigation techniques that a 5G RAN may implement

according to how fast they perform. Thus, we distinguish between static, semi-static,
dynamic, and interference-canceling techniques.

Static interference mitigation

The simplest way of reducing inter-cell interference is to statically assign different fre-
quency bands to each neighboring cell, which is called frequency reuse. As a result,
interference can be completely avoided, at the expense of reduced maximum data
rates, since each cell has only access to a piece of the whole available spectrum. Since
this disadvantage may counter the benefits of interference avoidance, more sophis-
ticated techniques such as fractional frequency reuse (FFR) have been proposed. With
FFR, the whole band is used for UEs camping on the inner region of the cell, which
is less affected by interference coming from other cells, whereas those UEs at the cell
edge are served with smaller bands that are chosen not to collide with neighbor trans-

14
 2.2. Centralized 5G RAN architectures

missions [Ham+13]. If the power allocated to these smaller bands is configured with
fine granularity, we refer to it as soft fractional frequency reuse (SFFR), which allows for
a greater control over the interference levels [Ham+13].
The main drawback of these techniques is that the allocation of frequency bands is
fixed, regardless of the instantaneous interference situation. That is, in some cases, a
cell may benefit from using a larger or a smaller piece of the frequency band if there is
less or more inter-cell interference than that expected at network design, respectively
[Bou+09]. Nonetheless, an important advantage of static interference mitigation tech-
niques is that they do not require any actual coordination among the cells.

Semi-static interference mitigation

If cells are able to communicate with each other, they can exchange information about
the channels that they would like to use to serve a UE, or the channels on which their
UEs are experiencing the most interference. Therefore, they can use this information
to try to minimize their mutual interference. Namely, they could send periodic re-
ports indicating their preferred frequency bands and use received reports from other
cells to agree on the bands that every cell should use. This technique, which is used
extensively in 4G networks, is called inter-cell interference coordination (ICIC) [Sor+17].
Although more dynamic than FFR or SFFR, ICIC still bases on relatively infrequent
coordination between neighbor cells. That is, cells still take their own scheduling
decisions, although they are influenced by the interference information sent to and re-
ceived from neighbor cells. Consequently, ICIC achieves better interference reduction
than static techniques [Ham+13], but it is outperformed by faster techniques [Sor+17].

Dynamic interference mitigation

We define dynamic interference mitigation as those techniques which use real-time in-
formation from a set of cells in order to allocate the best transmission and reception
resources for each individual cell. The most simple such technique is called coordi-
nated link adaptation, which consists in taking independent time-frequency allocation
decisions during scheduling and sharing them with neighbor cells, so that the mod-
ulation and coding schemes are adapted preemptively according to the predicted in-
terferences [MA+18].
A more sophisticated approach is coordinated scheduling, in which scheduling decisions
are taken jointly by all neighbor cells, either by distributed or centralized agreement
[Nar+18]. In either case, fast communication between the cells is strictly required,
since the scheduling interval in 5G networks ranges from 62.5 µs to 1 ms [3GP21d].
Similarly to coordinated scheduling, coordinated beamforming relies on joint agreement
to decide on which beams to use to serve UEs, in order to minimize spatial interfer-
ence [Che+16b].
These dynamic interference-mitigating techniques can be more effective than static or

15
2. The Dynamic Functional Split Selection Problem

semi-static ones, since they are able to deal with the instantaneous interference situ-
ation. Nonetheless, they cannot completely prevent interference, specially when the
RAN is congested. In those cases, there may not be enough non-overlapping resources
to schedule the users, therefore interference would still be present [Nar+18].

Interference cancellation

The most sophisticated technique to mitigate interference is interference cancellation,

that is, the removal of interference in the sheer radio or base-band signal. This can
be performed in two different ways. For the downlink, if a cell knows the interfering
signal coming from a neighbor cell before it is transmitted, it can subtract it in advance
to its next transmission so that the received signal at the UE resembles the originally
intended signal without interference. This technique is known as joint transmission1
[Jun+14; Dav+13]. For the uplink, cells may share their received signals and compare
them against one another in order to extract the individual contribution of interfering
UEs. This technique is known as joint reception [HP17].
Interference cancellation techniques may be, in theory, the most effective interference
mitigation techniques, specially for congested networks. Nonetheless, they require
very low-latency and high-throughput communication between cooperating base sta-
tions, otherwise their performance in practice may be poor [NMH14]. As a result, cen-
tralized architectures are a requirement for the implementation of these techniques,
such as those presented in the next section.

2.2.3. The Cloud-RAN architecture

From the previous description of interference techniques, it follows that advanced

interference mitigation is only possible when the processing functions of all coordi-
nating base stations can communicate fast with one another. This practically rules out
distributed architectures, since the switching, transmission, and propagation delays
introduced by medium or large networks may prevent the execution of dynamic co-
ordination techniques. As a result, centralized architectures, in which the processing
functions of the RAN are close together are necessary.
The Cloud-RAN architecture [Che+14a], or simply C-RAN, builds upon a simple idea:
all the processing performed by all base stations in the RAN, with the only exception
of analog RF processing, should be virtualized and moved to a single data center.
Virtualization implies that former hardware processing is replaced by software func-
tionality, which is abstracted from the underlying computational platform and there-
fore can be easily deployed, scaled, and modified. As a result, each base station is
divided into two units: a remote radio head (RRH), in charge of filtering, amplifying,
and transmitting radio signals; and a baseband unit (BBU) which is in charge of the
1
The term joint transmission can also refer to transmitting the same signal from two different cells si-
multaneously. However, in this work we henceforth interpret joint transmission as previously stated.

16
 2.2. Centralized 5G RAN architectures

RRH
BBU

(a) Distributed RAN (b) Centralized RAN (C-RAN)

RU
CU
DU

(c) Partially centralized RAN

Figure 2.1.: Depiction of a simple radio access network featuring a distributed archi-
tecture (a), a centralized architecture (b), and a partially centralized archi-
tecture implementing a CU/DU functional split (c).

remaining processing, including user- and control-plane signals from both high and
low layers. A depiction of the C-RAN architecture is shown in Fig. 2.1b.
The C-RAN architecture features four main advantages to motivate their implemen-
tation in actual 5G RANs [Che+14a]:
• Traffic adaptability and scalability. It is well-known that mobile traffic exhibit
time-varying patterns, owing to the movements and the activity changes of the
users during the day or the week. For instance, the traffic load can be 20 to 50
times higher in the mornings and afternoons than at night in most areas [Xu+16].
In addition, the difference in traffic load among area types (residential, business,
entertainment, etc.) may also be large depending on the time and day of the
week [Xu+16]. If the operator wants to guarantee that the peak traffic load is
satisfied the whole time and in all areas, it needs to deploy base stations with
large processing capabilities, which will nonetheless be underutilized a large
portion of the time. Conversely, if they opt for a more efficient use of the net-
work resources, then they may fail to satisfy user demands in the peak hours.
However, if the processing of all base stations of different area types is pooled
into a single data center, the operator can take advantage of the multiplexing
gain to ensure user satisfaction while requiring less network resources.
• Energy and cost reductions. Owing to the multiplexing gain of combining the
processing of all base stations into a single location, the operator can also reduce
energy consumption. The reason for this is twofold. First, the energy of cooling
a single, big data center can be substantially lower than that of cooling individ-
ual base stations. Second, function virtualization allows to turn off computing
servers if they are not needed. This energy reduction translates into cost savings,
which are furthered by the simpler management and installation of centralized
computing servers with respect to distributed base stations.

17
2. The Dynamic Functional Split Selection Problem

• Simpler maintenance and upgrades. Since the whole operation of all base sta-
tions is centralized, realizing maintenance tasks or upgrades is faster and less
costly than on distributed architectures.
• Support for advanced interference management. As mentioned before, the
C-RAN architecture allows for fast communication between processing func-
tions, which enables the implementation of advanced interference mitigation
techniques, such as the aforementioned coordinated scheduling and beamform-
ing, joint transmission and reception, etc.
Nevertheless, the C-RAN architecture also has important drawbacks [Che+14a]:
• High-capacity, low-latency transport network. Since the only distributed func-
tion in the C-RAN architecture is the final RF processing, quantized samples of
the base-band radio signals need to be sent through the transport network con-
necting the RRHs and the BBU. The capacity required for this is directly propor-
tional to the channel bandwidth, the number of MIMO layers, the quantization
levels, and the number of antenna sectors, hence it can be very large compared
to the actual user throughput. For instance, it is estimated that for a 100 MHz
channel, 32 antenna ports, and 16-bit quantization, the required capacity for a
single antenna sector would be 157.3 Gb/s, even when the actual user data rate
does not exceed 4 Gb/s [3GP20b]. If the channel bandwidth is 200 MHz and the
number of antenna ports is 256, the required capacity would be 2.56 Tb/s per
antenna sector [IT18]. In addition, the transport network must also guarantee
very low latency and jitter between RRHs and BBU to enable not only flawless
operation, but also the implementation of interference cancellation techniques.
Estimated required values for one-way latency and jitter are 250 µs and 0.5 µs,
respectively [3GP20b; Che+14a].
• BBU coordination and clustering. In order to exploit the benefits of function
centralization, dedicated procedures and interfaces to coordinated the central-
ized functions at the BBU need to be designed. The deployment of the BBUs in
a data center must be secure, since any security breach would affect the entire
network, and the connections with the RRH must be reliable. Moreover, the lo-
cation of the data center must be carefully selected so as to minimize power con-
sumption and network latency and ensure user satisfaction. Finally, the BBU has
to be prepared to cope with potentially high processing loads resulting from the
combined operation of multiple base stations, which also depend on the number
of antennas, resource blocks, modulations, and coding schemes [Nik15].
• Virtualization platform. The whole idea of function centralization relies on be-
ing able to replace hardware with software functions, that is, function virtu-
alization. This requires using a software platform to instantiate, manage, and
possibly migrate virtual functions over the physical computing servers. Owing
to the high throughput, high reliability, and low latency that is required to op-
erate each base station, choosing an appropriate virtualization platform can also
be an important challenge.

18
 2.2. Centralized 5G RAN architectures

Although the aforementioned advantages of C-RAN would allow an efficient and

cost-effective management of a 5G RAN, its disadvantages, specially those concerning
the high capacity of transport network between RRHs and BBU, render it infeasible
in many realistic scenarios. Indeed, current fiber-optical deployments of transport
networks connecting remote locations with a data center often feature links not ex-
ceeding 2 Tb/s in capacity [GS+18a], which severely limit the amount of base stations
that can be centralized. In the light of this, the current 3GPP 5G RAN specifications
recommend instead a partially centralized architecture featuring a functional split in the
processing function chain of each base station.

2.2.4. Functional Split in 5G networks

If full centralization is not possible, the next best approach is to centralize as many
network functions as possible. Namely, for each base station (gNodeB) a subset of
functions is deployed at the centralized unit (CU), whereas the remaining functions
remain at the distributed unit (DU), which is located close to the remote unit (RU)
containing the radio equipment. The CU can be a commodity data center, which hosts
the centralized functions of multiple gNodeBs. Moreover, the creation of separate
centralized and distributed units leads to the definition of a fronthaul2 network, which
is in charge of carrying user data and all the signaling traffic between centralized and
distributed functions.
In theory, we could arbitrarily choose how functions are defined within a base station.
In practice, however, functions are usually defined based on the 5G protocol stack
[Döt+13], which clearly states a sequence of functional layers. From top to bottom,
these layers are:
• Radio Resource Control (RRC): This layer terminates the control-plane mes-
sages between the base station and the UEs.
• Service Data Adaptation Protocol (SDAP): The SDAP layer is in charge of the
mapping between a quality-of-service (QoS) flow and data radio bearers.
• Packet Data Convergence Protocol (PDCP): This layer performs header com-
pression and implements ciphering and integrity protection.
• Radio Link Control (RLC): The main goal of this layer is to ensure reliable and
in-sequence delivery of data streams in downlink and uplink.
• Medium Access Control (MAC): This layer is in charge of scheduling transmis-
sions, prioritizing and multiplexing logical channels, and controlling the hybrid
automatic repeat request (HARQ) function.
• Physical layer (PHY): The physical layer comprises the low-level operations
2
Other authors use the term fronthaul network to refer to the network connecting DUs and RUs,
whereas that connecting DUs and CU is called midhaul network. Since we do not make a functional
distinction between DUs and RUs, we prefer the term fronthaul network as opposed to the backhaul
network connecting CUs with the core network.

19
2. The Dynamic Functional Split Selection Problem

High Low High Low High Low

RRC PDCP RF
RLC RLC MAC MAC PHY PHY

Data

Option 1 Option 2 Option 3 Option 4 Option 5 Option 6 Option 7 Option 8

High Low High Low High Low

RRC PDCP RF
RLC RLC MAC MAC PHY PHY

Data

Figure 2.2.: Depiction of the possible functional split options in a 5G network. The
upper processing chain belongs to the downlinks, whereas the lower pro-
cessing chain represents the uplink. The SDAP layer is not depicted, but
merged with the PDCP layer.

such as coding, modulation, IFFT/FFT, etc.

Since these functional layers are arranged as a function chain, when deciding which
functions to centralize it is sensible to simply perform a “cut” and divide the chain
into two subchains so that the interfaces between the functions are not affected. Thus,
the subchain containing the topmost layers is centralized, whereas the other subchain
is distributed. This “cut” is known as the 5G functional split. The resulting division
of functions is known as the centralization level, which is said to be higher the more
functions are centralized.
Deciding on the most appropriate functional split is not a simple task. The higher
the centralization level, the more opportunities there are regarding interference miti-
gation and the lower the operating and capital cost, but also the higher the required
capacity at the fronthaul. According to [3GP17], we identify eight main options to im-
plement a functional split, which are described in the following. A depiction of these
options is shown in Fig. 2.2.
• Option 1 (RRC–PDCP or RRC–SDAP split). With this split, only the RRC
function, belonging to the control plane of the access stratum, is centralized,
whereas the whole user plane is distributed. This allows for a clear control-
and user-plane separation, which may be exploited for edge computing appli-
cations where centralized control is beneficial and low latency is required in the
user plane. Nonetheless, the very limited centralization level of this split does
not allow for advanced interference mitigation techniques nor substantial cost
reductions.
• Option 2 (PDCP–RLC split). In this split option, both the control-plane RRC
function and the user-plane PDCP function are centralized, whereas the remain-
ing functions are deployed in the distributed unit. The main benefit of this split
with respect to the previous one is that the user traffic aggregation is done in
the centralized unit, thus enabling easier management of the traffic load. From a
standardization perspective, the implementation of this split is straightforward,

20
 2.2. Centralized 5G RAN architectures

as it is a direct continuation of the interfaces defined for LTE Dual Connectiv-

ity [3GP21i]. In fact, this split is already included in the specified architecture
for the next-generation RAN (NG-RAN) as the default option for the functional
split between CU and DU [3GP21a]. Although it features a higher level of cen-
tralization than the RRC-PDCP split, it is still not enough to enable advanced
interference mitigation techniques.
• Option 3 (Intra RLC split). This split option can be implemented in two differ-
ent ways, depending on how the high and low sublayers of RLC are defined.
Namely, the low RLC sublayer can either refer to the segmentation function or
the downlink RLC transmission entities, whereas the high RLC sublayer com-
prises the remaining RLC functions (automatic repeat request, uplink functions,
etc.). In either case, this split benefits from higher flow control and leverages
better resource pooling than previous splits without being as sensitive to latency
constraints as more centralized options. However, a new interface would need
to be defined so as to accommodate all the information exchange between the
two sublayers.
• Option 4 (RLC–MAC split). In this option, the whole RLC layer is centralized
alongside the RRC and PDCP layers, whereas the MAC, PHY, and RF layers are
distributed. Although the interface between the RLC and MAC layers is already
precisely defined, thus allowing for simple implementation, there are no special
benefits of using the RLC–MAC split over the Intra RLC split, apart from a small
multiplexing gain due to the higher centralization level.
• Option 5 (Intra MAC split). The MAC layer can be divided into a high MAC
layer, comprising mainly the scheduler function, and a low MAC layer, which
includes the hybrid automatic repeat request (HARQ) function, random access
control, and channel measurements, among others. If the high MAC layer is de-
ployed at the CU, the RAN can benefit from enhanced interference coordination
and coordinated scheduling or beamforming among all base stations, leading
to a potential reduction in the interference levels. Nonetheless, the interface be-
tween high and low MAC layers may be complex to define and implement.
• Option 6 (MAC–PHY split). In this option, the whole MAC layer is centralized,
along with the RRC, PDCP, and RLC layers, whereas only the PHY and RF lay-
ers are distributed. Since the MAC layer is centralized, it also enables the use
of coordinated scheduling and beamforming, and even simple forms of coordi-
nated transmission. This comes at the price of increased traffic at the fronthaul
network [MAGVK19b], since the communication between MAC and PHY layers
involves the exchange of a fair amount of signaling commands. The complexity
of the interface between MAC and PHY layers is, nonetheless, lower than in the
previous option. Indeed, there already exist proposed descriptions of it, such as
the 5G nFAPI interface [SCF21].
• Option 7 (Intra PHY split). This split also relies on dividing the physical layer
into two sublayers, which can be performed in multiple ways. For example,

21
2. The Dynamic Functional Split Selection Problem

the IFFT/FFT and cyclic-prefix removal/addition could be distributed, whereas

the encoder and modulator are centralized. In general, since the PHY layer is
computationally-intensive and deals with low-level signals, the main advantage
of implementing an Intra PHY split is to leverage cost reductions by resource
pooling and enable advanced interference-mitigation techniques. The obvious
drawback is that the more functions are centralized, the more capacity is re-
quired at the fronthaul network.
• Option 8 (PHY–RF split or C-RAN). This last option is equivalent to the original
idea of C-RAN: all functions except those dealing with analog RF processing are
hosted at the centralized unit. Although, as mentioned before, implementing
C-RAN may not be feasible in current 5G networks, there are already interface
specifications between the PHY and RF layers that can be used as a reference for
future implementations, such as the eCPRI interface [EAN19]. In this section, we
describe the considered network and present the concepts required to formulate
the functional split selection problem (FSSP).
From the above description of the layers, it is clear that the more centralized functions,
the more advanced the interference mitigation techniques that can be applied, but
also the higher the capacity required on the fronthaul network. In other words, if the
network operator wants to maximize the throughput of the users, it needs to centralize
as many functions as possible, but this is constrained by the capacity of the fronthaul
network. This trade-off defines the objective and constraints of the functional split
selection problem (FSSP).

2.3. Functional split selection problem

In this section, we introduce the basic notation and modeling assumptions that we
use in the formulation of the FSSP. The purpose of formulating the FSSP is to find the
optimal functional split option for every gNodeB in the network such that the overall
user experience and/or network revenue is maximized. Solving the FSSP, however,
entails a series of challenges, not only due to the limitations imposed by the fronthaul
network, but also because of the everchanging nature of mobile networks.

2.3.1. System modeling

General network description

We consider a 5G RAN consisting of G gNBs, including macro and small cells. The
operation of each cell is divided into a chain of software functions, which is split into
a DU and a CU. The DUs are deployed close to remote units (RU) containing the an-
tennas and radio equipment, whereas the CUs are all located in a single data center,
resulting in G different locations for the DUs and a single CU location. Communi-

22
 2.3. Functional split selection problem

DU with macro cell DU with small cell Fronthaul switch CU

Figure 2.3.: Depiction of an example network with G = 11 gNBs (including macro and
small cells) and eight fronthaul switches.

cation between CUs and DUs is accomplished by means of a packet-switched fron-

thaul network [GS+18a], which consists of a set of links and a set of layer-2 or layer-3
switches. We assume that this network is dynamically configurable, which implies
that the operator is able to reroute and divide the end-to-end flows during runtime,
for instance by utilizing software-defined networking (SDN) techniques.
We model the fronthaul network as a directed graph D = hN, Ei, where N is the set of
network nodes (including switches, DUs and CU) and E is the set of network links.
The total number of nodes and edges is denoted by N and E, respectively. The node
corresponding to the CU is referred to as n0 , whereas a DU node is represented by
ng , where g ∈ G and G , {1, ..., G}. We show a simple example of a RAN featuring
G = 11 gNBs and eight fronthaul switches in Fig. 2.3.
The number of simultaneously active user equipments (UEs) inside the coverage area
of all cells is denoted by U . Each UE u ∈ U, where U , {1, ..., U }, is connected to a
serving gNodeB, whose index is referred to as hu . We consider a densely-deployed
RAN in which all gNodeBs operate in the same frequency bands. Although a more
complicated frequency reuse strategy is not precluded, we intend to highlight the
feasibility and performance of a network featuring dynamically-enabled interference
mitigation techniques. We mainly address downlink communication throughout this
work, but an extension of the analysis to include the uplink is straightforward.

Modeling of functional splits

The operation of each gNB can be represented as a software function chain, where
functions are usually defined with each of the layers in the RAN protocol stack. These

23
2. The Dynamic Functional Split Selection Problem

CU
DU

CU PDCP-RLC RLC-MAC MAC-PHY C-RAN DU

PDCP RLC MAC PHY RF

Less fronthaul capacity required Interference management possible

Figure 2.4.: Representation of the considered network, example gNB functions, and
scheme functional splits.

functions can be hosted at either the CU or DU, as long as the overall chain structure is
kept. The number of possible functional split options, also referred to as centralization
levels, is denoted by M . For example, in Fig. 2.4 the network operator can choose be-
tween M = 4 centralization levels (PDCP-RLC, RLC-MAC, MAC-PHY, and C-RAN)
corresponding to five different functions, which are also the layers in the gNB pro-
tocol stack. The centralization level implemented by a gNodeB g at any given time
is denoted by ag , where ag ∈ M and M , {0, ..., M − 1}. For convenience, we say
that ag = 0 represents the lowest centralization level, i. e., the functional split option
with the lowest amount of centralized functions. Conversely, ag = M − 1 denotes the
highest centralization level, i. e., the functional split with the most centralized func-
tions. For instance, in Fig. 2.4, ag = 0 corresponds to the PDCP-RLC split, whereas
ag = 3 corresponds to the C-RAN split. We define the centralization vector as the vec-
tor containing the centralization levels of all gNBs at any given time, and denote it as
a , [a1 · · · aG ].
The motivation for choosing a one centralization level over another is twofold. On
the one hand, low centralization levels require less fronthaul capacity than high cen-
tralization levels, thus they can be implemented more easily without congesting the
fronthaul network. On the other hand, high centralization levels enable faster and
simpler communication between gNB functions, which can be exploited to enhance
transmission and reception coordination and thus to reduce their mutual interference.
However, this comes at the price of increased requirements regarding fronthaul capac-
ity [3GP17]. This is trade-off is also depicted in Fig. 2.4.

Modeling of the fronthaul network

We model the capacity required by centralization level ag implemented by gNB g by

means of the function ν(ag ). The output of this function ranges from a few Gb/s for
low centralization levels (such as the PDCP-RLC split) to hundreds of Gb/s for high
centralization levels, as previous research has shown [Döt+13; MAGVK19b; 3GP17].
For simplicity, we consider that all gNBs feature the same maximum user data rate,
therefore ν(ag ) does not depend directly on g.

24
 2.3. Functional split selection problem

We define ϑe as the capacity of each fronthaul link e ∈ E. For a downlink flow gener-
ated by gNB g between its DU and CU, let feg be the fraction of this flow that conveyed
g g g
by link e. We  flows-per-link vector as f , [f1 · · · fE ] ∀g ∈ G and the flows
 1defineGthe
vector f , f · · · f .

Modeling of the network state

We often consider the flows vector f and the centralization vector a together using the
following notation:
s , ha, f i . (2.1)
We refer to s as the state vector of the network, since it includes all variables that the
network operator is able to change to deal with changes in the environment. In ad-
dition, we denote the optimal state, centralization, and flow vectors by s∗ , a∗ , and f ∗ ,
respectively, such that:
s∗ , ha∗ , f ∗ i . (2.2)

Modeling of the interference mitigation

The activity of neighboring cells causes downlink interference on nearby UEs, hence
reducing user data rates. Proposed techniques to mitigate this interference (such as
coordinated scheduling, coordinated beamforming, joint transmission, etc.) require
some level of coordination between the involved gNB functions. As result, each
interference-mitigation technique requires a minimum centralization level to be ap-
plied, depending on which functions need to be centralized for the technique to op-
erate properly. For instance, coordinated scheduling requires the centralization of the
MAC layer [Nar+18], whereas joint transmission also requires the centralization of the
physical layer [Zha+17].
Based on the analysis shown in [PM17], we model the effectiveness of an interference
mitigation technique between two gNBs as a constant factor multiplying their average
received interference power. We relate each centralization level a with the interference
cancellation factor of the most effective interference mitigation technique that can be
applied by means of function И(a). The codomain of function И(a) is [0, 1], that is, it
ranges from 0 (full interference cancellation) to 1 (no interference cancellation).
Since centralization levels are defined incrementally along the function chain, the
higher the centralization level a, the lower its related interference-cancellation fac-
tor И(a). Moreover, an interference-mitigation technique can only be used by two
gNBs if both of them are operating at the required centralization level or higher. As
a consequence, the resulting interference-cancellation factor between gNBs g and g 0
is И(min(ag , ag0 )), that is, the gNB with the lowest centralization level is the bottle-
neck to interference mitigation. Knowing this fact, we can compute the expected total

25
2. The Dynamic Functional Split Selection Problem

interference power Iu experienced by UE u from all gNBs as:

G
X
Iu (a) = iu,g · И(min(ahu , ag )), (2.3)
g=1

where iu,g is the interference power received by UE u from gNB g and iu,hu , 0, as the
UE is not interfered by its serving gNB. Note that Iu is a function of the centralization
vector a, hence selecting the right values of a can be used to reduce overall interfer-
ence. For all UEs in U, we define the interference vector I(a) , [I1 (a) · · · IU (a)].

2.3.2. Static FSSP formulation

The objective of the network operator when selecting the functional split is to max-
imize user data rates, minimize operating costs, or optimize a combination of both.
We denote a generic objective function as Γ(a, I(a), f ), where a is the centralization
vector, I(a) is the interference vector, and f is the flow vector. Without loss of gener-
ality, we assume that our goal is to maximize the value of Γ(a, I(a), f ), subject to the
fronthaul network constraints. Consequently, we can formulate the generic functional
split selection problem (FSSP) as:

max Γ(a, I(a), f ), (P1a)

s

subject to

X X 0
 ∀n ∈ N \ {n0 , ng }
feg − g
fe = ν(ag ) for n = n0 ∀g ∈ G, (P1b)
−

e∈E+ (n) e∈E (n) 
−ν(ag ) for n = ng
G
X
feg ≤ ϑe ∀e ∈ E, (P1c)
g=1

feg ≥ 0 ∀e ∈ E, ∀g ∈ G, (P1d)
G
s∈M . (P1e)

where E− (n) denotes the set of edges entering node n and E+ (n) is the set of edges
leaving node n. Constraint (P1b) is the flow conservation constraint, since it guarantees
that flow is created at the CU, consumed at the DUs, and conserved in the interme-
diate nodes. Constraint (P1c) is the link capacity constraint, as it ensures that the link
capacities are never exceeded. Solving (P1) allows us to obtain the optimal state vec-
tor s∗ , and thus the optimal centralization vector a∗ and flow vector f ∗ , for a given
objective function Γ(a, I(a), f ).
Since the objective function Γ(a, I(a), f ) depends on the interference vector I(a), every
time this vector changes the FSSP (P1) will also change. This means that solving (P1)

26
 2.4. Solving the FSSP dynamically

yields an optimal vector state that is only instantaneously valid. However, finding the
optimal state for a given instant is not sufficient to operate a 5G RAN efficiently. This
is because changes in the positions and activity of the UEs, as well as variations in the
wireless channel, transform the problem over time, so that an optimal solution at the
present time may be highly suboptimal in the future. In the next section, we introduce
the concepts required to tackle the dynamic FSSP problem.

2.4. Solving the FSSP dynamically

In order to operate profitably, the RAN has to be able to adapt to changes in the en-
vironment, such as changes in the interference vector I(a). This can be realized by
leveraging recent advances in network softwarization, which allow for fast and cost-
efficient reconfiguration of communication networks. In fact, these new possibilities
have spurred the analysis and modeling of the ability of communication networks to
adapt to environmental changes, which is known as network flexibility.
In this section, we introduce the sofwarization technologies that enable dynamic
changes in communication networks, and thus can be applied to a 5G RAN featur-
ing a dynamically-selected functional split. Then, we discuss the new entities that are
needed in such a RAN and their implications. Finally, we present the formulation for
a dynamic FSSP.

2.4.1. Flexibility in softwarized communication networks

Network softwarization technologies

In the past, changing the network configuration dynamically was difficult to realize,
and it was instead preferred to design the RAN to optimize for the average load or
to support the peak demand [Li11]. Nowadays, recent advances in network soft-
warization allow us to replace stiff hardware equipment, which is hard to operate
dynamically, with software functions. These functions can be reconfigured, relocated,
and scaled with ease, which enables prompt and cost-efficient adaptation to changes
in the network environment and demands. Namely, there are three main novel tech-
nologies that offer network softwarization:
• Software defined networking (SDN) can be described as designing, deploy-
ing, and operating communication networks such that the forwarding rules in
switches and routers are programmed on a central server [Sta15]. Thus, switch-
ing devices are not configured individually, but their operation is defined by
their interaction with a central server, which can be reconfigured at will. In prac-
tice, this results in the separation of the control plane, consisting in the configu-
ration of forwarding decisions, from the user plane, dealing with user packet for-
warding. The main benefits of SDN are better adaptability, automation, mobility,

27
2. The Dynamic Functional Split Selection Problem

maintainability, scaling, and security than traditional, distributed approaches

for networks facing ever-increasing demands, such as 5G radio access networks
[Ope14; BK16].
• Network functions virtualization (NFV) is defined as the replacement of net-
work hardware functions by software functions with the intention of running
them in virtual environments, such as containers or virtual machines [Sta15]. Of-
ten in combination with SDN, it can be used to substitute proprietary hardware
platforms by commodity platforms, increase the efficiency of function manage-
ment, allow for function relocation, and enhance network scaling and maintain-
ability [Sta15].
• Network virtualization (NV): This technology consists in creating logically-
isolated network slices over a shared physical network, with the goal of enabling
the simultaneous use of virtual networks by multiple tenants. Physical network
resources can be aggregated or split into single or multiple virtual resources, re-
spectively, so that the virtual network as seen by the tenant is abstracted from the
underlying network [IT12]. The main advantages of NV are increased flexibility,
reduced operational and capital costs, and enhanced scalability [Sta15].
We conclude from the definitions of SDN, NFV, and NV that they can be applied to
make the network more adaptable to changes in the demands, that is, more flexible.
Indeed, these technologies motivate a novel interest in modeling network flexibility,
which is presented in the following.

Definition of network flexibility

In [He+19; Bab+20], network flexibility is defined as the timely and cost-efficient support
of changes in the network requirements. The concept of network requirements refers to any
environmental or user-related demand that affects network operation or profitability
but cannot be directly changed by the network operator. As a result, every time there
is a change in the network requirements, which is henceforth referred to as a request,
the current network configuration may be suboptimal, and thus adaptation may be
required.
There are two different manners in which a softwarized network may satisfy requests.
First, the network could be designed and deployed with enough resource overhead
so that it accommodates requests without reconfiguring itself. In our case, this would
imply an over-dimensioned 5G RAN whose fronthaul network supports full central-
ization. The main drawback of this approach is, however, the potentially high cost
of deploying and operating such a network. Second, the network may actively adapt
its topology, flows, functions, or resources to match the new network requirements.
This may reduce the deployment and operating costs with respect to the first option,
although timely adaptation is now of utmost importance. In this work, we investi-
gate whether this second approach is applicable to 5G RANs and whether it is more
adequate than non-adaptive implementations.

28
 2.4. Solving the FSSP dynamically

Table 2.1.: Technical Concepts and their support of flexibility in networks. (X: main
target). This table is extracted from [He+19].
Category Aspect SDN NFV NV

Flow Configuration: flow steering X - -

Function Configuration: function program- - X -
Adapt configuration ming
Parameter Configuration: change function pa- - X X
rameters
Function Placement: distribution, placement, - X X
Locate functions
chaining
Resource and Function Scaling: processing and X X X
storage capacity, number of functions
Scale
Topology Adaptation: (virtual) network adap- - - X
tation

Flexibility categories and aspects

According to [He+19], a softwarized network can satisfy a request in three different

manners: (i) by reconfiguring its current flows, functions, and parameters; (ii) by re-
locating software functions; and (iii) by scaling resource and functions or modifying
its (virtual) topology. These are known as flexibility categories, since they can be used
to classify a flexible network. Within each category, we identify one or more flexibil-
ity aspects, which narrow down the specific network feature that is changed during an
adaptation. We can use flexibility categories and aspects to identify the softwarization
technology that is required to implement a flexible network. This is because we can
relate each flexibility aspect with the softwarization technology that enables it, as it is
shown in Table 2.1
In order to implement a dynamic functional split, the 5G RAN needs to feature at least
three flexibility aspects. First, the RAN needs to be able to move virtual functions in
and out CU and the DUs so as to operate in the selected functional split. This can be
identified with the function placement aspect. Second, the RAN has to scale its comput-
ing resources according to the instantaneous load and also the current functional split.
This is the resource and function scaling aspect. Finally, the flows within the fronthaul
network may be rerouted every time the functional split changes, since the capacity
required between CU and DUs is different for each functional split. Thus, the flow
configuration aspect is also required. As a result, we see from Table 2.1 that a 5G RAN
featuring a dynamically-adapting functional split requires an SDN fronthaul network

29
2. The Dynamic Functional Split Selection Problem

2.4.2. Flexible functional split platform

Owing to both the advantages and disadvantages of function centralization in the

RAN, it is clear that being able to change the centralization level in a dynamic manner,
according to the instantaneous state of the network, may be beneficial for increasing
network performance and reducing operating costs. Nevertheless, implementing a
dynamically-configurable functional split is not a trivial task. Indeed, we need to
define two new network entities: (i) a decision-making entity, which is in charge of
monitoring the network and deciding when to change the centralization level, and (ii)
a migration platform, which realizes changes in the functional split without stopping
the operation of the network.

Decision-making entity

The decision-making entity is in charge of implementing a strategy to tackle the dy-

namic optimization problem of selecting the best functional split. There are abundant
previous works regarding dynamic optimization techniques, such as those provided
by the field of control theory. This field comprises the study and control of dynamic
systems, whose evolution is commonly modeled by differential equations. As classi-
cal demonstration of this type of dynamic control problem is the control of an inverted
pendulum [FYK91], where the goal is to keep a pendulum in the upright position
by controlling a moving base. The theoretical evolution of the movement and posi-
tion of the pendulum is modeled by well-known equations, which are combined with
measurements of the actions and resulting errors to swing up the pendulum. This
approach is, nonetheless, difficult to apply to our problem, since the evolution of a
mobile network depends on the random roaming of hundreds or even thousands of
UEs, which renders a theoretical description of its evolution intractable. However, we
may still apply discrete dynamic programming to our problem [Bel66], as long as the
future evolution of the system is known or can be accurately predicted. This possibil-
ity is discussed in Chapter 6. In any case, predicting the behavior of a large number
of UEs is a very challenging task. Furthermore, if the employed dynamic optimiza-
tion technique relies on mobility predictions, it would be arguable whether they are
accurate enough for any conclusion to hold in real networks. Consequently, in this
work we do not assume that the network operator has access to predictions about the
behavior of the UEs beyond simple statistics, and thus we use the results provided by
dynamic programming techniques only as references for other strategies.
An alternative approach is to use evolutionary dynamic optimization techniques
[NYB12]. Evolutionary algorithms, such as the genetic algorithm, are often adequate
to tackle continuously changing problems, since they can be configured to follow
changes in the environment instead of converging. Moreover, previous work uses a
genetic algorithm to successfully tackle a simplified formulation of the FSSP [MAK19a].
In this work, however, we address a more complete FSSP formulation that is not as
suitable for evolutionary algorithms, due to the stringent constraints that define our
solution set. The reason is that the continuous-valued flows vector f and the discrete-

30
 2.4. Solving the FSSP dynamically

valued centralization vector a have to be always in agreement and satisfying the net-
work constraints, rendering a random exploration of the solution set inefficient.
As a consequence, in this work we use simple, yet effective approach for the decision-
making entity. First, the decision-making entity periodically monitors the network
and solves the static FSSP for each new state, so as to obtain an updated network con-
figuration that would maximize the revenue of the network. Note that, since the FSSP
is NP-Hard [MAJK21], we cannot tell whether the network still features an optimal
configuration after a change in the problem unless we run the optimization prob-
lem again. After a new optimal configuration is known, the decision-making entity
chooses whether to perform an adaptation and move to the optimal configuration or
stay in the current state. This decision is taken based on a set of rules that are influ-
enced by the cost of operating in the current state, the cost of changing to the optimal
state, and the cost of operating at the optimal state. We refer to this set of rules as
the adaptation strategy, which are discussed in Sec. 6.5. The design and selection of the
adaptation strategies are supported by the application of the so-called cost-of-flexibility
framework [Bab+20; MA+21], which is used to separately model all cost components,
decide about the optimal monitoring period, and compare different adaptation strate-
gies.

Migration platform

Regarding the migration platform, the network operator needs to be able to change
the functional split without stopping the network operation. There are already NFV
platforms in the state of the art featuring such uninterrupted operation. In Chapter 5,
we present a framework that can change the functional split of a single base station in
less than 20 ms with no packet losses. Alternatively, the operator may employ virtual
machines or containers to host the RAN functions, since off-the-shelf frameworks also
exist to dynamically migrate between physical hosts [GA18].

2.4.3. Dynamic FSSP formulation

Assuming that the 5G RAN implements an SDN- and NFV-based platform that is able
to dynamically change the functional split and flow configuration, we still need to
decide when to trigger this change, and which state to move to. From the perspective
of a mobile network operator, the objective of performing this dynamic adaptation is
to maximize the network revenue, or, equivalently, to minimize the total cost. Clearly,
this total cost is a function of the performance experienced by the users, the operating
cost of running the network on a fixed state, and the cost of changing the functional
split. A thorough analysis of all these components is addressed in the next chapter.
Let us define τ as the time-ordered discrete index representing the time instant corre-
sponding to the end of a monitoring interval. We can use this index to make the time
dependence explicit on the state, centralization, and flow vectors as s(τ ), a(τ ), f (τ ), re-

31
2. The Dynamic Functional Split Selection Problem

spectively. In addition, we define C(τ ) as the cost associated with changing the state
at time τ , if applicable, and K(s) as the instantaneous cost associated with operating
on state s. Note that K(s) can be defined similarly to Γ(a, I(a), f ) if the latter is already
a cost function. Finally, the objective of the dynamic functional split selection problem
is to find the optimal sequence of states s(τ ) so that the total RAN cost is maximized
without compromising the fronthaul network:
τmax
X  
min lim
max
K(s(τ )) − C(τ ) , (P3a)
s(τ ) τ →∞
τ =0

subject to

X X 0
 ∀n ∈ N \ {n0 , ng }
g g
fe (τ ) − fe (τ ) = ν(ag (τ )) for n = n0 ∀g ∈ G, (P3b)

e∈E+ (n) e∈E− (n) 
−ν(ag (τ )) for n = ng
G
X
feg (τ ) ≤ ϑe ∀e ∈ E, (P3c)
g=1

feg (τ ) ≥0 ∀e ∈ E, ∀g ∈ G, (P3d)
s(τ ) ∈ MG . (P3e)

Compared to the static FSSP (P1), (P3) may be considerably more challenging to ad-
dress, especially if the future evolution of the time-varying parameters cannot be accu-
rately estimated. However, owing to its similarities, the same techniques that are ap-
plied to solve (P1) can be used in combination with dynamic optimization approaches
for tackling (P3). This is fully addressed in Chapter 6.

2.5. Summary
In Chapter 2, we introduce the problems of optimally and dynamically selecting the
functional split for a 5G radio access network. We first discuss the objectives and
requirements of 5G networks and delve into the issue of managing inter-cell interfer-
ence, which may hinder the viability of cell densification. We explore the different
RAN architectures that are proposed to cope with this issue and discuss their details,
advantages, and disadvantages. We conclude that, whereas a fully centralized archi-
tecture is hardly feasible for current networks, partially centralized architectures can
be leveraged to both improve user experience (by reducing interference) and limit
operating costs.
Consequently, we propose a model for a 5G RAN featuring a configurable functional
split, including mathematical descriptions of the overall state, the underlying fron-
thaul network, and the interference-mitigation capabilities. We then use this model to
formulate the static functional split selection problem (FSSP), which aims at finding

32
 2.5. Summary

the optimal functional splits and flow configurations that maximize a generic objec-
tive function. In following chapters, this generic function is replaced with concrete
definitions reflecting the user data rates, operating cost, and network revenue.
Finally, owing to the variable nature of mobile networks, we discuss the challenges
of solving the FSSP dynamically. We present recent advances in the field of network
softwarization and flexibility, which may help us to address the dynamic FSSP. After
discussing possible options for implementation options, we conclude by proposing a
generic formulation of the dynamic FSSP, which addressed in later chapters.

33
3. Modeling the Cost of Flexible
Communication Networks

3.1. Introduction

3.1.1. Motivation, scope, and challenges

In the last chapter, we introduce the problem of dynamically selecting the optimal
functional split for a 5G RAN. This problem features two main layers of complexity
from a theoretical perspective. On the one hand, we need an efficient algorithm to
find an optimal solution for an instantaneous situation, which is not a trivial task. On
the other hand, we have to devise a strategy to dynamically decide when and how to
reconfigure our network. In other words, our network has to be adaptable enough to
support live reconfiguration.
It is indeed a common trend in the design of 5G radio access networks to aim at swift
adaptation to traffic patterns and user mobility, in order to achieve high resource ef-
ficiency [NTT20; Tal+20]. Nonetheless, mobile RANs are not the only example of
communication networks requiring higher flexibility and adaptability. In fact, this
field has attracted a substantial amount of research effort in the recent past coming
from different angles. This is mainly due to the ever-increasing number of connected
devices as well as the emergence of new applications and use cases that many types
of networks face. One example is the emergence of the Internet of Things (IoT), where
the number of connected autonomous devices is estimated to grow from 12 billion
in 2015 to hundreds of billions by 2025 [Cis17]. These IoT-based networks will cover
a wide range of use cases, such as smart-city sensors, self-driving cars and platoon-
ing, industrial automation, etc [Che+14b]. Furthermore, increasing agility and pro-
grammability is also a major concern in data center networks, so as to efficiently deal
with new applications [Cas18].
In order to increase their adaptability, virtually every type of communication net-
works can benefit from network softwarization techniques, namely, from SDN, NFV,
and NV [Afo+18]. Exploiting the advantages of network softwarization is, nonethe-
less, still a greenfield in many regards. Indeed, owing to the novelty of the ability to
quickly adapt the network to changes in the demands, even dedicated research has
emerged to explore this ability. For example, in [Kel+18] the authors define a network
flexibility measure with the intention of providing quantitative assessment of possible
improvements in the adaptability of the network. The main purpose of this new mea-
sure is to evaluate how costly and quickly a softwarized network can adapt to changes

35
3. Modeling the Cost of Flexible Communication Networks

in the demand. As a result, network designers and operators can use it to compare
the performance and cost of adaptive solutions.
In principle, network flexibility is a desirable feature that network operators would
like to maximize. It not only reflects how good the network responds to changes in
the demand, but it can also be used to advertise the network over competitors and
attract potential users. Moreover, network flexibility may also indicate how likely it is
that the network supports changes in future demand trends. Nonetheless, increasing
network flexibility also comes at a price. As many other performance metrics, flexi-
bility is affected by a trade-off: the higher its value, the higher the revenue, but also
the higher the associated costs. The reason for this is twofold. On the one hand, in-
creasing network flexibility often implies investing in better equipment or consume
more energy resources. On the other hand, high flexibility is correlated with better
performance: faster demand satisfaction, more supported users, less relevance of link
failures, etc. These conflicting trends motivate the formulation of classical network
design problems: maximizing performance for a fixed cost, minimizing cost for fixed
performance, or find out the best cost-performance combination on the Pareto frontier.
As with most engineering systems, the standard approach to model the cost of a com-
munication network is to divide the total cost of ownership (TCO) into capital ex-
penses (CAPEX) and operating expenses (OPEX) [Bet+17]. If the environment within
which the network operates varies slowly, then the OPEX can be estimated from the
expected operating states and their corresponding revenues and resource consump-
tion rates [Hue+08; Zan97]. However, modeling the OPEX of a network dealing with
frequent changes in its environment is considerably more challenging. On the one
hand, estimating the expected operating state may be difficult, since it is the result of
a sequence of changing demands. On the other hand, the adaptation itself may incur
in additional cost, which has to be then included into the OPEX estimation.
Consequently, in order to characterize a fast-adapting communication network, we
require a more powerful model for its operating cost, such that adaptations and lack of
adaptations are taken into account. The intention of this model is to accurately predict
the total operating cost of a network facing changes in the demands whose durations
are in the same order of magnitude as the time required for adapting. That is, the
model should be able to estimate the cost of operating a network in “race conditions”
between environmental changes and adaptations. As a result, this cost model can be
utilized to select the optimal adaptation strategy, test alternatives for profitability in
future scenarios, or identify limiting factors.

3.1.2. Key contributions

The main content and contributions presented in this chapter are based on [Bab+20]
and specially [MA+21]. In summary, these contributions are as follows:
1. We present a cost model for flexible, dynamically-adaptive communication net-
works that is based on a probability theory frameworks and designed to capture

36
 3.2. Related Work

the internal trade-offs affecting cost and flexibility. Thus, it provides a more
elaborated point of view than conventional TCO analysis.
2. This model can be employed to make cost predictions in adaptive networks fac-
ing frequent demand changes.
3. Moreover, this cost model can be used to assist in decisions regarding the net-
work design, such as the selection of the least costly deployment option.
Our main intention when proposing this model is to use it to estimate the cost and
select the best adaptation strategy for a 5G RAN featuring a dynamically-adapted
functional split, which are the main topics of Chapters 4 and 6. Nonetheless, the model
is general and rigorous enough to be used with multiple types of communication
networks, as we show in Sec. 3.5 with two application examples.
The rest of this chapter is organized as follows. In Sec. 3.3 we present the system
model, including a definition of the network flexibility measure. In Sec. 3.4 we in-
troduce the complete cost model. Sec. 3.5 contains two application examples so as to
show how the model can be applied to real networks. Finally, Sec. 3.6 concludes the
chapter.

3.2. Related Work

Owing to the promising features of network softwarization technologies, there is con-
siderable interest in investigating their associated deployment and operating costs.
Moreover, the ability of softwarized networks to dynamically deal with demand changes
also receives dedicated attention. As a result, we classify the relevant related work to
this chapter into three categories. First, we discuss static cost models, which serve to
estimate the cost of deploying or operating a softwarized network without explicitly
including the cost of dynamic reconfiguration. Second, we survey dynamic cost mod-
els in which reconfiguration cost is indeed a main component. Finally, we comment
on flexibility models, which mainly address the evaluation of reconfiguration abilities
along with their associated cost.

3.2.1. Static cost models

As a consequence of their novel architecture and improved features, there are sev-
eral recent works estimating the TCO of softwarized 5G networks. For instance, in
[BNP16] the authors present a detailed cost model for the CAPEX and OPEX required
to deploy and operate, respectively, a 5G radio access and core networks featuring
SDN and NFV. Nonetheless, this model is intended to help in the selection of deploy-
ment alternatives and it does not include dynamic components. Similarly, in [TBK15]
the authors address the problem of selecting the optimal location of virtual network
functions (VNF) within a 5G core network. They describe a cost model with the ob-
jective of minimizing the operating cost, which is estimated from the number of in-

37
3. Modeling the Cost of Flexible Communication Networks

stantiated functions, while also ensuring user satisfaction. Although their approach
can be used dynamically, this paper lacks a model for the cost of VNF reconfigura-
tion. In [SRF15], a detailed model of the deployment cost of a C-RAN heterogeneous
network is presented. The main intention is to use the model to compare the deploy-
ment cost of C-RAN and distributed LTE RANs. Although there are cost components
addressing operating and processing costs, the model is intended to represent a static
RAN, not a dynamically-adapting one. Finally, in [Bou+18] and [BKM20], the au-
thors present cost models of 5G networks featuring Multiple Input Multiple Output
(MIMO), Distributed Antenna System (DAS), Cognitive Radio and a SDN RAN. How-
ever, in spite of the complexity of their models, they are not intended to be used for
guiding dynamic adaptations, but rather to evaluate alternative deployment options.
Regarding other types of networks, in [Bou+15] a simple cost model is derived for the
power consumption, network usage, and license fees of operating an NFV network.
The objective is to utilize this model so as to find the optimal location of virtual Deep
Packet Inspection (DPI) functions. Once again, albeit this problem may be solved at
runtime to trigger dynamic reconfigurations, this aspect is not covered by the cost
model. Finally, in [EPP19], the authors study the trade-off between flexibility and cost
in the field of SDN optical transport networks. They propose a cost model comprising
multiple cost components with the objective of formulating several cost-related opti-
mization problems. Nonetheless, they do not deal directly with dynamic adaptations.

3.2.2. Dynamic cost models

In comparison to static models, literature regarding cost models in which the cost of
reconfiguration is taken into account is less abundant. Nevertheless, there are still rel-
evant works in this field. For example, in [Gha+15], the authors propose an approach
to tackle the problem of elastically placing VNFs within a cloud network, with the
intention of minimizing the operating cost. They model multiple cost components,
including not only the cost of running VNFs and carrying traffic, but also the cost
of reconfiguring the network and migrating VNFs. Nonetheless, it is assumed that
changes in the demands that trigger reconfiguration are sufficiently spaced and that
reconfigurations are fast enough not to influence the operating cost. In [SYCP18], an
adaptive approach to optimize the monitoring and orchestration processes of a Cloud
Management System (CMS) is presented. The authors base on the concept of Qual-
ity of Decisions (QoD) [You+16] to propose a dynamic adaptation of the frequency of
sample points, when relevant performance parameters are monitored, and decisional
points, when the CMS takes management decisions. However, the link between mon-
itoring and orchestration frequencies and actual operating cost is not described in
detail. Finally, in [HW10], the authors describe a strategy to dynamically select be-
tween push and pull updates in the context of cloud management. This is done with
the intention of reducing operating costs, although the actual cost model translating
push and pull updates into actual cost is a simple sketch.

38
 3.3. System Model

3.2.3. Flexibility models

In contrast to static and dynamic cost models, we can also find related work in the lit-
erature linking the ability of a softwarized network to reconfigure itself with the cost
thereof. For instance, the Chair of Communication Networks at the Technical Uni-
versity of Munich proposes to use the concept of network flexibility as a well-defined
metric, as presented in [Kel+18]. This definition takes into account the cost that takes
to adapt the network to changes in the demands. Moreover, in [Bab+20] we extend
this metric and formulate it as a mathematical measure. In addition, multiple types of
costs associated with flexible adaptation are presented. Since flexibility and dynamic
adaptation is an usual feature of many engineering systems beyond communication
networks, we can also find related work in other technological fields. For instance,
in [Lin+13] the authors propose a model of the cost of manufacturing flexibility, for
which they use dedicated flexibility metrics.
Finally, to the best of our knowledge, [MA+21], upon which we base this chapter, is
the first work that addresses in detail the modeling of all cost components in a flexible,
dynamically-adapting softwarized network. This includes not only the operating cost
in stable conditions, but also the cost of reconfiguring the network and the cost of
operating in outdated states.

3.3. System Model

In this section, we introduce the concepts required to derive the cost model of a flexible
network.

3.3.1. Network states and demands

We consider a scenario consisting of a softwarized, configurable communication net-

work managed by a network operator to achieve a profitable purpose, such as providing
connectivity to users, carrying information within a data center, or managing virtual
network slices. The instantaneous configuration of the network is referred to as the
state s ∈ S of the network, where S is the set of all possible states that can be achieved.
For example, the routing tables in the network switches, the location of virtual func-
tions, or the physical resources allocated to a network slice can be used as the state.
The conditions on which the network operates are modeled by the demand d ∈ D,
where D is the set of all possible demands. The demand includes all parameters that
affect the network’s profitability but cannot be modified by the network. These pa-
rameters can describe the external environment (such as the number of connected
users or the requested virtual flows), but they also include any internal configuration
that may change out of the network’s control (such as the topology graph of active
nodes and links in a resilient network).

39
3. Modeling the Cost of Flexible Communication Networks

We say that a demand is satisfied if the network state is able to fulfill the expecta-
tions that this demand generates. For instance, a demand consisting of a flow request
between two network points is satisfied if the intermediate nodes can forward the
packets correctly between these points. Those states satisfying a given demand are
called valid states, whereas any other state is an invalid state. We define the function
f(d) to relate demand d to its set of valid states:

f(d) : D 7→ ℘(S), (1)

where ℘(S) is the power set of S. If f(d) = ∅, we say that the demand d is unsatisfiable.
In a flexible network, demands and states are subject to change over time. From the
definition of demand, it follows that the network cannot accurately predict nor pre-
vent demand changes. We model a sequence of demands within time interval (0, τ max )
as a discrete stochastic process {Dj }j∈Z on the sample space D, where j is an arbitrary
time-ordered integer index. We also define the sequence of states {Sj }j∈Z on the sam-
ple space S. We model the duration of each demand Dj = dj by the random variable
Tj on the sample space R+ , and hence we define the stochastic process {Tj }j∈Z as the
sequence of durations of each demand. Assuming that {Tj } is stationary, it can be de-
scribed by its marginal cumulative distribution function (CDF) FT (t). Processes {Dj }
and {Tj } fully describe the demands over time, as every observed demand Dj = dj is
associated with a duration Tj = tj .
A change in the network state is the result of a conscious network decision, which is
taken to address a demand change. Hence, the sequence of states is determined by
the sequence of demands. For convenience, we introduce the following notation to
represent a demand and a state change, respectively:

e j , hdj , dj+1 i ,
d esj , hsj , sj+1 i . (2)

We consider that an adaptation consists of a demand change de j and its corresponding

state change esj . From a modeling point of view, we associate every demand change
with a state change, although in practice it may happen that sj = sj+1 if there is no
effective state change.

3.3.2. The adaptation process

After a noticing a demand change d e j , a flexible network needs to perform two tasks.
First, it needs to run an adaptation algorithm to find the most appropriate state sj+1 to
satisfy the new demand dj+1 . Formally, we model the outcome of this algorithm by
means of the adaptation function:

χ(d) : D 7→ S, (3)

so that sj+1 = χ(dj+1 ). Since finding this new state may be computationally hard,
we need to account for the time and cost required to do this, as they may impact the

40
 3.3. System Model

Server Clients Node Firewall

Readiness phase Proaction phase Reaction phase Readiness phase

States sequence

Demands sequence Time

Time

Figure 3.1.: Adaptation phases traversed by a flexible network. The network operator
is in charge of providing connectivity between server and clients while
ensuring that all packets go through the firewall, whose location can be
dynamically chosen.

overall profitability. We refer to the former as proaction time zjP and to the latter as
proaction cost cPj . In addition, once the network has found the new state, it has to move
from the old state to the new one. We refer to the time and cost required to change the
state as reaction time zjR and reaction cost cR
j , respectively.

Overall, the time difference between a demand change d e j and its corresponding state
change esj is the action time zj = zjP + zjR . As with the sequences of states and demands,
we define the discrete stochastic process {Zj }j∈Z on the sample space R+ to model the
sequence of action times Zj = zj for every time index i. Assuming stationarity, this
process can be characterized by its marginal CDF FZ (z), i. e., the distribution of the
durations of each adaptation. We define in the same manner processes {ZjP }j∈Z and
{ZjR }j∈Z for the proaction and reaction times, respectively. Similarly, we denote the
action cost as cj = cPj + cR
j , which reflects the total effort of addressing demand change
e j and realizing state change esj , and define the discrete stochastic process {Cj }j∈Z on
d
the sample space R+ to model action times Cj = cj , whose marginal CDF is FC (c).
If at any time instant τ the network is satisfying the current demand, we say that the
network is in the readiness phase. The cost per time unit associated with operating the
system at this phase is referred to as the readiness cost kj for an arbitrary time index j,
which can be expressed as a function of the active demand and state. This is explained
in detail in Sec. 3.4.2. The readiness cost is affected by the amount of resources con-
sumed in the current state and the revenue obtained from demand satisfaction. As a
result, not being able to satisfy a demand mainly affects this cost component.
In order to clarify the meaning of the aforementioned definitions, we present an exem-
plary adaptation timeline in Fig. 3.1. This figure shows the observed demands and the
states implemented by a network providing connectivity between a server and a set

41
3. Modeling the Cost of Flexible Communication Networks

of clients. This connectivity is provided while enforcing a security policy: any packet
between the server and the clients must go through a virtual firewall, whose location
can be changed during runtime. The optimal firewall location is the one that mini-
mizes the number of links traversed by all packets, thus minimizing latency. Between
τ3 and τ4 , the network is operating in the readiness phase: demand d0 (clients blue
and green) is being satisfied by state s0 with minimal latency. The revenue obtained
from satisfying this demand and the cost of using network resources (links, nodes,
CPU, etc.) is reflected by the readiness cost. At time τ4 , the demand changes to d1 : the
red client connects to the network. After noticing the demand change, the network
realizes that firewall location may not be optimal anymore. As a result, it triggers the
adaptation algorithm to find out the optimal state for demand d1 . The time during
which the adaptation algorithm is running is the proaction phase, and the additional
cost associated to it (due to higher resource consumption) is the proaction cost. At time
τ5 , the adaptation algorithm has converged and returned a state s1 6= s0 featuring a
new firewall location. Therefore, the network starts the procedure to migrate the fire-
wall, hence starting the reaction phase, which lasts until the migration is completed at
τ6 . Any additional cost associated to this phase is reflected by the reaction cost. The
union of the proaction and reaction phases is the action phase, during which the active
state is delayed with respect to the current demand.

3.3.3. Flexibility measure

As mentioned in Sec. 3.2.3, the ability of a network to adapt to a changing environ-

ment has been tackled to some extent by previous literature. For instance, in [Klü+19;
Bab+20] a mathematical framework for a rigorous definition of network flexibility is
provided. In particular, for a given demand sequence, network flexibility Φ(z, c) is
defined as the ratio of satisfied demands within time limit z and cost limit c to the
total number of demands. This definition can be easily connected with the present
cost model, resulting in a more complete mathematical framework.
The intention of defining network flexibility is to measure the frequency of non-ideal
responses to a demand change. Ideally, every demand change should result in a state
change leading to a valid state. In real life, adaptation algorithms are not perfect and
demands may be unsatisfiable, thus it could happen that the network cannot find a
valid state for a new demand. As a result, we can split the sequence of demands
{Dj } into two non-overlapping sequences of satisfied demands {Dj∈ } and unsatisfied
demands {Dj∈/ } based on whether χ(dj ) ∈ V(dj ) or not. From these sets, we can define
the maximum flexibility ϕ of the network as

{Dj∈ }
ϕ = lim , (4)
j→∞ |{Dj }|

where the operator | · | yields the total length of a sequence. The maximum flexibility
ϕ is thus the ratio of satisfied demands to total demands, in the absence of the cost

42
 3.4. Cost model of a flexible network

and time constraints.

3.4. Cost model of a flexible network

In this section, we analyze the components of the total cost in a flexible network and
relate them to the flexibility framework defined in the previous section.

3.4.1. General definitions

As introduced in Sec. 3.3.2, our cost model consists of three independent components:
readiness, proaction, and reaction costs. These components can be straightforwardly
combined to obtain the total cost of operating a network.

Definition 3.4.1. The total cost Q of operating a flexible network over a long time interval
(0, τ ) is
P R
Q=K +C +C , (5)
P R
where K, C , and C are the mean readiness cost, proaction cost, and reaction cost.

For notation convenience, these components reflect cost over time (in arbitrary mone-
tary units per time unit), rather than absolute cost. Hence, the absolute cost of operat-
ing a network over interval (0, τ ) is Qτ .
In order to achieve a more powerful model, the total cost Q includes not only ex-
penses, but also revenue coming from providing service to users. This revenue is
modeled as negative cost, hence we say that the network is profitable over interval
(0, τ ) if and only if Q < 0. Although a network provider could charge users when
they specifically request a service, nowadays a subscription-based revenue, in which
users pay a flat rate for a service, is the dominant strategy [ZWW14; KDV08]. Thus,
P R
we model revenue as a part of the readiness cost, resulting in C > 0, C > 0 and
K < 0 in a profitable network.

3.4.2. Readiness cost

The instantaneous readiness cost K(s, d) is the cost of operating a network in state s
under demand d. Formally:

K(s, d) : S × D 7→ R+ . (6)

In words, K(s, d) reflects how well the network is satisfying demand d. It includes
both cost and revenue of operating a state: resource consumption, user payment via
subscriptions, penalizations for unsatisfied demands, etc. Thus, it is the only cost

43
3. Modeling the Cost of Flexible Communication Networks

component that can take negative values, which implies that the network operator
obtains a profit from operating in the current state. This fact leads to the definition of
the optimal adaptation function χ∗ (dj ), which returns the valid state that minimizes the
readiness cost for demand dj :

χ∗ (dj ) = arg min K(s, dj ), (7a)

s
s.t. s ∈ V(dj ). (7b)

In real scenarios, however, finding the optimal solution to this problem may be too
time consuming. We thus consider a more general definition of the adaptation func-
tion χ(dj ), which approximates χ∗ (dj ) but may return suboptimal states or may fail to
find a valid state. The ability of the adaptation function to return a (possibly subopti-
mal) valid state is captured by the maximum flexibility ϕ as defined in (4) in Sec. 3.3.3.
When the network adapts to a sequence of demands {Dj } via a sequence of states
{Sj }, this results in a sequence of readiness costs that can be modeled as the stochastic
process {Kj }j∈Z for every different demand-state pair. Since {Dj } and {Sj } are sta-
tionary, it follows that {Kj } is also stationary. Therefore, the mean readiness cost can
be defined as the expected value of this sequence:

K , E{K}. (8)

Note that we use a different variable to index the elements of process {Kj } with
respect to processes {Dj } and {Sj }. This is due to the possible presence of multi-
ple demand-state combinations that result in different readiness costs. To explain
this, let us consider a system facing demand dj by implementing state sj = χ(dj ).
The resulting readiness cost in this situation is kj = K(sj , dj ) (with slight abuse
of notation). When a new demand dj+1 is requested, the readiness cost changes
to kj+1 = K(sj , dj+1 ), as state sj may not satisfy the new demand, leading to de-
graded performance and higher cost. At this point, there are multiple possibilities for
the next readiness cost value. It could happen that the network finds a valid state
sj+1 = χ(dj+1 ) before the demand changes again, leading to kj+2 = K(sj+1 , dj+1 ) after
state change esj . Conversely, the system may be unable to find a valid state or a new
demand may appear before the new state is implemented, leading to a new readiness
cost value of kj+2 = K(sj , dj+2 ).
In order to model the cost resulting from the offset between demands and valid states,
we define the state delay δ(τ ) at time instant τ as the index difference between current
demand d(τ ) = dj and current state s(τ ) = sj , such that sj = χ(dj−δ(τ ) ). Similarly to
{Sj } and {Dj }, the sequence of state delays can be modeled by the discrete stochastic
process {∆j }j∈Z . The instantaneous state delay resulting from a sequence of demands
and states is shown in Fig. 3.2. We denote the marginal pmf of {∆j } as f∆ (δ), which
yields the overall probability of the network operating with state delay x.

44
 3.4. Cost model of a flexible network

0 2 4 6 8 10

0 2 4 6 8 10

0 2 4 6 8 10

Figure 3.2.: Demands, states, and state delays experienced by an action-interrupting

flexible network. The state delay is the instantaneous difference between
demand and state indices (numbers in circles).

We define the readiness degradation function:

K∆ (δ) , E{K|∆ = δ} (9)

as the mean readiness cost when the state delay is ∆ = δ. This function character-
izes the performance of the network when dealing with delayed states. An example
of such a function is shown in Sec. 3.5.2. In a properly designed network, the mean
readiness cost is lowest when state delay is ∆ = 0, that is, when the network imple-
ments a valid state. Moreover, the cost of a state should monotonically grow with
the state delay, reflecting that the demand becomes, on average, increasingly different
from the last satisfied demand. Formally, this implies that K∆ (δ2 ) ≥ K∆ (δ1 ) if and
only if δ2 ≥ δ1 . This leads to the following conclusion.

Lemma 3.4.2. A necessary condition for a network to be profitable is K∆ (0) < 0.

The proof for Lemma 3.4.2 is trivial, as it implies that a profitable network (Q < 0)
requires at least that operating in valid states is profitable. Knowing this fact, the
following lemma provides a method to compute the mean readiness cost.

Lemma 3.4.3. The mean readiness cost K of a flexible network can be calculated in terms of
K∆ (δ) as:
∞
X
K= K∆ (δ)f∆ (δ). (10)
δ=0

Proof. Eq. (10) follows directly from the application of the law of total expectation
[Bil95].

45
3. Modeling the Cost of Flexible Communication Networks

As mentioned before, K∆ (δ) is a characteristic function of the analyzed network and

has to be measured, simulated, or theoretically derived for each case. The pmf f∆ (δ)
can be obtained from FT (t) and FZ (z), that is, from the distributions of demand du-
ration and action time. Nonetheless, the resulting expression for f∆ (δ) is affected by
the behavior of the network when a new demand change appears during the action
phase, that is, while looking for or moving to a new state.
We refer to a network as action-persistent if its action phases are not interrupted by
a new change in the demand. That is, an action-persistent network carries on with
the action phase of the demand change that originated it, and eventually realizes
the associated state change, regardless of any new demand change that may occur
in the meantime. As a result, the new state may not be optimal from the start if
there are other demand changes after the first demand change. Conversely, an action-
interrupting network stops and resets its action phase if a new demand appears dur-
ing the action phase. As a consequence, an action-interrupting network only realizes
state changes leading to valid, non-delayed states. Fig. 3.2 is an example of an action-
interrupting network behavior. We can observe, for instance, that the action phase
associated with demand change d e 3 is interrupted by d
e 4 , which starts a new action
phase that eventually leads to the implementation of s∗5 and optimally satisfies d5 . We
argue that action-interrupting networks are the best option for operators who prefer
to implement valid states rather than not to interrupt adaptations. Therefore, in this
thesis we only show the derivation of f∆ (δ) for action-interrupting networks, leaving
the analysis of action-persistent networks for future work.
In our path to calculate f∆ (δ), we define the random variable R as the time difference
between any instant in the considered interval (0, τ max ) and the most recent demand
change. The probability density function (pdf) of this variable is provided in the fol-
lowing lemma.

Lemma 3.4.4. The pdf fR (r) of R is

1 − FT (r)
fR (r) = , (11)
T
where T , E{T }.

Proof. We introduce the intermediate random variable T 0 to model the duration of the
active demand at any uniformly-selected instant. By the law of the total expectation:
Z ∞
fR (r) = fR|T 0 (r|t)fT 0 (t)dt, (12)
0

where fT 0 (t) is the pdf of T 0 and fR|T 0 (r|t) is the conditional pdf of R when the most
recent demand is known. The probability of randomly selecting a demand is directly
proportional to its duration. From this fact and the law of total probability it follows

46
 3.4. Cost model of a flexible network

that
t · fT (t) t · fT (t)
fT 0 (t) = R ∞ = . (13)
0
ξ · fT (ξ)dξ T
The conditional pdf fR|T 0 (r|tj ) yields the probability density of selecting an instant
that is r time units after the start of demand dj , given that the active demand is dj .
Since there must be no bias when selecting these instants, it is clear that fR|D0 (r|tj ) = t1j
if 0 ≤ r < tj and fR|D0 (r|dj ) = 0 otherwise. Combining (13) with this fact results in:
Z ∞
fT (t)
fR (r) = dt, (14)
r T
which directly leads to (11).

Using Lemma 3.4.4 we can directly calculate an expression for f∆ (δ), as shown in the
following lemma.

Lemma 3.4.5. The pmf f∆ (δ) of the state delay of an action-interrupting network is
(
αϕ if δ = 0,
f∆ (δ) = δ−1
(15)
(1 − αϕ)(1 − βϕ) βϕ if δ > 0,

where Z ∞
1
α, FZ (t) (1 − FT (t)) dt (16)
T 0

and Z ∞
β, FZ (t)fT (t)dt. (17)
0

Proof. An action-interrupting network resets its action phase every time the demand
changes. Hence, for any instant, the probability of operating with state delay ∆ = 0 is
the probability of being able to find a valid state and surpassing the action time for the
most recent demand. The probability of the former event is given by the maximum
flexibility ϕ, whereas the latter event is derived as follows [Dur10]:
Z ∞
f∆ (0) = ϕ Pr{Z ≤ R} = ϕ FZ (r)fR (r)dr, (18)
0

which leads to the first case of (15) after substituting (11). If this is not the case, with
probability (1 − αϕ), the probability of reaching a state delay δ > 0 is the probability
of being able to find a valid solution after x unsuccessful attempts. This event follows
a geometric distribution of parameter p:
Z ∞
p = ϕ Pr{Z ≤ T } = ϕ FZ (r)fT (r)dr, (19)
0

which is the probability of obtaining a valid solution within an action phase that is

47
3. Modeling the Cost of Flexible Communication Networks

shorter than the duration of a demand. Given the pmf of a geometrically-distributed

random variable as p(1 − p)δ−1 (for δ > 0), (15) is finally obtained.

With an expression for f∆ (δ), we calculate the resulting mean readiness cost in the
following theorem.

Theorem 3.4.6. The mean readiness cost K of an action-interrupting network is

K = αϕK∆ (0) + (1 − αϕ)βϕK

bβ , (20)

where ∞
X
K
bβ , K∆ (δ)(1 − βϕ)δ−1 . (21)
δ=1

Proof. Eq. (20) is the result of combining (10) and (15).

The expression in Theorem 3.4.6 allows us to calculate the mean readiness cost of
an adaptive system from its demand duration distribution, action time distribution,
readiness degradation function, and maximum flexibility. In order to find out if a
network is profitable, the following corollary can be used.

Corollary [Link]. A necessary condition for a flexible network to be profitable is

α K
bβ
< . (22)
(αϕ − 1)β K∆ (0)

given that K∆ (0) < 0.

Proof. A profitable network must fulfill K < 0, which implies that K∆ (0) < 0 (Lemma 3.4.2).
After applying these relations in (20), we reach (22).

Corollary [Link] provides us with a simple method to rule out non-profitable network
configurations. The left side of (22) reflects the frequency of demand changes and the
swiftness and effectiveness of the adaptation, whereas the right side is influenced by
the quality of the solutions. As a result, it delimits a border for finding profitable
configurations within the speed-quality tradeoff. An example application of Theo-
rem 3.4.6 and Corollary [Link] is presented in Sec. 3.5.2.
Once we have a closed-form expression of K, we can derive some interesting prop-
erties that may be useful when analyzing flexible communication networks. In order
to highlight the dependence of K on the maximum flexibility ϕ, we use the notation
K(ϕ) in the following.

Property 3.4.7. As ϕ → 0, the value of the mean readiness cost K(ϕ) tends to the limit of the
RDF K∆ (δ) when δ → ∞. That is:

lim K(ϕ) = lim K∆ (δ) (23)

ϕ→0 δ→∞

48
 3.4. Cost model of a flexible network

Proof. (23) is a consequence of the final value theorem. This theorem states the follow-
ing equivalence [Opp+97]:
∞
X
lim K∆ (δ) = lim(z − 1) K∆ (δ)z−δ , (24)
δ→∞ z→1
δ=0

where the right-hand term is the Z-transform (using z as variable) of K∆ (δ). We per-
form the variable change z = (1 − βϕ)−1 , which leads to:
∞
X
lim K∆ (δ) = lim(z − 1) K∆ (δ)z−δ (25)
δ→∞ z→1
δ=0
∞
X
= lim βϕ K∆ (δ)(1 − βϕ)δ−1 (26)
ϕ→0
δ=0

= lim βϕK
b β (ϕ) (27)
ϕ→0

= lim αϕK∆ (0) + (1 − αϕ)βϕK

b β (ϕ) (28)
ϕ→0

= lim K(ϕ), (29)

ϕ→0

which is the identity stated in (23).

The implications of Property 3.4.7 are those intuitively expected: as the maximum
flexibility decreases, the mean readiness cost in the network approaches that cor-
responding to an infinite state delay. Thus, in a real network, K∆ (δ) converges to
the mean cost of operating in a state that is totally uncorrelated with the demand as
ϕ → ∞. The next property deals with the opposite extreme case.

Property 3.4.8. As ϕ → 1, the value of the mean readiness cost K(ϕ) is given by:
∞
X
lim K(ϕ) = αK ∆ (0) + (1 − α)β K∆ (δ)(1 − β)δ−1 (30)
ϕ→1
δ=1

Proof. This identity can be obtained by simply introducing ϕ = 1 into K(ϕ) as pre-
sented in (20).

By comparing Properties 3.4.7 and 3.4.8, we observe that, although ϕ = 0 guarantees

operating at the worst possible readiness cost, ϕ = 1 is not sufficient to operate at the
minimum readiness cost. Indeed, the RDF K∆ (δ) and the distributions of the dura-
tions of demands and states, by means of parameters α and β, define the minimum
achievable cost.
Once we know the mean readiness cost for extreme values of ϕ, the following proper-
ties address the shape of K(ϕ).

49
3. Modeling the Cost of Flexible Communication Networks

Property 3.4.9. For an increasing RDF K∆ (δ), the mean readiness cost K(ϕ) is a decreasing
function of ϕ. That is:

K(ϕ1 ) ≤ K(ϕ2 ) ⇐⇒ ϕ1 ≥ ϕ2 , ∀ϕ1 , ϕ2 ∈ [0, 1]. (31)

dK(ϕ)
Proof. We prove Property 3.4.9 by showing that dϕ
≤ 0 ∀ϕ ∈ [0, 1]. This is done in
Appendix A.1.

Property 3.4.10. For a bounded and increasing RDFK∆ (δ), the mean readiness cost K(ϕ) is
a convex function of ϕ. That is:

K (sϕ1 + (1 − s)ϕ2 ) ≤ sK (ϕ1 ) + (1 − s)K (ϕ2 ) ∀s, ϕ1 , ϕ2 ∈ [0, 1] (32)

d2 K(ϕ)
Proof. We prove Property 3.4.10 by showing that dϕ2
≥ 0 ∀ϕ ∈ [0, 1]. This is done in
Appendix A.2.

Properties 3.4.9 and 3.4.10 lead to important implications in the relationship between
the mean readiness cost K(ϕ) and the maximum flexibility ϕ. Namely, they allow us
to conclude not only that the mean readiness cost increases as the maximum flexibility
decreases, but also that the lower the maximum flexibility ϕ, the faster the growing
rate of K(ϕ), regardless of any other parameter. In addition, the convexity of K(ϕ)
will allow us to identify the value of ϕ that leads to the global minimum total cost Q,
since the action cost is a linear function of ϕ, as we show in Sec. 3.4.4.
The shape of four exemplary K(ϕ) functions for multiple RDF shapes is depicted
in Fig. 3.3 and Fig. 3.3. In Fig. 3.3, we show RDFs with the same range, −10 ≤
K∆ (δ) ≤ 10, in four different shapes: step, hyperbolic, linear, and logistic growth.
Their corresponding mean readiness cost functions K(ϕ) are shown in Fig. 3.3, along-
side experimental results that are obtained after simulating a system with those RDFs,
exponentially-distributed demand durations of mean 1, i. e., T ∼ Exp(1), and uni-
formly-distributed action times between 0 and 0.5, i. e., Z ∼ U nif (0, 0.5), for a total
simulated time of 20,000 demand durations. We observe that the shape of K(ϕ) is al-
ways convex and decreasing, regardless of the shape of its associated RDF, which may
be non-convex and non-concave. We also see that, as expected from Properties 3.4.7
and 3.4.8, K(0) ≈ 10 but K(1) 6= −10. Finally, we conclude that that the model and
the simulation results are an almost perfect match.

3.4.3. Proaction cost

The proaction cost C P (d

e j ) of a flexible network represents the cost of consuming com-
putational resources so as to find the new state after a demand change. This cost com-
ponent can be formally defined as a function mapping a demand change to a cost
value:
C P (d
e j ) : D2 7→ R+ . (33)

50
 3.4. Cost model of a flexible network

Figure 3.3.: Four exemplary readiness degradation functions (RDFs) featuring the
same range but different shapes.

Figure 3.4.: Mean readiness cost K(ϕ) against maximum flexibility ϕ corresponding
to the four RDFs described above, for exponentially-distributed demand
durations of mean 1, and uniformly-distributed action times between 0
and 0.5. Theoretical and simulated values.

The exclusive dependency on d e j implies that this cost is present every time there is
a demand change, even if there is no eventual state change. We identify two factors
contributing to the proaction cost. On the one hand, a demand change may imply
an instantaneous time-independent cost C0P , for instance resulting from the activation
of new capabilities to solve the adaptation problem. On the other hand, while the
adaptation problem is being solved, additional resources (CPU, memory, etc.) are
consumed during the proaction phase, incurring in a cost of CzP monetary units per
time unit. As a result, we can express the proaction cost as a function of the proaction

51
3. Modeling the Cost of Flexible Communication Networks

time of an action-interrupting network as follows:

1
C P (d C0P + CzP · min(zjP , tj+1 ) ,


ej ) , (34)
tj+1

where the minimum operator guarantees that the proaction phase is stopped if the
1
demand changes and the term tj+1 normalizes the cost to the duration of the demand.
From (34) and the sequence of proaction times {ZjP } we define a new stochastic pro-
cess {CjP }j∈Z to model the sequence of proaction costs, such that:

1
CjP = C0P + CzP · min(ZjP , Tj+1 ) .


(35)
Tj+1

The mean of the variable above is presented in the following.

Theorem 3.4.11. The mean proaction cost C P of an action-interrupting network is:

P
!
P C0P βP Z
C = + + 1 − β P CzP , (36)
T T

where Z ∞
P
β , FZ P (t)fT (t)dt, (37)
0
P
and Z , E{Z P }.

Proof. After applying the expectation operator to (35), we need to calculate

E{min(ZjP , Tj+1 )}.

The random variable within the brackets takes the same values as ZjP when ZjP ≤
Tj+1 . From stationarity and the law of total expectation:
P
E{min(Z P , T )} = Pr{Z P ≤ T } · Z + (1 − Pr{Z P ≤ T }) · T (38)

The probability Pr{Z P ≤ T } is derived in the same way as (17) to yield (37).

In Sec. 3.5.3, we show an example network where we apply Theorem 3.4.11 to find
out the optimal number of CPU cores to be used during the proaction phase.

Corollary [Link]. A necessary condition for an action-interrupting network to be profitable

is
P T K + C0P
Cz < P
. (39)
(β P − 1)T − β P Z

Proof. This relation follows directly from the fact that a profitable network must fulfill
P
K + C < 0.

52
 3.4. Cost model of a flexible network

Corollary [Link] provides us with an upper bound on the maximum number of ad-
ditional resources that a flexible network is allowed to utilize in order to cope with
demand changes before it turns unprofitable.

3.4.4. Reaction cost

The reaction cost C R (esj ) reflects the effort of performing the state change required for
an adaptation, after this has been selected in the proaction phase. Therefore, we define
it as a function of the state change esj :

C R (esj ) : S2 7→ R+ . (40)

Note that C R (esj ) = 0 if the demand change d e j results in no adaptation, that is, if
sj+1 = sj . This can happen either if sj is already optimal, or if the adaptation algorithm
could not find a better state which satisfying dj .
Based on the same rationale as with the proaction cost, we identify two factors con-
tributing to the reaction cost. First, an instantaneous time-independent cost C0R mod-
els the activation of state-changing procedures (such as memory allocation for vir-
tual migrations [MAGVK19a], for instance). Second, we define a constant rate of CzR
monetary units per time unit to characterize the usage of additional resources when
changing the network state. Consequently, we formulate the reaction cost as:

1
C R (esj ) , C0R + CzR · min(zjR , tj+1 − zjP )


(41)
tj+1

whenever tj+1 ≤ zjP and the demand is satisfiable. Otherwise, C R (esj ) = 0 as no new
state has been generated. We define the stochastic process {CjR }j∈Z to model a time-
ordered sequence of reaction costs as:

1
CjR = C0R + CzR · min(ZjR , Tj+1 − ZjP ) ,


(42)
Tj+1

for every index i whenever Tj+1 ≤ ZjP and CjR = 0 otherwise.

R
Theorem 3.4.12. The mean reaction cost C of an action-interrupting network is:

R ϕβ P
C0R + CzR βZ(1 − β)T − Z P .



C = (43)
T

Proof. The equality (42) occurs when ZjP ≤ Tj+1 with probability β P . By the law of
total expectation, we just need to figure out the value of E{min(ZjR , Tj+1 − ZjP )}. The
random variable within the brackets takes the same values as ZjR when Zj ≤ Tj+1 ,
that is, with probability β. After some straightforward algebra, (43) is obtained.

53
3. Modeling the Cost of Flexible Communication Networks

Figure 3.5.: Topology of the example network implementing random flow requests.
The topology is that of AT&T North America [ATT] taken from the Inter-
net Topology Zoo [Thea].

Corollary [Link]. The optimal maximum flexibility ϕ∗ that minimizes the total cost Q is
that satisfying the following equation:
∞
− αϕ∗
 
∗1
X
∗
αK∆ (0) + β K∆ (δ) 1 − 2αϕ − βϕ ∗
(δ − 1) (1 − βϕ∗ )δ−1 =
δ=1
1 − βϕ
βP
C0R + CzR βZ(1 − β)T − Z P . (44)



−
T

Proof. The total cost Q can be computed as a function of ϕ as:

P R
Q(ϕ) = K(ϕ) + C + C (ϕ). (45)
R
From Property 3.4.10 and Theorem 43, we know that K(ϕ) and C (ϕ) are convex over
ϕ. Thus, the value of ϕ that globally minimizes K(ϕ) satisfies:

dQ(ϕ)
= 0. (46)
dϕ
ϕ=ϕ∗

After differentiating (20) and (43) and some straightforward algebra, (44) is obtained.

3.5. Application example

The main goal of the cost model presented in this chapter is to help in the definition
of the FSSP, as well as in solving it dynamically. These issues are addressed in de-
tail in Chapters 4 and 6. However, this cost model is also general enough to be used

54
 3.5. Application example

for a wide variety of use cases. Owing to this, in this section we show an example
application to a different problem. Namely, we present the problem of designing a
flow-embedding network facing frequent requests. After proposing several alterna-
tives to solve it, and use the derived cost model to select the best implementation
options.

3.5.1. Network description

We consider a communication network with the topology shown in Fig. 3.5. The objec-
tive of this network is to provide virtual connectivity between pairs of nodes, that is,
to embed flow requests between any two nodes. These flows have to be implemented
on single paths, i. e., there can be no fractional flows. This can be accomplished by
dynamically reserving links and reconfiguring the routing tables in the intermediate
nodes. Providing this service yields a revenue for the network operator, but manag-
ing and reconfiguring the network is costly. As a result, the operator wants to serve
users with the minimum operational cost, so that the net revenue is maximized.
Each flow request consists of a source-destination pair and a required throughput. For
simplicity, we assume that all nodes have the same probability of becoming source or
destination, regardless of their geographical position, so that every source-destination
pair is equally likely. We define the instantaneous demand d as the set of all active
flow requests. As a consequence, every time a new flow request arrives, or a flow is
not needed any more, the demand d changes. The instantaneous network state s is
defined as a vector containing the state of each link (active or inactive) and the list
of flows that are assigned to it, if any. Inactive links cannot carry any flows, but they
consume less power. When a link is active, multiple flows can be assigned to it as long
as the capacity of the link is not exceeded. Without loss of generality, we normalize all
link capacities to 1. Moreover, we model the throughput requested by each flow with
a random uniform distribution between 0 and 1. We deem that a demand is satisfied if
the current network state allows the embedding of all flows contained in the demand
without exceeding any link capacity.
The duration of the demands T (in seconds) follows a Pareto Type II distribution, also
called Lomax distribution [Lom54]:
 σ
t
FT (t) = 1 − 1 + , (47)
λ
λ
for t ≥ 0. We set λ = 10 and σ = 2.25 so that the mean demand duration is T = σ−1 =8
seconds. This distribution is selected since it is commonly observed in the interarrival
time between internet bursts, file sizes, transfer times, etc [Dow05]. Moreover, we
choose a mean demand duration of 8 s so that it is comparable to the action time of
reconfiguring the network. Hence, it is unclear whether a network can operate prof-
itably in this situation when using conventional cost models. Note that this selection
of parameters is due to illustrative purposes, but other values or distributions can be

55
3. Modeling the Cost of Flexible Communication Networks

used without affecting the effectiveness of the model.

The problem of providing the intended connectivity can be formulated as an instance
of the integer min-cost multicommodity flow problem [Tom66]. This problem is known
to be NP-Hard [EIS75], thus the network operator relies on an approximation algo-
rithm rather than on an exact approach. In our example, the operator uses a genetic
algorithm as the adaptation algorithm [FBB19]. In a nutshell, the operation of the ge-
netic algorithm is as follows [MTK01]. First, a certain number of random solutions,
called the population size p, are generated. Each solution contains a flag per link indi-
cating if this link is active or not. Then each solution is evaluated to assess how close
it is to satisfy the current demand and how many links it uses. After all solutions have
been evaluated, the worst ones are discarded whereas the best ones are kept for the
next generation. These are then combined to each other and randomly modified to
produce the next generation. These steps are repeated until a convergence criterion is
satisfied, which in our case is the absence of improvements after 25 generations.

3.5.2. Selection of a profitable population size

Given the high number of parameters in a genetic algorithm, the operator is inter-
ested in finding the right ones so that the network is profitable. For example, they
want to select the population size p. Selecting the right p is not trivial, since it affects
the speed-accuracy trade-off of the algorithm. On the one hand, a large p increases
the probability of eventually finding an optimal solution with minimum readiness
cost. On the other hand, the higher p, the more solutions have to be evaluated, which
increases proaction time and cost.
The readiness cost kj at any point is the combination of three factors. First, subscribed
users provide a constant revenue of 71 normalized cost units (ncu). Second, active
links have a cost of 11 ncu/s, whereas inactive links do not cost anything. Finally,
whenever a requested flow is not being satisfied, the operator has to pay a compensa-
tion of 10 ncu/s to each affected user. Thus, the average readiness cost of state s and
demand d is:
K(s, d) = −71 + 11l(s) + 10v(s, d) ncu, (48)
where l(s) is the number of used links in state s and v(s, d) is the number of unsatisfied
flows for demand d and state s.
The operator is considering to use a population size in the set:

P = {250, 750, 1250, 1750, 2250, 2750} .

After evaluating the performance of the genetic algorithm in a dedicated simulator,

we observe that the action time can be modeled by a uniform distribution such that
Z ∼ U(0, Z),
b where Zb = 0.016 · p. We also measure the maximum flexibilities

Φ = {0.35, 0.55, 0.64, 0.7, 0.73, 0.75}

56
 3.5. Application example

Figure 3.6.: Readiness degradation function K∆ (δ) of a network employing a genetic

algorithm to implement integer multicommodity flows for six different
population sizes.

for each p in the same order as Π, representing the frequency of demand-satisfying

solutions. Finally, we measure the readiness degradation function K∆ (δ) for state
delays 0 ≤ δ ≤ 30 for each population size p, which is shown in Fig. 3.6. We observe
that the mean readiness cost steadily increases with the state delay, as a result of higher
compensations due to unsatisfied flows and unnecessarily active links. It is also clear
that the evolution of the cost for the different population sizes is very similar, although
low population sizes lead to higher link usage, which increases the cost.
From the above measurements and the distributions of T and Z, we are able to com-
pute parameters α and β as shown in (16) and (17). With this and the readiness degra-
dation function K∆ (δ), we can apply Corollary [Link] to figure out if any of the consid-
ered population sizes may lead to a profitable network. A graphical representation of
the result is shown in Fig. 3.7, where left and right sides of inequality (22) are depicted
as horizontal and vertical axes, respectively. We observe that three of the considered
population sizes, p ∈ {1250, 1750, 2250}, lie on the profitable region. An interesting
behavior is captured by the model, as those populations that are lower than 1250 or
greater than 2250 lead to unprofitable networks. The explanation is that, when the
population is small, the quality of the states yielded by the genetic algorithm is not
good enough to properly address the demands. Conversely, when the population is
large, the action time is so high that the network cannot properly cope with frequent
demand changes.

3.5.3. Optimal parallelization level

Let us now consider that the network operator has the ability to dedicate multiple
CPU cores to solving the adaptation problem in the proaction phase. Increasing the

57
3. Modeling the Cost of Flexible Communication Networks

Figure 3.7.: Graphical representation of inequality (22) for testing the profitability of
different population sizes.

number of cores reduces the proaction time, thus decreasing the state delay, which
may lead to higher revenue. Nevertheless, utilizing more cores also increases the
proaction cost, which may counter the revenue increase and result in higher total cost.
In order to find out the optimal level of parallelization, we can combine Theorems
3.4.6 and 3.4.11 to predict the evolution of readiness and proaction costs as the number
of cores assigned to the adaptation algorithm increases.
Let us denote the parallelization level, i. e., the number of additional CPU cores used
in the proaction phase, as g. For simplicity, let us assume that the reaction cost and
time are negligible so that Z P ≈ Z, hence Z P ∼ U(0, Z).b Clearly, the value of Zb is a
Z
decreasing function of g. Namely, we define it as Zb = Υ(g) , where Zb0 is the maximum
b0

proaction time with a single core and Υ(g) is the time reduction factor for g cores.
Ideally, Υ(g) = g if the load can be perfectly shared among all cores. Nonetheless, in
real scenarios it is observed that Υ(g) grows linearly at first, but eventually saturates
due to imperfections in load division [Glo15; Ber+14]. To capture this, we use the Υ(g)
depicted in Fig. 3.8.
Regarding the proaction cost components, let C0P = 0 (no additional cost for starting
the proaction cost) and CzP = 1 · g ncu/s, where g is the number of assigned CPU cores,
that is, the parallelization level. This value of CzP means that the network consumes
1 ncu/s per used CPU core during the proaction phase, in addition to the readiness
cost. Finally, based on the previous results, we select a population size of p = 1750,
which implies a maximum flexibility of ϕ = 0.7 and Zb0 = 28.
We can calculate the relationship between the mean proaction cost C P and the paral-
lelization level g by feeding the aforementioned expressions into (36), in Theorem 3.4.11.
The result is shown in Fig. 3.9, where an interesting behavior can be observed. Up to
around g = 8, the proaction cost increases rapidly, since the duration of the proaction
phase is limited by the demand duration. Indeed, when g = 1, the average demand
duration is T = 8 s, whereas the mean proaction time is, Z P = Z20 = 14 s. As a
b

consequence, the network is almost always in the proaction phase, and thus increas-

58
 3.5. Application example

Figure 3.8.: Reduction factor of the proaction time for different parallelization levels.
The dashed line has unit slope.

P
Figure 3.9.: Mean proaction cost C as a function of the parallelization level.

ing g only increases the proaction cost without affecting the duration of the proaction
phases. Nevertheless, as g grows, eventually the proaction time becomes lower than
the demand duration, thus allowing the network to leave the proaction phase and
leading to a less steep cost increase.
Finally, in Fig. 3.10 we show the combined readiness and proaction cost for this sce-
nario, which can be achieved via Theorems 3.4.6 and 3.4.11. We clearly observe a min-
imum point at g = 7 cores, which is thus the optimal parallelization level. Before this
value the proaction cost is lower, but the network cannot cope with demand changes
fast enough, resulting high readiness cost due to large state delays. For g > 7 the
readiness approaches its minimum value but the proaction cost increases, resulting in
higher combined cost.

59
3. Modeling the Cost of Flexible Communication Networks

P
Figure 3.10.: Combined mean readiness and proaction costs K + C as a function of
the parallelization level.

3.6. Summary
In this chapter, we propose a comprehensive cost model for a flexible, dynamically-
adapting network. We use probability-theory tools to gain a deep understanding of
the multiple phases and cost components that a flexible network may feature. Namely,
we identify three main adaptation phases (readiness, proaction, and reaction phases)
and three corresponding cost components (readiness, proaction, and reaction cost).
These cost components can be combined in order to calculate the total operating cost.
In addition, we provide several expressions and relationships that can be used to ac-
curately predict cost and take design decisions. Finally, we apply the cost model to
realistic examples to show how this can be done in a generic scenario.

60
4. Optimal Functional Split Selection in 5G
Radio Access Networks

4.1. Introduction

4.1.1. Motivation, scope, and challenges

In Chapter 2, we discuss the advantages of a centralized 5G RAN architecture, which

are mainly twofold. On the one hand, function centralization enables improved func-
tion coordination, which allows the implementation of advanced techniques for mit-
igating interference, such as interference coordination [Sor+17], or joint transmission
and reception [Jun+14]. This increases the signal quality at both the UEs and the
gNodeBs, thus improving user data rates. On the other hand, centralization entails
resource pooling and the use of large-scale computing servers, which reduces the op-
erating cost of the whole network. Owing to these two advantages, we conclude that
the mobile network operator should aim at the highest possible level of function cen-
tralization.
Nonetheless, centralization is usually constrained by the fronthaul network connect-
ing CU and DUs, since the capacity required to support high centralization levels may
exceed that of currently deployed networks [IT18; GS+18a]. That is, the operator may
only be able to centralize a subset of processing functions, but not all of them. The se-
lection of these centralized functions needs to be done carefully so as to improve user
performance and reduce the operating cost as much as possible without surpassing
the capacity limits.
However, the performance and the operating cost of a partially centralized RAN ar-
chitecture depend not only on the network state as configured by the operator, but on
instantaneous network conditions, such as the geographical distribution and activity
of the users. Intuitively, if UEs are more concentrated or active on a certain area, we
would expect that centralizing the processing functions of the gNodeBs located in that
area is more beneficial that a homogeneous centralization of functions over the whole
network. As a result, we need to take these instantaneous conditions into account in
order to find the optimal centralization and flow vectors for any instant.
In our path to dynamically operate an adaptive 5G RAN, the goal of this chapter is to
find the network state that optimizes user data rates, operating cost, or network rev-
enue for a fixed time instant. That is, we start by considering a single “snapshot” of the
RAN in which all variables (such as UE locations, channel quality, etc.) are fixed. We
refer to this as instantaneous optimization, since the network state is selected to optimize

61
4. Optimal Functional Split Selection in 5G Radio Access Networks

performance for a single instant. For this, we build upon the static FSSP formulation
that is presented in Sec. 2.3.2. In Chapter 6, we extend this instantaneous formula-
tion so as to include time variation and support dynamic adaptation to changes in the
network demands.
Finding the instantaneous optimal RAN state entails a series of challenges. The first
one is the modeling the user performance, operating cost, and readiness cost as a
function of the functional split of each gNodeB. This modeling includes the effect of
interference coordination on the user throughput and the impact of resource pooling.
In addition, we need to formulate an optimization problem that reflects the instan-
taneous distribution and activity of the UEs along with the fronthaul network con-
straints. Solving this problem may be also difficult, since it may involve non-linear
components and problem instances may be rather large, since a complete RAN may
contain hundreds of gNodeBs and thousands of UEs. In spite of this, it is not sufficient
to find an approach that yields optimal or near-optimal state vectors, but this has to be
achieved in a very limited time interval. Indeed, since the problem is only fully valid
for a single time instant and our ultimate intention is to be able to adapt the network
state dynamically, the optimization approach must converge quickly to a solution.
In this chapter, all these challenges are addressed and our proposed approaches are
evaluated.

4.1.2. Key contributions

In relation to the aforementioned challenges, this chapter bases on our previous works
[MAJK21], [MAK21], and [MAPK21] to feature the following contributions:
1. We model a 5G RAN featuring a dynamically-adapting functional split and pro-
pose three different objective functions to find the optimal network state. The
first function represents the proportionally-fair user throughput, which is esti-
mated after modeling the effect of the functional split on the interaction between
pairs of gNodeBs. The second functions reflects the operating cost associated
with each centralization and flow vectors, which is based on the network model
presented in Chapter 3.
2. Based on these objective functions, we formulate three main optimization prob-
lems to find the instantaneously optimal network state, building upon the generic
formulation shown in Chapter 2.
3. We propose multiple reformulations and approximations to these main opti-
mization problems, with the intention of finding an optimization approach that
yields good-quality solutions in a reasonable time and can be tackled by off-the-
shelf solvers.
4. We simulate a city-sized dense 5G RAN so as to evaluate the achieved spectral
efficiencies, operating cost, readiness cost, and convergence times of our pro-
posed approaches. We also compare these approaches to static architectures and
show that they lead to substantial performance improvement and cost reduc-

62
 4.2. Related work

tion.
The remainder of this chapter is structured as follows. In Section 4.2 we present the
related work on this topic. Section 4.3 briefly summarizes the system model. In Sec-
tion 4.4, we formulate the performance maximizing version of the functional split se-
lection problem and derive multiple approximations to tackle it efficiently. Section 4.5
describes the derivation of the operating-cost minimizing version of the functional
split selection problem. In Section 4.6 we discuss our performance-to-revenue func-
tions and present the readiness-cost minimizing version of the functional split selec-
tion problem. Section 4.7 presents the experimental evaluation of these approaches
regarding convergence time, spectral efficiency, operating cost, and readiness cost.
Finally, Section 4.8 concludes the chapter.

4.2. Related work

Previous work already tackles, to some extent, the problem of statically selecting the
optimal functional split of each gNB. This previous research can be divided into two
categories: those dealing mainly with theoretical aspects and those focusing on the
implementation. In the former category, [Mae+14] is one of the first works to propose
that the functional split could be selected differently for each gNB, basing this rea-
soning on the limitations of the fronthaul network. The authors argue that function
centralization is desirable to reduce interference and cost, but limited by the fronthaul
capacity. A similar idea is developed further in [Sab+13], where a more complete
framework is presented. Nevertheless, neither work proposes an actual optimization
problem to find the best RAN configuration, but they both limit to describe the reasons
why the functional split should not be homogeneous over a 5G RAN and describe a
high-level platform supporting this feature.
Continuing with this idea, in [GS+18a] the authors formulate the problem of select-
ing the optimal functional split for the deployment phase. Their objective is to min-
imize network and computing costs while centralizing as many functions as possi-
ble. In order to estimate the required fronthaul capacity, the expected average traffic
of each gNB is used. The authors of [DGA19] face a similar problem with a differ-
ent objective: minimizing traffic delay. In [HR18b] the idea of dynamically changing
the functional split is introduced with the intention of allocating new slices within a
virtual RAN framework. Inter-cell interference reduction and fronthaul bandwidth
minimization are the main objectives when selecting the functional split, although
this selection is not updated once the slice is implemented. Finally, our earlier work
[MAK19a] present for the first time the idea of adapting the functional split dynami-
cally to cope with the instantaneous interference situation. Nonetheless, they tackle a
simplified version of the problem and focus on confirming that the network changes
slowly enough so that dynamic adaptation is possible, without providing a detailed
strategy on how to select the functional split.
Regarding the implementation aspects, there are two main works that focus on realiz-

63
4. Optimal Functional Split Selection in 5G Radio Access Networks

Variable Definition
G Number of gNodeBs
G Set of gNodeB indices
U Number of UEs
U Set of UE indices
hu Index of gNodeB serving UE u ∈ U
N Number of fronthaul network nodes
N Set of fronthaul network nodes
n0 Node index of CU, n0 ∈ N
ng Node index of gNodeB g ∈ G, ng ∈ N
E Number of fronthaul network links
E Set of fronthaul network links
ϑe Capacity of link e ∈ E
M Number of functional split option (centralization levels)
ν(m) Required link capacity of centralization level m ∈ M
M Set of functional split indices
a Vector of functional split indices for each gNodeB, a ∈ MG
f Vector of allocation of network flows
s State vector of centralization and flow vectors, s , ha, f i
И(m) Interference cancellation factor of centralization level m ∈ M
iu,g Interference power received by UE u ∈ U from gNodeB g ∈ G
Iu (a) Total interference power received by UE u ∈ U, a ∈ MG

Table 4.1.: System modeling variables.

ing a flexible functional split. In [Cha+17b], a comprehensive description of a platform

supporting multiple functional splits is presented, although the capability of chang-
ing during runtime is not included. Conversely, in [MAGVK19a] we present a pioneer
framework that enables to change the functional split of a gNB without stopping its
operation or dropping packets. However, the motivation to trigger such a change is
not studied.
To the best of our knowledge, our work [MAJK21], on which this chapter is based,
is the first work to address in detail the problem of finding the instantaneously opti-
mal functional splits of all gNBs based on its experienced interference. In addition,
[MAK21] is the first work that addresses this problem while trying to maximize the
total network revenue.

4.3. System model

In Chapter 2, Sec. 2.3.1, we present a detailed description of the system model that
is used throughout this thesis. For the sake of clarity, in this section we briefly sum-

64
 4.3. System model

marize those concepts required to formulate the dynamic functional split selection
problem. We describe the network components, explain the considered functional
split, and present the adaptation framework. In Table 4.1, we show a list of the most
relevant variables used in this chapter.

4.3.1. Network description

The considered network consists of G gNBs, whose operation is divided into a DU and
a CU. The CUs of all gNBs are deployed in a single data center, whereas the DUs are
located close to the radio equipment of the cells. As a result, there are G different DU
locations and a single CU location. CUs and DUs are connected by means of a packet-
switched fronthaul network that includes layer-2 or layer-3 switches. We assume that
these switches are able to steer and divide the incoming flows as configured by a
central controller at the CU, thus allowing for reconfigurable fractional flows. We
model this fronthaul network via a directed graph D = hN, Ei, where N is the set of
network nodes (DUs, switches, and CU) and E is the set of network links. We denote
by n0 the node corresponding to the CU and by ng the node corresponding to DU g,
such that g ∈ G , {1, ..., G}.
There are U simultaneously active UEs within the coverage area of all cells. Each
UE u ∈ U , {1, ..., U } is connected to a serving gNodeB, which is denoted by hu .
Throughout this chapter, we focus on the downlink data rates and on the downlink
interference as perceived by the UEs. Nonetheless, an extension of the analysis to
include the uplink is straightforward. Finally, we assume that all gNodeBs operate
in the same frequency bands, that is, a frequency reuse factor of 1. This is done
to highlight the performance of the interference mitigation enabled by the dynamic
functional split, although using a different reuse factor is not precluded.

4.3.2. Functional splits

The processing chain of every gNB can be divided into functions, which are often
identified with the layer or sublayers of the RAN protocol stack [Döt+13]. For each
pair of consecutive functions we define a functional split option. We denote by M
the number of possible functional split options, also referred to as centralization lev-
els. The instantaneous centralization level of a gNodeB g is denoted by ag , such that
xg ∈ M , {0, ..., M − 1}. We consider that ag = 0 denotes the lowest centralization
level, that is, the functional split option for which the least amount of functions are
centralized. Conversely, ag = M − 1 denotes the highest centralization level. The
centralization vector of all centralization levels is defined as a , [a1 · · · aG ]. Low cen-
tralization levels require less fronthaul capacity, but their interference-mitigation ca-
pabilities are limited. Conversely, gNBs implementing high centralization levels are
able to coordinate with each other to reduce the interference they cause to each other,
at the expense of requiring higher fronthaul capacity [3GP17].

65
4. Optimal Functional Split Selection in 5G Radio Access Networks

4.3.3. Interference mitigation

The mutual inter-cell interference that base stations cause to each other can be miti-
gated by using well-known techniques, such as coordinated scheduling, coordinated
beamforming, joint transmission, etc. The application of these techniques results in
improved user data rates, but they all require some level of coordination between the
involved gNB functions. Owing to potentially high processing, switching, and prop-
agation latency offered by the fronthaul network, sophisticated coordination is often
deemed infeasible for distributed functions. For example, in order to employ inter-
ference cancellation, gNBs need to generate, communicate, and apply the interference
cancellation algorithm to their transmission slots in less than the duration of a 5G time
slot [FLF19], which can be as short as 62.5 µs when using high numerologies [3GP21e].
Consequently, in order to implement an interference-mitigation technique we need
to guarantee a minimum centralization level for coordinating gNBs. For example,
joint transmission also requires the centralization up to the physical layer [Zha+17],
whereas coordinated scheduling only requires the centralization of the MAC layer
[Nar+18].
Based on the results shown in [PM17], we model the effect of an interference-mitigation
technique as a constant factor multiplying the average received interference power be-
tween two coordinating gNBs. We define И(a) as the function relating centralization
level a with the associated interference-mitigation factor of the most effective tech-
nique that it can support. The value of function И(a) ranges from 0 (full interference
cancellation) to 1 (no interference cancellation). Function И(a) is a decreasing func-
tion of a, since the higher the centralization level a, the more effective the interference
mitigation. Using И(a), the expected total interference power Iu experienced by UE u
from all gNBs can be computed as:
G
X
Iu (a) = iu,g · И(min(ahu , ag )), (1)
g=1

where iu,g is the interference power received by UE u from gNB g and iu,hu , 0, as the
UE is not interfered by its serving gNB. Note that the min( · ) operator ensures that an
interference-mitigation technique can only be used by two gNBs if both of them are
operating at the required centralization level.

4.3.4. Fronthaul network

The capacity required for a fronthaul link connecting the DU and CU of a gNB de-
pends on its centralization level, that is, on its functional split. Namely, previous
research has shown that high centralization levels, such as the Intra-PHY split or
full centralization, require large link capacities (in the order of hundreds of Gb/s),
whereas low centralization levels (such as the PDCP-RLC split) require capacities
barely larger than the user data rate (in the order of a few Gb/s) [Döt+13; MAGVK19b;

66
 4.4. Performance-maximizing FSSP

3GP17]. Formally, we model the capacity required by gNB g with centralization level
ag as the function ν(ag ). For the sake of simplicity, we assume that all gNBs offer
the same maximum user data rate, hence ν(ag ) does not depend explicitly on g. If
required, extending ν(ag ) to include this dependency is straightforward.
Finally, we define ϑe as the capacity of each fronthaul link e ∈ E. For each gNB g
producing a downlink flow between its DU and the CU, we denote by feg the fraction
of this flow that is carried over link e. For notation convenience, we also define f g ,
[f1g · · · fEg ] ∀g ∈ G and f , f 1 · · · f G as the vectors of the flow generated by gNB g


and all flows, respectively.

4.4. Performance-maximizing FSSP

The main objective of this chapter is to derive a fast approach to find good-quality
solutions to the instantaneous functional split selection problem (FSSP), in which all
problem parameters are fixed. For this, we need to decide on the objective function
that we want to optimize by changing the centralization and flow vectors. In Chap-
ter 3, we argue that a mobile operator is interested in minimizing the total cost of
managing the network, which comprises readiness and action cost. Action cost is,
nonetheless, only present when the network state changes, so that it is only useful
for optimizing sequential adaptations over time. As a result, in this chapter we focus
on minimizing the readiness cost, whereas action cost is fully taken into account in
Chapter 6.
We can intuitively identify two components contributing to the instantaneous readi-
ness cost K(s) of any communication network resulting from operating in state s: the
operating cost Koper (s) and the revenue Krev (s) associated to providing service to the
users. As a result:
K(s) = Koper (s) + Krev (s) (2)
Revenue Krev (s) is modeled as negative cost, therefore Krev (s) ≤ 0 ∀s and K(s) < 0
if and only if the revenue is higher than the operating cost. Note that, since we are
dealing with instantaneous cost, we can drop the dependency of the readiness cost on
the demand d, as this can be interpreted as fixed problem parameters instead of input
variables. Thus, in this chapter we use simply K(s), Koper (s), and Krev (s) instead of
K(s, d), Koper (s, d), and Krev (s, d). In general, in this chapter we focus on the influence
of state s in the RAN cost and performance. In Chapter 6, we recover the explicit
dependency on demand d to address the problem of time-dependent functional split
adaptation.
The operating cost can be computed in a relatively straightforward manner, since
there are already models in the literature for such a task, as we discuss in Sec. 4.5.
Modeling the revenue as a function of the network state is, however, considerably
more challenging, since operators rarely disclose any details regarding this internal
information. Nonetheless, we can safely assume that the revenue should be an in-

67
4. Optimal Functional Split Selection in 5G Radio Access Networks

creasing function of the user performance, which reflect the utility that users obtain
from a given collection of user throughputs. Consequently, we decouple the problem
of selecting the optimal functional split into three stages, since they exhibit very dis-
tinct features, as we show in the following. First, we tackle the FSSP with the intention
of maximizing this user performance, which is addressed in this section. Second, we
solve the FSSP so as to minimize the operating cost, which is the matter of Sec. 4.5.
Finally, we combine both approaches into a single formulation to minimize the total
readiness cost, which is shown in Sec. 4.6.

4.4.1. Proportionally-fair formulation

From the point of view of user performance, we could set the objective of selecting
the optimal functional split to maximizing the data rate of all users, which can be
accomplished by smartly reducing interference. The data rate ρu (a) achieved by UE u
for a centralization vector a can be calculated as:

ρu (a) = µu ηu (a), (3)

where µu is the bandwidth allocated to UE u and ηu (a) denotes its downlink spectral
efficiency. We can use Shannon’s formula to estimate the latter as follows [Sha49]:
 
pu
ηu (a) = log2 1 + , (4)
ς + Iu (a)

where pu is the signal power received by UE u from its serving gNB hu , Iu (a) is the
experienced interference as defined in (1), and ς is thermal noise power (assumed
constant over all UEs). Using (4), we could formulate an optimization problem to find
the centralization vector a∗ that maximizes the sum of user data rates:
U
X
∗
a = arg max ρu (a), (5)
a∈A
u=1

where A is the set of vectors a whose required link capacities are supported by the
fronthaul network. From the operator’s point of view, obtaining a∗ as defined in (5)
would lead to an efficient use of the network resources, since it guarantees working
at maximum capacity. However, (5) may lead to unfair situations, since data rates of
users with good signal-to-interference-and-noise ratio (SINR) may be prioritized over
those with poor SINRs [SY14]. In order to prevent that, we have to maximize the sum
of the individual utilities associated with each data rate:
U
X
∗
a = arg max U(ρu (a)), (6)
a∈A
u=1

68
 4.4. Performance-maximizing FSSP

where U( · ) is an arbitrary rate utility function. This function should reflect some level
of risk aversion, that is, small data rate increments are more valuable when the initial
data rate is low than when the data rate is high, thus leading to a concave shape. This
risk aversion can be also interpreted as fairness, since it translates into giving higher
priority to those UEs facing bad interference situations.
There are several manners of achieving fairness in communication networks. Two of
the more usual ones are max-min fairness, in which the minimum data rate is maxi-
mized [BGH92], and proportional fairness, which maximizes the sum of relative data
rate increments with respect to the capabilities of each UE [MMD91]. In practice,
max-min fairness is more strict than proportional fairness, since it may give absolute
priority to those UEs featuring lower rates [SL05]. Therefore, in this work we select
proportional fairness to define our rate utility function, which is indeed common in
problems dealing with dynamic rate allocation, such as in time-frequency scheduling
[KH05]. According to [KMT98], a rate vector ρ = [ρ1 · · · ρU ] is a proportionally-fair
allocation of data rates if and only if
U
X ρ0u − ρu
≤0 (7)
u=1
ρu

for any other feasible rate vector ρ0 = [ρ01 · · · ρ0U ]. This requirement can be easily
fulfilled by using a logarithmic utility function U(ρ) = log(ρ), as it is proven in [SL05].
Thus, we define the proportionally-fair optimal centralization vector a∗ as:
U
X
a∗ = arg max log (ρu (a)) (8)
a∈A
u=1
XU
= arg max log (µu ηu (a)) (9)
a∈A
u=1
XU
= arg max log(ηu (a)) . (10)
a∈A
u=1

Note that we can remove µu from the formulation since it becomes an adding term that
does not depend on a. We refer to the problem of finding the centralization vector a∗
as defined in (10) as the proportionally-fair FSSP.
The objective function in (10) is directly related to the geometric mean ηe(a) of the spec-
tral efficiency over all UEs, which is defined as:

U
!1 U
!
U
Y 1 X
ηe(a) , ηu (a) = exp log (ηu (a)) . (11)
u=1
U u=1

Therefore, since the exponential function is monotonically increasing, an equivalent

69
4. Optimal Functional Split Selection in 5G Radio Access Networks

definition of the optimal, proportionally-fair centralization vector is:

a∗ = arg max ηe(a). (12)

a∈A

This equivalence allows us to use ηe(a) as performance indicator when comparing al-
ternative solutions using the same units as ηu (a).
From (4) and the definition of a∗ in (10), we can obtain a closed-form expression for
the objective function of the FSSP. Regarding the constraints of the FSSP, it is clear that
the validity of a solution a is limited by the topology and capacity of the fronthaul
network. In other words, a solution a is valid if and only if there exists a vector of
flows f that satisfies the flow requirements for every gNB, as mentioned in Sec. 2.3.1,
and can be implemented on the fronthaul network without exceeding the capacity of
any link. As a result, we can formulate the FSSP as follows:
U
!!
X pu
max log log2 1 + , (P4a)
ς+ G
P
g=1 iu,g · И(min(ahu , ag ))
a,f
u=1

subject to

X X 0
 ∀n ∈ N \ {n0 , ng }
feg − g
fe = ν(ag ) for n = n0 ∀g ∈ G, (P1b)
−

e∈E+ (n) e∈E (n) 
−ν(ag ) for n = ng
G
X
feg ≤ ϑe ∀e ∈ E, (P1c)
g=1

feg ≥ 0 ∀e ∈ E, ∀g ∈ G, (P1d)
G
s∈M . (P1e)

where E+ (n) is the set of edges leaving node n, and E− (n) is the set of edges entering
node n. As introduced in Sec. 2.3.2, constraint (P1b) is the flow conservation constraint,
which ensures that the flow leaving the CU and entering the DU is ν(ag ) for each gNB
g. In addition, (P1c) enforces the link capacity constraint for each link e.
The FSSP as formulated in (P4) is a mixed integer non-linear problem (MINLP), which
are, in general, NP-Hard. Moreover, the non-standard expression of the objective
function (P4a) prevents the direct utilization of state-of-the-art techniques. In order
to make it more tractable, we present two reformulations that simplify the problem
structure at the expense of introducing additional variables.
We start with the following variable change, which allows us to replace the discrete
functions ν( · ) and И( · ) by 0-1 polynomial expressions:
(
0 if ag < m
bm
g , ∀m ∈ M,
b (13)
1 if ag ≥ m

70
 4.4. Performance-maximizing FSSP

b , {1, ..., M − 1} For compactness, we define

where M
−1
bg , b1g · · · bM
 
g ∀g ∈ G (14)

and
b , [b1 · · · bG ] . (15)
For example, we can use (13) to convert ag = 2 into bg = [1, 1, 0] when M = 4. There is
a bijection between bg and ag ∀g ∈ G, and thus a bijection between b and a, by means
of the following reciprocal conversion:
M
X −1
ag = bm
g . (16)
m=1

As a result, b is an alternative, equivalent representation of centralization vector a,

which can be also expressed as:

a = b · (IG ⊗ 1M −1 ), (17)

where Ig is the identity matrix of size g, 1m is an all-ones column vector of size m,

and ⊗ is the Kronecker product. The purpose of this variable change is to convert the
integer variables ag into binary variables in a particularly useful fashion. In fact, (13) is
not the conventional manner of performing an integer-to-binary conversion in integer
programming, which usually consists in the binary representation of the numbers
from 0 to M and thus requires dlog2 (M )e new variables per original variable. Instead,
conversion (13) requires M − 1 new variables, but this increment in the number of
additional variables is very limited (since M ≤ 8 in real deployments [Döt+13]), and
m0
it is compensated by its useful implications. Namely, since it follows that bm g ≥ bg if
and only if m ≤ m0 , the minimum operator between a variables can be converted into
a polynomial function of y variables:
M
X −1
min(ag , ak ) = bm m
g bk . (18)
m=1

This property can be exploited to rewrite the integer non-polynomial function

И(min(ag , ak )) as a 0-1 polynomial:
M
X −1
И(min(ag , ak )) = И(1) − %(m)bm m
g bk , (19)
m=1

where %(m) , И(m − 1) − И(m). As a result, we can reformulate the FSSP as:
  
U
X pu
max log log2 1 + P   , (P5a)
b,f
P G M −1 m m
u=1 ς + Iu − g=1 iu,g m=1 %(m)bg bhu

71
4. Optimal Functional Split Selection in 5G Radio Access Networks

subject to

X X 0
 ∀n ∈ N\{n0 , ng }
g g
fe − fe = νb(bg ) for n = n0 ∀g ∈ G, (P5b)

e∈E+ (n) e∈E− (n) 
−b ν (bg ) for n = ng
b1g ≥ b2g ≥ ... ≥ bgM −1 ∀g ∈ G, (P5c)
b ∈ {0, 1}G , (P5d)
(P1c), and (P1d),

where Iu = G
P
g=1 И(1)iu,g is the interference power received by UE u when the lowest
centralization level is in operation on its serving gNodeB, and function νb( · ) is defined
as follows:
M
X −1
νb(bg ) , ν(1) − (ν(m − 1) − ν(m)) bm
g , (20)
m=1

such that νb(bg ) = ν(ag ).

Formulation (P5) replaces the integer variables and discrete functions ν( · ) and И( · )
of (P4) by binary variables and polynomial functions. As a result, linearization tech-
niques can be now applied to improve the tractability of the FSSP. Namely, the product
of two bmg variables can be linearized via the following variable change:

m
vg,k , bm m
g bk , (21)

which can be enforced with additional linear inequalities [GW74]. This leads to the
following reformulation:
  
U
X pu
max log log2 1 + P   (P6a)
b,v,f
PG M −1 m
u=1 ς + Iu − g=1 iu,g m=1 %(m)v g,hu

subject to
m
2vg,k ≤ bm m
g + bk ∀m ∈ M,
b ∀k ∈ G, ∀g < k, g ∈ G, (P6b)
m
1 + vg,k ≥ bm
g + bk
m
∀m ∈ M,
b ∀k ∈ G, ∀g < k, g ∈ G, (P6c)
G
v ∈ {0, 1}(M −1)( 2 ) , (P6d)
(P1c), (P1d) and (P5b) − (P5d),

where  m
v , vg,k ∀m ∈ M,
b ∀k ∈ G, ∀g < k, g ∈ G. (22)
Note that constraints (P6b)–(P6d) enforce (21), so that we can replace b, and thus
the
centralization vector a, with v. The number of additional v variables is (M − 1) G2 =
O(G2 ), as one additional variable is required for every pair of gNBs and consecutive

72
 4.4. Performance-maximizing FSSP

splits. In Section 4.4.2, we exploit the characteristics of the network to reduce the
number of these additional variables.
Formulation (P6) is still a MINLP, but its simpler objective function admits further
analysis. Indeed, we observe that the continuous relaxation of the objective function
is convex on v, but since the FSSP is a maximization problem, this implies that we
are in the realm of concave optimization. Thus, there may be multiple local maxima,
making the problem hard to tackle. It is still possible to use exact global optimiza-
tion techniques for concave MINLPs, such as those presented in [Hor86], but these
mainly consist on applying branch-and-bound or branch-and-cut algorithms, whose
convergence time may be high. Since the FSSP is a real-time problem, we opt instead
for deriving increasingly simpler approximations to the original FSSP, until a suitable
approach within the speed-quality trade-off is found.

4.4.2. Fractional approximations

The main obstacles when tackling formulations (P4)–(P6) are the logarithmic func-
tions, which prevent the application of simplifying reformulations. Fortunately, we
can exploit the slow growth rate of these functions, as its combination can be very
well approximated by a rational function:

ℵrat
log(log2 (1 + σ)) ≈ + ‫ג‬rat . (23)
irat + σ
Parameters ℵrat , irat and ‫ג‬rat can be obtained from rational fitting within the desired
interval. In our case, we choose the interval 0.1 ≤ σ ≤ 100, that is, an SINR ranging
from −10 dB to 20 dB. After applying (23) to (P6) and simplifying the expression, we
obtain the following reformulation:
  
U
X pu
arg max log log2 1 + P   (P6a)
b,v,f
P G M −1 m
u=1 ς + Iu − g=1 iu,g m=1 %(m)v g,hu
U
X ℵrat
≈ arg max pu + ‫ג‬rat . (24)
b,v,f irat + PG PM −1
u=1 ς+Iu − (
g=1 iu,g
m
m=1 %(m)vg,hu )
 P 
PG M −1 m
U
X ℵrat ς + Iu − i
g=1 u,g m=1 %(m)v g,hu
= arg max P  (25)
b,v,f
P G M −1 m
i
u=1 rat (ς + Iu ) + p u −i rat i
g=1 u,g m=1 %(m)v g,hu
 P 
PG M −1 m
rat ς + Iu − g=1 iu,g m=1 %(m)vg,hu + pu − pu
XU i
= arg max PG  PM −1  (26)
b,v,f m
u=1 irat (ς + Iu ) + pu −irat g=1 iu,g m=1 %(m)v g,hu

73
4. Optimal Functional Split Selection in 5G Radio Access Networks

U
X pu
= arg min P  (27)
b,v,f
PG M −1 m
u=1 irat (ς + Iu ) + pu −irat g=1 iu,g m=1 %(m)vg,hu

Note that in (26), constant ℵrat is replaced with irat so as to eventually remove the
polynomial from the numerator, since this replacement does not affect the location of
the optimal point. As a result, we formulate the fractional FSSP as an approximation
to the original proportionally-fair FSSP as:
U
X pu
min G M −1
! (P7a)
b,v,f
u=1
X X
m
irat (ς + Iu ) + pu −irat iu,g %(m)vg,hu
g=1 m=1

subject to
(P1c), (P1d), (P5b) − (P5d), and (P6b) − (P6d).

Problem (P7) is now a multiple-ratio fractional mixed-integer optimization problem,

which can be tackled with state-of-the-art techniques. Indeed, since the continuous
variables f do not appear on the objective function, we can directly apply existing
techniques to reformulate it into an MILP. We present two such techniques: the Li-
Wu-Tawarmalani transformation and the Borrero-Gillen-Prokopyev transformation.

Li-Wu-Tawarmalani transformation

The Li-Wu-Tawarmalani transformation (in short, LWT transformation) reformulates

a 0-1 multiple-ratio fractional program into an MILP by introducing continuous vari-
ables via the following variable changes [Li94; Wu97; TAS02]:

1
wu , G M −1
! ∀u ∈ U, (28)
X X
m
πu − irat iu,g %(m)vg,h u
g=1 m=1

xm m
u,g , wu vg,hu ∀m ∈ M,
b ∀u ∈ U, ∀g ∈ G, (29)

where πu , irat (ς + Iu ) + pu . For compactness, we define the following vectors to

contain the new LWT variables:

w , [wu ] ∀u ∈ U, (30)
x , xm
 
u,g ∀m ∈ M.
b (31)

74
 4.4. Performance-maximizing FSSP

Identities (28) and (29) can be enforced by additional constraints, resulting in the fol-
lowing equivalent formulation:
U
X
min pu w u (P8a)
b,v,w,x,f
u=1

subject to
−1
G M
!
X X
m
πu wu − irat iu,g %(m)vg,hu
=1 ∀u ∈ U, (P8b)
g=1 m=1

Wu− vg,h
m
u
≤ xm
u,g ≤ Wu+ vg,h
m
u
∀m ∈ M,
b ∀g ∈ G, ∀u ∈ U, (P8c)
−
xm m


u,g ≤ wu + Wu vg,hu − 1 ∀m ∈ M,
b ∀g ∈ G, ∀u ∈ U, (P8d)
xm + m


u,g ≥ wu + Wu vg,hu − 1 ∀m ∈ M,
b ∀g ∈ G, ∀u ∈ U, (P8e)
(P1c), (P1d), (P5b) − (P5d), and (P6b) − (P6d),

where Wu− and Wu+ are lower and upper bounds for wu , respectively, which can be
obtained straightforwardly by setting the v variables in (28) to all zeros and all ones:

1
Wu− = ∀u ∈ U, (32)
πu
1
Wu+ = P  ∀u ∈ U. (33)
PG M −1
πu − irat g=1 iu,g m=1 %(m)

Formulation (P8) is an MILP that requires U + 4(M − 1)GU new constraints, U addi-
tional w variables and (M −1)(G−1)U additional x variables with respect to (P7). Note
that, in the general case, the number of required x variables would be (M − 1) G2 U

as these variables originate from the product of v and w variables. However, in our
case it is clear that the interference coefficient corresponding to a triple hu, g, ki, u ∈ U,
g, k ∈ G is zero unless g = hu or k = hu and g 6= k, hence we can remove the variables
indexed by those triples. As a result, the number of variables of this reformulation
grows with O(G2 ), assuming that U scales linearly with G [3GP20b], instead of O(G3 ).

Borrero-Gillen-Prokopyev transformation

The Borrero-Gillen-Prokopyev transformation (BGP transformation) is a recent im-

provement on the LWT transformation, which aims at reducing the number of re-
quired variables and constraints by approximating all coefficients in the objective
function with integers [BGP16]. This is accomplished by introducing the three new
sets of variables with respect to (P7). In order to avoid notational clutter, we reuse
and redefine variables w and x, which are firstly defined for the LWT transforma-
tion, since BGP and LWT transformations are independent approaches that cannot be

75
4. Optimal Functional Split Selection in 5G Radio Access Networks

combined.
Variables w , [wu ] ∀u ∈ U are redefined as:

ð
wu , PG PM ∀u ∈ U, (34)
ou,0 + g=1 m=1 oqu,g vg,h
m
u

where
 
ðirat (ς + Iu )
ou,0 = ð+ , (35)
pu
 
q irat ðiu,g %(m)
ou,g = − ,
pu

ð is a constant factor used to scale the integer coefficients into the desired range, and
b · e represents the rounding operation to the nearest integer. In addition, new binary
variables x , [xu,l ], xu,l ∈ {0, 1} are defined via the following identity:
Lu
X G X
X M
2l−1 xu,l = Ou + oqu,g vg,h
m
u
∀u ∈ U, ∀l ∈ {1, ..., Lu }, (36)
l=1 g=1 m=1

where
G X
X M
Ou , − oqu,g (37)
g=1 m=1

and
Lu , blog2 (Ou )c + 1. (38)
Finally, new variables y , [yu,l ] are defined as:

yu,l , xu,l wu ∀u ∈ U, ∀l ∈ {1, ..., Lu }. (39)

The resulting reformulation, including additional constraints to enforce (34)–(39), is

the following [BGP16]:
X U
min wu (P9a)
b,v,w,x,y,f
u=1

subject to
Lu
X
(ou0 − Ou )wu + 2l−1 xu,l = −ð ∀u ∈ U, (P9b)
l=1
G X
X M Lu
X
m
ou,g vg,hu
− 2l−1 xu,l = −Ou ∀u ∈ U, (P9c)
g=1 m=1 l=1

Wu− xu,l ≤ yu,l ≤ +

Wu xu,l ∀u ∈ U, ∀l ∈ {1, ..., Lu }, (P9d)

76
 4.4. Performance-maximizing FSSP

yu,l − wu ≤ Wu− xu,l − Wu− ∀u ∈ U, ∀l ∈ {1, ..., Lu }, (P9e)

yu,l − wu ≥ Wu+ xu,l − Wu+ ∀u ∈ U, ∀l ∈ {1, ..., Lu }, (P9f)
(P1c), (P1d), (P5b) − (P5d), and (P6b) − (P6d),

where Wu− and Wu+ are lower and upper bounds to wu , respectively, whose value can
be computed as:

−ð
Wu− = (40)
ou,0
−ð
Wu+ = . (41)
ou,0 − Ou
P P
Formulation (P9) is an MILP that requires 2U +4 u∈U L Pu new constraints, U + u∈U Lu
additional continuous variables (vectors w and
P y) and u∈U Lu additional binary vari-
ables (vector x) with respect to (P7). Since u∈U Lu = O(U log(G)) [BGP16] and as-
suming again U = O(G), this implies that the number of additional variables and
constraints compared to (P7) grows with O(G log(G)), at the expense of losing accu-
racy in the problem coefficients. Nonetheless, the overall size of the problem instances
still grows with O(G2 ), due to the presence of variables v.

Punctured transformations

The fractional reformulation (P7) relies on the addition of v variables to be tractable

by the LWT and the BGP transformations. These variables replace the product of b
variables by single binary variables, which eventually enables these MILP reformula-
tions. As there must be a vg,k variable for each pair of gNBs hg, ki, their number grows
quadratically with the number of gNBs G.
However, in our problem not every pair of gNBs is worth considering. The interfer-
ence between two gNBs that are far apart is negligible, so any variable modeling it
contributes little to the overall solution. Knowing this fact, we can remove unnec-
essary variables so that the problem size is reduced without noticeably affecting the
optimal solution. To do so, we define ı̂g,k as the combined interference caused by gNBs
g and k: X X
ı̂g,k = iu,k + iu,g (42)
u∈Hg u∈Hk

where
Hg = {u ∈ U | hu = g, hu , g ∈ G} (43)
is the set of the UE indices served by gNB g. Now we sort coefficients ı̂g,k and remove
those gNB pairs hg, ki whose combined interference is below a configurable thresh-
old. For instante, in our experiments, we remove those gNB pairs with the smallest
ı̂g,k such that their addition contributes less than 5% to the total interference. This re-
moves the related vg,k variables and all additional variables and constraints that are

77
4. Optimal Functional Split Selection in 5G Radio Access Networks

defined from them, hence simplifying the problem. Since removing these variables
may impact the performance of the obtained solutions, we evaluate them separately
for the LWT and BGP transformations, and refer to them as punctured LWT and BGP
transformations, respectively.

4.4.3. Quadratic formulation

Instead of the approximation shown in (23), we can consider a simpler fractional ap-
proximation of the log(log2 (1 + σ)) function:

isim
log(log2 (1 + σ)) ≈ ℵsim − . (44)
σ
This approximation is less tight than (23), but in return it produces a much simpler
problem formulation. Indeed, after combining (44) with (P5), we arrive at the follow-
ing equivalent problem:
M −1
U X G
!
X iu,g X
max %(m)bm m
g bhu , (P10a)
b,f
u=1 g=1
p u m=1

subject to

(P1c), (P1d) and (P5b) − (P5d),

As this reformulation is a quadratic integer problem (QIP), we refer to it as the quadratic

formulation. Its standard form enables the use of off-the-shelf solvers to tackle it.
Nonetheless, its structure can be further exploited to reformulate it into a simple MILP.
For this, we first introduce these new coefficients:
  

 X iu,k X iu,g 
m
%(m)  + if g 6= k,
`g,k , u∈Hg
p u
u∈Hk
p u ∀m ∈ M,
b ∀g, k ∈ G (45)


0 if g = k,


These coefficients allow for the following alternative reformulation of (P10):

M XG G
!
X X
max bqg `m m
g,k bk , (P11a)
b,f
m=1 g=1 k=1

subject to

(P1c), (P1d) and (P5b) − (P5d).

78
 4.4. Performance-maximizing FSSP

Finally, we define a new set of variables

r , rgm
 
∀g ∈ G, ∀m ∈ M,
b (46)

by means of the following variable change:

G
!
X
rgm , ygm `m m
g,k yk ∀m ∈ M,
b g ∈ G, (47)
k=1

which can be enforced into our optimization problem by adding three new sets of
constraints, as follows [Glo75]:
M
X −1 X
G
max rgm (P12a)
b,r,f
m=1 g=1

subject to

0 ≤ rgm ≤ Rgm bm
g ∀m ∈ M,
b ∀g ∈ G, (P12b)
G
X
rgm ≥ `m m m m
g,k bg − (1 − bg )Rg ∀m ∈ M,
b ∀g ∈ G, (P12c)
k=1
X
rgm ≤ `m m
g,k bg ∀m ∈ M,
b ∀g ∈ G, (P12d)
k
(P1c), (P1d), and (P5b) − (P5d).
PG
where Rgm , m
k=1 `g,k .

Formulation (P12) requires (M − 1)G additional variables and 4GM additional con-
straints with respect to formulation (P5). As a result, problem instances grow with
O(G), leading to substantially smaller problems when compared to the previous MILP
formulations, which grow with O(G2 ). The drawback of this approach is that the ap-
proximation (44) is less tight than (23), which may impact the quality of the solutions.

4.4.4. Heuristic approaches

The previous reformulations approximate the proportionally-fair FSSP into MILPs,

which can be tackled by dedicated off-the-shelf software. Nonetheless MILPs are still
NP-Hard, thus exact techniques to solve them may be slow, exhibiting exponential
worst-case performance. This is actually shown in Sec. 4.7.2, along with other re-
lated measurements. Since we are interested in solving the problem as fast as possible
so as to dynamically adapt to changes in the interference situation, we propose two
heuristics with the intention of finding good-quality solutions with low convergence
time. The first technique exploits the correlation between the interference caused by
a gNB and its centralization level. The second technique is a local-search approach to

79
4. Optimal Functional Split Selection in 5G Radio Access Networks

improve the solutions of the quadratic reformulation.

Heuristic 1

As the objective of centralizing gNBs is to mitigate interference, it intuitively follows

that, given an optimal solution a∗ , the gNBs with the highest centralization levels may
tend to be those causing the most interference. From this, we can derive a heuristic
rule that assigns the centralization level of a gNB g based on the total interference Ibg
that it causes to all UEs in the absence of interference-mitigation:
U
X
Ibg , iu,g . (48)
u=1

Such a heuristic must not only produce a solution that is proportional to Ibg ∀g ∈ G,
but it must also satisfy constraints (P1b)–(P1d) and be as centralized as possible.
We propose the following method to accomplish these objectives. First, we define the
accumulated centralization level A
b given a (possibly infeasible) solution a as:

G
X
A
b, ag . (49)
g=1

An upper bound A b+ ≥ A b for all feasible a can be obtained by solving the follow-
ing MILP, which maximizes the accumulated centralization level without taking into
account the resulting cost or performance:
G
X
b+ = max
A ag (P13a)
a,f
g=1

subject to

(P1b) − (P1d).

Note that (P13a) only depends on the fronthaul network configuration via constraints
(P1b)–(P1d), therefore it can be solved offline before the network is put into operation.
In addition, a lower bound A b− ≤ A b for all feasible a can be also easily found, for
example as:  
− mine∈E (ϑe )
A =
b , (50)
И(M )
which is the maximum number of fully-centralized gNBs that can be supported by
the weakest link.
Now, given an initial guess of Ab within these bounds, we need a method to assign a
value to each ag ∀g ∈ G in accordance to the values of Ibg . We use the Webster/Sainte-

80
 4.4. Performance-maximizing FSSP

Laguë method (WSL method) to do this, an algorithm originally designed to propor-

tionally allocate seats in party-list voting systems [Gal91]. This method is selected
since it preserves the proportionality of the original interference levels better than al-
ternative approaches [Sch+03], yet it is simple to implement. In a nutshell, the WSL
method starts from a = 0, where 0 is an all-zeros vector, finds the index

g ∗ = arg max Ibg (51)

g∈G

belonging to the gNodeB the maximum interference Ibg , increments ag by 1 to increase

its centralization level, updates Ibg to 2aIgg+1 (or to −∞ if ag = M ) and repeats the process
until the desired value of Ab is achieved.
After running the WSL method, we have a candidate centralization vector a for the
desired A.
b At this point, we can solve the following minimization problem to find the
corresponding flow vector f :
XG X E
min feg (P14a)
f
g=1 e=1

subject to

(P1b) − (P1d).

Note that vector a is not present in the objective function, but only in constraint (P1b).
This problem is a linear program (LP) with E variables, and thus it can be tackled
very efficiently by modern solvers. However, it may happen that our initial guess
of Ab yields a vector a such that (P14) is infeasible. In that case, we need to find a
different value of A
b and try again until a feasible value of A
b is found. This process can
be performed efficiently by exploiting the properties of our objective function and the
WSL method via a binary search of the largest feasible A. b
As И( · ) is a monotonically decreasing function, it can be trivially proven that ηe(a) ≥
ηe(a0 ) whenever ag ≥ a0g ∀g ∈ G, and vice-versa. In words, this means that increasing
(decreasing) the centralization level of any gNB can only increase (decrease), on aver-
age, the mean spectral efficiency, as intuitively expected. Similarly, given constant Ibg
∀g ∈ G and two accumulated centralization levels A b and A b0 such that A b>A b0 , it can
0
be easily shown that the resulting centralization vectors a and a yielded by the WSL
algorithm fulfill ag ≥ a0g ∀g ∈ G. Finally, it is also clear that if a is feasible, then a0 must
be as well. We conclude that there is a single maximum value of A b such that for all
Ab0 < A b the WSL method returns feasible but lower-performance solutions, and for all
Ab00 > Ab the WSL method returns infeasible solutions. In the light of the above, we can
implement a binary search of the highest feasible value of A. b
The summarized operation of Heuristic 1 is shown in Algorithm 1. Since this algo-
rithm deals with the centralization vector a and the flow vector f directly, its com-
plexity grows with O(G).

81
4. Optimal Functional Split Selection in 5G Radio Access Networks

Algorithm 1: Heuristic 1 (Webster/Saint-Laguë method and binary search).

Input: Ab+ , Ab− , Ibg ∀g ∈ G
Output: a, f
1 ag ← 0 ∀g ∈ G

2 Ab←A b+
3 repeat // Binary search
4 repeat // WSL assignment
∗
5 g ← arg maxg {Ig | g ∈ G}
b
6 ag ∗ ← ag ∗ + 1
7 if ag∗ < M then
I ∗
8 Ig∗ ← 2xgg∗ +1
9 else
10 Ig∗ ← −∞
11 end
PG
12 g=1 ag = A
b
13 if (P14) feasible then // Feasibility check
14 Ab− ← A b
15 f ← f ∗ (a) as in (P14).
16 else
17 Ab+ ← A b−1
18 end
19 A+
b =A b−

Heuristic 2

This heuristic exploits the properties of the FSSP by using local search so as to improve
solutions provided by a previous algorithm. We start with an initial solution s = ha, f i
provided by the quadratic approach (P12) and we compute the following parameters
from it:
1 X b
Im = Ig (52)
|Gm | g∈G
m

where Gm = {g ∈ G | ag = m} ∀m ∈ M. The value of I m is the average interference

power caused by those gNBs whose centralization level is m. We then calculate devi-
ations ∆Ibg = Ibg − I ag ∀g ∈ G, which represent how far the total interference caused
by gNB g is from the average value among those with the same centralization level.
Based on these deviations, we can identify two types of gNBs: those producing rel-
atively high interference (which may benefit from a higher centralization level) and
those producing relatively low interference (which might accept a lower centraliza-
tion). Consequently, we select pairs of gNBs [k, k 0 ] ∈ {1, ..., M − 1} × {2, ..., M } such
that k belongs to the former type and k 0 belongs to the latter, and generate a new

82
 4.5. Operating-cost-minimizing FSSP

candidate solution a0 whose elements are as follows:


ag + 1 if g = k,

0
ag = ag − 1 if g = k 0 , (53)

ag otherwise.


Then, the mean spectral efficiency ηe(a0 ) of the solution is evaluated. If ηe(a0 ) > ηe(a), the
feasibility of a0 is evaluated with (P14). If a0 is both feasible and its spectral efficiency
is higher than that of a, it is taken as new initial solution and the procedure repeats
until no better solution is found.

4.5. Operating-cost-minimizing FSSP

In this section, we formulate the FSSP that focuses on minimizing the cost of operating
the network in a given state, regardless of the performance experienced by the UEs.
This problem is, thus, the counterpart to the performance-maximizing FSSP, both of
which are eventually combined in Sec. 4.6.
We base on the model presented in [GS+18b] to characterize the operating cost Koper (s)
of a mobile RAN featuring a configurable functional split. This operating cost can be
divided into three components:

Koper (s) = Kinst + Kcomp (a) + Krout (f ). (54)

The first component Kinst is the cost of instantiating functions at the DUs and CU,
which can be calculated as:

Kinst = (Kinst,CU + Kinst,DU ) G, (55)

where Kinst,CU is the cost of instantiating the gNB functions at the CU and Kinst,DU is
the cost of instantiating the gNB functions at the DU. The second component, the
computational cost Kcomp (a) of running centralized and decentralized functions, can
be expressed as a non-linear function of the state vector a as follows:
G
X

Kcomp (a) = Kcomp,CU (ag )γCU + Kcomp,DU (ag )γDU ρg , (56)
g=1

where Kcomp,CU (a) and Kcomp,DU (a) denote the computational effort in reference cores–
seconds per Gb/s (RC · s/Gb/s) required to process traffic at the CU and DU (respec-
tively) with centralization level a, γCU and γDU are the cost per computational effort at
the CU and DU (respectively) in ncu per RC · s, and ρg is the downlink mobile traffic
at gNB g in Gb/s. In order to be consistent with previous research, a RC · s is defined
as utilizing a single core of an Intel i7-4770 at 100% load during 1 second [GS+18b]. Fi-
nally, the third component, the cost Krout (f ) of routing the resulting flows is calculated

83
4. Optimal Functional Split Selection in 5G Radio Access Networks

as: X
Krout (f ) = ωfeg (57)
e∈E+ (n g)

where ω is the average cost in ncu per Gb/s of carrying traffic on a link and E+ (n) is
the set of edges leaving node n.
Therefore, the operating-cost-minimizing FSSP can be alternatively formulated as:

min Kinst + Kcomp (a) + Krout (f ) (P15a)

a,f

subject to

(P1b) − (P1d).

Problem (P15) is non-linear owing to the non-linearity of Kcomp (a), as it is evident

from (56). Nonetheless, we can replace again variable vector a with variable vector b
as shown in (13)–(16) to obtain a linear function of b. In order to do this we define the
following cost function:
G M
X X −1  
K
b comp (b) , K
b comp,0 + K b comp,DU (m)γDU ρg bm ,
b comp,CU (m)γCU + K (58)
g
g=1 m=1

where
G
X

K
b comp,0 , Kcomp,CU (0)γCU + Kcomp,DU (0)γDU ρg , (59)
g=1

and
b comp,CU (m) , Kcomp,CU (m) − Kcomp,CU (m − 1),
K (60)
b comp,DU (m) , Kcomp,DU (m) − Kcomp,DU (m − 1).
K (61)
It can be easily shown that K
b comp (b) = Kcomp (a) as long as (13) holds, but K
b comp (b) is
now a linear function of b.

Koper (s) = Kinst + K

b comp (b) + Krout (f ). (62)

Therefore, the cost-minimizing FSSP can be alternatively formulated as:

min Kinst + K
b comp (b) + Krout (f ) (P16a)
b,f

subject to

(P1c), (P1d) and (P5b) − (P5d),

Problem (P16) is equivalent to that presented in [GS+18b], using an edge formulation

instead of a path formulation for flow modeling. Note that (P16) is an MILP that

84
 4.6. Readiness-cost-minimizing FSSP

utilizes the same variables as (P12), except for the absence of auxiliary w variables. As
a result, it not only can be tackled directly by off-the-shelf solvers, but its complexity
grows with O(G), leading to reasonably smaller problems than those resulting from
the LWT and BGP reformulations of the performance-maximizing FSSP.

4.6. Readiness-cost-minimizing FSSP

In this section, we combine the performance-maximizing and operating-cost-mini-

mizing approaches shown in previous sections into a single formulation that mini-
mizes the average readiness cost. That is, the objective now is to minimize the whole
readiness function, not its individual components:

min K(s) = min Koper (s) + Krev (s). (P17a)

s s

subject to

(P1b) − (P1d).

From (62) we know that we can express the operating cost Koper (s) as a linear function
of the centralization vector b and flow vector f . The average revenue Krev (s), however,
may lead to more complex expressions.
We can assume that Krev (s) is an increasing function of the proportionally-fair objec-
tive function, log(ηu (a)). The reason is that, if this is the case, maximizing this function
would also imply maximizing the revenue, which seems a reasonable approach for
any operator to follow. Formally, we can express this relationship as:
U
!
X
Krev (s) = K
b rev log(ηu (a)) , (63)
u=1

where Kb rev (η) is an increasing function of η. For notational consistency, we define this
alternative version of the spectral efficiency function using b variables instead of a
variables:
ηbu (b) , ηu (a) ⇐⇒ (13). (64)
We can then use (64) to formulate Krev (s) as a function of the alternative centralization
vector b: !
X U
Krev (s) = K
b rev log(b
ηu (b)) , (65)
u=1

After combining (P17), (P16), and (65), we formulate the following optimization prob-
lem: !
U
X
min Kinst + Kb comp (b) + Krout (f ) + K
b rev log(b
ηu (b)) (P18a)
b,f
u=1

85
4. Optimal Functional Split Selection in 5G Radio Access Networks

subject to

(P1c), (P1d) and (P5b) − (P5d).

Problem (P18) is, in general, a mixed-integer non-linear problem (MINLP), owing

to the potential non-linearity of K b rev ( · ). Nonetheless, if K
b rev ( · ) were indeed linear,
then (P18) could be approximated as an MILP using the same techniques presented in
Sec. 4.4. In order to appreciate this more clearly, we can derive the following approxi-
mation of log(b ηu (b)) after combining (4) and (23):

U  
X ℵrat
ηu (b)) ≈ ‫ג‬rat +
log(b U
u=1
irat
U
ℵrat X pu
− P  (66)
irat u=1 i (ς + I ) + p −i PG M −1 m
rat u u rat i
g=1 u,g m=1 %(m)v g,hu .

Note that (66) is just a linear transformation of the objective function (P7a), and thus if
Kb rev ( · ) were linear we could transform (P18) into an MILP by using the LWT or BGP
transformations. Alternatively, if we use (44) instead to approximate log(b ηu (b)), the
following expression arises:
U U M −1 G
X X ς + Iu isim X X m
ηu (b)) ≈ ℵsim U −
log(b isim + rg . (67)
u=1 u=1
p u 2 m=1 g=1

Once again (67) is a linear transformation of the objective function (P12a), which al-
ready belongs to an MILP. As a result, we conclude that if we model K b rev ( · ) as a linear
function, this would result in a readiness-cost-optimizing MILP of a size similar to
those presented in Sec. 4.4, thus it could be solved in the same manner. Unfortunately,
in general we cannot guarantee that this conversion is linear, since the relationship
between user performance and revenue characterizes a central part of the operator
business, which is rarely disclosed.
Nevertheless, there are arguments to justify that Kb rev ( · ) may be indeed a quasi-linear
function. In order to show this, we first base on the following identity:
U
X
log (ηu (a)) = U log(e
η (a)), (68)
u=1

where ηe(a) is the geometric mean of the spectral efficiency over all UEs as defined
in (11). Previous research has shown that the value of ηe(a) ranges from ηe(a) ≈ 2
b/s/Hz in RANs whose functional split is chosen to optimize performance, to ηe(a) ≈
1 b/s/Hz in non-optimized RANs [MAJK21]. This very limited range translates into
an almost linear relationship between η̌(a) and ηe(a), thus we can formulate the fol-

86
 4.6. Readiness-cost-minimizing FSSP

Figure 4.1.: Comparison between function U log(e

η ) and its best linear approximation
for U = 3000 UEs and 1 ≤ ηe ≤ 2.

lowing approximation:
U
X
ηe(a) ≈ ℵlin + ilin log (ηu (a)) . (69)
u=1

Parameters ℵlin and ilin can be obtained by linear fit, for instance by least-squares min-
imization. A depiction of this approximation for U = 3000 UEs is shown in Fig. 4.1.
Furthermore, the limited range of ηe(a) also implies that the average UE data rates ex-
perience a roughly twofold increase in performance-optimizing RANs with respect to
non-optimizing RANs. Consequently, there is little room for severe non-linearities in
Kb rev ( · ) for actual networks.

In any case, even if we assume that K b rev ( · ) is linear, we cannot provide a single func-
tion that fits all networks. Indeed, operators may utilize multiple commercial strate-
gies and these may change over time. Therefore, even if we knew the function con-
verting performance into revenue for one operator within a time period, it would
be difficult to extrapolate this information to other operators and time periods. The
approach we follow to overcome this modeling difficulty is to provide not a single
function but instead a family of performance-to-revenue functions, which are charac-
terized by two parameters. The first parameter is the reference revenue-to-operating-
cost ratio ξ:
Ke rev (eηref )
ξ=− ref
, (70)
Koper
where  
ηe − ℵlin
K
e rev (e
η) , K η )) ≈ K
b rev (U log(e b rev (71)
ilin
is a reformulation of the performance-to-revenue function that takes the geometric
mean of the spectral efficiency ηe as input, ηeref is the geometric mean of the spectral ef-

87
4. Optimal Functional Split Selection in 5G Radio Access Networks

ref
ficiency achieved in a reference scenario, and Koper is the operating cost at a reference
scenario. We select the scenario where all UEs are uniformly distributed over the cov-
ered area as the reference scenario. The reason is that, in such a scenario, the operating
cost and performance achieved by all optimization approaches are approximately the
same, regardless of whether they are operating-cost-minimizing or performance-max-
imizing [MAJK21]. This can be observed in the results shown in Sec. 4.7.3. In addition,
the spectral efficiency achieved in this scenario is a lower bound to that of clustered
scenarios. As a result, the revenue of all optimization approaches is also similar, yield-
ing a consistent reference point. For our experiments, we consider milestone values
of ξ ∈ {0.5, 1, 2}, meaning that the revenue at the reference scenario may be half, the
same, or twice as much as the operating cost.
The second parameter υ is the revenue growth rate, defined as the ratio between the
revenue when the spectral efficiency is twice that of the reference scenario and the
reference revenue. Formally, we formulate it as follows:

K
e rev (2eηref )
υ= , (72)
Ke rev (e
ηref )

Parameter υ reflects how the revenue grows as the spectral efficiency doubles, thus
characterizing the slope of the performance-to-cost function. We select example val-
ues υ ∈ {1.5, 2, 3}, reflecting that the revenue increases 50%, 100%, and 200% with
respect to the reference revenue when the current spectral efficiency is twice the refer-
ence spectral efficiency.
As a result, we can formulate the performance-to-revenue conversion function K e rev (e
η)
between the geometric average of the spectral efficiency ηe and the revenue as:
 
ref η
η ) ≈ ξKoper 2 − υ + (υ − 1)
e
K
e rev (e . (73)
ηeref

In Fig. 4.2, we show the shape of these functions for the selected range of parame-
ters. It is clear that, even though we assume that our performance indicator and its
associated revenue are linearly related, we provide multiple options of such linear re-
lationships so as to cover a wide range of cases. If the actual relationship happens to
be non-linear, our proposed conversion functions can still be used as either upper or
lower bounds.
Finally, after applying the performance-to-cost function (73) and (66) to (P18) and
some straightforward simplifications, we obtain the following readiness-cost-mini-

88
 4.6. Readiness-cost-minimizing FSSP

Figure 4.2.: Comparison between nine performance-to-revenue functions K η ) for

e rev (e
revenue-operating cost ratios ξ ∈ {0.5, 1, 2} and revenue growth rates υ ∈
{1.5, 2, 3}.

mizing FSSP that can be tackled by means of LWT and BGP formulations:

min K
b comp (b) + Krout (f )
b,v,f
U
ref ilin ℵrat X pu
− ξKoper (υ − 1) P 
ηeref irat u=1 i (ς + I ) + p −i P G
i M −1
%(m)v m
rat u u rat g=1 u,g m=1 g,hu .
(P19a)

subject to

(P1c), (P1d), (P5b) − (P5d), and (P6b) − (P6d). (P19b)

If we instead combine (73) and (P18) with the quadratic approximation shown in (67),
we arrive to this alternative formulation of the readiness-cost-minimizing FSSP, which
can be tackled by the quadratic approach described in Sec. 4.4.3:
M −1 X
G
b comp (b) + Krout (f ) + ξK ref (υ − 1) ilin isim
X
min K oper rgm (P20a)
b,r,f 2eηref m=1 g=1

subject to

(P1c), (P1d), (P5b) − (P5d), and (P12b) − (P12d).

As we show in the following sections, the quadratic approach usually outperforms the
LWT and BGP reformulations. Therefore, we take (P20) as the final, reference MILP

89
4. Optimal Functional Split Selection in 5G Radio Access Networks

whose solving yields the optimal centralization vector a∗ (after the appropriate con-
version from vectors b∗ and v∗ ) and flows vector f ∗ that minimizes K(s∗ ) = K(a∗ , f ∗ ).

4.7. Experimental evaluation

After having presented three different interpretations of the FSSP, including the fi-
nal formulation minimizing the average readiness cost, in this section we evaluate
the performance of these approaches. We first focus on evaluating the convergence
times of the alternative formulations, which allow us to preliminarily discard those
formulations that take an excessive amount of time. Then, we assess the spectral effi-
ciency, operating cost, and readiness cost that all approaches achieve. This allows us
to conclude that the final readiness-cost-minimizing formulation is indeed adequate
and also observe the differences with the approaches that only focus on either user
performance or operating cost.
We use a MATLAB simulator to generate the interference coefficients iu,g , required by
all presented approaches, based on simulated UE and gNB positions. Then, we eval-
uate the algorithms on operator-grade hardware consisting of six computing servers
and 48 CPU cores [Bas+17], by using Gurobi [Gur], a commercial optimization solver.
We depict observed values of convergence time and spectral efficiency by means of
boxplots, with the intention of representing their distribution. In order to provide as
much information as possible in little space, we use a compact version of standard
boxplots, whose interpretation is as follows: the central dot represents the median,
the box contains the data between the first and third quartiles, and the whiskers ex-
tend to the lowest and highest values contained in 1.5 times the inter-quartile range.
Occasionally, we show the (arithmetic) mean of the distribution as a horizontal bar.
For the sake of clarity, outliers are not shown.

4.7.1. Simulation setup

Simulated mobile coverage

In order to produce realistic instances of the FSSP, we follow the recommendations for
simulating dense urban scenarios as described in 3GPP TS38.193 [3GP20b]. According
to this specification, gNBs are divided into two layers: macro and micro layer. The
macro layer follows an hexagonal layout with an inter-site distance of 200 m. This
results in a density of around 29 macro gNBs per square kilometer. In addition, the
number of micro gNBs should be three times that of the macro gNBs. As a result,
the average cell density is roughly 115 gNBs/km2 . This implies that a RAN with
G = 300 gNBs covers an area of 2.6 km2 , which approximately corresponds to the
center of a medium-sized or large city. For reference, the area of the City of London is
approximately 2.9 km2 .

90
 4.7. Experimental evaluation

1250
1250 1250
1000

Distance (m)
1000
Distance (m)

1000

Distance (m)
750
750 750
500
500 500
250
250 250
0
0 0 500 1000 1500 0
0 500 1000 1500 0 500 1000 1500
Distance (m) Distance (m)
Distance (m)
10 50

1250 200
1250 1250
8 40
1000 1000 1000
Distance (m)

150
Distance (m)

Distance (m)
6 30
750 750 750
100
4 20
500 500 500
50
250 2 250 10 250

0 0 0 0 0 0
0 500 1000 1500 0 500 1000 1500 0 500 1000 1500
Distance (m) Distance (m) Distance (m)

(a) θ = 0.5 (b) θ = 0.75 (c) θ = 0.95

Figure 4.3.: Visualization of UE concentration index for 2000 UEs. The red dots on the
left figure represent the positions of each UE on the considered area, which
is shown in light blue. The right figure is a color map of the 2-dimensional
empirical distribution of the UEs, showing the UE density of each 50 × 50
m square bin.

Regarding the UEs, 3GPP recommends to consider 10 UEs per gNB on average when
simulating 5G dense urban scenarios [3GP20b], which results in 1150 UEs/km2 . This
means, for instance, that a RAN with G = 300 gNBs serves, on average, 3000 simul-
taneously active UEs. Since the UE distribution may impact the performance of the
algorithms, we derive a dedicate metric to model it. We divide the considered net-
work area into square bins of side 50 m. Then, we count the number of UEs in each
bin to compute its 2-dimensional empirical distribution. From this distribution, we
compute its Gini coefficient and use it as the UE concentration index θ. It is observed
that θ = 0.5 corresponds to a uniformely-distributed random distribution of UEs, as it
can be seen in Fig. 4.3a. Higher values correspond to higher UE clustering, (Fig. 4.3b
and Fig. 4.3c), with a maximum value of θ = 1 when all UEs are within the same bin.
Once the position of the gNBs and the UEs is generated, we compute the received sig-
nal and interference power by using the log-distance path model for urban scenarios
[SS10].

Fronthaul network

For our performance evaluation, we set the number of functional split options to M =
4 (such as C-RAN, Intra-PHY, MAC-PHY, PDCP-RLC). Based on the analytical and

91
4. Optimal Functional Split Selection in 5G Radio Access Networks

(a) ψ = 2 (b) ψ = 3.5 (c) ψ = 5

Figure 4.4.: Visualization of the average fronthaul network degree for the same gNB
distribution. Black lines represent links, red stars represent macro DUs,
red triangles are micro DUs, blue square are fronthaul switches, and green
diamonds are CUs.

experimental results shown in [PM17], we use the following interference-cancellation

vector: c = h1, 0.6, 0.2, 0.01i, such that there is no interference mitigation when using
the lowest centralization level (as with PDCP-RLC) and a cancellation factor of 20 dB
when using the highest centralization level. Note, however, that other values or split
options are also possible, since they do not affect the problem formulation.
Regarding the fronthaul capacity vector, we use the values r = h4, 8, 80, 160i Gb/s, as
provided in [3GP17]. In order to simulate realistic fronthaul network layouts, we fol-
low the descriptions provided in [GS+18a], which are based on real mobile networks
on Italy, Romania, and Switzerland. Accordingly, we set the maximum link capacities
to 0.5, 1 or 2 Tb/s (the higher value that makes full centralization infeasible, to prevent
trivial results). Furthermore, we simulate several types of fronthaul networks by con-
trolling the average fronthaul network degree ψ, defined as the ratio of links to fronthaul
switches. According to [GS+18a], the average degree of a typical fronthaul network
ranges from ψ = 2 (a tree graph) to ψ = 5. In Fig. 4.4 we show some exemplary layouts
resulting from varying these parameters.

Computing platform

After creating the interference coefficients and the fronthaul network with the MAT-
LAB simulator, we have all the required components to run our adaptation algo-
rithms. Since the convergence time of this algorithms is very relevant to decide on
their viability, we use an operator-grade hardware platform consisting of 48 Intel Xeon
E5 cores distributed over six computing servers [Bas+17]. As optimization solver, we
use the commercial Gurobi software [Gur]. This software is able to divide large MILP
instances and efficiently process them in parallel.

92
 4.7. Experimental evaluation

Parameter Value Units Source

Kinst,CU 1 ncu [GS+18b, Table I]
Kinst,DU 0.5 ncu [GS+18b, Table I]
Kcomp,CU (a) [1 1.8 3.4 5] RC · s/Gb/s [GS+18b, Table I]
Kcomp,DU (a) [4 3.2 1.6 0] RC · s/Gb/s [GS+18b, Table I]
γCU 0.017 ncu/RC [GS+18b, Table I]
γDU 1 ncu/RC [GS+18b, Table I]
ωDU Variable ncu/Gb/s [GS+18b, Table I]

Table 4.2.: Summary of operating cost parameters.

Operating cost model

In order to apply the operating cost model presented in Sec. 4.5, we need to select
appropriate values to represent realistic 5G RAN deployments. With the intention of
providing comparable results, we take the values presented in [GS+18b] as our main
reference, which are summarized in Table 4.2.

4.7.2. Convergence time

The applicability of the FSSP approaches presented in this chapter is heavily influ-
enced by their convergence times, that is, the time required for the approaches to
reach a (near-)optimal solution. This is due to the fact that our ultimate intention is to
dynamically solve the FSSP during runtime, which is addressed in Chapter 6. Conse-
quently, it is possible that the solutions yielded by the algorithms are outdated if their
convergence time is too long. Previous work has shown that mobile traffic is highly
variable, and may often sharply deviate from average patterns [TGD19]. In addition,
in certain region types (such as entertainment or transport areas), even the average
patterns may be fast-changing [Xu+16]. In fact, the analysis of recent mobile traffic
traces show that the user traffic experienced by a 5G RAN may abruptly change in
few minutes [MAK19b; PL15]. As a result, approaches with long convergence times
are probably not useful for dynamic adaptation. For our experiments, we configure
the solver to allow for a maximum convergence time of 15 minutes, since solutions
taking longer are unlikely to be usable. Indeed, in Chapter 6 we show that the conver-
gence time should actually be in the order of a few seconds for dynamic adaptation to
be feasible.
We first focus on the convergence times of the MILP reformulations of the perfor-
mance-maximizing FSSP. In Fig. 4.5 we show the distributions of the convergence
time of the five MILP reformulations (quadratic, LWT, BGP, and punctured LWT and
BGP formulations) as a function of the number of gNBs, which in turn reflects the total
size of the RAN. The solver is configured for a maximum running time of 15 minutes
with a relative gap tolerance of 0.01% and each boxplot represents at least 100 runs.

93
4. Optimal Functional Split Selection in 5G Radio Access Networks

Figure 4.5.: Convergence time of the MILP reformulations of the performance-maxi-

mizing FSSP.

We observe that the convergence times of unpunctured LWT and BGP formulations
reach the 15-minutes limit with only 40 gNBs, and with 50 gNBs all instances take
longer or equal than this limit. Assuming a cell density of 115 gNBs/km2 , this implies
that they may be suitable only for areas smaller than 0.4 km2 . The running time of the
punctured LWT and BGP formulations is noticeably smaller, although the 15-minute
limit is once again reached with only 60 or 70 gNBs, corresponding to an area of ca.
0.6 km2 . As a result, both punctured and unpunctured LWT and BGP formulations
can only be applied to very small deployments. Moreover, since they only offer a very
slight performance improvement with respect to the quadratic approach (as we show
in Sec. 4.7.3), these approaches are not suitable for actual RAN optimization.
In contrast, the convergence time of the quadratic approach remains below 1 s for
areas with G = 80 gNBs or less. In fact, in most runs the quadratic approach yields
a solution in less than 0.1 s even when G = 80. In Fig. 4.6 we show an expanded
range of gNBs and include Heuristics 1 and 2 in the comparison. We can see that the
almost all instances of the quadratic approach take less than 10 seconds to converge
for the whole considered interval. This includes the case of G = 400, which translates
into a covered area of 3.5 km2 and 4000 simultaneously active UEs, which is enough
to cover the densely populated centers of most cities. We can also observe that the
convergence time of Heuristic 2 is at first similar to that of the quadratic approach,
but when G ≥ 300 this heuristic takes over ten times longer to converge than the
quadratic approach. This is due to the fact that Heuristic 2 is a local search approach,
and thus it may fail to converge if the explored area is large and no better local minima
are found. Finally, all instances of Heuristic 1 converge in less than 1 s for the whole
shown range, being thus the fastest algorithm to solve the performance-maximizing
version of the FSSP. We conclude, thus, that the most promising algorithms in terms
of convergence time are the quadratic approach and Heuristic 1.

94
 4.7. Experimental evaluation

Figure 4.6.: Convergence time of the quadratic reformulation and Heuristics 1 and 2
for the performance-maximizing FSSP..

Owing to its good convergence time, we select the quadratic approach to tackle the
readiness-cost-minimizing version of the FSSP, that is, formulation (P20). In Fig. 4.7
we show at last a direct comparison of the convergence times of the quadratic ap-
proach to the performance-maximizing FSSP, the operating-cost-minimizing FSSP, and
the quadratic approach of the readiness-cost-minimizing FSSP. We observe that the
operating-cost-minimizing FSSP, as formulated in (P16), is the fastest of the three
approaches, converging in less than 0.1 s on average for G ≤ 450 gNBs. The sec-
ond fastest is, somehow counter-intuitively, the readiness-cost-minimizing approach
(P20), whose convergence times are consistently lower than those of the performance-
maximizing approach for the whole interval, although the objective function of the
former contains that of the latter. The reason is for this is the influence of the operat-
ing cost component in the (P20), which leads to formulations that are easier to solve.
As a result, we conclude that with operator-grade equipment, the readiness-cost-min-
imizing FSSP can be solved in less than 2 s on average when G ≤ 300 gNBs, and in
less than 3 s for G ≤ 450 gNBs.

4.7.3. Spectral efficiency and cost evaluation

After evaluating the convergence time of the approaches, we measure their perfor-
mance in terms of the achieved geometric mean of the spectral efficiency ηe(a), as it is
defined in (11), as well as their achieved operating and readiness cost. We first evalu-
ate ηe(a) for all considered performance-maximizing approaches over a set of extreme
cases so as to select the best approach out of them. Then, we analyze the impact of
the configuration of the fronthaul network and the user concentration on the most
promising approaches. Finally, we assess the spectral efficiency, operating cost, and
readiness cost of all three interpretations of the FSSP.

95
4. Optimal Functional Split Selection in 5G Radio Access Networks

Figure 4.7.: Convergence time of the quadratic approach to the performance-maximiz-

ing FSSP, the operating-cost-minimizing FSSP, and the quadratic approach
of the readiness-cost-minimizing FSSP.

Selection of best performance-maximizing approach

In order to compare the performance of all performance-maximizing approaches, we

evaluate their spectral efficiency ηe(a) on the same scenario. Since the LWT and BGP
formulations are only applicable to small networks, we choose G = 50 gNBs for this
comparison, corresponding to U = 500 UEs and an area of approximately 0.4 km2 . We
now generate four types of scenarios to cover a wide range of interference cases: (i)
a sparse fronthaul (ψ = 2) with uniform population (θ = 0.5); (ii) a dense fronthaul
(ψ = 5), with uniform population (θ = 0.5); (iii) a sparse fronthaul (ψ = 2), with con-
centrated population (θ = 0.95); and (iv) a dense fronthaul (ψ = 5), with concentrated
population (θ = 0.95).
The simulation results after 200 runs are shown in Fig. 4.8, with scenarios (i)–(iv) be-
ing used in Fig. 4.8a–4.8d, respectively. Apart from the proposed approaches, we
also include in the comparison the spectral efficiencies of a fully distributed solution
(ag = 0 ∀g ∈ G), a static solution (in which the optimal s is precomputed assuming a
perfectly uniform population), and a fully centralized solution (ag = M ∀g ∈ G). The
fully distributed solution represents a network in which all processing is distributed,
thus providing a lower bound to all solutions. The static solution is the one proposed
in works such as [GS+18a; HR18b], in which the functional split of every gNB is cal-
culated for the average traffic and not adapted to the instantaneous interference situ-
ation. The fully centralized solution, in which all gNBs are centralized, is infeasible in
all cases, but it serves as a upper bound for the other approaches.
From these results we can observe two general trends. First, the denser the fronthaul
network, the higher the spectral efficiency of all solutions, but also the better the per-
formance of our proposed solutions with respect to the static solutions. Second, the
more concentrated the population, the higher the variance and the mean spectral effi-

96
 4.7. Experimental evaluation

Quadr.

(a) Sparse fronthaul, uniform population (b) Dense fronthaul, uniform population

(c) Sparse fronthaul, concentrated popula- (d) Dense fronthaul, concentrated popula-
tion tion

Figure 4.8.: Distributions of the mean spectral efficiency ηe(a) achieved by the perfor-
mance-maximizing approaches, a fully distributed approach, a static ap-
proach, and a fully centralized approach in four extreme scenarios.

ciency achieved by our solutions with respect to the static solutions. As a result, when
the fronthaul is dense and the population is concentrated, our approaches achieve
similar spectral efficiencies to that achieved by full centralization.
Apart from these general trends, we can compare the performance of our proposed
performance-maximizing approaches. We can conclude that, with the exception of
Heuristic 1, all these approaches perform very similarly in all cases. Under close ex-
amination, we can see that the best performance is achieved by LWT and BGP trans-
formations and Heuristic 2, followed closely by the quadratic approach and the punc-
tured LWT and BGP transformations. Nonetheless, the maximum difference in their
average spectral efficiency is less than 2% in all scenarios.
Given its good-quality solutions and its low convergence time, we conclude that the
quadratic formulation is the most efficient approach, and hence the most suitable to
use in combination with the operating-cost-minimizing approach to optimize the av-
erage readiness cost. Nonetheless, Heuristic 1 may be still adequate for large networks
(G ≥ 400) in which convergence time needs to be very short, in the sub-second range.
Similarly, the punctured and unpunctured LWT and BGP transformations may be ap-
plicable to small networks (G ≤ 40) in which maximizing the spectral efficiency is of
utmost importance, such as industrial or ultra-reliable networks.

97
4. Optimal Functional Split Selection in 5G Radio Access Networks

Quadr.

Figure 4.9.: Average spectral efficiency achieved by the quadratic approach, Heuristics
1 and 2, a static and a fully distributed network for G = 300 and a UE
concentration index of 0.75 as the fronthaul average degree varies.

In Fig. 4.8 we can see how the density of the fronthaul network affects the quality of
the solutions achieved by all approaches for a deployment with G = 50 gNBs. When
using a sparse fronthaul network of ψ = 2, few centralization vectors are feasible be-
cause of the limited number of paths, which leads to mean optimal spectral efficien-
cies of around 1.57 b/s/Hz for dispersed UEs and 2.14 b/s/Hz for concentrated UEs.
With a denser fronthaul network of ψ = 5, the average spectral efficiency grows up
to around 2.7 b/s/Hz for dispersed UEs and 3.31 b/s/Hz for concentrated UEs. With
the intention of observing this trend more clearly, we perform another experiment on
a larger network (G = 300, 2.6 km2 ), with partially concentrated UEs (θ = 0.75) and
let the density of the fronthaul vary from ψ = 2 to ψ = 5. The results are shown in
Fig. 4.9. We conclude that with sparse, tree networks (ψ = 2) the benefits of imple-
menting a performance-optimized solution are marginal, improving only from 1.07
b/s/Hz (static solution) to 1.18 b/s/Hz (quadratic approach), a 10% improvement.
With an average degree of ψ = 3 this improvement increases up to 28% (quadratic
approach), and with ψ = 5 it reaches 36% (quadratic approach).
Implementing an adaptive functional split is specially beneficial when the UEs tend
to be clustered in time-varying clusters. This can be observed once again in the ex-
periment shown in Fig. 4.8: as the UE concentration index changes from θ = 0.5 to
θ = 0.95, the average spectral efficiency increases from 1.57 b/s/Hz to 2.7 b/s/Hz
for ψ = 2, and from 2.14 b/s/Hz to 3.31 b/s/Hz for ψ = 5. Conversely, the average
spectral efficiency of the static solution barely changes. This is due to the fact that,
when UEs are concentrated around the same spots, an adaptive network can mitigate
their interference more efficiently than when they are spread apart. In order to evalu-
ate the effect of UE concentration in more detail, we perform another experiment on a
larger network (G = 300, 2.6 km2 ) with a constant fronthaul average degree of ψ = 3.5
and let the UE concentration index vary from θ = 0.5 to θ = 0.95. The results are
shown in Fig. 4.10. We conclude that an adaptive approach may achieve substantially

98
 4.7. Experimental evaluation

Quadr.

Figure 4.10.: Average spectral efficiency achieved by the quadratic approach, Heuris-
tics 1 and 2, a static and a fully distributed network for G = 300 and a
fronthaul average degree of 3.5 the fronthaul average degree varies as the
UE concentration index varies.

better spectral efficiency when UEs are concentrated with respect to a static solution.
Indeed, when θ = 0.8, the mean spectral efficiency can be improved from 1.18 b/s/Hz
to 1.75 b/s/Hz, a 48% improvement. For θ = 0.95, this improvement reaches almost
90%. Interestingly, the mean spectral efficiency of a static solution is barely affected
by the UE concentration, although its variance does increase. This is because clusters
form at any point of the covered area with equal probability, combined with the fact
that the static solution explicitly optimizes for the average UE position, regardless of
the instantaneous UE concentration.
From these evaluations of the convergence times and achieved spectral efficiencies,
we conclude once again that the quadratic approach as formulated in (P12) is the
most promising option for solving the performance-maximizing FSSP. Consequently,
we also use this approach, in combination with the operating-cost-minimizing for-
mulation, to find the best state vector s that minimizes the readiness cost. That is,
we opt for formulation (P20) when dealing with the readiness-cost-minimizing FSSP.
Nonetheless, since the heuristic approaches yield good quality solutions in a relatively
short time, we still include them in subsequent comparisons.

Comparison between performance-maximizing and operating-cost-minimizing

approaches

Although the performance-maximizing and the operating-cost-minimizing FSSPs have

different objectives, their solutions may not be radically different. Indeed, there may
be a positive correlation between operating cost minimization and spectral efficiency
maximization as a result of the influence of computational costs K b comp (b). Since the

99
4. Optimal Functional Split Selection in 5G Radio Access Networks

Distributed Quadratic Heuristic 1

Fully centralized Operating cost Heuristic 2

Figure 4.11.: Relative operating cost achieved by fully distributed, fully centralized,
quadratic, operating-cost-minimizing, and heuristic approaches when
G = 300, average fronthaul degree ψ = 3.5, and UE concentration in-
dex θ = 0.75.

cost of running mobile functions at the CU is much smaller than at the DU (γCU = 0.017
ncu/cycle vs. γDU = 1 ncu/cycle, according to [GS+18b]), minimizing this cost com-
ponent entails centralizing as many functions as possible, which is also desired by the
performance-maximization approach. Nonetheless, the routing costs Krout (f ) have the
opposite effect, since function centralization increases network usage. As a result, the
value of ω, the transport cost of carrying traffic on a link, greatly influences how ex-
ceedingly costly a performance-maximizing solution is, and how much throughput is
lost by a operating-cost-minimizing solution.
Fig. 4.11 shows the impact of ω on the operating cost of solutions provided by the per-
formance-maximizing quadratic approach and Heuristics 1 and 2, along with that of
the operating-cost-minimizing solution from (P16). For reference, we also include the
operating cost of fully distributed and fully centralized networks, although the latter
is always infeasible. We observe that when ω = 0, the operating cost of our proposed
approaches is very similar to that of the operating-cost-minimizing approach. In fact,
the average cost difference between the operating-cost-minimizing and quadratic ap-
proaches is less than 1%. As ω grows, the operating-cost-minimizing solution con-
verges to the distributed solution, since centralization incurs in high routing costs.
Conversely, the operating cost of our proposed approaches increases linearly with ω,
as this parameter is not taken into account for solution selection. As a consequence, at
ω = 0.5, the average operating cost difference between the operating-cost-minimizing
and quadratic approaches is 22%, and if ω = 1 this difference increases up to 45%.
In Fig. 4.12 we show the influence of the routing costs ω on the spectral efficiencies
achieved by the same approaches as in Fig. 4.11. We observe that the operating-cost-
minimizing approach always achieves noticeably lower spectral efficiencies than per-

100
 4.7. Experimental evaluation

Distributed Quadratic Heuristic 1

Fully centralized Operating cost Heuristic 2

Figure 4.12.: Average spectral efficiency achieved by fully distributed, fully central-
ized, quadratic, operating-cost-minimizing, and heuristic approaches
when G = 300, average fronthaul degree ψ = 3.5, and UE concentration
index θ = 0.75.

formance-maximizing approaches, as we might expect. At ω = 0, our quadratic ap-

proach achieves a 15% higher spectral efficiency than the operating-cost-minimizing
approach, this increases to 45% at ω = 0.5, and finally to 86% for ω ≥ 0.75, as the op-
erating-cost-minimizing approach converges to the distributed solution. We conclude
that the additional cost of performance-maximizing approaches translates into pro-
portionally higher improvements in the spectral efficiency. If the operator is able to
profit from these improvements, then performance-maximizing approaches may lead
to a higher revenue with respect to static or operating-cost-minimizing approaches.
With the intention of providing a detailed cost analysis, we also study how the fron-
thaul network density and the UE concentration influence the trade off between spec-
tral efficiency and cost in two scenarios. First, we investigate the case where routing
costs are negligible (ω = 0), and thus the operator is motivated to centralize as many
functions as possible, either pursuing minimal operating cost or maximum spectral
efficiency. The results for this scenario are depicted in Fig. 4.13. We observe that, in
this case, higher fronthaul densities lead to higher spectral efficiencies (Fig. 4.13a) and
lower costs (Fig. 4.13c) for all approaches, since more functions can be maximized.
Nevertheless, our proposed performance-maximizing approaches take more advan-
tage of dense fronthaul networks, achieving better spectral efficiency than the oper-
ating-cost-minimizing approach while having comparable cost. Indeed, when ψ = 5,
the quadratic approach achieves a 16.5% higher spectral efficiency while being only
1.3% more costly than the operating-cost-minimizing approach. A similar trend can
be observed as the UEs become more concentrated: if θ = 0.95, the quadratic approach
achieves a 15.1% higher spectral efficiency while being only 6.3% more costly.
Finally, we examine a scenario where routing costs are not negligible, but still low

101
4. Optimal Functional Split Selection in 5G Radio Access Networks

(a) Spectral efficiency, θ = 0.75 (b) Spectral efficiency, ψ = 3.5

(c) Relative operating cost, θ = 0.75 (d) Relative operating cost, ψ = 3.5

Figure 4.13.: Spectral efficiency and relative operating cost achieved by fully dis-
tributed, quadratic, operating-cost-minimizing and heuristic approaches
when G = 300 and ω = 0 ncu · s/Gb.

enough so that a fully distributed configuration is not the least costly option. We
set the routing cost to ω = 0.5 ncu · s/Gb to illustrate this scenario, which is the
midpoint between the value where the performance-maximizing approach becomes
more costly than the fully distributed configuration (ω ≈ 0.25 ncu · s/Gb), and the
value where the cost optimal approach converges to the fully distributed configura-
tion (ω ≈ 0.75 ncu · s/Gb). The results are depicted in Fig. 4.14. For this scenario, we
observe that the cost and spectral efficiency gaps between the operating-cost-minimiz-
ing and performance-maximizing approaches are noticeably larger than in the previ-
ous scenario, since the former is now very limited by the routing costs. When ψ = 5,
the quadratic approach achieves a 54.4% higher spectral efficiency (Fig. 4.14a) while
being 47.5% more costly than the operating-cost-minimizing approach (Fig. 4.14c). A
similar trend can be observed as the UEs become more concentrated: if θ = 0.95, the
quadratic approach achieves a 116.7% higher spectral efficiency (Fig. 4.14b) while be-
ing 90.7% more costly (Fig. 4.14d). These results suggest that, when the routing costs

102
 4.7. Experimental evaluation

(a) Spectral efficiency, θ = 0.75 (b) Spectral efficiency, ψ = 3.5

(c) Relative operating cost, θ = 0.75 (d) Relative operating cost, ψ = 3.5

Figure 4.14.: Spectral efficiency and relative operating cost achieved by fully dis-
tributed, quadratic, operating-cost-minimizing and heuristic approaches
when G = 300 and ω = 0.5 ncu · s/Gb.

are high, performance-maximizing approaches can only compete with the operating-
cost-minimizing one as long as the higher spectral efficiency translates into higher
profit.

Readiness cost evaluation

Finally we compare the spectral efficiency, operating cost, and readiness cost achieved
by the performance-maximizing, operating-cost-minimizing, and readiness-cost-min-
imizing interpretations of the FSSP. In Fig. 4.15 we show this comparison for G = 300,
ψ = 3.5, negligible routing costs (ω = 0 ncu · s/Gb), and three selected performance-
to-revenue functions: a low-intercept, small-slope relationship hξ, υi = h0.5, 1.5i; a
medium-intercept, medium-slope relationship hξ, υi = h1, 2i; and ; a large-intercept,
large-slope relationship hξ, υi = h2, 3i. For all cases, we observe that the spectral ef-
ficiency achieved by the readiness-cost-minimizing approach is similar to that of the

103
4. Optimal Functional Split Selection in 5G Radio Access Networks

4 4 4

3 3 3

2 2 2

1 1 1
0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95

(a) ξ = 0.5, υ = 1.5 (b) ξ = 1, υ = 2 (c) ξ = 2, υ = 3

3.5 3.5 3.5

3 3 3
2.5 2.5 2.5
2 2 2
1.5 1.5 1.5
1 1 1

0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95

(d) ξ = 0.5, υ = 1.5 (e) ξ = 1, υ = 2 (f) ξ = 2, υ = 3

2 0
1 -2 -10

0 -4
-20
-6
-1
-8 -30
-2
-10
-3 -40
0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95

(g) ξ = 0.5, υ = 1.5 (h) ξ = 1, υ = 2 (i) ξ = 2, υ = 3

Figure 4.15.: Spectral efficiency, operating cost, and readiness cost achieved by perfor-
mance-maximizing (quadratic), operating-cost-minimizing, and readi-
ness-cost-minimizing approaches when G = 300, ψ = 3.5, and ω =
0 ncu · s/Gb.

performance-maximizing approach when the UE concentration index θ is small or

moderate, and up to 5% lower on average for high UE concentrations. This implies
that the achieved spectral efficiency is consistently higher than that of the operating-
cost-minimizing approach. In addition, the operating costs of all approaches is very
similar in all cases. Indeed, the maximum average difference between the operating
cost of the readiness-cost-minimizing approach and the operating-cost-minimizing
approach is less than 2%. Finally, we conclude that, when ω = 0, the readiness cost
of the performance-maximizing and readiness-cost-minimizing approaches is rather
similar, although the latter achieves a 3% to 7% lower readiness cost when the UEs

104
 4.7. Experimental evaluation

3.5 3.5 3.5

3 3 3
2.5 2.5 2.5
2 2 2
1.5 1.5 1.5
1 1 1

0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95

(a) ξ = 0.5, υ = 1.5 (b) ξ = 1, υ = 2 (c) ξ = 2, υ = 3

6 6 6

5 5 5

4 4 4

3 3 3

2 2 2
0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95

(d) ξ = 0.5, υ = 1.5 (e) ξ = 1, υ = 2 (f) ξ = 2, υ = 3

2 0
4
0
-10
2 -2

-4 -20
0 -6
-30
-8
0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95 0.5 0.65 0.8 0.95

(g) ξ = 0.5, υ = 1.5 (h) ξ = 1, υ = 2 (i) ξ = 2, υ = 3

Figure 4.16.: Spectral efficiency, operating cost, and readiness cost achieved by perfor-
mance-maximizing (quadratic), operating-cost-minimizing, and readi-
ness-cost-minimizing approaches when G = 300, ψ = 3.5, and ω =
0.5 ncu · s/Gb.

are highly concentrated. Overall, it is a remarkable result that, for negligible routing
costs, the readiness-cost-minimizing approach behaves similarly to the performance-
maximizing approach regardless of which performance-to-revenue function is in use
(see Fig. 4.2).
In Fig. 4.16, we repeat the same experiments as in Fig. 4.15 but for non-negligible
routing costs, that is, we set ω = 0.5 ncu · s/Gb, as previously discussed. With respect
to the previous experiments, we notice two main differences. First, the spectral effi-
ciency, operating cost, and readiness cost of the readiness-cost-minimizing approach

105
4. Optimal Functional Split Selection in 5G Radio Access Networks

is not consistently similar to those of the performance-approach anymore. Second, the

shape of the performance-to-revenue function does play a role in the behavior of the
readiness-cost-minimizing approach in this case. Namely, for a low-intercept, small-
slope performance-to-revenue function hξ, υi = h0.5, 1.5i, the readiness-cost-minimiz-
ing approach resembles the operating-minimizing-approach more than the perfor-
mance-maximizing approach. This is consistent with our previous observation that,
when ω = 0.5, the performance and cost gaps between the performance-maximizing
and operating-minimizing-approaches are more similar than when ω = 0. Conversely,
for large-intercept, large-slope performance-to-revenue function hξ, υi = h2, 3i, the im-
proved performance plays a more relevant role in the total revenue, and thus in that
case the readiness-cost-minimizing approach resembles the performance-maximizing
approach. In conclusion, this confirms that our readiness-cost-minimizing approach
can indeed adapt to a wide range of network conditions and yield lower readiness
cost than any alternative approach.

4.8. Summary
In this chapter, we tackle the problem of finding the optimal functional split that opti-
mizes an instantaneous network scenario, building upon the static FSSP formulation
presented in Chapter 2. We provide three different optimization objectives, which re-
sult in three FSSP formulations: a performance-maximizing FSSP, an operating-cost-
minimizing FSSP, and a readiness-cost-minimizing FSSP that combines the other two.
We first tackle the performance-maximizing FSSP, which leads to a non-linear opti-
mization problem. We present four fractional reformulations, which can be converted
into MILPs: the punctured and unpuntured versions of the LWT and BGP transfor-
mations. In addition, we propose a simpler reformulation into a quadratic program,
which is then linearized too. We also provide two heuristic approaches: Heuristic 1,
which exploits the correlation between inter-cell interference and centralization level,
and Heuristic 2, which improves upon the quadratic approach.
The operating-cost-minimizing FSSP is simpler to address, since it can be directly ex-
pressed as an MILP. However, for the readiness-cost-minimizing FSSP we need to
translate network performance into revenue. In order to cover a wide range of cases,
we use a family of linear functions for this translation, whose intercept and slope can
be configured. After this, we combine previous performance-maximizing and operat-
ing-cost-minimizing MILP approaches to obtain the readiness-cost-minimizing FSSP.
In order to assess the viability and performance of the proposed formulations and
solving approaches, we simulate multiple realistic scenarios using the size of the net-
work G, average fronthaul degree ψ, and the UE concentration θ as control variables.
We first focus on evaluating the trade-off between achieved spectral efficiency and
convergence time for each approach, since our ultimate intention is to extend these
approaches and use them for dynamic adaptation. We observe that the fractional
reformulations of the performance-maximizing approach lead to the best spectral ef-

106
 4.8. Summary

ficiency, but this is less than a 2% improvement over the quadratic approach in all
scenarios. Moreover, they may take 15 minutes or more to converge for networks
with more than G = 80 gNodeBs, thus rendering these approaches unsuitable for dy-
namic adaptation. Conversely, Heuristic 1 converges in less than 200 ms even for large
networks (G = 300), but the spectral efficiency achieved by this approach can be up
to a 20% smaller than the quadratic approach, specially when the fronthaul network
is sparse (ψ = 2). As a result, the quadratic approach seems specially promising since
it often requires less than 5 s to converge, even when G = 300, and provides a spectral
efficiency comparable to that of fractional approaches. Regarding the other FSSP for-
mulations, the operating-cost-minimizing formulation converges in less than 30 ms
on average, whereas the readiness-cost-minimizing approach takes between 0.7 to 2.5
seconds to converge when using the quadratic approach.
We conclude that for sparse, uniformly-populated networks (ψ = 2, θ = 0.5) the ben-
efits of implementing a performance-optimized solution are marginal, since the spec-
tral efficiency improvement of the quadratic approach over the static reference is only
ca. 10%. Nonetheless, for dense fronthaul networks (ψ = 5) this improvement reaches
36%. If the routing costs are moderate (ω < 1), we observe that performance-max-
imizing approaches can leverage the benefits of fronthaul density better than both
static and operating-cost-minimizing approaches, since they achieve better spectral
efficiency than both of them while requiring an operating cost comparable to that of
the operating-cost-minimizing approach. However, when the routing costs are high
(ω ≥ 1), performance-maximizing approaches can only compete with the operating-
cost-minimizing one as long as the higher spectral efficiency translates into higher
profit. Something similar happens with the concentration of UEs. We conclude that
a performance-maximizing approach may achieve substantially better spectral effi-
ciency when UEs are concentrated with respect to a static solution, reaching a 48%
improvement when θ = 0.8 and a 90% improvement for θ = 0.95.
Finally, we observe that the additional cost of performance-maximizing approaches
over the operating-cost-minimizing approach translates into proportionally higher
improvements in the spectral efficiency. As a result, if the operator is able to profit
from these improvements, then performance-maximizing approaches may lead to a
higher revenue with respect to static or operating-cost-minimizing approaches. This
is confirmed after evaluating the performance and operating cost of the readiness-
cost-minimizing approach. We conclude that, when the routing costs are negligible,
the readiness costs of the performance-maximizing and readiness-cost-minimizing
approaches are rather similar, although the latter achieves a 3% to 7% lower readi-
ness cost when the UEs are highly concentrated. For non-negligible routing costs, the
readiness-cost-minimizing approach may still resemble the performance-maximizing
approach if the revenue function is steep, although it converges to the operating-cost-
minimizing approach if this is not the case. In conclusion, this confirms that our readi-
ness-cost-minimizing approach can indeed adapt to a wide range of network condi-
tions and yield lower readiness cost than any alternative approach, while still requir-
ing a very low convergence time. Therefore, the readiness-cost-minimizing approach
is a perfect candidate for being used in dynamic optimization.

107
5. Implementation of an Adaptive Functional
Split

5.1. Introduction

5.1.1. Motivation, scope, and challenges

In Chapters 2, 3, and 4 we motivate, propose, formulate, and present strategies to

solve the problem of dynamically selecting the optimal functional split for a instan-
taneous network configuration, with the objective of maximizing the experience of
the users without overloading the fronthaul network. We show the relevance of this
problem, observe that there are efficient strategies to tackle it, and conclude that a
network featuring an optimally adaptive functional split outperforms non-adaptive
deployments in terms of experienced user data rates and may even lead to lower
operating costs. Nonetheless, any advantages that an adaptive functional split may
have depend on a critical feature: changing the functional split during runtime must
be feasible, fast, and cost-efficient. That is, if changing the functional split were too ex-
pensive, took too long, or were technically infeasible, it would be impossible to benefit
from dynamically adapting the functional split.
Being specific to 5G and late 4G networks, the concept of a functional split in the RAN
is, on its own, rather novel. Indeed, in Sec. 2.2.4 we mention that, although it is gen-
erally agreed that a functional split should be beneficial in terms of user performance
and operating cost, there is relatively little standardization work regarding how to
realize the interfaces between separated functions. Consequently, with the current
state of the art, it would be intractable to assess the feasibility of adapting from and to
every possible functional split, since this would require defining new, dedicated pro-
tocols modeling all interactions between RAN functions. Nonetheless, we can indeed
focus on the most promising functional split options, which receive the most atten-
tion from the standardization and research communities, and build upon previous
work to propose a proof-of-concept RAN that can switch between two distinct func-
tional split options. This proof-of-concept RAN can be used to show that changing
the functional split during runtime is actually feasible and also to identify potential
problems in more complicated systems, thus paving the way for advanced functional
split adaptations.
Even if we only consider a small subset of functional split options in our proof-of-
concept, assessing the feasibility of being able to implement several of these options
and also supporting live changes in the centralization level without severely disturb-

109
5. Implementation of an Adaptive Functional Split

ing normal operation is a very challenging task. Namely, we first have to devise a
manner to support two or more functional split options so that they can work inde-
pendently without interfering with each other. Second, we must design a procedure to
migrate from one functional split option to the other during runtime. Finally, we need
to ensure that there is no noticeable downtime, and packet losses and additional la-
tency are minimized during the functional split migration. In this chapter, we present
a pioneer implementation of a proof-of-concept 5G RAN that is able to switch be-
tween the PDCP-RLC and MAC-PHY splits during runtime with minimal impact on
the network operation.

5.1.2. Key contributions

In order to address the aforementioned challenges, this chapter bases on our previous
works [MAGVK19a] and, to a lesser extent, [MAGVK19b] to feature the following
contributions:
1. We evaluate the possibilities to implement a proof-of-concept RAN featuring a
dynamically adaptive functional split, and select the most promising options.
Then, we combine and extend previous works to allow for a simultaneous im-
plementation of multiple functional split options within a single gNodeB.
2. We identify and discuss the objectives, challenges, and possible strategies to
change the functional split during runtime with minimal impact on normal op-
eration.
3. We explain a detailed procedure to migrate between the selected functional split
options that overcomes the identified challenges by defining and transferring
the instantaneous state of the base station from the CU to the DU, and vice
versa. A description of the additional required functions and their impact on
the migration procedure is also included.
4. We implement the proposed migration procedure on an actual 5G RAN testbed
and provide measurements of the internal operation of the buffers during a mi-
gration, as well as the impact of migrating on end-user latency and packet losses,
so as to ensure that the migration does not severely affect the operation of the
RAN.
The rest of this chapter is organized as follows. Sec. 5.2 summarizes the current state
of the art on the topic, including academic research and existing 4G/5G RAN imple-
mentations, upon which we build our proof-of-concept RAN. In Sec. 5.3, we briefly
discuss the functional split options and select those that are used in our proof-of-
concept. Sec. 5.4 presents the objectives and challenges of implementing an adaptive
functional split, as well as a detailed description of our proposed migration proce-
dure. In Sec. 5.5, we describe our implementation of the proof-of-concept RAN and
provide measurement results. Finally, Section 5.6 summarizes the contributions and
findings of this chapter.

110
 5.2. Related work

5.2. Related work

In this section, we divide previous works related to the content of this chapter into two
categories. On the one hand, we present a summary of relevant academic research that
considers or discusses 5G RAN designs whose functional split can be changed, either
from a theoretical or a practical point of view. On the other hand, since this chapter
focuses on implementation details, we describe the most relevant software platforms
that can be currently used to implement a 5G RAN, since our proof-of-concept RAN
actually builds upon one such platform.

5.2.1. Academic research

The advantages and challenges of functional splits in the architecture of the 5G RAN
have triggered the research on their optimal definition and selection. As mentioned
in previous chapters, we can find abundant work on the characteristics of diverse
functional splits, such as [Döt+13], [LCC18], [Bar+15], or [Val+16]. Owing to their
differences, the authors of these works suggest that the functional split should be
adapted to the external conditions that the network experiences. To this end, [Mae+16]
presents a high-level overview of a flexible 5G RAN, which includes the ability to
support multiple functional splits, in order to adapt the RAN configuration to the
expected user traffic. Building upon this, in [Cha+17a] the authors propose a RAN
architecture that simultaneously supports different functional splits for each DU in the
network. However, none of these works explicitly consider changing the functional
split on-the-fly, but rely on a priori statistics of the network to select the optimal, static
functional split.
Conversely, there are works which do tackle, to a greater or lesser extent, the change
of RAN functions at runtime. For instance, in FlexRAN [Fou+16], a novel imple-
mentation of a software-defined RAN is presented, featuring a virtualized RAN con-
troller whose location and configuration can be dynamically changed. In FlexCRAN
[Cha+17b] the authors propose a framework for a partially centralized RAN that sup-
ports on-the-fly changes of the functional split, although they do not elaborate on this
feature. In [HR18a], the problem of dynamically selecting the functional split for each
cell is addressed. The authors present an algorithm that allows virtual mobile opera-
tors to change the functional splits every time a new virtual network is added. Sim-
ilarly, [Mar+18] proposes an architecture for a 5G RAN implementing an algorithm
to dynamically select the functional split of a cell. This work puts special emphasis
on enabling changes at runtime. Nonetheless, it only covers the optimal selection of
the functional split, regardless of the feasibility of the required changes. Finally, in
[Alf+18] the authors present a pioneer platform that can switch between functional
splits at runtime. However, the options are limited to low-layer, intra-physical splits,
and the details about performing the switching are not explained in detail.
In the light of these previous works, we conclude that our proof-of-concept RAN is, to
the best of our knowledge, the first work addressing the implementation of a network

111
5. Implementation of an Adaptive Functional Split

that can adapt its functional split at runtime. As a result, we also pioneer the discus-
sion of objectives and challenges, and the provision of actual measurement results.

5.2.2. 4G/5G Software Platforms

Implementing a mobile RAN whose functional split can be changed during runtime
requires an underlying 4G/5G software platform to realize the remaining elements
of the mobile network. Fortunately, in recent years, some remarkable initiatives have
emerged with the intention of providing full-stack implementations of 3GPP mobile
networks, such as OpenAirInterface [Oped], srsRAN [SRSa], OpenLTE [Ben], and
OpenBTS [Ran]. In this section, we briefly introduce the two most relevant ones: Op-
enAirInterface and srsRAN.

OpenAirInterface

OpenAirInterface [Nik+14; Oped] is an open-source software platform that imple-

ments a complete 3GPP mobile network, including radio access and core networks,
which can be used on commercial off-the-shelf equipment, such as standard PCs,
software-defined radios (SDRs), smartphones and LTE dongles, etc. It provides a com-
plete set of C/C++ libraries and binaries that emulate the operation the most relevant
functions, protocols, and units of a mobile network, so that a full mobile network can
be replicated and experimented with in standard laboratory conditions. For instance,
by using OpenAirInterface it is possible to operate two conventional PCs and SDRs
as if they were a UE and a 3GPP base station, with support of the most important in-
ternal protocols and interfaces. This allows to evaluate potential modifications in the
operation of any component in the mobile network with ease and without requiring
to invest in expensive, dedicated equipment.
OpenAirInterface is developed by the OpenAirInterface Software Alliance (OSA),
which was founded in 2014 by EURECOM and comprises more than 85 members,
including mobile operators (such as Orange), hardware manufacturers (such as Qual-
comm, Xilinx, and Fujitsu), research companies (such as Nokia Bell Labs) and uni-
versities (such as the Sorbonne University and the Technical University of Munich)
[Opef]. The OSA is responsible for deciding on the development roadmap and mile-
stones of OpenAirInterface, providing quality control, and promoting the produced
software in both academic and industrial communities.
The OpenAirInterface project is divided into two sub-projects: one focusing on the
core network and another focusing on the RAN. Early versions of the OpenAirIn-
terface core network implement the legacy LTE components of the Evolved Packet
Core (EPC) as software functions, including the Mobility Management Entity (MME),
Home Subscriber Server (HSS), Serving Gateway (S-GW), and PDN Gateway (P-GW).
Standard interfaces connecting these functions, such as the S11 interface between the
MME and S-GW, the S6a interface between the HSS and MME, and the S5 interface be-

112
 5.2. Related work

tween S-GW and P-GW are also supported, thus facilitating development and provid-
ing interoperability with other EPC implementations. Moreover, in September 2020,
the OSA presented a new, open implementation of the service-based 5G core archi-
tecture, with the intention of eventually supporting a stand-alone deployment of a 5G
network. This implementation of the 5G core comprises the Access and Mobility Man-
agement Function (AMF), the Authentication Server Management Function (AUSF),
the Network Repository Function (NRF), the Session Management Function (SMF),
the Unified Data Management function (UDM), the Unified Data Repository (UDR)
and the User Plane Function (UPF). Although this implementation is, by September
2021, still under development, it already provides the majority of the core functions
required to run a simple 5G core network [Opec; Opee].
The OpenAirInterface RAN is structured into three parts, which roughly follow the
protocol stack separation in 4G and 5G networks. The first part complies with the
physical layer specification described in 3GPP LTE Release 8.6 and includes some fea-
tures of Release 10. This translates into a maximum of 20 MHz channel bandwidth,
two transmission and reception antennas, and a throughput of 70 Mb/s [Opea]. In
addition, this first part also includes the required code to operate off-the-shelf SDRs,
such as the Eurecom EXMIMO II, Ettus B210/X300, BladeRF, and LimeSDR. The sec-
ond part includes MAC, RLC, PDCP, RRC layers as described in 3GPP LTE Releases
8.6, 9.3, 10.1, and 14.3, respectively. The MAC layer comprises a scheduler function
implementing a proportionally fair allocation of radio resources and fully supports
hybrid ARQ (HARQ), link adaptation, and power control [Opea]. The RLC layer im-
plements all three transmission modes (transparent, unacknowledged, and acknowl-
edged), whereas the PDCP layer supports packet sequencing, integrity check, and
encryption using AES or Snow3G algorithms, but not header compression [Opea].
The RRC layer supports generation and broadcast of system information blocks 1, 2,
3, and 13 as well as standard RRC connection establishment, reconfiguration, release,
and re-establishment procedures, but handover and paging procedures are still in de-
velopment [Opea]. Finally, the third part deals with implementing the S1 application
protocol between the base station and the MME as well as the user-plane Non-Access
Stratum (NAS) GPRS Tunneling Protocol (GTPV1-U) for both base stations and UEs
as specified in 3GPP LTE Release 10.
The OpenAirInterface RAN currently supports splitting a base station into three units:
the Radio-Cloud Center (which can be identified with the CU in 5G networks, since
they share the same objective), the Radio-Access Unit (equivalent to the DU), and the
Remote Radio Unit (similar to the RU). A partial support for the MAC-PHY split be-
tween the Radio-Access Unit and the Remote Radio Unit is included, based on the
nFAPI interface proposed by the Small Cell Forum [Sma17]. It is also planned to im-
plement PDCP-RLC functional split between the Radio-Cloud Center and the Radio-
Access Unit as specified in interface F1 of 3GPP Release 15 [3GP21j], although this
feature is not fully supported yet [Opea].
Since the research community is currently interested in testing and developing fifth-
generation features, it is planned that by end of 2021 OpenAirInterface will support

113
5. Implementation of an Adaptive Functional Split

a completely stand-alone implementation of a 5G RAN [Opeb]. This will entail im-

provements and additions to the physical layer in order to support new radio (NR)
transmission and reception, as well as modifications to the MAC and RLC layers so
as to support bandwidth parts, among other novelties. Finally, the implementation of
the nFAPI interface enabling the MAC-PHY split will be completed so as to allow for
combined exchange control and data information, and also the support of improved
MIMO communications.

srsLTE/srsRAN

srsRAN [SRSa], formerly known as srsLTE, is an open-source software suite which, as

OpenAirInterface, intends to provide a full-stack implementation of a 4G/5G mobile
network, including core network, RAN, and UEs for off-the-shelf hardware equip-
ment. The srsRAN suite contains a collection of libraries, modules, and tools written
in C/C++ that can be compiled on standard Linux PCs to use alone or in combina-
tion with third-party 4G/5G software. Like OpenAirInterface, srsRAN can be used to
replicate a complete mobile network using only conventional PCs and SDRs.
srsRAN is developed by Software Radio Systems (SRS) and started in 2015 as a full-
stack UE application [SRSc]. In 2017 and 2018, SRS developed eNodeB and EPC
implementations, respectively, to complement the UE application, thus producing a
complete implementation of a 4G RAN. Since then, srsRAN has incorporated several
advanced features, such as MIMO, mobility management, and carrier aggregation.
Although software development is mainly performed by SRS, in contrast to the multi-
ple contributors to OpenAirInterface, the srsRAN suite is actively used and improved
by third-party organizations, such as National Instruments, Analog Devices, Nokia,
and the Massachusetts Institute of Technology [SRSa].
The srsRAN suite consists of three main components: srsUE (implementing the UE
protocol stack), srseNB/srsgNB (implementing the base station), and srsEPC (imple-
menting core network functions). The first component, srsUE, runs as a Linux PC
application to replicate the behavior and internal operation of a 3GPP LTE Release 10
UE, while also including some features of a 3GPP NR Release 15 UE. As OpenAirIn-
terface, srsUE supports a maximum channel bandwidth of 20 MHz for a single carrier,
but it is also able to use carrier aggregation to combine the capacity of up to two carri-
ers. It also includes support for evolved multimedia broadcast and multicast service
(eMBMS) [SRSb].
The second component, srseNB/srsgNB, implements a base station in accordance to
the specifications of 3GPP LTE Release 10. In contrast to OpenAirInterface, its support
of carrier aggregation enables a downlink throughput of up to 150 Mb/s, albeit it
also features some limitations like implementing frequency-division duplex (FDD)
but not time-division duplex (TDD) [SRSb]. Although it does not include out-of-the-
box support for functional splits, in srsRAN the interfaces between PHY, MAC, RLC,
PDCP, and RRC layers are well-defined and implemented as dedicated classes in the
source code, which facilitates the experimentation with custom functional splits.

114
 5.3. Considered functional splits

The third component, srsEPC, is a lightweight implementation of an LTE EPC, includ-

ing MME, HSS, S-GW, and P-GW. Whereas in OpenAirInterface each function requires
a dedicated binary, in srsEPC all functions are implemented as a single binary [SRSb].
This allows for a simpler operation of the core functions while still providing clear
distinction among functions in the source code.

5.3. Considered functional splits

As mentioned in Chapter 2, the 3GPP envisions up to eight different options for
the functional split in a 5G RAN [3GP17]. These correspond to the interfaces of the
five layers of the protocol stack (RRC-PDCP, PDCP-RLC, RLC-MAC, MAC-PHY, and
PHY-RF), plus three internal splits within the RLC, MAC, and PHY layers. How-
ever, as mentioned in Sec. 2.2.4, not all options are equally beneficial in actual im-
plementations. For instance, the RLC-MAC split is deemed almost useless, as it is
more complex than PDCP-RLC but does not bring any additional benefit [3GP17].
Similarly, the RRC-PDCP or RRC-SDAP splits have almost identical requirements
to PDCP-RLC without providing any major advantage. The Intra-RLC and Intra-
MAC splits may, conversely, outperform the PDCP-RLC split since they would feature
higher flow control and better resource pooling, and enable interference coordination.
Nevertheless, currently there are no well-established initiatives to standardize these
split options. Finally, the Intra-PHY and C-RAN splits promise good performance
and count with standardizing initiatives [EAN19], but their timing and capacity con-
straints are so high that render practical experimentation very difficult. As a result,
in this chapter, we focus on an adaptive 5G RAN that can switch between PDCP-RLC
and MAC-PHY splits. The selection of these two splits is based on their advantages
regarding low fronthaul utilization and good interference-coordination capabilities,
respectively, their relative simplicity, and the fact that there are already initiatives to
standardize their required interfaces, which simplifies development [Mak+17]. In the
following sections, we present them and summarize their features.

5.3.1. PDCP-RLC

In this split, the PDCP function is centralized in the CU, whereas the RLC, MAC,
PHY, and RF functions are located in the DU. This has three main advantages. First, it
reduces the operating cost with respect to a distributed architecture, since the PDCP
function is in charge of ciphering, which may be a computationally intensive task.
Second, the fronthaul traffic is very similar to the user traffic, as only a small PDCP
header is added to each IP packet. Thus, the fronthaul traffic is also comparable to
the backhaul traffic in LTE, which enables reutilization of backhaul networks. That
is, operators could reuse their former backhaul infrastructure as fronthaul network.
Finally, the standardization effort required to implement this split is low, given that it
has already been considered in the past for LTE Dual Connectivity [3GP17].

115
5. Implementation of an Adaptive Functional Split

Owing to its advantages, PDCP-RLC is the split currently considered by the 3GPP
Releases 15 and 16 for the NR specifications [3GP21j]. Nonetheless, the PDCP-RLC
lacks enough centralized functions to perform any kind of advanced coordination
technique with other gNBs, and the DU is still required to implement the computa-
tionally intensive MAC and PHY layers. Therefore, a more centralized split would be
a better option if the fronthaul capacity allows it.

5.3.2. MAC-PHY

In this split, the PDCP, RLC, and MAC functions are centralized in the CU, whereas
the PHY, and RF functions are located in the DU. More centralized functions mean
less operating costs with respect to PDCP-RLC. Furthermore, coordination techniques
such as coordinated scheduling or coordinated link adaptation are possible with a
centralized MAC [3GP17], while the fronthaul traffic is still considerably lower than
those of Intra-PHY or C-RAN splits [MAGVK19b].
Due to its characteristics, the MAC-PHY split has been selected by the Small Cell Fo-
rum for their envisioned network, and standardized in the nFAPI initiative [Sma17].
Nevertheless, the MAC-PHY split has also disadvantages. On the one hand, the pres-
ence of additional headers and control signals between MAC and PHY layers in-
creases the fronthaul capacity and latency requirements with respect to PDCP-RLC
[MAGVK19b]. On the other hand, its coordination capabilities are limited with re-
spect to C-RAN or Intra-PHY splits.

5.4. Adaptive functional split

In this section, we provide an overview of the objectives and challenges that an adap-
tive 5G RAN faces. Furthermore, we introduce the details of our solution, which
implements an adaptive functional split between PDCP-RLC and MAC-PHY.

5.4.1. Objectives and challenges

Given that a centralized architecture outperforms a distributed one in terms of cost

and coordination capabilities, the objective of an adaptive 5G RAN is to make sure
that each gNB in the network operates at the most centralized functional split that is
supported by the fronthaul network. Since the load of the fronthaul network depends
on the user traffic, which may change over time, the functional split should also be
able to change at runtime. The main difference between functional splits is the loca-
tion of the functions of the 5G processing chain, as explained in Sec. 2.2.4. As a result,
switching between functional splits at runtime is equivalent to live migrate functions
from the CU to the DU, or vice versa.

116
 5.4. Adaptive functional split

There are at least two main obstacles when live migrating RAN functions: increased
fronthaul traffic and function downtime. The former refers to the additional informa-
tion that needs to be exchanged between CU and DU during the migration, which
leads to an increase in the fronthaul traffic. The latter is the time during which the
functions being migrated are not available, owing to a possible need of halting these
functions in order to complete the migration. These two obstacles lead to two sec-
ondary objectives. On the one hand, the overhead traffic during the migration should
be minimized, as the sheer motivation of the migration may be the reduction of the
traffic on the fronthaul. That is, if the user traffic increases and the centralization level
has to be reduced in order to decrease the fronthaul load, a migration that produces
high overhead would be counterproductive. On the other hand, the downtime of the
migration should be minimized as well, since it may negatively impact the experience
of the user. This is specially concerning for the communication between low layers,
as any interruption may result in missing a whole subframe. For instance, in our case,
a downtime between PDCP and RLC layers translates into a delayed transmission
of PDCP packet data units (PDUs), which may be a problem for low-latency users.
Moreover, an interruption of the communication between MAC and PHY layers can-
not be higher than the scheduling interval, or else entire transmission and reception
slots would be wasted. In 5G, the slot duration ranges from 1 ms to 62.5, µs, so very
small downtimes are tolerated.

5.4.2. Migration strategy

As mentioned above, any change in the functional split implies moving functions
from the CU to the DU, or vice versa. In our case, the difference between MAC-PHY
and PDCP-RLC is the location of the MAC and RLC functions. In the MAC-PHY split,
these functions are located in the CU, whereas in PDCP-RLC split they are located in
the DU. Hence, when switching from PDCP-RLC to MAC-PHY, we need to move the
MAC and RLC functions from the DU to the CU, and vice versa. In order to do this,
we need an underlying migration strategy that fits the characteristics of the functions
and the requirements of the network.
The live migration of functions is a well-studied topic in the field of network func-
tion virtualization (NFV). Indeed, a common strategy to live migrate a function is to
deploy it on a virtual machine and then migrate the virtual machine [CC14]. A vir-
tualization platform is hence needed to manage the migration, since the hardware of
CU and DU may be different. For instance, platforms like OpenStack, based on hy-
pervisors such as KVM [KVM] or Xen [Theb], allow for such live migrations [Thec].
However, this virtualization-based approach conflicts directly with the two limitations
presented in the previous section. To the best of our knowledge, no existing virtual-
ization platform can offer a downtime comparable to the scheduling interval of a 5G
network (1 ms or less) [BNK16; Bas+19; Tor+21]. Moreover, live migrating a virtual
machine often entails copying its disk and memory to the destination, thus producing
a high traffic overhead until the migration is completed [Tor+21].

117
5. Implementation of an Adaptive Functional Split

A faster, lighter type of live migration is therefore needed to change functional splits in
a 5G RAN. We propose a replication-based approach, in which the MAC and RLC func-
tions are simultaneously deployed in both DU and CU. That is, at every instant there
is one active set of MAC and RLC functions at either the DU or the CU, and an inactive
set at the other unit. When the migration is performed, the roles are exchanged: the
active functions are disabled, and the inactive functions are enabled. This approach
can be considerably faster and produces much less overhead during the migration
than the virtualization-based approach. Its main drawback is that it requires to have
MAC and RLC processes running simultaneously on both units, even when they are
not used. This creates additional memory and CPU consumption on the inactive unit,
which should be taken into account as operating expenses. This drawback is dis-
cussed in more detail in Chapter 6. Furthermore, in order to provide uninterrupted
service to the users, the MAC and RLC functions cannot be just turned on or off. In
fact, there are multiple state variables and data structures that are created, modified,
and used by both functions during runtime. These have to be carefully transferred to
the destination before completing the migration. In the next section, we address the
mater of transferring the state of a base station to switch functional splits.

5.4.3. State transfer

In a replication-based migration, there are three basic tasks to accomplish: (i) transfer
the state from the old set of functions to the new one, (ii) deactivate the old set of
functions, and (iii) activate the new set of functions. That is, before being able to
toggle the two sets of functions on and off, we must guarantee that they are in the
same state. In this work, we define the state of the MAC and RLC functions as the
set of parameters that are required for the correct operation of these functions and are
susceptible to change over time. Note that this state is defined within the scope of the
internal operation of a base station and should not be confused with the concept of
network state presented in Chapter 3.
We can classify the parameters defining the state of the MAC and RLC functions into
three types, according to their origin and updating frequency. First, we have the exter-
nal parameters, which are used by the MAC and RLC layers but not created or modi-
fied by them. These are mostly details of the connected UEs and defined radio bearers
provided by either the UEs or by higher layers of the gNB. Some examples of exter-
nal parameters are logical channel IDs (LCIDs), Radio Network Temporary Identifiers
(RNTIs), or timer values. These parameters do not change frequently, as they are only
created or modified when a UE connects or disconnects, or when the higher layers
decide so. Second, the internal parameters are those created or modified by the MAC
and RLC layers themselves, such as the RLC sequence number, frame and subframe
numbers, and the list of active HARQ processes. In general, these parameters change
every slot or subframe. Finally, we consider the content of the buffers as the last type of
state parameters, including the RLC transmission and retransmission buffers, and the
HARQ retransmission buffer. The content of these buffers varies every time slot, after

118
 5.4. Adaptive functional split

transmission and reception, or when new data is received.

Given their different characteristics, the three types of parameters should be handled
differently in order to minimize the overhead and downtime of the migration. For
instance, the external parameters can be forwarded to both active and inactive RLC
and MAC functions every time they are updated, e.g., when a new UE connects or
disconnects. This results in a very small overhead traffic following these events, but
since it is performed ahead of time, it reduces the amount of data exchanged during
the migration. Conversely, the internal parameters, owing to their fast updating fre-
quency, have to be transmitted during the migration, as only the active MAC and RLC
layers are aware of the most updated values. The amount of overhead traffic caused
by transferring these internal parameters during a migration is, however, almost neg-
ligible, as only a few bytes per UE are needed to store them. Namely, up to two bytes
for the RLC sequence number, two bytes for frame and subframe numbers, and one
byte for the HARQ processes [3GP21b; 3GP21f].
In contrast, transferring the content of the buffers is a more challenging task. Accord-
ing to the 3GPP specifications for 5G [3GP21k], the RLC retransmission buffer should
store from 50 to 160 RLC PDUs in order to account for the maximum expected ac-
knowledgment delay. This implies that, if the content of this buffer is transferred in
one scheduling interval (in order to avoid downtime), the overhead traffic produced
during the migration would be between 50 and 160 times higher than the downlink
data rate of the air interface, which is actually comparable to the capacity needed for
full centralization. Indeed, in [3GP17] the capacity required for the C-RAN split of in
one gNB is estimated to be around 40 times larger than the aggregated user traffic.
Hence, an adaptive functional split would be pointless with such a high overhead,
since, if it could be supported, full centralization would probably also be supported,
thus removing the need for performing a functional split.
In order to reduce the overhead, a solution would be to put both active and inactive
buffers into a common pre-defined state just before the migration, instead of transfer-
ring the content. As the packets stored in the buffer cannot be modified, this can only
mean to empty the buffers before completing the migration. There are two options to
achieve this: either to drop the content of the old buffers, or to redirect new packet
arrivals to the new buffers while the old buffers drain normally. The former, or hard
migration, provides the fastest migration with no overhead traffic, although it implies
packet losses. The latter, or soft migration, prevents packet losses, but it may introduce
delay while the old buffers are emptied, and produces an overhead traffic equal to the
arrival rate of new user packets. In addition, a combination of both migration options,
or custom migration, can be also defined. For this option, the old buffers are given a
fixed amount of time to empty, after which all remaining content is dropped. There-
fore, a custom migration provides an adjustable trade-off between latency and packet
losses.

119
5. Implementation of an Adaptive Functional Split

Migration controller Migration controller

MP MP
lib lib MAC RLC
PHY PDCP
PR PR
MAC RLC lib lib

DU CU
Figure 5.1.: Functional architecture of the two binaries supporting an adaptive func-
tional split. The blue trapezoid symbolizes the ability of the migration
controller to steer the flow of data.

5.4.4. Migration platform

With the objective of testing the viability of the aforementioned migration scheme, we
implemented it using a modified version of srsRAN as software platform [GM+16].
As commented in Sec. 5.2.2, both OpenAirInterface and srsRAN would be good op-
tions to implement our proof-of-concept. OpenAirInterface, for example, already pro-
vides partial implementation of the MAC-PHY and PDCP-RLC splits in its original
code. Nonetheless, since it features multiple advanced characteristics, realizing mod-
ifications on its code can occasionally be more challenging than on that of srsRAN,
which is relatively simpler. Although srsRAN does not provide functional splits out
of the box, its code is clearly organized into layers, thus facilitating the implementa-
tion of functional splits. As a result, even though OpenAirInterface would also be a
valid option to realize a similar proof-of-concept, in this work we choose srsRAN as
the software platform. We split the original, monolithic srsRAN binary that contains
all the RAN layers into two different 5G binaries: one containing the PHY, MAC, and
RLC functions for the DU, and another hosting the MAC, RLC, PDCP, and RRC func-
tions for the CU (see Fig. 5.1). The code implementing the MAC and RLC functions
can be disabled during runtime in both binaries, thus providing the software basis for
switching functional splits.
In order to make these two binaries work with each other, interfaces for the PDCP-
RLC and MAC-PHY splits need to be defined. For the sake of simplicity, these inter-
faces are implemented as libraries whose purpose is to convert the original function
calls between layers into Ethernet packets. This is done by means of Google Protocol
Buffers [Goo], which is a tool that helps transforming C++ objects into serialized data.
The two libraries are known as the PR library for the PDCP-RLC split, and the MP
library for the MAC-PHY split. The PR library performs a pure serialization of the ob-
jects exchanged by the srsRAN function calls, which already resembles the application
protocol of the F1 interface in 5G [3GP21j], whereas the MP library is explicitly cus-
tomized to follow the general structure of nFAPI [Sma17]. Both libraries are logically
connected in pairs through the fronthaul network. This is also shown in Fig. 5.1.

120
 5.4. Adaptive functional split

Apart from the RAN layers and their interfaces, two migration controllers are needed
to orchestrate the migration in both units. These functions can activate or deactivate
the MAC and RLC functions of their unit, and steer the user traffic accordingly. Fur-
thermore, they have East-West interfaces to communicate with each other, in order to
coordinate the migration, and a northbound interface to receive the migration com-
mand from upper layers. In this chapter, we neglect the origin of this migration com-
mand, which can either result from an automatic decision or manual intervention. In
Chapter 6, however, we discuss this topic further by proposing multiple strategies to
automatically trigger the migration at the CU.

5.4.5. Proposed migration procedure

In this section, we describe our migration procedure between PDCP-RLC and MAC-
PHY splits. Based on what is explained in the previous section, we treat the synchro-
nization of external and internal parameters differently. We synchronize the external
parameters before the migration by duplicating configuration messages at two points.
First, when a UE performs a successful random access procedure, the data structures
defining the established radio bearers are copied to the inactive MAC and RLC func-
tions. Second, after the core has successfully registered or updated a UE, the RRC
function forwards its configuration to both active and inactive MAC and RLC func-
tions before generating the messages RRC Connection Setup or Reconfiguration.
The internal parameters and the content of the buffers are synchronized by means
of a five-step migration procedure. For the sake of brevity, we focus on a migration
from MAC-PHY to PDCP-RLC with only downlink flows. However, extending this
procedure to the other direction or to the uplink is straightforward, as this does not
affect the essential operation of MAC and RLC layers. The procedure is as follows:
1. CU request handling: The CU receives the command for a soft, hard, or custom
migration through its northbound interface. In the case of custom migration, it
also receives the maximum time allocated for the draining of RLC and HARQ
buffers.
2. CU traffic steering: The migration controller function redirects the flow of ar-
riving PDCP PDUs to the DU. As a consequence, the RLC buffer at the CU stops
receiving new arrivals and the RLC buffer at the DU starts receiving them.
3. Buffer synchronization: This step may consist of up to three stages, depending
on the type of migration.
3.1) RLC draining: Only for soft and custom migrations. The MAC scheduler
continues its normal operation until the RLC buffers are empty, or until the
allocated time runs out (in a custom migration).
3.2) MAC draining: Only for soft and custom migrations. The MAC layer at
the CU waits until the acknowledgment of the last active HARQ process is
received, or until the allocated time runs out (in a custom migration). In a

121
5. Implementation of an Adaptive Functional Split

RLC buffer MAC scheduler Used HARQ Process IDs

Before migration During migration After migration

PDCP PDCP PDCP PDCP PDCP

PHY PHY PHY PHY PHY

CU RLC buffer

DU RLC buffer Time

CU traffic steering RLC draining + MAC draining DU traffic steering

Figure 5.2.: Evolution of the internal operation of RLC and MAC layers during a soft
migration from the MAC-PHY split to the PDCP-RLC split.

soft migration, this step guarantees that no packets are lost.

3.3) MAC and RLC reset: Only for hard migrations, and custom migrations
after the allocated time. The content of the RLC and HARQ buffers is
dropped and the list of active HARQ processes is reset.
4. MAC and RLC synchronization: The current frame, subframe, and RLC se-
quence numbers are sent from the CU to the DU, and the RLC and MAC func-
tions at the DU are updated accordingly.
5. DU traffic steering: The migration controller activates the MAC and RLC func-
tions at the DU, so they can start processing the PDUs stored in the RLC buffers.
In Fig. 5.2, we show a representation of the state of the RLC buffers, HARQ PIDs, and

122
 5.5. Implementation results

user packet routes before, during, and after a soft migration from the MAC-PHY split
to the PDCP-RLC split. Before the migration, in the leftmost figure, the RLC buffer
at the CU contains four user packets and four HARQ PIDs are in use. The MAC
scheduler at the CU decides, based on information provided by the UEs, when to let
packets at the RLC buffer proceed to the MAC layer, and which HARQ PIDs to use
and release. Right after receiving the migration command (step 1), the CU disables
the route connecting outgoing PDCP packets to its RLC buffer, and instead forward
them to the PR library to be directly sent to the RLC buffer at the DU (step 2). After
this is performed, the RLC buffer at the CU keeps draining normally as decided by
the MAC scheduler (step 3) until this buffer is empty (step 3.1) and no HARQ PIDs
are in use (step 3.2). At this point, the remaining internal parameters are sent from the
MAC and RLC functions at the CU to those at the DU (step 4). Finally, the CU disables
its MAC and RLC functions and those at the DU are enabled (step 5), thus allowing
packet flow from the MAC function at the DU to the PHY layer.
This procedure produces no downtime, as it guarantees that there are always active
MAC and RLC functions during the migration. This is achieved by letting the MAC
and RLC functions at the DU store the new arriving PDUs (step 2), while the functions
at the CU are still processing those in the buffer when the migration starts. However,
the absence of downtime does not imply that the migration does not cause additional
latency to the users. Indeed, step 3.2 (MAC draining) is basically a waiting step, which
may delay the processing of PDUs arriving at that time. This delay is prevented by
hard migrations or limited by custom migrations, at the cost of potential packet losses
(step 3.3).

5.5. Implementation results

In this section, we present the details of our flexible platform implementing an adap-
tive 5G RAN, as well as measurement results. This platform is able to switch from
PDCP-RLC to MAC-PHY, or vice versa, at runtime without interrupting ongoing user
traffic. In order to do this, it follows the aforementioned procedure for replication-
based migrations of the three types: hard, soft, and custom.

5.5.1. Hardware platform description

The hardware equipment consists of four off-the-shelf Intel i7 PCs for the UE, DU, CU,
and core network. The DU and CU are connected by means of a fronthaul link, which
is realized with a 1 Gb/s Ethernet link. The same type of link is used for the backhaul,
which connects the CU to the core network. For the radio interface between the UE
and the DU, two programmable USRP B200 are used to transmit a single-carrier 10
MHz LTE signal, which implies a maximum user data rate of 31.7 Mb/s. Although
only one UE is used in this setup, the results are applicable to any multi-UE which
produces the same joint downlink rate.

123
5. Implementation of an Adaptive Functional Split

Buffer contents during soft migration

RLC Tx buffer (CU)

25
HARQ buffer (CU)
Buffer content (kB)

20 RLC Tx buffer (DU)

15

10

0
0 5 10 15 20
Time (ms)
Figure 5.3.: Content of the RLC and HARQ buffers during a soft migration. The in-
put traffic is a constant-rate 20 Mb/s stream, and the network downlink
capacity is 31.7 Mb/s. The migration starts at t = 0 ms.

5.5.2. Measurement results

In order to show the performance of our implemented 5G RAN, in this section we

present the most relevant experimental results. In Fig. 5.3, we can observe an ex-
emplary measurement of the evolution of the content of the RLC and HARQ buffers
when a soft migration from MAC-PHY to PDCP-RLC is performed, that is, when the
MAC and RLC functions are migrated from the CU to the DU. For this experiment,
a constant-rate 20 Mb/s downlink stream is used to represent the user traffic. The
service rate of the base station is 31.7 Mb/s, corresponding to the downlink data rate
supported by the air interface of a UE experiencing the highest channel quality in a
10 MHz carrier. Thus, the RAN operates at 63% capacity, on average. The data shown
in the figure is extracted directly from execution logs, which report the status of the
buffers every time they are modified, thus ensuring an update periodicity of one mil-
lisecond or less. The migration command is received by the migration controller at
time t = 0, which performs the CU traffic steering step immediately and starts drain-
ing the RLC buffer. We can see that the RLC draining step lasts a small fraction of a
millisecond, after which the RLC buffer at the CU is already empty. Then, the HARQ
buffer, which is full when the migration starts, takes around 8 ms to drain. During
that time, the RLC buffer at the DU receives PDCP PDUs from the CU, but it is not
yet ready to pass them on to the MAC layer. Shortly after the HARQ buffer at the CU
is empty, implying that all HARQ PIDs are unused, the migration is finished and the

124
 5.5. Implementation results

Figure 5.4.: Cumulative distribution functions of the RLC/MAC draining times for
different inter-arrival times ε of incoming downlink PDCP PDUs. Each
curve corresponds to 100 soft migrations.

RLC buffer at the DU starts to be emptied by the MAC layer. In this measurement,
we can clearly see how the handover between old and new MAC and RLC functions
is performed, and how this affects the normal operation of the buffers.
As mentioned in Sec. 5.4.5, the time it takes for the RLC and HARQ buffers to drain
is important, as it impacts the end-user latency. Therefore, a specific experiment is
performed in order to find out the distribution of this RLC/MAC draining time as a
function of the inter-arrival time ε of the incoming PDCP packets. The results, shown
in Fig. 5.4, allow us to conclude that the lower the inter-arrival time, the higher the
average draining time. This is to be expected, as the lower the inter-arrival time, the
more HARQ processes will be active, leading to a slower buffer draining time. In
addition, we see that the maximum RLC/MAC draining time is around 10 ms, which
corresponds to 1 ms to empty the RLC buffer and 8–9 ms to empty the HARQ buffer
(containing up to 8 MAC PDUs).
In Fig. 5.5 we show boxplots representing the distributions of the additional laten-
cies experienced by users packets that are affected by migration events. We consider
eleven types of migrations: soft, hard, and nine custom migrations whose deadlines
range from 1 to 9 ms. In these experiments, the RAN is dealing with a downlink traf-
fic of periodic packets transmitted every 0.3 ms. Besides, the interference conditions
produce a 10% packet error rate, that is, 1 out of 10 MAC PDUs has to be retransmit-
ted. We observe that hard migrations cause a negligible impact on end-user latency,
whereas soft migrations may cause an additional delay of 14 ms, with a median of
around 7 ms. Custom migrations can be used to finely tune the experienced latency
down to the acceptable values between these two extremes. Nonetheless, we conclude

125
5. Implementation of an Adaptive Functional Split

Figure 5.5.: Additional end-user latency for soft, hard, and nine custom migrations
with allocated times ranging from 1 to 9 ms. Each point represents 50
migrations from MAC-PHY to PDCP-RLC.

Figure 5.6.: Probability of experiencing a migration with packet losses for soft, hard,
and nine custom migrations with allocated times ranging from 1 to 9 ms.
Each point represents 50 migrations from MAC-PHY to PDCP-RLC, and
vertical bars represent 95%-confidence intervals.

that even soft migrations lead to very small additional latencies, which may only be a
problem for ultra-low latency use cases.
Finally, in Fig. 5.6 we show the probability of experiencing packet losses during a mi-
gration. We conclude that hard migrations are prone to suffer from packet losses, since
95% of the observed migrations resulted in packet losses. Conversely, soft migrations
can be used to guarantee no packet losses, at the cost of introducing additional end-
user latency as shown in Fig. 5.6. The probability of experiencing packet losses when
using custom migrations lay in the middle between these two extremes, showing a

126
 5.6. Summary

roughly linear decay as the deadline grows.

5.6. Summary
In previous chapters, we explore the advantages of a 5G RAN featuring the ability
of changing the functional splits of its base stations during runtime. However, this
ability is presently neither included in any deployed network, nor described in cur-
rent specifications, and only recently considered in previous research work. Owing
to this, it is unclear whether changing the functional split is actually feasible and, if it
is, how fast and costly it can be. In this chapter, we address the challenge of provid-
ing an actual implementation of a 5G RAN that is able to change its functional split
without interrupting its normal operation. Although it is developed using experimen-
tal research equipment in a simplified scenario, this proof-of-concept implementation
shows that live adaptation of the functional split is indeed possible.
After reviewing the state of the art and selecting the most adequate subset of func-
tional splits, we enumerate the challenges that changing the functional split entails.
Based on these challenges, we conclude that relying on off-the-shelf virtualization
platforms is not the best option, and instead we propose our own replication-based
migration strategy. This includes a detailed description of the state of a base station,
consisting of those parameters that need to be transferred during a migration. We
describe the hardware and software platforms that we use to implement this proof-
of-concept and list the steps that are required to perform a change in the functional
split with minimal packet losses and impact on end-user latency.
Finally, we provide actual measurements of the performance of our flexible 5G RAN
during a migration. We observe that the buffer draining time, which is directly re-
lated to the total migration time, does not exceed 10 ms in a base station providing a
10 MHz channel and working at high load. This translates into a maximum additional
end-user latency of 14 ms when using soft migrations, although this additional latency
can be completely prevented by using hard migrations. The main disadvantage of a
hard migration is, however, increased packet losses, which occur in 95% of the ob-
served experiments. If this is undesired, custom migrations can be used to select the
appropriate operating point within the trade-off between packet losses and end-user
latency, or we can resort to soft migrations to completely remove packet losses.

127
6. Dynamic Functional Split Adaptation in
Real Time

6.1. Introduction

6.1.1. Motivation, scope, and challenges

In Chapter 2 we introduce the motivation for deploying a 5G RAN that can adjust its
functional split during runtime. We base this on the different characteristics that each
option features in terms of required fronthaul capacity and interference-coordination
capabilities. Nonetheless, implementing such an adaptive network requires us to care-
fully monitor the cost associated to both operating the network in a stable state and
changing from an obsolete state to a more appropriate one. That is the reason why
we introduce a novel, dedicated cost model in Chapter 3 that captures the most rele-
vant components of the total cost in a flexible network. In addition, deciding on the
optimal functional split for a given instant is not a trivial task, but it entails solving
non-linear mixed-integer optimization problem as quickly as possible. In Chapter 4,
we lay out multiple strategies to tackle this problem, evaluate them, and finally se-
lect the most promising one. However, even if we trust that adapting the functional
split leads to enhanced performance and/or reduced cost, it is still unclear whether
realizing this adaptation in actual networks is possible. In order to demonstrate that
this adaptation is indeed feasible, in Chapter 5 we describe and evaluate a proof-of-
concept implementation of a 5G RAN that is able to dynamically change its functional
split with minimal impact on normal operation. Therefore, at this point we know that
a 5G RAN may benefit from adapting its functional split and that the migration to
the optimal functional split is possible and reasonably fast. Nevertheless, thus far we
have focused on modeling and selecting the functional split for fixed time instants,
neglecting the potential degradation of its optimality over time and the influence of
the cost of dynamically reconfiguring the network.
Mobile networks face large variability in their demands, owing mainly to the move-
ment of their users, changes in the channel quality, and user traffic variations (which
can be part of well-known patterns or result from unexpected events). Thus, the net-
work state that is optimal for a given instant may perform very suboptimally after
some time. As a result, the network operator must be able to not only select the
instantaneously optimal network configuration, but also keep track of the network
evolution and decide when moving to a new configuration is desirable. Ideally, the
operator should try to always operate in the optimal network state. Nonetheless, there

129
6. Dynamic Functional Split Adaptation in Real Time

are two reasons why this is not possible in actual deployments. On the one hand, find-
ing the optimal state requires solving an NP-Hard problem, as shown in Chapter 4.
Therefore, the only way of checking the optimality of a state is solving the problem
anew [Aro09], which may be too computationally intensive to perform at every time
instant. On the other hand, changing the network state may imply a non-negligible
action cost, as anticipated in Chapter 3 and observed in Chapter 5. Thus, it may not
be worth migrating to the optimal network state if the current state is just minimally
suboptimal, since the additional revenue of operating in the optimal state may not
compensate the action cost.
In this chapter, we propose and evaluate multiple adaptation strategies that a 5G RAN
operator may follow in order to decide in real time when to move from its current
state to the optimal state, based on the cost and benefit of doing so. Investigating
these strategies entails facing a series of challenges. First, we have to decide an ap-
propriate manner in which the network operator keeps track of the evolution of the
optimal state. Second, we must identify the time-varying parameters that may affect
the optimality of our current state. Third, we need to take into account all cost com-
ponents that contribute to the total cost of operating and changing the network state
in real time. Fourth, we have to use these information so as to find good-behaving
adaptation strategies that are implementable in a real 5G RAN and lead to minimal
total cost. Finally, we must ensure that these strategies perform correctly in realistic
network conditions.

6.1.2. Key contributions

With the intention of addressing the aforementioned challenges, this chapter bases on
our previous work [MAK21] to feature the following contributions.
1. We propose a simple but effective rule for the network operator to monitor the
optimal state of a 5G RAN from the available UE and channel information. We
base on the cost-of-flexibility model to properly configure the parameters of this
monitoring rule.
2. We model the demand of a 5G RAN in order to represent the influence of the
most relevant time-varying parameters. This demand can be used in the sub-
sequent analysis to investigate and evaluate the performance of the adaptation
strategies.
3. Based on the cost-of-flexibility model and the implementation results shown in
Chapter 5, we present a detailed cost model of an adaptive 5G RAN, which
includes the cost of selecting new states and migrating to them.
4. We derive multiple adaptation strategies with the intention of exploring a wide
range of adaptation features. Some of these strategies are very simple to imple-
ment but may result in relatively poor performance, other strategies exhibit a
good behavior but they may be difficult to implement, whereas the remaining
strategies feature balanced characteristics.

130
 6.2. Related work

5. In order to find the most appropriate adaptation strategies, we simulate their

performance against realistic 5G RAN conditions and evaluate them in terms of
achieved total cost and simplicity.
The remaining of this chapter is organized as follows. In Sec. 6.2, we summarize the
previous work related to the problem of selecting an optimal adaptation strategy in
a 5G RAN that can change its functional split. Sec. 6.3 discusses the monitoring rule
that the network operator may use in order to keep track of the optimal network state
with sufficient accuracy. In Sec. 6.4, we present the time-dependent cost modeling that
we use to derive and analyze adaptation strategies, including a model for the time-
varying demand. Sec. 6.5 lays out and discusses seven different adaptation strategies,
which cover a wide range of trade-off features that may be desirable for the network
operator. In Sec. 6.6 we apply the cost-of-flexibility model to the adaptation strategies
in order to find critical frontiers points for the action cost that the operator may use to
simplify the selection of the adaptation strategy. Sec. 6.7 shows a detailed evaluation
of the performance of the adaptation strategies, as well as the accuracy of the esti-
mated action cost frontiers. Finally, Sec. 6.8 summarizes and concludes this chapter.

6.2. Related work

We can divide the work related to this chapter into two categories. First, we have
those works tackling, to some extent, the selection of a functional split dynamically,
which have been also presented in previous chapters. For example, in [Mae+14;
Sab+13] the concept of flexible functional split is introduced, in [GS+18a; GS+18b]
the authors tackle the problem of optimally selecting the functional split at the de-
sign phase, [DGA19; DHA20; Die+21] tackle the delay characterization of a network
featuring a changing functional split, in [Cha+17b] the authors explain a platform fea-
turing reconfigurable functional splits, and in [HR18b] a dynamic functional split is
also discussed with the purpose of allocating slices in a virtual RAN framework. Nev-
ertheless, none of these works deal with the issue of optimally deciding the moment
of changing the functional split.
Since, to the best of our knowledge, there are no previous works dealing directly with
this decision, the second category of related work comprises those works addressing
similar problems or generic approaches to dynamic problems. For example, similar
problems can be found in the field of optimal decision theory, which deals with the
selection of the decisions taken by a generic agent so as to optimize an expected out-
come [DeG69; Cen00]. One such problem is optimal stopping, which addresses the
problem of selecting when to stop a certain action in order to maximize earnings or
minimize losses [AS07; BCJ19]. In our case, we can interpret that the network opera-
tor has to decide when to stop using the current state and move to the optimal state,
and thus we can use the concepts and techniques employed in this field.
In addition, our problem can be considered a dynamic optimization problem, since
our intention is to find the sequence of solutions that optimizes a dynamically-varying

131
6. Dynamic Functional Split Adaptation in Real Time

problem [CGP11]. There are several approaches to tackle a dynamic optimization

problem, such as dynamic programming [Bel03], in which the dynamic problem is
decomposed into increasingly simpler subproblems in order to find the optimal se-
quence of actions. Dynamic programming is a rather popular strategy that has been
also applied to 5G networks problems, such as for dynamic content placement
[Aye+18] or resource allocation in network slicing [Zha+18]. Alternative approaches
to solving dynamic problems include evolutionary optimization [NYB12], which uses
evolutionary algorithms to tackle hard dynamic problems, such as non-linear prob-
lems and those whose future parameters are difficult to estimate. One example of
this is our previous work [MAJK20], which uses a genetic algorithm to find a near-
optimal functional split of a simplified network. Another popular approach to tackle
dynamic problems is reinforcement learning, in which an agent automatically learns
the action rules that maximize a reward function [Sut18]. Reinforcement learning has
been successfully applied to several dynamic configuration problems in 5G networks
[Xio+19], such as resource scheduling [Com+18] or caching [SSG17].
In this chapter, we base on the concepts and techniques presented in the aforemen-
tioned works to tackle the problem of optimally adapting the functional splits in real
time for 5G RANs. To the best of our knowledge, our contribution is the first work
that addresses the dynamic selection of the optimal functional split, taking into ac-
count the continuous degradation of implemented functional splits and the cost of
reconfiguring them.

6.3. Optimal state monitoring

In Chapter 2, Sec. 2.4.2, we introduce the decision-making entity, which is in charge

of monitoring the correspondence between the current network state and the instan-
taneous demand. If the network state optimally satisfies the demand, that is, if s(τ ) =
s∗ (τ ) = χ(d(τ )) at time instant τ , where χ( · ) is the adaptation function as defined in
Sec. 3.3.2, then no special action is required. However, if s(τ ) 6= s∗ (τ ) = χ(d(τ )), the
decision-making entity has to decide whether the network should move to the optimal
state s∗ (τ ) or the current state s(τ ) is still near-optimal enough not to motivate a state
migration. As a result, this decision-making entity has two important tasks [Kel+19]:
(i) keep track of the instantaneous optimal state and (ii) decide on state adaptations.
Monitoring the evolution of the demand is a relatively simple task, since the CU,
which contains the decision-making entity, is periodically updated about the location,
traffic load, and channel quality of its served UEs [DPS18, Sec. 8.2]. Nonetheless,
attempting to figure out the corresponding optimal state for every demand may be
rather challenging. Indeed, in Chapter 4 we show that in order to obtain the optimal
state of a 5G RAN for fixed conditions we need to solve the FSSP, which is an NP-
Hard problem. This hinders the achievement of a continuously-adapted sequence of
optimal states, since tackling an NP-Hard problem is time consuming and computa-
tionally intensive. In addition, NP-Hardness implies that we cannot tell whether a

132
 6.3. Optimal state monitoring

Demands sequence Time

Time

States sequence

Figure 6.1.: Example diagram of the sequences of demands and states in the con-
sidered network, along with demand-sampling instants. The optimal
states are related to their corresponding demand as s∗i = χ(di ), where
i ∈ {A, B, C, D, E, F}.

previously optimal state is still optimal after a change in the optimization problem
unless we solve it again.
In this thesis, we suggest a simple, yet effective approach for monitoring and control-
ling the state of the network. First, the decision-making entity periodically samples
the network demand and solves the FSSP for each sample, so as to obtain an updated,
instantaneous network configuration that optimizes the revenue. We refer to the mon-
itoring period between demand samples as Zmon . The length of Zmon plays an important
role in the performance of the functional split adaptation. If the period between sam-
ples is too long, the network may spend a large portion of the time in severely obsolete
states. If it is too short, we may waste computational resources. In Sec. 6.7.3, we apply
the cost-of-flexibility model to obtain an appropriate value for Zmon , taking into ac-
count both factors. Once the new optimal state is known, the decision-making entity
chooses whether to stay in the current state or adapt to the optimal state. In Sec. 6.5
we discuss several adaptation strategies that the decision-making entity may use to
select the right option.
In Fig. 6.1 we show a diagram of possible sequences of demands and states over time
for an abstract RAN. Demands change at unforeseeable instants, and the network is
first aware of the change at the next monitoring instant, which are homogeneously
spaced with period Zmon . These instants are indexed via τ ∈ {1, ..., 5}. At these times,
the operator can decide to either set s(τ ) = s∗ (τ ) or stay at s(τ ) = s(τ − 1), depending
on the adaptation strategy. The duration of action phases between demand and state
changes are denoted as {z1 , ..., z5 }. Note that, in some cases, the demand may change
multiple times before the next monitoring instant. In fact, in an actual deployment,
we may argue that the demand changes in a continuous manner, since it is affected by
the location and activity of the UEs, as we explain in Sec. 6.4.1. This is, however, taken
into account by the cost-of-flexibility framework used in previous and subsequent
analysis.
In Chapter 3 we assume that the network can instantaneously perceive changes in
the demands, and it responds to them by immediately starting the action phase so
as to find and implement the new optimal state. The duration of this action phase is
modeled by the random variable Z. In our case, however, we do not detect demand

133
6. Dynamic Functional Split Adaptation in Real Time

changes immediately, but at a periodic rate. This is not a problem from the modeling
point-of-view, since we can simply merge the actual action phase with the difference
between demand changes (occurring at unknown instants) and the next monitoring
instants. In fact, since the time required to solve the FSSP is in the order of 1 − 2 s
(see Sec. 4.7.2), and the time to migrate the network state is in the order of 1 − 10 ms
(see Sec. 5.5.2), this demand-to-monitoring delay may dominate Z for a sufficiently
large Zmon . In that case, it is clear that the duration of the action phase Z follows
a uniform distribution ranging from 0 to Zmon . The reason is that, since monitoring
and demand changes are independent processes (due to our ignorance about when
demand changes occur), the difference between a demand change and the next mon-
itoring instant can be modeled as a uniformly distributed random variable ranging
from 0 to Zmon . Consequently:
(
1
if 0 ≤ z ≤ Zmon ,
Z ∼ U (0, Zmon ) ⇔ fZ (z) = Zmon (1)
0 otherwise.

6.4. Time-dependent cost modeling

With the intention of providing a consistent basis for the derivation and analysis of
adaptation strategies, in this section we introduce our model of time-dependent cost
components. We first define the time-dependent demand of a 5G RAN featuring an
adaptive functional split, then we present the mathematical notation to refer to time-
dependent variables, and finally we discuss the definition of time-dependent readi-
ness and action costs.

6.4.1. Demand modeling

In Chapter 3, we define the network demand as the set of all parameters that affect
the profitability of the network but cannot be directly modified by the operator. A
precise definition of demand is not specially relevant for the instantaneous problem
formulations shown in Chapter 4, since in that case the demand can be interpreted as
fixed problem parameters. Nonetheless, when addressing the evolution of cost and
performance of an adaptive 5G RAN over time, it is important to clearly state which
problem parameters are part of the demand, so as to track their variation and properly
adapt to them.
In principle, we could associate to the demand every parameter playing a role in the
performance-maximizing, operating-cost-minimizing, or readiness-cost-minimizing
FSSPs, as formulated in (P4), (P15), and (P17), respectively (see Chapter 4, Sec. 4.4,
4.5, and 4.6). However, some of these parameters describe fixed network features,
such as the sets of fronthaul edges E and nodes N, the function modeling capacity
required by each functional split ν( · ), etc. These parameters are unlikely to change
during normal network operation, and thus for simplicity we neglect them as part

134
 6.4. Time-dependent cost modeling

of the demand. Nevertheless, it is straightforward to include these parameters in the

definition of demand if this is required, since this does not affect the validity of the
subsequent modeling and analysis.
There are four different sets of variables in our FSSP formulations, as presented in
(P4), (P15), and (P17), whose values are susceptible to change over time. First, the
received signal power pu received by UE u ∈ U from its serving gNB can change, due
either to UE mobility or to changes in the channel quality. Second, the association hu
between UE u ∈ U and its serving gNBs may also vary because of handover proce-
dures. Third, the interference power iu,g received by UE u ∈ U from all gNBs is likely
to change frequently, owing again to mobility and channel fluctuations. Finally, the
aggregated downlink traffic ρg for all gNBs could also change, due to variations in re-
ceived signal quality of their served UEs. As a result, we formally define the demand
d of a dynamically adapting 5G RAN as:

d , hp, h, i, ρi , (2)

where p , [pu ] ∀u ∈ U is the vector of all signal powers, h , [hu ] ∀u ∈ U is the vector
of all gNB-to-UE associations, i , [iu,g ] ∀ hu, gi ∈ U × G is the vector of all interference
powers, and ρ , [ρg ] ∀g ∈ G is the vector of all downlink traffics. We thus use d in
our subsequent formulations to emphasize the dependency of performance or cost on
these variable parameters.

6.4.2. Time dependency modeling

Since the values of the parameters which make up the demand may change over time,
we use the following notation to refer to the value of demand d at time instant τ :

d(τ ) , hp(τ ), h(τ ), i(τ ), ρ(τ )i . (3)

This time dependency can be transmitted to every other function or variable that de-
pends on d. For example, vectors p, h, i affect the spectral efficiency ηu that each UE
u can achieve as defined in Sec. 4.4.1. Hence, we can include explicit dependency on
d(τ ) as follows:
!
pu (τ )
ηu (a, d(τ )) = log2 1 + , (4)
ς+ G
P
g=1 iu,g (τ ) · И(min(ahu , ag ))

From ηu (a, d(τ )), we can calculate the time-dependent geometric mean of spectral
efficiencies as follows:

U
!1 U
!
U
Y 1 X
ηe(a, d(τ )) , ηu (a, d(τ )) = exp log (ηu (a, d(τ ))) . (5)
u=1
U u=1

135
6. Dynamic Functional Split Adaptation in Real Time

Similarly, the variation of ρ influences the computational cost of running centralized

and decentralized functions, as discussed in Sec. 4.5. Therefore we use the notation
Kcomp (a, d(τ )) to refer to the instantaneous computational cost at time τ and, by ex-
tension, Koper (s, d(τ )), Krev (s, d(τ )), and K(s, d(τ )) to denote the operating cost, the
revenue, and readiness costs resulting from applying state s at time τ , respectively.
The same notation applies for any other cost-related function.
In principle, the time index τ could take any continuous value. In Sec. 6.3, how-
ever, we state that our preferred approach for a dynamically adapting RAN implies
sampling the network at periodic intervals, instead of continuous monitoring. Con-
sequently, we henceforth assume that the time index τ is discrete and takes values
within the set τ ∈ {1, 2, ..., τ max }, where τ max is the last considered instant. Since the
state s of the network may also change as a result of conscious adaptations at those
points, we use the notation s(τ ) to reflect the time dependency of the network state,
as it is already done in Chapter 3. In addition, we define

S , [s(1) · · · s(τ max )] (6)

as the sequence vector containing all network states in the interval τ ∈ {1, ..., τ max }.
Therefore, S is a finite realization of the stochastic sequence of states {Sj }j∈Z that is
defined and discussed in Sec. 3.3.1. Given the optimal network state s∗ (τ ) at time τ ,
we also define
S∗ , [s∗ (1) · · · s∗ (τ max )] (7)
as the sequence vector containing all optimal network states. Note that the elements
of S and S∗ may differ if the network decides not to operate in the optimal state, but
both their values are always known up to the present value of τ . Similarly, we can
define the sequence of demands

D , [d(1) · · · d(τ max )] (8)

as a finite realization of the stochastic sequence of demands {Dj }j∈Z that is defined in
Sec. 3.3.1.
Finally, we use the following shorthand notation to refer to the readiness cost of the
optimal state s∗ (τ ) with respect to current demand d(τ ) at time τ :

K ∗ (τ ) , K(s∗ (τ ), d(τ )) = K(χ(d(τ )), d(τ )), (9)

where χ( · ) is the adaptation function as defined in Sec. 3.3.2.

6.4.3. Mean readiness and action costs

In this section, we provide time-dependent estimations of the mean readiness, proac-

tion, and reaction costs over time, following the notation described above. In the end,
we combine these three cost components into a single total cost, which can be used as

136
 6.4. Time-dependent cost modeling

the minimization objective for our adaptation strategies.

Readiness cost

The readiness cost K(s, d) depends on time in two different manners. On the one
hand, the state s(τ ) of the RAN can be dynamically selected by the adaptation strategy
in order to minimize readiness cost. On the other hand, even if the state remains
constant, the demand d(τ ) of the network does change over time, which also impacts
the resulting readiness cost. Thus, the empirical mean readiness cost K(S, D) for a
finite sequence of states S and a finite sequence of demands D is calculated as:
τ max
1 X
K(S, D) , K(s(τ ), d(τ )) (10)
τ max τ =1

We can compute the instantaneous readiness cost associated to arbitrary state s and
demand d as described in Sec. 4.5 and 4.6, which can be summarized as:

K(s, d) = Koper (s, d) + Krev (s, d) (11)

= Kinst + Kcomp (a, ρ) + Krout (f ) + K
e rev (e
η (a, d)), (12)

where
G
X

Kcomp (a, ρ) = Kcomp,CU (ag )γCU + Kcomp,DU (ag )γDU ρg (13)
g=1

now includes explicit dependency on ρ. Assuming that {Sj }j∈Z and {Dj }j∈Z are sta-
tionary processes, we can define the theoretical mean readiness cost as:
τ max
1 X
K , max
lim K(s(τ ), d(τ )). (14)
τ →∞ τ max τ =1

We can also employ the cost-of-flexibility framework presented in Sec. 3.4.2 to esti-
mate the mean readiness cost K of a dynamically adapting network as:

K = αϕK∆ (0) + (1 − αϕ)βϕK

b β (ϕ), (15)

where ∞
X
K
b β (ϕ) , K∆ (δ)(1 − βϕ)δ−1 , (16)
δ=1
Z∞
1
α, FZ (t) (1 − FT (t)) dt, (17)
T
0

137
6. Dynamic Functional Split Adaptation in Real Time

Z∞
β, FZ (t)fT (t)dt, (18)
0

ϕ is the maximum flexibility parameter, which reflects the ratio of demands that could
be eventually satisfied assuming infinite action time, K∆ ( · ) is the readiness degrada-
tion function (RDF), a characteristic function of the network, FZ ( · ) is the cumulative
distribution function of the action phase duration Z, FT ( · ) is the cumulative distri-
bution function of the demand duration T , and fT ( · ) is the probability distribution
function of the demand duration T .
The utility of using (15) to compute K is twofold. On the one hand, since we can
relate the distribution of Z to the monitoring period Zmon , as stated in (1), we can
use this expression to observe the relationship between K and Zmon and thus estimate
an adequate value of Zmon . The monitoring period Zmon should be short enough to
closely keep track of the changes in the optimality of the state vector, but long enough
to be feasible (and profitable) in a real RAN. This selection is addressed in Sec. 6.7.3.
On the other hand, the dependency of (15) on ϕ allows us to model the total cost of
certain adaptation strategies, as it is discussed in Sec. 6.5. In addition, we can use this
modeling to draw conclusions for all adaptation strategies.
In order to use (15), we need to find the RDF and the distribution of T in a RAN featur-
ing a dynamically adapted functional split. We can obtain this information from either
previous experimental data or new simulations. Although it is thus required to per-
form preliminary measurements or simulations to enable the use the cost-of-flexibility
framework, once the RDF and the distribution of T are known the theoretical model
can be used to predict the readiness cost for any distribution of Z and any value of
ϕ. This removes the necessity for dedicated measurements or simulations to test the
validity of multiple adaptation strategies or Zmon values.
For estimating the distribution of T , we can simply simulate our network with a fine
time granularity and observe how frequently an optimal state vector s∗ changes over
time. This has been already performed in our previous work [MAJK20], where we
present an approximation to the actual duration of the demands with a time granu-
larity of 1 s. This can be used directly as estimates of FT (t) and fT (t), although more
advanced techniques can be also applied to yield a better approximation to the origi-
nal distribution of T . We show one such technique in Appendix A.3, which can also be
used to estimate FT (t) and fT (t) from simulations featuring coarser time granularity.
Regarding the estimation of the RDF K∆ (δ), we can also measure it directly or in-
fer it from related functions, such as the evolution of the cost of an optimal solution
over time since the moment it is first implemented (also provided in [MAJK20]). If
τ
we denote this time-based degradation function as K∆ (τ ), the following expression
holds: ∞
X
τ
K∆ (τ ) = K∆ (δ) Pr{∆ = δ|τ }, (19)
δ=0

where Pr{∆ = δ|τ } is the probability that the state delay is δ after τ time units since the

138
 6.4. Time-dependent cost modeling

optimal state was implemented. In other words, it is the probability of experiencing

∆ = δ demand changes in a time interval of length τ . We present a procedure to
τ
estimate K∆ (δ) from K∆ (τ ) in Appendix A.4.
∗
For notational convenience, we also define K as the mean readiness cost for the se-
quence of optimal states:
∗
K , max lim K(S∗ , D). (20)
τ →∞

Since the optimal state is the one that minimizes the readiness cost, it is clear the mean
readiness cost of the optimal sequence is less or equal than the mean readiness cost of
any other sequence, that is:
∗
K ≤ K. (21)

Proaction cost
P
The mean proaction cost C is the additional computational cost of finding a new
optimal state vector s∗ (τ ) every monitoring interval Zmon . We define ζ as the average
computational effort of solving the FSSP, in RC · s (as defined in Sec. 4.5). We can con-
vert this computational effort into proaction cost by using the following expression:

P ζ · γCU
C = , (22)
Zmon
where γCU is the computational cost per reference core at the CU, since we can safely
assume that solving the FSSP occurs at the CU, and Zmon is the monitoring period.
P
In Sec. 6.7.4, we provide an estimation of the value of C for our simulated parameters
and those provided by previous work. We observe that the value of this cost compo-
nent is almost negligible when compared to the readiness or the reaction cost, thus
playing little role in the profitability of the network or the selection of the adaptation
strategy.

Reaction cost

The instantaneous reaction cost C R (sini , sfin , d) reflects the effort or inconvenience of
changing the state vector from an initial state sini to a final state sfin during runtime.
This cost component is the most challenging to estimate, since the technology for
changing the functional split of a RAN during runtime is not yet widespread, thus
little details about the actual reaction cost are known. Therefore, instead of consider-
ing a single estimation of the reaction cost, we cover a wide range of values with the
intention of exploring the profitability limits of dynamic adaptation as a function of
the reaction cost.
R
As with the readiness cost, we define the empirical mean reaction cost C (S, D) of

139
6. Dynamic Functional Split Adaptation in Real Time

operating with finite state sequence S and demand sequence D as:

τ max
R 1 X
C (S, D) , C R (s(τ − 1), s(τ ), d(τ )). (23)
τ max τ =2

Assuming that {Sj }j∈Z and {Dj }j∈Z are stationary processes, we can define the theo-
retical mean reaction cost as:
τ max
R 1 X
C , max
lim C R (s(τ − 1), s(τ ), d(τ )). (24)
τ →∞ τ max τ =2

R∗
In addition, we define C as the mean reaction cost of the sequence of optimal states:
R∗ R
C lim C (S∗ , D).
, max (25)
τ →∞

The sequence of optimal states S∗ is the one featuring the highest number of state
changes out of all sensible strategies, since it makes no sense to adapt more often
than the frequency of the changes in the optimal states. Therefore it follows that the
reaction cost of always operating in the optimal state, that is, the impatient strategy, is
an upper bound to any other reasonable strategy:
R R∗
C ≤C . (26)
R∗
It is worth noting that C can also be interpreted as the average cost per migration,
since S∗ includes all potential migrations that any strategy may feature.
We identify two main sources of reaction cost: the migration cost Cmigr R
(sini , sfin , d) di-
rectly related to a state migration (including computational efforts and flow rerouting)
and the penalization cost CpenR
(sini , sfin , d) coming from penalizations due to potential
service interruptions. Hence:

C R (sini , sfin , d) = Cmigr

R
(sini , sfin , d) + Cpen
R
(sini , sfin , d) (27)

We can estimate the former subcomponent based on our previous experimental re-
sults. In Chapter 5, we present an implementation of a RAN featuring a dynamically
configurable functional split. The migration between two functional splits is accom-
plished by means of function redundancy, that is, the function to be migrated runs at
both the DU and CU simultaneously while the transmission and reception buffers are
being emptied. As a result, we model the additional instantaneous cost of migrating
from an initial state vector sini to a final state vector sfin by means of the following

140
 6.4. Time-dependent cost modeling

function:
G h
Zmigr X
R
Cmigr (sini , sfin , d) = Kcomp,CU (max(aini fin
g , ag ))γCU
Zmon g=1
!
i
+ Kcomp,DU (min(aini fin fin
g , ag ))γDU ρg − Kcomp (a ) , (28)

Z
where Zmigr is the migration length, assumed constant for simplicity. The factor Zmigr mon
scales the migration cost according to its duration, since it may only apply for a very
short interval. Indeed, all experimental migrations shown in Chapter 5 conclude in
less than 20 ms. The term −Kcomp (afin ) prevents counting the readiness cost of the final
R
state vector twice, which implies that Cmigr (s, s, d) = 0, that is, there is no migration-
related reaction cost if there is no change in the state vector. Regarding additional
routing cost, the migration strategy described in Chapter 5 does not incur in an in-
creased network usage, apart from negligible additional signaling. As a result, we do
not include an explicit dependence on f ini or f fin in Cmigr
R
(sini , sfin ).
An estimation for the value of the penalization cost Cpen R
(sini , sfin , d) is rather difficult
to obtain, since operators rarely disclose these operational details. In order to com-
R
pensate for the lack of information about Cpen (sini , sfin ) and also potential inaccuracies
R∗
R
in our estimation of Cmigr (sini , sfin ), we use C , the average cost per migration, as a
R∗
variable in our experiments. In order to decide on an appropriate range for C , we
R∗
define C I as the known contribution of the migration costs to the reaction cost:
τ max
R∗ 1 X
CI , R
Cmigr (s∗ (τ − 1), s∗ (τ )), (29)
τ max τ =2

R∗ R∗
Since C I only includes migration costs, it is clearly a lower bound of C . As a refer-
R∗
ence upper bound, we can simply multiply C I by a large value. In our experiments
R∗
we consider the following range of values for C :
R∗ R∗ R∗
CI ≤ C ≤ 2 · 106 · C I . (30)

As we show in Sec. 6.7, this wide range is chosen since it allows us to clearly see
the influence of the reaction cost on the overall profitability and identify important
frontiers that impact the selection of the adaptation strategy.

141
6. Dynamic Functional Split Adaptation in Real Time

Total cost

The mean total cost Q of a RAN featuring a dynamically-adapting functional split is

the sum of the mean readiness, proaction, and reaction costs.
P R
Q , K + C + C . (31)

This theoretical definition can be approximated for long sequences of states S and
demands states D as:
P
Q(S, D) , K(S, D) + C + C R (S, D) ≈ Q. (32)

Since Q combines all cost components in the network (operating costs, revenue, and
cost of adapting), we use it as the main indicator of the profitability of the network
when comparing scenarios and adaptation strategies.

6.5. Adaptation strategies

In Chapter 4, we present a method to find the state vector that minimizes the in-
stantaneous readiness cost, given a fixed demand. This method is applied periodi-
cally, namely every Zmon seconds, in order for the network to be updated of potential
changes in the environment that lead to a change in the optimal state. Thus, the net-
work knows at each monitoring interval whether the current state is optimal or not,
along with the readiness cost associated with the current, the optimal, and past states.
Since both operating in a suboptimal state and moving to the optimal state are costly,
we need a set of rules to decide whether staying at the current state is more beneficial
than moving to the optimal state, or vice-versa. We refer to this set of rules as the adap-
tation strategy. We make a distinction between static strategies, which feature a state
vector that does not change over time, and dynamic strategies, whose state vector may
be adapted to changes in the environment. Although it may seem counter-intuitive to
consider static strategies as adaptation strategies, in some cases the optimal rule may
be not adapting at all.
If we could accurately predict the future, we could find the optimal adaptation strat-
egy by using dynamic programming, and thus there would be no need to consider
other strategies, unless it could not be implemented efficiently. Nonetheless, in this
thesis we assume that no sophisticated method for predicting the evolution of the net-
work is available, apart from having access to basic statistics. This is done with the
intention of not having to rely on the accuracy of our predictions to justify the feasibility
and profitability of a dynamic functional split adaptation.
In this section, we present seven adaptation strategies to compare its performance
against each other. While some of them are described for reference purposes, such as
the dynamic programming strategy, others are implementable strategies that could be

142
 6.5. Adaptation strategies

used in a real RAN.

6.5.1. Uniform-static strategy

The uniform-static strategy is a static strategy in which the state vector

s(τ ) = sufst ∀τ ∈ {1, ..., τ max } (33)

is constant and minimizes the readiness cost for a uniform distribution of UEs within
the covered area. The state vector sufst can be obtained by firstly calculating the aver-
age ratios between interference and signal powers ipu,gu
∀u ∈ U, ∀g ∈ G for a uniform
UE distribution (e. g. via a straightforward simulation), then computing coefficients
`m
g,k ∀s ∈ S, ∀g, k ∈ G from these ratios, and finally use them to obtain the optimal
state vector in (P20).
The total cost of obtained when using this approach serves as a upper bound for that
of any profitable adaptation strategy, since it features no action cost and its readiness
cost is only optimal for a uniform distribution of UEs average situation. Although
probably suboptimal, this strategy can be always implemented in real RANs and re-
quires almost no information from the environment.
Since the state vector never changes, there is no action cost component for this strat-
egy. Hence, the mean total cost of the uniform-static strategy is:
ufst
Q ≈ K(Sufst , D), (34)

for long sequences of states Sufst = [sufst · · · sufst ] and demands D.

6.5.2. Mean-static strategy

Even if the network operator prefers to implement a static strategy, it may happen
that the average UE concentration in the RAN is not uniform, but instead UEs tend to
cluster around known places. In that case, it would be reasonable to take this infor-
mation into account when calculating the optimal static solution. That is, instead of
assuming a uniform distribution of UEs, the operator may use a different distribution
of UEs that better represents their average location. Thus, we define the mean-static
strategy as the static strategy in which the state vector is:

s(τ ) = smnst ∀τ ∈ {1, ..., τ max }, (35)

where smnst is the state vector that minimizes the readiness cost for the actual, average
distribution of UEs.
mnst ufst
The readiness cost Q of the mean-static strategy is lower or equal than Q if the
average distribution of users is estimated correctly, as can be trivially shown. Indeed,

143
6. Dynamic Functional Split Adaptation in Real Time

the mean-static strategy is the static strategy that yields the lowest possible readiness
cost in a well-characterized network. As a result, it is a tighter upper bound for prof-
itable dynamic strategies than the uniform-static strategy, and a lower bound for static
strategies.
Similar to that of the uniform-static strategy, the mean total cost of the mean-static
strategy is:
mnst
Q ≈ K(Smnst , D), (36)
for long sequences of states Smnst = [smnst · · · smnst ] and demands D.

6.5.3. Impatient strategy

The impatient strategy is the simplest of the dynamic strategies, as it consists in chang-
ing the state vector every time a new optimal one is detected, that is:

s(τ ) = s∗ (τ ) ∀τ ∈ {1, ..., τ max }. (37)

As extensively discussed in Chapter 4, and especially in Sec. 4.6, we define the optimal
state s∗ (τ ) as that state minimizing the instantaneous readiness cost at time τ . The
optimization problem that has to be solved to achieve this is:
M −1 X
G
∗ ∗ ∗ ref ilin isim X
hb , r , f i = arg min K b comp (b) + Krout (f ) + ξKoper (υ − 1) rgm (P20a)
b,r,f 2eηref m=1 g=1

subject to
G
X
feg ≤ ϑe ∀e ∈ E, (P1c)
g=1

feg ≥ 0 ∀e ∈ E, ∀g ∈ G, (P1d)

X X 0
 ∀n ∈ N\{n0 , ng }
g g
fe − fe = νb(bg ) for n = n0 ∀g ∈ G, (P5b)
+ −

e∈E (n) e∈E (n) −bν (bg ) for n = ng


0 ≤ rgm ≤ Rgm bm
g ∀m ∈ M,
b ∀g ∈ G, (P12b)
G
X
rgm ≥ `m m m m
g,k bg − (1 − bg )Rg ∀m ∈ M,
b ∀g ∈ G, (P12c)
k=1
X
rgm ≤ `m m
g,k bg ∀m ∈ M,
b ∀g ∈ G, (P12d)
k
b1g ≥ b2g ≥ ... ≥ bgM −1 ∀g ∈ G, (P5c)
G
b ∈ {0, 1} . (P5d)

144
 6.5. Adaptation strategies

From optimal variables b∗ and f ∗ we can obtain the optimal instantaneous state s∗ as:

s∗ , ha∗ , f ∗ i = hb∗ · (IG ⊗ 1M −1 ), f ∗ i , (38)

where Ig is the identity matrix of size g, 1m is an all-ones column vector of size m, and
⊗ is the Kronecker product.
From solving (P20) every monitoring interval, we obtain the sequence of optimal
∗ R∗
states S∗ , from which we can estimate K and C as discussed in Sec. 6.4.3. There-
impa
fore, the mean total cost Q of the impatient strategy can be approximated from a
long optimal state sequence S∗ as:
impa ∗ P R∗
Q =K + C + C ≈ Q(S∗ , D). (39)

We can see that the readiness cost achieved by this impatient strategy is always mini-
mal, while the action cost is always maximal, as we can deduce from (21) and (26).
Alternatively, we can also use the cost-of-flexibility framework of Chapter 3 to predict
impa
Q , since its readiness cost is that of a dynamic network with maximum flexibility
ϕ = 1. Therefore, we can use (15) to formulate the mean total cost as:
impa
Q b β (1) + C P + C R∗ .
= αK∆ (0) + (1 − α)β K (40)

6.5.4. Random deferral strategy

Whereas with the impatient strategy the state vector s(τ ) changes every time there
is a change in the optimal state vector, with the random deferral strategy the decision
of moving to the optimal state is taken randomly with probability ϕ. That is, the
resulting state sequence of one random realization can be formulated as:

Srds = [srds (1) · · · srds (τ max )], (41)

where (
s∗ (τ ) with probability ϕ,
srds (τ ) = rds
(42)
s (τ − 1) otherwise.

rds
The advantage of this strategy, apart from its simplicity, is that its mean total cost Q
accepts a direct modeling from the cost-of-flexibility framework, as a function of the
adaptation probability ϕ (which is equivalent to the concept of maximum flexibility)
and the mean reaction cost of the optimal state sequence:
rds
 R∗

Q ϕ, C = αϕK∆ (0) + (1 − αϕ)βϕK b β (ϕ) + ϕC R∗ . (43)

R∗
In addition, for a given C , the value of ϕ may not be arbitrary, but we can select

145
6. Dynamic Functional Split Adaptation in Real Time

ϕ = ϕords so that the total cost is minimized. We refer to this strategy as the optimal
random deferral strategy. The value of ϕords can be obtained after solving the following
equation:
∂ rds R∗

Q ϕ, C = 0. (44)
∂ϕ ϕ=ϕords

Equation (44) does not accept a closed-form solution, but it can be easily solved by
numerical methods.

6.5.5. Dynamic programming strategy

The dynamic programming strategy uses dynamic programming to find the optimal se-
quence of states given full information about the optimal states in the future, along
with their readiness and action costs. Since we assume that a real network may not
have access to sophisticated predictions, we only use the results of this strategy to
provide a lower bound to the total cost that any feasible strategy can achieve.
In order to use dynamic programming, we model the system in the following way.
We assume that we already know the full vector of optimal states S∗ for the whole
time interval {1, ..., τ max }. This optimal state vector contains τ max states, which are all
the states that our dynamic programming algorithm can choose from. In theory, we
could not only provide those states in S∗ but also the optimal states for each possi-
ble combination of monitoring intervals, or even all possible states in the state space
SG . However, either option may lead to a combinatorial explosion even for moderate
values of G or τ max , which is known as the curse of dimensionality in dynamic pro-
gramming [Pow07]. Fortunately, since our goal when using this technique is only to
provide a lower bound to the cost of non-prediction-based strategies, using just the
optimal states for every time instant is enough, since these are also the only informa-
tion that is available to the other techniques.
We define the total cost-to-go function Λ(τ, s∗ (τ 0 )) as the minimum total cost that the
network may produce when starting from time instant τ and state s∗ (τ 0 ) till τ max . From
this definition, we can define Bellman’s equation as [Bel66]:
n o
Λ(τ, s∗ (τ 0 )) = minmax Λ(τ + 1, s∗ ()) + C R (s∗ (τ 0 ), s∗ (), d(τ )) + K(s∗ (), d(τ + 1))
∈{1,...,τ }
(45)
The term C (s (τ ), s (), d(τ )) represents the reaction cost of moving from state s (τ 0 ),
R ∗ 0 ∗ ∗

which is an arbitrary state, to the next selected state s∗ (). Conversely, the term
K(s∗ (), d(τ + 1)) reflects the readiness cost of operating in the next selected state.
Equation (45) can be efficiently solved using backwards induction [Bel66]. The result
DP
is a vector of states that minimizes the total cost Q , which can be calculated as:
DP
n o
Q = minmax Λ(1, s∗ ()) . (46)
∈{1,...,τ }

146
 6.5. Adaptation strategies

This total cost is a lower bound to any non-prediction-based strategies, but it cannot be
implemented in actual RANs unless advanced prediction techniques are used, since
it requires knowledge about future optimal states.

6.5.6. Greedy strategy

Although a real 5G RAN cannot predict future optimal states, it can indeed be aware
of current past optimal states, which can be used to decide when it is a good time
to move to an optimal state. The greedy strategy follows this principle by triggering
the adaptation at local minima in the accumulated cost gap between current and the
optimal states.
Let us define function Д(τ, ) for  ∈ {0, 1, 2, ...} as:

X
Д(τ, ) , [K(s∗ (τ ), d(τ + σ)) − K ∗ (τ + σ)] , (47)
σ=0

This function is the accumulated readiness cost gap between operating at an obsolete
optimal state and the optimal readiness cost. Let us consider a simple case in which
Д(τ, ) is deterministic, fully known, and wide-sense stationary such that it only de-
pends on the gap  but not on the initial instant τ . Then, with slight abuse of notation,
we can drop the dependency on τ :

Д(τ, ) = Д() ∀τ, . (48)

Since Д() is the accumulated readiness cost gap with respect to the optimal, minimal
readiness cost, it clear that Д() is an increasing function, and thus it is easy to show
that the optimal strategy would be to change the state with a fixed periodicity min
such that:
bR
Д(min ) + C Д() + CbR
≤ , ∀ ∈ {0, 1, ...}, (49)
min 
where
R∗
bR , P max τ max C
C τ (50)
∗ ∗
τ =2 [s (τ ) 6= s (τ − 1)]

is the average reaction cost of non-zero migrations, and [ · ] is the Iverson bracket,
bR
which returns 1 if the expression inside is true. This follows from the fact that Д()+

C

is the mean total cost for a fixed , which includes the accumulated cost of not chang-
ing the state for  intervals and the cost of eventually adapting. As a result, the net-
bR
work only needs to keep track of the evolution of Д()+ 
C
every monitoring interval
and perform a state change whenever an increase is detected from the previous value,
thus following a greedy behavior. A detailed description of the procedure to set the

147
6. Dynamic Functional Split Adaptation in Real Time

Algorithm 2: Greedy strategy to dynamically select the optimal functional

split.
bR (estimated)
Input: C
1 τ ←1 // Time index
grdy ∗
2 s (τ ) ← s (τ ) // Initial adaptation
3 τlast ← τ // Time of last adaptation
4 repeat
5 τ ←τ +1
6 s∗ (τ ) ← from solving (P20) // Calculate new optimal state
7 Д0last ∗ ∗
← K(s (τlast ), d(τ )) − K (τ ) // Cost gap from previous state
Дlast +C bR Дlast +Д0last +C
bR
8 if τ −τlast −1
≤ τ −τlast
then
grdy ∗
9 s (τ ) ← s (τ ) // If relative cost gap increases, adapt
10 τlast ← τ
11 Дlast ← 0
12 else
13 Дlast ← Дlast + Д0last // If not, stay and accumulate cost gap
14 end
15 end

sequence of the state vectors

Sgrdy , [sgrdy (τ ) · · · sgrdy (τ max )] (51)

is shown in Algorithm 2.
Since the operation of the greedy strategy depends on the specific events that the
network is facing, we can only model its total cost by applying definition (32):
grdy
Q = Q(Sgrdy , D). (52)

6.5.7. Adaptive threshold strategy

The greedy strategy assumes that there is a single, deterministic accumulated readi-
ness cost gap function Д(τ, ). This is a rather unrealistic assumption, since the cost
gap between current and optimal solutions can evolve in multiple ways depending on
the initial state vector and the evolution of the network. The adaptive threshold strat-
egy strategy relaxes these assumptions and makes them closer to the actual system
dynamics, while providing a method for iterative adjustment. While being simple to

148
 6.5. Adaptation strategies

implement, if offers very promising results with respect to the other approaches.
Let us suppose that instead of a single Д() function, we have an arbitrary number
Л of them. That is, the readiness cost of a state can degrade in Л different ways
after it stops being optimal. We denote each of these functions as Дi (i ), i ∈ {1, ..., Л}.
Without loss of generality, we assume that all Дi (i ) are equally likely, i. e., they appear
at similar rates. We denote by i the time during which these functions apply, which
we can modify by performing or delaying adaptations. Then, after time Л
P
i=1 i , the
mean total cost can be computed as:
PЛ bR
i=1 Дi (i ) + ЛC
Q= P0 (53)
i=1 i

If we relax the i variables and assume they are continuous, we can find the minimum1
by setting their partial derivatives to 0:
PЛ bR
∂ i=1 Дi (i ) + ЛC
PЛ = 0, ∀i ∈ {1, ..., Л}, (54)
∂i i=1 i

which leads to the following identities:

Д01 (1 ) = Д02 (2 ) = ... = Д0Л (Л ) (55)

Л
! Л
X X
Д01 (1 ) · (i − Дi (i ) = ЛCbR (56)
i=1 i=1

where Д0i (i ) , ∂∂ i Дi (i ). The first set of equations leads to an interesting conclusion:
all derivatives of Дi (i ) reach the same value for an optimal selection of continuous i .
In our model, variables i are discrete, thus we cannot actually compute derivatives.
Nonetheless, we can approximate the continuous first derivative by the discrete first
difference:

Д0i (i ) ≈ Дi (i ) − Дi (i − 1) (57)

= K(s∗ (τ ), d(τ + i )) − K ∗ (τ + i ) (58)

If we combine (55) and (56), we can define the threshold value Д0thres at which we reach
the optimal selection of i ∀i ∈ {1, ..., Л} as:

bR + Д
C
Д0thres , Д01 (1 ) = , (59)


1
Since we model the first derivatives Д0i (i ) as positive increasing functions, it can be easily shown
that Дi (i ) and (53) are convex functions.

149
6. Dynamic Functional Split Adaptation in Real Time

where
D
1 X
Д, Д (i ) (60)
D i=1 i
and
D
1 X
, i . (61)
D i=1

That is, Д is simply the average accumulated readiness cost gap right before an adap-
tation, and  is the average duration of holding an obsolete state before an adaptation.
Both components can be easily observed by the network, and they can be used to it-
eratively find the optimal threshold Д0thres . A complete description of the procedure to
set the values the sequence of the state vectors

Sadth , [sadth (τ ) · · · sadth (τ max )] (62)

is shown in Algorithm 3.
In order to kick-start this algorithm, we can initially use the greedy algorithm to obtain
the first estimations of CbR and Д0 . Possible extensions to this algorithm may be to
thres
use an exponential moving average to compute Д and , so that the threshold responds
faster to changes in the UE distribution or channel conditions. In addition, the average
non-zero reaction cost C bR can also be iteratively calculated from the observed state
changes if a good estimation is not available.
As with the greedy strategy, we can only model the total cost of the adaptive threshold
strategy by applying definition (32):
adth
Q = Q(Sadth , D). (63)

In Table 6.1, we present a comparison of all proposed adaptation strategies, where we

show their main features and the optimality of their readiness and action costs.

6.6. Reference action cost frontiers

In a 5G RAN featuring a dynamically-adapted functional split, the relationship be-
tween the readiness cost of operating at a possibly suboptimal state and the action cost
of performing and state change is crucial. If the action cost is negligible, then there is
no objection to adapt whenever possible. Nonetheless, as the action cost grows with
respect to the readiness cost, adaptation should be more careful. Indeed, if the action
cost is high, dynamic adaptation may not be worthwhile at all, thus being static states
desirable instead. As a result, it is useful for an operator to have a set of reference action
cost frontiers in order to decide whether dynamic adaptation is worthwhile or not.
R∗
In addition to the lower bound of the reaction cost C I presented in Sec. 3.4.2, we

150
 6.6. Reference action cost frontiers

Algorithm 3: Adaptive threshold strategy to dynamically select the optimal

functional split.
bR , Д0thres (estimated)
Input: C
1 τ ←1 // Time index
2 sadth (τ ) ← s∗ (τ ) // Initial adaptation
3 τlast ← τ // Time of last adaptation
4 m←0 // Adaptations counter
5 repeat
6 τ ←τ +1
7 s∗ (τ ) ← from solving (P20) // Calculate new optimal state
8 Д0last ← K(s∗ (τlast ), τ ) − K ∗ (τ ) // Cost gap from previous state
9 Дlast ← Дlast + Д0last // Accumulate total cost gap
10 if Д0last ≥ Д0thres then // Threshold check
11 sadth (τ ) ← s∗ (τ ) // Adaptation
12 τlast ← τ
13 m←m+1
−τlast
14  ←  m+1m
+ τm+1
m 1
15 Д ← Д m+1 + Дlast m+1
b R +Д
16 Д0thres ← C 
// Update threshold
17 Дlast ← 0
18 end
19 end

Strategy Type Implementable Readiness cost Action cost

Uniform-static Static Yes Suboptimal None

Minimal
Mean-static Static Yes None
(of static strategies)
Impatient Dynamic Yes Minimal Maximal
Random deferral Dynamic Yes Suboptimal Suboptimal
Dynamic programming Dynamic No Optimal Optimal
Greedy Dynamic Yes Suboptimal Suboptimal
Adaptive threshold Dynamic Yes Suboptimal Suboptimal

Table 6.1.: Comparison of proposed adaptation strategies. The term “optimal” implies
that the readiness or action costs are such that the total cost is optimal.

151
6. Dynamic Functional Split Adaptation in Real Time

Adaptation strategies

Mean-static Impatient Optimal random deferral

Impatient frontier Dynamic frontier Static frontier

Figure 6.2.: Location of the reference action cost frontiers.

propose three theoretically-derived reference points for the reaction cost: the impatient
R∗ R∗ R∗
frontier C II , the dynamic frontier C III , and the static frontier C IV . We only focus on the
reaction cost as the main component of the action cost, since the proaction cost is
probably negligible, as we show in Sec. 6.7.4.
R∗ R∗ R∗
The theoretical values of C II , C III , and C IV can be calculated from FT (t), the RDF
K∆ (δ), and a reference upper limit to the total cost, such as that of the mean-static
mnst
strategy Q . Although we still need to simulate the behavior of our 5G RAN to
obtain accurate characterizations of these components, its use removes the need for
simulating the adaptation strategies individually. Owing to its good modeling char-
acteristics, we use the optimal random deferral strategy as the reference dynamic strategy
to calculate these frontiers.

6.6.1. Impatient frontier

R∗ R∗
The impatient frontier C II is defined as the last value of C where the impatient and
R∗
the optimal random deferral strategies are equivalent. In other words, up to C II , the
R∗
mean reaction cost C of always adapting is so low that the total cost of impatient
adaptation is near-optimal, and thus adapting the state every time the optimal state
changes is the best strategy.
The cost-of-flexibility framework presented in Chapter 3 allows us to directly com-
pute the point where the random deferral strategy first departs from the impatient
strategy. We know from Properties 3.4.9 and 3.4.10 that (15) is a decreasing and con-
vex function on ϕ (see Sec. 3.4.2 and Appendices A.2 and A.1). As a result, (43) is also

152
 6.6. Reference action cost frontiers

convex, which implies that there may be at most one value of ϕ that produces a global
minimum. Nonetheless, since ϕ ∈ [0, 1] for realistic strategies, if the ϕ that minimizes
(43) is greater or equal than 1, the ϕ that actually minimizes (43) must be ϕ = 1, in
R∗
other words, the impatient strategy. Therefore, C II must satisfy this identity:
rds R∗
∂Q (ϕ, C II )
= 0. (64)
∂ϕ
ϕ=1

R∗
After solving (64) for C II , we obtain the following expression:
∞  
R∗ X 1−α
C II = −αK∆ (0) + β K∆ (δ) 1 − 2α − β (δ − 1) (1 − β)δ−1 . (65)
δ=1
1 − β

In Fig. 6.2, we show the location of the impatient frontier in an abstract representation
of the possible relationships between the total cost and the average cost per migration
for four adaptation strategies. We see that the total cost of the mean-static strategy
R∗
does not depend on C , whereas the impatient and optimal random deferral strate-
R∗
gies exhibit linear and concave dependencies on C , respectively. It is clear that if
R∗ R∗
0 ≤ C ≤ C II the impatient strategy is optimal, hence the operator may select this
simple strategy in this region.

6.6.2. Dynamic frontier

R∗ R∗
The dynamic frontier C III is defined as the value of C where the total cost of the
impatient strategy intersects a reference upper bound, such as the total cost of the
mnst
mean-static strategy Q . This implies that the mean total cost resulting from using
R∗ R∗
dynamic strategies with a mean action cost between C II and C III is still lower than
that of static strategies, but the impatient strategy is not optimal in this region any-
R∗ R∗
more. Therefore, we define C III as the value of C that fulfills the following identity:

impa mnst
Q =Q . (66)
R∗ R∗
C = C III

R∗
After solving (40) for C III , we obtain:
R∗ mnst
C III = Q − αK∆ (0) − (1 − α)β K
b β (1). (67)

In Fig. 6.2, the dynamic frontier is represented as the point where the mean-static and
R∗ R∗
impatient strategies intersect. If C < C III , the impatient strategy may be suboptimal
R∗ R∗
but is still less costly than the mean-static strategy. However, if C ≥ C III only

153
6. Dynamic Functional Split Adaptation in Real Time

advanced dynamic strategies may result in less total cost than static strategies.

6.6.3. Static frontier

R∗ R∗
The static frontier C IV is defined as the value of C beyond which the total cost of
using dynamic adaptation strategies is comparable to the cost of implementing the
R∗ R∗
mean-static strategy. Thus, for C > C IV , using a static strategy may be the best
option, since dynamic strategies may perform just marginally better or even worse.
R∗
One possible approach to estimate C IV would be to find the mean action cost such
that:
rds R∗
∂Q (ϕ, C IV )
= 0. (68)
∂ϕ
ϕ=0
R∗ R∗
Similarly to the derivation of C II , solving (68) for C IV would give us the mean re-
action cost such that the random deferral strategy is as costly as no adaptation at all
(ϕ = 0). However, this approach has two important drawbacks. On the one hand, this
would yield the maximum reaction cost for which the random deferral approach is,
on average, less costly than an obsolete static state. We want, however, to compare dy-
namic strategies with the mean-static strategy, which is a simple but reasonable strat-
R∗
egy. On the other hand, the accuracy of solving (68) for C IV may also be low, since
a measurement- or simulation-based RDF K∆ (δ) may include few points for large δ,
as they correspond to large intervals without adaptations, which are relatively rare
events.
R∗
With the intention of both improving the accuracy and making sure that C IV is a
R∗
lower bound to the maximum admissible reaction cost, we define C IV as the point
where the cost of random deferral strategy equals the cost of the mean-static strategy
for a reference flexibility ϕ = ϕref :
rds R∗
Q (ϕref , C IV ) = Qmnst . (69)
R∗
After solving for C IV :

R∗ Qmnst
C IV = − αK∆ (0) − (1 − αϕref )β K
b β (ϕref ) (70)
ϕref

The value of ϕref should be low enough to obtain a tight lower bound, but cannot be
lower than the accuracy that we have for the RDF. Based on our simulation results,
we suggest using ϕref = 0.1, as it is shown in the next section.
In Fig. 6.2, we depict the static frontier as the point where the difference between the
optimal random deferral and the mean-static strategies surpasses a small predefined
R∗ R∗
threshold. As a result, if C > C IV , the benefits of operating a dynamically adaptive
RAN are low, thus the operator may consider implementing a static strategy instead.

154
 6.7. Experimental results

6.7. Experimental results

In this section, we simulate a realistic 5G RAN to evaluate the adaptation strategies
proposed in Sec. 6.5. In addition, we apply the cost-of-flexibility framework to derive
an adequate monitoring period for the network, as discussed in Sec. 6.4.3, and to
predict the reference action cost frontiers, as presented in Sec. 6.6.

6.7.1. Simulator description

We use a MATLAB simulator to generate the gNBs, the fronthaul network, and the
UE positions from which we obtain all required parameters to formulate the FSSP as
in (P20). This problem is tackled using a commercial optimization solver on operator-
grade hardware, and then another MATLAB simulator is used to assess the cost of the
adaptation strategies.

Simulated mobile coverage

We follow the recommendations for simulating dense urban scenarios provided in

3GPP TS38.193 [3GP20b] so as to produce simulation results as realistic as possible.
Consequently, gNBs are divided into a macro and a micro layers. The macro gNBs
are located at the nodes of a hexagonal grid with an inter-site distance of 200 m. The
micro gNBs are randomly (and uniformly) distributed over the covered area. This
results in a density of around 29 macro gNBs per square kilometer. There are three
times more micro gNBs than macro gNBs, which leads to an the average cell density
of approximately 115 gNBs/km2 . In our experiments, we set the total number of gNBs
to be G = 300, thus the covered area is 2.6 km2 , which approximately corresponds
to the center of a medium-sized or large city. Regarding the number of UEs, it is
recommended to consider 10 active UEs per gNB on average, that is, U = 3000 and a
UE density of 1150 UEs/km2 .

Time-dependency and UE mobility

Each simulation run consists of τ max = 960 monitoring intervals, which for a monitor-
ing period of Zmon = 30 s (as we estimate in Sec. 6.7.3) corresponds to a simulated time
of 8 hours per simulation run. For each monitoring interval τ , a new UE distribution is
generated as an evolution from that of τ − 1, and then (P20) is solved to obtain a new
s∗ (τ ). The distribution and mobility of UEs during this simulated time correspond
to that described in [Xu+16] for transport, entertainment, and comprehensive areas,
since those are the area types that we can most frequently find in city centers. Namely,
UEs are divided into two layers of simulated mobility. A subset of UEs roam across
the covered area without any preference for the traversed places, while the remaining
UEs move around randomly-positioned UE clusters.

155
6. Dynamic Functional Split Adaptation in Real Time

The number of UE clusters for a given simulation run is denoted by parameter $ ∈

{0, ..., 8}. This allows us to evaluate slow-varying scenarios with no UE clusters (ex-
cept for those occasionally formed by the random roaming of UEs) when $ = 0, and
fast-varying scenarios with one UE cluster per hour, on average, when $ = 8. The
rising and falling times of each cluster are randomly selected between 25 and 45 min-
utes, while their peak durations range from 15 minutes to 4 hours, in accordance with
the mobility data presented in [Xu+16]. Since the number of UE clusters $ reflects
how variable the UE distribution is, it is used as a variable in the experiments.

Fronthaul network

Based on the results of our previous work presented in Chapter 4, we set the number
of possible centralization levels to M = 4, which in our case corresponds to the PDCP-
RLC, MAC-PHY, Intra-PHY, and C-RAN functional split options. In accordance with
these options, we use the interference-cancellation vector c = h1, 0.6, 0.2, 0.01i as sug-
gested in [PM17]. This implies that there is no interference cancellation with PDCP-
RLC split, whereas interference power is reduced by 20 dB when using C-RAN.
Regarding the fronthaul network, we use the capacities vector r = h4, 8, 80, 160i Gb/s,
as provided in [3GP17]. The link capacity is set to ϑe = 1 Tb/s according to [GS+18a],
where the authors describe actual fronthaul networks on Italy, Romania, and Switzer-
land. Moreover, we base on the results shown in Chapter 4 [MAJK21] to generate the
fronthaul network with an average fronthaul network degree, defined as the ratio of links
to fronthaul switches, of 3.5 by means of a Waxman topology model [Wax88].

Computing platform and simulation time

With the intention of providing realistic simulation results, we use an operator-grade

hardware platform featuring 44 Intel Xeon E5 cores [Bas+17]. For solving the FSSP,
we use the commercial Gurobi MILP solver [Gur].
We repeat each simulation run (consisting of τ max = 960 data points) 100 times to en-
sure statistically tight results. Consequently, for each simulation experiment we have
a total of 96,000 data points, corresponding to approximately 33 days of simulated
time.

6.7.2. Derivation of K∆ (δ) and FT (t)

After performing simulations for $ ∈ {0, ..., 8} UE clusters, ξ = {0.5, 1, 2} reference

revenue-operating cost ratios, and υ = {1.5, 2, 3} revenue growth rates (81 simula-
tion instances in total, one per combination of parameters), we apply the techniques
described in Sec. 6.4.3, Appendix A.3, and Appendix A.4 to estimate the readiness
degradation function K∆ (δ) and the distribution of demand durations FT (t).

156
 6.7. Experimental results

(a) RDF for ξ = 0.5, υ = 1.5 (b) FT (t) for ξ = 0.5, υ = 1.5

(c) RDF for ξ = 1, υ = 2 (d) FT (t) for ξ = 1, υ = 2

(e) RDF for ξ = 2, υ = 3 (f) FT (t) for ξ = 2, υ = 3

Figure 6.3.: Estimated RDF K∆ (δ) and FT (t) for revenue growth rates and revenue-
operating cost ratios (ξ, υ) = {(0.5, 1.5), (1, 2), (2, 3)}, and cluster sizes $ ∈
{0, 4, 8} from simulation data. The values of the RDFs are depicted in
thousands of normalized cost units (kncu).

In Fig. 6.3, we show the resulting K∆ (δ) and FT (t) functions for a selection of pa-
rameters. We observe noticeable differences among the RDFs depending on the pa-
rameters $, ξ, and υ. As the number of UE clusters $ increases, the mean readi-
ness cost decreases for almost all state delays, which is consistent with the fact that a
higher UE concentration leads to higher achieved spectral efficiencies as observed in
Sec. 4.7. Conversely, there is an inverse relationship between $ and the degradation
rate. When there are no UE clusters ($ = 0), the degradation of the readiness cost is
much less noticeable than when $ = 4 or $ = 8. This makes intuitive sense, since
old states are less likely to operate correctly when the UE distribution changes in a
clustered scenario.

157
6. Dynamic Functional Split Adaptation in Real Time

The distribution of the duration of the demands FT (t), that is, the time between
changes in the optimal state, seems remarkably consistent for all values of $, ξ, and
υ. The explanation for this is twofold. On the one hand, we observe that the FSSP
is rather sensitive to small changes in the distribution of UEs, leading to frequent
changes in the optimal solutions. The average time between these changes, in the or-
der of one minute, is small when compared to the evolution of the UE clusters, whose
shortest raising or falling times are 25 minutes. Since short-term variations are similar
for all scenarios, FT (t) is little affected by the value of $. On the other hand, since
the performance-to-revenue function is a linear function and the values of (ξ, υ) are
within a limited range, their specific values do not substantially change the shape of
these short-term variations either, leading to a negligible influence of these parame-
ters.

6.7.3. Selection of monitoring period Zmon

Once we have K∆ (δ) and FT (t), we can use them to select an adequate value of Zmon .
As we discuss in Sec. 6.4.3, the time between a demand change, i. e., a change in
the optimal state vector, and the next opportunity for a state change is modeled by
means of the random variable Z, which follows a uniform distribution for periodic
monitoring. That is, Z ∼ U (0, Zmon ). From this and (15), we can formulate the mean
readiness cost as a function of Zmon as:

K(Zmon ) = α(Zmon )K∆ (0) + (1 − α(Zmon ))β(Zmon )K

b β(Zmon ) (1), (71)

where
Z∞
Z 
Zmon
1 t (1 − FT (t))
α(Zmon ) =  dt + (1 − FT (t)) dt , (72)
T Zmon
0 Zmon

and
Z
Zmon
tfT (t)
β(Zmon ) = 1 − FT (Zmon ) + dt. (73)
Zmon
0

Note that we have set ϕ = 1 in (71) since this is the value that minimizes the readiness
cost.
In Fig. 6.4 we show the evolution of K(Zmon ) as Zmon ranges from 0.1 s to 1000 s for
multiple values of $, ξ, and υ. We observe that the mean readiness cost stays relatively
constant for all cases until Zmon ≈ 100 s, where the long monitoring period leads to a
noticeable cost increase, specially for high values of $, ξ, and υ. Thus, we conclude
that monitoring periods in the range 0 ≤ Zmon . 100 s do not to affect substantially
the mean readiness cost that can be achieved in the simulated network. In our case,
we henceforth choose Zmon = 30 s as the monitoring interval for all simulations.

158
 6.7. Experimental results

(a) ξ = 0.5, υ = 1.5 (b) ξ = 0.5, υ = 3

(c) ξ = 2, υ = 1.5 (d) ξ = 0.5, υ = 3

Figure 6.4.: Evolution of the mean readiness cost achievable by the network as a func-
tion of the monitoring interval Zmon .

Figure 6.5.: Distribution of solving times of the FSSP as the number of UE clusters $
varies.

159
6. Dynamic Functional Split Adaptation in Real Time

Parameter Value Units Source

Kinst,CU 1 ncu [GS+18b, Table I]
Kinst,DU 0.5 ncu [GS+18b, Table I]
Kcomp,CU (a) [1 1.8 3.4 5] RC · s/Gb/s [GS+18b, Table I]
Kcomp,DU (a) [4 3.2 1.6 0] RC · s/Gb/s [GS+18b, Table I]
γCU 0.017 ncu/RC [GS+18b, Table I]
γDU 1 ncu/RC [GS+18b, Table I]
ωDU 0 ncu/Gb/s [GS+18b, Table I]
Zmon 30 s Sec. 6.7.3
Zmigr 20 ms Sec. 5.5

Table 6.2.: Summary of cost and simulation parameters.

Figure 6.6.: Average computational effort of solving the FSSP and its associated proac-
tion cost.

P
6.7.4. Proaction cost C estimation

We can compute the proaction cost from (22) as discussed in Sec. 6.4.3. Equation (22)
depends on Zmon , γCU whose values can be found in Table 6.2, and the mean computa-
tional effort ζ, which we need to measure from our simulations. In Fig. 6.5 we show
the distribution of the solving times of (P20), which is directly related to the mean
computational effort. Namely, in order to convert solving times into ζ, we just need
to multiply the mean solving time by the number of used cores (44 cores) and scale
by the consumed power, since our Intel Xeon E5 require 1.7 W per core whereas the
reference Intel i7-4770 cores in [GS+18b] require 3.5 W per core [Int].
P
The resulting computational effort ζ and its associated proaction cost C are shown
in Fig. 6.6. If we compare its value with the ranges that we deal with for the readiness
cost (see for example Fig. 6.4), we conclude that the proaction cost has a negligible
impact on the overall cost for the considered scenarios.

160
 6.7. Experimental results

(a) $ = 0, ξ = 0.5, υ = 1.5 (b) $ = 0, ξ = 2, υ = 3 (c) $ = 4, ξ = 0.5, υ = 1.5

(d) $ = 4, ξ = 2, υ = 3 (e) $ = 8, ξ = 0.5, υ = 1.5 (f) $ = 8, ξ = 2, υ = 3

Figure 6.7.: Comparison of the mean total cost Q of static and dynamic strategies as the
R∗
average action cost C grows for $ ∈ {0, 4, 8} UE clusters and readiness
cost parameters (ξ, υ) ∈ {(1, 1), (3, 3)}.

6.7.5. Comparison of adaptation strategies

In Fig. 6.7 we show the mean total cost achieved by all adaptation strategies consid-
ered in Sec. 6.5. In order to observe the influence of the reaction cost, we vary the
R∗
mean reaction cost C of the optimal state sequence, which represents the mean cost
of potentially changing the state after each monitoring interval. We explore values of
R∗
C in the following range:
R∗ R∗ R∗
CI ≤ C ≤ 2 · 106 C I , (74)
R∗ R∗
where C I is a lower bound of C defined in (29).
At first glance, we observe that the relative performance of all adaptation strategies is
rather consistent for all scenarios. That is, the dynamic programming strategy always
achieves the lowest total cost, followed by the adaptive threshold and the optimal
random deferral strategies. Moreover, the total cost achieved by these three strate-
R∗
gies is a concave function of C in the explored range, which is clearly a desirable
feature, and the adaptive threshold strategy is noticeably closer to the lower bound

161
6. Dynamic Functional Split Adaptation in Real Time

provided by dynamic programming than any other strategy. The cost of the greedy
R∗
strategy seems to grow linearly with C and is almost always worse than that of the
adaptive threshold and optimal random deferral strategy, with few exceptions. As we
R∗
expected, the cost of the impatient strategy also grows linearly with C and features
the steepest slope out of all strategies.
R∗
If we look into the values of C at which the optimal random deferral and adap-
tive threshold strategies meet the mean-static strategy, we can draw interesting con-
clusions regarding the impact of the number of UE clusters $ and the shape of the
performance-to-revenue. For instance, in Fig. 6.7a, which represents an scenario with
no UE clusters and a rather flat performance-to-revenue function (υ = 1.5), the mean
reaction cost at which the random deferral and the adaptive threshold strategies stop
R∗ R∗
being less costly than the mean static approach are C ≈ 91 ncu and C ≈ 130 ncu,
respectively. These are 5% to 8% of the readiness cost of using the mean-static strategy
mnst
(Q ≈ 1720 ncu), which implies that the network is only able to benefit from a dy-
namic adaptation if the cost of changing the functional split is 5−8% of the cost of run-
ning the network during a single monitoring interval. Nevertheless, when $ = 8 UE
clusters with the same performance-to-revenue function (Fig. 6.7e), the points where
the random deferral and adaptive threshold strategies meet the mean-static readiness
R∗ R∗
cost are C ≈ 2.1 kncu and C ≈ 4.6 kncu, respectively, which are 2.6 and 5.1 larger
mnst
than the total cost of running the mean-static strategy (Q ≈ 890 ncu). For $ = 8
and a steep performance-to-revenue function (υ = 3) as shown in Fig. 6.7f, the max-
imum mean reaction costs that we can tolerate for the random deferral and adaptive
R∗ R∗
threshold strategies are C ≈ 39 kncu and C ≈ 77 kncu, respectively, which is
almost 20 times larger compared to the previous case. We thus conclude that only in
scenarios where the UEs are uniformly distributed and the performance-to-revenue
function is almost flat, static strategies may be less costly than dynamic strategies,
whereas in the remaining cases the dynamic strategies clearly outperform the static
ones.
R∗
In Fig. 6.8 we show the same results as in Fig. 6.7 for a lower range of C , in order
to better appreciate the behavior of the dynamic strategies when the mean reaction
cost is low. We observe that the impatient, and random deferral strategies behave in
R∗ R∗
the same way when C is very small, such as C . 20 ncu in all but the case where
($, ξ, υ) = (0, 0.5, 1.5). This implies that the impatient strategy is preferable over the
greedy or adaptive threshold strategy for those cases. Conversely, it becomes clear
that the greedy strategy is never less costly than all other implementable strategies,
R∗
which allows us to discard it as a desirable option for any C .

6.7.6. Reference action-cost frontiers

Finally, we test the accuracy of the action-cost frontiers that are presented in Sec. 6.6.
R∗ R∗ R∗
In Fig. 6.9 we show the theoretical values of C II , C III , and C I (calculated from the

162
 6.7. Experimental results

(a) $ = 0, ξ = 0.5, υ = 1.5 (b) $ = 0, ξ = 2, υ = 3 (c) $ = 4, ξ = 0.5, υ = 1.5

(d) $ = 4, ξ = 2, υ = 3 (e) $ = 8, ξ = 0.5, υ = 1.5 (f) $ = 0, ξ = 2, υ = 3

Figure 6.8.: Comparison of the mean total cost Q of dynamic strategies as the average
R∗
action cost C grows for $ ∈ {0, 4, 8} UE clusters and readiness cost
parameters (ξ, υ) ∈ {(1, 1), (3, 3)}.

mnst
estimated FT (t), RDF K∆ (δ), and Q ), alongside the corresponding values obtained
R∗
by individually simulating the random deferral strategy for the C range shown in
(74), $ = {0, ..., 8} UE clusters, ξ = {0.5, 1, 2}, and υ = {1.5, 2, 3}. We observe a
close correspondence between the theoretical and simulated values, thus validating
the accuracy of the proposed model and the derivation of the RDF.
R∗
The value of C II is also depicted in Fig. 6.8, indicating the predicted points where the
random deferral strategy start producing less total cost than the impatient strategy.
We can see that this frontier is a good predictor of the point beyond which the impa-
tient strategy is clearly not optimal for all simulation instances. Thus, its value can be
used to decide whether always adapting is an adequate option or a more advanced
strategy, such as the adaptive threshold strategy, should be used.
R∗ R∗ R∗
Frontiers C III and C IV are also shown in Fig. 6.7, marking the predicted values of C
where the impatient strategy meets the mean-static strategy, and providing a lower
bound to the intersection between the random deferral strategy and the mean-static
strategy, respectively. We can see that these predictions are indeed fairly accurate,
R∗
specially for $ > 0. C III can be used to estimate the maximum mean action cost
R∗
beyond which an advanced adaptation strategy is mandatory. Nonetheless, C IV is,

163
6. Dynamic Functional Split Adaptation in Real Time

(a) ξ = 0.5, υ = 1.5 (b) ξ = 0.5, υ = 2 (c) ξ = 0.5, υ = 3

(d) ξ = 1, υ = 1.5 (e) ξ = 1, υ = 2 (f) ξ = 1, υ = 3

(g) ξ = 2, υ = 1.5 (h) ξ = 2, υ = 2 (i) ξ = 2, υ = 3

Figure 6.9.: Comparison between theoretical and simulated values of the reference
R∗ R∗ R∗
action-cost frontiers C II , C III , and C IV for $ = {0, ..., 8}, ξ = {0.5, 1, 2},
and υ = {1.5, 2, 3}.

as expected, a rather conservative frontier for dynamic strategies. This is due to two
R∗
main reasons. On the one hand, C IV is a lower bound to the actual intersection be-
tween the optimal random deferral strategy and the mean-static strategy. On the other
hand, advanced adaptation strategies such as the adaptive threshold strategy are able
to provide a total cost lower than that of the mean-static strategies even when the op-
timal random deferral strategy yields a higher cost, as we can observe in Fig. 6.7. As
R∗ R∗
a result, if C ≤ C IV we can be confident that using the adaptive threshold strategy
R∗ R∗
results in lower total cost than impatient or static strategies, whereas if C > C IV the
adaptive threshold strategy may still be a valid option, although further analysis is
required.

164
 6.8. Summary

6.8. Summary
In previous chapters, we motivate the implementation of a 5G RAN whose functional
split can be dynamically adapted arguing that it may increase the revenue of the oper-
ator. We present a comprehensive cost model to capture all the adaptation dynamics,
an approach to select the instantaneously optimal functional split, and a proof-of-
concept implementation to show that a dynamic functional split is technically feasi-
ble. In this chapter, we finally cope with the issue of estimating the revenue resulting
from a dynamic functional split adaptation.
We propose a simple strategy to monitor changes in the network demand and, after
introducing the corresponding notation, we provide a detailed description of all cost
components that need to be considered in a dynamically adapting network. We show
how the readiness degradation function can be obtained from a limited number of
measurements or simulations in order to provide a simplified estimation of the readi-
ness cost. This reduces number of simulations needed to characterize a dynamic 5G
RAN and enables using the cost-of-flexibility model to predict its behavior in other
conditions. We model the proaction and reaction costs using our proof-of-concept im-
plementation as a reference, although a wide range of possible values is included so
as to cover unforeseen cost components.
We present, describe, and compare seven adaptation strategies that a network oper-
ator may consider for running a dynamic 5G RAN. Two of these are static strategies,
which represent networks which do not change the functional split dynamically but
still intend to maximize revenue. The impatient strategy reflects a naive approach in
which the network always attempts to operate in the optimal state, regardless of the
cost of doing so. The dynamic programming strategy serves as a lower bound for all
other strategies, since it relies on unrealistic future knowledge to provide the optimal
sequence of state changes. The optimal random deferral strategy bases on the cost-
of-flexibility model to select the optimal frequency of random state adaptation. The
greedy strategy provides a small enhancement over the impatient strategy, since the
state is not always changed, but only when immediate reward is detected. Finally, the
adaptive threshold strategy improves upon the greedy strategy to provide a simple,
yet promising approach to decide when the functional split should be adapted.
Taking these strategies as a reference, we derive three action cost frontiers. The first
two, the impatient and dynamic frontiers, predict the points at which the impatient
strategy stops being optimal and better than static strategies (owing to its high action
cost), respectively. The third, static frontier, estimates the mean value of the action
cost beyond which static strategies may be desirable over dynamic strategies.
Based on the previous analysis, we finally simulate a dynamic 5G radio access net-
work in order to test the accuracy of our models and the viability of a dynamic func-
tional split adaptation. We observe that we can afford a monitoring period of up to
100 s without noticeably degrading network performance, thus allowing more than
enough time for solving the FSSP dynamically. We conclude that the proaction cost
is negligible when compared to the other cost components, owing to the short con-

165
6. Dynamic Functional Split Adaptation in Real Time

vergence time of the algorithms solving the FSSP and the low cost of computational
resources at the CU. We evaluate our proposed adaptation strategies and realize that
dynamic strategies clearly outperform static strategies in all but extreme cases. When
R∗
the action cost is low, that is, below frontier C II , the impatient strategy is enough to
achieve between 1% to 200% higher revenue (or lower cost) than static strategies. For
R∗ R∗ R∗
higher reaction costs, namely C II ≤ C ≤ C IV , the adaptive threshold approach
still provides near-optimal performance in all considered scenarios, specially in those
cases where the UEs are not uniformly distributed over the covered area. Indeed,
there is a large range of the action cost values that ensure that dynamic operation
yields higher revenue than static operation. Finally, we successfully validate the ac-
curacy of our estimated action cost frontiers by comparing it with the results of dedi-
cated simulations.

166
7. Conclusion and Outlook

Every new generation of mobile networks features a common improvement over its
predecessor: increasing user data rates. In 5G, multiple technology enhancements
have been and are being developed to achieve this goal, such as the introduction of
new modulation and coding schemes, the allocation of additional spectrum, and cell
densification. Although cell densification is often considered among the most promis-
ing options to provide better service to the users, it also entails increasing inter-cell
interference. This may hinder the achievement of the ambitious data rates that 5G
networks promise to their customers.
In order to counter the effects of inter-cell interference, mobile networks may use
interference-mitigation techniques, such as coordinated scheduling or joint transmis-
sion and reception. These techniques, however, often require high-throughput and
low-latency communication among base stations. Centralized RAN architectures may
enable this, but in reality they are difficult to implement, owing to the high require-
ments they that pose on the fronthaul networks connecting centralized and remote
units. As a result, partial function centralization is the most promising option for
implementing a 5G RAN.
Designing and deploying a partially centralized architecture implies deciding which
RAN functions should be centralized and which functions may remain distributed.
The optimality of this decision, however, depends on the instantaneous location of
the UEs, their activity, and the channel quality. Since these aspects are constantly
changing in actual networks, implementing a static functional split may result into
highly suboptimal performance. Conversely, if the RAN is able to adapt its functional
split to the instantaneous situation, it can not only improve user data rates, but also
reduce operating cost.
In this thesis, we investigate the motivation, feasibility, cost, and performance of im-
plementing a functional split that can be dynamically adapted to the instantaneous
RAN conditions. We model the implications that a functional split has over the re-
quired fronthaul capacity, interference-mitigation capabilities, operating cost, and to-
tal revenue of a 5G RAN. We formulate the selection of the best functional split as an
optimization problem and propose efficient algorithms to tackle it. In addition, we
present of a proof-of-concept implementation in order to demonstrate the feasibility
of dynamic adaptation. Finally, we derive multiple adaptation strategies so as to show
that dynamic operation is indeed effective and desirable in a wide range of realistic
conditions.

167
7. Conclusion and Outlook

7.1. Summary

This thesis presents four main contributions to the problem of optimally selecting
and adapting the functional split in 5G radio access networks. The first contribution,
discussed in Chapter 3, comprises the definition of a cost model of flexible communi-
cation networks. The second contribution, explained in Chapter 4, deals with the for-
mulation and optimal resolution of the functional split selection problem. The third
contribution, presented in Chapter 5, demonstrates the feasibility of a dynamically
adapted 5G RAN by means of a practical implementation. The fourth and final con-
tribution, described in Chapter 6, tackles the selection of the most adequate functional
splits over time. In this section, we briefly discuss the details and main conclusions of
each contribution.
Derivation of a cost model for flexible networks. Network sofwarization technolo-
gies, such as network function virtualization and software-defined networking, offer
enhanced flexibility and scalability to all types of communication networks, includ-
ing radio access networks. They allow the opportunity for networks to dynamically
adapt its operational state to changes in the demand, thus promising superior per-
formance. As a result, this feature can be exploited to improve the profitability of
the network. However, finding the appropriate state that satisfies a demand, real-
izing a state change, and spending transitional time in obsolete states may lead to
increased operating costs. In this thesis, we present a novel cost model for flexible
networks that takes into account all these cost components and combines them into a
single cost metric. We use a probability-theory framework to derive the relationships
among three main components: readiness, proaction, and reaction cost. In addition
to providing generic examples of possible applications of this model, we apply it to
the problem of dynamically selecting the optimal functional split in a 5G radio access
network. Namely, we use it to calculate an adequate sampling rate for monitoring
the evolution of the demand and to estimate the behavior of our proposed adaptation
strategies.
Proposal and evaluation of optimization approaches to select the best functional
split. The increased cell density required by 5G RAN deployments unavoidably leads
to aggravated inter-cell interference. A fully centralized RAN architecture, in which
all the processing of the base station is performed at a single data center, may be able
to mitigate this additional interference, but the current limitations of fronthaul net-
works render it infeasible. A partially centralized architecture, in which only a subset
of the processing functions is centralized, is thus the only viable option. Nonetheless,
selecting the most adequate division between centralized and distributed functions
is not a trivial task. In this thesis, we analyze three different approaches to select
the optimal functional split that maximizes proportionally-fair user data rates, min-
imizes operating cost, and minimizes readiness cost (which combines operating cost
and performance-related revenue). After deriving and assessing multiple problem
formulations, we present a readiness-cost-minimizing approach that yields near opti-
mal results with a very low convergence time.

168
 7.2. Future work

Implementation of a proof-of-concept adaptive 5G RAN. By comparing the perfor-

mance of optimally selected functional splits with that of static deployments, we are
able to show that a dynamically adaptive functional split is beneficial for the prof-
itability of a 5G RAN. Nevertheless, performing live changes in the location of the
processing functions is unprecedented for mobile networks. In order to demonstrate
that this feature is indeed feasible, in this thesis we show a detailed description of a
proof-of-concept 5G RAN implementation that is able to switch between two func-
tional split options during runtime. We build upon existing software that implements
the 4G/5G protocol stack on conventional equipment and realize the required mod-
ifications to allow dynamic operation. We describe a migration strategy that allows
changing the functional split of a base station with almost no disturbance to its nor-
mal operation. Indeed, packet losses can be fully prevented if a minimal end-to-end
additional delay (in the order of 10 ms) is permitted. Conversely, migrations can be
instantaneous if packet losses can be afforded. This implementation not only shows
that dynamically adapting the functional split is indeed feasible, but it can be used to
provide actual operational data for selecting an adequate adaptation strategy.
Proposal and evaluation of dynamic adaptation strategies. From the previous con-
tributions, we have an efficient approach to select the instantaneously-optimal func-
tional split and we are confident that the functional split can be adapted in a timely
and affordable manner. However, it is still unclear when the network operator should
trigger an adaptation of the functional split. On the one hand, the network demand
changes continuously, but this does not guarantee that these changes are substantial
enough to motivate an adaptation. On the other hand, adaptations of the functional
split may still be costly, owing to additional resource consumption and potential pe-
nalizations related to packet losses or increased end-user delay. In this thesis, we
apply the cost-of-flexibility model and dynamic optimization techniques to derive
an efficient adaptation strategy. After comparing several options, we observe that
there are indeed dynamic adaptation strategies that lead to remarkable readiness cost
reductions with respect to static approaches. We thus conclude that a dynamic func-
tional split adaptation is a feasible and very promising option to increase the efficiency
and profitability of 5G radio access networks.

7.2. Future work

In this section, we present some promising research directions for future continuation
of this work.
Extension of the cost-of-flexibility model. The current cost-of-flexibility model, as
presented in Chapter 3, only considers action-interrupting networks. These networks
interrupt their ongoing action phases if there is a new demand change, so that when
a new state is implemented it is guaranteed to be optimal. We argue that this is
probably the best strategy that a communication network facing varying demands
should follow, since it ensures dealing with the most updated information about the

169
7. Conclusion and Outlook

demands. Nevertheless, in some cases it may be possible to prefer an action-persistent

network, for which action phases are not interrupted. For example, if the demand
changes frequently and operating in obsolete states is not particularly harmful, an
action-persistent network may perform better than an action-interrupting network.
However, this behavioral difference fundamentally changes the probabilistic analy-
sis on which the cost-of-flexibility model bases. Therefore, extending the model to
include action-persistent networks may be an interesting research direction.
Realization of a more complete dynamic 5G proof-of-concept. The proof-of-concept
of a 5G RAN implementing a dynamically-adaptive functional split that we present
in Chapter 5 has several limitations. On the one hand, it only features two functional
split options, which hinders the experimentation about the features of other central-
ization levels. On the other hand, our proof-of-concept consists in two base stations
providing service to up to three UEs. This results in a very simple fronthaul network,
whose management can be tackled by uncomplicated algorithms. Therefore, a more
realistic deployment featuring multiple base stations and centralization levels may be
used to assess the effectiveness of our proposed adaptation strategies.
Use of reinforcement learning to improve adaptation strategies. In Chapter 6, we
present a collection of adaptation strategies, ranging from static and naive strategies to
more complicated approaches. Although none of these adaptation strategies requires
an accurate estimation of the future evolution of the RAN, they all rely on information
about simple network statistics, such as the average UE distribution or the expected
degradation of the readiness cost. Whereas some strategies require this information
as an external input, the adaptive threshold strategy is able to autonomously update
these statistics from its own experience. This is similar to what reinforcement learning
is able to do, and indeed there are many examples of using reinforcement learning to
improve performance of communication networks. Hence, future work could explore
the application of reinforcement learning as an alternative adaptation strategy, and
compare its advantages and disadvantages with those of our current strategies.

170
A. Proofs and derivations

A.1. Proof of Property 3.4.9

Proof. If we differentiate (20), we obtain:

∞
dK(ϕ) X
= αK∆ (0) + β(1 − 2αϕ) K∆ (δ)(1 − βϕ)δ−1
dϕ δ=1
∞
X
2
− β ϕ(1 − αϕ) K∆ (δ)(δ − 1)(1 − βϕ)δ−2 (1)
δ=1

Now we define D∆ (δ) , K∆ (δ) − K∆ (0). Since K∆ (δ) is an increasing function, D∆ (δ)
is positive and also increasing. After replacing D∆ (δ) into (1):
"∞ #
dK(ϕ) X K ∆ (0)
= αK∆ (0) + β(1 − 2αϕ) D∆ (δ)(1 − βϕ)δ−1 +
dϕ δ=1
βϕ
" ∞ #
X K ∆ (0)
− β 2 ϕ(1 − αϕ) D∆ (δ + 1)δ(1 − βϕ)δ−1 + 2
, (2)
δ=1
(βϕ)

This comes as a consequence of the following identities, whose proof is straightfor-

ward: ∞
X K∆ (0)
K∆ (0)(1 − βϕ)δ−1 = , (3)
δ=1
βϕ
∞
X K∆ (0)
K∆ (0)x(1 − βϕ)δ−1 = . (4)
δ=1
(βϕ)2
Based on the following trivial identity:

K∆ (0) K∆ (0)
αK∆ (0) + β(1 − 2αϕ) − β 2 ϕ(1 − αϕ) = 0, (5)
βϕ (βϕ)2

171
A. Proofs and derivations

we can simplify (2) into:

" ∞ #
dK(ϕ) X
= β(1 − 2αϕ) D∆ (δ)(1 − βϕ)δ−1
dϕ δ=1
" ∞ # (6)
X
− β 2 ϕ(1 − αϕ) D∆ (δ + 1)δ(1 − βϕ)δ−1
δ=1
∞
" #
X βϕ(1 − αϕ)
=β D∆ (δ) (1 − 2αϕ) − (δ − 1) (1 − βϕ)δ−1 (7)
δ=1
1 − βϕ

Each addend in the summation of (7) has three factors: D∆ (δ), which is a positive and
increasing function of δ, (1 − βϕ)δ−1 , which is a positive and decreasing function of δ,
and
βϕ(1 − αϕ)
V(δ, α, β, ϕ) , (1 − 2αϕ) − (δ − 1), (8)
1 − βϕ
which is a negative function, if 1 − 2αϕ < 0, or, otherwise, positive from 1 ≤ δ ≤ δ̃ + ,
and then negative, for some value of δ̃ + . In the former case, it is clear that dK(ϕ)
dϕ
≤ 0. In
the latter case, we divide the summation into two parts, one positive and one negative:

δ̃ +
dK(ϕ) X
=β D∆ (δ)V(δ, α, β, ϕ)(1 − βϕ)δ−1
dϕ δ=1
∞
X
+β D∆ (δ)V(δ, α, β, ϕ)(1 − βϕ)δ−1 (9)
δ=δ̃ + +1

In order for dK(ϕ)

dϕ
to be positive, the former (positive) summation in (9) has to be larger
in absolute value than the latter (negative) one. For fixed α, β, ϕ, the largest value that
D∆ (δ) could take in the range 0 ≤ δ ≤ δ̃ + is D∆ (δ̃ + ), since D∆ (δ) is an increasing
function. Similarly, a lower bound to the minimum value that D∆ (δ) can take when
δ > δ̃ + is also D∆ (δ̃ + ). As a result:
∞
dK(ϕ) X
≤ βD∆ (δ̃ + ) V(δ, α, β, ϕ)(1 − βϕ)δ−1 (10)
dϕ
δ=1 
+ 1 − 2αϕ βϕ(1 − αϕ)
= βD∆ (δ̃ ) − (11)
βϕ (βϕ)2
= −αD∆ (δ̃ + ) (12)

dK(ϕ)
Since D∆ (δ̃ + ) ≥ 0, it follows that dϕ
≤ 0.

172
 A.2. Proof of Property 3.4.10

A.2. Proof of Property 3.4.10

Proof. If we derive (7), we obtain:

∞
d2 K(ϕ)

X 2α 2β(1 − 2αϕ)
2
=β D∆ (δ) − − (δ − 1)
dϕ δ=1
1 − βϕ (1 − βϕ)2
β 2 (1 − αϕ)ϕ

+ (δ − 1)(δ − 2) (1 − βϕ)δ (13)
(1 − βϕ)3

There are three factors in each addend of the summation in (13). D∆ (δ) and (1 − βϕ)δ
are always positive, whereas

2α 2β(1 − 2αϕ) β 2 (1 − αϕ)ϕ

W(δ, α, β, ϕ) , − − (δ − 1) + (δ − 1)(δ − 2) (14)
1 − βϕ (1 − βϕ)2 (1 − βϕ)3

may be positive or negative. Namely, the first addend in (14) is negative, the third
addend is positive, and the second addend may be positive or negative depending on
the value of δ. Nonetheless, the value of the third addend grows with δ 2 , whereas the
second and first addends are linear and constant terms, respectively. Thus, it is clear
that there exists a δe+ such that W(δ, α, β, ϕ) > 0 ∀δ > δe+ , although W(0, α, β, ϕ) < 0.
This allows us to split (13) into the sum of a negative-valued and a positive-valued
summations:

δ̃ +
d2 K(ϕ) X
= β D∆ (δ)W(δ, α, β, ϕ)(1 − βϕ)δ
dϕ2 δ=1
∞
X
+β D∆ (δ)W(δ, α, β, ϕ)(1 − βϕ)δ (15)
δ=δ̃ + +1

2
In order for d dϕ
K(ϕ)
2 to be negative, the former (negative) summation in (15) has to be
larger in absolute value than the latter (positive) one. For fixed α, β, ϕ, the largest
value that D∆ (δ) could take in the range 0 ≤ δ ≤ δ̃ + is D∆ (δ̃ + ), since D∆ (δ) is an
increasing function. Similarly, a lower bound to the minimum value that D∆ (δ) can
take when δ > δ̃ + is also D∆ (δ̃ + ). As a result:
∞
d2 K(ϕ) +
X
≥ βD ∆ (δ̃ ) W(δ, α, β, ϕ)(1 − βϕ)δ (16)
dϕ2
δ=1
2α 2β(1 − 2αϕ) 2β 2 ϕ(1 − αϕ)

+
= βD∆ (δ̃ ) − − + = 0, (17)
βϕ (βϕ)2 (βϕ)3

which proves that K(ϕ) is a convex function of ϕ.

173
A. Proofs and derivations

A.3. Estimation of FT (t) from sampled durations

The distribution of the duration of demands in the network is modeled by FT (t). We

assume that this distribution is unknown, but we sample the demand every Zmon sec-
onds and denote the duration of the sampled demands by the random variable Tb ,
whose units are sampling intervals ofl length k a certain value T = t, the cor-
m Zjmon . For
t t
responding value of Tb can be either or , depending on whether the first
Zmon Zmon
sampling instant lies on the first t − btc seconds or not, respectively. As a result, we
can define the conditional probability mass function of Tb when T = t as:
 l m
t
t − btc
 if t̂ = Zmon ,
fTb |T (t̂ | T = t) = j k (18)
1 − t + btc if t̂ = t

Zmon
.

Applying the law of total probability, the probability mass function of Tb can be for-
mulated as:
Z∞
fTb (t̂) = fTb |T (t̂ | T = t)fT (t)dt. (19)
−∞

After combining (18) and (19) and some straightforward algebra, we obtain the fol-
lowing expression for fTb (t̂) as a function of FT (t):
R Zmon (t̂+1) R Zmon t̂
Zmon t̂
FT (t)dt − Zmon F (t)dt
(t̂−1) T
fTb (t̂) = R Zmon . (20)
Zmon − 0 FT (t)dt

Now, for some arbitrary t̂max , we define the following vector:

hR R Zmon t̂max i
Zmon
F= 0 FT (t)dt · · · 0 FT (t)dt , (21)

the following coefficient matrix:

 
−2+fTb (1) 1+fTb (2) fTb (3) · · · fTb (t̂max )

 1 −2 1 0 ···  
 0 1 −2 1 · · · 
H= , (22)
 
 0 0 1 −2 ···  

 0 0 0 1 ···  
.. .. .. .. ..
. . . . .

and the following coefficient vector:

f = [fTb (1) · · · fTb (t̂max )]. (23)

174
 A.4. Estimation of RDF from time-based cost degradation

We can combine F, H, and f in the following way according to (20):

F| · H = Zmon f| (24)

It is clear that if we had F, we could directly calculate FT (t) by taking derivatives.

From (24), we can find an estimate F by either computing:

F = Zmon f| H+ , (25)

where H+ denotes the Moore–Penrose inverse of H, or by minimizing kF| · H−Zmon f| k2

with additional constraints if more information is known about FTb (t).

A.4. Estimation of RDF from time-based cost

degradation
We can establish the following relationship between the time-based degradation func-
τ
tion K∆ (τ ) and the RDF K∆ (δ), as mentioned in Sec. 6.4.3:
∞
X
τ
K∆ (τ ) = K∆ (δ) Pr{∆ = δ|τ }. (26)
δ=0

The probability Pr{∆ = δ|τ } of the state delay being δ after τ time units can be written
as: ( δ+1 )
X X δ
Pr{∆ = δ | τ } = Pr Ti ≥ τ Ti < τ , (27)
i=1 i=1

since a state delay ∆ = δ at time τ since the optimal state was implemented implies
that there has been exactly δ changes in the demand since then. Let TΣδ , δi=1 Ti , we
P
can calculate (27) as:
Zτ
Pr{∆ = δ | τ } = (1 − FT (τ − t)) fTΣδ (t)dt, (28)
0

where the density function fTΣδ (t) can be obtained by convoluting fT (t) δ times with
itself. Now, for a long, arbitrary τ max , we define the following vectors:
τ
Kτ = [K∆ τ
(1) · · · K∆ (τ max )] (29)
K = [K∆ (1) · · · K∆ (τ max )] (30)
D = [Pr{∆ = δ | 1} · · · Pr{∆ = δ | τ max }] (31)

Combining (29)–(31) with (27), we derive the following relationship:

K τ = K · D| (32)

175
A. Proofs and derivations

We can solve for an estimation of K by either computing:

K = Kτ · D+ (33)

where D+ denotes the Moore–Penrose inverse of D, or by minimizing kKτ − K · D| k2

with additional constraints if more information is known about K∆ (τ ).

176
Bibliography

Publications by the author

Journal publications
[Bab+20] P. Babarczi, M. Klügel, A. Martínez Alba, M. He, J. Zerwas, P. Kalm-
bach, A. Blenk, and W. Kellerer. “A mathematical framework for mea-
suring network flexibility”. In: Computer Communications 164 (2020),
pp. 13–24.
[Die+21] L. Diez, A. Martínez Alba, W. Kellerer, and R. Agüero. “Flexible Func-
tional Split and Fronthaul delay: A queuing-based model”. In: IEEE
Access (2021). [Under review].
[He+19] M. He, A. Martínez Alba, A. Basta, A. Blenk, and W. Kellerer. “Flexi-
bility in softwarized networks: Classifications and research challenges”.
In: IEEE Communications Surveys & Tutorials 21.3 (2019), pp. 2600–
2636.
[Kel+18] W. Kellerer, A. Basta, P. Babarczi, A. Blenk, M. He, M. Klugel, and
A. Martínez Alba. “How to Measure Network Flexibility? A Proposal
for Evaluating Softwarized Networks”. In: IEEE Communications Mag-
azine (2018).
[MAJK21] A. Martínez Alba, S. Janardhanan, and W. Kellerer. “Enabling dy-
namically centralized RAN architectures in 5G and beyond”. In: IEEE
Transactions on Network and Service Management (2021). Early access.
[MAK21] A. Martínez Alba and W. Kellerer. “Dynamic Functional Split Adap-
tation in Next-Generation Radio Access Networks”. In: IEEE Transac-
tions on Network and Service Management (2021). [Under review].

Conference publications
[MA+18] A. Martínez Alba, A. Basta, J. H. Gómez Velásquez, and W. Kellerer.
“A realistic coordinated scheduling scheme for the next-generation
RAN”. In: IEEE Global Communications Conference (GLOBECOM). 2018.
[MA+21] A. Martínez Alba, P. Babarczi, A. Blenk, M. He, P. Krämer, J. Zerwas,
and W. Kellerer. “Modeling the Cost of Flexibility in Communication
Networks”. In: IEEE Conference on Computer Communications (INFO-
COM). IEEE. 2021.

177
Bibliography

[MAGVK19a] A. Martínez Alba, J. H. Gómez Velásquez, and W. Kellerer. “An adap-

tive functional split in 5G networks”. In: IEEE Conference on Com-
puter Communications Workshops (INFOCOM WKSHPS). IEEE. 2019,
pp. 410–416.
[MAGVK19b] A. Martínez Alba, J. H. Gómez Velásquez, and W. Kellerer. “Traffic
characterization of the MAC-PHY split in 5G networks”. In: IEEE
Global Communications Conference (GLOBECOM). IEEE. 2019, pp. 1–6.
[MAJK20] A. Martínez Alba, S. Janardhanan, and W. Kellerer. “Dynamics of
the flexible functional split selection in 5G networks”. In: IEEE Global
Communications Conference (GLOBECOM). IEEE. 2020, pp. 1–6.
[MAK19a] A. Martínez Alba and W. Kellerer. “A Dynamic Functional Split in 5G
Radio Access Networks”. In: IEEE Global Communications Conference
(GLOBECOM). IEEE. 2019, pp. 1–6.
[MAK19b] A. Martínez Alba and W. Kellerer. “Large-and Small-Scale Modeling
of User Traffic in 5G Networks”. In: International Conference on Network
and Service Management (CNSM). IEEE. 2019, pp. 1–5.
[MAPK21] A. Martínez Alba, S. Pundt, and W. Kellerer. “Comparison of performance-
and cost-optimal functional splits in 5G and beyond”. In: IEEE Global
Communications Conference (GLOBECOM). IEEE. 2021.

General publications
[3GP17] 3GPP. Study on new radio access technology: Radio access architecture and
interfaces. Technical Report (TR) 38.801. Version 14.0.0. 3rd Generation
Partnership Project (3GPP), Mar. 2017.
[3GP20a] 3GPP. Evolved Universal Terrestrial Radio Access Network (E-UTRAN);
X2 general aspects and principles. Technical Specification (TS) 36.420.
Version 16.0.0. 3rd Generation Partnership Project (3GPP), July 2020.
[3GP20b] 3GPP. Study on scenarios and requirements for next generation access tech-
nologies. Technical Report (TR) 38.913. Version 16.0.0. 3rd Generation
Partnership Project (3GPP), July 2020.
[3GP21a] 3GPP. 5G; NG-RAN; Architecture description. Technical Specification
(TS) 38.401. Version 16.6.0. 3rd Generation Partnership Project (3GPP),
July 2021. URL: http : / / www . 3gpp . org / \ - DynaReport / \ -
[Link].
[3GP21b] 3GPP. 5G; NR; Medium Access Control (MAC) protocol specification. Tech-
nical Specification (TS) 38.321. Version 16.5.0. 3rd Generation Partner-
ship Project (3GPP), July 2021.
[3GP21c] 3GPP. 5G; NR; Multiplexing and channel coding. Technical Specification
(TS) 38.212. Version 16.6.0. 3rd Generation Partnership Project (3GPP),
June 2021.

178
[3GP21d] 3GPP. 5G; NR; NR and NG-RAN Overall description; Stage-2. Technical
Specification (TS) 38.300. Version 16.6.0. 3rd Generation Partnership
Project (3GPP), July 2021.
[3GP21e] 3GPP. 5G; NR; Physical channels and modulation. Technical Specifica-
tion (TS) 38.211. Version 16.6.0. 3rd Generation Partnership Project
(3GPP), June 2021.
[3GP21f] 3GPP. 5G; NR; Radio Link Control (RLC) protocol specification. Technical
Specification (TS) 38.322. Version 16.2.0. 3rd Generation Partnership
Project (3GPP), Jan. 2021.
[3GP21g] 3GPP. 5G; NR; User Equipment (UE) radio transmission and reception;
Part 1: Range 1 Standalone. Technical Specification (TS) 38.101-1. Ver-
sion 16.6.0. 3rd Generation Partnership Project (3GPP), June 2021.
[3GP21h] 3GPP. 5G; NR; User Equipment (UE) radio transmission and reception;
Part 2: Range 2 Standalone. Technical Specification (TS) 38.101-2. Ver-
sion 16.6.0. 3rd Generation Partnership Project (3GPP), June 2021.
[3GP21i] 3GPP. LTE; Evolved Universal Terrestrial Radio Access (E-UTRA) and
Evolved Universal Terrestrial Radio Access Network (E-UTRAN); Overall
description; Stage 2. Technical Specification (TS) 36.300. Version 16.6.0.
3rd Generation Partnership Project (3GPP), July 2021.
[3GP21j] 3GPP. NG-RAN; F1 Application Protocol (F1AP). Technical Specifica-
tion (TS) 38.473. Version 16.6.0. 3rd Generation Partnership Project
(3GPP), Aug. 2021.
[3GP21k] 3GPP. NR; User Equipment (UE) radio access capabilities. Technical Spec-
ification (TS) 38.306. Version 16.5.0. 3rd Generation Partnership Project
(3GPP), July 2021.
[Afo+18] I. Afolabi, T. Taleb, K. Samdanis, A. Ksentini, and H. Flinck. “Network
slicing and softwarization: A survey on principles, enabling technolo-
gies, and solutions”. In: IEEE Communications Surveys & Tutorials 20.3
(2018), pp. 2429–2453.
[Alf+18] Y. Alfadhli, M. Xu, S. Liu, F. Lu, P. Peng, and G. Chang. “Real-Time
Demonstration of Adaptive Functional Split in 5G Flexible Mobile
Fronthaul Networks”. In: Optical Fiber Communications Conference and
Exposition (OFC). 2018, pp. 1–3.
[Ali+17] E. Ali, M. Ismail, R. Nordin, and N. F. Abdulah. “Beamforming tech-
niques for massive MIMO systems in 5G: overview, classification, and
trends for future research”. In: Frontiers of Information Technology &
Electronic Engineering 18.6 (2017), pp. 753–772.
[AMO13] R. Ahuja, T. Magnanti, and J. Orlin. Network Flows: Theory, Algorithms,
and Applications. Always learning. Pearson, 2013. ISBN: 9781292042701.
[Aro09] S. Arora. Computational complexity: a modern approach. Cambridge New
York: Cambridge University Press, 2009. ISBN: 9780521424264.

179
Bibliography

[AS07] A. Aries and A. Shiryaev. Optimal Stopping Rules. Stochastic Mod-

elling and Applied Probability. Springer Berlin Heidelberg, 2007. ISBN:
9783540740100.
[Aye+18] T. M. Ayenew, D. Xenakis, N. Passas, and L. Merakos. “Dynamic pro-
gramming based content placement strategy for 5G and beyond cellu-
lar networks”. In: IEEE 23rd International Workshop on Computer Aided
Modeling and Design of Communication Links and Networks (CAMAD).
IEEE. 2018, pp. 1–6.
[Bae+19] J. H. Bae, A. Abotabl, H.-P. Lin, K.-B. Song, and J. Lee. “An overview
of channel coding for 5G NR cellular communications”. In: APSIPA
Transactions on Signal and Information Processing 8 (2019), e17.
[Bar+15] J. Bartelt, P. Rost, D. Wubben, J. Lessmann, B. Melis, and G. Fettweis.
“Fronthaul and backhaul requirements of flexibly centralized radio
access networks”. In: IEEE Wireless Communications 22.5 (2015), pp. 105–
111.
[Bas+17] A. Basta, A. Blenk, K. Hoffmann, H. J. Morper, M. Hoffmann, and W.
Kellerer. “Towards a cost optimal design for a 5G mobile core net-
work based on SDN and NFV”. In: IEEE Transactions on Network and
Service Management 14.4 (2017), pp. 1061–1075.
[Bas+19] D. Basu, X. Wang, Y. Hong, H. Chen, and S. Bressan. “Learn-as-you-
go with megh: Efficient live migration of virtual machines”. In: IEEE
Transactions on Parallel and Distributed Systems 30.8 (2019), pp. 1786–
1801.
[BCJ19] S. Becker, P. Cheridito, and A. Jentzen. “Deep optimal stopping”. In:
Journal of Machine Learning Research 20 (2019), p. 74.
[Bel03] R. Bellman. Dynamic Programming. Dover Books on Computer Science
Series. Dover Publications, 2003. ISBN: 9780486428093.
[Bel66] R. Bellman. “Dynamic programming”. In: Science. Dover Books on
Computer Science Series 153.3731 (1966), pp. 34–37.
[Ber+14] D. Bergman, A. A. Cire, A. Sabharwal, H. Samulowitz, V. Saraswat,
and W.-J. van Hoeve. “Parallel combinatorial optimization with deci-
sion diagrams”. In: International Conference on AI and OR Techniques in
Constriant Programming for Combinatorial Optimization Problems. Springer.
2014, pp. 351–367.
[Bet+17] M. Bettner, S. Haka, J. Williams, and J. Carcello. Financial & Managerial
Accounting. McGraw-Hill Education, 2017. ISBN: 9781259692406.
[BGH92] D. P. Bertsekas, R. G. Gallager, and P. Humblet. Data networks. Vol. 2.
Prentice-Hall International New Jersey, 1992.

180
[BGP16] J. S. Borrero, C. Gillen, and O. A. Prokopyev. “A simple technique
to improve linearized reformulations of fractional (hyperbolic) 0–1
programming problems”. In: Operations Research Letters 44.4 (2016),
pp. 479–486.
[Bil95] P. Billingsley. Probability and Measure. Wiley Series in Probability and
Statistics. Wiley, 1995. ISBN: 9780471007104.
[BK16] N. Bizanis and F. A. Kuipers. “SDN and virtualization solutions for
the Internet of Things: A survey”. In: IEEE Access 4 (2016), pp. 5591–
5606.
[BKM20] C. Bouras, A. Kollia, and E. Maligianni. “Techno-economic Compar-
ison of Cognitive Radio and Software Defined Network (SDN) Cost
Models in 5G Networks”. In: Wireless Personal Communications (2020).
[BNK16] C. H. Benet, K. A. Noghani, and A. J. Kassler. “Minimizing Live VM
Migration Downtime Using OpenFlow Based Resiliency Mechanisms”.
In: IEEE International Conference on Cloud Networking (Cloudnet). 2016,
pp. 27–32.
[BNP16] C. Bouras, P. Ntarzanos, and A. Papazois. “Cost modeling for SDN/NFV
based mobile 5G networks”. In: International Congress on Ultra Modern
Telecommunications and Control Systems and Workshops (ICUMT). IEEE.
2016, pp. 56–61. ISBN: 9781467388184.
[Bou+09] G. Boudreau, J. Panicker, N. Guo, R. Chang, N. Wang, and S. Vrzic.
“Interference coordination and cancellation for 4G networks”. In: IEEE
Communications Magazine 47.4 (2009), pp. 74–81.
[Bou+15] M. Bouet, J. Leguay, T. Combe, and V. Conan. “Cost-based placement
of vDPI functions in NFV infrastructures”. In: International Journal of
Network Management (2015).
[Bou+18] C. Bouras, S. Kokkalis, A. Kollia, and A. Papazois. “Techno-economic
comparison of MIMO and DAS cost models in 5G networks”. In:
Wireless Networks (2018).
[Cas18] B. Casemore. How Network Disaggregation Facilitates Datacenter and IT
Modernization. White Paper. IDC, 2018.
[CC14] W. Cerroni and F. Callegati. “Live migration of virtual network func-
tions in cloud-based edge networks”. In: IEEE International Conference
on Communications (ICC). 2014, pp. 2963–2968.
[Cen00] N. Cencov. Statistical Decision Rules and Optimal Inference. Translations
of mathematical monographs. American Mathematical Society, 2000.
ISBN: 9780821813478.

[CGP11] C. Cruz, J. R. González, and D. A. Pelta. “Optimization in dynamic

environments: a survey on problems, methods and measures”. In: Soft
Computing 15.7 (2011), pp. 1427–1448.

181
Bibliography

[Cha+17a] O. Chabbouh, S. B. Rejeb, N. Agoulmine, and Z. Choukair. “Cloud

RAN Architecture Model Based upon Flexible RAN Functionalities
Split for 5G Networks”. In: International Conference on Advanced Infor-
mation Networking and Applications Workshops (WAINA). 2017, pp. 184–
188.
[Cha+17b] C. Chang, N. Nikaein, R. Knopp, T. Spyropoulos, and S. S. Kumar.
“FlexCRAN: A flexible functional split framework over ethernet fron-
thaul in Cloud-RAN”. In: IEEE International Conference on Communica-
tions (ICC). 2017, pp. 1–7.
[Che+14a] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S.
Berger, and L. Dittmann. “Cloud RAN for mobile networks–A tech-
nology overview”. In: IEEE Communications surveys & tutorials 17.1
(2014), pp. 405–426.
[Che+14b] S. Chen, H. Xu, D. Liu, B. Hu, and H. Wang. “A vision of IoT: Ap-
plications, challenges, and opportunities with China perspective”. In:
IEEE Internet of Things journal 1.4 (2014), pp. 349–359.
[Che+16a] A. Checko, A. P. Avramova, M. S. Berger, and H. L. Christiansen.
“Evaluating C-RAN fronthaul functional splits in terms of network
level energy and cost savings”. In: Journal of Communications and Net-
works 18.2 (2016), pp. 162–172.
[Che+16b] S. Chen, S. Sun, Q. Gao, and X. Su. “Adaptive beamforming in TDD-
based mobile communication systems: State of the art and 5G re-
search directions”. In: IEEE Wireless Communications 23.6 (2016), pp. 81–
87.
[Cis17] Cisco. Prepare to succeed with the Internet of Things. White Paper. 2017.
[Cis19] Cisco public. Cisco visual networking index: global mobile data traffic fore-
cast update, 2017–2022. White Paper. 2019.
[Cis20] Cisco public. Cisco annual internet report (2018–2023). White Paper.
2020.
[Com+18] I.-S. Comşa, S. Zhang, M. E. Aydin, P. Kuonen, Y. Lu, R. Trestian, and
G. Ghinea. “Towards 5G: A reinforcement learning-based scheduling
solution for data traffic management”. In: IEEE Transactions on Net-
work and Service Management 15.4 (2018), pp. 1661–1675.
[Dav+13] A. Davydov, G. Morozov, I. Bolotin, and A. Papathanassiou. “Evalua-
tion of joint transmission CoMP in C-RAN based LTE-A HetNets with
large coordination areas”. In: IEEE Globecom Workshops (GC Wkshps).
IEEE. 2013, pp. 801–806.
[DeG69] M. DeGroot. Optimal statistical decisions. New York: McGraw-Hill, 1969.
ISBN: 0070162425.

182
[DGA19] L. Diez, V. Gonzalez, and R. Agüero. “Minimizing Delay in NFV 5G
Networks by Means of Flexible Split Selection and Scheduling”. In:
IEEE Vehicular Technology Conference (VTC). IEEE. 2019, pp. 1–6.
[DHA20] L. Diez, C. Hervella, and R. Agüero. “Understanding the performance
of flexible functional split in 5G vRAN Controllers: A Markov Chain-
based model”. In: IEEE Transactions on Network and Service Manage-
ment 18.1 (2020), pp. 456–468.
[Döt+13] U. Dötsch, M. Doll, H.-P. Mayer, F. Schaich, J. Segel, and P. Sehier.
“Quantitative analysis of split base station processing and determi-
nation of advantageous architectures for LTE”. In: Bell Labs Technical
Journal 18.1 (2013), pp. 105–128.
[Dow05] A. B. Downey. “Lognormal and Pareto distributions in the Internet”.
In: Computer Communications 28.7 (2005), pp. 790–801.
[DPS18] E. Dahlman, S. Parkvall, and J. Skold. 5G NR: The Next Generation
Wireless Access Technology. Elsevier Science, 2018. ISBN: 9780128143247.
[Dur10] R. Durrett. Probability: Theory and Examples. Cambridge Series in Sta-
tistical and Probabilistic Mathematics. Cambridge University Press,
2010. ISBN: 9781139491136.
[EAN19] N. C. Ericsson AB Huawei Technologies Co. Ltd and Nokia. Common
Public Radio Interface: eCPRI Interface Specification. Interface Specifica-
tion. Version 2.0. May 2019.
[EIS75] S. Even, A. Itai, and A. Shamir. “On the complexity of time table
and multi-commodity flow problems”. In: 16th Annual Symposium on
Foundations of Computer Science (sfcs 1975). IEEE. 1975, pp. 184–193.
[Eng+98] S. Engstrom, T. Johansson, F. Kronestedt, M. Larsson, S. Lidbrink, and
H. Olofsson. “Multiple reuse patterns for frequency planning in GSM
networks”. In: VTC’98. 48th IEEE Vehicular Technology Conference. Path-
way to Global Wireless Revolution (Cat. No. 98CH36151). Vol. 3. IEEE.
1998, pp. 2004–2008.
[EPP19] A. Eira, J. Pedro, and J. J. Pires. “Modeling Cost Versus Flexibility
in Optical Transport Networks”. In: Journal of Lightwave Technology
(2019).
[Eur17] European Union. Flash Eurobarometer 456. SMEs, resource efficiency and
green markets. Report. 2017.
[Eur21] European Union. Special Eurobarometer 513. Climate Change. Report.
2021.
[FBB19] N. Farrugia, J. A. Briffa, and V. Buttigieg. “Solving the Multi-Commodity
Flow Problem using a Multi-Objective Genetic Algorithm”. In: IEEE
Congress on Evolutionary Computation (CEC). IEEE. 2019, pp. 2816–2823.

183
Bibliography

[FLF19] C.-H. Fang, P.-R. Li, and K.-T. Feng. “Joint interference cancellation
and resource allocation for full-duplex cloud radio access networks”.
In: IEEE Transactions on Wireless Communications 18.6 (2019), pp. 3019–
3033.
[Fou+16] X. Foukas, N. Nikaein, M. M. Kassem, M. K. Marina, and K. Konto-
vasilis. “FlexRAN: A flexible and programmable platform for software-
defined radio access networks”. In: Proceedings of the 12th International
on Conference on emerging Networking EXperiments and Technologies (CoNEXT).
2016, pp. 427–441.
[FYK91] K. Furuta, M. Yamakita, and S. Kobayashi. “Swing up control of in-
verted pendulum”. In: IECON. Vol. 91. 1991, pp. 2193–2198.
[GA18] K. Govindaraj and A. Artemenko. “Container live migration for la-
tency critical industrial applications on edge computing”. In: IEEE
23rd International Conference on Emerging Technologies and Factory Au-
tomation (ETFA). Vol. 1. IEEE. 2018, pp. 83–90.
[Gal91] M. Gallagher. “Proportionality, disproportionality and electoral sys-
tem”. In: Electoral studies 10.1 (1991), pp. 33–51.
[GAO17] R. A. Gustav A. Oertzen. On the Technical Future of the Telecommunica-
tions Industry. White Paper. Oliver Wyman, 2017.
[Gha+15] M. Ghaznavi, A. Khan, N. Shahriar, K. Alsubhi, R. Ahmed, and R.
Boutaba. “Elastic virtual network function placement”. In: IEEE 4th
International Conference on Cloud Networking, CloudNet 2015. 2015. ISBN:
9781467395014.
[Glo15] G. Glockner. Parallel and distributed optimization with gurobi optimizer.
2015.
[Glo75] F. Glover. “Improved linear integer programming formulations of non-
linear integer problems”. In: Management Science 22.4 (1975), pp. 455–
460.
[GM+16] I. Gomez-Miguelez, A. Garcia-Saavedra, P. D. Sutton, P. Serrano, C.
Cano, and D. J. Leith. “srsLTE: an open-source platform for LTE evo-
lution and experimentation”. In: Proceedings of the Tenth ACM Interna-
tional Workshop on Wireless Network Testbeds, Experimental Evaluation,
and Characterization. ACM. 2016, pp. 25–32.
[GS+18a] A. Garcia-Saavedra, X. Costa-Perez, D. J. Leith, and G. Iosifidis. “Flu-
idRAN: Optimized vRAN/MEC orchestration”. In: IEEE Conference
on Computer Communications (INFOCOM). IEEE. 2018, pp. 2366–2374.
[GS+18b] A. Garcia-Saavedra, G. Iosifidis, X. Costa-Perez, and D. J. Leith. “Joint
optimization of edge computing architectures and radio access net-
works”. In: IEEE Journal on Selected Areas in Communications 36.11 (2018),
pp. 2433–2443.

184
[GW74] F. Glover and E. Woolsey. “Converting the 0–1 polynomial program-
ming problem to a 0–1 linear program”. In: Operations research 22.1
(1974), pp. 180–182.
[Ham+13] A. S. Hamza, S. S. Khalifa, H. S. Hamza, and K. Elsayed. “A survey
on inter-cell interference coordination techniques in OFDMA-based
cellular networks”. In: IEEE Communications Surveys & Tutorials 15.4
(2013), pp. 1642–1670.
[Hor86] R. Horst. “A general class of branch-and-bound methods in global
optimization with some new approaches for concave minimization”.
In: Journal of Optimization Theory and Applications 51.2 (1986), pp. 271–
291.
[HP17] A. Hajisami and D. Pompili. “Dynamic joint processing: Achieving
high spectral efficiency in uplink 5G cellular networks”. In: Computer
Networks 126 (2017), pp. 44–56.
[HR18a] D. Harutyunyan and R. Riggio. “Flex5G: Flexible Functional Split in
5G Networks”. In: IEEE Transactions on Network and Service Manage-
ment 15.3 (2018), pp. 961–975.
[HR18b] D. Harutyunyan and R. Riggio. “Flex5G: Flexible functional split in
5G networks”. In: IEEE Transactions on Network and Service Manage-
ment 15.3 (2018), pp. 961–975.
[Hue+08] R. Huelsermann, M. Gunkel, C. Meusburger, and D. A. Schupke. “Cost
modeling and evaluation of capital expenditures in optical multilayer
networks”. In: Journal of Optical Networking 7.9 (2008), pp. 814–833.
[HW10] H. Huang and L. Wang. “P&P: A combined push-pull model for re-
source monitoring in cloud computing environment”. In: IEEE 3rd In-
ternational Conference on Cloud Computing (CLOUD). 2010. ISBN: 9780769541303.
[Ips20] Ipsos. 5G Awareness and Needs. European Study. 2020.
[IT12] ITU-T. Framework of network virtualization for future networks. Recom-
mendation Y.3011. Telecommunication Standardization Sector of ITU,
Jan. 2012.
[IT18] ITU-T. Transport network support of IMT-2020/5G. Technical Report GSTR-
TN5G. Telecommunication Standardization Sector of ITU, Feb. 2018.
[Jun+10] V. Jungnickel, A. Forck, S. Jaeckel, F. Bauermeister, S. Schiffermueller,
S. Schubert, S. Wahls, L. Thiele, T. Haustein, W. Kreher, et al. “Field
trials using coordinated multi-point transmission in the downlink”.
In: IEEE 21st International Symposium on Personal, Indoor and Mobile
Radio Communications Workshops. IEEE. 2010, pp. 440–445.
[Jun+14] V. Jungnickel, K. Manolakis, W. Zirwas, B. Panzner, V. Braun, M. Los-
sow, M. Sternad, R. Apelfröjd, and T. Svensson. “The role of small
cells, coordinated multipoint, and massive MIMO in 5G”. In: IEEE
communications magazine 52.5 (2014), pp. 44–51.

185
Bibliography

[KDV08] G. Kesidis, A. Das, and G. de Veciana. “On flat-rate and usage-based

pricing for tiered commodity internet services”. In: Annual Conference
on Information Sciences and Systems. IEEE. 2008, pp. 304–308.
[Kel+19] W. Kellerer, P. Kalmbach, A. Blenk, A. Basta, M. Reisslein, and S.
Schmid. “Adaptable and data-driven softwarized networks: Review,
opportunities, and challenges”. In: Proceedings of the IEEE 107.4 (2019),
pp. 711–731.
[KH05] H. Kim and Y. Han. “A proportional fair scheduling for multicar-
rier transmission systems”. In: IEEE Communications letters 9.3 (2005),
pp. 210–212.
[Klü+19] M. Klügel, M. He, W. Kellerer, and P. Babarczi. “A Mathematical Mea-
sure for Flexibility in Communication Networks”. In: IFIP Networking
Conference (IFIP Networking). IEEE. 2019, pp. 1–9.
[KMT98] F. P. Kelly, A. K. Maulloo, and D. K. H. Tan. “Rate control for commu-
nication networks: shadow prices, proportional fairness and stabil-
ity”. In: Journal of the Operational Research society 49.3 (1998), pp. 237–
252.
[Kos+12] C. Kosta, B. Hunt, A. U. Quddus, and R. Tafazolli. “On interference
avoidance through inter-cell interference coordination (ICIC) based
on OFDMA mobile systems”. In: IEEE Communications Surveys & Tu-
torials 15.3 (2012), pp. 973–995.
[KS17] G. P. Koudouridis and P. Soldati. “Spectrum and network density
management in 5G ultra-dense networks”. In: IEEE Wireless Commu-
nications 24.5 (2017), pp. 30–37.
[LCC18] L. M. Larsen, A. Checko, and H. L. Christiansen. “A survey of the
functional splits proposed for 5G mobile crosshaul networks”. In: IEEE
Communications Surveys & Tutorials (2018).
[Li11] X. Li. Radio Access Network Dimensioning for 3G UMTS. Springer, 2011.
[Li94] H.-L. Li. “A global approach for general 0–1 fractional programming”.
In: European Journal of Operational Research 73.3 (1994), pp. 590–596.
[Lin+13] L. Lingitz, C. Morawetz, D. T. Gigloo, S. Minner, and W. Sihn. “Mod-
elling of flexibility costs in a decision support system for midterm
capacity planning”. In: Procedia CIRP. 2013.
[Lin16] J. C. Lin. “Human exposure to RF, microwave, and millimeter-wave
electromagnetic radiation [Health Effects]”. In: IEEE Microwave Mag-
azine 17.6 (2016), pp. 32–36.
[Lom54] K. S. Lomax. “Business failures: Another example of the analysis of
failure data”. In: Journal of the American Statistical Association 49.268
(1954), pp. 847–852.

186
[Mac+13] G. R. MacCartney, J. Zhang, S. Nie, and T. S. Rappaport. “Path loss
models for 5G millimeter wave propagation channels in urban micro-
cells”. In: IEEE Global Communications Conference (GLOBECOM). IEEE.
2013, pp. 3948–3953.
[Mae+14] A. Maeder, M. Lalam, A. De Domenico, E. Pateromichelakis, D. Wübben,
J. Bartelt, R. Fritzsche, and P. Rost. “Towards a flexible functional split
for cloud-RAN networks”. In: European Conference on Networks and
Communications (EuCNC). IEEE. 2014, pp. 1–5.
[Mae+16] A. Maeder, A. Ali, A. Bedekar, A. F. Cattoni, D. Chandramouli, S.
Chandrashekar, L. Du, M. Hesse, C. Sartori, and S. Turtinen. “A scal-
able and flexible radio access network architecture for fifth generation
mobile networks”. In: IEEE Communications Magazine 54.11 (2016), pp. 16–
23.
[Mak+17] N. Makris, P. Basaras, T. Korakis, N. Nikaein, and L. Tassiulas. “Exper-
imental evaluation of functional splits for 5G cloud-RANs”. In: 2017
IEEE International Conference on Communications (ICC). IEEE. 2017, pp. 1–
6.
[Mar+18] A. Marotta, D. Cassioli, K. Kondepu, C. Antonelli, and L. Valcarenghi.
“Efficient Management of Flexible Functional Split through Software
Defined 5G Converged Access”. In: IEEE International Conference on
Communications (ICC). 2018, pp. 1–6.
[MMD91] R. Mazumdar, L. G. Mason, and C. Douligeris. “Fairness in network
optimal flow control: Optimality of product forms”. In: IEEE Transac-
tions on Communications 39.5 (1991), pp. 775–782.
[MT90] S. Martello and P. Toth. Knapsack Problems: Algorithms and Computer
Implementations. Wiley Series in Discrete Mathematics and Optimiza-
tion. Wiley, 1990. ISBN: 9780471924203.
[MTK01] K. Man, K. Tang, and S. Kwong. Genetic Algorithms: Concepts and De-
signs. Advanced Textbooks in Control and Signal Processing. Springer
London, 2001. ISBN: 9781852330729.
[Nar+18] G. Nardini, G. Stea, A. Virdis, A. Frangioni, L. Galli, D. Sabella, and
G. M. Dell’Aera. “Practical feasibility, scalability and effectiveness of
coordinated scheduling algorithms in cellular networks towards 5G”.
In: Journal of Network and Computer Applications 106 (2018), pp. 1–16.
[NGM15] NGMN Alliance. “5G white paper”. In: Next generation mobile net-
works, white paper 1 (2015), pp. 1–125.
[Nik+14] N. Nikaein, M. K. Marina, S. Manickam, A. Dawson, R. Knopp, and C.
Bonnet. “OpenAirInterface: A flexible platform for 5G research”. In:
ACM SIGCOMM Computer Communication Review 44.5 (2014), pp. 33–
38.

187
Bibliography

[Nik15] N. Nikaein. “Processing radio access network functions in the cloud:

Critical issues and modeling”. In: Proceedings of the 6th International
Workshop on Mobile Cloud Computing and Services. 2015, pp. 36–43.
[NMH14] G. Nigam, P. Minero, and M. Haenggi. “Coordinated multipoint joint
transmission in heterogeneous networks”. In: IEEE Transactions on
Communications 62.11 (2014), pp. 4134–4146.
[NTT20] NTT DOCOMO. 5G Evolution and 6G. White Paper. Jan. 2020.
[NYB12] T. T. Nguyen, S. Yang, and J. Branke. “Evolutionary dynamic opti-
mization: A survey of the state of the art”. In: Swarm and Evolutionary
Computation 6 (2012), pp. 1–24.
[Ope14] Open Data Center Alliance. Open Data Center Alliance: Software-Defined
Networking Rev. 2.0. Technical Report. 2014.
[Opp+97] A. Oppenheim, A. Willsky, S. Nawab, w. Hamid, and I. Young. Signals
& Systems. Prentice-Hall signal processing series. Prentice Hall, 1997.
ISBN: 9780138147570.
[PL15] R. K. Polaganga and Q. Liang. “Self-similarity and modeling of LTE/LTE-
A data traffic”. In: Measurement 75 (2015), pp. 218–229.
[PM17] H. Paixão Martins. “Analysis of CoMP for the Management of Inter-
ference in LTE”. In: Master Thesis (2017).
[Pow07] W. Powell. Approximate Dynamic Programming: Solving the Curses of
Dimensionality. Wiley Series in Probability and Statistics. Wiley, 2007.
ISBN: 9780470182956.
[Rup+18] A. Ruprecht, D. Jones, D. Shiraev, G. Harmon, M. Spivak, M. Krebs,
M. Baker-Harvey, and T. Sanderson. “Vm live migration at scale”. In:
ACM SIGPLAN Notices 53.3 (2018), pp. 45–56.
[Sab+13] D. Sabella, P. Rost, Y. Sheng, E. Pateromichelakis, U. Salim, P. Guitton-
Ouhamou, M. Di Girolamo, and G. Giuliani. “RAN as a service: Chal-
lenges of designing a flexible RAN architecture in a cloud-based het-
erogeneous mobile network”. In: Future Network & Mobile Summit.
IEEE. 2013, pp. 1–8.
[SCF21] SCF. 5G nFAPI specifications. Document 225.2.0. Small Cell Forum (SCF),
May 2021.
[Sch+03] K. Schuster, F. Pukelsheim, M. Drton, and N. R. Draper. “Seat biases
of apportionment methods for proportional representation”. In: Elec-
toral Studies 22.4 (2003), pp. 651–676.
[Sey+16] T. Seyama, D. Jitsukawa, T. Kobayashi, T. Oyama, T. Dateki, H. Seki,
M. Minowa, T. Okuyama, S. Suyama, and Y. Okumura. “Study of Co-
ordinated Radio Resource Scheduling Algorithm for 5G Ultra High-
Density Distributed Antenna Systems–Performance Evaluation of Joint
Transmission Multi-User MIMO”. In: IEICE Technical Report; IEICE
Tech. Rep. 115.472 (2016), pp. 181–186.

188
[Sha49] C. E. Shannon. “Communication in the presence of noise”. In: Proceed-
ings of the IRE 37.1 (1949), pp. 10–21.
[SL05] G. Song and Y. Li. “Cross-layer optimization for OFDM wireless networks-
part I: theoretical framework”. In: IEEE transactions on wireless commu-
nications 4.2 (2005), pp. 614–624.
[Sma17] Small Cell Forum. FAPI and nFAPI specifications. Document 082.09.05.
Release 9.0. May 2017.
[Sor+17] B. Soret, A. De Domenico, S. Bazzi, N. H. Mahmood, and K. I. Ped-
ersen. “Interference coordination for 5G new radio”. In: IEEE Wireless
Communications 25.3 (2017), pp. 131–137.
[SRF15] V. Suryaprakash, P. Rost, and G. Fettweis. “Are heterogeneous cloud-
based radio access networks cost effective?” In: IEEE Journal on Se-
lected Areas in Communications 33.10 (2015), pp. 2239–2251. arXiv: 1503.
03366.
[SS10] P. K. Sharma and R. Singh. “Comparative analysis of propagation
path loss models with field measured data”. In: International Journal
of Engineering Science and Technology 2.6 (2010), pp. 2008–2013.
[SSG17] A. Sadeghi, F. Sheikholeslami, and G. B. Giannakis. “Optimal and
scalable caching for 5G using reinforcement learning of space-time
popularities”. In: IEEE Journal of Selected Topics in Signal Processing 12.1
(2017), pp. 180–190.
[Sta15] W. Stallings. Foundations of Modern Networking: SDN, NFV, QoE, IoT,
and Cloud. Always Learning. Networking. Pearson, 2015. ISBN: 9780134175393.
[Sut18] R. Sutton. Reinforcement learning: an introduction. Cambridge, Massachusetts
London, England: The MIT Press, 2018. ISBN: 9780262039246.
[SY14] R. Srikant and L. Ying. Communication Networks: An Optimization, Con-
trol and Stochastic Networks Perspective. Cambridge University Press,
2014. ISBN: 9781107036055.
[SYCP18] V. Sciancalepore, F. Z. Yousaf, and X. Costa-Perez. “Z-TORCH: An
Automated NFV Orchestration and Monitoring Solution”. In: IEEE
Transactions on Network and Service Management (2018). arXiv: 1807.
02307.
[Tal+20] T. Taleb, R. L. Aguiar, I. Grida Ben Yahia, B. Chatras, G. Christensen,
U. Chunduri, A. Clemm, X. Costa, L. Dong, J. Elmirghani, et al. White
paper on 6G networking. White paper. 6G Research Visions no. 6, Uni-
versity of Oulu, 2020.
[TAS02] M. Tawarmalani, S. Ahmed, and N. V. Sahinidis. “Global optimization
of 0–1 hyperbolic programs”. In: Journal of Global Optimization 24.4
(2002), pp. 385–416.

189
Bibliography

[TBK15] T. Taleb, M. Bagaa, and A. Ksentini. “User mobility-aware Virtual

Network Function placement for Virtual 5G Network Infrastructure”.
In: IEEE International Conference on Communications. 2015. ISBN: 9781467364324.
[TGD19] H. D. Trinh, L. Giupponi, and P. Dini. “Urban anomaly detection by
processing mobile traffic traces with LSTM neural networks”. In: IEEE
International Conference on Sensing, Communication, and Networking (SECON).
IEEE. 2019, pp. 1–8.
[Tom66] J. Tomlin. “Minimum-cost multicommodity network flows”. In: Op-
erations Research 14.1 (1966), pp. 45–51.
[Tor+21] R. Torre, R.-S. Schmoll, F. Kemser, H. Salah, I. Tsokalo, and F. H. Fitzek.
“Benchmarking Live Migration Performance Under Stressed Condi-
tions”. In: IEEE 18th Annual Consumer Communications & Networking
Conference (CCNC). IEEE. 2021, pp. 1–2.
[Uni21] I. I. T. Union). Measuring Digital Development: ICT Price Trends 2020.
2021.
[Upd21] I. Update. Ericsson Mobility Report. Tech. rep. Ericsson, June 2021.
[Val+16] L. Valcarenghi, K. Kondepu, F. Giannone, and P. Castoldi. “Require-
ments for 5G fronthaul”. In: International Conference on Transparent Op-
tical Networks (ICTON). 2016, pp. 1–5.
[Wax88] B. M. Waxman. “Routing of multipoint connections”. In: IEEE journal
on selected areas in communications 6.9 (1988), pp. 1617–1622.
[Wu97] T.-H. Wu. “A note on a global approach for general 0–1 fractional pro-
gramming”. In: European Journal of Operational Research 101.1 (1997),
pp. 220–223.
[Xio+19] Z. Xiong, Y. Zhang, D. Niyato, R. Deng, P. Wang, and L.-C. Wang.
“Deep reinforcement learning for mobile 5G and beyond: Fundamen-
tals, applications, and challenges”. In: IEEE Vehicular Technology Mag-
azine 14.2 (2019), pp. 44–52.
[Xu+16] F. Xu, Y. Li, H. Wang, P. Zhang, and D. Jin. “Understanding mobile
traffic patterns of large scale cellular towers in urban environment”.
In: IEEE/ACM transactions on networking 25.2 (2016), pp. 1147–1161.
[You+16] F. Z. Yousaf, C. Goncalves, L. Moreira-Matias, and X. C. Perez. “RAVA
- Resource aware VNF agnostic NFV orchestration method for virtu-
alized networks”. In: IEEE International Symposium on Personal, Indoor
and Mobile Radio Communications, PIMRC. 2016. ISBN: 9781509032549.
[Zan97] J. Zander. “On the cost structure of future wideband wireless access”.
In: IEEE 47th Vehicular Technology Conference. Technology in Motion. Vol. 3.
IEEE. 1997, pp. 1773–1776.

190
[Zha+17] Y. Zhang, J. Ding, M.-W. Kwan, J. Ni, E. K. Tsang, Y.-N. R. Li, and
J. Li. “Measurement and Evaluations of Coherent Joint Transmission
for 5G Networks”. In: IEEE 85th Vehicular Technology Conference (VTC
Spring). IEEE. 2017, pp. 1–5.
[Zha+18] P. Zhao, H. Tian, S. Fan, and A. Paulraj. “Information prediction and
dynamic programming-based RAN slicing for mobile edge comput-
ing”. In: IEEE Wireless Communications Letters 7.4 (2018), pp. 614–617.
[ZWW14] L. Zhang, W. Wu, and D. Wang. “Time dependent pricing in wireless
data networks: Flat-rate vs. usage-based schemes”. In: IEEE Confer-
ence on Computer Communications (INFOCOM). IEEE. 2014, pp. 700–
708.

Cited websites
[ATT] ATT North America. Next-Generation IP MPLS Backbone. URL: https:
//[Link]/Common/merger/files/pdf/wired-network/
Domestic\_0C-768\_Network.pdf (visited on 01/14/2021).
[Ben] Ben Wojtowicz. OpenLTE. URL: [Link]
openlte/ (visited on 09/06/2021).
[Goo] Google Developers. Google Protocol Buffers. URL: [Link]
[Link]/protocol-buffers (visited on 09/06/2021).
[Gur] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. URL :
[Link] (visited on 07/20/2021).
[Int] Intel Corporation. Intel Processors for PC, Laptops, Servers, and AI. URL:
https : / / intel . com / content / www / us / en / products /
details/[Link].
[KVM] KVM. Kernel Virtual Machine. URL: https : / / www . linux - kvm .
org/ (visited on 09/06/2021).
[Opea] OpenAirInterface Software Alliance. Open Air Interface RAN Feature
Set. URL: [Link]
blob/master/doc/feAture_sEt.md (visited on 09/06/2021).
[Opeb] OpenAirInterface Software Alliance. OpenAirInterface 5G Radio Access
Network Project. URL: https : / / openairinterface . org / oai -
5g-ran-project/ (visited on 09/06/2021).
[Opec] OpenAirInterface Software Alliance. OpenAirInterface AMF Feature Set.
URL : [Link] cn5g-
amf/-/blob/master/docs/fEature_Set.md (visited on 09/06/2021).
[Oped] OpenAirInterface Software Alliance. OpenAirInterface Main Page. URL:
[Link] (visited on 09/06/2021).

191
Bibliography

[Opee] OpenAirInterface Software Alliance. OpenAirInterface SMF Feature Set.

URL : [Link] cn5g-
smf/-/blob/master/docs/fEature_Set.md (visited on 09/06/2021).
[Opef] OpenAirInterface Software Alliance. OpenAirInterface Software Alliance
Members. URL: [Link]
(visited on 09/06/2021).
[Ran] Range Networks. OpenBTS. URL : [Link] (visited
on 09/06/2021).
[SRSa] SRS. srsRAN - Your own mobile network. URL: [Link]
(visited on 09/06/2021).
[SRSb] SRS. srsRAN 21.04 Documentation. URL : [Link]
com/ (visited on 09/06/2021).
[SRSc] SRS. The srsLTE project is evolving. URL: https : / / www . srslte .
com/srslte-srsran (visited on 09/06/2021).
[Thea] The Internet Topology Zoo. URL: [Link] zoo.
org/ (visited on 09/06/2021).
[Theb] The Linux Foundation Projects. Xen Project. URL: [Link]
org/ (visited on 09/06/2021).
[Thec] The OpenStack Fundation. Open Source Cloud Computing Infrastructure
– OpenStack. URL: https : / / www . openstack . org/ (visited on
09/06/2021).

192
Acronyms and abbreviations

3G Third generation of mobile networks

3GPP Third Generation Partnership Project
4G Fourth generation of mobile networks
5G Fifth generation of mobile networks
AES Advanced Encryption Standard
AMF Access and Mobility Management Function
ARQ Automatic Repeat Request
AUSF Authentication Server Function
BGP Borrero-Gillen-Prokopyev
BBU Baseband unit
BPSK Binary Phase-Shift Keying
CAPEX Capital Expenditure
CDF Cumulative Distribution Function
CMS Cloud Management System
CPU Central Processing Unit
CQI Channel Quality Indicator
C-RAN Cloud-RAN
CU Centralized Unit
DAS Distributed Antenna System
DPI Deep Packet Inspection
DRAN Distributed RAN
DU Distributed Unit
eCPRI Enhanced Common Public Radio Interface
eMBB Enhanced Mobile Broadband
eMBMS Evolved Multimedia Broadcast Multicast Service
eNodeB Evolved NodeB
EPC Evolved Packet Core
FDD Frequency Domain Duplex
FFR Fractional Frequency Reuse
FFT Fast Fourier Transform
FR1 Frequency Range 1
FR2 Frequency Range 2
FSSP Functional Split Selection Problem
gNB Next-generation NodeB
GPRS General Packet Radio Service
GTP GPRS Tunneling Protocol
HARQ Hybrid ARQ

193
Bibliography

HSS Home Subscriber Server

ICIC Inter-cell Interference Coordination
IFFT Inverse Fast Fourier Transform
IoT Internet of Things
IP Internet Propotcol
KVM Kernel-based Virtual Machine
LCID Logical Channel ID
LDPC Low-density Parity-check Code
LP Linear Problem
LTE Long Term Evolution
LWT Li-Wu-Tawarmalani
MAC Medium Access Control
MILP Mixed-Integer Linear Problem
MIMO Multiple Input Multiple Output
MINLP Mixed-Integer Non-Linear Problem
MME Mobility Management Entity
mMTC Massive Machine-Type Communication
NAS Non-Access Stratum
nFAPI Network Functional Application Platform Interface
NFV Network Function Virtualization
NG-RAN Next-Generation RAN
NR New Radio
NRF Network Repository Function
NV Network Virtualization
OPEX Operating Expenses
PC Personal Computer
PDCP Packet Data Convergence Protocol
PDN Public Data Network
PDU Packet Data Unit
P-GW PDN-Gateway
PHY Physical layer
PID Process ID
QAM Quadrature Amplitude Modulation
QIP Quadratic Integer Problem
QoD Quality of Decisions
QoS Quality of Service
RAN Radio Access Network
RDF Readiness Degradation Function
RF Radiofrequency
RLC Radio Link Control
RNTI Radio Network Temporary Identifier
RRC Radio Resource Control
RRH Remote Radio Head
RU Remote Unit
SDAP Service Data Adaptation Protocol

194
SDN Software Defined Networking
SDR Software Defined Radio
SFFR Soft Fractional Frequency Reuse
S-GW Serving Gateway
SINR Signal to Interference and Noise Ratio
SMF Session Management Function
TCO Total Cost of Ownership
TDD Time Domain Duplex
UDM Unified Data Management Function
UDR Unified Data Repository
UE User Equipment
UPF User Plane Function
URLLC Ultra-Reliable Low-Latency Communication
VNF Virtual Network Function
WSL Webster/Sainte-Laguë

195
List of Figures

1.1. Architecture of distributed, fully centralized, partially centralized, and

dynamically centralized radio access networks. . . . . . . . . . . . . . . . 3
(a). Distributed RAN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
(b). Fully centralized RAN. . . . . . . . . . . . . . . . . . . . . . . . . . 3
(c). Partially centralized RAN. . . . . . . . . . . . . . . . . . . . . . . . . 3
(d). Dynamically centralized RAN. . . . . . . . . . . . . . . . . . . . . . 3
1.2. Diagram of the outline of this thesis. . . . . . . . . . . . . . . . . . . . . . 9

2.1. Simple radio access network featuring a distributed architecture. . . . . 17

(a). Distributed RAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
(b). Centralized RAN (C-RAN) . . . . . . . . . . . . . . . . . . . . . . . 17
(c). Partially centralized RAN . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2. Possible functional split options in a 5G network. . . . . . . . . . . . . . 20
2.3. Example network with G = 11 gNBs and eight fronthaul switches. . . . 23
2.4. Considered network, example gNB functions, and scheme functional
splits. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.1. Adaptation phases traversed by a flexible network. . . . . . . . . . . . . 41

3.2. Demands, states, and state delays experienced by an action-interrupting
flexible network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3. Four exemplary readiness degradation functions (RDFs) featuring the
same range but different shapes. . . . . . . . . . . . . . . . . . . . . . . . 51
3.4. Mean readiness cost K(ϕ) against maximum flexibility ϕ. . . . . . . . . . 51
3.5. Topology of the example network implementing random flow requests. 54
3.6. Example of seadiness degradation function. . . . . . . . . . . . . . . . . . 57
3.7. Graphical representation of inequality (22). . . . . . . . . . . . . . . . . . 58
3.8. Reduction factor of the proaction time for different parallelization levels. 59
3.9. Mean proaction cost as a function of the parallelization level. . . . . . . . 59
3.10. Combined mean readiness and proaction costs as a function of the par-
allelization level. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.1. Comparison between function U log(e η ) and its best linear approxima-
tion for U = 3000 UEs and 1 ≤ ηe ≤ 2. . . . . . . . . . . . . . . . . . . . . . 87
4.2. Comparison between nine performance-to-revenue functions K e rev (e
η)
for revenue-operating cost ratios ξ ∈ {0.5, 1, 2} and revenue growth
rates υ ∈ {1.5, 2, 3}. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.3. Visualization of UE concentration index for 2000 UEs. . . . . . . . . . . . 91
(a). θ = 0.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

197
List of Figures

(b). θ = 0.75 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
(c). θ = 0.95 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.4. Visualization of the average fronthaul network degree for the same
gNB distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
(a). ψ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
(b). ψ = 3.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
(c). ψ = 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.5. Convergence time of the MILP reformulations of the performance-max-
imizing FSSP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.6. Convergence time of the quadratic reformulation and Heuristics 1 and
2 for the performance-maximizing FSSP.. . . . . . . . . . . . . . . . . . . 95
4.7. Convergence time of the quadratic approach to the performance-max-
imizing FSSP, the operating-cost-minimizing FSSP, and the quadratic
approach of the readiness-cost-minimizing FSSP. . . . . . . . . . . . . . . 96
4.8. Distributions of the mean spectral efficiency ηe(a) achieved by the per-
formance-maximizing approaches, a fully distributed approach, a static
approach, and a fully centralized approach in four extreme scenarios. . . 97
(a). Sparse fronthaul, uniform population . . . . . . . . . . . . . . . . . 97
(b). Dense fronthaul, uniform population . . . . . . . . . . . . . . . . . 97
(c). Sparse fronthaul, concentrated population . . . . . . . . . . . . . . 97
(d). Dense fronthaul, concentrated population . . . . . . . . . . . . . . 97
4.9. Average spectral efficiency achieved by the quadratic approach, Heuris-
tics 1 and 2, a static and a fully distributed network for G = 300 and a
UE concentration index of 0.75 as the fronthaul average degree varies. . 98
4.10. Average spectral efficiency achieved by the quadratic approach, Heuris-
tics 1 and 2, a static and a fully distributed network for G = 300 and a
fronthaul average degree of 3.5 the fronthaul average degree varies as
the UE concentration index varies. . . . . . . . . . . . . . . . . . . . . . . 99
4.11. Relative operating cost achieved by fully distributed, fully centralized,
quadratic, operating-cost-minimizing, and heuristic approaches when
G = 300, average fronthaul degree ψ = 3.5, and UE concentration index
θ = 0.75. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.12. Average spectral efficiency achieved by fully distributed, fully central-
ized, quadratic, operating-cost-minimizing, and heuristic approaches
when G = 300, average fronthaul degree ψ = 3.5, and UE concentra-
tion index θ = 0.75. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.13. Spectral efficiency and relative operating cost achieved by fully dis-
tributed, quadratic, operating-cost-minimizing and heuristic approaches
when G = 300 and ω = 0 ncu · s/Gb. . . . . . . . . . . . . . . . . . . . . . 102
(a). Spectral efficiency, θ = 0.75 . . . . . . . . . . . . . . . . . . . . . . . 102
(b). Spectral efficiency, ψ = 3.5 . . . . . . . . . . . . . . . . . . . . . . . . 102
(c). Relative operating cost, θ = 0.75 . . . . . . . . . . . . . . . . . . . . 102
(d). Relative operating cost, ψ = 3.5 . . . . . . . . . . . . . . . . . . . . . 102

198
 List of Figures

4.14. Spectral efficiency and relative operating cost achieved by fully dis-
tributed, quadratic, operating-cost-minimizing and heuristic approaches
when G = 300 and ω = 0.5 ncu · s/Gb. . . . . . . . . . . . . . . . . . . . . 103
(a). Spectral efficiency, θ = 0.75 . . . . . . . . . . . . . . . . . . . . . . . 103
(b). Spectral efficiency, ψ = 3.5 . . . . . . . . . . . . . . . . . . . . . . . . 103
(c). Relative operating cost, θ = 0.75 . . . . . . . . . . . . . . . . . . . . 103
(d). Relative operating cost, ψ = 3.5 . . . . . . . . . . . . . . . . . . . . . 103
4.15. Spectral efficiency, operating cost, and readiness cost achieved by per-
formance-maximizing (quadratic), operating-cost-minimizing, and readi-
ness-cost-minimizing approaches when G = 300, ψ = 3.5, and ω =
0 ncu · s/Gb. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(a). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(b). ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(c). ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(d). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(e). ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(f). ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(g). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(h). ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
(i). ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.16. Spectral efficiency, operating cost, and readiness cost achieved by per-
formance-maximizing (quadratic), operating-cost-minimizing, and readi-
ness-cost-minimizing approaches when G = 300, ψ = 3.5, and ω =
0.5 ncu · s/Gb. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(a). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(b). ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(c). ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(d). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(e). ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(f). ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(g). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(h). ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
(i). ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.1. Functional architecture of the two binaries supporting an adaptive func-

tional split. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.2. Evolution of the internal operation of RLC and MAC layers during a
soft migration from the MAC-PHY split to the PDCP-RLC split. . . . . . 122
5.3. Content of the RLC and HARQ buffers during a soft migration. . . . . . 124
5.4. Cumulative distribution functions of the RLC/MAC draining times for
different inter-arrival times ε of incoming downlink PDCP PDUs. . . . . 125
5.5. Additional end-user latency for soft, hard, and nine custom migrations
with allocated times ranging from 1 to 9 ms. . . . . . . . . . . . . . . . . 126
5.6. Probability of experiencing a migration with packet losses for soft, hard,
and nine custom migrations with allocated times ranging from 1 to 9 ms. 126

199
List of Figures

6.1. Example diagram of the sequences of demands and states in the con-
sidered network, along with demand-sampling instants. . . . . . . . . . 133
6.2. Location of the reference action cost frontiers. . . . . . . . . . . . . . . . . 152
6.3. Estimated RDF K∆ (δ) and FT (t) for (ξ, υ) = {(0.5, 1.5), (1, 2), (2, 3)},
and $ ∈ {0, 4, 8} from simulation data. . . . . . . . . . . . . . . . . . . . . 157
(a). RDF for ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . 157
(b). FT (t) for ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . 157
(c). RDF for ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
(d). FT (t) for ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
(e). RDF for ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
(f). FT (t) for ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.4. Evolution of the mean readiness cost achievable by the network as a
function of the monitoring interval Zmon . . . . . . . . . . . . . . . . . . . 159
(a). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
(b). ξ = 0.5, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
(c). ξ = 2, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
(d). ξ = 0.5, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.5. Distribution of solving times of the FSSP as the number of UE clusters
$ varies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.6. Average computational effort of solving the FSSP and its associated
proaction cost. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.7. Comparison of the mean total cost Q of static and dynamic strategies
R∗
as the average action cost C grows for $ ∈ {0, 4, 8} UE clusters and
readiness cost parameters (ξ, υ) ∈ {(1, 1), (3, 3)}. . . . . . . . . . . . . . . 161
(a). $ = 0, ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . 161
(b). $ = 0, ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
(c). $ = 4, ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . 161
(d). $ = 4, ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
(e). $ = 8, ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . 161
(f). $ = 8, ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.8. Comparison of the mean total cost Q of dynamic strategies as the aver-
R∗
age action cost C grows for $ ∈ {0, 4, 8} UE clusters and readiness
cost parameters (ξ, υ) ∈ {(1, 1), (3, 3)}. . . . . . . . . . . . . . . . . . . . . 163
(a). $ = 0, ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . 163
(b). $ = 0, ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
(c). $ = 4, ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . 163
(d). $ = 4, ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
(e). $ = 8, ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . 163
(f). $ = 0, ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.9. Comparison between theoretical and simulated values of the reference
R∗ R∗ R∗
action-cost frontiers C II , C III , and C IV for $ = {0, ..., 8}, ξ = {0.5, 1, 2},
and υ = {1.5, 2, 3}. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
(a). ξ = 0.5, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
(b). ξ = 0.5, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

200
 List of Figures

(c). ξ = 0.5, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

(d). ξ = 1, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
(e). ξ = 1, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
(f). ξ = 1, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
(g). ξ = 2, υ = 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
(h). ξ = 2, υ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
(i). ξ = 2, υ = 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

201
List of Tables

2.1. Technical Concepts and their support of flexibility in networks. . . . . . 29

4.1. System modeling variables. . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4.2. Summary of operating cost parameters. . . . . . . . . . . . . . . . . . . . 93

6.1. Comparison of proposed adaptation strategies. . . . . . . . . . . . . . . . 151

6.2. Summary of cost and simulation parameters. . . . . . . . . . . . . . . . . 160

203