kth.diva-portal.orgkth.diva-portal.org/smash/get/diva2:1291162/fulltext01.pdf · iii abstract the...

Agile, Resilient and Cost-efficient Mobile BackhaulNetworks

Fundamentals of Network Design and Adaptation

FOROUGH YAGHOUBI

PhD Thesis in Information and Communication TechnologySchool of Electrical Engineering and Computer Science

KTH Royal Institute of TechnologyStockholm, Sweden, 2019

TRITA-EECS-AVL-2019:17ISBN 978-91-7873-106-0

KTH School of Electrical Engineering andComputer ScienceSE-164 40 Kista

SWEDEN

Akademisk avhandling som med tillstånd av Kungl Tekniska högskolan framläg-ges till offentlig granskning för avläggande av teknologie doktorsexamen i datalogimåndag den 18 March 2019 klockan 10.00 i Ka-Sal A, Electrum, Kungl Tekniskahögskolan, Kistagången 16, Kista.

© Forough Yaghoubi, March 2019

Tryck: Universitetsservice US AB

iii

Abstract

The exponentially increasing traffic demand for mobile services requiresinnovative solutions in both access and backhaul segments of 5th generation(5G) mobile networks. Whilst substantial research efforts address the accesssegment, the backhaul part has received less attention and still falls short inmeeting the stringent requirements of 5G in terms of capacity and availability.

Ease of deployment and cost efficiency motivate the use of microwavebackhauling that supports fiber-like capacity with millimeter-wave commu-nications. However, these carrier frequencies are subject to weather distur-bances like rain that may substantially degrade the network throughput andavailability performance. To meet the stringent 5G requirements, in this the-sis we develop a complete framework for network design and online adaptationin the presence of weather-based disruptions.

For topology design, we investigate the trade-off between the path diver-sity and link budget to meet the high availability requirements. We proposeseveral efficient algorithms for joint optimization of cost and power to satisfythe availability, differential delay and data rate requirements. The resultsshow that joint optimization of link budget and cost leads to more power-efficient solutions. Moreover, we characterize the correlation among failureevents and incorporate its impact in the topology design problem. Perfor-mance evaluation results verify that considering correlation increases the net-work robustness under weather-based failures.

For network adaption, we develop a fast and accurate rain detection algo-rithm that triggers a network-layer strategy, e.g., rerouting. The rain impactcan be alleviated by regular rerouting using a centralized approach realized bysoftware defined networking (SDN) paradigm. However, careless reconfigura-tion may impose inconsistency due to asynchrony between different switches,which leads to a significant temporary congestion and limits the gain of rerout-ing. To address this, we propose a consistency-aware rerouting frameworkthat considers the cost of reconfiguration. At each time slot, the central-ized controller may either take a rerouting decision to increase the networkthroughput while accepting the switching cost, or choose not to reroute atthe expense of a decreased throughput due to route sub-optimality. We use amodel predictive control algorithm to provide an online sequence of decisionpolicies to minimize the total data loss. Compared to regular rerouting, ourproposed approach reduces the throughput loss and substantially decreasesthe number of reconfigurations.

In the thesis, we also study which backhaul options are the best from atechno-economic perspective. Fiber-based solutions provide high data rateswith robust connectivity under different weather conditions, whereas wirelesssolutions offer high mobility at low installation costs with lower data rate andavailability. We develop a comprehensive framework to calculate the totalcost of ownership of the backhaul segment and analyze the profitability interms of cash flow and net present value. The evaluation results highlight theimportance of selecting proper backhaul solution to increase profitability.

Keywords: 5G, topology design and control, software defined network-ing, rain disturbance, techno-economic framework, network consistency.

v

Sammanfattning

Det exponentiella trafikökningen som efterfrågan på mobila tjänster med-för, kräver innovativa lösningar inom både access- och backhaul-segmenteni 5:te generationens mobila nät (5G). Medan omfattande forskningsinsatseradresserar accessegmentet har backhauldelen fått mindre uppmärksamhet ochklarar fortfarande inte att uppfylla de stränga krav som ställs på 5G-nät närdet gäller genomströmning och tillgänglighet.

Enkel installation och kostnadseffektivitet motiverar användningen avmikrovågsbaserad backhaul som utnyttjar mikrovågsöverföring med en ka-pacitet motsvarande optisk fiber. Dessa bärvågsfrekvenser påverkas dock avväderförhållanden, såsom regn vilket avsevärt kan försämra genomströmning-en och tillgängligheten. För att uppfylla de stränga 5G-kraven utvecklar vi, ettkomplett ramverk för nätverksdesign och anpassning av nätet för att hanteraväderbaserade störningar. För design av topologi undersöker vi avvägningenmellan väg-diversitet och länkbudget för att uppfylla de höga tillgänglighets-kraven. Vi föreslår flera effektiva algoritmer för samtidig optimering av kost-nad och energiförbrukning för att tillfredsställa tillgängligheten, differentiellfördröjning och datahastighetskraven. Resultaten visar att samtidig optime-ring av länkbudget och kostnad leder till mer energieffektiva lösningar. Vida-re beskriver vi sambandet mellan felhändelser och införlivar deras inverkan itopologidesignproblemet. Utvärdering av prestanda visar att användning avdenna korrelation ökar nätverkets robusthet vid väderbaserade störningar.

För omkonfigurering av nätverket utvecklar vi en snabb och noggrann al-goritm för regndetektering som triggar en anpassning av nätverkslagret, t exomdirigering. Regninverkan kan lindras genom regelbunden omdirigering medhjälp av en centraliserad strategi baserad på SDN (software defined network).Då uppdateringsprocessen mellan olika växlar inte alltid är synkroniserad,kan vårdslös omkonfigurering medföra inkonsistenser i nätet vilket kan ledatill en betydande tillfällig överbelastning och begränsa fördelarna med omdiri-geringen. Vi föreslår en konsekvens-medveten omkonfigureringsmetod som tarhänsyn till omkonfigureringens kostnad. Vid varje tidpunkt kan den centralastyrenheten antingen ta ett omdirigeringsbeslut för att öka genomströmning-en och samtidigt acceptera omkonfigureringskostnaden, eller välja att avståfrån omkonfigurering på bekostnad av en minskad genomströmning på grundav det suboptimala vägvalet. Jämfört med konventionell omdirigering, redu-cerar vår föreslagna metod genomströmnings-förlusten och minskar väsentligtantalet omkonfigureringar genom ett lämpligt tidval för omdirigeringar.

I avhandlingen studerar vi också vilka backhaul-alternativ som är bästfrån en tekno-ekonomiskt synvinkel. Fiberlösningar ger höga datahastighetermed robust anslutning under olika väderförhållanden medan trådlösa lösning-ar erbjuder hög mobilitet och låga installationskostnader men med lägre da-tahastighet och tillgänglighet. Vi utvecklar en modell för att beräkna totalaägandekostnaden för backhaul-segmentet och analyserar lönsamheten när detgäller kassaflöde och nuvärde. Våra resultat markerar vikten av att välja rättbackhaul-lösning för att öka lönsamheten.

Nyckelord: 5G, topologidesign och styrning, programvarustyrt nätverk,regninverkan, tekno-ekonomisk modell, nätverkskonsistens.

To my late mother Forouzan: because I owe it all to you.

vii

Acknowledgements

Firstly, I would like to express my sincere gratitude to my advisors, Professor LenaWosinska, and Dr. Marija Furdek for accepting me as their PhD student, alwaysbelieving in me, and for all their guidance, patience, motivation, immense knowledgeand support during these years. I also want to offer my special thanks to mytalented co-supervisor Associate Professor Jiajia Chen for the continuous supportof my Ph.D study and related research. Their guidance helped me throughout myresearch and writing of this thesis. I could not imagine having better advisors andco-advisor for my Ph.D study.

I would like to thank Professor Carlo Fischione for the advance review of my PhDthesis and Professor Danijela Cabric for reviewing my PhD proposal with insightfulcomments and encouragement, but also for the hard question which incented meto widen my research from various perspectives. I am also grateful of ProfessorDipak Ghosal for accepting the role of opponent for my thesis and to membersof my grading committee, Professor Marina Petrova, Professor Michal Pioro andProfessor Tommy Svensson.

I would also like to express my appreciation to everyone working in the Kista5G Transport Lab (K5) project for their support and sharing their knowledge. Aspecial thanks should be given to Dr. Ahmad Rostami and Dr. Peter Öhlén forthe fruitful discussions and their guidance that helped me to publish several papersand a patent.

I also thank all my friends in the Optical Network Lab (ONLab) for creatingsuch friendly environment and bringing joy in my daily work.

I am also grateful to my friends for their endless kindness and support, whichmakes my life more lovely even far from home, Meysam, Sara, Milad, Maryam,Farshin, Amir, Srwa, Farshad, Yuhanna, Elahe, Sajed, Zahra, Ehsan, Hamid,Alireza, Forough, Maryam, Hamed, Behdad, Kaveh, Mozghan, Matteo, Sibel andCarlos.

Last but not least, I would like to thank my family for all their love and en-couragement. To my parents Reza and Forouzan who raised me with a love ofscience and supported me in all my pursuits. To my sister Sahar, who supportedme spiritually throughout my entire life. And most of all to my loving, supportive,encouraging, and patient husband Hossein, whose faithful support during all stagesof my life is so appreciated. Thank you.

Forough Yaghoubi,Stockholm, March 2019.

Contents

Contents ix

List of Figures xi

List of Tables xiii

List of Acronyms xv

List of Papers xvii

1 Introduction 11.1 Contributions of the Thesis . . . . . . . . . . . . . . . . . . . . . . . 6

1.1.1 Offline Topology Design . . . . . . . . . . . . . . . . . . . . . 71.1.2 Online Network Adaptation . . . . . . . . . . . . . . . . . . . 91.1.3 A Techno-economic Framework for 5G Backhaul Networks . 10

1.2 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Topology Design for Reliable Networks under Correlated Failures 132.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . 142.2.2 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . 152.2.3 Link Failure Probability Model . . . . . . . . . . . . . . . . . 17

2.3 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.3.1 Topology Design under Independent Failures . . . . . . . . . 182.3.2 Topology Design under Correlated Failures . . . . . . . . . . 19

2.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 212.4.1 Synthetic Network . . . . . . . . . . . . . . . . . . . . . . . . 222.4.2 Realistic Deployed Network . . . . . . . . . . . . . . . . . . . 23

3 Joint Topology Design and Control for Reliable Networks 273.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.2 Problem Description and Formulation . . . . . . . . . . . . . . . . . 28

ix

x CONTENTS

3.3 Topology Design with Joint Power and Data Rate Optimization . . . 303.3.1 Subproblems . . . . . . . . . . . . . . . . . . . . . . . . . . . 313.3.2 Master Problem . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.4 Topology Design with Delay Consideration . . . . . . . . . . . . . . 343.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 35

4 Accurate Rain Detection for Improved Network Performance 394.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.2.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . 424.2.2 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.3 Rain Detection Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 444.4 Impact of Detection Error on Network Performance . . . . . . . . . . 464.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 46

5 Consistency-aware Rerouting Framework 515.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.2.1 Switching Cost Minimization . . . . . . . . . . . . . . . . . . 535.2.2 Total Data Loss Minimization . . . . . . . . . . . . . . . . . . 56

5.3 Online Algorithm for Efficient Rerouting . . . . . . . . . . . . . . . . 575.3.1 Rain Attenuation Prediction . . . . . . . . . . . . . . . . . . 585.3.2 Finite Time Horizon Definition . . . . . . . . . . . . . . . . . 585.3.3 The Optimal Control Action with Prediction (OCAP) Algo-

rithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.4.1 Synthetic Network . . . . . . . . . . . . . . . . . . . . . . . . 615.4.2 Real Network . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

6 A Techno-economic Framework for 5G Backhaul Networks 696.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 706.2 Techno-economic Framework . . . . . . . . . . . . . . . . . . . . . . 716.3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 74

7 Conclusions and Future Work 797.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 797.2 Open Challenges and Future Work . . . . . . . . . . . . . . . . . . . 81

Bibliography 83

List of Figures

1.1 Percentages of different technology providing backhaul connectivity . . . 2

2.1 The network topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.2 Average availability and total deployment cost for different values of the

weight factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3 Realistic wide-area network topology . . . . . . . . . . . . . . . . . . . . 232.4 Average availability and total deployment cost for a different number of

gateways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.5 Run time for a different number of gateways . . . . . . . . . . . . . . . . 24

3.1 Realistic network topology . . . . . . . . . . . . . . . . . . . . . . . . . . 363.2 Total link budget for different value of the maximum link budget . . . . 373.3 Total link budget for different values of the data rate threshold . . . . . 373.4 Total link budget for different values of the maximum number of added

links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.5 Total link budget for different values of the maximum differential delay 38

4.1 The overall assumed network architecture . . . . . . . . . . . . . . . . . 414.2 Network topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424.3 Flowchart for the proposed rain detection algorithm. . . . . . . . . . . . 454.4 The impact of different design parameters on rain detection algorithm . 474.5 The impact of detection error on network performance . . . . . . . . . . 484.6 Performance of our rain detection algorithm using real data and topology 49

5.1 Illustrative example for congestion imposed during network transtion . . 545.2 Total data loss for synthetic network . . . . . . . . . . . . . . . . . . . . 625.3 Sensitivity analysis considering different flow update time . . . . . . . . 625.4 Average data loss for synthetic network . . . . . . . . . . . . . . . . . . 635.5 Sensitivity analysis of average data loss considering different rain intensity 645.6 A part of a microwave backhaul topology deployed in Sweden . . . . . . 645.7 Total data loss for real network . . . . . . . . . . . . . . . . . . . . . . . 655.8 Total aggregated throughput and number of reconfigurations for differ-

ent rerouting policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

xi

xii List of Figures

5.9 Sensitivity analysis of data loss considering different time horizons forreal network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

6.1 Techno-economic framework . . . . . . . . . . . . . . . . . . . . . . . . 716.2 Total cost of ownership classification . . . . . . . . . . . . . . . . . . . . 736.3 Network topology of case study . . . . . . . . . . . . . . . . . . . . . . . 746.4 Cost evalution for different scenarios . . . . . . . . . . . . . . . . . . . . 766.5 NPV evalution for different scenarios . . . . . . . . . . . . . . . . . . . . 76

List of Tables

2.1 Simulation parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.1 Simulation parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.1 Simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

6.1 Scenarios used in the case study. . . . . . . . . . . . . . . . . . . . . . . 746.2 Input values used for TCO calculation. . . . . . . . . . . . . . . . . . . . 75

xiii

List of Acronyms

5G 5th generation of networkAMC Adaptive modulation and codingAWG Arrayed waveguide gratingsBER Bit error rateBS Base stationCA Correlated areaCAPEX Capital expendituresCC Centralized controllerCF Cash flowCSG Close subscriber groupDC Data centerDP Dynamic programmingDCA Difference of convex approachEPC Evolved packet coreHetNet Heterogeneous networkILP Integer linear programmingLTE Long term evolutionMBS Mobile base stationMILP Mixed integer linear programmingMNO Mobile network providerMPC Model predictive controlNPV Net presented valueOLT Optical line terminalONU Optical network terminalOPEX Operational expendituresOSG Open subscriber groupOSPF Open shortest path firstPON Passive optical networkP-WARP Predictive weather assisted routing protocolQoS Quality of serviceRAN Radio access pointRHS Receding horizon control

xv

xvi LIST OF ACRONYMS

SC Switching costSDN Software defined networkingTCO Total cost of ownershipTG Throughput gapTWDM Time division multiplexingWAN Wide area networkWMN Wireless mesh networkXl-OSPF Cross-layered open shortest path first

List of Papers

Papers Included in the Thesis

Paper I. F. Yaghoubi, M. Furdek, A. Rostami, P. Öhlén and L. Wosin-ska,“Reliable Topology Design of Wireless Networks under Corre-lated Failures,” Proc. IEEE International Conference on Commu-nications (ICC), pp.1-6, May 2018.

Paper II. F. Yaghoubi, M. Furdek, A. Rostami, P. Öhlén and L. Wosin-ska,“Impact of Correlated Failures on Design and Reliability Perfor-mance of Wireless Networks,” submitted to IEEE Transactions onCommunication, 2018.

Paper III. F. Yaghoubi, M. Furdek, P. Öhlén and L. Wosinska,“Joint Topol-ogy Design and Control for Reliable Wireless Networks,” submittedto IEEE Transactions on Communication, 2019.

Paper IV. F. Yaghoubi, J. Chen, A. Rostami, and L. Wosinska, “Mitigation ofRain Impact on Microwave Backhaul Networks,”Proc. IEEE Inter-national Conference on Communications (ICC), pp. 134-139, May2016.

Paper V. F. Yaghoubi, M. Furdek, A. Rostami, P. Öhlén and L. Wosinska,“Consistency-aware Weather Disruption-tolerant Routing in SDN-based Wireless Mesh Networks,” IEEE Transactions on Network andService Management, vol. 15, no. 2, pp.582-595, Jan. 2018.

Paper VI. F. Yaghoubi, M. Mahloo, L. Wosinska, P. Monti, F.S. Farias,J.C.W.A. Costa, and J. Chen, “A Techno-Economic Framework for5G Transport Networks,” IEEE Wireless Communications, vol.25,no.5, pp.56-63, Oct. 2018.

Patents and Papers not Included in the Thesis

Patent I. F. Yaghoubi, A. Rostami, J. Chen, L. Wosinska, and P. Öhlén,“Routing control in a communication network,” U.S Patent applica-

xvii

xviii LIST OF PAPERS

tion, WO2017101984 A1, 2017.

Paper VII. F. Yaghoubi, A. Abbasfar, B. Maham, “Energy-efficient RSSI-based localization for wireless sensor networks,” IEEE Communi-cations Letters, Vol. 18, no. 6, pp.,973-976. Jun, 2014.

Chapter 1

Introduction

Fifth generation mobile networks, commonly referred to as 5G, aim at supportinga large number of mobile users per unit area and providing ultra-high capacity,low latency, and ultra-high reliability. The exponentially increasing bandwidthdemand for mobile data services requires a massive network densification whichrelies on the idea of introducing low-cost and low-power small cells to upgrade thearea capacity while keeping the macro cells to ensure coverage [1]. This evolutiontowards densified cell deployment in access networks puts enormous strain on thebackhaul segment, i.e., the part that connects the users to the network core [2].With ultra-demanding 5G access, backhaul network should evolve further to be ableto guarantee the expected performance [2]. The challenges posed to the backhaulsegment in 5G deployment include stringent requirements in terms of capacity tosupport such huge amount of traffic, end-to-end delay, availability, energy and costefficiency [3].

Due to the benefits of microwave technology such as ease of deployment and lowcost, there is a growing consensus in both academia and industry that microwavetechnology will play an important role in the backhauling of next generation radioaccess networks (RAN). Fig. 1.1 shows the portions of installed backhaul technolo-gies providing backhaul connectivity worldwide (excluding China, Japan, Korea,and Taiwan). It is forecasted that fiber and microwave technologies will becomethe two dominant mobile backhaul solutions. Specifically, by 2020, it is foreseenthat more than 65% of all sites will be connected with microwave links using dif-ferent frequency bands [4], expected to achieve fiber-like capacity. For instance,Ericsson envisions a need for 10 Gbps data rate backhaul connections for extremecapacity radio sites [5]. Moreover, NGMN report claims that 5G should enable99.999% network availability, including robustness against climatic events andguaranteed services at low energy consumption [6]. On the one hand, achievingfiber-like capacity requires shifting to higher carrier frequency such as millimeterwave communication. On the other hand, higher carrier frequency solutions are sus-ceptible to weather disturbances that may substantially degrade network through-

1

2 CHAPTER 1. INTRODUCTION

Microwave Fiber Copper

2008 2010 2015 2020

100%

80%

60%

40%

20%

0%

Inst

alle

d co

nnec

tions

Figure 1.1: Percentages of different technologies providing backhaul connectivitythroughout time [4].

put and availability [7], hindering the aforementioned requirements. Therefore, inorder to meet the availability, capacity and delay requirements, the evaluation andimprovement of network design and adaptation in the presence of failures, especiallyweather-based disruptions, becomes essential.

Network topology design and control are critical in satisfying the objectivesrelated to the reliability, throughput, energy and cost efficiency. Traditionally,existing backhaul topology takes form of a tree, which ensures connectivity in acost-efficient manner. However, due to inherent vulnerabilities of the wireless envi-ronment, these networks may fail to support the stringent requirements of prolif-erating services. Indeed, a well-planned and optimized network topology combinedwith efficient survivable routing algorithms can significantly improve the networkperformance in terms of throughput and availability at a slightly higher cost. Thiscreates the need for a complete study that can (i) address the problem of design-ing the network topology to push the network to its maximum performance witha minimum power and cost budget, and (ii) study the behavior of the networkunder varying conditions and provide adaptive and efficient solutions to meet theperformance requirements in real time.

Topology design determines the subgraph of the network and the underlyingphysical layer topology that satisfies the predefined criteria. In wireless networks,the most noticeable source of performance degradations in terms of throughput andconnectivity is poor network design [8]. Therefore, it is vital to focus on this prob-lem since a well-planned topology can often provide better performance at the sameinfrastructure cost. Topology control focuses on optimizing the transmit power, an-tenna gain and direction to achieve the considered performance [9]. Note, however,that the terms topology design and topology control are used interchangeably in

3

the majority of literature. Cost and power budget are the two fundamental factorsin deployment and management of future networks that should be considered dur-ing network planning. A good network design solution essentially involves a carefulchoice of link locations while guaranteeing performance metrics such as reliabilityand throughput at a minimum cost and power consumption. Different performanceparameters can drive the topology design. In this thesis, we consider the networkavailability and throughput which are strongly affected by weather disruptions.Many studies focus on the cost-efficient topology design to meet the reliability per-formance requirements [9–13]. These works argue that adding redundant networkinfrastructure such as links or nodes may increase the reliability, which also sub-stantially increases the deployment cost. However, by ignoring the possibility offailure correlation, which is a dominant characteristic of weather-based disruptions,the algorithms proposed therein fall inefficient due to their modeling inaccuracies.In this thesis, we show that the topology design with the assumption of indepen-dent failures fails to satisfy the reliability threshold under correlated failures suchas rain.

To satisfy the network reliability performance requirements by careful networkdesign, connection availability can be enhanced by (i) establishing more links toensure multipath connectivity (diversity), or (ii) increasing the availability of theexisting links by allocating a higher link budget (transmit power + antenna gain)to compensate for the attenuation caused by rain. Diversity (multipath routing)as a standalone solution for meeting availability requirements may lead to an ex-pensive topology since it ignores the possibility of deploying highly reliable links byadjusting the link budget. Many studies focus on optimal power allocation alone forachieving certain network connectivity [14–16]. Such approaches yield high powerconsumption for achieving the required reliability due to single path connectivity.To cope with these problems, the two distinct approaches were combined in [17,18]by allocating link budget so that k disjoint paths and the path diversity requiredfor multipath routing are guaranteed. However, existing approaches for joint linkbudget allocation and topology design from the literature cannot be applied to theproblems considered in this thesis due to following reasons. Firstly, the term con-nectivity considered therein does not quantify end-to-end availability, while the QoSrequirements in 5G specifications typically refer to 5- or 6-nine end-to-end avail-ability performance [6]. Secondly, the existing fault tolerant approaches provision afixed, predetermined number of disjoint paths, k, for all sources, which may lead toinefficient solutions in cases when the required availability for some sources can beobtained by using a lower number of paths with a lower link failure probability. Tothe best of our knowledge, the problem of jointly considering path diversity and linkbudget adjustment for achieving end-to-end availability has not been investigatedbefore. Hence, we develop an approach that finds an efficient topology and linkbudget allocation for each established link in order to meet the availability and thedata rate requirements with minimum cost and power.

Topology design approaches typically optimize the network based on averagestatistical properties of the environment. However, due to external dependencies of


wireless medium degradations, there is no guarantee to satisfy the requirements inreal time. Therefore, online adaptation algorithm running on top of a well-plannedtopology is needed. Link layer adaptation such as adaptive modulation and cod-ing (AMC) has already been employed in commercial systems to overcome thestochastic behavior of channel attenuation [2, 19]. Main events that may requirelink adaptation include rain, fog, snow, multipath fading, path obstruction andinterference. These events can affect the network for shorter or longer times, andefficient mitigation of their negative effects requires actions at different levels. Theformer can typically be addressed locally, on a per link basis, while the latter canpersist for a relatively long period (e.g., half an hour) and may require network-wideadaptation. Moreover, the conditions and the extent of impact may vary over thisperiod, requiring more frequent re-adaptation. To be able to react most efficientlyto each type of events and mitigate their impact according to their specifics, weshould first be able to distinguish between long-term and short-term fading. In thisthesis, we consider rain as one of the long-term events that has a severe impacton links with higher carrier frequency and may affect several links simultaneously.Detecting the presence of rain using microwave technology and rainfall attenua-tion on each link has been studied for many years [20, 21]. The primary objectiveof existing approaches is to increase the efficiency of rainfall mapping for hydro-meteorological purposes. However, the long detection time, the need to use extralinks with different frequency bands or wide angular separation, as well as highcomplexity make them inapplicable for our approach. The need for quick restora-tion from rain-related failures (e.g., 50 ms [22]) and low complexity, i.e., simple andfast rain detection with high accuracy, were guiding the design of our rain detectionalgorithm. The algorithm takes advantage of both spatial and temporal correlationof the rain attenuation to be able to quickly and accurately detect the presence ofrain and prompt adaption mechanisms.

Once our proposed rain detector estimates the presence of rain, network-layeraction such as routing is triggered. Rerouting is known to be an effective approachfor mitigating the rain impact. The existing approaches to compute and apply newroutes for traffic flows based on the changing network conditions can be catego-rized into distributed and centralized solutions. Distributed algorithms allow foreach node to make rerouting decisions locally, forgoing the need to communicatewith a central entity, which increases the scalability and fault tolerance. In cen-tralized routing algorithms, the global information such as topology and link stateinformation are gathered from each network element and maintained at a central-ized unit responsible for computing the routing policy, while network elements carryout simplified execution tasks. The distributed solutions may lead to underutiliza-tion of network resources due to only local solution optimality [23], or may beinapplicable for services with strict delay requirements due to the delay caused bytheir slow convergence. Existing centralized approaches cope with these challengesby preventing oscillation between different routes [24], ensuring globally optimalsolutions [25], and increasing the network throughput [26,27]. The rapidly expand-ing paradigm of software-defined networking (SDN) facilitates the development of

5

centralized approaches by shifting the forwarding intelligence and management toa centralized controller and keeping the network elements as simple as possible.SDN as a centralized control platform is a promising solution for providing networkprogrammability and for facilitating dynamic quality-of-service provisioning [28].Although the SDN concept was primarily proposed for wired networks, integrationof SDN and wireless mesh network (WMN) backhaul solutions has been the subjectof several studies [29–31]. The SDN-enabled global view eases reconfiguration andmanagement of the whole backhaul network, which is particularly important in caseof large-scale disturbances of varying intensity and coverage due to weather.

Upon the detection of rain, routing should be done frequently (e.g., each 5min) due to fluctuations of link attenuation throughout the duration of rain totightly fit the traffic to the changing network states. In SDN architecture, thecentralized controller carries out the reconfigurations of the data plane by updatingthe network elements’ flow tables to reflect the newly computed routes. Due toa delay in receiving the update messages from the controller, the update processacross different switches is not synchronized. Therefore, it can happen that packetsin some switches are forwarded based on the new rules while other switches stillprocess the traffic using the old rules. This causes inconsistency which leads thenetwork to unwanted transition states. Properties that hold in the old and in thenew network state might not hold in a transition state, possibly violating networkinvariants such as link capacity. The violation of invariants deteriorates the networkperformance including throughput and delay. Therefore, successful use of SDN inimplementing centralized routing requires not only approaches to compute feasibleroutes in the target state but also methods to facilitate transition to that state ina way that maintains the desired consistency properties.

Various types of consistencies were introduced in the literature, imposing dif-ferent rules that should be held during transition [32]. For instance, loop freedomconsistency ensures that no loops are generated during state transition. Congestion-free updates ensure that the summation of flows traversing a link does not exceedthe link capacity during transition. A possible approach to avoid congestion duringtransition is to sequentially update flows, where the flows are migrated accordingto a precomputed ordering. For congestion-free consistency, a rule update thatbrings a new flow to a link must occur after an update that removes an existingflow if the link cannot support both flows simultaneously. However, finding thesequence of updates that eliminates transient congestion may not always be possi-ble. Congestion-free updates are only achievable if the network supports splittingof traffic flows [33], or if the network is under-utilized and there is sufficient ex-tra capacity [26]. Otherwise, even judicious sequential reconfiguration can leadto congestion. In this case, a possible approach is to accept the congestion whileorganizing the updates so as to minimize the resulting congestion. The conges-tion imposed during reconfiguration may limit the effectiveness of frequent networkupdates and should be considered in the rerouting process during rain. As this con-gestion depends on the initial and the final computed routes, the time for applyingnew routes should be chosen wisely such that the gain of rerouting outweights the


transient-imposed congestion. Existing routing approaches for mitigating rain ef-fects in wireless mesh networks, such as [22] and [34], ignore the cost of adaptationand its deleterious effect on network performance. To address this issue, we firstmodel the switching cost as the minimum imposed data loss due to flow reconfigura-tion, and then, knowing the cost and the gain of rerouting, we formulate a problemto select the best time for applying the rerouting action in order to minimize thetotal data loss during rain.

Aside from alleviating the impact of rain on the wireless backhaul networks, wealso investigate the problem of selecting the most appropriate technology for thebackhaul according to different criteria. A network operator has many options indesigning a proper backhaul segment. One option are fiber-based solutions whichprovide relatively large bandwidth and high reliability. The other option are wire-less solutions with fast deployment and low installation cost, which are especiallyappealing for backhaul in areas where fiber deployment is not possible. The cost andeconomic viability of each option are among the most important parameters consid-ered by a mobile network operator (MNO) when designing the backhaul segment.Some existing works (e.g., [35]) assess the revenue and cost of different heteroge-neous networks (HetNet) deployment and management strategies. However, suchworks only focus on the RAN segment and do not take into account the backhaulnetwork infrastructure which aggregates the traffic from each cell to the evolvedpacket core (EPC) of the operator. It should be noted that the introduction ofsmall cells in RANs can significantly affect the design of the backhaul segment [36].Therefore, this thesis includes a techno-economic analysis of the overall mobile net-work deployment (considering both the backhaul and the RAN), which is crucialin finding the most economically viable solution from the MNO’s point of view.

1.1 Contributions of the Thesis

The contributions of this thesis are categorized into three main parts. In the firstpart, we focus on the design and control of a network topology that can satisfy thestringent requirements on future backhaul networks with a minimum cost and powerbudget. We consider network reliability performance as the guiding design criterionand develop approaches for computing cost- and power-efficient deployments. Ourproposed algorithms can be distinguished from the existing works in two mainaspects. Firstly, the algorithms take into account the correlation among failures,which is shown to have a severe impact on network reliability. Secondly, our designalgorithms consider trade-off between the path diversity and the link budget, andjointly optimize them, thus meeting the reliability threshold at a drastic reductionof the power and cost budget. In the second part of the thesis, we study ouroptimal wireless backhaul network in operation. We investigate the environmentaleffects such as rain that deteriorate the network performance and try to mitigatetheir impact. Our study shows that rerouting at carefully selected times underthe control of one central unit during rain improves the network performance in

1.1. CONTRIBUTIONS OF THE THESIS 7

terms of throughput, compared to non-scheduled rerouting. As the adaptationof the network starts with a rain event, we develop a rain detection algorithmthat accurately determines the presence of rain by observing and correlating thereceived power for each link. Although frequent reconfiguration improves networkperformance in terms of, e.g., throughput, this gain does not come for free. Eachreconfiguration imposes congestion to the network defined as switching cost, whichlimits the efficacy of reconfiguration. Therefore, we propose a framework thatconsiders the cost of reconfiguration during network adaptation and selects thebest policy for applying flow routing updates to minimize the total data loss duringrain.

Apart from quantifying the benefits of deploying a huge number of small cells interms of spectrum and energy efficiency, the assessment of the economic viability isan important factor for a network operator when deciding among different backhaulsolutions. Although the HetNet paradigm increases the spectral efficiency in thewireless access network, i.e., the segment between mobile users and access points,it poses extra costs on the backhaul network due to the requirement of providingtransport for a large number of small cells. Costly backhaul solutions that yieldlower income will limit the efficiency of deploying a huge number of small cells. Inthe third part of the thesis, we evaluate the cost and profitability of different back-haul solutions for 5G to aid the operator’s decision on: (i) what type of transporttechnology (e.g., fiber or microwave) is the most cost-efficient for a specific RANdeployment (e.g., homogeneous or heterogeneous), and (ii) what is the best time toinvest in a new mobile network deployment in order to provide sufficient capacitywhile maximizing the profit in the long run. The main contributions of this thesisare summarized in the following.

1.1.1 Offline Topology DesignIn this part, we explain two contributions related to offline network design solutionsfor improving the performance of wireless backhaul networks.

1.1.1.1 Topology Design for Reliable Networks under CorrelatedFailures

Existing approaches for topology design aimed at satisfying predefined criteria suchas reliability neglect the impact of correlated failures on topology design, whichmakes these algorithms inefficient under correlated scenarios such as disturbancescaused by rain. In this thesis, we address this issue by formulating a topology de-sign problem capable of considering the failure correlation. We propose a methodto formulate the spatial correlation in two parts, 1) intra-path, i.e., the correla-tion between links along a path, captured by modifying the failure probabilityof each path using the joint probability distribution of rain attenuation betweentwo consecutive links, and 2) inter-path, i.e., the correlation among different pathsbetween a pair of nodes, captured in a newly defined penalty cost. Considering


the correlation, we formulate the topology design problem as a quadratic integerprogramming which is solved to optimality using CPLEX for a smaller probleminstance, and develop two efficient heuristic algorithms for a larger, realistic net-work topology. The heuristics are based on continuous and Lagrangian relaxationthat provide near-optimal solutions in polynomial time. The results presented inPaper I and Paper II capture the trade-offs between the cost of deployment andachievable path availability under correlated failures and show that our proposedtopology design approach meets the reliability requirement even under correlatedfailures at a slightly higher cost that the correlation-agnostic variant due to selec-tion of uncorrelated paths. We evaluate the performance of our proposed heuristicalgorithms and show that Lagrangian relaxation provides closer approximation tothe optimal solution compared to continuous relaxation, at the expense of highercomplexity.

1.1.1.2 Joint Topology Design and Topology Control for ReliableNetworks

Increasing the number of links to provide diversity has been shown to be an effectivesolution to satisfy the availability requirements. However, if links are not reliabledue to a low link budget, the number of links needed to meet the availability re-quirement significantly increases imposing a very costly topology design solution.Conversely, if only link budget is optimized to meet the availability requirement, itmay impose a high total link budget due to a lack of path diversity. Considerablenumber of works optimize either the cost of added links or the link budget to satisfythe availability requirements, which leads to inefficient solutions, while the ones thatsolve the two problems jointly do not capture the end-to-end availability requiredto support a targeted quality of service. We close this gap by formulating a jointoptimization problem for cost and link budget (transmit power + antenna gain) al-location, that considers the trade-offs between these two parameters to satisfy thereliability performance requirements for end-to-end connection. The objective is tominimize the total link budget with a constraint on the availability, the numberof added links, and the requested data rate. As rain is one of the main causes offailure in wireless backhaul networks, we consider the applicability of the proposedapproach to a network affected by rain where each link can fail with certain failureprobability that depends on the link budget. However, it should be noted that ouroptimization framework is quite general and can be used for other causes of linkfailures as well. We model the joint problem as a mixed integer nonlinear programwhich requires a logical central controller that takes as input the location of nodes,their destinations and rain attenuation characteristics. As output, it finds the linkbudget allocation for each node, as well as the location and the number of linksthat should be established in the network. We show that the formulated problemis NP-complete, hence, we propose a light-weight heuristic algorithm to provide asub-optimal solution with lower complexity. We characterize the differential de-lay imposed by our optimal solution and modify our topology design algorithm to

1.1. CONTRIBUTIONS OF THE THESIS 9

consider the delay constraint. Through simulations, we asses the impact of maindesign parameters, including the link budget limitations, the target reliability per-formance level, the number of added links, the maximum differential delay, and theminimum requested data rate on the performance of our proposed algorithm. Theresults of this study, presented in Paper III, show that jointly optimizing topologyand link budget for a targeted reliability substantially reduces the deployment costand improves energy efficiency of the network. Moreover, the results indicate thatthe power efficient solutions impose high differential delay due to multipath routingover paths with different lengths, while limiting the delay increases the total linkbudget.

1.1.2 Online Network Adaptation

In this section, we describe our two contributions for adapting the network onlineto reduce the effect of weather disturbances.

1.1.2.1 Accurate Rain Detection for Improved Network Performance

The centralized routing solution in SDN architecture for mitigating the impact ofrain is triggered by rain presence detection. Upon rain detection, one can use anestablished toolkit of multi-commodity flow theory to calculate a new network con-figuration optimized for resource usage efficiency under new channel conditions.Updating the network based on newly computed routes results in throughput gainwhen the disturbing event lasts long enough. Otherwise, triggering rerouting forshort-term events such as multipath fading can in fact increase the disturbance andcause oscillation between optimal solutions. Therefore, it is necessary to distin-guish the short-term events such as multipath fading from long-lasting fading suchas rain. The existing rain detection algorithms either have long detection time orimpose extra cost to be able to detect the rain, which makes them inapplicable forour purpose. We address this problem by introducing a new algorithm to distin-guish long-lasting fading events (such as rain fading) from short-term ones (suchas multipath fading) by relying on the temporal and spatial correlation betweenreceived signal samples. We apply our generalized detector to rain fading and showthat a very low detection error can be achieved by analyzing only a few samplesof the received signals. We also analyze the impact of the error probability of therain detection algorithm on network performance. False detection of a long-termevent may decrease network throughput or pose extra overhead on the networkby triggering unnecessary rerouting. Using our accurate rain detection algorithmreduces the loss of network throughput and the rerouting overhead. The proposedalgorithm is simple and can be readily implemented on top of existing protocols forwireless backhaul networks. Our work leading to this contribution is presented inPaper IV as well as in Patent I.


1.1.2.2 Consistency-aware Rerouting Framework

In SDN architecture, updating the data plane, i.e., establishing new rules in switches’flow tables, may impose congestion. Several works study the congestion imposedduring adaptation and try to avoid or minimize it by sequentially updating theflows [25,26,37–39]. However, these solutions are only applicable to networks wherethe initial state is feasible. In a wireless network affected by rain, the initial net-work state is rendered infeasible due to link capacity reduction and the longer ittakes to transit to the target state, the higher the data loss is. In this condition,the imposed congestion during transition is caused by two factors: the asynchronybetween switches and the delay in transferring each flow to its target state. In thisthesis, we propose a model for calculating the congestion due to these two factors.Then we develop a linear optimization approach to sequentially update the flowssuch that the congestion, defined as switching cost, is minimized. Knowing thegain and cost of rerouting at each time, we try to determine the optimal time forapplying the newly computed routes in order to minimize the total data loss duringrain. To do so, we formulate the problem of computing the best update policy usingdynamic programming (DP). To solve the problem with DP algorithm, the networkstate should be known for present and future time samples. Due to huge complex-ity and network causality, DP is not applicable for online solutions. Therefor, weconvert our problem to short term planning problem by limiting the time horizon.When designing such approach, the update policy can greatly benefit from the factthat rain attenuation follows a Lognormal distribution [40], making it possible topredict the severity of attenuation over the upcoming observation period. As weuse prediction for future time samples, model predictive control (MPC) algorithmis an appropriate choice to solve our DP as an online algorithm. The solution toour problem gives us an update policy that combines the cross-layer information onthe physical layer including capacity of wireless links and prediction of rain atten-uation on each link, as well as the network layer, including traffic flow informationto decide the best time for applying the computed routes. According to the resultspresented in Paper V, using our proposed framework to choose the rerouting deci-sion at each time sample decreases the transition loss during rain period comparedto the regular rerouting. By wisely selecting the appropriate time for applying thecomputed optimal routes, our proposed policy is able to minimize the total dataloss as well as the number of reconfigurations.

1.1.3 A Techno-economic Framework for 5G Backhaul Networks

Inefficient and costly backhaul design can decrease the benefits of HetNet deploy-ment. Therefore, it is necessary to take into account the backhaul network solutiontogether with the access segment in order to obtain a realistic estimation of thetotal network cost. This needs a complete framework for calculating the total costof ownership (TCO) of the backhaul network. The design of cost efficient back-haul solutions for both homogeneous and heterogeneous deployments was addressed

1.2. ORGANIZATION OF THE THESIS 11

in [13,41–43]. However, accurate cost evaluation requires a proper cost model cov-ering different aspects such as capital expenditures (CAPEX) (e.g., infrastructureand equipment cost ) as well as operational expenditures (e.g., maintenance, faultmanagement and energy cost) that may affect TCO. Such a comprehensive techno-economic model is still missing in the literature. Moreover, only TCO evaluation isnot sufficient to provide the operator an insight into the profitability of the network.To address this problem, we extend our cost model with the profitability analysis,making it possible to evaluate cash flow and net present value of the mobile back-haul networks. The proposed techno-economic framework can be applied for bothfiber and microwave backhaul deployments. A case study is carried out to find themost profitable and cost-efficient backhaul solution for given scenarios. The resultspresented in Paper VI clarify the significant contribution of the backhaul segmentto the total cost and profitability of solutions. Specifically, it is shown that fibertechnology is more efficient than wireless solution in terms of cost and profitabilityfor ultra-high density access networks.

1.2 Organization of the Thesis

The thesis is organized as follows:

• Chapter 2 proposes topology design approaches that aim to find the optimalnetwork configuration such that reliability threshold is satisfied while the costof network deployment is minimized. We investigate the design problem undercorrelated failures by introducing a new model for considering their spatialcorrelation. The evaluation results on both synthetic data and data mea-sured in a real deployed network verify that our topology design significantlyimproves the reliability performance.

• Chapter 3 extends our topology design problem to jointly optimize the numberof added links and link budget in order to satisfy the reliability, delay anddata rate requirements of all terminals. We investigate the performance of ourdesign algorithm and show that joint optimization can substantially decreasethe total link budget compared to the benchmarking approaches.

• Chapter 4 introduces an efficient detection algorithm for distinguishing therain fading from multipath fading as a prerequisite for improving the networkthroughput during rain. The results of the detection are used to trigger thererouting procedure for rain mitigation. After presenting the assumed modelfor the SDN controller, network and channel model in the presence of rain,we evaluate the performance of the proposed detection algorithm and itsimpact on network throughput. The results indicate that rerouting upon raindetection significantly increases throughput compared to adaptive modulationand coding.


• Chapter 5 provides the consistency-aware rerouting framework to mitigaterain which considers the switching cost. We first calculate the imposed re-configuration cost and then determine the best rerouting policy by modelingthe problem using dynamic programming. The evaluation results on bothsynthetic data and data measured in a real deployed network indicate theefficiency of our proposed algorithm in terms of the total required number ofreconfigurations and the imposed data loss.

• Chapter 6 presents our proposed economic framework in order to evaluatethe benefits of different deployment options in terms of cost and profitability.The profitability evaluation for different scenarios reveals the best choice forbackhauling in highly dense areas.

• Chapter 7 draws concluding remarks and lessons learned from this thesis.Moreover, it outlines the remaining challenges and possible extensions of thework.

Chapter 2

Topology Design for ReliableNetworks under CorrelatedFailures

Inherent vulnerability of wireless backhauling to random fluctuations of the wirelesschannel, such as those caused by rain, complicates the design of reliable backhaulnetworks. In the presence of such disturbances, providing redundant paths betweena given source and destination can significantly improve network reliability. Manystudies deal with modifying and designing the network topology to meet the reli-ability requirements in a cost-efficient manner. However, these studies ignore thecorrelation among link failures, particularly those caused by weather disturbances.Consequently, the resulting topology designs may fail to satisfy the network reli-ability requirements under correlated failure scenarios. To address this issue, inthis chapter we focus on the design of cost-efficient and reliable wireless backhaulnetworks under correlated failures considering rain disturbances. We first proposea new model to consider the pairwise correlation among links along a path. Themodel is verified on real data, indicating a closer approximation to reality than theexisting independent model. Secondly, we consider the correlation among differentpaths by defining a penalty factor in the objective function. Higher value of thepenalty factor corresponds to a larger inter-path correlation and steers the approachtowards selecting less correlated paths. Using the newly formalized link and pathcorrelation, we formulate the network topology design problem as a quadratic in-teger program to find the optimal solutions for smaller problem instances. As theproblem is shown to be NP-hard, two lightweight heuristic algorithms are developedto find near-optimal solutions for larger instances in reasonable time. Performanceevaluation shows that correlation-aware design substantially improves the resiliencyunder rain disturbances at a slightly increased cost compared to approaches thatconsider only independent failures.

13

14CHAPTER 2. TOPOLOGY DESIGN FOR RELIABLE NETWORKS UNDER

CORRELATED FAILURES

2.1 Related Work

Topology design aimed at ensuring or maximizing network reliability has been inthe focus of many studies [9–13]. In [10], the authors evaluate the impact of in-troducing redundant nodes on the network reliability performance for both staticand random wireless multi hop networks. The authors in [11] study the reliabletopology design problem in a time-evolving environment. Their proposed heuristicalgorithms significantly reduce the total network topology cost while maintainingthe reliability threshold over a predefined period. The authors in [12] investigatethe topology design problem to maximize the algebraic connectivity, defined as thesecond smallest eigenvalue of the Laplacian matrix. They propose a method toincrementally add links so that in each step the algebraic connectivity increasesthe most. In [9], the topology design problem is considered from a game theo-retic perspective, where each node increases its connectivity in a non-cooperativeway by adding links to other nodes, constrained by the delay and link cost. Thework presented in [13] addresses the cost-efficient topology design problem with theobjective of satisfying the fluctuating traffic demands and meeting the reliabilityrequirements. However, all the aforementioned works apply a common assumptionof uncorrelated link failures, which does not hold in many scenarios such as weatherdisturbances, e.g., caused by rain, where a set of links may fail simultaneously andthe failures exhibit strong spatial correlation. In this scenario, the above approachesmay lead to inefficient solutions due to their modeling inaccuracy.

Modeling correlated failures for reliability assessment has been studied in [44,45].In these works, spatial correlation between failures leads to the concept of a regionfailure. This model can be categorized into two types: i) deterministic failureswhich consider circular regions formed by spatially correlated areas, where thefailure probability is equal to one inside, and to zero outside the region, and ii)probabilistic failures where the failure probability inside the region monotonouslydecreases with the distance from the failure epicenter [46]. Such models may notbe suitable for rain disturbance due to the circular representation of the failureregion and the constant probability of failure between two consecutive concentricannuluses.

2.2 System Model

In this section, we explain the network and channel model considered in Chapters 2and 3.

2.2.1 Network Model

We consider a typical wireless network that consists of multiple macro base sta-tions (BSs). In such deployment, traffic is transferred from each BS to one of thenetwork gateways using single or multiple hops. The BSs can communicate using a

2.2. SYSTEM MODEL 15

Core

Network

Fiber or leased line

backhaul traffic to the

core network

Figure 2.1: An example network topology.

highly directional antenna providing point-to-point line-of-sight links with negligi-ble mutual interference [7]. We model the physical deployment with graph H(V, E),where each vertex v ∈ V represents a BS, and each edge i ∈ E denotes a wirelesscommunication link. The subset of BSs that serve as gateways is denoted by G.

2.2.2 Channel Model

We consider a network that is partially or completely affected by rain. Dependingon the weather conditions, each link may experience different channel attenuation:path loss in clear sky or rain attenuation plus path loss during rain. Under clear sky,the standard Friss transmission equation for mm waves gives the received power indB of wireless link i as follows:

P ri = P ti +Gri +Gti − ali, (2.1)

where Gti and Gri show the transmitter and receiver antenna gain, respectively, P tiis the transmitted power, and ali is the path loss attenuation given by:

ali = 20 log(

4πdiλ

), (2.2)

where λ is the wavelength of the carrier and di denotes the length of link i.Under the rain disturbance, the total received power in dB is given by:

P ri = P ti +Gri +Gti − ali − ai, (2.3)


CORRELATED FAILURES

where, ai denotes the rain attenuation in dB for a link in clear sky, following theLognormal distribution [47]:

ai ∼ lnN (mi, σi) , (2.4)

where mi and σi are the mean and standard deviation of ln(ai), respectively. Fromnow on, we define the link budget as the summation of the transmit power and theantenna gains in transmitter and receiver of link i, denoted with si = P ti +Gri +Gti.

The average rain attenuation on link i is formulated as:

mi = γrηdi, (2.5)

where r [mm/hour] shows the rainfall intensity, and γ and η are constants thatdepend on carrier frequency [48]. These rain statistics can be derived from long-term measurements during several years or estimated from Recommendations ITU-R P.530 [49].

The rain attenuation shows both temporal and spatial correlation. In order togenerate the time series of rain attenuation, the model should capture both timeand space correlation [47, 48, 50, 51]. The time variability is generated by passingthe white Gaussian noise with zero mean and unit variance from a low pass filterH (z) that has the impulse response given by the Maseng-Bakken model [47]:

H (z) =√

1− µ2

1− µz−1 (2.6)

whereµ = exp (−β∆t) , (2.7)

where β describes the temporal variation of rain attenuation on one link, and ∆tis the sampling time. To create the Lognormal attenuation ai with mean mi andvariance σi at each time sample, the filtered signal should be fed into a memorylessnon-linear transformation:

ai = exp (mi + σix) (2.8)where x is the output of the low pass filter H (z). Note that this model does nottake into account the transitions between the clear and the rainy sky and it can beonly applied during the presence of rain.

In order to model the space correlation of rain attenuation between differentlinks, the independent white Gaussian noise for each link is combined with a spacecorrelation matrix R. Entry (i, j) of the correlation matrix, denoted by κij , showsthe correlation coefficient between links i and j and is calculated as:

κij = ρij√ρiiρjj

, (2.9)

where ρij describes the spatial correlation between links i and j, and ρii shows thespatial correlation of points along link i. To calculate ρij , we first need to computethe rainfall rate correlation between two points, modeled as:

ϕ = exp (−αd) , (2.10)


where d expresses the distance between the two points and α−1 is the correlationdistance, defined as the distance at which ϕ = e−1 [47]. By knowing the rainfall ratecorrelation between any two points on each path, the spatial correlation betweentwo paths is:

ρij =∫ Li

0

∫ Lj

0ϕd`id`j , (2.11)

Similarly, ρii for each link is calculated by double integrating (2.10) along linki. More detailed explanations can be found in [47,52].

The joint probability distribution between the rain attenuation of link i, ai, andlink j, aj , also follows a lognormal distribution that can be expressed as [47]:

f (ai, aj) = 1

2aiajσiσjπ√

1− κ′ij2

exp (−g(ai, aj)) , (2.12)

where

g (ai, aj) = 12 (1− κij ′2)

[(ln ai −mi)2

σ2i

+ (ln aj −mj)2

σ2j

−2κ′ij(ln ai −mi)(ln aj −mj)

σiσj

].

(2.13)In the equation, mi and σi are the mean and standard deviation of ln(ai), and κ′ijshows the correlation coefficient between two links that depends on κij computedfrom (2.9) as follows:

κ′ij =ln(κij√

exp(σ2i )− 1

√exp(σ2

j )− 1 + 1)

σiσj. (2.14)

2.2.3 Link Failure Probability ModelLink failure probability parameter abstracts the ability of the physical layer ofwireless systems to reliably transmit packets [10]. A link is considered to fail whenits received power falls below a certain threshold necessary to achieve a certain biterror rate (BER) that is required to maintain a given quality of service level.

Let us define pth as the threshold corresponding to a link failure event. There-fore, the failure probability of link i is equal to the probability that its receivedpower denoted by P ri is lower than pth. In this work, as we assume that only rainattenuation causes failure, the link failure probability, denoted by pf (si), can becomputed as follows:

pf (si) = Pr (P ri ≤ pth) = Pr(ai ≥ si − ali − pth

)= pr

2 −pr2 erf

(ln(si − ali − pth)− ln(mi)

σi√

2

) (2.15)

where pr is the probability of a rain event.


CORRELATED FAILURES

2.3 Problem Definition

In this section, we formulate the topology design problem first by assuming in-dependent failures, and then we extend the problem formulation to capture thecorrelated failures.

2.3.1 Topology Design under Independent Failures

Given a set of BSs in the network, the topology design problem is to determinethe network topology with minimum cost so as the reliability requirement is met.The reliability metric used in this thesis is defined as the probability of successfultransmission between each node v ∈ V \G and g, possibly using multiple paths. Forsake of simplicity, the optimization problem presented here is for one gateway, butit can be extended to multiple gateways using the method described in Paper II.We assume that each source uses disjoint paths to meet the reliability constraint, asin [45]. This provides an upper bound on the topology cost as it may overestimatethe number of links required to satisfy the connectivity. Throughout this chapter,we formulate the problem for a given source, which is applicable for other terminals.Considering v as a given source, we assume that the set of candidate paths betweenv and a given gateway g is known and denoted by Qv. We define z as the pathselection variable to connect source v to gateway g, where zk is 1 if path qk ∈ Qvis selected to be established between (v, g), and 0 otherwise. Based on the set ofall possible paths, we generate Dk as the set of links included in path k and a pathrouting matrix denoted by C, where element [C]ik is equal to 1 if path qk ∈ Qvcontains link i, and 0 otherwise. Let us define pf =

{p1f , p

2f , ..., p

|Qv|f

}as the path

failure probability vector, where the kth element pkf is the failure probability of pathqk ∈ Qv. With the assumption of uncorrelated failures, pkf is given by:

pkf = 1−∏i∈Dk

(1− Pr (ai ≥ ath)) , (2.16)

where Pr (ai ≥ ath) is the failure probability of link i computed from Eq. (2.15).

Let w = [w1, ..., wL] denote the cost vector where wi captures the deploymentcost of link i. wi is equal to zero for already established links, and takes a non-zerovalue for the potential links to be added to the network. Assuming this cost vector,the objective of the problem is to compute the minimum-cost topology capable ofsatisfying the reliability threshold between the given source v and destination gunder rain condition. Hence, we can formulate the benchmarking topology design

2.3. PROBLEM DEFINITION 19

problem assuming independent failures (TD-IF) as follows:

TD-IF : minz

wTCz (2.17a)

Cz ≤ 1, (2.17b)

ln(pf)T z ≤ ln(1− ε), (2.17c)

z ∈ {0, 1}|Gg| . (2.17d)

Equation (2.17a) expresses the total cost of links in the network, whose minimiza-tion is the objective of our approach. Constraint (2.17b) ensures that the selectedpaths do not share any common links, and constraint (2.17c) enforces the reliabilityrequirement. It should be noted that we have used the ln (.) operator to derive alinear constraint in (2.17c).

2.3.2 Topology Design under Correlated FailuresSo far, we have neglected the impact of correlation in reliability calculations as wellas our topology design. In this section, we aim to modify our TD-IF formulationto capture the spatial correlation of failure events such as rain. To do so, we firstcompute the intra-path correlation by recalculating pf using the joint probabilitydistribution of adjacent links along one path; and inter-path correlation by defininga matrix where each element denotes the pairwise correlation among different paths.We then reformulate the TD-IF problem to capture these two types of correlation.

2.3.2.1 Path Failure Probability

The path failure probability under correlated failures is computed based on jointfailure probability of all links included in that path. Consider a path qk betweena given source and destination. We define the attenuation vector for the set oflinks Dk included in qk as a =

(a1, a2, ..., a|Dk|

). As link failures are correlated, the

failure probability pkf is calculated as:

pkf =1− pks=1− Pr

(a1 ≤ ath, a2 ≤ ath, .., a|Dk| ≤ ath

), (2.18)

where pks is the success probability of path qk, i.e., the probability that all links alongthat path are available. To calculate the success probability based on the aboveequation, we should know the joint probability distribution of their correspondingattenuation ai, ..., ai+|Dk|. However, relying on the existing literature, we only knowthe pairwise joint probability distribution of rain attenuation between different links(see Eq. (2.12)). Hence, we approximate the equation in 2.18 with the followingformulation which considers the pairwise distribution of all consecutive link pairsincluded in the path:

pks ≈ Pr (a1 ≤ ath)∏

i∈|Dk|−1

Pr (ai+1 ≤ ath|ai ≤ ath) . (2.19)


CORRELATED FAILURES

The marginal Pr (ai ≤ ath) and pairwise probability distribution Pr (ai+1 ≤ ath, ai ≤ ath)can be computed from (2.4) and (2.12), respectively.

To capture the correlation among different paths connecting source v to gatewayg, we define T as the path correlation matrix where [T]kl = tkl quantifies the cor-relation between paths qk and ql. The element tkl is computed from the correlationcoefficient between each pair of links in the paths given by:

tkl =∑i∈Dl

∑j∈Dk

κij , (2.20)

where Dl and Dk are link sets included in path ql and qk, respectively, and κij isdefined by Eq. (2.9).

2.3.2.2 Problem Formulation

In this subsection, we transform the TD-IF optimization problem to consider bothintra-path and inter-path correlation. To capture the former term, the path fail-ure probability pkf from (2.17c) is approximated using Eqs. (2.18) and (2.19). Toinclude the latter term, we add a penalty part in the objective function that isproportional to an estimate of the inter-path correlation. With this penalty term,the optimization is steered towards the set of solutions with low correlation. Letτ12...m denote the penalty cost of simultaneously choosing paths {q1, q2, ..., qm} forconnection between source v and gateway g. We define it based on the pairwisecorrelation coefficient between different paths [T′]kl = t′kl, as follows:

τ12...m =m∑k

m∑k<l

t′kl. (2.21)

This penalty should be paid if the path selection variable zk corresponding toeach selected path qk is equal to 1, so the cost can be formulated as τ12...mz1z2...zm.The penalty term has a combinatorial nature which makes the optimization prob-lem computationally prohibitive. Hence, we simplify the penalty formulation byconsidering the sums all possible path pairs given by:

τ12...mz1z2...zm ≈m∑k

m∑k<l

t′klzkzl. (2.22)

The objective function of our topology design is a convex combination of total de-ployment cost (2.17) and penalty cost (2.22). Defining this objective, our topologydesign with correlated failures (TD-CF) problem is formalized as follows:

2.4. PERFORMANCE EVALUATION 21

TD-CF : minz

µwTCz + (1− µ) zTT′z (2.23a)

Cz ≤ 1 , (2.23b)

ln(pf)T z ≤ ln(1− ε) , (2.23c)

z ∈ {0, 1}|Qg| . (2.23d)

In the equations, t′kl is the (k, l)-th element of path correlation matrix T′ computedin (2.20), and µ is a constant in [0, 1], which allows for tuning the relevance of costand failure correlation during optimization. The optimization problem above isa mixed integer quadratic program which has been shown to be NP-hard in gen-eral [53], implying poor scalability. Therefore, we develop two heuristic algorithmsto provide near-optimal solutions at much lower complexity. These two algorithmsare based on Lagrangian and continuous relaxation, two common approaches foralleviating the complexity of integer programming. We show that the continuousrelaxation significantly reduces the complexity, while the one based on Lagrangianrelaxation gives a solution closer to optimal at the expense of higher complexity.The detailed explanation of each heuristic algorithm and their complexity analysesare presented in Paper I.

2.4 Performance Evaluation

In this section, we evaluate the performance of our proposed TD-CF algorithms andcompare the results with the TD-IF topology design. The considered algorithmsare as follow:

• TD-IF, where we obtain optimal solution to the TD-IF problem formulatedin (2.17) using the CPLEX solver.

• TD-CF, where we obtain optimal solution to the TD-CF problem formulatedin (2.23) using CPLEX solver.

Table 2.1: Simulation parameters.

Parameters Description Valueγ Rain-related parameter 0.15η Rain-related parameter 1.04r Rain rate 20 [mm/hour]β Temporal variation 1/9 min−1

σ Variance of rain attenuation 0.9α De-correlation factor 0.2 km−1

ath Failure event threshold −60 [dB]


CORRELATED FAILURES

0 0.2 0.4 0.6 0.8 1

µ

0.9988

0.99885

0.9989

0.99895

0.999

0.99905

0.9991

0.99915

Averageconnectionavailability TD-IF

TD-CF

TD-CF-L

TD-CF-R

ǫ = 0.999

(a)

0 0.2 0.4 0.6 0.8 1

µ

26

28

30

32

34

36

38

Cost[m

]

TD-IF

TD-CF

TD-CF-L

TD-CF-R

(b)

Figure 2.2: (a) Average availability, and (b) total deployment cost obtained by thefour algorithms for different values of the weight factor µ.

• TD-CF-R, where the heuristic algorithm based on continuous relaxation isused to solve the TD-CF problem.

• TD-CF-L, where the heuristic algorithm based on Lagrangian relaxation isused to solve the TD-CF problem.

To investigate the reliability of each designed topology under correlated failuressuch as those caused by rain, we subject the (sub-)optimal topology obtained byeach algorithm to simulated rain. To do so, we generate 10000 random rain eventsconsidering both temporal and spatial correlation using the model presented inSection 2.2.2. The parameters used in simulations are summarized in Table 2.1.We assume that the links fail due to rain attenuation and their failure probabilitiescan be computed from Eq. (2.15). We investigate the performance of our approachesfor a small synthetic network as well as a realistic deployed topology in the followingsections.

2.4.1 Synthetic NetworkIn the first part, we consider a synthetic small network where |V| = 8 nodes arerandomly distributed in the |V| × |V| plane. To eliminate the dependency of theperformance evaluation to nodes location, the results for each parameter settingare averaged over 50 random topologies. In each test case, the gateway is pickedrandomly among the network nodes.

Fig. 2.2 shows the effect of the weight factor µ on the average path availabilityand the deployment cost of the topologies obtained by four proposed algorithmswhen reliability threshold is set to ε = 0.999. The average availability results shownin Fig. 2.2a show that the TD-IF approach fails to satisfy the reliability thresholdunder correlated failures in all cases, while our proposed TD-CF algorithms meetthe requirement for the weight factor µ ≤ 0.18. This is because for smaller valueof µ, i.e., higher penalty cost, the TD-CF algorithms choose spatially uncorrelated


-5 0 5 10 15 20 25

x [km]

-45

-40

-35

-30

-25

-20

-15

-10

y[km]

1

2

3

4

Source

Gateway

Figure 2.3: Realistic wide-area network topology used in the simulations.

paths which provides higher availability under correlated failures. The cost resultsin Fig. 2.2b illustrate that the TD-IF approach yields the lowest deployment cost.For the TD-CF approaches, the cost increases for a small value of µ as the spa-tially uncorrelated paths require addition of more links. For instance, to reach thereliability threshold when µ = 0.18, the optimal topology obtained by the TD-CFapproach results in 18% higher deployment cost than TD-IF. Comparing the re-sults of the two heuristic algorithms with the optimal TD-CF results, it can beseen for µ ≤ 0.18 that the TD-CF-L solution meets the reliability threshold withlower cost and performs more closely to the optimal solution, while TD-CF-R incurshigher cost. In particular, for value µ = 0.18, the cost of the solutions obtainedby TD-CF-R and TD-CF-L is 4% higher and 4% lower than the TD-CF optimum,respectively.

2.4.2 Realistic Deployed Network

In this part of our evaluation, a fixed-wireless network is considered that is currentlydeployed as a backhaul network in Sweden [54]. The topology is shown in Fig. 2.3and includes 30 BSs and multiple gateways. The number next to each gatewaydefines the order of using that gateway, where gateways are sequentially addedto the network. For instance, |G| = 3 means that nodes 1, 2, 3 are considered asgateways, while node 4 is assumed to be a source.

Fig. 2.4 illustrates the impact of the number of gateways on the network avail-ability and cost obtained by four algorithms when reliability threshold is set toε = 0.999. The results in Fig. 2.4a show that a higher number of gateways improvespath availability under correlated failures for both TD-IF and TD-CF approaches,as it increases the chance of finding uncorrelated paths due to the possibility ofchoosing different gateways to connect to. However, it can be seen that TD-IF stillcannot reach the reliability threshold even for a higher number of gateways, whileour TD-CF approaches satisfy it when |G| ≥ 3. It should be noted that, when


CORRELATED FAILURES

1 2 3 4

Number of gateways |G|

0.996

0.9965

0.997

0.9975

0.998

0.9985

0.999Averageavailability

TD-IF

TD-CF

TD-CF-L

TD-CF-R

ǫ = 0.999

(a)

1 2 3 4

Number of gateways |G|

20

40

60

80

100

120

Cost

TD-IF

TD-CF

TD-CF-L

TD-CF-R

(b)

Figure 2.4: (a) Average availability, and (b) total deployment cost obtained by thefour algorithms for a different number of gateways.

1 2 3 4

Number of gateways G

0

200

400

600

800

1000

1200

Runtime[s]

TD-IF

TD-CF

TD-CF-L

TD-CF-R

Figure 2.5: Run time obtained by four algorithms for a different number of gate-ways.

the number of gateways is low, i.e. |G| ≤ 2, the chance of finding highly uncorre-lated paths decreases, which causes some terminals to fail to meet the availabilityrequirement. The cost assessment in Fig. 2.4b shows that a higher number of gate-ways reduces the cost of achieving the same reliability threshold. The comparisonof the TD-CF and TD-IF algorithms at |G| = 3 shows that TD-CF imposes 60%higher deployment cost compared to TD-IF to guarantee the selection of uncorre-lated paths. Moreover, the results verify that the TD-CF-L solutions cause 10%higher cost due to sub-optimality, while this value reaches 30% for the TD-CF-Rapproach. The small optimality gap of TD-CF-L comes at the expense of highercomplexity which is analyzed in the following.

In order to estimate the computational complexity of each proposed algorithm,we record their run time. All algorithms are executed on a workstation runningRed Hat Enterprise Linux with an 8-core 16-thread Intel Xeon processor clockedat 3 GHz and 64 GB of RAM. The corresponding results are presented in Fig. 2.5which verifies that the heuristic algorithms based on continuous relaxation substan-tially reduce the complexity compared to the TD-CF formulation. The TD-CF-R


approach exhibits the strongest run time reduction compared to TD-CF. TD-CF-Lexhibits longer run time compared to TD-CF-R, which leads to a better approxi-mation of the optimal solution. Compared to TD-CF, its run time reduction is themost significant for a greater number of gateways.

Chapter 3

Joint Topology Design and Controlfor Reliable Networks

Besides deciding on the set of necessary links to be established in the network, inves-tigated in the previous chapter, another way of achieving a pre-defined availabilitythreshold can be by tuning the link budget to compensate for rain attenuation.A number of existing studies are related to link budget adjustment for meetingthe reliability performance requirements. However, these approaches either ignorethe impact of diversity or consider k − edge connectivity as the only metric foravailability, which makes their solutions inefficient or inapplicable to 5G wirelessbackhaul network scenarios considered in this thesis. In this chapter, we performtopology design with the objective of minimizing the total link budget so that thenumber of added links is limited, while the availability and data rate constraintsare met. To ensure fairness in resource allocation, we make sure that a prede-fined minimum data rate is supported for each source. We formulate the jointoptimization problem as a mixed integer non-linear program. As this is shown tobe NP-complete, we develop a heuristic algorithm by applying the decompositionmethod to the Lagrangian relaxation of the problem. The evaluation results showthat our joint topology design can meet the reliability performance requirementwith a significantly lower link budget compared to the benchmarking algorithmswhich rely either on link budget or topology adaptation only.3.1 Related WorkOur work is on the intersection of topology design and control as well as multipathrouting. In the following, we provide a survey of the related works in these areas.

Indeed, splitting the traffic among multiple (disjoint) paths (multipath routing)not only provides high network availability, but can also increase the data rate. Con-siderable amount of work in the literature, such as [55,56], has studied the problemof multipath routing. The work in [55] introduces a new metric called effectivebandwidth that provisions the required data rate taking into account the connec-tion availability over multiple paths which are not necessarily disjoint. Moreover,the authors extended their work with cost considerations while providing reliable

27

28CHAPTER 3. JOINT TOPOLOGY DESIGN AND CONTROL FOR

RELIABLE NETWORKS

connections. An adaptive multipath routing assuring the requested data rate anddifferential delay is proposed in [56]. The authors provision multiple disjoint pathsto satisfy the requested data rate and guarantee that at least a fraction of traffic isprotected from any single point of failure. Essentially, in these works the link fail-ure probability is fixed and the only parameter to achieve reliability performanceis multipath diversity. However, in real wireless networks the antenna gain andtransmit power can also be controlled to change the link failure probability.

Topology control refers to the problem of determining the transmit power, an-tenna gain and direction for achieving the predetermined network performancesuch as availability and energy efficiency. Topology control focusing on differentobjectives, for example availability, has been extensively studied in a wide bodyof literature [14–18]. For instance, the work in [14] proposes two topology con-trol algorithms that minimize the maximum transmission power of each node andguarantee the network connectivity. The joint problem of routing and topologycontrol is investigated in [15] where the authors optimized the network topologyand routes between the BS and gateway to satisfy the traffic requirement whilebalancing the load across all links. In [16], the authors proposed a multi-objectivetopology control using a game-theoretic algorithm to optimize the transmissionpower and end-to-end delay. These mentioned works consider 1-edge connectivity,namely robustness in the presence of single edge failures, which may fail to guaran-tee the reliability performance under multiple failures. This problem is mitigatedin [17, 18] where the authors proposed a fault-tolerant algorithm by guaranteeingk-edge connectivity during topology control. A graph is k-edge connected if it isstill connected after the deletion of any k random edges. The work in [17] proposesa reliable topology control to preserve the k-edge connectivity to maximize the en-ergy efficiency. This algorithm is extended for networks with dense BS (vertex ofgraph) deployments in [18] by proposing a local tree-based reliable topology algo-rithm. In [9] the authors study the connectivity performance with respect to thecost of establishing links, delay and interference from a game-theoretic perspective.The algorithms developed in the above works assure k-edge connectivity and donot capture the availability requirement specified in the most QoS requirements.Hence, there is a need for an efficient algorithm that optimizes the topology designsatisfying the end-to-end availability requirement.

3.2 Problem Description and FormulationIn this section, we first formulate the topology design problem to minimize the totallink budget considering the trade-off between the path diversity (number of addedlinks) and the total link budget so that the reliability performance threshold issatisfied for all network terminals. We consider independent link failures and leavethe correlated failure scenarios for future work. We show that our optimizationproblem is a mixed integer nonlinear program and provide a sub-optimal solutionbased on Lagrangian relaxation and decomposition.

To formulate the problem, we generalize the topology design problem presentedin Section 2.3.1 to jointly consider all sources together. Therefore, all the assump-

3.2. PROBLEM DESCRIPTION AND FORMULATION 29

tions and notations are the same. Recall thatQv denotes the set of all possible pathsbetween source v and a gateway, denoted by g. Although the optimization problemis presented here for one gateway for simplicity, it can be extended to multiple gate-ways using the method described in Paper II. Path selection matrix is denoted withZ =

[z1, z2, ..., z|V\G|

], where element [Z]kv is equal to 1 if path k ∈ Qv is selected

to be established between (v, g′), and 0 otherwise. Let F =[f1, f2, .., fv, .., f|V\G|

]represent the flow matrix where vector fv is the corresponding flow for each pathbetween source v and gateway. For each source v, we define path routing matrixCv, set of all links Dv, and the cost vector w = [w1, ..., wL] similar to the problemformulation in section 2.3.1. Let us denote the path failure probability matrix asPf =

[p1,p2, ...,p|V\G|

], where pv represents the vector of path failure probability

for each path k ∈ Qv originating from source v given by:

[pv]k = ln

1−∏i∈Dv

k

(1− pf (si))

, (3.1)

where si is the link budget and pf (si) is the failure probability of link i computedfrom Eq. (2.15). It should be noted that we use ln (.) operator to derive a linearequation.

In addition to path selection variable Z, in this chapter, we define the linkselection vector x, where xi is equal to 1 if link i is selected to be included inthe network, and 0 otherwise. Considering s = [s1, s2, ..., sL] as the link budgetvector where si denotes the budget of link i, our objective is to minimize the totallink budget needed to satisfy the reliability performance threshold and requesteddemand for each source when a limited number of links is added (cost constraint).We denote the maximum number of added links by lth, availability threshold by ε,and the requested demand by Fth. The optimization problem with the objective ofminimizing the total network link budget can then be formulated as follows:

TD-PA : minx,Z,F,s

sTx (3.2a)

s.t. xTw ≤ lth (3.2b)Cvzv ≤ 1, ∀v ∈ V \ G, (3.2c)∑v

Cvzv ≤Mx, (3.2d)

pv (s) zv ≤ ln (1− ε) , ∀v ∈ V \ G, (3.2e)s ≤ PmaxI, ∀i ∈ E , (3.2f)∑v

Cvfvzv ≤ ci, ∀i ∈ E , (3.2g)

fvI ≥ F vth, ∀v ∈ V \ G. (3.2h)θF vth ≤ fv, ∀v ∈ V \ G. (3.2i)zv,x ∈ {0, 1} , ∀v ∈ V \ G. (3.2j)


RELIABLE NETWORKS

where ci is the capacity of link i, and M is an arbitrary large number. Let Bdenote the available bandwidth for each link, then the value of ci is computed fromShannon theory given by:

ci = B log(

1 + 10(si−mi)/10

N0B

)(3.3)

where si and mi are the link budget and the average of the rain attenuation for linki, respectively. Equation (3.2a) shows the total link budget for the selected links,whose minimization is the objective of our approach. Constraint (3.2b) assures thatthe number of total links deployed in the network is limited to lth. Constraint (3.2c)enforces that the selected paths do not share any common links. Constraint (3.2d)makes sure that link i is selected if at least one of the sources uses that link to toconnect to the gateway g, and constraint (3.2e) enforces the reliability performancerequirement for each source. Constraint (3.2f) limits the maximum link budget,and constraint (3.2g) ensures that the total flow traversing link i does not exceedits capacity. Constraint (3.2h) enforces the summation of flow allocated to sourcev to a value higher than the requested demand, and constraint (3.2i) ensures thateach path connecting source v to gateway carries at least a portion θ of the totalrequested flow, denoted by θF vth. By controlling θ, we can adjust the minimumactivity of every link.

3.3 Topology Design with Joint Power and Data RateOptimization

The problem formulated in Eq. (3.2) is a mixed integer nonlinear problem shownto be NP-complete [57]. Applying decomposition method on dual function can beused to solve this type of problem [58]. In such approach, the bigger problem ispartitioned into smaller sub-problems which are solved in parallel or sequentiallywith much lower complexity compared to the original one. We propose a heuristicalgorithm by applying the decomposition method to its Lagrangian dual function.If vectors λ and α show the Lagrangian multiplier, the dual function is formulatedas follows:

g (λ,α) = minx,Z,s,F

L (λ,α, s,x,F,Z) (3.4a)

s.t. (3.2b), (3.2c), (3.2d), (3.2f), (3.2h), (3.4b)

where

L (λ, s,x,Z) = sTx +∑v

λvpvzv −∑v

λv ln(1− ε) (3.5)

+∑i

αi∑v

Cvfvzv −∑i

αiB log(

1 + 10(si−mi)/10

N0B

). (3.6)

3.3. TOPOLOGY DESIGN WITH JOINT POWER AND DATA RATEOPTIMIZATION 31

Fixing the path and link selection variables, zv and x, the dual function can bedecomposed into two disjoint problems of solving the link budget and flow assign-ment as follows:

h1 =mins

sTx +∑v

λvpvzv −∑i

αiB log(

1 + 10(si−mi)/10

N0B

)(3.7a)

s.t. (3.2f). (3.7b)

and,

h2 =minfv

∑i

αi∑v

Cvfvzv (3.8a)

s.t. fv ≥ θF vth, ∀v ∈ V \ G. (3.8b)

Then the original problem, also called the master problem, is equivalent to:

g (λ,α) =minx,Z

sTx +∑v

λvpvzv +∑i

αi∑v

Cvfvzv (3.9a)


Cvzv ≤Mx, (3.9d)

To find the solution for the problem in (3.4), the decomposition method firstfinds the optimal values for the sub-problems (3.7) and (3.8), and then solves themaster problem by substituting the value for s and fv. In the following, we explainthe solution for each problem.

3.3.1 SubproblemsThe second sub-problem, formulated in (3.8), is a linear program which can besolved in linear time. However, the first one, formulated in (3.7), is a nonlinear,non-convex optimization problem. Fortunately, we can prove that h1 can be writtenas difference of convex (DC) in Paper III. Some existing approaches such as branchand bound and cutting plane solve the DC problem to the global optimum but havevery high complexity, while some convex-based algorithms such as difference ofconvex approach (DCA) significantly reduce the complexity but only reach a localoptimum. In DCA, a number of regularization and starting-point selection methodsexist that can help the algorithm to yield a globally optimal solution [59–61].

DCA defines two sets of primal and dual variables, denoted by sn, yn, so that theformer meets the local optimum of the primal problem, s∗, and the latter convergesto the local optimum of the dual problem y∗. The algorithm first initializes theprimal variables by choosing a random and feasible point s0. In iteration n, the


RELIABLE NETWORKS

dual variable yn is computed from the following optimization problem:

yn = argmaxy

{sTny− g∗cvx(y)

}, (3.10)

where g∗cvx is the conjugate function of gcvx that can be computed by:

g∗cvx(y) = supt∈R{tTy− gcvx(t)

}. (3.11)

We show that if gcvx is differentiable, the problem in (3.10) can be simplified asfollows:

yn = g′cvx(sn) (3.12)

Due to duality, the primal variable s at next step n+ 1 is updated by:

sn+1 = argmins∈R

{hcvx(s)− sTyn

}. (3.13)

The algorithm continues iterating until the change in sn or yn is lower than prede-fined error, denoted by ξ.

3.3.2 Master ProblemAfter computing the optimal value of s and fv from each subproblem, the masterproblem formulated in (3.9) should be solved. This problem is a linear integerprogram and there exist efficient heuristic algorithms [62] to obtain a near to optimalsolution.

The solution to the above problem provides a lower bound to problem (3.2).Hence, maximizing the Lagrangian dual function with respect to Lagrangian mul-tiplier provides the closest lower bound:

maxλ,α

g (λ,α) . (3.14)

To minimize the dual function with respect to Lagrangian multiplier, we use itera-tive sub-gradient method as follow:

λn+1 = max(

0,λn − tλ∂g(λ)∂λv

), (3.15)

where tλ is step size and ∂g(λ)∂λv

shows the derivation of g(λ) relative to λ:

∂g(λ,α)∂λv

= gvzv − ln(1− ε). (3.16)

andαn+1 = max

(0,αn − tα

∂g(α)∂αi

), (3.17)

3.3. TOPOLOGY DESIGN WITH JOINT POWER AND DATA RATEOPTIMIZATION 33

Algorithm 1 Power-aware topology design with path diversity (PA-TD-D)1: Given: All source and destination pairs, set of all paths Qv from all sources v ∈ V \G

to the master gateway g′, the link cost vector w, the reliability performance thresholdε, the requested demand from each source Fth and the path failure probabilities matrixP.

2: Initialization: Set m = 0, step size tα ≥ 0, tλ ≥ 0, and ξ = 0.01. Initialize the linkand path selection variables, x0 and Z0, as well as Lagrangian multipliers, α0 ≥ 0and λ0 ≥ 0.

3: while ‖λm+1 − λm‖ ≥ ξ or ‖αm+1 − αm‖ ≥ ξ do4: Compute the optimal link budget sm and the optimal data rate Fm for the optimal

topology Zm by solving the sub-problems formulated in (3.7) and (3.8), respectively.5: Update the path and link selection variables Zm+1 and xm+1 by solving the master

problem in (3.9).6: Update λm+1 = max

(0,λm − tλ ∂g(λ)

∂λv

),

7: Update αm+1 = max(0,αm − tα ∂g(α)

∂αi

),

8: Project the unfeasible solution s to feasible one.9: m← m+ 1

10: end while

where tα shows the step sizes and ∂g(λ)∂αi

expresses the derivative of g(α) with respectα, given by:

∂g(λ,α)∂αi

=∑v

Cvfv −B log(

1 + 10(si−mi)/10

N0B

). (3.18)

The pseudo-code of the described approach is explained in Algorithm 1. Thealgorithm takes as inputs the set of all paths from each source Qv to the gatewayand their corresponding failure probability matrix P, the weight vector w, therequested demand, Fth and the reliability performance threshold ε. The algorithmis initialized by setting the iteration variablem to 1, assuming a feasible topology x0,and Z0, and choosing a random starting point for the Lagrangian multiplier λ0, andα0, as well as choosing their step size (line 2). At iterationm, the algorithm uses theoptimal value of path and link selection variable Zm and xm from previous iterationas well as λm, αm to solve the subproblems in (3.7) and (3.8) (line 4). Then thenetwork topology is recomputed based on the allocated optimal link budget, smand data rate Fm by solving the master problem in (3.9) (line 5). At the next step,the algorithm updates the Lagrangian multipliers using the sub-gradient method(line 6-7). If the achieved primal solution s is not feasible, i.e., does not satisfythe constraints in (3.2e), (3.2f) and (3.2g), we project the infeasible solution to afeasible one by finding the closest value that satisfies the original constraint (line 8).Finally, this process will iterate as long as the changes in the Lagrangian multiplierare higher than ξ (line 3-9).


RELIABLE NETWORKS

3.4 Topology Design with Delay Consideration

The algorithm developed in the previous section for topology design may establishmultiple paths between each source and gateway to reach the availability threshold.In this section, we evaluate the imposed differential delay due to this multipathprovisioning. To do that, we first study the characteristics of optimal solution of ourtopology design problem and then propose a modified topology design formulationconsidering possible constraint on differential delay.

For a single source and gateway, let Q represent the set of paths, |qj | denote thethe number of hops, and fj capture the data rate of path j. We state two followingLemmas.

Lemma 3.4.1 Considering s as the link budget of a link to achieve the minimumdata rate Fth, we have:

1. The objective function of our topology design problem formulated in (3.2),∑i si is given by:

ψ (|qj |, fj) =∑j

sj |qj | =∑j

|qj | (1 + s)fjFth − |qj |, (3.19)

2. ψ(|qj |, fj) is Schur concave with respect to the length of their chosen paths|qj |, and it is minimized when the solution contains paths whose lengths areunevenly distributed and the shorter paths are assigned higher data rates thanlonger paths, or ψ(|qj |, fj) ≤ ψ(|qk|, fk), if |qk| ≺ |qj | when fk ≤ fj for|qk| ≤ |qj |.

A proof is presented in III, Appendix A.

Lemma 3.4.2 If the failure probabilities of all links are the same, the failure prob-ability of a connection between a source and the gateway is Schur concave withrespect to |qj |.

A proof is presented in [45].Based on Lemmas 3.4.1 and 3.4.2, it can be inferred that a power efficient solu-

tion leads to high differential delay. To cope with this issue, we modify the topologydesign problem by adding a differential delay constraint. Let vector łv show thenumber of hops in all paths between source v and the gateway g′. Assuming dthas the differential delay threshold, we assure that the difference between the max-imum and minimum number of hops of chosen paths does not exceed dth. Hence,the delay constraint is given by:

max{

łTv zv}−min

{łTv zv

}≤ dth. (3.20)


Before adding the above constraint to our topology design, we linearize (3.20)by introducing two variables, t1 = max

{łTv zv

}, t2 = min

{łTv zv

}. Considering this,

the optimization problem is modified as follows:

g (λ,α) =minx,Z

sTx +∑v

λvpvzv +∑i

αi∑v

Cvfvzv (3.21a)


Cvzv ≤Mx, (3.21d)

t1 − t2 ≤ dth, (3.21e)łTv zv ≤ t1,∀v ∈ V \ G, (3.21f)łTv zv ≥ t2,∀v ∈ V \ G. (3.21g)

We can use the developed heuristic approach in Algorithm 1 to solve the topologydesign problem with the delay constraint, the only change is adding the (3.21e),(3.21f) and (3.21g) to the master problem.


In this section, we evaluate the performance of our Power-Aware Topology Designwith Path Diversity (PA-TD-D) and compare it with two benchmarking approaches.The considered algorithms are listed in the following:

• Equal Power Topology Design (EP-TD): where the topology design considersonly diversity and optimizes the multipath routing to reach the availabilityand data rate requirements. In this approach the link budget is allocatedequally so that the total link budget that can satisfy both the data rate andavailability requirements is minimized.

• Power-Aware Topology Design without Path Diversity (PA-TD-WoD): wherethe topology design optimizes the link budget allocation while consideringonly one possible path between each source and gateway in the network (treetopology).

• Power-Aware Topology Design with Path Diversity (PA-TD-D): where thetopology design jointly optimizes the link budget and multipath routing usingour developed heuristic algorithm. To reach the optimal solution, we run theAlgorithm 1 for 100 times and choose the point with the minimal total linkbudget.

For simulation we use a part of network deployed as backhaul in Sweden thatcovers a metropolitan area. Our topology consists of 14 BSs which can connect to


RELIABLE NETWORKS

-4 -2 0 2 4 6 8

x [km]

-45

-40

-35

-30

-25

-20

-15

-10

y[km]

Source

Gateway

Figure 3.1: Network topology with 14 BSs and 2 gateways.


Parameters Description ValuePmax Maximum link budget 100 [dB]ε Availability threshold 0.9999lth Maximum added links 20θ Minimum data rate 0B Bandwidth 28[MHz]Fth Data rate threshold 0.1 [Gbps]pr Rain event probability 0.06ath Failure threshold -60 [dB]

two gateways shown in Fig. 3.1. We consider a scenario where rain attenuation isthe only cause of failures and the link failure probability is obtained from Eq. (2.15).The default simulation parameters are reported in Table 3.1.

Fig. 3.2 presents the total link budget for different values of the availabilitythreshold. The results show that PA-TD-D performs the best and significantlyreduces the link budget compared to both EP-TD and PA-TD-WoD approaches,with the gain increasing for higher values of ε. Specifically at ε = 0.9999, it canbe seen that jointly solving the topology design and link budget problem reducesthe link budget by 44% compared to PA-TD-WoD and 66% compared to topologydesign using equal link budget allocation (EP-TD). Moreover, the results expressthe inherent trade-off between the path diversity and link budget adjustment toachieve the reliability performance threshold. For instance, in PA-TD-WoD up-grading the network from 3-nine to 4-nine availability requires increasing the totallink budget by 68%. On the other hand, if we allow path diversity by adding 6 morelinks (PA-TD-D), the value is reduced to 42% compared to PA-TD-WoD. Besides,the reliability performance of ε = 0.99999 is not achievable using PA-TD-WoD


0.999 0.9999 0.99999

ǫ

0

200

400

600

800

1000

Total

linkbudget[dB] PA-TD-D

PA-TD-WoD

EP-TD

Figure 3.2: Total link budget for different value of maximum link budget usingEP-TD, PA-TD-D and PA-TD-WoD approaches.

0 0.05 0.1 0.15 0.2 0.25

Fth [Gbps]

0

200

400

600

800

1000

Total

linkbudget[dB]

PA-TD-D

PA-TD-WoD

EP-TD

Figure 3.3: Total link budget for different values of the data rate threshold usingPA-TD-D, PA-TD-WoD and EP-TD approaches.

with Pmax = 100 dB while our proposed algorithm PA-TD-D meets the requiredthreshold with 65% lower link budget compared to EP-TD.

Fig. 3.3 presents the total link budget obtained by PA-TD-D, PA-TD-WoD andEP-TD approaches for different values of data rate threshold. As expected, higherdata rate threshold increases the link budget required to meet the demand. Theresults show that our proposed joint topology design PA-TD-D always outperformsthe PA-TD-WoD and EP-TD approaches and the difference between the equalpower topology design EP-TD and the other two approaches, PA-TD-D, and PA-TD-WoD, significantly increases for higher data rate requirements. Moreover, theresults denote that PA-TD-WoD cannot satisfy the required minimum data rateafter a certain value, in our case Fth = 0.2 [Gbps], due to Pmax limitation on eachlink.

Fig. 3.4 shows the total link budget for a varied number of added links lth. Theresults demonstrate that link budget for both PA-TD-D and EP-TD continues todecrease when the deployment of more links is allowed, since it provides greaterdiversity to reach the reliability performance threshold. However, the improvementgain due to such diversity in PA-TD-D diminishes after a certain value of lth, in


RELIABLE NETWORKS

14 16 18 20 22 24 26 28 30

lth

0

100

200

300

400

500

600

Total


EP-TD

Figure 3.4: Total link budget for different values of the maximum number of addedlinks using PA-TD-D and EP-TD approaches.

1 2 3 4 5

dth

0

50

100

150

Total


Figure 3.5: Total link budget for different values of the maximum differential delayusing PA-TD-D approach.

this case for lth = 20.Fig. 3.5 assesses the impact of the maximum differential delay, dth, on the total

link budget for our PA-TD-D approach. The results verify Lemma 3.4.1 claimingthat tolerating higher differential delay reduces the link budget for achieving thepredefined availability threshold. Specifically in our scenario, the link budget canbe decreased by 89% if a service can bear the differential delay increase of 1 hop,i.e. changing dth from 1 to 2.

Chapter 4

Accurate Rain Detection forImproved Network Performance

Wireless backhauling increases the flexibility of backhaul networks and supportsplug-and-play mobile backhauling, where new base stations can be easily addedor removed [63]. However, due to the fluctuations of the wireless channels, highcapacity for backhauling the traffic to the core network cannot be guaranteed at alltimes. Severe link quality degradation caused by different factors can create bot-tlenecks and diminish network performance. Rain is a typical, relatively frequentlong-term event that can reduce the network throughput [2]. In the worst case thismay imply a complete failure of communication links bringing about further insta-bility problems of end-to-end paths. Routing is an effective approach for mitigatingthe impact of long-term events such as rain. In order to initiate the right action forrain mitigation it is important to distinguish between the short-term and long-termevent. In this regard, after reviewing the existing works in rain mitigation as well asrain detection, we provide a detailed system model and propose our rain detectionalgorithm. We evaluate the performance of our algorithm through simulations. Weapply the proposed algorithm to trigger the centralized rerouting process using thereal data measurements on real topology and examine the impact of detection erroron network performance.

4.1 Related Work

Different rain mitigation methods are studied in the literature. The traditionallink-layer adaptation such as adaptive modulation and coding (AMC) [19] maypartially compensate for the link quality degradation locally, but is not sufficient foraddressing long-term events such as rain. AMC adjusts the transmission rate at thephysical layer to maintain a certain quality of service level, which is usually definedin terms of the Bit-Error-Rate (BER) [2]. Site diversity as another approach for rainmitigation is studied in [47,50,64]. In site diversity solutions each node individually

39

40CHAPTER 4. ACCURATE RAIN DETECTION FOR IMPROVED

NETWORK PERFORMANCE

ensures finding an alternative, less affected link to replace a more degraded link.Some of the solutions provide site diversity by using different frequency band, forinstance [50,64], while others such as [47] choose a link with high angular separationcompared to the affected link. The work in [65] changes the network topologyutilizing adaptive antenna alignment to avoid the rain affected area. However,these approaches are applied locally (e.g., at a certain link) or pose extra cost tothe network.

Weather disruption-tolerant routing has been investigated as an effective ap-proach for mitigating the harmful effects of rain [22, 34]. A distributed routing al-gorithm is proposed in [34] to overcome the rain attenuation such that the quality ofservice for different users is satisfied while the end-to-end delay is minimized. Thework in [22] introduces two routing algorithms, i.e. the XL-OSPF (Cross-LayeredOpen Short Path First) and P-WARP (Predictive Weather Assisted Routing Pro-tocol) which modify the existing routing algorithm called open shortest path first(OSPF) by taking into account rain attenuation on each link and avoiding theseverely affected links. XL-OSPF improves the performance of OSPF by defininga cost metric that is proportional to the bit error rate (BER) of each link. InP-WARP, the external information of weather radar reflectivity data modeled inreal-time is used to compute the BER in advance which reduces the response delayof the algorithm. Most of these routing algorithms update the network in a dis-tributed manner without a centralized view and control of the network. From theperspective of network update, distributed updating schedules can only generate lo-cally optimal solutions [23]. Moreover, distributed solutions have slow convergenceand it is quite challenging to design a fast converging distributed algorithm.

Detecting the rain using a wireless mesh network topology has been studiedin several works [20, 21, 66]. In [20], the authors propose a rain detection methodbased on compressed sensing of rainfall attenuation. They use three links crossinga single area to improve the accuracy of rain detection. The work in [21] usestwo channels with different frequencies at each node to be able to identify a rainevent. Such approaches are limited due to their specific requirements, such asthree crossing microwave links, or two links with different frequency bands, whichmay be inapplicable for many practical scenarios. In order to distinguish the dryfrom the rain period, the authors in [66] measure the cross-correlation between theattenuation of two links on the same path, which is expected to be high in thepresence of rain. However, the approach requires sampling over a 15-minute timehorizon to reliably determine the presence of rain. Considering the high capacity ofbackhaul links, such a long delay can introduce huge traffic losses. Hence, there is aneed for accurate and fast rain detection algorithm that is applicable for triggeringthe rerouting process.


Application 3Application 2Application 1

Flow ManagerRouting Computation

Statistic Manager

Centralized Controller

Infrastructure Layer

Application Layer

Network Layer

Data Plane

API API API

OpenFlow

Figure 4.1: The overall network architecture.

4.2 System Model

We consider a wireless mesh network with an SDN-based centralized controller,shown in Fig. 4.1. The controller is connected through north bound interface tothe application layer, while the computed routes to satisfy requested demand aredispatched to the network element using SDN protocol such as OpenFlow via south-bound interface. The main components of our envisioned SDN controller are: 1)Statistics manager, 2) Routing computation, and 3) Flow manager. The statisticsmanager periodically gathers the information from the flow level and the physicallayer of the network. This information is fed to the routing computation componentwhere the optimal routes are computed based on the network conditions includingcapacity of each link and traffic demand of each source. Flow manager starts ap-plying the new optimal routes immediately after receiving the output of routingcomputation. The data plane for our network is assumed to be a wireless meshconnecting multiple base stations (BSs) where each BS gathers its own user dataor can be a relay for other BSs’ traffic. Each BS has a large number of antennaelements, making it feasible to create beams with very narrow beamwidth, knownas pencil-beams [7]. We assume that the connections between every pair of BSs arerealized with point-to-point line of sight communications with pencil-beams hav-ing negligible mutual interference among BSs [67]. We analyze our network in thepresence of rain which affects a part of the network as illustrated in an exampletopology of Fig. 4.2.


NETWORK PERFORMANCE

Core

Network

Fiber or leased line

backhaul traffic to the

core network

Figure 4.2: An example network topology.

4.2.1 Network Model

We consider a mesh topology consisting of multiple Base Stations (BSs) servingsubscribers in a given region and a number of gateways that transfer the aggregateddata to the core of the network. Throughout this chapter, we convert the continuoustime to discrete time samples with sampling frequency of 1/∆t. Let us assume ournetwork is affected by rain which lasts for N time samples from t0 to t0 +(N−1)∆t,where ∆t is the sampling period. We model the network as a graph denoted byH(V, E), where each vertex v ∈ V represents a BS or a gateway and each edge(u, v) ∈ E denotes a communication link between BSs (gateways) u and v. Wedenote the set of BSs and gateways with S and G, respectively. The traffic generatedat each BS is destined to a particular gateway, which transfers the traffic to thenetwork core. Suppose that there are K requested flows with demand {dnk}Kk=1 attime sample i, called commodities, which should be transferred to their destination.The flow corresponding to the k-th commodity passing through the edge of (u, v)at n-th time sample is represented by xnk (u, v). The total amount of flows passingthrough a link (u, v) is limited by the link capacity at time sample n denoted ascn (u, v). The goal is to choose the proper route for serving the requested demands,in order to maximize a predefined utility function defined as Φ, subject to thelink capacity and demand satisfaction constraints. The above problem can beformulated as follows:


maximizeXn

Φ (Xn)

subject to∑u∈V

xnk (j, u)−∑u∈V

xnk (u, j) =

0 otherwisednk if j ∈ S−dnk if j ∈ G

, ∀k, n , (4.1a)

∑k

xnk (u, v) ≤ cn (u, v) , ∀ (u, v) ∈ E , ∀ n = 1, . . . , N , (4.1b)

xnk ≥ 0, dnk ≥ dthk , ∀ n = 1, 2, .., N . . . , ∀ k = 1, 2, ..,K .(4.1c)

Constraint (4.1a) ensures the flow conservation at each node, Constraint (4.1b)specifies that the accumulated flow passing through one link is limited to the ca-pacity of that link, and Constraint (4.1c) models fairness among the sources byensuring a minimum amount of each request denoted as dthk is served. The utilityfunction in this section is assumed to maximize throughput of the whole system ateach time n. Consequently, the objective function can be defined as:

Φ (Xn) =∑k

dnk . (4.2)

where dnk represents the demand for k-th commodity at n-th time sample.

4.2.2 Channel ModelDepending on the weather conditions, i.e., clear or rainy sky, the channels betweenBSs can be affected by multipath or rain fading. Let P rc denote the received powerin dB for a link in clear sky, which is affected by multipath fading, following theNakagami distribution [68]:

f (P r,c) = 2mm

MΓ (m) P r,cexp

(2mP r,c

20 log(e) −m

P r,cexp 2P r,c

20 log(e)

), (4.3)

where m and P r,c represent the shape parameter (directionality) and the averagereceived power for that link, respectively. The average received power follows thepath loss model that depends on the length of the link, transmitted power, andcarrier frequency [69].

As we assume that our network may partially be affected by rain, some links,besides multipath fading, may also experience rain attenuation. Recall that ai isthe rain attention in dB on a link i, which follows a Lognormal distribution:

ai ∼ lnN (mi, σi) , (4.4)where mi and σi are mean and standard deviation of ln(ai), and depend on thelength of link i affected by rain and rainfall intensity [48]. This information can be


NETWORK PERFORMANCE

derived from long-term measurements during some years or estimated from Rec-ommendations ITU-R P.530 [49]. The rain attenuation shows both temporal andspatial correlation, Section 2.2.2 explains the detailed model used to generate thesecorrelation.

Another model for showing the spatial correlation between different links is thecorrelated area (CA), defined as a geographical area wherein the wireless channelgain has the same statistical distribution [51]. These CAs can be characterized byusing the spatial correlation factor of the rain, shown in equation (2.9). Each CAconsists of a set of links whose spatial correlation κij is greater than a predefinedthreshold. Accordingly, we can divide the entire network into several disjoint CAs.We define the CR+ and CR− as the set of CAs in which rain is present and sky isclear, respectively. In the following, we assume that these CA sets are given.

4.3 Rain Detection Algorithm

In this section, we propose an efficient rain detection algorithm that uses both timeand space correlation between rain attenuation samples to distinguish between theclear and rainy sky.

To detect the rain presence in each correlated area, we introduce a binary hy-pothesis Hj for each correlated area CAj , taking the value of 1 in the presence, and0 in the absence of the rain in that correlated area:

Hj ={

0 , if CAj ∈ CR− ,1 , if CAj ∈ CR+ .

(4.5)

Let pr,lj (i) denote the received power at the receiver of link l in CAj at timesample i. Based on each hypothesis, the received power at each link located in CAj

will be different as follows:

P rij (n)(dB) ={P r,cij (n) , if CAj ∈ CR− ,P r,cij (n)− aij (n) , if CAj ∈ CR+ .

(4.6)

where P r,cij (n) is the received power in clear sky for link i located in CAj at n-thtime sample shown in equation (4.3) and aij (n) shows the rain attenuation on linki located in CAj at time sample n modeled as equation (4.4).

In equation (4.5), the problem of rain detection is expressed as a binary testinghypothesis. The aim is to develop a decision rule (namely, rain detection algorithm)that maps the observations to one of those hypotheses, H0 or H1. In particular,each receiver periodically samples the received power in time intervals bigger thanthe coherence time of multipath fading at frequency 1/∆t. Since the coherencetime of multipath fading is much smaller than for the rain fading [70], the sampleswould be correlated in the presence of rain. Therefore, we can develop a simplerain detection algorithm that is executed at the receiver of each link independentlyto capture the correlation between power samples. Each receiver keeps M latest

4.3. RAIN DETECTION ALGORITHM 45

Record the received power of M non-overlapping samples

Add a bad sample in the sample set if the received power is below the

thershold

Current detection= Presence of the rain

Report the detection to the centralized controller.

Past detection=Current detection

Current detection= Clear sky Past detection= Current

detection

Centralized controller combine the links local decision located in

the same CA to make a global decision

The number of bad samples >

ρ M

Current detection=Past

detection?

NO

YES

YES NO

Figure 4.3: Flowchart for the proposed rain detection algorithm.

samples of the received power from each incoming link and if ρM , where ρ is a designparameter, samples are below a predefined power threshold T , called "bad sample",the presence of rain is declared since the temporal correlation between the sampleswas detected. We run our rain detection algorithm in each base station. Theyreport their local detection results to the CC, which then makes the global decisionon rain presence. Once each node makes the local detection it reports the decisionto the central controller (CC). The CC then identifies the spatial correlation of rainsamples based on the rain detection statuses of links belonging to the same CA.The "k out of n" method is used to correlate the link statuses obtained by localdetection within a CA and make the global decision. In this method, if at leastk of n links report the presence of rain, defined as voting links, the CC identifiesthe event as rain and triggers the rerouting. The flowchart for our proposed raindetection algorithm is shown in Fig. 4.3.

The rain detection process is subject to errors that may be caused by two mainfactors, i.e., false alarm and misdetection. False alarm probability, denoted as Pfa,refers to the likelihood of declaring the rain state while being in the no-rain state,whereas misdetection probability, denoted by Pmd, refers to being in the rain statebut declaring the no-rain state. The detailed computation of these two componentsis presented in Paper IV. The design parameters such as N and ρ can be chosento minimize both error probabilities.


NETWORK PERFORMANCE

4.4 Impact of Detection Error on Network Performance

Once rain is detected, the SDN controller keeps recalculating the optimal rout-ing by solving (4.1) and updating the network periodically throughout the rainduration. Since triggering the rerouting depends on the decision of the rain detec-tion algorithm, both misdetection and false alarm can affect network performance.Rain misdetection may lead to a throughput loss due to the failure to trigger anetwork-level reaction and reroute the traffic when it is required. In this case, theperformance would be the same as AMC. On the other hand, false alarms maypose unnecessary routing overhead on the centralized controller. The detectionerror in each CR− , which is equal to the false alarm probability, would introduceextra overhead on the centralized controller, so the total overhead is equal to:

λe =∑

j∈CR−

Pfa,j , (4.7)

where Pfa,j shows the false alarm probability of CAj .AMC is considered as an approach where just link adaptation is applied when

rain happens. Based on this, the throughput gain of our approach is defined asthroughput improvement compared to the AMC approach when the rain detectionis used to trigger the rerouting, and can be computed as follows:

TG = 1N∆t−M∆t

∫ N∆t

M∆t

1−∑

j∈CR+

Pmd,j

R(t)−RAMC(t) dt , (4.8)

where N∆t andM∆t denote the rain duration and time required for rain detection,respectively. Pmd,j represents the misdetection probability of CAj , More detailedinformation about this work can be found in Paper IV and Patent I of the thesis.


In this section, the performance of the proposed algorithm is evaluated by custom-built MATLAB simulator. The received power is sampled every ∆t = 10ms, whichis longer than multipath fading coherence time. The carrier frequency of 30 GHzis assumed according to the commonly used values in the microwave backhaulnetworks [71]. The mean and variance of rain attenuation are obtained accordingto [40].

We first evaluate the impact of different parameters on the efficiency of our raindetection algorithm via simulations. We consider both design-related parameterssuch as ρ, the number of samples M , and the number of voting links in a CA, aswell as environment-dependent parameters such as rain rate r. Secondly, by usingour rain detection algorithm to trigger the rerouting process, we investigate theimpact of the detection error on network throughput and the imposed overhead tothe controller.


Number of samples0 5 10 15 20 25 30 35

Pe=

1 2(P

md+Pfa)

10−3

10−2

10−1

100

ρ = 0.7, r = 20ρ = 1, r = 10ρ = 0.7, r = 10ρ = 1, r = 20

(a)

Number of voting links1 2 3 4 5 6

Pe=

1 2(P

md+Pfa)

10−5

10−4

10−3

10−2

10−1

M = 10M = 20

(b)

Figure 4.4: (a) The total error probability as a function of the number of samples.(b) The impact of the number of voting links in one CA to the total error probability.In both figures the probabilities of rain and clear sky are equal to 0.5. (Reprintedfrom [54] ©IEEE 2016, reused with permission.)

The total error probability is defined as Pe = z1Pfa+z2Pmd, where 0 ≤ z1, z2 ≤ 1are two constants representing the probability of clear or rainy sky, respectively.This information can be extracted from yearly measurement statistics. In this thesisit is assumed that the probability of rain and clear sky is equal, z1 = z2 = 0.5. Inour simulations, we consider a wireless mesh network where some links are affectedby rain while others have only multipath fading.

Fig. 4.4a shows the variation of the total error probability with time dependentparameter M for different rain rate r and design parameter ρ. Using a higher


NETWORK PERFORMANCE

10−3 10−2 10−1

Misdetection Probability

0

5

10

15

20

25

Throughputgain

%

(a)

10−4 10−3 10−2 10−1

False alarm probability

0

0.1

0.2

0.3

0.4

0.5

λe[1/s]

(b)

Figure 4.5: The impact of detection error on (b) throughput gain, and (b) overhead.(Reprinted from [54] ©IEEE 2016, reused with permission.)

number of samples (M), which implies longer detection time, improves the detectionaccuracy. In addition, the results show that by choosing ρ = 0.7 the detection errordecreases significantly, while ρ = 1 may cause a large error. This is because in ρ = 1,rain is announced when all samples are accounted as bad samples, however, due tomultipath fading variation in received signal the probability that all samples becomebelow certain threshold is very small. For instance with the same number of samplesM = 15, the error probability drops about 40% when rain rate is r = 20mm/hourjust by choosing a proper value for ρ.

Fig. 4.4b shows the behavior of the total error probability as a function of thenumber of voting links (k) that are considered in the global detection. The resultsshow that cooperative decision with an appropriate number of voting links improvesthe detection accuracy up to 95%. In our simulation, the best detection result isachieved when CC declares the presence of rain if at least 3 links out of 6 links in aCA detect rain. In this cooperative decision process, higher number of voting linksdecreases the likelihood of false alarm while increasing the misdetection probabil-ity. On the other hand, a lower number of voting links causes lower misdetectionand higher false alarm probability. Therefore, to minimize the summation of bothprobabilities with the same weight, e.g., 0.5, the best performance is achieved whenhalf of the total voting links are considered.

To validate the effect of the detection error on network performance, especiallythe system throughput, we consider a backhaul network with 7 BSs and one gate-way, partially affected by rain, as shown in Fig. 4.2. Our rain detection algorithmis used to trigger the rerouting.

Fig. 4.5a shows the throughput gain obtained by our approach, i.e., the through-put achieved by network-layer adaptation compared to the one achieved by onlyAMC, that verifies the importance of maintaining a low detection error. The resultsindicate the linear dependency of throughput gain and misdetection error as cap-tured by equation (4.8). Similarly, the overhead evaluation in Fig. 4.5b verifies that


00:00 03:00 06:00 09:00 12:00 15:00 18:00 21:00 00:00

Time [h]

4

4.3

4.6

4.9

5.2

5.5

5.8

6

Through

put[G

bps]

Regular reroutingAMCActive reroutingRain detection interval

Figure 4.6: Throughput comparison when different adaptation schemes are applied.

increasing the false alarm probability linearly increases the unnecessary rerouting.However, it is shown that a small error rate (e.g., 10−2) caused by our proposedrain detection algorithm is acceptable, still leading to a high throughput gain whileminimizing the network overhead.

We run our detection algorithm and rerouting procedure on a realistic microwavebackhaul network deployed in Sweden. The topology considered in the simulationconsists of 41 MBS and 5 gateways which are connected with mesh topology. Thereal data set includes the real received power measurements sampled in each 10sprovided by a network operator. To evaluate the throughput gain by using our raindetection algorithm, denoted as active rerouting, we consider two approaches thatprovide lower and upper bounds on the network throughput. The upper bound onthroughput is achieved when the network adaptation is triggered very frequently (inour simulation it is carried out every 5 min), denoted as regular rerouting. In sucha scheme, a short-term multipath fading is also considered as a cause for rerouting.The AMC approach with no rerouting provides a lower bound on the throughputas it does not support any rerouting.

Fig. 4.6 presents the throughput of the three considered schemes. The resultsverify that rerouting the traffic very frequently, even considering the multipathfading, does not provide throughput gain. However, re-optimizing the routes whenrain is detected is sufficient to achieve good network throughput, which is about 14%higher than throughput obtained by AMC approach in the worst rain attenuationcondition. With 41 base station deployed in our network, this gain means thateach MBS can support on average 20 = 800/41 Mbs more data rate using reroutingadaptation compared to AMC approach, which is a significant improvement as 20Mbp lower data rate may interrupt the service level agreement.

Chapter 5

Consistency-aware ReroutingFramework

The dynamic nature of network requires online adaptation in response to varyinglink capacities due to weather disturbances, which is not an easy task even underone central unit such as SDN controller. Reconfigurations in SDN are performedthrough updating the switches’ flow tables by the centralized controller, where eachnetwork element updates its flow table individually and without synchronizing withothers [25]. The asynchrony between different switches may lead to unwanted inter-mediate states and network inconsistency. The imposed inconsistencies create flowdisturbances or packet loss, which deteriorates network performance such as delayand network throughput. Therefore, frequent reconfigurations at predefined times,as considered in the previous chapter, carried out without considering their adverseimpact on the network may not provide the expected gain due to inconsistency. Inthis chapter, we present a model to compute the minimum imposed inconsistencyduring reconfiguration which is defined as the cost that must be paid for adaptingthe network to the new situation. Knowing the throughput gain and the cost ofrerouting, we propose a rerouting policy that determines the best time for applyingthe optimal computed routes to the network with the objective to minimize the dataloss during rain. Due to the temporal correlation of rain attenuation, the decisionat each time interval depends on the future network states. To predict the futurerain attenuation, our proposed method uses an unbiased estimator and considersthat estimation in the rerouting approach. The simulation results obtained on bothsynthetic and real data show that using our rerouting policy significantly decreasesthe data loss compared with applying regular rerouting at predefined times.

5.1 Related Work

Depending on the application using SDN architecture, various types of consistencyare defined in the literature including memory limit, packet drop freedom, loop free-

51

52 CHAPTER 5. CONSISTENCY-AWARE REROUTING FRAMEWORK

dom, packet coherency, and congestion-free consistency [25,26,32,37–39,72]. Eachof these consistency properties implies certain rules in applying newly computedroutes. To make sure that the transition is memory-limit consistent, the old rulesshould be deleted before new rules are updated in a flow table. The authors in [32]propose an update algorithm for loop freedom consistency which assures that noneof the packets traverse a loop in intermediate states. Packet-coherence consistencyis studied in [72] by proposing an algorithm that ensures packet- and flow-level con-sistency while minimizing the transition overhead. This type of consistency impliesthat each packet is routed either according to the new rule or the old one at allnodes, but not their combination. To do so, the authors in [72] tag a certain packetwith the old rules or the new rules. Many studies, including [25,26,37–39], proposevarious congestion-free update algorithms. In congestion-free update, one shouldensure that the amount of flows passing any link during transition satisfies the ca-pacity constraint. A simple approach for ensuring this is to avoid one shot updatesand sequentially move out the exiting flows before adding new ones to each link toavoid transient overflow. The work in [26] introduced SWAN algorithm to adaptto dynamically changing traffic demands with zero transition loss with a sequenceof intermediate states. In this method, to avoid congestion, new traffic distributionand tunnels for all flows are computed for each intermediate state. The zUpdatealgorithm proposed in [37] supports zero-loss update for data center network bycalculating the sequence of traffic matrices that should be applied to the switches’flow tables. To avoid congestion, both aforementioned approaches assume that allnetwork elements support flow splitting at any proportion and rely on creating in-termediate traffic distribution. Introducing the intermediate states during updatesmay complicate the update process or even disturb the users’ QoS due to larger de-lay experienced in intermediate states than in the initial and final states. Moreover,it is shown that these approaches require extra capacity on each link to guaranteethe feasibility of congestion-free update, which results in a great waste of capacity.Consequently, in networks with unsplittable flows and high link utilization, conges-tion cannot be avoided. In such scenarios, the only possible approach is to acceptthe congestion that will occur and try to minimize it.

The work in [39] tries to minimize both the update time and the imposed con-gestion just by manipulating the sequence of flow updates without creating anyintermediate paths or traffic matrices. In a similar way, the authors in [25] proposeDionysus algorithm to optimize the update time and calculate the fastest updatesequence. However, they consider dynamic switching delay for each OpenFlowswitch as they show that the switching time for each network element varies overtime depending on their load and type of update. In case when there is no feasiblesolution for congestion-free update, the algorithm limits the rate of some flows. Thework in [38] considers the trade-off between the imposed congestion and the updatedeadline, and designs a sequential update algorithm based on user-specific require-ments. However, the approaches for computing and minimizing congestion in theaforementioned studies cannot be applied to the wireless network where the initialcondition of the network becomes infeasible because of the rain degradation. In

5.2. PROBLEM DESCRIPTION 53

such degraded network, a delay in removing the affected flows as a response to linkdegradation produces data loss that should be taken into account when definingthe timing and sequence of network updates.

5.2 Problem Description

The rain attenuation fluctuations may make the solution applied in one time inter-val not optimal in the next one. Adapting the network in each time interval duringrain offers the best network performance in each period. However, unplanned in-termediate states during transition may increase the data loss. Therefore, frequentadaptations without considering the cost of rerouting may create huge data losssince it might happen that the loss during rerouting outweighs the rerouting gain.This inherent tradeoff between the gain and the cost of rerouting leads us to searchfor the most efficient policy for organizing the rerouting process. At the beginningof each time interval, upon calculating the alternative routes which would yieldmaximum throughput, two options are available: 1) let the network operate usingexisting, suboptimal paths during that time interval and accept the throughputloss, or 2) apply the new routes and accept the switching cost related to adaptingthe network to the new routing solution. Our objective is to find the best actionpolicy at each time so that the total data loss is minimized. In this section, wefirst compute the switching cost by minimizing the imposed congestion at each timesample using mixed integer linear programming (MILP). Then by knowing the costof rerouting, we formulate our problem to find the best sequence of decisions forminimizing the total data loss.

5.2.1 Switching Cost Minimization

The transient congestion incurred during network reconfiguration can be consideredas the cost of keeping the network operating as close to the optimal throughputas possible. By applying sequential updates, the amount of congestion duringtransition can be minimized, but it also creates delay. If the adaptation is performedin response to a failure such as a link degradation below a certain threshold becauseof rain, the extra delay causes data loss. In this situation, one shot flow update, i.e.,updating all the flows at once in non-sequential manner, may produce less transitionloss caused by asynchrony between different switches, compared to a data loss dueto the delay of sequentially updating the flows to their final state. Therefore, incase of a wireless network undergoing capacity degradation, two events contribute tothe data loss during reconfiguration: 1) the asynchrony between different switchesminimized by sequential, rather than one shot flow updates, and 2) the delay inupdating the affected flows that is minimized by removing them from the degradedlinks as soon as possible. In sequential updates, sometimes the data loss due to thelatter issue may exceed the former one, so the affected flows should be moved tothe final state even if their destination links are still occupied with the old flows.


B1

B5

B4

B3

B2B1

B5

B4

B3

B2F3:5

B1

B5

B3

B2F3:5

B1

B5

B4

B3

B2 B1

B5

B4

B3

B2

1) Initial state 5) Final state

B4

(2)

Loss: 8 units (B1-B4)

(4)


(3)


(a)

B1

B5B4

B3

B2F3:5

B1

B5

B4

B3

B2F3:5

B1

B5

B3

B2

(2)


(3)


B4

(4)


(b)

Figure 5.1: A network is updated from initial state (1) to final state (5) through4 steps. (a) shows the flow migration according to policy [F4 → F3 → F2 → F1],and (b) shows the flow migration considering the initial network state, namely[F2 → F4 → F3 → F1]. (Reprinted from [73] ©IEEE 2018, reused with permission.)

In the following, we first provide an illustrative example to clarify the causes ofimposed congestion and then we compute the minimum imposed congestion due tothese factors by formulating the problem as a MILP problem.

The impact of the flow update order on the total loss is shown in an illustrativeexample in Fig. 5.1. Each edge has the capacity of 10 units except (B1, B4) withthe capacity of 13 units. There are 4 flows in the network, denoted as F1, F2,F3, and F4, with initial routes depicted in the figure. Once the network is affectedby rain the capacity of link (B1, B4) decreases to 5 units. To minimize the dataloss due to rain, the traffic flows are updated from the initial (step 1) to the finalstate (step 5) through 4 consecutive steps. One possible sequential update thattries to minimize congestion but does not consider the initial state of the networkand rain degradation is [F4 → F3 → F2 → F1]. Using this plan, we can minimizethe transient congestion due to switch asynchrony, but 8 + 3 + 3 = 14 units of dataare lost due to the delay in response to the rain degradation. However, taking the

5.2. PROBLEM DESCRIPTION 55

initial condition of the network into account and trying to minimize the total lossyields solution [F2 → F4 → F3 → F1], where F2 is evacuated as soon as possibleand the resulting data loss is equal to 3 + 3 + 3 = 9 units.

Let us assume that Xn represents the set of all flows that exist in the networkat the beginning of the current time sample n, and Xno is the set of optimal flowroutes in the present time interval, obtained by solving the optimization problemfrom equation (4.1). In this section, we formulate an optimization problem tofind the best sequence for updating flows such that the total transition loss isminimized and compute the minimum transient congestion imposed to the networkat n-th time sample if the network state changes from Xn to the new routes Xno .The optimization problem provides a solution that considers both the data lossgenerated because of asynchrony between switches and the delay in updating theaffected flows. The flows are assumed to be unsplittable during transition. Letxk (u, v) ∈ Xn denote the amount of flow k that passes through link (u, v) inthe current state, similarly x′k (u, v) ∈ Xno is the updated flow for flow k on link(u, v), computed according to the new network condition. We assume that the linkcapacity does not change during transition since the flow update time, denoted asα, is typically by an order of magnitude shorter than the sampling time. Let Udefine the set of all flows affected by the link capacity deterioration. It should benoted that if the updated solution just limits the rate of a flow but does not changeits route, this flow is not denoted as updated and is not included in U . Let usdivide the total flow update time α into R rounds. In each round, defined as s,one or more flows migrate to their final states. We define a binary matrix Y as amigration matrix, where element yk,s = 1 if flow k is updated to its final state atround s. As no intermediate state is allowed during the update, each flow shouldbe transferred to its final state once, which poses the following constraint on theoptimization problem:

R∑s=1

yk,s = 1 . (5.1)

At round s, each link (u, v) may be carrying two types of flows because of asyn-chrony between different switches. The first type is background traffic, comprisingflows that are not transferred to their final destination before time s, and the secondtype is the new incoming flow that is transferred at round s. The correspondingconstraint is formulated as follows:

fk (u, v) +∑k

yk,sIk (u, v) ≤ c (u, v) , (5.2)

where fk (u, v) is the background traffic on link (u, v) in round s, and can becalculated as:

K∑k=1

xk (u, v)s−1∑l=1

yk,l +K∑k=1

x′k (u, v)(

1−s−1∑l=1

yk,l

)(5.3)


and Ik (u, v) represents the amount of added flow k at link (u, v), given by:

Ik (u, v) = max {0, x′k (u, v)− xk (u, v)} (5.4)

To compute the amount of data lost during transition, we define bs (u, v) as avariable that models the amount of link capacity required to satisfy constraint (5.2).The sum of the required capacity bs (u, v) over all links translates to the total dataloss. Therefore, the objective is to find the best migration matrix such that the totalamount of required capacity for all links during R rounds is minimized. Increasingthe total number of rounds R may allow for a decrease in the transition loss dueto asynchrony, while it may also increase the transition loss because of the delay inupdating the affected flows. In this thesis, we assume a fixed value of R such thatthe update time (R) is low, however it can be optimized to obtain less transitionloss. The following optimization provides the best migration matrix in terms oftransition loss:

minimizeY,B

R∑s=1

∑(u,v)

bs (u, v)

subject toR∑s=1

yk,s = 1, ∀k,

(5.5a)

fk (u, v)+∑k

yk,sIk (u, v) ≤ c (u, v)+bs (u, v) , ∀ (u, v) , s

(5.5b)yk,s ∈ {0, 1} , bs (u, v) ≥ 0, ∀k, s, (u, v) (5.5c)

Constraint (5.5a) ensures that each flow is transferred to its final state just once,and Constraint (5.5b) limits the total amount of flows on each link to its capacity.

5.2.2 Total Data Loss MinimizationAfter modeling the switching cost and determining the optimal update sequencethat minimizes the imposed congestion during transitions, we proceed to computethe overall gain of rerouting and formulating the total data loss.

Recall that Xno contains all optimal traffic flows in time sample n , and Xn is theset of all traffic flows that already exist in the network, i.e., the state of the networkat time sample n. If the network is not updated to the optimal solution Xno , thedata loss rate due to solution sub-optimality over n-th time interval [n∆t, (n+ 1)∆t]will amount to TG (n). Therefore, TG (n) expresses the gap between the optimalthroughput that would be obtained by applying the optimal routing solution Xno ,and the network throughput obtained by keeping the existing flow state Xn. Hence,the optimal gap can be written as:

TG (n) = T (Xno , Cn)− T (Xn, Cn) (5.6)

5.3. ONLINE ALGORITHM FOR EFFICIENT REROUTING 57

where T (Xn, Cn) and T (Xno , Cn) represent the network throughput obtained byapplying routes Xn and Xno , respectively, for link capacities Cn.

Knowing the gain and the cost of rerouting at each time interval allows us toformally define the problem of finding the optimal rerouting policy. The policy,denoted as π, comprises a sequence of rerouting decisions at each time interval i,denoted as un. It is a binary variable whose value of 1 denotes the decision toreroute, while 0 means no rerouting takes place.

Assuming that rain will last for N time samples, we define π = {u0, un, ..., uN}to be rerouting policy, i.e., a sequence of rerouting decisions in each of the N timesamples. The objective of our approach is to minimize the total data loss duringadaptation to the rain, formulated as follows:

Jπ =N∑i=0

TG (n) ∆t (1− un) + SC (n)αun (5.7)

Depending on the decision at each time interval, the total data loss will bedifferent. Choosing un = 0 means that the existing routes remain applied to thenetwork in the next time interval. The data loss in this case arises only due toapplying suboptimal solution and is equal to the throughput gap TG (n) ∆t duringn-th time interval [n∆t, (n+ 1)∆t]. Setting un to 1 will change the network stateto the optimal one at the price of the switching cost, which is equal to SC (n) duringflow update time α computed by solving the optimization problem (5.5).

An optimal policy π∗ is the one that minimizes the total cost over all combina-tions of π, given by:

Jπ∗ = minπ

Jπ (5.8)

The problem formulated in (5.8) can be considered as a dynamic programmingproblem since, firstly, its state transition in the n-th time sample depends on thecontrol variable un which must be chosen from a finite set {0, 1} and, secondly, thecost accumulates over time and its total, final value depends on the visited states,namely TG (n) and SC (n), and the control policy un at each time interval n [74].In the next section, we present the online solution for solving our optimizationproblem.

5.3 Online Algorithm for Efficient Rerouting

While solving the proposed DP problem, we must consider two key points. Firstly,the policy at time sample n depends on the future steps and requires knowledge offuture states. This challenge is solved by using prediction of the rain attenuation forfuture time samples. We propose a naive estimator with low complexity that lever-ages on the Lognormal distribution of rain attenuation and provides a predictionof the future states. The second point is related to the choice of the time horizonthat we should consider for our optimization, denoted as w. The time horizon w


depends on many factors and it will affect the complexity and performance of ouralgorithm. In this section after describing the procedure for predicting the rainattenuation, we provide some guidelines for choosing an appropriate value for thetime horizon by considering the trade-offs between the complexity and performanceof our algorithm. Finally, we describe the details of the proposed online algorithmfor minimizing the total data loss during rain.

5.3.1 Rain Attenuation PredictionThe temporal model of rain attenuation explained in Chapter 2 shows that rainattenuation, denoted by a, behaves as a first order Markov process. Hence, knowingthe attenuation at time sample n provides sufficient information to generate itsstatistics in the upcoming time samples. The probability of rain attenuation infuture time samples based on the attenuation at time sample n is modeled with thefollowing distribution, as described in [40]:

fa (n∆t|a (t) = a0) = lnN (p1 (a0) , p2, α0) (5.9)

wherep1 (a0) = m1−exp(−β|n∆t|).a

exp(−β|n∆t|)0

p2 = σ2 (1− exp (−2β |n∆t|))

Parameters m and σ are mean and standard deviation of ln a, respectively. Thetime dependence is described by β.

A naive and simple estimator of a random variable with known probabilitydistribution is:

a (n∆t) = E {fa (n∆t)} (5.10)

where E [.] shows the expectation operator applied to rain attenuation. To evaluatethe performance of our estimator, we define the prediction error as:

Pne = Pr {|a (n∆t)− a (n∆t)| ≥ ε} (5.11)

where Pne denotes the probability of having the absolute error higher than prede-fined threshold, denoted as ε, at n-th time samples. It is shown in Paper V thatthe Pne increases exponentially with n.

5.3.2 Finite Time Horizon DefinitionTo choose an appropriate value for the time horizon w, we should consider thefollowing points:

• As the model for rain attenuation does not model the presence and absence ofrain, the time for applying our DP algorithm cannot exceed the rain durationN , w ≤ N .

5.3. ONLINE ALGORITHM FOR EFFICIENT REROUTING 59

• As proved in Paper V, the estimation error increases exponentially with n.In the interval [n, n+ w], the prediction of the w-th time sample produces thehighest error. Therefore, we should limit w so that the estimation error forthe w-th time sample is lower than a predefined threshold. More explanationcan be found in Paper V.

• The complexity of finding the solution for the total loss minimization problemin (5.7) grows exponentially with w, as shown in Paper V.

Considering the explanations above, there is a trade-off between the optimalityof the solution, computational complexity, and prediction error that should be notedwhen selecting a proper value for w.

5.3.3 The Optimal Control Action with Prediction (OCAP)Algorithm

By choosing a proper value for w our problem is translated to a short term plan-ning problem. A wide variety of algorithms have been proposed for online convexoptimization problems with a finite time horizon. We use a well-known process con-trol method called Model Predictive Control (MPC), also referred to as RecedingHorizon Control (RHC), to solve our problem. MPC relies on dynamic modelingof processes which allows for the optimization of the current time slot taking thefuture time slots into account as well. The ability of MPC to consider the anticipa-tion of future events and take control actions accordingly makes it very applicablefor solving our problem. In the rest of this section, we apply the MPC algorithmto solve our optimization problem.

The rerouting control action at each time is determined by the proposed al-gorithm which we refer to as Optimal Control Action with Prediction (OCAP),summarized in Algorithm 2. The set of existing applied routes which correspondsto the initial state of the network as well as the time horizon w are given. To initial-ize our algorithm, we set n = 0 and compute the throughput gap TG and switchingcost SC at time 0. At time sample n, the algorithm has the exact informationabout the current state of the network and based on that, calculates TG (n) andSC (n). Let us suppose that Q(Xn) is an estimator that will provide the informationabout future states during [n+ 1, n+ w] based on current attenuation informationfor each link using equation (5.10). Applying this estimator, the algorithm has allthe required inputs to solve the problem given in equation (5.8) by using the MPCalgorithm for each of the w intervals starting from current time n. By knowing theexact value of TG and SC in the current time and predicting the future state ofthe network, the stochastic optimization problem shown in equation (5.8) is trans-formed into a deterministic optimization problem, which can be solved relativelyeasy. Due to a finite number of states during w time frames, equation (5.8) can berepresented by a transition tree where each arc is labeled with the cost of transitionfrom the previous to the next state depending on the decision policy. By introduc-ing one virtual end state, the shortest path algorithm can be used to calculate the


Algorithm 2 OCAP algorithm for computing optimal rerouting policy during rain

1: Given: Time horizon w, network H, initial routing X 0, rain duration N , link capac-ities C0.

2: Initialization: Set time sample to 0 and measure the rain attenuation for each link.Based on that compute the initial values of TG (0) and SC (0).

3: while n+ w ≤ N do4: Predict the rain attenuation for each link during interval [n+ 1, n+ w].5: Compute TG (k) and SC (k) for each time sample during interval [n+ 1, n+ w].6: Find the control sequence {un, un+1, ..un+w} that solves the deterministic problem

from equation (5.7) for interval [n, n+ w], substituting TG (k) and SC (k).7: Update un with calculated policy just for the n-th time sample.8: if un = 1 then9: Applied the new computed route in Xno .

10: Update the network state Xn+1 = Xno .11: end if12: n← n+ 1.13: end while

path with the lowest cost from the initial state to the final state [74]. The optimalsolution to the optimization problem provides the set of optimal rerouting controlactions which minimize the total data loss. We apply the control action for eachinterval n individually and repeat the algorithm for the next time slot. In PaperV we compute the complexity of our proposed algorithm as a function of the rainduration N , the number of base stations V and the length of the time horizon w. Itis shown that the computational complexity of our proposed algorithm is equal toO(Nw2w +NV 2 + 2NwV 3) which increases exponentially with the time horizon

w. For a small fixed value of w, the complexity increases with the cube of networksize V .


In this section, we evaluate the performance of our proposed optimal policy byapplying it to a synthetic and to a city-wide topology deployed in Sweden. MAT-LAB is used as simulation platform to generate all the results. In the first partof our evaluation, the results are generated for one instance of rainfall based onthe time series model explained in Chapter 2. To generalize the evaluation, we runthe algorithm for 200 different randomly generated rain instances and obtain theaverage performance. Finally, we use the real topology and real data measurementsto verify the applicability of our algorithm in realistic scenarios. For benchmarkingpurposes, we implement 3 rerouting policies and compare them to our proposedOCAP approach:

1. Regular action (RA): it reconfigures the network upon every change of thechannel conditions in each time frame ∆t;



Parameter Description Value∆t Sampling time 10 sα Flow update time 2 sN Rain duration 50 minw Time horizon 30 sr Rain rate 20 mm/hourσr Rain attenuation standard deviation 1.14 dBβ Temporal variation of rain attenuation 1/9 min−1

2. Greedy action (GA): rerouting depends only on the current network con-dition without considering the future states, namely w = 0;

3. Optimal Control Action with perfect knowledge (OCA): a modifica-tion of our rerouting policy assuming that the perfect knowledge of all futurestates of the network is available. Note that the OCA with large enoughtime period (i.e.,w = N) policy provides an upper bound on the achievableperformance.

5.4.1 Synthetic NetworkThe topology and the network model assumed in this section is the same as the onein Chapter 2. To recall briefly, we consider a wireless mesh network that consist of 6BSs, where 6 traffic flows are generated from each BS and directed to one of the corenetwork gateways. The objective is to provide the routes for each flow such that thethroughput is maximized. We generate the rain samples based on the time seriesmodel explained in Chapter 2, where the sampling interval is ∆t = 10. The short-duration rainfall (the duration less that 1 hour [75]) which occurs more frequentlyin summer is considered for our simulation, however, the proposed algorithm isapplicable for arbitrary rain duration. The assumed simulation parameters for themean and standard deviation of rain attenuation as well as the flow update time αare reported in Table 5.1.

Fig. 5.2 shows a comparison of the total data loss during rain duration obtainedby our online OCAP and the three benchmarking rerouting policies. The resultsindicate that, when switching cost is considered, applying the RA without tak-ing into account the network conditions increases the transition data loss, whichhighlights the non-negligible impact of the switching cost. The proposed OCAPpolicy minimizes the transition data loss during network adaptation to the changedconditions, yielding 79% lower total loss than RA. Moreover, the OCAP algorithmapproaches the lower bound on the loss obtained by OCA with less than 8% devi-ation, indicating its robustness to the error in estimating the future states of the


0 200 400 600 800 1000 1200 1400 1600 1800 2000

Time [s]

0

2

4

6

8

10

12

Data

loss

[Gb]

GA

OCA

RA

OCAP

Figure 5.2: The total data loss obtained by the OCAP, OCA, RA and GA reroutingpolicies. (Reprinted from [73] ©IEEE 2018, reused with permission.)

0 200 400 600 800 1000 1200 1400 1600 1800 2000

Time [s]

0

1

2

3

4

5

6

7

8

Data

loss

[Gb]

RA, α=0.5OCAP, α=0.5RA, α=1.5OCAP, α=1.5RA, α=2.5OCAP, α=2.5

Figure 5.3: Data loss variation for different flow update time α obtained by theproposed OCAP policy and the RA policy. (Reprinted from [73] ©IEEE 2018,reused with permission.)

network. Another interesting conclusion from this figure is the poor performanceof GA compared to the online algorithm with prediction. This indeed implies theimportance of taking rerouting decisions based on future network conditions, evenin the presence of prediction error.

The parameters that affect the performance can be divided into two categoriesincluding environment-related factors such as rain rate r and flow update time α,and design-related parameters such as time horizon w. To analyze their impacton the algorithm performance, we run a set of simulations by varying one of theparameters while fixing the others to the same values as in Table 5.1.

Fig. 5.3 shows the impact of the flow update time α on the performance of


0 200 400 600 800 1000

Time [s]

0

1

2

3

4

5

Data

loss

[Gb]

OCAPOCAGARA

Figure 5.4: The average data loss obtained by OCAP, OCA, RA and GA.(Reprintedfrom [73] ©IEEE 2018, reused with permission.)

our algorithm. The results show that our OCAP has a superior performance, andthe performance gain improves with the increase of the flow update time valueswhich correspond to higher switching costs. Furthermore, increasing the switchingcost increases data loss in RA, whereas our OCAP is substantially less sensitiveto dramatically longer the flow update times. This highlights the applicability ofour algorithm to support guaranteed QoS levels for networks with varying flowupdate time. More results regarding the sensitivity analysis for synthetic data arepresented in Paper V.

To evaluate the average performance of our algorithm, we apply it to 200 ran-domly generated rainfall series and record the average data loss. Fig. 5.4 depictsthe average data loss obtained by all four rerouting policies. The results clearlygeneralize those of Fig. 5.2 and indicate the superior performance of our proposedalgorithm. On average, OCAP reduces the total data loss by 67% compared toRA. Moreover, it obtains only 9% greater data loss compared to the OCA approachwhich uses perfect knowledge about the future rain attenuation.

Fig. 5.5a illustrates the effect of rain rate on the average total data loss. Asexpected, heavier rain causes greater data loss in the beginning, however it reachesthe steady state sooner than for the medium and light rain rate (similarly to [76],we consider r ≥ 20 as medium rain and r ≥ 40 as heavy rain). In the steadystate of the network no further improvement is achieved by further rerouting. Theresults also indicate that the total data loss in the steady state of the network isapproximately the same for medium and heavy rain. This is because after a certainlevel of rain attenuation, the link capacity is heavily deteriorated and many linkscannot be used for rerouting. Even though the transition loss for medium andheavy rain seems to converge to the same value, the total network throughput forhigher rate rain decreases significantly, as shown in Fig. 5.5b. From Fig. 5.5b it canbe inferred applying our proposed algorithm OCAP under heavy rain rate provides


0 100 200 300 400 500 600 700 800 900 1000

Time [s]

1.5

2

2.5

3

3.5

Data

loss

[Gb]

OCAP, r = 10RA, r = 10OCAP, r = 30RA, r = 30OCAP, r = 50RA, r = 50

(a)

10 30 50

Rain rate [mm/h]

0

5

10

15

20

25

30

Totalthroughput[G

bps]

OCAPAMC

(b)

Figure 5.5: The impact of rain intensity on (a) average data loss and (b) totalnetwork throughput. (Reprinted from [73] ©IEEE 2018, reused with permission.)

-5 0 5 10 15 20

x [km]

-45

-40

-35

-30

-25

-20

-15

-10

y[km]

Figure 5.6: A part of a microwave backhaul topology deployed in Sweden. Thebase stations are shown with empty circles and filled black circles are the possiblegateways. (Reprinted from [73] ©IEEE 2018, reused with permission.)

more gain. The sensitivity analysis for the rest of the parameters is reported inPaper V.

5.4.2 Real Network

In order to verify the performance of our algorithm in realistic scenarios, we applyit on an actual deployed network with real data measurements. The topologyconsidered for this analysis is a part of the tree-like backhaul topology deployedin a city in Sweden, modified by adding links to turn it into a mesh and make itsuitable for applying our algorithm. The mentioned topology is shown in Fig. 5.6comprising multiple BSs considered as traffic sources and several gateways that can


0 200 400 600 800 1000 1200 1400 1600 1800 2000

Time [s]

0

30

60

90

120

150

Data

loss

[Gb]

RAGAOCAOCAP

Figure 5.7: Total data loss over time obtained by OCAP, RA, and GA. (Reprintedfrom [73] ©IEEE 2018, reused with permission.)

0

rerouting

0

2

4

6

8

10

12

14

16

18

20

Number

ofreconfiguration

RA

OCAP

OCA

(a)

100 200 300 400 500 600 700 800 900

Time [s]

0

5

10

15

20

25

30

Throughputgain

[Gbps]

OCAPRAOCA

(b)

Figure 5.8: (a) Number of reconfigurations and (b) network throughput gain fordifferent policies. (Reprinted from [73] ©IEEE 2018, reused with permission.)

be used as a destination by different BSs. The data set contains the received powerfor each BS sampled every 10s.

Fig. 5.7 shows the total data loss obtained by applying the OCAP, OCA, RAand GA approaches to the network inFig. 5.6. The results indicate significantimprovements obtained by OCAP compared to GA and RA. For instance, rainprediction for only 3 future samples decreases the transition data loss by 64%compared to GA. The result for GA policy shows that just looking at present timemay reduce the data loss at current time, but causes the same data loss as RA afterhalf an hour (2000s). Comparing the results of OCAP to OCA clarifies that evenwith the presence of prediction error, our algorithm performs closely (e.g., with in9% ) to the OCA policy.

Fig. 5.8a shows the total number of reconfigurations during 1000s and Fig. 5.8b


0 200 400 600 800 1000 1200

Time [s]

0

20

40

60

80

100

120

140

Data

loss

[Gb]

OCA, w = 1

OCAP, w = 1

OCA, w = 3

OCAP, w = 3

OCA, w = 5

OCAP, w = 5

Figure 5.9: The impact of different time horizon for our OCAP and OCA.(Reprinted from [73] ©IEEE 2018, reused with permission.)

represents the total throughput gain compared to adaptive modulation and cod-ing (AMC) approach for OCA, OCAP, and RA. The results indicate that OCAPachieves higher throughput gain with 2 times less reconfigurations than RA (shownin 5.8a) by wisely picking the time for triggering the rerouting process. The num-ber of reconfigurations for OCAP approach is equal to 8 which is higher than OCAdue to prediction error, but still is half of the reconfiguration in RA decision policy.Regular action reconfigures the network 19 times without achieving any gain dueto ignoring the switching overhead. The results in Fig. 5.8b show that RA achievesthe poorest performance, obtaining the lowest throughput gain due to data losscaused by careless reconfigurations. OCA performs the best and is closely followedby OCAP that eventually accumulates slightly greater loss, which is in line withits higher number of reconfigurations due to prediction error.

Fig. 5.9 analyzes the sensitivity of the total data loss to the time horizon dura-tion. The results show that for both OCA and OCAP algorithms, increasing timehorizon w decreases the data loss. However, the difference in the total data lossobtained for different time horizons w diminishes at higher values of w. Chang-ing w = 1 to w = 3 reduces the OCAP losses by 61%, while further increasingw = 3 to w = 5 results in much lower performance improvement (only 13%). Asthe complexity of our proposed algorithm increases exponentially with w, the gainachieved for higher values of w can be neglected after a certain value of w. Thisshows that our approach already performs very well for relatively low values of w,where it obtains a beneficial trade-off between the computational complexity andperformance. Another interesting conclusion that can be drawn from these resultsis that the OCA algorithm is more sensitive to the time horizon w compared toOCAP which uses prediction for future samples. For instance, increasing w from3 to 5 in case of OCAP reduces the amount of data loss by 13%, while in OCAapproach the reduction gain is 33%. The main reason is that larger time horizons


are associated with greater prediction errors. Our OCAP algorithm performs betterfor higher w but worse for less accurate predictions of the future network states.These two effects cancel each other, and consequently higher w may not providesimilar gain for OCAP compared to the OCA approach, but exponentially increasesthe complexity of the rerouting algorithm in both cases. Therefore, an appropriatevalue for w in case of prediction-based OCAP must be less than the one chosen inOCA algorithm due to estimation error.

To conclude, the evaluation results indicate that OCAP algorithm is a promis-ing solution that minimizes the total data loss during network adaptation by wiselypicking the time for applying the computed routes. Moreover, OCAP algorithm notonly minimizes the number of reconfigurations, but increases the network through-put.

Chapter 6

A Techno-economic Framework for5G Backhaul Networks

While the work presented in previous sections focuses on optimization of the mi-crowave backhaul segment in the presence of rain, the decision on the backhaulingtechnology in terms of profitability and the overall cost optimization requires adetailed techno-economic analysis. Knowing that already the cost of the backhaulsegment is an unnegligible part of the total cost of ownership (TCO) in homoge-neous wireless networks [77], it can be expected that, with an increasing numberof small cells (i.e., as in a HetNet deployment) the impact of the backhaul segmenton the TCO of radio access networks (RANs) will become even more critical [78].Hence, in order to assess the TCO for mobile network both the access and the back-haul segment should be considered. While many studies focus on providing costefficient backhaul solutions for HetNets, e.g., [13, 41–43], none of them provides acomplete and general framework for TCO estimation that can be applied for bothheterogeneous and homogenous networks with both fiber and wireless technology.Apart from TCO, it is also of paramount importance for a network operator tounderstand whether a certain mobile network deployment plan is profitable or not.Pure TCO calculation of a given mobile deployment solution is not sufficient tounderstand its profitability, which depends on many other factors, such as initialinvestment, user penetration, revenues during network operating phase, competi-tors, and regulations. Moreover, a dynamic analysis, which can take into accounthow these parameters vary over time, is vital for the economic viability assessment.This is because both the yearly cash flow and the net present value (NPV) (i.e., twokey parameters in assessing business viability) are time-dependent [79]. Our workaddresses this issue by providing a complete techno-economic framework which in-cludes both the TCO calculation and an economic viability analysis for any typeof mobile access network deployment as well as various backhaul technologies. Theframework includes: 1) a detailed breakdown of the capital expenditures (CAPEX)and operational expenditures (OPEX), which are parts of TCO, for fiber and wire-

69

70CHAPTER 6. A TECHNO-ECONOMIC FRAMEWORK FOR 5G BACKHAUL

NETWORKS

less backhaul technologies, both of which are expected to provide high capacity forfuture mobile networks [80], and 2) the NPV analysis for profitability estimationbased on the yearly cost and cash flow calculation. We carry out a case study withdifferent scenarios focusing on both fiber and microwave technology assuming homo-geneous and heterogeneous access deployment. The obtained results indicate thatfiber may be the most cost-efficient technology to provide a high-capacity backhaulsolution for heterogeneous network deployments in high density area. Moreover,the NPV results show that a lower TCO does not always lead to higher profits.

6.1 Related Work

A large number of studies focus on the cost evaluation of different backhaul tech-nologies, such as [13,41–43]. In [13], different microwave-based backhaul topologies(including mesh and tree) are compared with respect to their total cost. The au-thors consider only the cost related to CAPEX and reach the conclusion that meshtopology is the most cost efficient option under different requirement of traffic de-mand and reliability. The authors in [43] design a cost efficient passive opticalnetwork (PON) architecture to backhaul the small cell traffic and compare it withpoint-to-point fiber deployment. The utilized cost model considers both CAPEXand OPEX, but focuses on the fiber-based backhaul solutions and is not generalenough to be applied to other backhaul technologies, such as microwave solutions.The work presented in [42] compares the deployment cost of backhauling in a longterm evolution (LTE) homogeneous wireless network considering various backhaultechnologies. However, this study was done only for homogeneous wireless deploy-ment, and the impact of small cells has not been evaluated. In [41], the authorsinvestigate various wireless architectures (both homogenous and heterogeneous) andtry to assess the impact of backhaul on the total TCO. Unfortunately, the modelin [41] leaves many important details on cost calculation unspecified, making theresults hard to validate. None of the aforementioned references propose a completeand general techno-economic framework that is able to evaluate different backhaultechnologies and their impact on the TCO of various mobile network deployments,specifically considering the HetNet scenarios.

To assess the economic viability, many existing studies such as [53-55] considera dynamic analysis of cash flow and NPV. The work in [81] focuses on techno-economic aspects of different deployment strategies including macrocell, microcell,and small cell in dense suburban environment. It compares the capacity, coverage,cost and profitability of these deployment strategies from the indoor service providerperspective. According to their results, indoor small cell solution has the higherperformance in terms of cost and profitability. The authors in [82] try to financiallymanage the cash flow to gradually deploy the 4G on top of the existing 3G servicesso that the final profit is maximized. The results in [82] indicate that the operatormay not deploy full 4G coverage in areas with low user density or high deploymentcost. The NPV analyses of the mentioned studies are focused only on the RAN

6.2. TECHNO-ECONOMIC FRAMEWORK 71

Techno-economic Module

Total Cost of Ownership Module

Network Dimensioning Module

Mar

ket M

odul

e

A

rchi

tect

ure

Mod

ule

Topo

logy

Mod

ule

Bus

ines

s Mod

ule

Cos

t Mod

ule

Figure 6.1: Proposed techno-economic framework. (Reprinted from [84] ©IEEE2018, reused with permission.)

segment and do not consider the overall mobile network (i.e., backhaul + RAN). Thework in [83] analyzes the total cost and NPV for some specific scenarios dedicated tosparsely-populated areas. Consequently, the conclusions therein lack the generality,creating the need for the development of a comprehensive general model.

6.2 Techno-economic Framework

Fig. 6.1 shows the schematic view of our proposed framework which consists ofseveral modules, defined as follows.

• Architecture Module: This module defines the technology used in backhaulsegments and the types of components that should be installed to support aspecific architecture.

• Topology Module: This module defines the physical topology of the net-work which can be mesh, tree or ring. The module requires demographicaland geographical information of the region such as the region size, the numberof buildings, and user density as input. Based on this, it creates the topol-ogy, the number of necessary nodes (e.g, cabinets, central offices), and thedistance between different nodes, which can be used as input for the networkdimensioning module.

• Market Module: This module is responsible for estimating the possible rev-enues and the number of users that are expected to subscribe or unsubscribe.


NETWORKS

The input that is required for this module is user penetration rate, operatormarket share, user behavior, prices of the services and user churn rate.

• Network Dimensioning Module: Using the data received from the archi-tecture, topology, and market modules, this module calculates the amount ofnew infrastructure and components deployed in each year. This module alsocomputes labor activity related parameters such as traveling time to a certainsite.

• Cost Module: The time dependence of each cost parameter in TCO esti-mation is captured by this module. For instance, the costs related to humanresources such as salary will increase over time, while the equipment cost willtypically decrease. To model the cost variation, we use the following linearformula [79]:

Pi = P0 + τPi−1 , (6.1)

where Pi and P0 are the prices in year i and the initial year, respectively. Pa-rameter τ expresses the cost change factor, which can be negative or positive.Typically, τ has a negative value when the price variation of hardware compo-nents is considered, while it has a positive value when salaries or energy costare calculated. In general, τ might also vary in time. However, for simplicityτ is often assumed to be constant during the whole network operational time.

• Total Cost of Ownership: This module provides the detailed breakdownfor calculating the CAPEX and the OPEX of the backhaul segment. TheCAPEX covers all expenses related to installing the backhaul network ele-ments in place. The OPEX encompasses the expenses during network opera-tion. As shown in Fig. 6.2, both the CAPEX and the OPEX consist of manysubcomponents, whose detailed explanation is presented in Paper VI.

• Business Module and Scenarios: Network provider (NP), service provider(SP), and mobile network operator (MNO) are accounted as different actorsand this module defines the business relations between them such as thecooperation models between them and various governmental entities. Twoexamples of business cases related to backhaul deployments are listed below,where we assume that an MNO is also a SP.

1. MNO lease the backhaul network infrastructure.2. MNO deploys its own infrastructure.

In order to have the complete cost evaluation, the business model for ac-cess deployment should also be known. There are two main business modelsrelated to the access segments when small cells are deployed: the closed sub-scriber group (CSG) and the open subscriber group (OSG) [85]. In the lattercase, only a closed group of users can access the indoor cells (i.e., it is con-sidered as a private network for improving the service quality) and MNO is

6.2. TECHNO-ECONOMIC FRAMEWORK 73

Total cost of ownership (TCO)

CAPEX OPEX

Infrastructure Equipment

Purchasing

Fault Management

Spectrum & Fiber Leasing

Energy Cost

Floor Space Maintenance

Installation

Figure 6.2: Total cost of ownership classification. (Reprinted from [84] ©IEEE2018, reused with permission.)

not responsible for small cell deployment cost. In the former case where theMNO is responsible for the small cells deployment cost, the small cells can beaccessed by any users (i.e., regardless of their subscription).

• Techno-economic Module: In this module, the profitability of one specificbackhaul deployment can be achieved. The required input for computing thecash flow in each year is the calculated cost from the TCO module. Cash flowrefers to the difference in the amount of cash available at the beginning ofeach time period and the amount at the end of that period. The well-knownEquation (6.2) is used for computing the total net present value (NPV):

NPV =Ln∑i=0

CFi

(1 + r)i, (6.2)

where Ln and CFi denote the network operational time and cash flow of yeari, respectively. Discount rate r is the rate of the return used in discountedcash flow analysis to determine the present value of the future cash flowsby considering time value of the money and risk of income uncertainties infuture. Time value of money states that money available at present time isworth more than the same amount in the future due to its potential earningcapacity.


NETWORKS

(a) (b)

Figure 6.3: (a) Microwave based deployment. (b) Fiber based deployment.(Reprinted from [84] ©IEEE 2018, reused with permission.)

Table 6.1: Scenarios used in the case study.

Scenarios Deployment type Technology Utility ownership

Scenario 1 (Ho MW) Homogeneous Microwave OwnedScenario 2 (He MW) Heterogeneous Microwave OwnedScenario 3 (Ho Tr) Homogeneous Fiber OwnedScenario 4 (Ho Le) Homogeneous Fiber LeasedScenario 5 (He Tr) Heterogeneous Fiber OwnedScenario 6 (He Le) Heterogeneous Fiber Leased


In this section, to assess the impact of the backhaul segment on the total cost ofthe network and evaluate the deployment profitability, we perform a case studywhere we calculate the total cost and NPV for deploying and running both theradio access network (RAN) and the backhaul segment for 10 years consideringvarious scenarios (see Table 6.1).

The topology is assumed to be a 5 × 5 dense urban area the user populationdensity of 300 users/km2 which represents the European city. The area consists ofdifferent buildings with five floors placed according to Manhattan street model [86].We consider two technologies for backhaul segment including fiber and microwavewith both homogeneous and heterogeneous access architectures as shown in Fig. 6.3.A detail explanation of the architecture of each technology assumed in our casestudy is presented in Paper VI .

The main criteria in dimensioning the access segment is the required throughputper km2 in each year considering user penetration. We assume one small cell for


Table 6.2: Input values used for TCO calculation.

Component Price (Euro)

Technician salary [hour] 52Energy cost [kWh] 0.1

Indoor yearly rental fee [m2] 220Outdoor yearly rental fee [m2] 180Small/Large microwave antenna 500/2000

G-Ethernet switch 1800Microwave hub + installation 2000

Ethernet switch 150Yearly spectrum leasing per [MHz] 5OLT (4x10G array transceiver) 7000

ONU 150Power splitter (1:16/1:32) 170/340

Fiber [km] 80Trenching [km] 45000

Leasing upfront fee [km] 800Yearly fiber leasing fee 200

Macro base station and cell site 48000Small indoor base station 250

each building in the first year for indoor coverage as well as sufficient numberof macro cells to provide coverage to outdoor users. The TCO for deploymentand operation of studied network architectures is computed using our proposedTCO module shown in Fig. 6.2. The cost parameters are retrieved from differentliterature including [41,42,87] as shown in Table 6.2.

Fig. 6.4 shows the obtained TCO values for each of the considered scenarios.The results indicate that the cost of backhaul cannot be neglected, especially inHetNet deployment. It is evident from the results that decreasing the cost of RANfor all scenarios in HetNet deployment could lead to an increase in the cost of thebackhaul segment. Thus, it is necessary to jointly consider the backhaul and theRAN segment in order to design a cost efficient network.

Fig. 6.5 shows the profitability of each deployment scenario. It shows the resultsof the NPV analysis for all six scenarios based on a yearly cost evaluation assuming


NETWORKS

0

5

10

15

20

25

Ho_MW He_MW HO_Tr He_Tr Ho_Le He_Le

TCO(M

illions€)

BH_OPEX BH_CAPEX RAN

Figure 6.4: Total cost of RAN and backhaul segments in different scenarios.(Reprinted from [84] ©IEEE 2018, reused with permission.)

-4-3-2-10123456

Ho_MW He_MW Ho_Tr Ho_Le He_Tr He_Le

NPV

(million€)

Figure 6.5: NPV of different scenarios. (Reprinted from [84] ©IEEE 2018, reusedwith permission.)


an average monthly subscription fee of 30 e per user (for voice and data), a discountrate of 10%, and revenue depending on user penetration as shown in Paper VI.Except the HetNet deployment combined with microwave backhaul (He MW), allscenarios have a positive NPV and can be considered economically viable. Thenegative NPV means non profitable solution and the results show that (He MW)is not economically viable in very dense urban area due to high component cost andthe power consumed by the microwave links in ultra-dense area. However, fiber-based backhauling provides more efficient deployment even if an operator needs todeploy its own fiber infrastructure. This is because unlike the wireless solutions,the cost of fiber deployment does not increase linearly with the number of users.Moreover, the HetNet deployment with leased fiber infrastructure for backhauling(He Le) has the lowest TCO value among all scenarios, while the NPV analysisindicates that the Ho Le deployment has the highest profitability. This can beexplained by the fact that in He Le scenario, the largest part of the investmentsfor both backhaul and RAN segment takes place in the first years. Normally,the money spend later has a lower NPV due to the potential earning capacity,inflation, etc. Therefore, without bringing a sufficient income a big investment inHetNet deployment in earlier years is not profitable in the long run. The study alsoshows that the TCO and the NPV do not always exhibit the same trends, e.g., thetechnology with the lowest TCO might not be preferable for a long-term investment,indicating a strong need for a comprehensive techno-economic framework. A moredetailed explanation of the results obtained in this study is presented in PaperVI.

Chapter 7

Conclusions and Future Work

This chapter summarizes the contributions, outlines the main findings of the thesis,and draws concluding remarks. Then, it describes open challenges and remainingissues for possible future work.

7.1 Conclusions

Next generation wireless networks need a redesign of the backhaul segment to sup-port the capacity and availability requirements of emerging wireless services. Mi-crowave backhauling at high carrier frequencies, such as millimeter-wave bands, is apromising solution for small cells, particularly in areas where the fiber option is notavailable or not feasible. However, random fluctuations of the wireless channels,especially at the millimeter-wave, pose new challenges to provide high through-put with high reliability performance in backhaul network. Besides, to make anappealing case for mobile operators to upgrade their networks, the proposed solu-tions should be cost-efficient, which makes cost evaluation and economic viabilityof different backhaul solutions essential. This thesis addresses the aforementionedchallenges in three main threads.

The first thread of this thesis investigates the topology design problem in orderto optimize the network based on average network statistics. A well-designed net-work can substantially decrease the cost and power consumption necessary to meetthe reliability requirement. We optimize or modify the network topology consider-ing the cost of link addition to satisfy the network reliability. Our design approachconsiders the correlation among failures, whose neglecting may substantially reducethe network reliability under correlated failures such as rain. To do so, we developa new model to consider the spatial correlation of rain attenuation between differ-ent links along each path, and among different paths. The results clarify that ourapproach increases the backhaul network resilience under rain at the expense ofa higher deployment cost. Besides diversity (increasing the number of establishedpaths), connection availability can be improved by deploying more reliable links at

79

80 CHAPTER 7. CONCLUSIONS AND FUTURE WORK

the expense of higher link budget. Hence, we consider the link budget as anotherfactor in topology design problem. We formulate a joint optimization problem tooptimize the total link budget with constraint on number of added links to satisfythe reliability and data rate criteria. Moreover, we investigate the differential delayimposed by multipath provisioning, and modify our topology design to guaranteea certain value of delay depending on required QoS. The evaluation results showthat, compared to the benchmarking topology design approach, the new algorithmachieves the same reliability performance in a more cost- and power-efficient way.In addition, the results indicate that tolerating higher differential delay drasticallyreduces the required link budget to meet the availability and data rate threshold.

The second thread of this thesis focuses on improving the real-time performance(e.g., in terms of throughput) of microwave backhaul networks under weather dis-turbances. We first utilize the statistical difference between rain and multipathfading to design a lightweight rain detection algorithm, and show the impact ofdifferent design parameters such as the number of samples taken from the receivedsignal, the local decision rules at the base stations, and the global decision rules ofthe SDN controller on the performance of our algorithm. This algorithm feeds anintelligent network adaptation protocol that triggers a network-layer action (e.g.,rerouting) when the rain duration is sufficiently long. We evaluate the impact ofthe rain detection error, defined as false alarm and misdetection error, on the net-work performance in terms of the throughput and the overhead on the centralizedcontroller. The results show that intelligent adaptation substantially improves thenetwork throughput while reducing the overhead on the control plane, due to elim-inating unnecessary rerouting actions. We then focus on the problem of reroutinginconsistency which arises because of asynchrony when updating routing flow tablesof different switches. To address this problem, we develop an optimal sequence ofupdate decisions to minimize the throughput loss due to rerouting. Compared tothe regular rerouting policy, our proposed approach not only reduces the through-put loss, but also substantially decreases the number of reconfigurations by wiselypicking the appropriate time to apply the new routes. We validate the efficiency ofour approach on both synthetic and a realistic deployed network.

The third thread of this thesis develops a comprehensive techno-economic frame-work to find the best backhaul option for an operator based on the TCO of thebackhaul segment and profitability measured in terms of cash flow and net presentvalue. The case study carried out in the thesis focuses on (i) two types of tech-nologies for the backhaul network, i.e., microwave and fiber; and (ii) two types ofwireless network deployments, i.e., heterogeneous and homogeneous. We highlightthe importance of selecting the right backhaul technology in order to keep the costsavings and benefits brought by heterogeneous deployments, which is particularlyrelevant for future 5G mobile networks where a high capacity transport is required.We show that ultra-dense infrastructure deployments may benefit from fiber back-hauling in long-run. However, the microwave technology may still be necessary formobile backhauling and networks with lower density of small cells. In addition, thecase study results show that deployment solutions with a low TCO do not always

7.2. OPEN CHALLENGES AND FUTURE WORK 81

lead to high profits. It is because in a long-term project, the time period when aninvestment is made may significantly affect the total project profitability. Hence, inorder to have a profitable solution (i.e., high net presented value), the deployment,which does not request a large investment in the beginning of the project and bringsincome as soon as possible is recommended.

Detailed mathematical analysis and discussions of this thesis aim to provideoriginal and important insights on the design of agile, resilient and cost-efficientfuture backhaul networks. We demonstrate the importance of an SDN-based controlarchitecture to support reliability, high throughput, and availability in real-time.Some of the possible future research directions are explained in the next section.

7.2 Open Challenges and Future Work

In this work we have addressed some challenges related to efficiently designinga wireless backhaul network to satisfy the 5G requirements. However, successfulrealization of future high-performance backhaul networks mandates other aspectsto be investigated as well. In the following, we list four main possible directions forextension of the work presented in this thesis.

It would be interesting to extend the topology design work for two-tier backhauldeployment, where each small cell transfers its information to macro base stationthrough wireless link, that is known as an inevitable solution for future backhaulnetwork. Backhauling the information from a massive number of small cells toeach macro base station is quite challenging. Power and cost-efficient design forsuch deployments is even more critical due to the low power of the small cell andtheir massive number. Besides, due to dense deployment of small cells using omni-directional antennas, interference is another parameter that should considered inthe topology design.

Furthermore, as fiber and wireless will be dominant technologies in the future,it is important to consider a joint wireless and fiber deployment scenario for thetopology design. Fiber and microwave technology are complementing solutions fore-cast to dominate the backhaul segment. Although the fiber solution may entail ahigh deployment cost, it provides huge capacity as well as high robustness againstweather-based disturbances, which makes it suitable for rain mitigation. On theother hand, the microwave solution enjoys a substantially lower deployment cost atthe price of a lower throughput and a high sensitivity to rain disturbance. Decidingon the location and technology for the extra links added to the network in orderto upgrade a tree topology to a mesh one requires a general framework that willconsider the trade-offs between the gain and the cost of each technology as well asthe geographical constraints. In addition, the design of the fiber network topologywill impact the selection of the wireless network gateways, which should also beconsidered during joint design. In such a joint design problem, it is important toconsider a more realistic cost function. Currently, in the developed algorithms, weestimate the cost of each link according to its length to make the cost model math-

82 CHAPTER 7. CONCLUSIONS AND FUTURE WORK

ematically tractable. The cost evaluation framework for each link can be extendedto include the real CAPEX and OPEX of a particular deployment solution. Anexample of such model is developed in Paper VI, and could be incorporated intothe planning framework.

The proposed algorithm for consistent-aware adaptation assumes that all traffichas the same priority and ignores different application requirements. However, thenetwork is shared by numerous users while simultaneously carrying various typesof services. To make a universal tool for controllers, updating the flows should beextensible and easy to adapt to application-specific requirements. In this thesis,we have minimized the congestion level for a given update flow time α, i.e., updatedeadline. This congestion level may be still too high for many applications withhigh reliability requirements. In such circumstances, it may be preferable to havea smaller congestion level but for a longer period. Extending our framework toaddress the application-dependent trade-offs between the duration of the updateand the amount of imposed congestion is another relevant future direction.

Last but not least, recent advantages of machine learning approaches open newpossibilities for online adaptation. For instance, reinforcement learning can beapplied to the consistency-aware rerouting problem to automatically learn the en-vironment and quickly adapt the behavior to the disruption. Due to the similarityof our problem in Chapter 5 to a reinforcement learning setup, it would be in-teresting to investigate the possibility of applying this method to interact with atime-varying environment and adapt the rerouting actions. This work can also beextended to evaluate the impact of rainfall prediction on the performance of rein-forcement learning in terms of convergence rate and optimality of the final solution.

Bibliography

[1] J. Zhao, T. Q. Quek, and Z. Lei, “Heterogeneous cellular networks using wire-less backhaul: Fast admission control and large system analysis,” IEEE Jour-nal on Selected Areas in Communications, vol. 33, no. 10, pp. 2128–2143, Oct.2015.

[2] S. Little, “Is microwave backhaul up to the 4G task?” IEEE Microwave Mag-azine, vol. 10, no. 5, pp. 67–74, Aug. 2009.

[3] M. Fiorani, P. Monti, B. Skubic, J. Mårtensson, L. Valcarenghi, P. Castoldi,and L. Wosinska, “Challenges for 5G transport networks,” in Proc. Inter-national Conference on Advanced Networks and Telecommuncations Systems(ANTS). IEEE, 2014, pp. 1–6.

[4] Ericsson, “Microwave toward 2020,” Technical report, Sept. 2014.

[5] E. M. Outlook, “Trends and needs in the microwave industry,” Technical report,Oct. 2016.

[6] R. E. Hattachi and J. Erfanian, “NGMN 5G white paper,” NGMN Board,vol. 1, Feb. 2015.

[7] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, “Wide-band millimeter-wave propagation measurements and channel models for fu-ture wireless communication system design,” IEEE Transactions on Commu-nication, vol. 63, no. 9, pp. 3029–3056, Sept. 2015.

[8] D. Benyamina, A. Hafid, and M. Gendreau, “Wireless mesh networks design-A survey,” IEEE Communications Surveys & Tutorials, vol. 14, no. 2, pp.299–310, 2012.

[9] A. Nahir, A. Orda, and A. Freund, “Topology design and control: A game-theoretic perspective,” in Proc. IEEE Conference on Computer Communica-tions (INFOCOM), 2009, pp. 1620–1628.

[10] G. Egeland and P. E. Engelstad, “The availability and reliability of wirelessmulti-hop networks with stochastic link failures,” IEEE Journal on SelectedAreas in Commun., vol. 27, no. 7, Sept. 2009.

83

84 BIBLIOGRAPHY

[11] F. Li, S. Chen, M. Huang, Z. Yin, C. Zhang, and Y. Wang, “Reliable topologydesign in time-evolving delay-tolerant networks with unreliable links,” IEEETransactions on Mobile Computing, vol. 14, no. 6, pp. 1301–1314, June.

[12] H. Wang and P. Van Mieghem, “Algebraic connectivity optimization via linkaddition,” in Proc. Conference on Bio-Inspired Models of Network, Informationand Computing Sytems (BIONETICS), 2008, p. 22.

[13] F.-C. Kuo, F. A. Zdarsky, J. Lessmann, and S. Schmid, “Cost-efficient wirelessmobile backhaul topologies: An analytical study,” in Proc. IEEE Conferenceon Global Telecommunications (GLOBECOM), 2010, pp. 1–5.

[14] R. Ramanathan and R. Rosales-Hain, “Topology control of multihop wirelessnetworks using transmit power adjustment,” in Proc. IEEE Conference of theComputer and Communications Societies (INFOCOM), vol. 2, 2000, pp. 404–413.

[15] W. Wong, X. Chen, F. Long, and S.-H. G. Chan, “Joint topology control androuting assignment for wireless mesh with directional antennas,” in Proc. IEEEConference on Global Communications Conference (GLOBECOM), 2012, pp.5699–5704.

[16] J. Gui, L. Hui, and N. Xiong, “A game-based localized multi-objective topologycontrol scheme in heterogeneous wireless networks,” IEEE Access, vol. 5, pp.2396–2416, 2017.

[17] N. Li and J. C. Hou, “Localized fault-tolerant topology control in wirelessad hoc networks,” IEEE Transactions on Parallel and Distributed Systems,vol. 17, no. 4, pp. 307–320, 2006.

[18] K. Miyao, H. Nakayama, N. Ansari, and N. Kato, “LTRT: An efficient andreliable topology control algorithm for ad-hoc networks,” IEEE Transactionson Wireless Communications, vol. 8, no. 12, 2009.

[19] L. Bao, J. Hansryd, T. Danielson, G. Sandin, and U. Noser, “Field trial onadaptive modulation of microwave communication link at 6.8 GHz,” in Proc.IEEE European Conference on Antennas and Propagation (EuCAP), 2015, pp.1–5.

[20] G. D. Abrajano and M. Okada, “Compressed sensing based detection of local-ized heavy rain using microwave network attenuation,” in Proc.IEEE EuropeanConference on Antennas and Propagation (EuCAP), 2013, pp. 2383–2386.

[21] A. Rahimi, A. Holt, G. Upton, and R. Cummings, “Use of dual-frequencymicrowave links for measuring path-averaged rainfall,” Journal of GeophysicalResearch: Atmospheres, vol. 108, no. D15, Aug. 2003.

BIBLIOGRAPHY 85

[22] A. Jabbar, J. P. Rohrer, A. Oberthaler, E. K. Cetinkaya, V. Frost, and J. P.Sterbenz, “Performance comparison of weather disruption-tolerant cross-layerrouting algorithms,” in Proc. IEEE International Conference on ComputerCommunications (INFOCOM), 2009, pp. 1143–1151.

[23] S. Wang, D. Li, and S. Xia, “The problems and solutions of network updatein SDN: A survey,” in Proc. IEEE Conference on Computer CommunicationsWorkshops (INFOCOM WKSHPS), 2015, pp. 474–479.

[24] M. Caesar, D. Caldwell, N. Feamster, J. Rexford, A. Shaikh, and J. van derMerwe, “Design and implementation of a routing control platform,” in Proc.Conference on Symposium on Networked Systems Design & Implementation(NSDI). USENIX Association, 2005, pp. 15–28.

[25] X. Jin, H. H. Liu, R. Gandhi, S. Kandula, R. Mahajan, M. Zhang, J. Rexford,and R. Wattenhofer, “Dynamic scheduling of network updates,” in Proc. Con-ference on Applications, Technologies, Architectures, and Protocols for Com-puter Communication (SIGCOMM), vol. 44, no. 4. ACM, 2014, pp. 539–550.

[26] C.-Y. Hong, S. Kandula, R. Mahajan, M. Zhang, V. Gill, M. Nanduri, andR. Wattenhofer, “Achieving high utilization with software-driven WAN,” inACM SIGCOMM Computer Communication Review, vol. 43, no. 4. ACM,2013, pp. 15–26.

[27] S. Jain, A. Kumar, S. Mandal, J. Ong, L. Poutievski, A. Singh, S. Venkata,J. Wanderer, J. Zhou, M. Zhu et al., “B4: Experience with a globally-deployedsoftware defined WAN,” ACM SIGCOMM Computer Communication Review,vol. 43, no. 4, pp. 3–14, 2013.

[28] S. Sezer, S. Scott-Hayward, P.-K. Chouhan, B. Fraser, D. Lake, J. Finnegan,N. Viljoen, M. Miller, and N. Rao, “Are we ready for SDN? implementationchallenges for software-defined networks,” IEEE Communications Magazine,vol. 51, no. 7, pp. 36–43, July 2013.

[29] K. Seppänen, J. Kilpi, and T. Suihko, “Integrating WMN based mobile back-haul with SDN control,” Mobile Networks and Applications, vol. 20, no. 1, pp.32–39, Feb. 2015.

[30] P. Dely, A. Kassler, and N. Bayer, “Openflow for wireless mesh networks,”in Proc. IEEE International Conference on Computer Communications andNetworks (ICCCN), 2011, pp. 1–6.

[31] H. Huang, P. Li, S. Guo, and W. Zhuang, “Software-defined wireless mesh net-works: Architecture and traffic orchestration,” IEEE Network, vol. 29, no. 4,pp. 24–30, 2015.

86 BIBLIOGRAPHY

[32] R. Mahajan and R. Wattenhofer, “On consistent updates in software definednetworks,” in Proc. ACM Workshop on Hot Topics in Networks. ACM, 2013,p. 20.

[33] S. Brandt, K.-T. Förster, and R. Wattenhofer, “On consistent migration offlows in SDNs,” in Proc. IEEE Conference on Computer Communications (IN-FOCOM), 2016.

[34] M. P. Anastasopoulos, A. D. Panagopoulos, and P. G. Cottis, “A distributedrouting protocol for providing QoS in wireless mesh networks operating above10 GHz,” Wireless Communications and Mobile Computing, vol. 8, no. 10, pp.1233–1245, 2008.

[35] P. Lin, J. Zhang, Y. Chen, and Q. Zhang, “Macro-femto heterogeneous networkdeployment and management: From business models to technical solutions,”IEEE Wireless Communications, vol. 18, no. 3, 2011.

[36] F. S. Farias, P. Monti, A. Vastberg, M. Nilson, J. Costa, and L. Wosinska,“Green backhauling for heterogeneous mobile access networks: What are thechallenges?” in Proc. IEEE Conference on information, Communications andSignal Processing (ICICS), 2013, pp. 1–5.

[37] H. H. Liu, X. Wu, M. Zhang, L. Yuan, R. Wattenhofer, and D. Maltz, “zUp-date: Updating data center networks with zero loss,” in ACM SIGCOMMComputer Communication Review, vol. 43, no. 4. ACM, 2013, pp. 411–422.

[38] S. Luo, H. Yu, L. Luo, and L. Li, “Arrange your network updates as you wish,”http://shouxi. name/publication/cup-tr. pdf, Tech. Rep, 2015.

[39] L. Luo, H. Yu, S. Luo, and M. Zhang, “Fast lossless traffic migration for SDNupdates,” in Proc. IEEE International Conference on Communications (ICC),2015, pp. 5803–5808.

[40] T. Maseng and P. M. Bakken, “A stochastic dynamic model of rain attenua-tion,” IEEE Transactions on Communication, vol. 29, no. 5, pp. 660–669, May1981.

[41] A. A. W. Ahmed, J. Markendahl, C. Cavdar, and A. Ghanbari, “Study onthe effects of backhual solutions on indoor mobile deployment macrocell vs.femtocell,” in Proc. IEEE International Symposium on Personal Indoor andMobile Radio Communications (PIMRC), 2013, pp. 2444–2448.

[42] M. Paolini, “An analysis of the total cost of ownership of point-to-point, point-to-multipoint, and fibre options,” White paper on crucial economics for mobiledata backhaul, 2011.

BIBLIOGRAPHY 87

[43] C. Ranaweera, M. G. Resende, K. Reichmann, P. Iannone, P. Henry, B.-J.Kim, P. Magill, K. N. Oikonomou, R. K. Sinha, and S. Woodward, “Design andoptimization of fiber optic small-cell backhaul based on an existing fiber-to-the-node residential access network,” IEEE Communications Magazine, vol. 51,no. 9, pp. 62–69, Sep. 2013.

[44] J. Liu, X. Jiang, H. Nishiyama, and N. Kato, “Reliability assessment for wire-less mesh networks under probabilistic region failure model,” IEEE Transac-tions on Vehicular Technology, vol. 60, no. 5, pp. 2253–2264, June 2011.

[45] H.-W. Lee, E. Modiano, and K. Lee, “Diverse routing in networks with prob-abilistic failures,” IEEE/ACM Transactions on Networking, vol. 18, no. 6, pp.1895–1907, Dec. 2010.

[46] J. Rak, M. Pickavet, K. S. Trivedi, J. A. Lopez, A. M. Koster, J. P. Sterbenz,E. K. Çetinkaya, T. Gomes, M. Gunkel, K. Walkowiak et al., “Future researchdirections in design of reliable communication systems,” Telecommun. Systems,vol. 60, no. 4, pp. 423–450, December 2015.

[47] M. Cheffena, L. E. Braten, and T. Ekman, “On the space-time variations ofrain attenuation,” IEEE Transactions on Antennas and Propagation, vol. 57,no. 6, pp. 1771–1782, June 2009.

[48] S. Lin, “A method for calculating rain attenuation distributions on microwavepaths,” Bell System Technical Journal, vol. 54, no. 6, pp. 1051–1086, July-Aug.1975.

[49] I.-R. Recommendations, “Propagation data and prediction methods requiredfor the design of terrestrial line-of-sight systems.” Technical Report, ITU-R P.530-12, 2007.

[50] B. C. Grémont and M. Filip, “Spatio-temporal rain attenuation model forapplication to fade mitigation techniques,” IEEE Transactions on AntennasPropagation, vol. 52, no. 5, pp. 1245–1256, May 2004.

[51] M. Castro and G. S. Granados, “Cross-layer packet scheduler design of a multi-beam broadband satellite system with adaptive coding and modulation,” IEEETransactions on Wireless Communication, vol. 6, no. 1, pp. 248–258, Jan. 2007.

[52] J. Arnau, D. Christopoulos, S. Chatzinotas, C. Mosquera, and B. Ottersten,“Performance of the multibeam satellite return link with correlated rain at-tenuation,” IEEE Transactions on Wireless Communications, vol. 13, no. 11,pp. 6286–6299, Nov. 2014.

[53] M. K. Kozlov, S. P. Tarasov, and L. G. Khachiyan, “The polynomial solvabilityof convex quadratic programming,” USSR Computational Mathematics andMathematical Physics, vol. 20, no. 5, pp. 223–228, 1980.

88 BIBLIOGRAPHY

[54] F. Yaghoubi, J. Chen, A. Rostami, and L. Wosinska, “Mitigation of rain impacton microwave backhaul networks,” in Proc. IEEE International Conference onCommunications Workshops (ICC), 2016, pp. 134–139.

[55] S. Rai, O. Deshpande, C. Ou, C. U. Martel, and B. Mukherjee, “Reliable mul-tipath provisioning for high-capacity backbone mesh networks,” IEEE/ACMTransactions on Networking, vol. 15, no. 4, pp. 803–812, 2007.

[56] W. Zhang, J. Tang, C. Wang, and S. de Soysa, “Reliable adaptive multipathprovisioning with bandwidth and differential delay constraints,” in Proc. IEEEConference on Computer Communications (INFOCOM), 2010, pp. 1–9.

[57] F. Yaghoubi, M. Furdek, P. Öhlén and L. Wosinska, “Joint topology designand control for reliable wireless networks,” submitted to IEEE Transactions onCommunication, Feb. 2019.

[58] S. Boyd, L. Xiao, A. Mutapcic, and J. Mattingley, “Notes on decompositionmethods,” Notes for EE364B, Stanford University, pp. 1–36, 2007.

[59] M. Masoudi, B. Khamidehi, and C. Cavdar, “Green cloud computing for multicell networks,” in Proc. IEEE Conference in Wireless Communications andNetworking (WCNC). IEEE, 2017, pp. 1–6.

[60] R. Horst and N. V. Thoai, “Dc programming: overview,” Journal of Optimiza-tion Theory and Applications, vol. 103, no. 1, pp. 1–43, 1999.

[61] T. H. A. Le and P. D. Tao, “The dc (difference of convex functions) program-ming and dca revisited with dc models of real world nonconvex optimizationproblems,” Annals of operations research, vol. 133, no. 1-4, pp. 23–46, 2005.

[62] D. Li and X. Sun, Nonlinear integer programming. Springer Science & BusinessMedia, 2006, vol. 84.

[63] N. Wang, E. Hossain, and V. K. Bhargava, “Backhauling 5G small cells:A radio resource management perspective,” IEEE Wireless Communications,vol. 22, no. 5, pp. 41–49, Oct. 2015.

[64] I. Otung and S. Enoch, “Mitigation of rain attenuation in fixed WiMAX usingcell site diversity,” Electronics Letters, vol. 47, no. 5, pp. 349–351, 2011.

[65] J. Rak, “A new approach to design of weather disruption-tolerant wirelessmesh networks,” Telecommunication Systems, pp. 1–13, April 2015.

[66] O. Goldshtein, H. Messer, and A. Zinevich, “Rain rate estimation using mea-surements from commercial telecommunications links,” IEEE Transactions onSignal Processing, vol. 57, no. 4, pp. 1616–1625, Jan. 2009.

BIBLIOGRAPHY 89

[67] S. Singh, R. Mudumbai, and U. Madhow, “Interference analysis for highlydirectional 60-GHz mesh networks: The case for rethinking medium accesscontrol,” IEEE/ACM Transactions on Networking, vol. 19, no. 5, pp. 1513–1527, 2011.

[68] H. Suzuki, “A statistical model for urban radio propogation,” IEEE Transac-tions on Communication, vol. 25, no. 7, pp. 673–680, Jul. 1977.

[69] A. Goldsmith, Wireless communications. Cambridge university press, 2005.

[70] U.-C. Fiebig, “Modeling rain fading in satellite communications links,” in Proc.IEEE Vehicular Technology Conference (VTC), vol. 3, 1999, pp. 1422–1426.

[71] A. Bolle and A. Nascimbene, “Microwave transmission in mobile networks,”Ericsson Review, vol. 3, pp. 124–131, 2002.

[72] M. Reitblatt, N. Foster, J. Rexford, C. Schlesinger, and D. Walker, “Abstrac-tions for network update,” in Proc. ACM Conference on Applications, Tech-nologies, Architectures, and Protocols for Computer Communication (SIG-COMM), 2012, pp. 323–334.

[73] F. Yaghoubi, M. Furdek, A. Rostami, P. Öhlén and L. Wosinka, “Consistency-aware weather disruption-tolerant routing in SDN-based wireless mesh net-works,” IEEE Transactions on Network and Service Management, no. 99, pp.1–15, Jan. 2018.

[74] D. P. Bertsekas, Dynamic programming and optimal control. Athena ScientificBelmont, MA, 1995, vol. 1, no. 2.

[75] H. Chen, T. Zhou, R. Yu, and J. Li, “Short communication: Summer rain fallduration and its diurnal cycle over the us great plains,” International Journalof Climatology, vol. 29, pp. 1515–1519, Dec. 2009.

[76] G. D. Abrajano and M. Okada, “Determining rain location through microwavemesh network signal attenuation,” in Proc. IEEE Region Ten Conference(TENCON), 2012, pp. 1–6.

[77] M. Paolini, “Crucial economics for mobile data backhaul.”

[78] Skyfiber, “How to meet your backhaul capacity needs while maximizing rev-enue,” White paper, 2013.

[79] S. Verbrugge, K. Casier, J. Van Ooteghem, and B. Lannoo, “White paper:Practical steps in techno-economic evaluation of network deployment plan-ning,” UGent/IBBT: Gent, Belgium, p. 45, 2009.

[80] D. C. Chen, T. Q. Quek, and M. Kountouris, “Backhauling in heterogeneouscellular networks: Modeling and tradeoffs,” IEEE Transactions on WirelessCommunications, vol. 14, no. 6, pp. 3194–3206, June 2015.

90 BIBLIOGRAPHY

[81] S. F. Yunas, J. Niemelä, M. Valkama, and T. Isotalo, “Techno-economicalanalysis and comparison of legacy and ultra-dense small cell networks,” in Proc.IEEE Conference on Local Computer Networks Workshops (LCN Workshops),2014, pp. 768–776.

[82] Y. Chen, L. Duan, and Q. Zhang, “Financial analysis of 4G network de-ployment,” in Proc. IEEE Conference on Computer Communications (INFO-COM), 2015, pp. 1607–1615.

[83] S. M. Mishra, J. Hwang, D. Filippini, R. Moazzami, L. Subramanian, andT. Du, “Economic analysis of networking technologies for rural developingregions,” in in International Workshop on Internet and Network Economics(WINE). Springer, 2005, pp. 184–194.

[84] F. Yaghoubi, M. Mahloo, L. Wosinska, P. Monti, F. de Souza Farias, J. C.W. A. Costa, and J. Chen, “A techno-economic framework for 5g transportnetworks,” IEEE Wireless Communications, vol. 25, no. 5, pp. 56–63, 2018.

[85] Z. Frias and J. Pérez, “Techno-economic analysis of femtocell deployment inlong-term evolution networks,” Journal on Wireless Communications and Net-working (EURASIP), vol. 2012, no. 1, pp. 1–15, 2012.

[86] M. A. Marsan, G. Abertengo, A. Francese, E. Leonardi, and F. Neri, “All-optical bidirectional manhattan networks,” in Proc. International Conferenceon Communications (ICC). IEEE, 1992, pp. 1461–1467.

[87] S. Tombaz, P. Monti, F. Farias, M. Fiorani, L. Wosinska, and J. Zander, “Isbackhaul becoming a bottleneck for green wireless access networks?” in Proc.International Conference on Communications (ICC). IEEE, 2014, pp. 4029–4035.

kth.diva-portal.orgkth.diva-portal.org/smash/get/diva2:1291162/fulltext01.pdf · iii abstract the...

Documents