E volution of Mobile DD B ackhaul -otlaA

-otl

aA

DD

9

2/

710

2

+ddadh

a*GMFTSH

9 NBSI 3-3037-06-259-879 )detnirp( NBSI 6-2037-06-259-879 )fdp(

L-NSSI 4394-9971 NSSI 4394-9971 )detnirp( NSSI 2494-9971 )fdp(

ytisrevinU otlaA

gnireenignE lacirtcelE fo loohcS gnikrowteN dna snoitacinummoC fo tnemtrapeD

if.otlaa.www

+ SSENISUB YMONOCE

+ TRA

+ NGISED ERUTCETIHCRA

+ ECNEICS

YGOLONHCET

REVOSSORC

LAROTCOD SNOITATRESSID

ydz

orD

dápr

Á l

uahk

caB

elibo

M fo

noit

ulov

E y

tisr

evi

nU

otla

A

7102

gnikrowteN dna snoitacinummoC fo tnemtrapeD

eliboM fo noitulovE luahkcaB

ydzorD dáprÁ


seires noitacilbup ytisrevinU otlaASNOITATRESSID LAROTCOD 92 / 7102

luahkcaB eliboM fo noitulovE

ydzorD dáprÁ

fo rotcoD fo eerged eht rof detelpmoc noitatressid larotcod A eht fo noissimrep eht htiw ,dednefed eb ot )ygolonhceT( ecneicS

cilbup a ta ,gnireenignE lacirtcelE fo loohcS ytisrevinU otlaA hcraM 02 no loohcs eht fo 2UT llah erutcel eht ta dleh noitanimaxe

.21 ta 7102

ytisrevinU otlaA gnireenignE lacirtcelE fo loohcS


rosseforp gnisivrepuS dnalniF ,ytisrevinU otlaA ,rennaM akkuJ .forP

srenimaxe yranimilerP

dnalniF ,älyksävyJ fo ytisrevinU ,imeinatsiR inapaT .forP aeroK ,ytisrevinU iesnoY ,miK nuyL-gnoeS .forP

tnenoppO

dnalniF ,ygolonhceT fo ytisrevinU erepmaT ,amaklaV okkiM .forP

seires noitacilbup ytisrevinU otlaASNOITATRESSID LAROTCOD 92 / 7102

© ydzorD dáprÁ

NBSI 3-3037-06-259-879 )detnirp( NBSI 6-2037-06-259-879 )fdp(


:NBSI:NRU/if.nru//:ptth 6-2037-06-259-879

yO aifarginU iknisleH 7102

dnalniF

tcartsbA otlaA 67000-IF ,00011 xoB .O.P ,ytisrevinU otlaA if.otlaa.www

rohtuA ydzorD dáprÁ

noitatressid larotcod eht fo emaN luahkcaB eliboM fo noitulovE

rehsilbuP gnireenignE lacirtcelE fo loohcS

tinU gnikrowteN dna snoitacinummoC fo tnemtrapeD

seireS seires noitacilbup ytisrevinU otlaA SNOITATRESSID LAROTCOD 92 / 7102

hcraeser fo dleiF ygolonhcet gnikrowteN

dettimbus tpircsunaM 7102 yraurbeF 31 ecnefed eht fo etaD 7102 hcraM 02

)etad( detnarg hsilbup ot noissimreP 7102 yraunaJ 13 egaugnaL hsilgnE

hpargonoM noitatressid elcitrA noitatressid yassE

tcartsbAdenfieder dna demrofsnart yllacitamard evah skrowten sseleriw ,yrutnec a fo retrauq tsal eht nI reffo osla yeht ,yltnerruc ;ecivres enohpelet elibom dedivorp skrowten ralullec ,tsrfi tA .sevil ruo

-reve edivorp ot ecap gnihsinotsa na ta gnivlove neeb sah ygolonhcet sihT .ssecca tenretni elibomelihW .ssecca ycnetal rewol dna ,setar atad tenretni elibom gnisaercni ,seiticapac krowten rehgih ot ytivitcennoc gnidivorp krowten dexfi eht ,noitnetta tsom eht nward evah skrowten ssecca oidar oidar eht no snoitavonni sA .ecap ralimis a ta devlove sah—krowten luahkcab eht—snoitats esab eht

no snoitavonni ot del sah siht ,nrut ni ,luahkcab eht rof seussi etaerc ot deunitnoc evah ecafretni ti ;krowten luahkcab eht rof segnellahc xis setagitsevni noitatressid sihT .krowten luahkcab eht

.snoitalumis hguorht ecnamrofrep rieht setaulave dna segnellahc eseht ot snoitulos stneserpsllecotmeF .smetsys G5.3 htiw deraeppa hcihw ,tpecnoc llecotmef eht saw egnellahc hcus enO

.noitcennoc tenretni deriw yrartibra na yb deluahkcab era taht sllec llams erayrev ot del sah PIoV .)PIoV( locotorP tenretnI revo ecioV ot tfihs eht si egnellahc rehtonA

evissecxe yb denedrub era hcihw ,stekcap eciov tneuqerf dna llams fo noitatropsnart tneicfifeni dexelpitlum ro deldnub era stekcap llams eht fi decuder eb nac daehrevo sihT .daehrevo tekcap

smhtirogla gnixelpitlum evitpadA .daehrevo elgnis a erahs ot stekcap elpitlum rof redro ni rehtegot .ycneicfife evorpmi rehtruf taht desoporp era

a si hcihw ,)PMoC( tniopitluM detanidrooC rof tnemeriuqer luahkcab eht si eussi rehtruf A luahkcab yb detcirtser si ytilibacilppa stI .ycneicfife ecafretni oidar sesaercni taht G4 fo erutaef

luahkcab eseht sesaerced taht dohtem a sebircsed noitatressid sihT .stnemeriuqer ycnetal .soiranecs erom ni tnemyolpme sti swolla dna ,stnemeriuqer

eht neewteb ecnereffid eht erehw detneserp si tpecnoc metsys oidar evitingoc a ,eromrehtruF .derrulb si luahkcab dna ecafretni oidar

rehgih raf ot nrut ot evah skrowten sseleriw ,G5 yb desimorp setar atad eht edivorp oT .sezis llec mumixam llams yrev seilpmi siht ,revewoh ;shtgnelevaw ertemillim sa hcus ,seicneuqerf -itlum si etadidnac ygolonhcet tneicfife tsoc a ,stnemyolped esned hcus rof luahkcab gnidivorp roF

siht ni detaulave osla si hcihw fo ecnamrofrep eht ;luahkcab dnab-ni ,evaw-ertemillim ,poh .noitatressid

.dezitnauq osla si luahkcab poh-itlum evaw-ertemillim no gnidaf niar fo tceffe eht ,yllaniFhcae fo segnellahc luahkcab laudividni eht taht wohs noitatressid siht ni detneserp stluser ehT

.snoitulos luahkcab gnidnopserroc htiw emocrevo eb nac skrowten sseleriw fo noitareneg

sdrowyeK tekcap ,gnildnub tekcap ,sllecotmef ,G5 ,G4 ,G3 ,skrowten sseleriw ,luahkcab gnidaf niar ,luahkcab poh-itlum ,oidar evitingoc ,PMoC ,gnixelpitlum

)detnirp( NBSI 3-3037-06-259-879 )fdp( NBSI 6-2037-06-259-879

L-NSSI 4394-9971 )detnirp( NSSI 4394-9971 )fdp( NSSI 2494-9971

rehsilbup fo noitacoL iknisleH gnitnirp fo noitacoL iknisleH raeY 7102

segaP 091 nru :NBSI:NRU/fi.nru//:ptth 6-2037-06-259-879

ämletsiviiT otlaA 67000 ,00011 LP ,otsipoily-otlaA if.otlaa.www

äjikeT ydzorD dáprÁ

imin najriksötiäV oituulove nokrevotriis neimesaikutnilehupaktaM

ajisiakluJ uluokaekrok nakiinketökhäS

ökkiskY sotial nakiinketokkrevoteit aj -ennekiiloteiT

ajraS seires noitacilbup ytisrevinU otlaA SNOITATRESSID LAROTCOD 92 / 7102

alasumiktuT akkiinketokkrevoteiT

mvp neskutiojrikisäK 7102.20.31 äviäpsötiäV 7102.30.02

äviäpsimätnöym navulusiakluJ 7102.10.13 ileiK itnalgnE

aifargonoM ajriksötiävilekkitrA ajriksötiäveessE

ämletsiviiT.itsävättikrem teenuttuum tavo tokrev tamottagnal anakia nadasisouvsennäjlen nesiemiiV

tynyttihek no akkiinkeT .aatsiakajaaliliibom söym nisiykyn tavaojrat tokrevnilehupaktaM aipmaekrok ähy ,ajetteetisapakokkrev aipmeruus ähy netojrat alleduepon ällävättytsämmäh

teeneiv tavo tokrevajaalit tamottagnal akkiaV .ajessnetal aipmesiahla aj ,aiskueponotriis no—okkrevotriis .kn—niimesaikut nedyethy aaojrat akoj ,okkrev äetniik ,noimouh nammiruus

atietsaah teenattesa tavo assannipajaroidar toitaavonni aksoK .niithat naamas tynyttihek iiktut ajriksötiäv ämäT .assiokrevotriis söym niisieprat-oitaavonni tunathoj no ämät ,elliokrevotriis

äykyksutirous nediin ioivra aj niisietsaah nihiän ajusiaktar äättise ;attetsaah nokrevotriis attuuk .allioitaalumis

tulos-otmeF .ässimletsejräj G5.3 nööttyäk niitteto akoj ,etisäk ulos-otmef ilo etsaah neniallät iskY .ättyethy-tenretni nokrev näetniik näätetyäk änetyethyotriis nedioj ,ajulos äineip tavo

nivyh tunathoj no PIoV .assiulehup ottonööttyäk )PIoV( nillokotorp tenretni no etsaah nenioT neittekapinää neivutsiot niesu aj netneip nes neuthoj neetuukkohetotriisnodeit neesiahla

iat allamaakkap äätneneip naadiov attuumottohet ätäT .ätsätnekokkisto atseruus nesilleethus ämäT .ättnekokkisto neniethy isky no älliin ätte ,netis ,neethy ajettekap äineip allamioskelpitlum

.ajemtirogla-itnioskelpitlum aisiviitpada iskesimatnarap neduukkohet aattodhe ajriksötiävokkrevotriis tamattesa neetnimiot )PMoC( tnioPitluM detanidrooC äättyät no etsaah neniot särE

.attuukkohet nannipajaroidar aatnarap akoj ,suusillannimiot nokrev-G4 no PMoC .teskumitaav äättise ajriksötiäv ämäT .äöttyäk nes niknetiuk tavattiojar teskumitaaveviiv nokrevotriiS

nes aatsillodham netis aj ,aiskumitaav okkrevotriis aimattesa ni-PMoC äätneveil akoj ,nämletenem .nötyäk nammejaal

neniläv nokrevotriis aj -ätnytiil assoj ,itpesnok nämletsejräjoidar neniviitingok näälletise iskäsiL .tynyträmäh no ore

äytriis yytyät nejokkrev neimottagnal ,ätl:G5 näätetyllede atioj aiskueponotriis ätiin neeskatojraT aarues saat atsoj ,niiskuutipnollaa nirtemillim netuk ,niiskuujaat niipmaekrok nojlap ätsiykyn

atojrat isiov akoj ,iskaigolonket iskaakkohetsunnatsuk sakodhE .tookulos imiskam teneip nivyh neniäsisnatsiak ,neniujaatirtemillim ,neniyppyhinom no ,elliskunnesa elliehit ellisiallät nokrevotriis

.assajriksötiäv ässät utiosylana söym no ykyksutirous aknoj ;okkrevotriis .elliokrevotriis ellisiutipnollaairtemillim sutukiav neetas näälletirääm ,iskupoL

nevlopukusokkrev namottagnal nesiakoj ätte ,tavattioso teskolut tytetise assajriksötiäv ässäT .alliusiaktar allivipos äättivles naadiov teetsaah okkrevotriis tesiättisky tamattuehia

tanasniavA ,ulos-otmef ,G5 ,G4 ,G3 ,tokrev tamottagnal ,okkrevotriis neimesaikutnilehupaktam ,oidar neniviitingok ,PMoC ,itnioskelpitlum neittekap ,suakkapneethy neittekap

sunnemiavedas ,luahkcab neniyppyhinom )utteniap( NBSI 3-3037-06-259-879 )fdp( NBSI 6-2037-06-259-879

L-NSSI 4394-9971 )utteniap( NSSI 4394-9971 )fdp( NSSI 2494-9971

akkiapusiakluJ iknisleH akkiaponiaP iknisleH isouV 7102

äräämuviS 091 nru :NBSI:NRU/fi.nru//:ptth 6-2037-06-259-879

1

Preface

First and foremost, I would like to thank Csaba Vulkán. All Iknow about the subject of this dissertation I learnt from him.Additionally, he set an example of flawless work, which I couldonly try to follow. Without the countless hours he spent on me,this dissertation would never have been completed.

I am also indebted to the Finnish experts at Nokia, JoukoKapanen, Pekka Wainio, Juha Salmelin and Jyri Putkonen forthe many years of continued cooperation. Without their sup-port, I do not know how I could have managed.

My Hungarian colleagues and co-authors at Nokia must alsobe acknowledged: Balázs Héder, László Kőrössy, Zoltán Vincze,Attila Rákos, Csaba Deák, Péter Szilágyi, Lajos Bajzik, ZsoltLakatos, Norbert Radics, Tamás Kárász, Szabolcs Novácki, andZoltán Nagy, not to mention a few others, for the many fruitfulyears that we worked together.

I also wish to thank my professor, János Bitó, under whosetutelage I spent most of my post-grad years; he strived to sup-port me to the end.

Thanks should also be expressed to Péter Kántor for the workthat we did together.

For his kind-heartedness and essential role in the completionof this dissertation, I must mention Harri Okkonen.

Furthermore, I am especially obliged to my current professor,Jukka Manner, who not only invited me to Aalto University,thus providing me with the means to finish my dissertation, butalso believed in me.

Preface

2

One other person I wish to thank is Anya Siddiqi who hasbeen a blessing to this text by spending many an hour editingeach line with the author.

This work starts with 3G systems, continues with 4G systems,and concludes with future 5G systems. Simulations of theprevious 2G generation telecommunication systems,specifically GSM, were already covered by the previousgeneration, specifically my father, whom I also must thank.

Espoo, September 2016Árpád Drozdy

3

Contents

Preface .......................................................................................... 1

List of Publications ...................................................................... 7

Author’s Contribution ................................................................. 9

List of Abbreviations .................................................................. 11

Introduction ................................................................................17

1.1 Motivation .................................................................. 18

1.2 Scope ........................................................................... 18

1.3 Research Question and Methodology ....................... 19

1.4 Contribution and Results........................................... 21

1.5 Structure of Dissertation ........................................... 22

2. Mobile backhaul ............................................................. 23

2.1 Backhaul Costs ........................................................... 24

2.2 Dimensioning Backhaul ............................................. 25

2.3 Backhaul Technologies .............................................. 26

2.4 Backhaul Architecture Evolution .............................. 27

2.5 Summary ..................................................................... 31

3. Femtocells for 3.5G ........................................................ 33

3.1 Introduction to Femtocells ........................................ 34

3.2 Interference Issues ..................................................... 36

3.3 The Business Case for Femtocells ............................. 38

3.4 Femtocell Network Integration ................................. 39

3.5 Purpose of the Femtocell Simulations ...................... 41

3.6 Simulation Results ..................................................... 43

Contents

4

3.6.1 Voice Call Quality if DSQ is Available ...................... 43

3.6.2 Voice Call Quality if only BE Queuing is Available .. 44

3.6.3 Jitter of Timing Packets ............................................ 45

3.7 Summary .................................................................... 46

4. Improving Backhaul Efficiency for 3.75G, 3.9G, and4G ........................................................................................ 47

4.1 Improving Backhaul Efficiency for VoIP in 3.75G and3.9G ..................................................................................... 48

4.1.1 The Large Overhead of VoIP Packets on the BackhaulNetwork ................................................................................. 49

4.1.2 Methods for Improving Backhaul Efficiency ............ 51

4.1.3 Performance of Bundling and Multiplexing in aneHSPA Network..................................................................... 55

4.1.4 Adaptive Multiplexing Methods ............................... 60

4.1.5 Evaluation of the Adaptive Algorithms in an LTENetwork ................................................................................. 63

4.1.6 Summary .................................................................... 66

4.2 Coordinated Multipoint Backhaul for 4G ................ 68

4.2.1 Introduction to Coordinated Multipoint (CoMP) .... 68

4.2.2 DL CoMP Overview ................................................... 72

4.2.3 Methods for Decreasing the Backhaul Load of Inter-eNB DL CoMP ........................................................................75

4.2.4 Data Sharing for CoMP JP ......................................... 77

4.2.5 Splitting the Data and Control Traffic on the X2Interface ................................................................................. 80

4.2.6 Technical description of the data sharingalgorithm ................................................................................81

4.2.7 Further benefits of the proposed method ................ 83

4.2.8 Evaluation .................................................................. 84

4.2.9 Summary .................................................................... 85

5. Backhaul for 5G (Beyond 4G) ....................................... 87

5.1 Cognitive Radio .......................................................... 88

5.1.1 Cognitive Radio and Television White Spaces ......... 88

Contents

5

5.1.2 Proposed System Concept and Use Cases ................ 89

5.1.3 Cognitive Management .............................................. 90

5.1.4 Summary ..................................................................... 91

5.2 Multi-hop Wireless Mesh Backhaul for 5G .............. 92

5.2.1 Millimetre-Wave Antennas ....................................... 93

5.2.2 Multi-Hop In-Band Backhaul ................................... 94

5.2.3 Millimetre-Wave Propagation ................................... 95

5.2.4 Combining In-Band and Dedicated BackhaulLinks ..................................................................................... 96

5.2.5 Interference ................................................................ 97

5.2.6 Link Outage .............................................................. 101

5.2.7 Ultra-low latency applications ................................ 104

5.2.8 Summary ................................................................... 105

5.3 The Effect of Rain Fading on 5G Backhaul............. 107

5.3.1 Rain Fading Measurements .................................... 107

5.3.2 Rain Fading Simulations ......................................... 108

5.3.3 Measurement Data Combined with SimulationData ................................................................................... 109

5.3.4 Summary .................................................................... 111

6. Conclusion and Discussion .......................................... 113

6.1 Conclusions from Deployed Technologies .............. 114

6.2 Discussion on the Adoption of FutureTechnologies ......................................................................... 115

References ................................................................................. 117

Publications ............................................................................. 125

6

7

List of Publications

This dissertation consists of an overview and of the followingpublications which are referred to in the text by their Romannumerals.

I Árpád Drozdy, László Kőrössy, Csaba Vulkán and János Bitó.Femtocell Quality of Service over DSL backhaul. InProceedings of the ICT Mobile and Wireless CommunicationsSummit (ICT-MobileSummit), Stockholm, Sweden, June 2008.

II Árpád Drozdy, Zoltán Vincze and Csaba Vulkán. Bundling andmultiplexing in packet based mobile backhaul. In Proceedingsof the 16th European Wireless Conference (EW), Lucca, Italy,April 2010.

III Árpád Drozdy, Attila Rákos, Zoltán Vincze and Csaba Vulkán.Adaptive VoIP Multiplexing in LTE Backhaul. In Proceedingsof the 73rd IEEE Vehicular Technology Conference (VTCSpring), Budapest, Hungary, May 2011.

IV Kamran Arshad, Richard MacKenzie, Ulrico Celentano, ÁrpádDrozdy, Stéphanie Leveil, Geneviève Mange, Juan Rico, ArturoMedela and Christophe Rosik. Resource Management for QoSSupport in Cognitive Radio Networks. IEEE CommunicationsMagazine, volume 52, issue 3, pp. 114-120, March, 2014.

List of Publications

8

V Árpád Drozdy, Péter Kántor and János Bitó. Effects of RainFading in 5G Millimetre Wavelength Mesh Networks. InProceedings of the 10th European Conference on Antennasand Propagation (EuCAP), Davos, Switzerland, April 2016.

VI Árpád Drozdy, Jouko Kapanen and Jukka Manner. User LevelPerformance Analysis of Multi-hop In-band Backhaul for 5G.Accepted for publication in Wireless Networks pending minorrevisions.

9

Author’s Contribution

Publication I: “Femtocell Quality of Service over DSLbackhaul”

The simulational implementation, simulational studies, andresearch work was done by Árpád Drozdy, under thesupervision of Csaba Vulkán and János Bitó, with assistancefrom László Kőrössy. The article was written by Árpád Drozdy,László Kőrössy, and Csaba Vulkán, most of the text was writtenby Árpád Drozdy.

Publication II: “Bundling and multiplexing in packet basedmobile backhaul”

The simulational implementation, simulational studies, andresearch work was done by Árpád Drozdy, under the leadershipof Csaba Vulkán, with the assistance of Zoltán Vincze. Thearticle was written by all three authors, most of the text waswritten by Árpád Drozdy.

Publication III: “Adaptive VoIP Multiplexing in LTEBackhaul”

The simulational implementation, simulational studies, andresearch work was done by Árpád Drozdy. Of the three ideas inthe article, the “predictive method” was proposed by AttilaRákos, the “average number method” by Árpád Drozdy, and thelater discarded “stochastic method” by Zoltán Vincze. The workwas done under the leadership and supervision of CsabaVulkán. All four authors contributed to the writing of the article,most of the text was written by Árpád Drozdy.

Author’s Contribution

10

Publication IV: “Resource Management for QoS Support inCognitive Radio Networks”

This research was done by the QoSMOS EU project. Theresearch and writing was done together, Árpád Drozdy workedin the project.

Publication V: “Effects of Rain Fading in 5G MillimetreWavelength Mesh Networks”

The research and simulations on a 5G mesh network was doneby Árpád Drozdy. Péter Kántor did the research on rain fading.The work was done under the supervision of János Bitó. Thetext was written by Árpád Drozdy, except for the section on linktransformation, which was written by Péter Kántor.

Publication VI: “User Level Performance Analysis of Multi-hop In-band Backhaul for 5G”

The work presented in this paper was done by Árpád Drozdy,who also wrote the text. Jouko Kapanen provided valuableassistance. Jukka Manner gave advice.

11

List of Abbreviations

1G The first generation of wireless mobiletelecommunications technology, which wereanalogue, e.g. NMT

2G The second generation of wireless mobiletelecommunications technology, e.g. GSM

3G The third generation of wireless mobiletelecommunications technology, e.g. UMTS

3.5G The enhanced version of the third generation ofwireless mobile telecommunications technology,e.g. HSDPA

3.75G A marketing name for a further enhanced versionof the third generation of wireless mobiletelecommunications technology, specificallyevolved HSPA

3.9G The specific generational number of LTEtechnology

3GPP 3rd Generation Partnership Project

4G The fourth generation of wireless mobiletelecommunications technology, e.g. LTE-A

5G The fifth generation of wireless mobiletelecommunications technology

AAL ATM Adaptation Layer

ADSL Asymmetric Digital Subscriber Line

AMR Adaptive Multi-Rate audio codec

AP Access Point


12

ATM Asynchronous Transfer Mode

BBU Baseband Unit

BS Base Station

BSC Base Station Controller

BTS Base Transceiver Station

CCDF Complementary Cumulative DistributionFunction

CM-RM Cognitive Manager for Resource Management

CM-SM Cognitive Manager for Spectrum Management

CPRI Common Public Radio Interface

C-RAN Cloud Radio Access Network, sometimesCentralized Radio Access Network

CS Circuit Switched

CS/CB Coordinated Scheduling / CoordinatedBeamforming

CSI Channel State Information

CoMP Coordinated Multipoint

DAS Distributed Antenna System

DCS Dynamic Cell Selection

DL Downlink

DSL Digital Subscriber Line, or Digital SubscriberLoop

EDGE Enhanced Data rates for GSM Evolution

eNB Evolved Node B

eNodeB Evolved Node B

ETSI European Telecommunications StandardsInstitute

GGSN Gateway GPRS Support Node


13

GNSS Global Navigation Satellite System

GPRS General Packet Radio Service

GPS Global Positioning System

GSM Global System for Mobile Communications,originally Groupe Spécial Mobile

GTP GPRS Tunnelling Protocol

GTP-U GPRS Tunnelling Protocol, user plane

HARQ Hybrid Automatic Retransmission Request

HetNet Heterogeneous Networks

HSDPA High-Speed Downlink Packet Access

HSUPA High-Speed Uplink Packet Access

HSPA High-Speed Packet Access, also known asHSDPA/HSUPA

IEEE Institute of Electrical and Electronics Engineers

I-HSPA Internet HSPA

IMT-Advanced International Mobile Telecommunications— Advanced

IoT Internet of Things

IP Internet Protocol

IPsec Internet Protocol Security

ISM Industrial, Scientific and Medical radio frequencyband

ITU International Telecommunication Union

JP Joint Processing

JT Joint Transmission

LAN Local Area Network

LOS Line Of Sight

LTE Long Term Evolution


14

LTE-A Long Term Evolution Advanced

M2M Machine to Machine

MAC Medium Access Control

MIMO Multiple Input Multiple Output

MSC Mobile Switching Centre

MTC Machine Type Communication

MTU Maximum Transmission Unit

NFV Network Function Virtualization

NGMN Next Generation Mobile Networks Alliance

NLOS Non Line Of Sight

NMT Nordic Mobile Telephone

Node B Not an acronym, similar in meaning to BS

PDCP Packet Data Convergence Protocol

PDH Plesiochronous Digital Hierarchy

PDU Payload Data Units

PHY Physical layer

PPP Point-to-Point Protocol

PPPmux Point-to-Point Protocol multiplexing

PRB Physical Resource Blocks

PS Packet Switched

QoS Quality of Service

QoSMOS Quality of Service and MObility driven cognitiveradio Systems

QPSK Quaternary Phase Shift Keying

RF Radio Frequency

RLC Radio Link Control


15

RNC Radio Network Controller

RRH Remote Radio Heads

RTP Real-time Transport Protocol

RTT Round Trip Time

SAE-GW System Architecture Evolution Gateway

SDH Synchronous Digital Hierarchy

SDN Software-Defined Network

SGSN Serving GPRS Support Node

SHDSL Single-Pair High-speed Digital Subscriber Line

SIM Subscriber Identity Module

SONET Synchronous Optical Networking

TCP Transmission Control Protocol

TDD Time Division Duplex

TDM Time Division Multiplexing

TTI Transmission Time Interval

UDP User Datagram Protocol

UE User Equipment

UHF Ultra High Frequency

UL Uplink

UMTS Universal Mobile Telecommunications System

UNI User-Network Interface

VDSL Very high bit rate Digital Subscriber Line

VHF Very High Frequency

VoIP Voice over Internet Protocol

VoLTE Voice over LTE

WAN Wide Area Network


16

WCDMA Wideband Code Division Multiple Access

Wi-Fi Institute of Electrical and Electronics Engineers'(IEEE) 802.11 standards

WiMAX Worldwide Interoperability for Microwave Access

17

Introduction

Wireless cellular networks, also known as mobile networks arenot only wireless, but also include fixed links which connectbase stations to a mobile core or public internet network. Thesefixed links that connect the cellular base stations to each otherand the core network are called backhaul links, which form thebackhaul network. This network is often overlooked. When in-vestigating the performance of a cellular network, often onlythe air interface is taken into account. While such calculationswere sufficiently accurate in earlier generation systems, inmodern networks, they lead to overestimation of the networkperformance. In the first digital cellular networks, backhaullinks required predetermined data rates and latency was not anissue; these networks were deployed with dedicated backhaullinks that had sufficient capacity. In contrast, when these net-works were upgraded to newer generation radio technologies,often the base stations were replaced while the backhaul linkswere not. In such scenarios, it would often be the backhaul linksthat limited the maximum data rates that could be offered. Inthese later generations, the data rate demand of a fully loadedbase station depends on the radio conditions of the mobilelinks. Consequently, costs can be saved if the backhaul is notdimensioned to support the maximum theoretical radio capac-ity, this limits the performance for only a fraction of the time.

Additionally, there are certain cases where backhaul is an es-pecially critical issue, such as for femtocells and CoordinatedMultipoint (CoMP), as will be described in later chapters.

Introduction

18

1.1 Motivation

In the past few decades, one of the biggest technological ad-vances has been mobile telecommunications. In the currentdecade, “traditional” phones are being swiftly replaced withsmartphones. This shift to smartphones creates a huge demandfor mobile data connectivity. Mobile networks are also compet-ing with fixed internet connections in terms of both cost anddata rates.

Cellular networks have evolved at an unprecedented speed.Approximately every decade, a completely new Radio AccessNetwork (RAN) generation has appeared. This rapid develop-ment has necessitated a wide range of research work. However,one of the more neglected research topics is mobile backhaul.Researching it requires extensive knowledge in both radio andpacket transport technologies.

The demand for ever higher mobile data rates will be satisfiedby a fifth generation standard, which has been promised for the2020 Olympic Games in Tokyo; thereby continuing the trendof ever increasing network capacities. 5G is currently a very‘hot’ research topic. The appearance of 5G will most certainlycreate more challenges for the backhaul; therefore, research inmobile backhaul will likely increase in the next few years.

1.2 Scope

This dissertation only considers systems standardized by the3rd Generation Partnership Project (3GPP). In the 2nd and 3rd

generation, multiple different standards were deployed. GSM,standardized by the European Telecommunications StandardsInstitute (ETSI), became the dominant 2nd generation standardglobally. 3GPP was founded with ETSI as a core organizationalpartner to create a 3G standard called Universal Mobile Tele-communications System (UMTS). UMTS has emerged as theleading 3G standard.The 4th generation 3GPP standard, namely, Long Term Evolu-tion (LTE), did initially have competition in the form of World-wide Interoperability for Microwave Access (WiMAX), alsoknown as IEEE 802.16, but it never posed a serious challenge.

Introduction

19

A WiMAX simulator was also implemented by the author in aseparate project, however, as this technology did not have alasting effect on the evolution of mobile networks, it is excludedfrom this dissertation.

Currently, there is consensus that 5G will be standardized by3GPP; organizations such as the Next Generation Mobile Net-work Alliance (NGMN), and the Small Cell Forum create rec-ommendations for 3GPP instead of competing standards.

This dissertation considers technology generations startingfrom 3.5G and ending with 5G, which is to be standardized inthe future. 1G, such as Nordic Mobile Telephone (NMT), is ex-cluded as it was analogue, not digital. 2G and basic 3G are notelaborated, as they have predetermined backhaul capacity de-mands; therefore, the issues investigated in this dissertation donot arise.

All technologies are evaluated based on the offered user levelperformance, in other words, the level of service that the userscan perceive.

1.3 Research Question and Methodology

Before the launch of any mobile network, network performancecannot be measured. Therefore, simulations are vital to predictthe behaviour of such networks before deployment. All of thedescribed research was conducted before the wide scale deploy-ment of the technology in question, and the technologies de-scribed in later chapters are only to be deployed in the future.

The overall research question was to evaluate the user levelperformance of mobile backhaul. The specific research ques-tions always focused on a critical backhaul issue of the giventechnology generation:

For third generation systems, the possibility ofproviding backhaul with arbitrary wired internet con-nections, such as a Digital Subscriber Line (DSL).Small cells with such backhaul are called femtocells.The feasibility of the femtocell concept was investi-gated. The specific research questions were whether adifferentiating scheduler was necessary, and whether

Introduction

20

clock synchronization is possible over the fixed inter-net network. (Chapter 3)

In evolved third generation, and early fourth genera-tion networks, an analysis was conducted of thetransport efficiency of Voice over Internet Protocol(VoIP). The goal was to find a more efficient methodof transporting VoIP packets on the backhaul. (Chap-ter 4.1)

For advanced fourth generation systems, the backhaulrequirements of Coordinated Multipoint (CoMP)were studied. The question was to quantify the back-haul latency and capacity requirements of coordinat-ing the transmission of multiple base stations. (Chap-ter 4.2)

One of the research questions for fifth generation net-works was to establish a framework for providingQuality of Service (QoS) with only opportunistic spec-trum use. (Chapter 5.1)

Another goal for fifth generation systems was to finda low cost backhaul architecture that can provide thecapacities necessary for millimetre-wave radio accessnetworks. For this, the feasibility and performance ofmulti-hop backhaul was analysed. A specific researchquestion was whether the interference conditions per-mit this multi-hop backhaul mesh to be in-band. Theoption of combining in-band links with dedicatedlinks was also studied. A further point was the sensi-tivity of the backhaul network to occasional link fail-ures. Additionally, ultra-low latency requirementswere considered. (Chapter 5.2)

The final question for fifth generation networks wasto quantify the probability of partial network discon-nection due to rain fading in a backhaul mesh net-work. (Chapter 5.3)

The research was based on simulations. For this, the authorof this dissertation had to implement specific functions, orcompletely new simulators for the technology in question. Forthe research presented in Chapter 3, the author had to imple-

Introduction

21

ment a femtocell simulator on an existing NS-2 simulation plat-form used by Nokia. The work elaborated in Chapter 4 requiredthe implementation of adaptive bundling and multiplexing,and Coordinated Multipoint (CoMP) on this same NS-2 simu-lator platform. The research presented in Section 5.2 necessi-tated the coding of a 5G simulator on the public NS-3 simula-tion platform. The work exhibited in Section 5.3 was carried outwith its own simulator written by the author himself. In all ofthese cases, implementing the simulator on the platform re-quired far more work hours than evaluating the results.

1.4 Contribution and Results

This dissertation summarizes most of the author’s researchwork carried out from 2007 to 2016. The coding of the neces-sary simulator features, the running and processing of simula-tions were conducted by the self-same author.

The results elaborated in this dissertation were used at thetime by the industry for the designing of the system in question,as at the time, the specific topics had not yet been investigatedin such a manner from the backhaul point of view. However,with the exception of the 5G research, the system concepts werealready defined at the time.

The main findings elaborated in this dissertation are: As discussed in Chapter 3, differentiated scheduling

has been found to be unavoidable in providing the ex-pected QoS for femtocells. Furthermore, clock syn-chronization over the internet network is also possibleonly if a scheduler is available.

As shown in Chapter 4.1, bundling and multiplexinghas been confirmed as a method to increase thetransport efficiency of VoIP packets. Moreover, twonovel, adaptive packet multiplexing algorithms wereinvented to further enhance efficiency. The adaptivealgorithms were internationally patented [2].

As described in Chapter 4.2, the backhaul require-ments of inter-eNB CoMP have been evaluated. Thework directly led to the invention of a mechanismwhich minimizes the backhaul requirements of inter-

Introduction

22

eNB CoMP. An international patent application hasbeen filed, and the patent for the proposed mecha-nism is pending [3].

As explained in Chapter 5.1, a framework for a cogni-tive radio architecture has been developed.

As elaborated in Chapter 5.2, a multi-hop, millimetre-wave backhaul mesh network concept has been eval-uated. From the study, it can be concluded that inter-ference will be a manageable issue, and that in-bandmulti-hop backhaul is feasible. The combination ofdedicated and in-band backhaul links has beendemonstrated to be efficient. It has also been proventhat occasional link failures on the backhaul meshcause minimal performance degradation providedthat appropriate rerouting algorithms are employed.Furthermore, results show that ultra-low latency canbe guaranteed with differentiated scheduling.

As presented in Chapter 5.3, it has been calculatedbased on measurement data, that rain fading willcause network failures in millimetre-wave backhaulmeshes in only the heaviest rainfalls.

1.5 Structure of Dissertation

The next five chapters of this dissertation are structured as fol-lows. Each of the main chapters presents one or more of theincluded publications.

Chapter 2 provides an introduction to the basic concepts ofmobile backhaul. Chapter 3 discusses femtocells for 3.5G net-works, and shows that backhaul connectivity is a major issuefor femtocells. Chapter 4 describes backhaul technology tech-niques for 3.75G, 3.9G and 4G. It discusses adaptive bundlingand multiplexing, followed by the backhaul requirements ofCoMP. Chapter 5 provides a description of possible 5G net-works, including cognitive networks and millimetre-wavemulti-hop networks. It also investigates the probability of rainfading disrupting 5G backhaul networks. Finally, Chapter 6concludes the dissertation.

23

2. Mobile backhaul

Wireless networks cannot operate without a wired infrastruc-ture. The users in each wireless cell are served by a base station,and the base stations require fixed data connections to the mo-bile core. The transport networks serving mobile networks arecalled mobile backhaul. It connects wireless base stations to ra-dio network controllers, gateways and servers. This chapterprovides an overview of mobile backhaul and its architecture ingeneral by describing basic concepts and issues which will beelaborated upon in further chapters.

As described by Metsälä and Salmelin [4], backhaul can bedivided into three tiers, the access, aggregation, and backbonetier, see Figure 1. At the base station end of the backhaul net-work, the so-called lower level, is the access tier. Characteristi-cally, the access tier connects the base stations to the aggrega-tion tier with the fewest possible, or close to fewest number oflinks. Therefore, its topology is usually a tree, or it may have afew additional links for redundancy. It is common for accesstier links to be fixed wireless microwave links due to their easeof deployment. In contrast, the aggregation tier topology isbased more on rings, thus ensuring service even if a single linkfails. At the upper level, the backbone tier connects the aggre-gation tier with the core nodes, and the core nodes with one an-other. It is always comprised of very high capacity fibre-opticrings.

Mobile Backhaul

24

Figure 1. Backhaul network tiers.

2.1 Backhaul Costs

While it is evident that the majority of an operator’s cost is re-lated to the mobile radio access network, nevertheless, the costefficiency of the backhaul network is business critical. Metsäläand Salmelin [4] estimate that the backhaul share of the entirenetwork related costs is 10–40%, depending on the density ofthe cellular network. 70–80% of the backhaul cost [4] is asso-ciated with the access tier, due to the larger number of links,and the least cost is associated with the backbone tier. Thus, thecost and transport efficiency of the access tier is the primarytarget of backhaul efficiency optimization. However, reliabilityis more critical at the upper levels of the backhaul network asan interruption in service affects a larger geographical area, andmore subscribers.

Since the dawn of cellular networks in the 1980s, cell sizeshave continuously decreased. In parallel, the cost of a base sta-tion device itself has also dropped exponentially in accordancewith Moore’s law. The cost of deploying the base station and

Gateway Core sites

Routers /switches

Basestations

Access tier Aggregation tier Backbone tier

Mobile Backhaul

25

establishing a backhaul connection has decreased much less,due to the associated labour and site costs. Future small cellswill require many more, short, high capacity backhaul links.Additionally, as the average revenue per user continues to de-cline while the number of base stations increases, the cost ofbackhaul connectivity will also have to decrease.

Backhaul can either be owned by the operator, or leased. It isthe specific scenario that determines the preference. In the caseof leased backhaul, a specific service level agreement is neces-sary which applies between the User-Network Interfaces(UNI). The capacity, Quality of Service (QoS) and cost of theleased lines depends on both the service level agreements andthe employed transport technology.

2.2 Dimensioning Backhaul

Adaptive modulation and coding was not featured in earlierdigital systems, such as 2G GSM, and 3G UMTS, also known asWideband Code Division Multiple Access (WCDMA). There-fore, the total data rate of a fully loaded cell was predeterminedand independent of the radio channel conditions; it wasstraightforward to determine the necessary capacities of the ac-cess tier backhaul links. This was suited to circuit switched mo-bile telephone service, which at first was the only service of-fered. UMTS was initially designed to also offer mobile internetaccess. However, after its standardization in 1999, it was real-ized that the constant bit rate channels it offered were very in-efficient for fluctuating data rate traffic, such as web browsing.Furthermore, UMTS always employed Quaternary Phase ShiftKeying (QPSK) modulation, even if the radio channel condi-tions would have allowed much higher order modulation and,therefore, higher data rates. Subsequently, already in 2002,High-Speed Downlink Packet Access (HSDPA) was standard-ized, followed by High-Speed Uplink Packet Access (HSUPA)in 2005. The combination of HSDPA and HSUPA has beennamed High-Speed Packet Access (HSPA). They introducedadaptive modulation and coding for the downlink (DL) and up-link (UL) directions, respectively. Furthermore, they also intro-duced fast scheduling, where the base station resets the data

Mobile Backhaul

26

rate of every user every few milliseconds, instead of only at thebeginning of the connection. In parallel, Enhanced Data ratesfor GSM Evolution (EDGE) was defined to allow adaptive mod-ulation for GSM. Adaptive coding was already introduced toGSM by General Packet Radio Service (GPRS).

Consequently, the data rate of a fully loaded base station de-pends on its radio channels. It is unlikely that every UserEquipment (UE) attached to a base station has an ideal radiochannel, which would enable it to use the modulation and cod-ing combination with the highest possible data rate. Therefore,in order to minimize cost, the capacity of the backhaul connec-tion serving a base station does not necessarily have to be asmuch as the absolute maximum data rate of the cell. However,it should be higher than the data rate of a fully loaded cell withaverage radio conditions, as backhaul capacity costs less thanradio access spectrum. When multiple last mile backhaul linksare combined together at upper levels, the capacity require-ment of the common links is also not trivial. It should be higherthan the sum of the data rates of the cells with average radioconditions in busy hours, but may be lower than the sum of theabsolute maximum data rates. A reasonable guideline is to cal-culate with the absolute maximum of one cell plus the averagefull load of the other cells. This way the advertised maximumdata rate is possible for at least some of the UEs most of thetime.

2.3 Backhaul Technologies

In the previous decade, the most common physical layerbackhaul link technologies were: Plesiochronous Digital Hier-archy (PDH), Synchronous Optical Networking (SONET) /Synchronous Digital Hierarchy (SDH); and for microwavelinks, Time Division Multiplexing (TDM). Furthermore, Asyn-chronous Transfer Mode (ATM) was the standard adopted byUMTS for communication between the base stations and theradio network controller (RNC). While these were previouslycommonly used standards, they are currently being completelyreplaced by Ethernet based technologies; by 2015, over three-quarters of backhaul links were Ethernet based [4]. Ethernet

Mobile Backhaul

27

has proven to be a very cost effective solution as its cost hasdecreased due to its ubiquity, and it can be employed over mi-crowave, copper and fibre-optic links. Though at first Ethernetwas only used for local area networks (LAN), to date, CarrierEthernet has become widespread for long distance datatransport. Multiprotocol Label Switching (MPLS) can be usedover Ethernet to replace Internet Protocol (IP) routing. Whileearlier physical layer protocols were primarily designed fortransporting Circuit Switched (CS) traffic, the shift to PacketSwitched (PS) traffic has also favoured Ethernet.

Conventional telephone calls are transported more easily in aCS network, as their data rate is mostly constant, and require agiven latency and maximum packet loss rate. However, packet-based data traffic, such as Transmission Control Protocol (TCP)or User Datagram Protocol (UDP) over IP, has far outgrownconventional voice traffic. Transporting IP packets is efficientonly in a PS network; thus, since the introduction of GPRS,there have been both CS and PS networks in parallel. However,it would save cost to entirely eliminate the CS network, and tohave a fully IP-based network. Current PS networks offer farlower latencies than those required by voice calls, thereby ena-bling Voice over Internet Protocol (VoIP) to be used for regularphone calls. VoIP fragments phone calls into packets and trans-ports only packets. The shift to all-IP networks is a long-termtransition that is still in progress.

2.4 Backhaul Architecture Evolution

At the time GSM was standardized, in the late 1980s, hard-ware for baseband digital processing was very expensive.Therefore, moving the digital processing of many cells to a sin-gle site saved cost; this node was called the Base Station Con-troller (BSC), which served multiple base stations called BaseTransceiver Stations (BTS). The BSCs were connected to a Mo-bile Switching Centre (MSC) in the core network. Afterwards,as packet switched networking was added to GSM in the formof GPRS, the CS traffic continues to be connected to the MSC.However, in the PS domain, the RNCs are connected to a Serv-

Mobile Backhaul

28

ing GPRS Support Node (SGSN), which is connected to a Gate-way GPRS Support Node (GGSN). For a depiction of the archi-tecture evolution, see Figure 2.

Figure 2. 3GPP architecture evolution, packet switched domain only.

UMTS continued to use this node structure [5], though thebase station was renamed to Node B, and the BSC was renamed

BSCBTSUE

SGSN GGSN

2.5GGSMGPRS

RNCNode BUE

SGSN GGSN

3GUMTS

RNCNode BUE SGSN

GGSN

3.5GHSPADirectTunneling

Node BUE SGSN

GGSN

3.75GI-HSPA

eNBUE MME

SAE-GW

3.9GLTE

BBUUE MME

SAE-GW

4GLTE-AC-RAN

fronthaul

Control PathData Path

RRH

Mobile Backhaul

29

to RNC; both nodes’ complexity increased. Shortly after thestandardization of UMTS, the issues with this concept becameevident. One of the key features of HSPA [6], fast scheduling,could not be implemented at the RNC, as the signalling roundtrip time (RTT) between the Node B and RNC is too high.Therefore, much of the Medium Access Control (MAC) func-tionalities had to be moved from the RNC to the Node B. Dueto the fluctuating data rates of HSPA, flow control and conges-tion control had to be implemented on the link between theNode B and RNC, which is referred to as the Iub interface.

Later versions of HSPA changed this node structure [7][8]. Inthe direct tunnelling, also known as one tunnel solution, theuser plane data traffic circumvents the SGSN and is trans-ported directly between the RNC and GGSN. Control plane traf-fic is routed via the SGSN as in previous versions. This directdelivery decreases the latency of packet delivery between thesetwo nodes.

An alternative to the direct tunnelling architecture, intro-duced in the same Release, number 7, is Internet HSPA (I-HSPA), which introduces a flat architecture [8]. In this solu-tion, not only does the data path bypass the SGSN, but the sep-arate RNC node is also completely eliminated. All of the RNCfunctionalities are moved to the Node B, essentially as if everyNode B would have its own RNC integrated into it. This greatlysimplifies the network, as the Iub interface is completely re-moved. It also saves the cost of deploying RNCs, and makes thenetwork more resilient as there are no more RNC disconnec-tions or failures. Furthermore, data packet traffic is more effi-cient over the Iu interface to the GGSN than over the Iub, dueto less overhead. Latency is also significantly reduced as thedata path is not obstructed between the Node B and the gate-way node. Additionally, it streamlines the transition to LongTerm Evolution (LTE) architectures.

LTE [9][10] has a similar structure as I-HSPA, with the dif-ference that the Node B is called Evolved Node B (eNodeB) orjust eNB. Furthermore, instead of the SGSN, there is the Mo-bility Management Entity (MME), and the gateway node iscalled System Architecture Evolution Gateway (SAE-GW),which is also known as the Serving Gateway (S-GW). The node

Mobile Backhaul

30

structure of LTE has rendered it in some ways simpler and lesscostly than legacy UMTS/HSPA, as there is no separate RNCnode.

A further twist in architectural development is the Cloud Ra-dio Access Network (C-RAN) [11]. With it, the architecturalevolution has come full circle. In GSM, the equipment near theantenna, the BTS, was relatively simple and most of the pro-cessing was moved to a more central site, the BSC. In UMTS,both the Node B and the RNC performed baseband processing.In a C-RAN network, the device at the antenna is the simplestpossible and all of the processing is performed in a more centralnode, the digital Baseband Unit (BBU). The difference betweenthe two concepts is that in a C-RAN network, the antennas areconnected to extremely simple Remote Radio Heads (RRH).RRHs do not perform digital signal processing, they only con-vert a sampled radio signal to the radio frequency. Alterna-tively, with Radio over Fiber (RoF) technology, the RRH is es-sentially just two amplifiers. The RRHs simply convert and for-ward a data stream with insignificant delay. The RRHs form aDistributed Antenna System (DAS). This architecture allowsdigital baseband equipment cost savings and enhanced coordi-nation among neighbouring cells. On the other hand, it requireshigh capacity, low latency links between the BBU and the RRHswhich are not interrupted by routers, and in some cases, it ne-cessitates fibre-optic cables. These links are called fronthaullinks, and they are not considered to be backhaul links.

The architecture of 5G networks has not yet been decided.Chapter 5 will investigate a multi-hop backhaul architecture. Itis likely though, that 5G networks will continue the flat archi-tecture concept and will have no RNC-like element between thebase stations and the core network. It is also possible that thespecific core network nodes will be replaced by virtual nodes.As the backhaul architecture of 5G is expected to differ signifi-cantly from that of 4G, further investigations in this field arenecessary, such as the research described in Chapter 5.

Mobile Backhaul

31

2.5 Summary

Though often overlooked when planning mobile networks,backhaul represents an increasing portion of total networkcosts. Therefore, dimensioning backhaul is a business criticalissue. In contrast to 2G and 3G networks, it is not optimal in3.5G and 4G networks to proportion access tier backhaul ca-pacities to match the maximum air interface capacities.

Ethernet has become the dominant physical layer transporttechnology, and is replacing all other standards. Fibre-opticand copper cables continue to be used in conjunction with fixedwireless links.

Backhaul architecture has changed every generation, withevery generation bringing about a different optimal level ofcentralization. With the introduction of C-RAN, this progresshas come full circle. The architecture of 5G has not yet beenstandardized, although multi-hop backhaul seems to be a likelycandidate.

32

33

3. Femtocells for 3.5G

In the 1990s, mobile cellular telecommunications networkswere comprised of cells designed to cover large areas, as cover-age was the foremost issue. As these systems attracted more us-ers, network capacity became increasingly important. To im-prove network capacity, the overlaying cell layout was comple-mented with smaller cells that covered smaller, more fre-quented areas, such as an office building or a shopping mall.These smaller cells were named micro cells, and the large cellswere named macro cells. Subsequently, as mobile telecommu-nication became more commonplace and even more capacitywas required, the term picocell was used to describe cells thatwere even smaller than microcells. Macrocells, microcells andpicocells differed in cell size, but utilized similar backhaul solu-tions. In the late 2000s, the word ‘femtocell’ had to be intro-duced for the new architecture with a completely new form ofbackhaul. With the further evolution and increased complexityof cellular networks, the broad term ‘small cells' has becomecommon for referring to microcells, picocells, and femtocells.Macrocells and small cells jointly comprise the architecture re-ferred to as Heterogeneous Networks (HetNet).

In this chapter, an overview of 3G femtocells is presented, anda series of femtocell backhaul simulations from 2007 is de-scribed. The details of the research can be found in PublicationI: At the time of publication, the term in use was 3G femtocell,but since the investigated technology is in fact HSDPA/HSUPAinstead of legacy UMTS, it is technically 3.5G, which currentlyis a much more favoured term than 3G due to competitive rea-sons.

Femtocells for 3.5G

34

Publication I won the best paper award at the ICT MobileSummit in 2008, and its Hungarian language version won abest paper award in the journal titled Híradástechnika [1].

3.1 Introduction to Femtocells

Deploying dedicated backhaul links has implied serious costsfor Telecom operators, who are always searching for lower costalternatives. One such alternative is to connect a base stationwith already existing internet connections. Fixed wireline in-ternet, such as Digital Subscriber Line (DSL), coaxial cable in-ternet access, and fibre to the home are technologies that arecompeting with broadband mobile internet. Despite this com-petition, fixed internet connections can be used for the benefitof mobile networks, as they are commercially available in verymany locations. The base station that uses an arbitrary internetconnection as backhaul is called a femto access point. It is alsoknown as an access point base station, a home base station, ahome node, a home cell base station, or for 3G systems only, a3G home gateway. It is a small, low cost, low power, scaleddown version of a Node B. The small cell it serves is called afemtocell, which typically covers a single home, or one floor ofan office building. It is designed to be typically used with a DSLor coaxial cable connection as backhaul. If a location has a fi-bre-optic internet access, the available bandwidth allows regu-lar Node Bs to be connected, not only femto access points. As afemtocell essentially enables mobile phone calls over a fixed in-ternet connection, it is one of the concepts comprising fixed-mobile convergence. Fixed-mobile convergence is the erosionof differences between fixed and mobile networks.

To further decrease the cost of a femtocell, deployment costis also minimized by simplifying the deployment to a “plug andplay” approach. In this concept the installation of the femto ac-cess point is simplified and cannot be more than plugging itinto an internet connection and a power outlet; it has to be self-configuring. Thus, femto access points can be given to the endusers who install it for themselves. Many end users have theirown wired internet connections at home or at work, which theycan access much more easily than professionals from a telecom

Femtocells for 3.5G

35

operator; furthermore, this implies no cost for the operator. Onthe other hand, this does not allow for the planning of the posi-tion of femtocells, in contrast to the careful planning of the po-sitions of macrocells, microcells and picocells. It is necessary tooptimize the layout of macro-, micro- and picocells in order tominimize inter-cell interference. This is not possible with arbi-trarily placed femtocells, therefore, other approaches areneeded to cope with inter-cell interference. Allocating a sepa-rate spectrum for femtocells would be prohibitively expensiveand spectrally inefficient. Therefore, the overlaying macrocelllayer and femtocells have to share the same frequency bands,in other words, they are in-band.

While allowing a macro Node B to use a DSL connection asbackhaul would simplify macrocell deployment and save costsfor operators, many DSL links do not provide the Quality ofService (QoS) required for a macro Node B. A macrocell re-quires a high data rate, low latency reliable connection forbackhaul. While the downlink data rate of DSL connections isoften more than what is required to serve a macrocell, DSL datarates are asymmetric. The uplink data rate may not be sufficientfor a macrocell, depending on the type of DSL connection andthe length of the DSL loop. Typically, an Asymmetric DigitalSubscriber Line (ADSL) uplink data rate is insufficient for amacrocell, while a Very High Bit Rate Digital Subscriber Line(VDSL) uplink rate is sufficient provided that the loop length isshort enough. Single-Pair High-speed Digital Subscriber Line(SHDSL) is also insufficient in the downlink. In contrast, afemtocell can provide satisfactory service even if its data rate isseverely limited by the backhaul connection. Additionally, DSLmay use interleaving which increases latency. Furthermore, aDSL connection does not guarantee that its nominal maximumdata rate is always available. This topic was studied in Publica-tion I for femtocells, and will be elaborated later on in thischapter.

The main competitor technology of femtocells is in fact notmacro-, micro- or picocells, but Wi-Fi. Wi-Fi is the most com-mon group of technologies for extending internet service froma fixed internet subscription wirelessly over a short distance. Itis supported by every present day smartphone and laptop, is

Femtocells for 3.5G

36

inexpensive, and can provide very fast data connections. Theservice it provides is very similar to that of femtocells, with onlytwo key differences. One difference is that Wi-Fi does not ena-ble conventional telephone calls. Thus, a user within Wi-Fi cov-erage, but isolated from the mobile cellular network, will missany incoming calls and will not be able to initiate phone calls.The other difference is that in the case of Wi-Fi, only the pro-vider of the fixed internet connection can charge the subscriberfor the data service, while for a femtocell both the mobile oper-ator and the fixed internet provider can charge the user. Bothdifferences make femtocells attractive to mobile network oper-ators; however, subscribers might have to pay twice for theirdata traffic.

3.2 Interference Issues

In the 21st century, people spend most of their time indoors,which is the reason for indoor coverage being perhaps more im-portant than outdoor coverage. However, providing ubiquitousindoor coverage from outdoor macro Node Bs is challengingdue to the penetration losses radio signals experience whenpropagating through walls, especially at higher frequencies.While penetration loss poses a challenge for macrocells, it isbeneficial for femtocells [12]. Femto access points are designedto be placed indoors and provide indoor coverage, as wired in-ternet connections are usually available indoors. The penetra-tion losses of building walls, therefore, separate femtocellsfrom the overlaying macro layer and decrease inter-cell inter-ference. When planning macrocell layouts, it is challenging, ifat all possible to provide coverage inside every building. Therewill usually be buildings at the edge of macrocells in whichthere is no macro layer coverage, or only near the windows. Ar-eas, such as basements or underground parking lots are evenmore difficult to cover. While this does cause some users to beextremely dissatisfied, deploying more macrocells to providecoverage in every basement is prohibitively expensive. There-fore, an important role of femtocells is to provide service insidebuildings that happen to be isolated from the macrocell layer.

Femtocells for 3.5G

37

This offers both cost savings for the operators and higher cus-tomer satisfaction.

While designing femtocells to be in-band with macrocellsdoes eliminate the need to procure new spectrum, it introduceschallenging interference issues. One important concept is to al-low User Equipment (UEs) to freely hand over amongfemtocells and macrocells [13][14]. This prevents an unassoci-ated UE and femtocell from causing each other serious inter-cell interference at very close proximities. However, from suchopen access, it also follows that all of the UEs in the area of thefemtocell will add to its total data traffic, but the backhaul costof this is borne by the subscriber of the internet connectionused as femtocell backhaul. Thus, this is not possible in somebusiness models.

To mitigate interference, femtocells should operate at thehighest possible frequency. The higher penetration loss in-creases the separation provided by the walls of buildings. Fur-thermore, if the macro Node Bs operate at multiple frequencies,it is possible that the macro coverage is continuous only at thelower frequencies, as higher frequencies do not propagate as faras from the same sites. At locations where there is only lowerfrequency macro coverage, femtocells can operate at higher fre-quencies without interfering with the macro layer. Addition-ally, femto access points are more likely to be placed in suchlocations.

Another important countermeasure against interference is tolimit the transmission power of femto access points [13][14].Limiting this transmission power mitigates the interferencethat femtocells can cause to macrocells in the downlink direc-tion. This also limits the size of femtocells, and the necessarytransmission powers of UEs attached to femtocells. In turn, thismitigates the uplink interference from femtocells to macrocells.

The typical maximum transmission power of a femto accesspoint is approximately 20 dBm [13][14]. This is even lower thanthe maximum transmission power of a UE, which is 24 dBm.As the femto access points have lower transmission powersthan macro Node Bs, the most expensive component, the RF

Femtocells for 3.5G

38

circuitry, also costs less. While the price of commercially avail-able femto access points is subject to change, it currently rangesbetween 85€ and 200€ for private customers [15].

3.3 The Business Case for Femtocells

Though the cost of a femto access point is low, this cost is notshared among many customers. A femtocell is designed to serveonly a few users, for example, those who work on the same of-fice floor or live in the same apartment. Therefore, the revenuefrom only a few customers, perhaps just a single customer, hasto cover the cost of a femto access point. In addition to this, thecustomers may have to pay the cost of the increased traffic ofthe fixed internet connection and the electricity consumption.This is in contrast to macro Node Bs, which are designed toserve a very large number of parallel connections. Additionally,since only a fraction of the subscribers within a macrocell gen-erate traffic at any given moment, the number of subscribersthat can be served within the area of a macrocell is far more.Therefore, while the procurement, deployment and operationof a macrocell costs the operator far more than a femtocell, anetwork of femtocells is not necessarily the cost efficient alter-native.

The business model for femtocells is less well defined than itstechnology. There are multiple competing business models onways to finance femtocells. While it is clear that the fixed inter-net connection, electricity, and premises has to be provided bythe mobile subscriber, the cost of the femto access point can becovered by the subscriber or the mobile network operator. Sub-scribers may wish to purchase femto access points if phone callsin the femtocells are at a discount price. Even free calls are pos-sible as femtocell traffic does not strain the macrocell network.However, this shifts the business model of mobile operatorsaway from the very successful subscription based model, to amodel that receives revenue from femto access point sales. Analternative of this business model is to charge a fixed monthlyfee for the use of femtocells. A completely different approach isthat the operator provides the femto access point for free if thesubscriber has no macrocell coverage at home or at work, in

Femtocells for 3.5G

39

this case, the regular usage prices apply. This approach gener-ates little extra revenue for the operators, as it only providesservice in more locations. However, it can greatly increase thesatisfaction of some users and prevent them from cancellingtheir subscriptions. Due to the difficulties of providing macro-cell coverage in every building, it is extremely expensive for op-erators to deploy such a number of macro Node Bs that all userswould have satisfactory indoor coverage; handing outfemtocells costs far less. As an example, this is the currentmodel in Japan. A further possibility is to charge the users foracquiring the femto access point and charging them for theirtraffic in the femtocells. This means that subscribers have topay for the femto access point, the data traffic on the fixed in-ternet connection, the traffic on the mobile link, and the elec-tricity consumption of the femto access point; as well as per-haps even a monthly fee for the femto access point. While sub-scribers will be reluctant to pay for a femtocell on such terms,they may be forced to comply, because none of the competingmobile operators provide coverage in their homes or work-places. While this business model will not lead to the wide-spread use of femtocells, this business model is very beneficialfor mobile operators.

A further issue is whether only the subscribers who pay forthe femto access point may use the femtocell, or anyone withinthe femtocell. In some of the aforementioned business cases, itwould decrease the operator’s revenue if more users could usethe femtocell with discounted prices or for free. This may leadto some users not being allowed to hand over to a femtocelleven if they are very close to the femto access point. As previ-ously mentioned, this would cause interference issues. Anotherissue is that allowing any user to hand over to a femtocell willgenerate additional data traffic on the fixed internet connec-tion, and which may impose additional fees for its subscriber.

3.4 Femtocell Network Integration

As the backhaul links of femtocells are arbitrary internet con-nections instead of secure, dedicated links, a completely newarchitecture is necessary for connecting the femtocells to the

Femtocells for 3.5G

40

core network. To ensure the security of the data sent over theinternet connection, additional security measures, such as In-ternet Protocol Security (IPsec) has to be used, which adds sub-stantial overhead. In order to enable femto access points frommultiple manufacturers to be used, the interface between thefemto access point and the core network has to be standardized.The entire system architecture does not have to be redesignedif an existing standard interface is repurposed.

One possible such interface is the Iub interface, which is orig-inally intended to connect Node Bs with Radio Network Con-trollers (RNC). In the architecture called Iub over IP, the femtoaccess points are connected over the internet to an RNC on thecore side, through an Iub interface. The primary issue with suchan architecture is that the RNC node controls many of the func-tions of a Node B. If the latency of the link between these twonodes is not guaranteed, or if there are packet losses, then com-plications may arise for which the interface was originally notdesigned to cope with. The delay between the two nodes mayalso decrease performance. Additionally, while conventionalRNCs were proportioned to handle only a small number ofNode Bs, the number of femtocells that would need to be con-trolled may be several orders of magnitude higher.

A different approach is the femto gateway model. This con-cept is based on Internet High Speed Packet Access (I-HSPA),which is also referred to as Evolved HSPA (eHSPA), HSPA+,3.5G, or 3.75G, and is essentially an evolutionary step forwardfrom the HSPA architecture [7][8]. The term I-HSPA is usedfor the solution from Nokia, which is the solution investigatedin this study. An I-HSPA Node B is responsible for both theNode B and RNC functionalities; thus, the air interface of thefemto access point is controlled completely locally, without la-tency issues. The femto access points are connected to a nodecalled a femto gateway through an Iu-IP interface, over the ar-bitrary internet connection. The femto gateway provides astandard Iu interface for the core network side, while acting asa gateway for the traffic of tens of thousands of femto accesspoints. Due to these advantages, and its popularity with ven-dors, it is the architecture considered in Publication I. The con-sidered femto gateway topology is shown in Figure 3. It can be

Femtocells for 3.5G

41

seen that the DSL line is shared between the femto access pointand conventional internet use, such as desktop computers orlaptops. The two competing traffic types are combined and splitat the Customer Premises Equipment (CPE) and the DSL ac-cess multiplexer.

Figure 3. Femto gateway topology

3.5 Purpose of the Femtocell Simulations

At the time of the study, the main question was whether a DSLline provides sufficient QoS in order to serve as backhaul for afemtocell. A DSL line has a limited maximum data rate, anadded latency, and may drop packets. If a DSL connection uti-lizes interleaving, it will add serious latency to the femtocell.Thus, in the simulations, the interleaving on the DSL link wasconsidered to be disabled.

The problem of limited bandwidth is exacerbated due to thepacket overheads, especially the large IPsec overhead. In thesimulation setup, the voice calls, using Voice over Internet Pro-tocol (VoIP), generated packets containing 32 bytes of userdata, corresponding to a data rate of 12.8 kb/s when not trans-mitting only silence frames. The overhead from all the protocollayers and IPsec inflates this size to such an extent that it can-not fit into less than four Asynchronous Transfer Mode (ATM)cells, which means that 4×53=212 bytes is reserved for a singlepacket containing only 32 bytes of useful data. This corre-sponds to a total data rate of 84.8 kb/s. While DSL connections

Femtocells for 3.5G

42

have higher downlink speeds, many DSL connections only havea 512 kb/s data rate available in the uplink, this implies thatsuch a DSL connection could support at best six parallel voicecalls. This strictly limits the number of simultaneous voice callsone femtocell can serve.

However, the most serious issue is that the voice traffic fromthe femtocell has to compete with other traffic sources for thebandwidth on the DSL line. As conventional mobile phone callsupport is the key distinguishing feature of femtocells com-pared to Wi-Fi, its quality of service has to be guaranteed. Thequality of the voice calls may degrade when the data rate of theDSL line is consumed and its buffers are filled by parallel con-nections, especially TCP connections. These parallel connec-tions may originate either from the fixed or mobile sources.One of the focuses of the simulations was to study this perfor-mance degradation.

The other investigated issue was whether it is possible to syn-chronize the internal reference clocks of the femto access pointsover the network with timing packets. For accurate time refer-ence, femto access points cannot rely on Global Navigation Sat-ellite Systems (GNSS), such as Global Positioning System(GPS), as the weak signal from the satellites may not be availa-ble to penetrate into the indoor location of the femtocell. Whilemacro Node Bs have very precise internal clocks for accuratetime and frequency reference, such a clock would have too higha cost for a femtocell. It costs much less to install a simpler clockin the femto access point, and to continuously synchronize itwith more accurate clocks over the data network with timingpackets [16][17]. The issue with this solution is that it is verysensitive to packet delays, as packet jitter decreases the preci-sion of the synchronization.

Scheduling AlternativesWhile some connections, such as VoIP, are sensitive to packetdelays, other connections, such as TCP based ones, are not. Ad-ditionally, voice calls are business critical as they are usuallybilled at a higher price, and subscribers expect reliable voiceservice. Thus, the quality of delay sensitive services can be im-proved by employing a scheduler that differentiates different

Femtocells for 3.5G

43

types of services. Differentiated Service Queuing (DSQ) assignsdelay sensitive packets a higher priority level at the scheduler,providing them with prioritized access to the link, in this caseto the DSL line. High priority is assigned to control plane pack-ets, timing packets, and VoIP packets, while the lowest priorityis assigned to TCP packets. However, DSQ is not always availa-ble, especially since the DSL link’s endpoint, the customerpremises equipment is not originally designed to hostfemtocells, and generally does not distinguish between the de-vices attached to it, such as a femto access point. If DSQ is notavailable at the DSL access multiplexer and CPE, then packetshave to be scheduled in a first-come first-served manner, whichis referred to as Best Effort (BE) queuing. The simulations com-pared the user level performance of these two alternativescheduling methods, in order to determine whether intelligentscheduling is necessary. At the time of this study, while it wassuspected that a sophisticated scheduling mechanism would bebeneficial, its extent had not yet been quantized until then, andthis was a critical and unanswered question for the industry.

3.6 Simulation Results

Simulations considered scenarios where DSQ is available, inaddition to scenarios where only BE queuing is possible. Fur-thermore, the delay variation of timing packets was also inves-tigated.

For further details on the simulation settings, readers are re-ferred to Publication I.

3.6.1 Voice Call Quality if DSQ is Available

The simulations revealed that if DSQ is available, then the qual-ity of voice calls is unaffected by all other traffic types. As longas the maximum number of voice calls that the DSL line cansupport is not exceeded, the call quality depends only on theproportion of dropped packets. This sensitivity to packet losswas quantified with the help of Conversational Mean OpinionScore (MOSc) [18], which is a quality of experience metric forphone calls. It is a subjective rating from 1 to 5, where 5 is the

Femtocells for 3.5G

44

best score. In the simulation, however, it is calculated with thePESQ algorithm [19][20], considering packet delays and losses,to attain an objective value equal to the expected average sub-jective value. The MOSc values for voice call quality, accordingto the ratio of dropped ATM cells on the DSL link, is shown inFigure 4. As previously mentioned, one small voice packet can-not fit into less than four ATM cells on the DSL link due to theextreme overhead, the voice call is four times as sensitive to celllosses due to the overhead. This is visible in Figure 4, which alsodisplays a purely hypothetical simulation case where the packetoverheads are omitted. In addition to the simulated values, atheoretical maximum calculation is also presented which takesinto account only the limitations of the speech compression andthe packet drops.

Figure 4. Average MOSc according to cell loss probability on the DSL link, inthe case of DSQ

3.6.2 Voice Call Quality if only BE Queuing is Available

If DSQ is not available and BE queuing is used, then the qualityof voice calls also depends on the volume of competing datatraffic. The simulations described in I provided evidence thatall voice calls will be dissatisfactory if there is just a single TCPbased connection also transferred over the DSL line. For exam-ple, in the traffic mix with four mobile voice calls and no otherconnections, the MOSc of the voice calls is 4.06 and 4.08 for

1

1.5

2

2.5

3

3.5

4

4.5

5

0.00001 0.0001 0.001 0.01 0.1 1

Aver

age

MO

Sc

Cell loss probability on the DSL link

theoretical maximum

simulated result

without overhead

ExcellentGood

Medium

Poor

Bad

Very bad

Femtocells for 3.5G

45

Downlink (DL) and Uplink (UL) directions, respectively, re-gardless of the available bandwidth. This represents clearly au-dible voice quality. In contrast, if there is just a single mobileTCP data connection both in the uplink and downlink direction,then voice quality is unacceptable. If the UL data rate of theDSL line is 512 kb/s, then the MOSc in the UL direction will be1.95-2.01, which indicates that the call is barely audible. If theUL data rate is 1 Mb/s, then the voice call quality scores 2.77 onthe MOSc scale, which means the conversation is difficult tounderstand. In the DL direction the situation is only better forhigher DL data rates. These values clearly indicate that satis-factory QoS cannot be provided with BE queuing when the DSLlink also serves TCP traffic. This is due to the inherent natureof TCP, as TCP increases its data rate until packets are droppedfrom a buffer.

3.6.3 Jitter of Timing Packets

The investigation of the delay and jitter of timing synchroniza-tion packets also considered both DSQ and BE scheduling. Inthe case where DSQ is available, the timing packets that arriveat the DSL link are always the first to be scheduled, as they havethe highest priority. Despite this, other traffic types do have aneffect on the end to end latency of the timing packets. If thereis a packet transmission already in progress when the timingpacket arrives, then the high priority timing packet cannot ac-cess the DSL link immediately, and is buffered. The ongoingpacket transmission has to be completed, and only then will thehigh priority packet be sent through the DSL link. While thiscauses only a minor extra delay, it is completely random, there-fore causing some jitter, and network synchronization is verysensitive to this jitter.

Simulation results revealed that if there are no parallel TCPor streaming connections active, then the end to end delay ofthe timing packets will be very steady. If there are other activehigh data rate applications, then the maximum delay can in-crease by maximum 2.5 ms, and at least 95% of timing packetswill have less than 1 ms added delay in all investigated scenar-ios. This is within the tolerance limit, according to Metsälä and

Femtocells for 3.5G

46

Salmelin [2], who state that the packet jitter should be less than5 ms. This assures us that clock synchronization over the net-work is possible if DSQ is available, and timing packets are pri-oritized.

On the other hand, if DSQ is not available, only BE queuing,then timing packets may be excessively delayed, depending onthe length of the packet queue. When any data connection fullyloads the DSL link, then the delay of timing packets can in-crease by several seconds. This clearly is several orders of mag-nitude more than that which is tolerable for precise synchroni-zation. Therefore, in the case of BE queuing, synchronizationover the backhaul network is only possible when the DSL linkis uncongested.

3.7 Summary

Femto access points are low cost Node Bs for providing indoorcoverage where there would otherwise be a coverage hole. Theycan use any wired broadband internet connection for backhaul.Femtocells can be considered to be a competitor to Wi-Fi, andtheir advantage is that they also offer business critical mobilephone service. Additionally, femtocells free users from switch-ing access technologies when moving indoors.

At the time of this study, these simulations provided noveland valuable results for the industry. They underlined the inef-ficiency of VoIP, which would be the key selling point offemtocells. The main conclusion that was drawn from these re-sults is that DSQ scheduling is strictly necessary, which was nottrivial at the time. The reason for this is twofold. One reason isthat QoS for mobile phone calls can only be guaranteed withDSQ, as otherwise anyone using the fixed home network orfemtocell will disrupt any phone call every time they click on anew web page. The other reason is that precise internal clocksynchronization over the backhaul network is not possiblewithout DSQ if the backhaul link is not idle. This implies that amore precise and therefore expensive internal reference clockwould be necessary.

Publications from this femtocell study received two best pa-per awards.

47

4. Improving Backhaul Effi-ciency for 3.75G, 3.9G, and4G

Generally, three main generations are distinguished for exist-ing, digital mobile networks, namely 2G, 3G, and 4G. However,the third generation includes a diverse set of different accessand backhaul technologies; therefore, more subtle generationaldistinctions are also made. UMTS is specifically 3G, while 3.5Grefers to HSPA. 3.75G represents Evolved HSPA. Though 3.75Gis not a technical term, it has been used to market EvolvedHSPA systems to distinguish it from HSPA. From a technicalpoint of view, Evolved HSPA would also be classified as 3.5G,however, this dissertation uses the term 3.75G as there is a sig-nificant difference in the backhaul architecture. The Interna-tional Telecommunication Union (ITU) RadiocommunicationSector defined in 2008 requirements for what can be labelledas 4G; these are the International Mobile Telecommunications-Advanced (IMT-A) requirements. Long Term Evolution (LTE)did not strictly meet all of these requirements; therefore, theprecise term for LTE is 3.9G. However, LTE was subsequentlyallowed to be marketed as 4G, and technical literature has cometo refer to LTE as 4G. The IMT-A requirements are fully met byLong Term Evolution Advanced (LTE-A); therefore, LTE-A isgenuinely 4G.

The air interfaces of 3.75G and 3.9G are incompatible, thereis a significant generational leap between them. In contrast,their streamlined backhaul has the same architecture, enablinga seamless generational transition. Therefore, the next chapter

Improving Backhaul Efficiency for 3.75G, 3.9G, and 4G

48

discusses the backhaul issues of 3.75G and 3.9G together. LTE-A introduces Coordinated Multipoint (CoMP), which in turn in-troduces completely new backhaul requirements; therefore,LTE-A is discussed separately.

The following results are published in Publications II and III.Publication III won best paper award at the European WirelessConference. Additionally, this work led to an accepted interna-tional patent [2] and a pending international patent [3].

4.1 Improving Backhaul Efficiency for VoIP in 3.75Gand 3.9G

Over the past two decades, mobile networks have undergone amajor shift from Circuit Switched (CS) to Packet Switched (PS)networks. 2G networks, such as GSM, were only designed fortelephone service. Therefore, they were circuit switched, as cir-cuit switched architecture is well suited to the constantly lowdata rate, delay sensitive voice connections. In contrast, inter-net traffic, which is continuously growing in volume, requirespacket switched architectures. Supporting both a circuit andpacket switched infrastructure implies higher costs; thus, theparallel circuit switched architecture is gradually being phasedout, and all telecommunication networks are moving towardsan all-IP structure. CS voice calls are being replaced by PS Voiceover Internet Protocol (VoIP) connections. In the past decades,the data traffic volume of IP connections has been increasing atan exponential rate; thus, voice traffic is only a small fractionof all traffic and is declining further in comparison. Due to this,the inefficiency of VoIP connections due to the large overheadshas not been the primary concern of the industry. However, thisdoes not mean that transport efficiency issues of VoIP shouldbe ignored, since voice traffic volume is not decreasing in abso-lute terms.

As this section will describe, Publications II and III have pro-posed methods with which the transport efficiency of VoIP con-nections can be greatly increased, at the cost of a tolerable levelof additional VoIP packet delay. These proposed methods havebeen patented internationally [2]. By saving bandwidth withthe presented methods, more voice connections can be served


49

with limited capacity backhaul links, and freed bandwidth canbe used by other connection types.

The backhaul efficiency improving techniques presented inthis section are applied to both Evolved HSPA (eHSPA) andLong Term Evolution (LTE) systems. eHSPA is referred to as I-HSPA by Nokia, and is also known as HSPA.

4.1.1 The Large Overhead of VoIP Packets on the Back-haul Network

The drawback of VoIP in mobile networks is the large overhead.For an Adaptive Multi-Rate audio codec (AMR) with a bit rateof 12.2 kbit/s, voice packets will have in the case of eHSPAbackhaul a bandwidth efficiency of 21% at best, and in the caseof LTE 15% at best. Note that 12.2 kbit/s is the data rate of thebest quality constant bit rate AMR codec standardized by3GPP, lower bit rate AMR codecs will have the same overheadwith even less useful data, and thus even worse efficiency.

When there is speech to encode, an AMR codec generates onepacket every 20 ms, for the high quality 12.2 kbit/s rate thispacket is 32 bytes. On the other hand, during silent periods, thecodec generates only a 9 byte silence frame every 160 ms. Thesesilence packets are necessary in order for the receiver to beaware that the connection has not been disconnected. Whilethey are generated less frequently, implying that the packetoverhead is also transmitted at a lower average data rate, thetransmitted information is in a sense only overhead. In a typicaltelephone conversation, only one person speaks at a time, andthe dialogue alternates between the two speakers. From this itfollows that only silence frames are generated close to half ofthe time in one direction.

In an eHSPA system, the data packets are transferred over thebackhaul links through the interface called the Gn user planeinterface or Gn-U. In the simulations, it is assumed that thetransport is Ethernet based. When the data packets are passedthrough the protocol layers, the protocol layers add their ownheader. Typically, prior to transmission through the Gn-U in-terface, data packets are stacked with headers from the Real-time Transport Protocol (RTP), User Datagram Protocol(UDP), Internet Protocol (IP), GPRS Tunnelling Protocol user


50

plane (GTP-U) layers, and then again from UDP and IP layers,and finally the Ethernet layer. Assuming IP version four, thetotal packet overhead is 118 bytes, see Table 1; this would beeven higher in the case of IP version six.

The protocol layers of an LTE S1 user plane interface are sim-ilar, see Table 2. However, in our studies, it was assumed thatLTE would necessitate IPsec for data security, further addingto the overhead. Thus, the total overhead increases to 188bytes.

Table 1. Overhead calculation for eHSPA VoIP packets over the Gn-U inter-face. 32 bytes of useful data is transferred with 106 bytes of packet overhead.The useful data is 21% of the total, the overhead is 79%. The inter packetgap is included in the Ethernet overhead.

Layer Overhead(bytes)

Total size(bytes)

AMR - 32RTP 12 44UDP 8 52IPv4 20 72

GTP-U 12 84UDP 8 92IPv4 20 112

Ethernet 38 150

Table 2. Overhead calculation for LTE VoIP packets over the S1-U interface.32 bytes of useful data is transferred with 188 bytes of packet overhead. Theuseful data is ~15% of the total, the overhead is ~85%. The inter packet gapis included in the Ethernet overhead. The IPsec overhead is an averagevalue.

Layer Overhead(bytes)

Total size(bytes)

AMR - 32RTP 12 44UDP 8 52IPv4 20 72

GTP-U 12 84UDP 8 92IPv4 20 112

IPsec 66(average value) 178

Ethernet 42 220

From the values it can be seen that fully IP based transportwas not designed for small VoIP packets. Most applications


51

generate larger packets, TCP packets are typically close in sizeto the maximum transmission unit. This enables much bettertransport efficiency, as the similar overhead is a smaller frac-tion of the larger total packet size. A typical Maximum Trans-mission Unit (MTU) sized TCP packet in eHSPA includes only10% overhead and in LTE 13% overhead.

4.1.2 Methods for Improving Backhaul Efficiency

The low efficiency of VoIP packet transport offers us the oppor-tunity to reduce the physical layer bit rate to a fraction of theoriginal, as only a fraction of the transmitted data is useful in-formation. Thus, if the capacities of backhaul links are limited,the backhaul network can be enabled to support more VoIPconnections, and the freed link capacity can be used by othertraffic types.

Header CompressionOne possible technique for decreasing overhead size is to com-press the headers of consecutive packets that change only occa-sionally during the lifetime of a connection. The full headerscan be sent only occasionally by the compressor at the senderto the de-compressor at the receiver; otherwise, only thechanged fields are transmitted. This allows the 40 byte longRTP/UDP/IP headers to be reduced to an average of 3 bytes.

In eHSPA and LTE networks header compression is per-formed in the Packet Data Convergence Protocol (PDCP) layerthat either uses IP Header Compression [21] or Robust HeaderCompression protocol [22]. The main issue is that since thePDCP layer is located in the Node B and the UE, the headercompression is normally carried out only on the air interfaceand is not extended to the backhaul.

Bundling and MultiplexingThe other possible technique is to aggregate multiple voicepackets into one aggregated frame instead of sending each oneseparately. This way, multiple packets will have one shared setof lower protocol layer headers. This aggregation scheme is re-ferred to as bundling when multiple consecutive voice packetsof the same user are bundled into one frame, see Figure 5.


52

When packets from multiple users are aggregated into oneframe, it is referred to as multiplexing, see Figure 6.

Figure 5. Concept of bundling, DL shown.

Figure 6. Concept of multiplexing, DL shown.

A typical voice codec, such as AMR, generates a small packetevery 20 ms during talk spurts. From a transport efficiencypoint of view, it would be more efficient if the codec generatedlarger packets at less frequent intervals. However, if, for exam-ple, during a talk spurt, every second packet is deliberately de-layed 20 ms, then pairs of packets can be sent simultaneously.These synchronously sent packets can be bundled into a single,larger frame for improved transport efficiency. If four packetsare bundled together, then every fourth packet is buffered for60 ms, the two following packets are buffered 40 ms and20 ms, and the last packet is not delayed. While bundling re-duces the relative overhead of lower layers to a fraction of theoriginal value, it is clear that its drawback is the additional VoIPpacket delay. The number of bundled packets is limited by themaximum allowed mouth-to-ear delay, which is 250ms, minusthe network end to end delay. This maximum mouth to ear de-lay limit is necessary in order for the people in the phone call to

GatewayBundling entityBuffer

Flow A queueAssembler

Flow X queue

Timer

Node B1Bundling entityDisassembler

Transportframes

Bundling entityDisassembler

Voiceframes

Flow A

Flow X

Flow B

Flow A

Flow X

Flow B

Voiceframes

Flow B queue

Node Bn


53

immediately notice when the other person has started talkingand to not start talking simultaneously. If any voice packet ar-rives at the receiver later than this, it is treated as a lost packet,and the listening user experiences a very short service interrup-tion. To satisfy the latency demands of some other traffic types,current mobile networks tend to have much lower end-to-enddelays than this, even for long distance connections. If the net-work latency is sufficiently low, bundling will not impair theperceived speech quality.

Multiplexing is more efficient than bundling, as it allows thepackets of multiple VoIP connections on the same connectionpath to be multiplexed together. Its efficiency depends on thenumber of parallel VoIP connections; more connections allowmore packets to be multiplexed together and lower addedpacket delays. In the typical use case, a multiplexing timer isalso employed to ensure that no packet is delayed more than apre-set limit. Upon the expiration of the timer the bufferedpackets are multiplexed and sent out regardless of the numberof packets present, and the timer is restarted when the nextpacket arrives. The maximum delays are in practice set to lessthan 20 ms; therefore, all the multiplexed packets will be fromdifferent connections. The amount of packets multiplexed intoone aggregated frame is limited not by the delay requirementas in the case of bundling, instead by the size of the MTU, whichis 1500 bytes in the case of Ethernet. When there are many par-allel VoIP connections, then the maximum number of packetscan be reached within a few milliseconds. In contrast, if thereis only one parallel VoIP connection, then multiplexing will of-fer no efficiency gain.

Even higher transport efficiency can be achieved if multiplex-ing is combined with header compression, thus the headers ofthe higher protocol layers are compressed, while the overheadfrom lower protocol layers is reduced by multiplexing.

The bundling and multiplexing investigated in the study wasperformed at the base station and the gateway; in the case ofeHSPA, these nodes are called Node B and GGSN, while in thecase of LTE, they are called eNB and SAE-GW. To enable bun-dling or multiplexing, additional capabilities have to be imple-


54

mented in these nodes. Thus, aggregated frames are trans-ported over the entire backhaul. It is not possible to extendmultiplexing to the other side of the core network or the air in-terface, as packets for different connections will have differentsources and destinations. Bundling also cannot be extended tothe other side of the core network, because it cannot be guar-anteed that the other end will support the bundling, as it is notstandardized. In contrast, bundling can be performed on the airinterface [23][24]. Header compression, bundling, or both onthe air interface can increase the number of voice calls that canbe served in a cell. In fact, if VoIP packets at the air interfaceare congested and delayed, multiple VoIP packets will automat-ically be transmitted together when a user is scheduled. If bun-dling is employed on the air interface, then it can be continuedon the backhaul up to the gateway (or started from the gateway)without any additional delay. This is an advantage of bundlingover multiplexing.

There are several standardized protocols for concatenatingmultiple VoIP packets into a single frame. Any protocol that canbe used for multiplexing can also be used for bundling. The ef-ficiency depends on which protocol layer the packet aggrega-tion is performed in, since only the layers below that layer willadd a single header per aggregated frame. The VoIP protocolstack of eHSPA is shown in Figure 7 and LTE in Figure 8. Pack-ets can be bundled at the RTP layer according to RFC 4867[25][26]. This is the most efficient bundling alternative; how-ever, multiplexing at the RTP layer is not possible since the RTPlayer is located at the UE. Packets can be multiplexed at theUDP layer according to 3GPP TS 29.414 [27] and TR 29.814[28]. Multiplexing can also be achieved according to the TMux[29] protocol, which can be implemented at the IP layer; hence,this solution is referred to as IP layer multiplexing. Further-more, multiplexing can also be implemented by adding a Point-to-Point Protocol (PPP) [30] layer to enable PPP multiplexing(PPPmux); the offered bandwidth gain is similar to that ofTMux; therefore, it was not separately investigated.


55

Figure 7. eHSPA protocol stack for VoIP.

Figure 8. LTE protocol stack for VoIP.

In eHSPA systems, due to the soft handover traffic in the up-link direction, there is significant cross traffic between theeHSPA Node Bs over the Iur interface. The load of this trafficcan also be reduced by applying multiplexing on the Iur inter-face; this was also investigated in our simulations. As bundlingintroduces higher extra delay, it is not feasible to apply bun-dling on the Iur interface.

4.1.3 Performance of Bundling and Multiplexing in aneHSPA Network

The performance of bundling and multiplexing for VoIP was in-vestigated in an eHSPA system. For details on the simulationsetup, readers are referred to Publication II. In the investigatedscenario, the Node Bs were each served by a 2 Mbit/s backhaullink that was a bottleneck. The transport efficiency increase dueto bundling can be seen in Figure 9, and multiplexing in Figure10. In these figures, it can be seen that the load on the backhaullinks is less if bundling or multiplexing is used, as these tech-niques decrease the redundant overhead. The most efficienttechnique is RTP bundling, as it is performed at a higher pro-tocol layer than the other methods, thus enabling more proto-col headers to be shared.

EthernetIP

EthernetIP

UDPGTPIP

EthernetIP

UDPGTP

IPPDCPRLCMACPHY

PDCPRLCMACPHY

UDPRTP

IP

Airinterface

UEeHSPA Node B

Gnuser plane

RouterGGSN

Remote entities

VoIP

UDPRTP

IP

VoIP

EthernetIP

EthernetIP

UDPGTPIP

EthernetIP

UDPGTP

IPPDCP39

RLCMAC39

PHY

PDCP39RLC

MAC39PHY

UDPRTP

IP

Airinterface

UEeNode B

S1user plane

RouterSAE-GW

Remote entities

VoIP

UDPRTP

IP

VoIP


56

Figure 9. Average link usage if bundling is used, including a reference casewithout bundling. The differences in throughput are the gains due to theefficiency improvement. Simulated with 20 VoIP connections.

Figure 10. Average link usage if multiplexing is used, including a referencecase without multiplexing. The differences in throughput are the gainsdue to the efficiency improvement. Simulated with 20 VoIP connections.

The simulations investigated the maximum number of con-nections the bottleneck backhaul links can support. To evaluatethis, the quality of voice connections was calculated on theMOSc scale [18][19][20]. When the bottleneck link becomescongested, packets will be queued in the IP buffers, and someof them will be dropped. This clearly disrupts the perceivedquality and the phone calls will become unintelligible. The

IP router eHSPA Node B link

0

200

400

600

800

1000

Nor

mal

2pac

kets

3pac

kets

4pac

kets

2pac

kets

3pac

kets

4pac

kets

2pac

kets

3pac

kets

4pac

kets

IP bundling UDP bundling RTPbundling

Max. no. of packets bundled

Avg

.Eth

erne

tthr

ough

put

[kbp

s]

IP router eHSPA Node B link

0

200

400

600

800

1000

Nor

mal

5ms

10m

s15

ms

5ms

10m

s15

ms

5ms

10m

s15

ms

5ms

10m

s15

ms

IP mux UDP mux IP + Iur mux UDP +Iur mux

Multiplexing timer value

Avg

.Eth

erne

tthr

ough

put

[kbp

s]


57

voice quality on the MOSc scale is a good indicator of the max-imum number of supported VoIP connections, and it also con-siders the effect of the added delays. The simulations were re-peated with different numbers of VoIP connections. The effectof UDP level bundling is shown in Figure 11, the maximumnumber of packets bundled together was simulated with differ-ent values. While the transport is more efficient when morepackets are bundled together, the largest performance differ-ence is between the cases with only two packets bundled to-gether and no bundling at all. The effect of UDP level multiplex-ing is shown in Figure 12. In this case, the maximum time apacket may be delayed was set to different values; the averagenumber of packets multiplexed together is proportional to thisparameter. It can be seen that multiplexing is far more effectivethan bundling at the same protocol layer. Furthermore, whenthe number of VoIP connections is sufficiently large to congesta link, then the frequency of VoIP packet arrivals is also high;therefore, a sufficient number of packets can be multiplexed to-gether even in the case of a short maximum added delay. Figure13 compares the efficiency of bundling performed at differentprotocol layers. It is clear that the high level bundling per-formed at the RTP layer is by far more efficient than bundlingperformed at lower protocol layers. Figure 14 compares the ef-ficiency of different multiplexing techniques, and reveals thatmultiplexing performed at a higher layer is more efficient. Ad-ditionally, multiplexing the soft handover traffic on the Iur in-terface that connects Node Bs is also found to improve perfor-mance.


58

Figure 11. Voice quality according to the number of parallel VoIP connec-tions on a 2 Mbit/s link when UDP level bundling is used. A referencecase without bundling is also shown. The backhaul link can support moreVoIP connections if bundling is used.

Figure 12. Voice quality according to the number of parallel VoIP connec-tions on a 2 Mbit/s link when UDP level multiplexing is used. A referencecase without multiplexing is also shown. The backhaul link can supportmore VoIP connections if multiplexing is used.

2

3

4

20 22 24 26 28 30No. of VoIP conns. / I-HSPA Node B

Ave

rage

MO

ScNormal

2 packets

3 packets

4 packets

2

3

4


Ave

rage

MO

Sc

Normal

5ms

10ms

15ms


59

Figure 13. Comparison of bundling performed at different protocol layers.Voice qualities according to the number of parallel VoIP connections ona 2 Mbit/s are shown. A reference case without bundling is also shown.RTP bundling, performed at a higher protocol layer, enables the mostparallel connections.

Figure 14. Comparison of different multiplexing techniques. Multiplexing maybe applied only on the gateway–Node B interface, or also to soft hando-ver traffic. Voice qualities according to the number of parallel VoIP con-nections on a 2 Mbit/s are shown. A reference case without multiplexingis also shown. The best performance is achieved by higher level multi-plexing on both interfaces.

Based on these eHSPA results, it can be stated that while RTPlevel bundling offers the highest bandwidth savings, the bestbandwidth–delay trade-off can be achieved with UDP levelmultiplexing, especially if multiplexing is extended to traffic onthe Iur interface. VoIP consumes considerable bandwidth whenthere are many parallel connections in parallel, thus in the mostrelevant cases, multiplexing will be efficient even if the VoIPpackets are delayed for only a few milliseconds.

2

3

4


Ave

rage

MO

Sc

Normal

IPbundlingUDPbundlingRTPbundling

2

3

4


Ave

rage

MO

Sc

Normal

IPmultiplexingUDPmultiplexingIP + IurmultiplexingUDP + Iurmultiplexing


60

4.1.4 Adaptive Multiplexing Methods

The efficiency of multiplexing with a certain protocol dependson two parameters: the number of packets that are multiplexedtogether, and the time these individual packets are delayed.Generally, multiplexing more packets together results in higherretention times; thus, there is a conflicting requirement. Aspreviously mentioned, the most common method for determin-ing the number of packets to be multiplexed together is settinga timer; this ensures that the maximum delay is limited. How-ever, as our research has shown, this method is suboptimal. InPublication III, superior algorithms have been proposed thatadapt to the VoIP traffic. These algorithms have also been pa-tented [2].

Average Number MethodLet us assume that packets arrive independently and randomlyat the multiplexing entity, according to a Poisson process. Inthis case, if, during its operation, the multiplexing entity waitsa further dt time for more packets to arrive, the number of ex-pected new packet arrivals at the multiplexing entity is propor-tional to dt. If there are n packets already waiting in the bufferof the multiplexing entity, then a further retention of dt willcause the n number of packets to be delayed by dt. Therefore,the total retention time of all the packets is proportional to∫ndt, and is the shaded area in Figure 15.


61

Figure 15. Operation of the multiplexer in the case of the average numberbased method.

The average number method proposed by the author statesthat the multiplexer should delay the packets in its bufferlonger if there are fewer packets there, as this causes less totalpacket delay. In contrast, it should send out the multiplexedframe sooner if there are more packets queued in its buffer, asin this case any further delay would cause more total packet de-lay. The network operator should choose an average delay pa-rameter; based on the observed traffic, the multiplexing entitycontinuously calculates the average number of packets that ar-rives in the specified time duration. When the number of pack-ets buffered in the multiplexer reaches this average number ofexpected packets, then the packets are to be multiplexed to-gether and sent out. Although this method changes only mini-mally the average time the multiplexer waits to assemble aframe, the distribution of these delays will be such that individ-ual packets will be delayed the least. As a further benefit, themultiplexed frame is sent out immediately when the last packetarrives, without any unnecessary waiting as in the case of thetimer-based method.


62

Predictive MethodAnother adaptive method, which the author has proven withsimulations to be efficient, utilises the inherent fact that thepackets of a VoIP connection arrive at regular intervals; in thecase of an AMR connection, a packet is generated periodicallyevery 20 ms. This method cannot be applied when this assump-tion is not true. Based on the previous packet arrivals, the adap-tive multiplexer can predict when future packets will arrive.The predictive multiplexing algorithm keeps track of a timeline,illustrated in Figure 16. Each time a speech frame arrives, thepredictive multiplexing algorithm assumes that in 20 ms an-other packet will arrive, and marks the time slot on the timelinethat is 20ms from the moment of the arrival. This algorithmalso starts a timer when the first packet of a new frame arrivesat the multiplexer. However, instead of always waiting for thetimer to expire, the algorithm identifies the last marked timeslot on the timeline preceding the timer expiry, the transmis-sion time is set to the end of this time slot. This means that ifthere are no more packet arrivals expected before the expira-tion of the timer, then there is no gain in waiting further, andthe frame is sent out earlier.

Figure 16. Principle of the predictive method.

Based upon their size, the algorithm can distinguish speechpackets from silence frames, which in the case of AMR codingarrive every 160 ms. When a silence frame arrives, it is assumedthat the next silence frame will arrive in 160 ms.

Note that this prediction is not perfect; there can be jitterwhich causes irregular packet arrivals, and voice connections


63

keep changing their states between speech transmission and si-lence frame sending. However, if the prediction of the algo-rithm is false, then the only harm done is that the retentiontime for one frame is a few milliseconds more than the optimal.

Stochastic MethodThe third investigated method is a variant of the average

number method which uses a soft threshold for sending out themultiplexed frame. If the number of buffered voice packetsreaches the target value, then the frame is sent out only with agiven probability p.

4.1.5 Evaluation of the Adaptive Algorithms in an LTENetwork

Similarly to the performance of VoIP bundling and multiplex-ing, the performance of adaptive multiplexing algorithms hasbeen quantified. However, they were evaluated in an LTE sys-tem. The backhaul architecture of eHSPA is designed to beidentical to that of LTE to enable a seamless evolutionary tran-sition. Both eHSPA and LTE have an all-IP architecture that isflat, there are no intermediary nodes between the NodeB/eNodeB and the gateway except for routers or switches.

The topology of the simulations is visualized in Figure 17. Inthis topology, three eNodeBs are connected to a router which isserved by a 10 Mbit/s link, which is the bottleneck link. Thesimulations were repeated with several different numbers ofVoIP users per eNodeB. For further details on the simulationsetup, readers are referred to Publication III. In the following,only the DL direction is shown.


64

Figure 17. LTE topology considered in the simulations, the protocol stack isalso shown.

In the simulations, the parameters of the adaptive algorithmswere set to values which corresponded to a similar averagenumber of packets multiplexed together, hence the bandwidthgain of the different algorithms was similar. This enabled thecomparison of the efficiency of the algorithms by simply com-paring the packet retention times, as presented in Figure 18. Itcan be seen that the added delays of the average number andpredictive methods is lower than those of the static multiplex-ing method. This is evidence that supports the claim that thesetwo adaptive multiplexing algorithms are superior to the com-mon timer-based method, therefore proving them to be benefi-cial innovations. The predictive method outperforms the aver-age number method; however, it has to be noted that it is ap-plicable only if the packets are generated periodically. Althoughthe stochastic method proposed by a co-author was interestingto investigate, the results showed that it was a failure, as its per-formance is worse than that of the static multiplexing method.


65

0

5

10

15

5 10 15 20

Del

ay[m

s]

Number of VoIP connections / eNB

Average numberStochasticPredictiveStatic

Figure 18. Average multiplexing delay of the adaptive algorithms, accordingto the number of parallel VoIP connections per eNodeB. The referencecase is the timer-based, static multiplexing. Note that the average num-ber and predictive methods outperform the static multiplexing.

To clarify the benefit of the adaptive algorithms, the distribu-tion of the delays experienced by packets is plotted in Figure 19.An advantage of the average number method is that a signifi-cant fraction of the packets is not delayed at all. When the lastpacket that is necessary to reach the target number of packetsarrives, the multiplexed frame is sent out immediately, and thelast packet is not delayed at all. Its disadvantage is that somepackets will be delayed longer than the strict limit of the timer-based method, when the average number of packets per multi-plexed frame is the same. The retention times in the case of thepredictive method are consistently less than for the static,timer-based method due to the predictive logic, which preventsunnecessary delays. As the predictive method considers 1 mstime slots, it delays many packets less than one millisecond,which means that these packets are delayed only until the endof the current time slot. This also means that no packets aresent forward immediately.


66

0

20

40

60

80

100

0 5 10 15 20

[%]

Multiplexing delay [ms]

Average numberStochasticPredictiveStatic

Figure 19. Cumulative density function of the multiplexing delay distributionfor 5 VoIP connections per eNodeB. Note that the average number andpredictive methods outperform the static multiplexing.

4.1.6 Summary

To summarise, bundling and multiplexing can save bandwidthon the backhaul links. If networks are to shift to the all-IP ap-proach, then circuit switched voice calls have to be replaced byVoIP connections. Despite data traffic being an ever increasingpart of the total global traffic, the volume of conventional phonecall traffic is still growing [31]. VoIP is a very inefficient proto-col due to the excessive packet overheads; therefore, there is anopportunity to possibly even double its transport efficiency.

The primary advantage of this is that if the backhaul link hasinsufficient capacity to serve as many VoIP connections as theair interface enables, then with bundling or multiplexing, thislimitation can be overcome, and no new phone calls will be re-jected due to the backhaul. While in most scenarios such an oc-currence is rare due to the proper dimensioning of the backhaullinks, there may be some cases where backhaul capacity is lim-ited. One such scenario is when at the site of a legacy base sta-tion, a newer generation Node B is installed, and the backhaullink is not upgraded. Another such scenario is the case offemtocells. As seen in the previous chapter, femtocells mayhave serious issues with the maximum number of phone callsthey can serve, and this problem can be solved with bundlingor multiplexing.


67

A decrease in the physical layer data rate of VoIP connectionsis beneficial even if their combined data rate is below the ca-pacity of the backhaul. The capacity freed this way can be usedby other traffic types. Therefore, in its typical use case, bun-dling and multiplexing increases the throughput of parallelTCP connections.

Naturally, bundling and multiplexing have the drawback ofadding to the end-to-end delay of the network. An exception tothis is when multiple packets are already transmitted or re-ceived synchronously over the air interface, such as when bun-dling is used on the air interface [23][24]. The simulations haveprovided evidence that supports that the added delay does notcause any noticeable degradation of perceived voice quality. Onthe contrary, when the backhaul link has insufficient capacityto serve the VoIP connections, then bundling or multiplexingimproves the perceived voice quality. As has been described,multiplexing causes less latency increase than bundling, there-fore, based on the results, preferring multiplexing over bun-dling is recommended.

The scientific novelties published in Publication III are theadaptive multiplexing algorithms. The two methods, whichhave been named the ‘average number’ and ‘predictivemethod’, offer better added delay to transport efficiency ratiosthan basic, timer-based multiplexing. When the arrival of pack-ets is periodical, as is the case for the currently used voice co-decs, we recommend the use of the predictive method, other-wise the average number method is recommended. When im-plementing multiplexing on a backhaul network, implementingone of these adaptive methods requires only a very slight effort.The only operational cost of these algorithms is a slightly in-creased calculation complexity. Therefore, when employingmultiplexing, the benefits of selecting an adaptive method faroutweighs its costs, even if the offered decrease in packet delaycompared to static multiplexing may not be noticed by the endusers.


68

4.2 Coordinated Multipoint Backhaul for 4G

Long Term Evolution Advanced (LTE-A) is an enhancement ofLTE. It introduces several new key features, the most importantbeing carrier aggregation, relay nodes, new Multiple InputMultiple Output (MIMO) options, and Coordinated Multipoint(CoMP). While in LTE, one carrier can have at most 20 MHz ofbandwidth, carrier aggregation allows the use of up to five non-contiguous carriers, for a total maximum bandwidth of100 MHz. This offers far higher data rates, and therefore, LTE-A meets all the IMT-Advanced requirements for 4G, thus beingappropriately called 4G. However, some operators now refer toLTE-A as 4G+, though this is not a technical term. Relay nodes,also known as self-backhauled evolved Node Bs (eNBs), do notneed separate backhaul, and provide increased coverage withminimal deployment effort and cost. In contrast, CoMP intro-duces new, very strict backhaul requirements, which this studyinvestigated in detail with simulations.

This section presents the different CoMP variants, their back-haul requirements and a novel method to decrease the back-haul requirements. The simulational studies also led to a valu-able scientific improvement for which the international patentis pending [3]. This section is based on an unpublished manu-script, its submission to a scientific forum has been delayed.

4.2.1 Introduction to Coordinated Multipoint (CoMP)

To satisfy the ever-increasing user demand for bandwidth, thecontinuous improvement of spectral efficiency is necessary.While approaching the theoretical limits of channels within asingle-cell, attention has yet again turned towards mitigatinginterference caused by neighbouring cells. In legacy cellularnetworks, inter-cell interference has been regarded as an ever-present limit on the overall capacity of the network. However,Long Term Evolution Advanced (LTE-A) systems propose toovercome this limitation by introducing a group of techniquesthat offer to decrease, eliminate or possibly even harness inter-cell interference by coordinating the operation of multiple sites.This promising feature which includes different levels of coop-eration is referred to as Coordinated Multipoint (CoMP), see


69

Figure 20. DL CoMP is referred to as CoMP transmission, anduplink UL CoMP is referred to as CoMP reception, though bothoffer significant gains, the two impose different challenges andrequire dedicated solutions. Since a user’s uplink signal inher-ently reaches the receiver antennas of multiple cells, the partialexploitation of this signal was already introduced in the softhandover feature of third generation systems. Downlink CoMPwas first introduced in fourth generation LTE-A systems[32][33], thereby imposing new challenges, including the re-quirement for fast backhaul communication among cooperat-ing transmission sites, which was the focus of the study.

Figure 20. CoMP technique categories

While the transmission of a single UE inherently reaches mul-tiple eNBs in UL CoMP, DL CoMP necessitates multiple trans-mitters to coordinate over the backhaul [34] prior to transmis-sion. Additionally, DL CoMP introduces a requirement foreNBs to have detailed and up to date Channel State Information(CSI) [32] of the mobile channels. This CSI, and possibly evenuser plane data has to be shared among transmitting nodesover the backhaul. This imposes strict latency and capacity re-quirements for the backhaul, which is not met in many LTE de-ployments.

An overview of the different standardized topological scenar-ios [33] is given by Lee et al. [35], see Figure 21.

CoordinatedMultipoint

Intra-site Inter-siteintra-eNB Inter-eNB

DownlinkUplink

CoordinatedScheduling /Beamforming

JointProcessing

Dynamic CellSwitching

JointTransmission

CoordinatedScheduling

JointProcessing

Dynamic CellSwitching

JointReception


70

Figure 21. Intra-site, inter-site intra-eNB and inter-eNB CoMP scenarios.

If the cooperating cells are the sectors of the same eNB, thenit is referred to as intra-site CoMP. This scenario has no extrabackhaul requirements as it requires no communication amongdifferent nodes; however, it fails to address the issue of inter-ference at the cell edges between different sites, where spectralefficiency is the worst.

If the antennas participating in the cooperative transmissionare at geographically separate locations, then it is referred to asinter-site CoMP; it offers a huge potential in increasing celledge spectral efficiency. Though the cooperating transmittingantennas may be at different sites, they do not necessarily be-long to different eNBs. A scenario where one eNB controls mul-tiple Remote Radio Heads (RRH, sometimes remote radiounits) forming a Distributed Antenna System (DAS) [36] is aninter-site intra-eNB scenario. This eNB may also be referredto as the baseband unit, baseband hotel, or equipment hotel.The architecture can be called Cloud Radio Access Network (C-RAN), or baseband pooling [37]. The RRHs may be low powerpico cells or high power macro cells. The low power RRHs mayeach have different cell IDs or may share the same cell ID,thereby in effect forming multiple small cooperating cells orone large cell. This is a topology which is different from alreadydeployed legacy infrastructures. It requires high capacity con-nections with guaranteed low latency between the eNB and theRRHs, which are referred to as fronthaul. Such a topology eas-ily enables any level of CoMP cooperation, as it is resolvedwithin a single eNB.

RemoteRadioHead

eNBeNB

Inter-site intra-eNBCoMP

Intra-siteCoMP

Inter-eNB CoMP


71

Inter-site inter-eNB CoMP is the cooperation among indi-vidual eNBs. This scenario offers improved cell edge perfor-mance at possibly any cell edge since it could be used to coor-dinate any neighbouring eNBs [39]. The eNBs are intercon-nected with the X2 interface, which has no capacity or latencyguarantees [40]. CoMP demands low latency coordination,therefore, its applicability is limited in legacy network topolo-gies without dedicated solutions [34].

In the case of inter-eNB UL CoMP, the reception points arenot co-located; therefore, the received signals are sampled andforwarded to a single Evolved Node B (eNB). This implies ahigh data rate, thus requiring very high capacity links [38].Diehm and Fettweis [41] quantify the negative impact of back-haul signalling delay on the performance of the scheduling. Thecombined decoding at the eNB is performed based on all of thereceived signals. Since an UL Hybrid Automatic Retransmis-sion Request (HARQ) retransmission has to occur strictly 8 msafter the initial failed transmission, the necessity of a retrans-mission has to be signalled back to the UE strictly 4 ms afterthe initial failed transmission [42]. Whether a retransmissionis required can only be precisely determined after the combineddecoding. This unavoidably implies transporting large volumesof data within a strict delay requirement. Consequently, eitherthe HARQ process has to be delayed or the applicability of ULCoMP is limited in inter-eNB scenarios [38]. However, the in-tention of the study was to show that inter-eNB DL CoMP doesnot necessarily require the transfer of large amounts of datawithin a strict time frame.

Since CSI describes the momentary conditions of fluctuatingchannels, it applies only for a brief period. Therefore, in DL in-ter-eNB cases, this CSI along with control plane informationhas to be shared among the cooperating nodes via the backhaulnetwork with low latency. It is critical for the efficiency of CoMPtransmission that the CSI sharing procedure is completedwithin a few milliseconds, or preferably within one Transmis-sion Time Interval (TTI), which is 1 ms, as mentioned by Irmeret al. [39] and described by Biermann et al. [34]. Cili et al. [43]calculate that for a low mobility scenario the performance of the


72

selected variant of CoMP starts degrading after 5 ms of CSI de-lay, whereas in a high mobility scenario, CSI delays of over asingle millisecond cause severe degradation due to the shortchannel coherence times. Furthermore, as described later on,in some versions of DL CoMP, not only control plane signallingbut also the user plane data has to be synchronously availablefor transmission at multiple nodes; this may imply a large vol-ume of data traffic. The backhaul signalling procedure de-scribed by Choi et al. [44] requires first CSI and then a largeamount of data to be moved over the X2 which is then pro-cessed within a single millisecond. Biermann et al. [45] de-scribe the extent to which the topology of the backhaul linkslimits feasibility.

4.2.2 DL CoMP Overview

As specified by 3GPP [33], the concept of CoMP transmissionincludes three different levels of cooperation; these differ at thelevel of the radio interface cooperation and amount of infor-mation exchanged among the eNBs. These variants are de-scribed in detail by Sawahashi et al. [46] and Lee et al. [47].


73

Figure 22. CoMP alternatives: a) Coordinated Beamforming, b) Dynamic CellSelection, c) Joint Transmission.

Coordinated Scheduling / Coordinated Beamform-ing (CS/CB) is the most basic form of CoMP, see Figure 22.Coordinated scheduling is an optimized cross-cell schedulingwhich coordinates the scheduling decisions of neighbouringcells to decrease cross-cell interference. If the eNB has multipletransmit antennas, it aims to avoid the collision of transmitbeams from neighbouring cells, and is referred to as coordi-nated beamforming. CS/CB requires the exchange of only CSIand scheduling information among the cooperating nodes, thisis only a few Mbit/s of traffic. Therefore, even in the case of in-ter-eNB CoMP, the capacity requirement is usually within thelimits of the X2 interfaces of legacy deployments [38]. How-ever, the latency requirement stated by Irmer et al. [39] andDiehm and Fettweis [41] of a few milliseconds still limits its ap-plicability.

Dynamic Cell Selection (DCS), also known as dynamictransmission point selection, adapts the system to changing ra-dio conditions; it always ensures that the cell with the best mo-mentary radio conditions transmits to the UE (Figure 22). It isa very fast switching of the transmitting cell, much faster than

Fast selection

Data exchange

b)

CSI exchange

a)

Simultaneoustransmission

Data exchange

c)


74

legacy handovers. For this, the DL user data also has to be readyfor transmission at multiple neighbouring cells, but it is onlytransmitted from the cell with the most favourable radio chan-nel. A variant of DCS which intends to decrease inter-cell inter-ference is referred to as dynamic point blanking or dynamicpoint muting. In this variant, the cell which is not selected totransmit to a certain UE is also ordered not to transmit to anyother UE on the corresponding Physical Resource Blocks(PRB).

Joint Transmission (JT) aims to harness what was previ-ously regarded as harmful inter-cell interference (see Figure22). Multiple neighbouring cells transmit the same DL userdata on the same PRBs synchronously. Not only is the receivedsignal strength at the UE antenna increased by the superposedradio waves, but no harmful interference is generated by neigh-bouring cells. The signals can be transmitted and combinedover the air non-coherently or coherently [32], where coherentjoint transmission offers more gain, but requires more accurateCSI. A serious issue of inter-eNB JT, especially for coherent JT,is that it requires accurate time and frequency synchronizationof the cooperating eNBs, this synchronization requirement isdescribed by Jungnickel et al. and Bladsjö et al. [49]; its impair-ment effects are investigated by Manolakis et al. [50].

DCS and JP can be referred to by the common name JointProcessing (JP). Both require the DL user data to be presentat multiple nodes. In the case of inter-eNB CoMP, this may pos-sibly imply X2 traffic with multiple times the data rate of the S1interface, which connects the gateway node and the eNBs. Sucha traffic volume may overload the backhaul network. In JP, allthe DL user data has to be precoded based on fresh CSI. Thisimplies a strict latency requirement, since the CSI, which is re-ported in the UL to the serving eNB, has to be available at thetransmitting eNBs by the time of the synchronous transmis-sion. Except if the alternative method proposed by Papado-giannis et al. [51] is used where the UE reports CSI in the UL tomultiple eNBs simultaneously.


75

4.2.3 Methods for Decreasing the Backhaul Load of Inter-eNB DL CoMP

Due to the strictness of the backhaul demands of inter-eNBCoMP, there are several approaches for decreasing the amountof information that has to be exchanged.

Though inter-eNB CoMP is an effective method for increasingcell edge spectral efficiency, there is a limit to the number ofcell edges where CoMP can be applied. Since cooperation is lim-ited to cells within close proximity, the standard [32][33] spec-ifies that neighbouring cells should be configured into clusterscalled cooperating sets. This should be done considering theposition of the nodes, the geographic layout of the cells, and thetransport topology [34]. CoMP cooperation is allowed amongcells of the same cooperating set, but not between cells of twodifferent cooperating sets. The CSI reports sent by the UE con-stitute a signalling overhead in the uplink, this overhead and itspossible mitigation is described by Papadogiannis et al. [52];Jungnickel et al. [53] propose a method to reduce the CSI feed-back overhead with frequency selectivity. To keep this CSI feed-back overhead reasonably low, the UE should report the meas-urements of only the best radio channels, not for all of the cellsin the cooperating set; this is also specified in the standard[32][33]. This group of cells for which CSI is reported is themeasurement set, it is a subset of the cooperating set. As furtherspecified in the standard [32][33], in the case of JP, fromamong these cells are selected the cells that participate in thetransmission to the UE; these cells form the transmission set.

The number of cell edges between different cooperating setsis trivially less if larger sets are used. Larger cooperating setsimply both higher maximum node to node backhaul latency,and extra backhaul bandwidth consumed by CoMP. Theamount of information to be shared among the nodes increaseswithin a group according to the size of the group, additionally,it also has to be forwarded to more nodes [38]. Although thiscan be overcome with higher capacity backhaul links, it re-mains challenging to guarantee that the backhaul latency re-quirement is met by the furthest two communicating nodes;therefore, the size of the cooperating sets has to be limited [34].


76

One possibility to enable CoMP at every cell edge is to allowoverlapping cooperating sets, which implies overlapping cross-cell scheduling decisions. To allow this, Zirwas and Mennerich[54] propose that the radio resources should be divided amongseveral so-called cover shifts, thus each cooperating set sched-ules only a fraction of the available PRBs; this is a variant offractional frequency reuse. Overlapping cooperating sets ena-ble CoMP at every cell edge where there is a suitable transportconnection between the cooperating nodes, but the drawbackof this method is that it limits the scheduling mechanism to se-lect from only a fraction of the radio resources. A differentmethod is proposed for enabling overlapping cooperating setsby D. Hui [55], who determines precoding matrices accordingto an iterative negotiation among eNBs.

An efficient solution to enable CoMP at every relevant celledge is for the system to adaptively group cells into cooperatingsets instead of grouping the cells during network planning andconfiguration. This allocation should be based on the momen-tary radio conditions and traffic demands of the UEs, also con-sidering the backhaul limitations. Several adaptive algorithmshave been proposed. Marsh and Fettweis [38] provide a com-prehensive summary of both non-overlapping and overlapping,in addition to static and adaptive clustering. The optimal place-ment of the controller node of the cooperating set is investi-gated by Biermann et al. [56]. Biermann et al. [34] also investi-gate the feasibility of cooperating set configurations accordingto different backhaul network densities. Draxler et al. [57] andBiermann et al. [58] propose a method to reconfigure the coop-erating set allocation considering the bottlenecks in the back-haul capacity. A self-organizing adaptive clustering method isdescribed by Weber et al. [59]. Choi et al. [60] present a proce-dure that considers before every scheduling decision whetherthe cooperating eNBs have enough backhaul capacity to be in-cluded in the transmission set. Scalia et al. [61] propose that theextent of cooperating sets should be limited considering thebackhaul network to decrease energy consumption. Choi et al.[44] describe a possible backhaul signalling procedure for a dy-namic wireless clustering procedure.


77

Since JP aims to eliminate or harness inter-cell interference,it is most beneficial where there is strong signal reception fromneighbouring cells, which is at the cell edge [38]. DCS willswitch the transmitting cells only if the received signal strengthof one cell occasionally exceeds that of the other cell. When aUE is in the middle of a cell, close to the transmitting node, thenJT offers very little gain. When all factors are taken into ac-count, it can be concluded that CoMP JP should be applied onlyto the cell edge UEs, and not to the cell centre UEs due to thefollowing reasons: the extra overhead imposed by the CSI feed-back both on the air interface (and in the case of inter-eNBCoMP also the backhaul), and also due to the inaccuracy of thechannel estimations.

Zhang et al. [62][63][64] describe a method with dynamicswitching between JP and CS/CB. Relieving the system of serv-ing every UE with inter-eNB CoMP JP greatly decreases the ex-tra backhaul load and the backhaul capacity requirement. Seifiet al. [65] calculate that in the case of very limited backhaul ca-pacity single cell transmission may be preferable in a two UEscenario.

It is also possible to employ so-called rate splitting [66][67]where the data stream of even a single user is divided betweenJT and single cell transmission to scale down the backhaul loadof a single UE.

4.2.4 Data Sharing for CoMP JP

There are several possibilities for distributing the DL data in aCoMP JP [68]. In the centralized, distributed antenna systemapproach there is one centralized baseband unit (the eNB) andmultiple RRHs, all data is processed at the eNB. Such systemsare described by Heath et al. [36] and also by You et al. [69].When the eNB performs coordinated scheduling and data pro-cessing, it has all the necessary information, allowing optimaldecisions without the need for synchronizing data flows or co-ordination information exchanges over the X2 interface. It doesnot require significant changes to the LTE radio protocol stack;the DL output of the lowest protocol layer, the physical layer is


78

forwarded to the RRHs [70]; this is marked as “existing cen-tralized solution” in Figure 23. However, this architecture hasthe very serious drawback that regardless of CoMP, it requiresexcessive bandwidth on the fronthaul connections. Onemethod to connect RRHs to the eNB is with the use of radioover fibre where analogue signals are transmitted over a direct,dedicated fibre-optic cable. The feasibility of this was demon-strated by Diehm et al. [71] and in a field trial by Nagate et al.[72]. Another option is with the digital sampling of the signal,either in the time or frequency domain; one such standard isCommon Public Radio Interface (CPRI) sampling [73]. Thedrawback of such methods is that the data rate of the sampleddata stream is multiple times the data rate of the original datastream. Due to these very high capacity requirements, this datacan be transported only over fibre, gigabit Ethernet, or high ca-pacity microwave links, and also require dedicated links thatguarantee low latency delivery. Thus, a centralized architecturedoes enable CoMP without additional requirements; however,even without CoMP, it demands backhaul topologies which aredifferent from already deployed legacy topologies [38].

Figure 23. Data flows in different CoMP architectures.

Without the use of dedicated backhaul solutions, CoMP JTcan be enabled in limited backhaul capacity scenarios only by

Overhead

increasesasdataisconveyed

downtheprotocolstack

Before SchedulingAfter Scheduling

ServingeNB

OtherTransmitting

Node

PDCP processing

RLC processing

MAC processing

RLC processing

MAC processing

Forward error coding

Precoding

Transmission

Existingcentralizedsolution

Existingdistributedsolution

Proposedsolution

Lenient delayrequirement

Strict delayrequirement

Strict delayrequirement

UE

Forward error coding

Precoding

Transmission


79

using a distributed, inter-eNB architecture. Such an inter-eNBapproach was proven to be feasible in testbeds in Berlin andDresden [39]. The Dresden testbed is presented by Irmer et al.[74], and described also by Holfeld et al. [75][76] and by Jung-nickel et al. [77]. The field trials in Berlin are reported by Jung-nickel et al. [78]. However, the critical issue with this distrib-uted architecture is the coordination of data flows at the coop-erating eNBs to ensure that different eNBs can transmit thesame data symbols on the same PRBs. In this approach, data isreceived from the gateway over the S1 interface at the servingeNB of each UE; the PDCP, RLC and MAC layer processing isperformed at this eNB. This processed, but not yet precodeddata is forwarded to the other cooperating eNBs over the X2interface in addition to the necessary precoding matrices. Thecooperating eNBs receive this data from the serving eNB, per-form the precoding, and transmit it synchronously. This ismarked as “existing distributed solution” in Figure 23. Oka-mawari et al. [79] implemented a similar architecture in a realsystem, where MAC Payload Data Units (PDU) are forwarded.

These distributed solutions require much less transport ca-pacity than centralized inter-site intra-eNB architectures; theamount of the forwarded data is of the same order of magnitudeas the DL user data received over the S1 [38]. During the criticaltime frame—from the reception of the fresh CSI and the trans-mission—not only is there control information exchangeamong cooperating eNBs, but also large amounts of data is for-warded over the X2 interface. If, due to some transient conges-tion, this data arrives late, synchronous multipoint transmis-sion fails and performance degrades. Legacy LTE backhaul net-works in general lack the necessary capacity for this.

Another proposal to enable a distributed architecture is tosynchronize the data processing of the eNBs as proposed byJungnickel et al. [48]. This same proposal is also considered byBrueck et al. [80]. According to this, the DL user data is sent toevery cooperating eNB before transmission, either by multi-casting the data on the S1 interface from the gateway node toevery eNB, or by forwarding the data from the eNB that re-ceived the data (the serving eNB of the corresponding UE) tothe other eNBs over the X2 interface. The PDCP, RLC, MAC and


80

physical layers would have to process the data in a synchro-nized manner to generate identical physical layer frames.Though this method generates less backhaul traffic, the draw-back of such an eNB synchronization is its high complexity.

Brueck et al. [81] briefly mention an option of storing the userdata in all transmission points; after this, only control infor-mation is sent from a master node to slave nodes, but it alsostates that it is outside the scope of the paper to investigate thebenefits of this approach and does not provide an enablingmethod.

4.2.5 Splitting the Data and Control Traffic on the X2 In-terface

The patent of the solution presented in this section is pendingat the time of writing this dissertation [3]. This solution is basedon the idea that as long as user data is queued at an eNB, thenits forwarding to another eNB is not delay critical. Therefore,the proposed algorithm shares the user plane data among theeNBs well before it is scheduled for transmission, while a copyof it is waiting in queue at a congested air interface. Thus, dur-ing the delay critical time frame—just prior to transmission—itrequires only a small amount of control plane data to be for-warded. This allows the low data rate, delay critical signallingand control traffic to be transported with higher priority, incontrast to the bulk of the traffic which is tolerant to delays onthe X2. If only a small portion of the X2 traffic is sent withhigher priority, it is easy to ensure that this higher priority traf-fic is not delayed due to congestion; thus, only the delay toler-ant, lower priority traffic may experience congestion. This ena-bles CoMP JT even in the case of limited backhaul capacity andcongestion on the X2 interface.

The main benefit of CoMP JT is its ability to increase the max-imum throughput of the air interface. If the air interface is notfully loaded, then CoMP JT is unnecessary. In contrast, whenthe air interface is congested, then data packets experiencequeuing delays at the eNBs, and the backhaul is not the bottle-neck; this enables the use of the proposed CoMP JT algorithm.While CoMP JT mitigates congestion on the air interface, itgenerates an extra load on the backhaul. If, due to the improved


81

air interface efficiency and additional backhaul load, the net-work bottleneck is shifted to the backhaul, then the algorithmcan adapt to this and decrease the ratio of traffic that is sentwith CoMP, and find the optimal balance depending on theavailable backhaul link capacities.

4.2.6 Technical description of the data sharing algorithm

Our solution does not require the complicated synchroniza-tion of the RLC or MAC layers of the eNBs as described by Jung-nickel et al. [48]. Instead, the cooperating eNBs simultaneouslygenerate the same transmission data based on previouslyshared, identical information. When a serving eNB receives DLuser data, it performs PDCP layer processing (sequence num-bering, header compression, integrity protection, and cipher-ing) and adds the PDCP header to form a PDCP PDU. If thereis congestion on the air interface, these PDCP PDUs are imme-diately forwarded to the other transmitting eNBs. This traffichas a large volume, but it is not delay critical. When the othertransmitting eNBs receive these PDUs, they buffer them andacknowledge them to the serving node. This provides an up-to-date measurement of the X2 Round Trip Time (RTT). The sys-tem considers that there is congestion on the air interface if thisRTT is less than the queuing delay at the PDCP layer. In thiscase, the data exchange over the X2 will be completed by thetime the corresponding data will be scheduled for transmission.Otherwise, if there is free air interface capacity (or if the ac-knowledgment does not return in time), then the eNB revertsback to single cell transmission, as the extra air interface capac-ity offered by CoMP would not be utilized.

When the eNBs receive the measured CSI from the UEs theyserve, they immediately forward this CSI to the cooperatingeNBs in addition to any information necessary for scheduling,including data queue lengths and scheduling metrics. This con-trol plane signalling message is sent over the X2 with higherpriority than the bulk data for fast delivery. When each eNB re-ceives this message from all the other eNBs in the cooperatingset, they perform a synchronous cross-cell pre-scheduling.


82

Based on the same shared information, they synchronously al-locate common radio resources (PRBs) to eNBs. These PRBscan be assigned to single-cell transmission or CoMP transmis-sion. Afterwards, each eNB performs proper scheduling, allo-cating only those PRBs that were assigned to that eNB. For itsscheduled data, the eNB performs RLC layer processing (in-cluding segmentation and concatenation, and adding of theRLC header), MAC layer processing (including resource map-ping and adding of the MAC header) and calculates the precod-ing matrices based on the received cross-cell CSI. This pro-cessing is done only at one eNB, which is the serving eNB of theUE, hence, it does not require the synchronization of RLC andMAC layers. Instead, the serving eNB gathers and forwards tothe other transmitting eNBs all the information necessary forthe replication of the generated MAC PDUs. This informationis the RLC and MAC headers, specific references to the seg-ments of previously shared PDCP PDUs, and added control in-formation. This is accompanied with the calculated precodingmatrices and the resource mapping information. Since this in-formation does not contain the actual user data, only specificreferences to it, this is a low number of bytes, therefore ena-bling it to be forwarded in a single higher priority packet,thereby ensuring timely delivery. When the other transmittingeNBs receive this information, based on the previously receivedPDCP PDUs, they can replicate identical MAC PDUs. This pro-cess is marked as “proposed solution” in Figure 23. The opera-tion of the physical layers is assumed to be synchronized, theidentical MAC PDUs are jointly transmitted after forward errorcoding and precoding at a predetermined TTI. During the delaycritical time frame, this signalling protocol involves only onemessage exchange over the X2 on both directions, the mini-mum for duplex exchange and coordination, thus providingmaximum tolerance for the X2 delay (see Figure 24). Unlike inthe UL, in the DL direction the HARQ process can be delayed;due to this, DL HARQ does not impose further delay require-ments.


83

Figure 24. Signalling procedure.

4.2.7 Further benefits of the proposed method

In this architecture, occasional queuing delays on the X2 in-terface do not impair the CoMP operation, therefore, an adap-tive algorithm can be employed that adjusts the data rate of theCoMP traffic to the available bandwidth on the X2. The eNBshave up-to-date measurements of the X2 RTT, therefore ena-bling detection of congestion on the X2 interface. In this case,the eNBs can intervene to decrease the volume of CoMP traffic.As mentioned previously, the largest CoMP JT gains are offeredfor the cell edge UEs, while UEs at the cell center have muchless to gain from CoMP JT. Thus considering that CSI feedbackis not ideal, and considering the overheads introduced byCoMP JT, it is optimal to serve cell edge UEs jointly from mul-tiple cells, while serving cell center UEs with single cell trans-mission. The serving eNB should regularly select the transmit-ting cells based on the received CSI and momentary traffic con-ditions. When the eNB detects a transient congestion on the X2interface, it should adaptively decrease the number of UEsserved with JT in order to resolve the X2 congestion, andshould only increase the number of UEs served with JT if theX2 RTT is sufficiently low. The X2 RTT value should be keptlower than the queuing delay at the PDCP buffer of the servingeNB; in this case the data sharing procedure over the X2 will becompleted by the time the corresponding DL user data is sched-uled.

Though the previous description explained the applicabilityof JT, the same method can be applied to DCS, with the onlydifference that even though the user data will be available atmultiple eNBs, only one of them will transmit to the UE.

time

CooperatingeNB

ServingeNB

UE

Data for UE via S1

ACK

PDCP PDUs

CSI

CSI

Synch.info

Joint Transmission

PDCPprocessing

Cross-cellScheduling,MAC & RLCprocessing

Pre-scheduling

MAC PDUReplication

Critical time frame

X2 interface

CSI


84

Though JT has better potential gains than DCS, in some sce-narios, only DCS is applicable. DCS is not possible if the accu-racy of the CSI feedback is insufficient, or the reference clocksof the eNBs are not sufficiently accurate for the required levelof time or frequency synchronization.

4.2.8 Evaluation

The concept of the proposed procedure was conceived when theauthor was implementing a CoMP simulator and performingsimulations on an NS-2 platform. The entire S1 and X2 protocolstack, and functions necessary for simulating CoMP JT was im-plemented in detail. This included the entire operation of theGTP, UDP, IPv4 with IPsec, and Ethernet layers for the S1 userplane interface; as well as the PDCP, RLC, MAC and physicallayers for the air interface. The simulated network topologyconsisted of seven eNBs with three cells each. Cooperating setswere statically allocated and were comprised of two or threeneighbouring cells. There were at most two hops between eNBsin the same cooperating set. The backhaul links formed a treetopology and their base delay was assumed to be 0.1 ms. Thelast mile backhaul links were simulated with multiple datarates. The simulated traffic was TCP file downloads.

The air interface was based on 3GPP TR 25.814 [82], with2×10 MHz bandwidth. The mobility model was random walkwith inter-site handover procedures. The channel quality wassimulated separately for each user considering distance loss,shadow fading, fast fading and penetration loss. Downlink in-ter-cell interference was calculated from the propagation loss,momentary transmit power, resource allocation of interferingeNBs, and the CoMP mode of operation.

Two CoMP architecture alternatives were implemented, eachwith detailed signalling exchanges over the backhaul network.One alternative was based on literature[39][74][75][76][77][78], where the precoding is distributed.The other was the new X2 delay tolerant alternative proposedby the author. In both alternatives, the synchronous transmis-sion of data occurred 1 ms—one TTI—after the CSI was meas-


85

ured; this imposed a strict time limit for the signalling proce-dure on the backhaul. The HARQ mechanism was also updatedto support CoMP JT.

The results demonstrated that the proposed method canachieve significant improvements in cell edge performance,even if there is very limited backhaul capacity, as illustrated inFigure 25. However, if ample backhaul capacity is available,then the gain depends only on the technique used on the airinterface. The simulations proved that the proposed method issuperior to existing solutions when considering the backhaulrequirements. It was confirmed that it imposes far less strictcapacity and latency requirements on the backhaul, while ena-bling larger cooperating sets.

Figure 25. Performance improvement of simulated methods, according tothe available backhaul capacity. The orange bars stand for a referencecase where CoMP JT is not used. The purple bars denote a referencecase where the precoding is distributed; in this case, system perfor-mance is degraded due to the congested backhaul links. The red barsrepresent the delay tolerant adaptive method described in this disserta-tion. Note that the proposed method offers significant cell edge perfor-mance improvement, even if the backhaul links have low capacities.

4.2.9 Summary

CoMP is a key feature of future LTE-A networks, as it increasesspectral efficiency. This chapter described the wide range ofpossibilities to coordinate the DL transmission of multiple cells

0

50

100

150

200

250

30 40 50 60 70 80 90 100

Ave

rage

FTP

thro

ughp

utof

celle

dge

UE

s[k

bit/s

]

Last mile link capacity [Mbit/s]

Without CoMP JT

Reference casewith distributedCoMP JT

Proposed X2delay tolerant JTmethod


86

with CoMP. The backhaul requirements of CoMP vary on aneven broader scale. While intra-site CoMP has no backhaul re-quirements at all, and CS/CB requires only the transfer of CSI,the backhaul capacity requirements of JP may even be severalorders of magnitude higher than the air interface throughput.In inter-site CoMP JP, the benefit of increased cell edge spec-tral efficiency has to be weighed against the cost of upgradinglegacy backhaul networks to meet the capacity and latency de-mands.

Therefore, a method has been proposed which separates thedelay critical messages from the bulk of the X2 traffic. It dis-tributes the user data over the backhaul efficiently with mini-mal overhead. It also enables inter-eNB CoMP JP over the X2with no minimum backhaul capacity requirement as it adaptsto the available capacity. This alleviates the requirement ofcostly upgrades or the redesigning of existing backhaul net-works to enable CoMP JP; only simple software upgrades at theeNBs are necessary. It requires low latency X2 links, as everyother version of inter-eNB CoMP. In contrast to other versions,it keeps the information exchange sequence over the X2 as sim-ple as possible (the time critical message traverses the X2 anda reply is returned), and requires only that the X2 RTT be lowerthan the fluctuation time of the air interface (which is unavoid-able). Therefore, the proposed method maximises the delay tol-erance of CoMP JP by as much as is theoretically possible. It isa completely distributed architecture that is compatible withany JP technique used on the air interface and with adaptiveclustering.

87

5. Backhaul for 5G (Beyond 4G)

5G is a very broad term, it has not yet been specifically defined.At the time the work was started on 5G, the term used for it was‘Beyond 4G’; as in 2012, 4G was the marketed new generationtechnology. Since then, almost everything new has been re-ferred to as 5G. Amidst all the confusion about the term, it isalready clear that 5G intends to cover a wide range of use cases,some of which are not covered by current 4G technology. Eachof the previous mobile system generations had one main targetuse case. 1G introduced mobile telephony, 2G improved mobiletelephony, 3G offered mobile internet, and 4G improved theperformance of mobile internet. Instead of having a single goalin mind, 5G aims to be an architecture which is a symbiotic in-tegration of multiple techniques in order to cover any possibleuse case. 5G ambitiously aims to provide 10 000 times morecapacity, offer multi-gigabit data rates, connect 100 times moredevices, have less than 1 ms latency, and increase battery life toup to ten years [83]. This would enable extremely broadbandmobile internet access, support for all devices comprising theInternet of Things (IoT), and also guarantee reliable and lowlatency service for cyber-physical systems. As the latency of 5Gradio access is comparable to the delay of nerve signals betweenthe brain and the hand, thus rendering it imperceivable, it isreferred to as ‘tactile internet’.

Therefore, to be specific, the topic of this chapter is ‘new ra-dio’, which refers to the radio interface of 5G. The next threesections focus on three aspects. One is cognitive radio, whichallows more efficient spectrum allocation. Another is multi-hop

Backhaul for 5G

88

backhaul for millimetre wavelength mobile networks. And fi-nally, the third is the sensitivity of millimetre-wave backhaul torain fading.

The results described in this chapter can also be found in Pub-lications IV, V, and VI.

5.1 Cognitive Radio

The QoSMOS project [84], or “Quality of Service and MObilitydriven cognitive radio Systems” project was an EU fundedFramework 7 Integrated Project; it began in 2010 and ran forthree years. It involved 14 industrial and academic partnersfrom across Europe and one from Japan. One of the academicpartners was the Budapest University of Technology and Eco-nomics, on whose behalf the author participated in the work.The objective of the project was to design a platform for spec-trally efficient radio access for future Software-Defined Net-works (SDN). One of the disseminations of the project was Pub-lication IV, which presents the proposed architecture, with anemphasis on QoS management.

5.1.1 Cognitive Radio and Television White Spaces

In recent years, spectrum shortage has emerged as a majorchallenge for telecommunications. However, when measuringspectrum occupancy, it can be seen that the majority of the ra-dio spectrum is underutilized most of the time [85]. Therefore,spectral efficiency can be multiplied with cognitive radios,which intend to utilize momentarily unused spectrum. They ap-ply dynamic spectrum management. Instead of being allocateda certain frequency band, they monitor the radio spectrum attheir location, and always select unused frequency bands fortheir operation. Software-defined radios provide the hardwarerequired for such a broadband operation. While cognitive ra-dios offer to end the current spectrum shortage by offeringmuch more efficient spectrum utilization, it remains a chal-lenge to correctly determine the momentarily available freespectrum.

Backhaul for 5G

89

While there are Industrial, Scientific and Medical (ISM) radiofrequency bands that any device may use, cognitive radios alsoaim to operate in licenced bands where other devices have anexclusive licence. The devices with the exclusive spectrum li-cence are called the incumbent devices, or primary devices. Thecognitive radios without the exclusive licence are called the op-portunistic, or secondary users. The users or operators of theincumbent technology pay for their undisturbed operation intheir allocated spectrum. In contrast, the opportunistic, cogni-tive users detect the activity of the incumbent users, and keepselecting a frequency band where they are certain that they willnot interfere with the operation of the incumbent users; inother words, they use only locally and momentarily vacant fre-quency bands. For this, they lack an exclusive spectrum licence;instead, they have a license to operate as secondary users. Re-gardless of this, the primary users are charged for the spectrum.This is a form of spectrum sharing, where the spectrum isshared according to unequal terms.

There is a large amount of spectrum where this sharing canbe allowed. The most notable frequency band is the televisionwhite space in the Very High Frequency (VHF) and lower UltraHigh Frequency (UHF) bands. Television broadcasting hasbeen allocated large swathes of very valuable spectrum. Ana-logue television broadcasting is being replaced with digitalbroadcasting; due to its efficiency, this has freed up a so-calleddigital dividend which can be refarmed for other purposes. De-spite this, television broadcasting still occupies more spectrumthan it utilises. Furthermore, television broadcasting is verypredictable, the television stations broadcast constantly at fixedfrequencies. This makes its spectrum ideal for cognitive radios.

5.1.2 Proposed System Concept and Use Cases

The QoSMOS project undertook the contradictory task ofproviding QoS for opportunistic users. By definition, opportun-istic users cannot be guaranteed the use of any frequency band;therefore, it is possible that there is momentarily no availablefrequency at all; this is the Achilles’ heel of any cognitive radionetwork. Thus, it is challenging at best to promise guaranteed

Backhaul for 5G

90

service quality. A further complicating factor was that the userswere considered to be mobile. It was envisioned that the pro-posed cognitive system would be sufficiently intelligent andsensitive to always find some appropriate spectrum.

A diverse set of use cases was defined. One scenario was thecellular extension in whitespace. In this scenario, a cellular net-work, such as LTE, could use the proposed cognitive function-alities to occasionally use additional spectrum. One benefit of-fered for cellular networks is the additional operating band-width. Another benefit is that the cellular network would haveaccess to lower frequencies, such as bands allocated for televi-sion. Radio propagation conditions are better at lower frequen-cies. In rural areas, this increases the range of base stations andextends coverage. Furthermore, due to the better penetrationcapabilities, this mitigates indoor coverage holes.

Another use case scenario was the cognitive femtocell, whichis very similar to femtocells described in Chapter 3. The differ-ence is that a cognitive capability can improve the performanceof the femtocell. The isolation provided by the penetration lossof buildings ensures the availability of ample high frequencyspectrum. The cognitive femtocell principle could be used inconjunction with both 3G/4G networks and Wi-Fi.

A third considered scenario was the cognitive ad-hoc net-work. Such ad-hoc networks would typically be formed in thecase of events with many participants which are limited in du-ration and area. Examples of this include reliable communica-tion in case of a larger emergency, a network established for abusiness meeting, and traffic offloading in the case of a largecrowd, such as a sporting event. The user devices would auto-matically establish a mesh of device-to-device links to provideservice even in the most unexpected or demanding conditions.

5.1.3 Cognitive Management

In the proposed system, the core of the cognitive functionalitiesis two entities: the Cognitive Manager for Spectrum Manage-ment (CM-SM), and the Cognitive Manager for Resource Man-agement (CM-RM). The CM-SM is responsible for keepingtrack of the available spectrum. It acquires context information

Backhaul for 5G

91

from regulatory databases, a common repository database, andspectrum sensing measurements. The CM-RM is in charge ofallocating the available radio resources. Both entities provideincumbent protection, thus it is implemented on two differentlevels for maximum protection. Additionally, there is a separatespectrum sensing entity.

Due to the multiple different use cases, the system has no uni-form topology; instead, multiple different topologies are de-fined. The nodes at which the entities are located also dependson the scenario. Thus, entities such as the CM-SM, CM-RM,and spectrum sensing entity may be implemented in differenttypes of nodes; and it is also possible that they are present inmultiple types of nodes in the same scenario. This is a form ofNetwork Function Virtualization (NFV), and software-definednetworks. The side of the air interface or backhaul on whichthese critical functionalities should be implemented is not de-termined. However, the architecture of the core network wasnot considered in the project.

Therefore, the distinction between the air interface and back-haul is blurred. While the cellular extension and cognitivefemtocell scenarios typically imply wired backhaul, the cogni-tive ad-hoc network employs multiple hops of the air interfaceto serve as backhaul, which connect directly to the core networkvia a gateway.

Based on these aforementioned principles, an architecturewas designed to provide QoS with only opportunistic spectrumaccess. While in theory it should accomplish this, the verifica-tion of this was not part of the objectives. Further details can befound in Publication IV and in the other disseminations of theproject [84].

5.1.4 Summary

This section presented the operating principle of a cognitiveradio system. Cognitive, dynamic spectrum access offers the re-use of spectrum that has been exclusively licenced to anothertechnology, without interfering with the primary users. Thispromises an increase of spectral efficiency, and the more wide-

Backhaul for 5G

92

spread availability of more valuable lower frequencies. In addi-tion to this, the goal of the work was to provide QoS guaranteesto all primary and secondary users.

5.2 Multi-hop Wireless Mesh Backhaul for 5G

Over the past decades, the volume of mobile traffic has beenincreasing exponentially, and is expected to keep continuouslyincreasing [31]. To increase mobile broadband capacity 10 000fold, 5G aims to increase spectral efficiency by a factor of ten,available spectrum by a factor of twenty, and cell density by afactor of fifty [86]. The spectrum allocated for cellular networkshas thus far been increasing incrementally. However, there isample free spectrum at very high frequencies, such as the so-called millimetre wavelength (millimetre-wave) spectrumabove 30 GHz [88]. While millimetre-waves offer more band-width than all of the spectrum below 6 GHz combined, theywere previously considered unsuitable to mobile networks.This was proven incorrect when propagation measurements[87] revealed that despite the minimal diffraction, millimetre-waves can be used for mobile networks, allowing multi-gigabitper second data rates.

Publication VI presents a concept for an ultra-dense multi-hop millimetre-wave network. As this system incorporates anew technological generation of air interface, and offers far su-perior data rates than any current mobile system, it is a fifthgeneration (5G) system in the truest sense. Using a simulatorwritten for this purpose, Publication VI demonstrates that theuser level performance of such a network is superior to that ofany current system. The efficiency of multi-hop backhaul iselaborated. It shows that interference is less of an issue at suchfrequencies. Furthermore, it also asserts that the effect of occa-sional link outages is manageable. For further details on thesimulation setup, readers are referred to Publication VI.

Backhaul for 5G

93

5.2.1 Millimetre-Wave Antennas

It is a common misconception that higher frequencies alwayssuffer greater free space attenuation. According to the Friistransmission equation, the free space attenuation is:

2

4

rGG

PP

rtt

r

where Pt and Pr are the transmission and reception powers, re-spectively, Gt and Gr are the transmitter and receiver antennagains, λ is the wavelength, and r is the distance between the an-tennas. From this equation, it may be concluded that shorterwavelengths correspond to less received signal power. How-ever, this holds true only if the considered antennas gains areconstant, which is the case when omnidirectional or short di-pole antennas are used. This is because at shorter wavelengthsthe effective width of a dipole antenna is narrower, and the ef-fective aperture area of an omnidirectional antenna is smaller.If aperture antennas with fixed aperture sizes are considered,then their gain is higher at higher frequencies:

24AG

where A denotes the effective antenna aperture size. When sub-stituting this into the previous formula, the free space path lossin the case of aperture antennas is:

22rAA

PP rt

t

r

where At and Ar are the effective aperture sizes of the transmit-ter and receiver antennas, respectively. One conclusion thatcan be drawn from this is that path loss is less at higher fre-quencies if aperture antennas are used. Another conclusion isthat at very high frequencies, such as millimetre-waves, direc-tional antennas are necessary to overcome link budget issues.Fortunately, at such high frequencies, even very small sized an-tennas can provide high gain. While aperture antennas can eas-ily be used for fixed links, they are unsuited to mobile links. Asolution to this problem is the beamforming antenna.

Beamforming antennas are phase controlled antenna arrays,which can be electronically steered to direct a narrow beam to-wards even a moving target. Their gain depends on the number

Backhaul for 5G

94

of antenna elements of which they are comprised. Their disad-vantage is that they require either one radio frequency phaseshifter or a digital/analogue converter for every antenna ele-ment; therefore, state of the art millimetre-wave beamformingantennas suffer from cost and power consumption issues. How-ever, as with all electronics, it is expected that over time theircost will decrease, and their efficiency will increase. It is ex-pected that even handsets will incorporate beamforming anten-nas [89].

As an alternative to beamforming antennas, beam steeringlens antennas could also be used at Access Points (AP). Suchantennas are comprised of an array of antenna elements behinda lens. Only one of the antenna elements is chosen for trans-mission. The radio waves originating from that antenna ele-ment are focused into a narrow beam by the lens. The directionof the beam can be steered by selecting another antenna ele-ment for transmission, which is located in a slightly differentposition behind the lens. Similarly, a receiver antenna selectsone antenna element for reception. Such an antenna system islow cost as it requires only one digital/analogue converter andno phase shifters. Additionally, its gain only depends on thesize of the lens. Its drawback is that the angle range in which itcan steer its beam is limited.

5.2.2 Multi-Hop In-Band Backhaul

Despite the use of directional antennas, the maximum range ofmillimetre-wave mobile links is expected to be short. Sulymanet al. [90] estimate it will be 220 m, and MacCartney et al. [91]expect it to be 200 m, in urban environments. This necessitatessmall cell sizes and ultra-dense deployments, which in turn en-ables even higher capacity. However, this limits millimetre-wave 5G to urban scenarios; therefore, millimetre-wave net-works will be complemented by another access technology op-erating at lower frequencies, which will provide rural coverage.

Due to the very large number of millimetre-wave APs, theircost and ease of deployment will be crucial. While the cost ofthe AP device is expected to fall, the labour costs of deploying itwill remain a serious issue, as well as the costs of providing it

Backhaul for 5G

95

with a backhaul connection. Laying down a fibre-optic cable inevery urban street would be prohibitively expensive. Conven-tional wireless point-to-point links would require their aper-ture antennas to be precisely aligned, and realigned every timethey are slightly nudged. Furthermore, they cannot be mountedon masts that sway in the wind. A solution to these problems isto provide wireless backhaul with beamforming antennas.

Since the mobile access links are served by highly directionalbeamforming antennas, these same antennas can also be usedto establish backhaul links. Thus, since an AP will have multiplesectors to provide omnidirectional coverage, the AP can estab-lish backhaul links in any direction without being mechanicallyaligned, and without the need for separate hardware for thebackhaul links. The backhaul links could also use the same fre-quency as the mobile access links, in other words, they could bein-band. APs could automatically establish such in-band wire-less backhaul links among themselves. By allowing multiplesuch hops for a backhaul connection, only a fraction of the ac-cess points would require expensive fibre-optic connectivity.Whether this in-band operation causes interference issues wasone of the focuses of the study, and will be elaborated in follow-ing sections.

It would be very easy to deploy such a multi-hop, self-back-hauled network where the APs essentially act as relay nodes. Asthe cells would be small, the APs would be placed below rooftoplevel, such as on lamp posts, traffic signs, and the sides of build-ings. As they would not need a fixed backhaul connection, theywould only need to be connected to a source of electricity, suchas a solar panel, or the electricity cable inside a lamp post.

5.2.3 Millimetre-Wave Propagation

There are several issues with the radio propagation propertiesof millimetre-waves. At such high frequencies, radio waves dif-fract only minimally. Despite this, Non-Line Of Sight (NLOS)radio links can be established over propagation paths where thesignal is reflected or scattered. Such links suffer from high pathlosses and are limited to very short ranges. In urban environ-ments, having Line Of Sight (LOS) to the AP is only probable if

Backhaul for 5G

96

the distance to the AP is short; this is another factor that de-creases cell sizes. Shadow fading is expected to be very high[92]. Due to the narrow Fresnel zone combined with minimaldiffraction, millimetre-wave links are very easy to obstruct.Even a human hand can block and disconnect a link to an AP.The only possibility in case a link is blocked is to hand over theUE to another AP. This will result in frequent handovers, andnecessitates somewhat overlapping cells. The interference is-sues of overlapping cells will be mitigated by the directionalityof the UE antennas.

Due to the high penetration losses, it will be challenging toprovide indoor millimetre-wave coverage with outdoor APs, ifat all possible. Indoor coverage can be provided with indoorAPs, but even an indoor AP can provide coverage only for a lim-ited area within a building.

To further increase the data rate of a radio link, spatial mul-tiplexing—a form of Multiple Input Multiple Output (MIMO)—can be used in combination with beamforming. However, thisimplies further multiplying the number of antenna elements. Amore efficient technique is to employ dual-polarized antennaarrays, where the beamforming antennas are comprised of a setof horizontally polarized, and another set of vertically polarizedantenna elements. This enables communication over two par-allel channels that are separated by polarization. This is possi-ble due to the high cross-polarization discrimination [93][94].

5.2.4 Combining In-Band and Dedicated Backhaul Links

Combined mobile access and in-band multi-hop backhaul is avery low cost solution, as the access and backhaul shares thesame spectrum, radio frequency (RF) devices and antennas.However, it has the disadvantage that the backhaul and accesslinks compete with each other for radio resources. Thus, back-haul links can only be allocated a fraction of the radio re-sources, which proportionally decreases their capacity. In thesystem considered in Publication VI, the network is Time Divi-sion Duplex (TDD). In this TDD system, not only do the AP sec-tors switch between serving DL and UL connections in a given

Backhaul for 5G

97

time slot, but may additionally keep switching between allocat-ing the time slots for access or backhaul. Thus, any backhaullink is active during only half of the time slots at most, andtransmitting in one direction of the backhaul link in a quarterof the time slots at most.

If separate RF devices and antennas are issued for backhaullinks, then such backhaul links do not need to share resourceswith mobile access links. On the one hand, this increases thecapacity of both the wireless backhaul and access links. On theother hand, it also greatly increases the cost of the device andthe deployment costs. Therefore, in the study presented in Pub-lication VI, the possibility of combining in-band backhaul linkswith dedicated backhaul links was considered. In the simula-tions, a few selected in-band bottleneck links were replacedwith dedicated links. Results showed that this greatly increasedthe system performance, while increasing the cost of the net-work only proportionally to the number of dedicated links.

The study also included an option of increasing the numberof fibre-optic connections to the wireless network. In exchangefor the additional cost, this increases network capacity, as it dis-tributes traffic more evenly on the network. Simulations re-vealed that capacity increase of an additional fibre-optic linkfrom the core network to the wireless network is similar to thecapacity gain of replacing a few in-band links with dedicatedlinks. Therefore, it may be more cost-effective to combine in-band links with dedicated links, instead of deploying multiplefibre-optic links and only in-band links. A further effect of de-ploying multiple fibre-optic connections is that the number ofwireless hops will also be reduced. The simulations revealedthat the performance of TCP connections is independent of thenumber of wireless hops. This is due to the very short Trans-mission Time Interval (TTI), which was 0.1 ms, and the possi-bly very long queuing delays, which could reach up to severalhundred milliseconds.

5.2.5 Interference

One of the objectives of the simulations was to determine theeffect of interference in an in-band multi-hop network. The

Backhaul for 5G

98

novelty of Publication VI is that while many studies have cal-culated the level of interference for simple traffic models, ourstudy calculates interference while considering a realistic, TCPbased traffic model, combined with a multi-hop, in-band net-work. When the purpose-built simulator calculates the interfer-ence on every wireless link, in each time slot, it takes into ac-count, the number of bytes of data transmitted on each individ-ual interfering link in that time slot. Other considered detailsinclude specific antenna patterns, transmission powers, geo-metric distances, path losses, TDD scheduling, and angles ofazimuth. The system employed adaptive modulation. To alwaysselect the appropriate modulation number and coding rate, asophisticated link adaptation algorithm was implemented,which required an estimate of the interference of each link inthe next time slot. Based on the simulated performance, it wasfound that it was efficient to estimate the interference level ofthe next time slot to be equal to the maximum interference ob-served in previous time slots.

The simulations were repeated with different beamformingantenna configurations. The number of beamforming antennaelements at the APs and UEs was varied, a higher number ofantenna elements corresponds to higher antenna gain and lessinterference. It was considered that each user downloads onefile during the simulation, the size of this file was either 10 MBor 100 MB. The base delay of the fibre-optic connection be-tween the wireless network and the server on the core networkside was also simulated with different values. User level perfor-mance results are shown in Figure 26, Figure 27, Figure 28, andFigure 29.

Backhaul for 5G

99

Figure 26. File download times according to different antenna array sizes,which correspond to different in-band interference levels. Simulated with10MB files. A hypothetical reference case without interference is alsoshown. Note that downloads are quicker when larger antenna arrays areused.

Figure 27. File download times according to different antenna array sizes,which correspond to different in-band interference levels. Simulated with100MB files. A hypothetical reference case without interference is alsoshown. Note that downloads are quicker when larger antenna arrays areused.

00.20.40.60.8

11.21.41.61.8

2

nointerference

8x8, 4x4 4x4, 2x2 2x2, 1x1

[sec

]

AP and UE antenna array size

Average File Download Time, 10 MB files

Delay to server:0.5ms

Delay to server:5ms

Delay to server:50ms

02468

101214161820

nointerference

8x8, 4x4 4x4, 2x2 2x2, 1x1

[sec

]


Average File Download Time, 100 MB files


Delay to server:5ms


Backhaul for 5G

100

Figure 28. Download goodputs according to different antenna array sizes,which correspond to different in-band interference levels. Simulated with10MB files. A hypothetical reference case without interference is alsoshown. Note that download speeds are higher when larger antenna ar-rays are used.

Figure 29. Download goodputs according to different antenna array sizes,which correspond to different in-band interference levels. Simulated with100MB files. A hypothetical reference case without interference is alsoshown. Note that download speeds are higher when larger antenna ar-rays are used, and that the system is more congested than in the casewith smaller file downloads.

The conclusion that can be drawn is that even in the case ofin-band backhaul, the effect of interference will be less of anissue in millimetre-wave networks than in previous generationsystems. This is due to the high directivity antennas which mit-igate interference. Results have also shown that in such a multi-hop network, interference is only an issue for the bottleneck

050

100150200250300350400

nointerference

8x8, 4x4 4x4, 2x2 2x2, 1x1

[Mb/

s]


Average File Download Goodput, 10 MB files


Delay to server:5ms


050

100150200250300350400

nointerference

8x8, 4x4 4x4, 2x2 2x2, 1x1

[Mb/

s]


Average File Download Goodput, 100 MB files


Delay to server:5ms


Backhaul for 5G

101

links, as there is excess link capacity when less loaded links se-lect lower data rate modulation and coding. Since only a fewbottleneck links can be fully loaded, the interference in much ofthe network has little impact on user level performance.

5.2.6 Link Outage

Millimetre-wave cells are expected to be small, and millimetre-waves diffract only very slightly. Therefore, in dense urban ar-eas, most millimetre-wave APs will not be placed above rooftoplevel; instead, they will be placed in positions with a good viewof a street canyon, or on a corner covering two streets. An idealplacement would be to deploy an AP on every second lamp post.However, placing APs at lower heights increases the chancethat a moving object blocks a wireless link. This can cause com-plete link failure due to the narrow Fresnel zones and inabilityto diffract around the object. To counter this, APs can be orga-nized into a partial mesh topology with redundant wirelessbackhaul links. This allows traffic to be diverted to the remain-ing links in case one backhaul link is blocked. This topology alsoenables efficient AP to AP communication within the mesh forfast handovers within the mesh; handovers are expected to befrequent between millimetre-wave small cells.

The study also simulated the effect of a blockage on the wire-less mesh. In the simulations, if no link is blocked, then packetsare routed on pre-calculated primary paths which have theminimum number of hops. If a primary path is blocked, then apre-calculated secondary path is used, see Figure 30. However,when the path is reselected, there may be some packets intransit on the wireless mesh. In the simulations, two alterna-tives are compared; in one, these packets are simply droppedwhen they arrive at the blocked link. In the other, the in-transitpackets are also rerouted.

Backhaul for 5G

102

Figure 30. Simulated topology with the occasionally blocked link. Normally,packets are routed over the paths marked with solid arrows, but whenthe link is blocked, then these packets are routed over the paths markedwith dashed arrows.

The simulator considered a worst case scenario for theblocked link to evaluate the rerouting mechanisms. The linkwas repeatedly available for one second, and then experiencedan outage for the following second. While this represents morefrequent outages than expected, it is useful as a reference worstcase scenario.

The simulated goodputs of the rerouted connections areshown in Figure 31 and Figure 32. The results indicate that thererouting mechanism works as planned, and the remainingwireless mesh links handle the rerouted traffic well; the userlevel performance drop is close to the ratio of lost link capacity.These results prove that the occasional loss and appearance ofa link does not cause serious problems, except for the decrease

800 m

800

m

Primary path Secondary path

Backhaul for 5G

103

of the network capacity, assuming that every AP remains con-nected. This justifies connecting the APs into a partial mesh to-pology in order to ensure robustness.

In contrast, if multiple links fail simultaneously, some APsmay be disconnected. The only ways to avoid this are to in-crease the connectivity of the partial mesh topology or deploymore reliable links. However, it is less likely that multiple linksare blocked at the same time, unless the link failures are due toheavy precipitation, as discussed in the following section.

Figure 31. Comparison of the goodputs of FTP downloads that are rerouted.The investigated scenarios are: a reference case where all links are al-ways available; a case where there are occasional blockages, and pack-ets arriving at the blocked link are dropped; a case where there are oc-casional blockages, and packets arriving at the blocked link are rerouted;and a reference case where the link is always blocked. The downloadedfiles are 10 MB in size.

0

50

100

150

200

250

300

350

reliable links blockages& drops

blockages& rerouting

less links

[Mb/

s]

Link blocking

Average File Download Goodput, 10 MB files,flows on possibly blocked path


Delay to server:5ms


Backhaul for 5G

104

Figure 32. Comparison of the goodputs of FTP downloads that are rerouted.The investigated scenarios are: a reference case where all links are al-ways available; a case where there are occasional blockages, and pack-ets arriving at the blocked link are dropped; a case where there are oc-casional blockages, and packets arriving at the blocked link are rerouted;and a reference case where the link is always blocked. The downloadedfiles are 100 MB in size.

For further details on the simulations, readers are referred toPublication VI.

5.2.7 Ultra-low latency applications

5G is expected to provide not only extremely high data rates,but also to support applications which require ultra-low latencypacket delivery. Some Machine Type Communication (MTC),or interactive real-time video applications may demand end-to-end delays of no more than a few milliseconds. Internet serviceoffering such low latencies is often referred to as “tactile inter-net” [95]. Ultra-low latency can only be guaranteed for trafficthat is scheduled with higher priority than bulk TCP traffic, asthe natural operation of TCP is to increase the connection datarate until buffers are filled, which unavoidably introducesqueuing delays. The timely delivery of VoIP packets is not anissue in 5G networks, as VoIP has far more lenient latency re-quirements than the delay critical applications 5G is designedto serve.

In the later phases of the simulations, in addition to TCP, ma-chine type traffic was also considered. The simulated applica-tion generated downlink UDP packets at regular intervals,

0

50

100

150

200

250

reliable links blockages& drops

blockages& rerouting

less links

[Mb/

s]

Link blocking

Average File Download Goodput, 100 MB files,flows on possibly blocked path


Delay to server:5ms


Backhaul for 5G

105

which were scheduled at every node with strictly higher prioritythan background TCP connections. The average end-to-end de-lay of the machine type traffic is shown in Figure 33. Resultsshow that due to the strictly higher priority scheduling, the la-tencies remain very low, thus revealing that the internet serviceis tactile. Furthermore, these delays are mostly independent ofthe volume of lower priority background traffic. This is due tothe large bandwidth; despite the short TTIs, multiple packetscan fit into a single time slot. Therefore, the second fragment ofa packet can be transmitted along with newly arrived higherpriority packets in the same time slot. However, in the rare casethat the high priority traffic in itself congests the network, QoScollapses.

Figure 33. Average delay of machine type traffic according to the MTCpacket size and background traffic. The average downlink delay from therouter of the wireless mesh network gateway to the UE is shown. TheMTC traffic is scheduled with strictly higher priority than the backgroundFTP based file downloads. In this scenario, there are 20 MTC, and op-tionally 20 FTP users. Each MTC user generates ten packets in everymillisecond.

5.2.8 Summary

Millimetre-wave cellular systems offer mobile broadbandspeeds that were previously deemed unattainable. However,due to the inherent challenges implied by such high frequen-cies, a radically new approach is necessary. Beamforming an-tennas have to be used at both the APs and the UEs to compen-

0.90

0.95

1.00

1.05

1.10

10 Bytes 100 Bytes

mea

nde

lay

[ms]

MTC packet size

Average high priority downlink packet latency

in an unloadednetwork

with 100 MB filedownloads in thebackground

Backhaul for 5G

106

sate for the unfavourable link budget. To overcome the cost is-sues of the inherent cell size limitation, multi-hop in-bandbackhaul mesh networks are a key enabler.

In the study presented in Publication VI, a packet-based sim-ulator was implemented on the NS-3 simulator platform tosimulate user level performance in the case of a TCP-based traf-fic model.

The simulations quantized the detrimental effect of interfer-ence on user level performance. The remarkable result is thateven in the case of in-band backhaul, the effect of interferenceis far less than that in previous generation systems. This is dueto the highly directed beamforming antennas. In most scenar-ios, the system will be noise limited instead of interference lim-ited.

While in-band backhaul links are low cost, and are very easyto deploy, dedicated RF devices for backhaul links offer highercapacities. The results presented in this paper show the benefitsof using a combination of the two techniques. A gradual taper-ing of capacity is recommended from the fibre-optic link to theedges of the mesh topology.

Due to the higher probabilities of a link being obstructed,simulations also considered the occasional outage of a criticallink. Results prove that even in the worst case scenario, due tothe fast rerouting, the effect on TCP traffic is tolerable, as longas only a single link is lost.

Furthermore, results also show that ultra-low end-to-end de-lay can be guaranteed for delay critical traffic types providedthat their packets are scheduled with strictly higher priority.

In the future, we intend to study handovers within the meshnetwork. Validating the simulation results by comparison toreal measurements will only be possible in a few years, as at thetime of writing, 5G has not yet been standardized.

Backhaul for 5G

107

5.3 The Effect of Rain Fading on 5G Backhaul

While the study presented in the previous section investigatedthe occasional loss of a single link on a millimetre-wave back-haul mesh network, it did not consider the possibility of multi-ple backhaul links failing simultaneously. The simultaneousfailure of backhaul links is possible due to bad weather. At suchfrequencies, radio waves are susceptible to rain fading, whichincludes sleet and snowfall. Fortunately, millimetre-waves canpropagate through fog, unlike free space optical links, whichare easily disconnected by fog.

While it is evident that the maximum capacity of a millimetre-wave mesh network is decreased by heavy rain, it is crucial toprevent the complete disconnection of APs and its associatedusers from the core network.

The work described in Publication V determines the proba-bility of the disconnection of a part of a millimetre-wave meshnetwork due to precipitation. The calculation is based on bothsimulations and measurement data.

5.3.1 Rain Fading Measurements

A rain attenuation measuring system has been operating in Bu-dapest, Hungary, since 1997. Using this system, the authors ofPublication V have processed several years of measurementdata. This measurement data in Publication V has been used tocalculate the Complementary Cumulative Distribution Func-tion (CCDF) of the precipitation attenuation on short millime-tre-wave links, see Figure 34. The considered link is 200 m, op-erates at 38 GHz, and is horizontally polarized. Note that theseresults only apply to one specific continental climate.

Backhaul for 5G

108

Figure 34. CCDF curve of the attenuation caused by rain calculated for a200 m link in Budapest, considering the weather of years 2003 to 2007.

5.3.2 Rain Fading Simulations

In addition to measurement data, a simulator was also used.The author implemented a simulator to calculate the probabil-ity of a part of a mesh network being disconnected due to rainfading. It calculated in dB/km the minimum rain attenuationlevel required to disconnect a part of the backhaul mesh.

The simulator deployed semi-randomly a set of nodes on a500 × 500 m area. One node, the wireless mesh network gate-way, also known as the sink node, was placed in the middle ofthe area. The wireless mesh network gateway was consideredto have a fibre-optic connection. After the random deployment,the connectivity of the network was investigated. Two nodeswere considered to be able to establish a link if they were within200 meters [91] of each other. It was ensured that every AP hada topological path to the wireless mesh network gateway node.After the topology was generated, the minimum rain attenua-tion level that causes link failure was calculated for every wire-less link. The link budget reserve of the wireless link was con-sidered to be proportional to the difference of the link lengthand maximum link length, as it was assumed that all wirelesslinks have the same transmission power and other parameters,except for the link length. An AP was considered disconnectedfrom the network when due to the rain attenuation no topolog-ical route remained between the AP and the wireless mesh net-work gateway. The simulations were repeated 10 000 times,

0 2 4 6 8 10 1210

-6

10-4

10-2

100

A [dB]

p

Backhaul for 5G

109

providing a ratio of the disconnected topologies, displayed inFigure 35 and Figure 36.

Figure 35. Probability that a part of the simulated mesh network becomesdisconnected, in function of the rain attenuation, simulated with 10nodes.

Figure 36. Probability that a part of the simulated mesh network becomesdisconnected, in function of the rain attenuation, simulated with 20nodes.

5.3.3 Measurement Data Combined with Simulation Data

In the following step, the measurement data is combined withthe simulation data. The measurement data determined theprobability of a certain level of rain fading, and the simulationdata determined the probability of a partial network disconnec-tion in the case of a certain level of rain fading. Therefore, theprobability of network disconnection can be calculated. Thenetwork disconnection probabilities for the continental climateare provided in Figure 37 and Figure 38.

0%

20%

40%

60%

80%

100%

0 10 20 30 40Rain attenuation [dB/km]

Network disconnection probability, 10 nodes

1 node disconnected

2 nodes disconnected


Average ratio ofdisconnected nodes

0%

20%

40%

60%

80%

100%

0 10 20 30 40Rain attenuation [dB/km]

Network disconnection probability, 20 nodes

1 node disconnected



10 nodesdisconnected

Average ratio ofdisconnected nodes

Backhaul for 5G

110

Figure 37. Probability that a part of the simulated mesh network graph be-comes disconnected in function of the rain attenuation exceedance prob-ability, simulated with 10 nodes.

Figure 38. Probability that a part of the simulated mesh network graph be-comes disconnected in function of the rain attenuation exceedance prob-ability, simulated with 20 nodes.

To clarify, Figure 37 reveals, for example, that one node is dis-connected at least 10-2 part of the time—which corresponds to3 days and 16 hours a year—in 20% of the random deploy-ments. Five nodes, meaning half of the network, is discon-nected for at least 10-3 part of the time—8 hours and 46 minutesa year—in 9% of the random topologies. For example, as seenin Figure 38, in a denser and thus more reliable 20 node sce-nario, one node is disconnected at least 10-2 part of the time—3days and 16 hours a year—in 7% of the random deployments.Half the nodes in the network, in this case ten nodes, are dis-connected at least 10-3 part of the time—8 hours and 46 minutesa year—in only 6% of the random topologies.

For specific details, readers are referred to Publication V.

10-610-510-410-310-210-10.1%

1%

10%

100%Network disconnection probability, 10 nodes

Ratio of time

1 node disconnected2 nodes disconnected5 nodes disconnectedAverage ratio ofdisconnected nodes

10-610-510-410-310-210-10.1%

1%

10%

100%Network disconnection probability, 20 nodes

Ratio of time

1 node disconnected2 nodes disconnected4 nodes disconnected10 nodes disconnectedAverage ratio ofdisconnected nodes

Backhaul for 5G

111

5.3.4 Summary

This section used data from rain fading measurements span-ning multiple years and combined then with a simulation casestudy. From the results, it can be concluded, that though thereis a common belief that rain fading disrupts the reliability ofmillimetre wavelengths, a partial mesh network remains fullyoperational most of the time, even if resilience is provided onlyby rerouting in the backhaul network. However, at high rain in-tensities, the ratio of the disconnected nodes can be considera-ble. It has to be noted, that these results are applicable to a con-tinental climate, and are accurate only in Hungary, where themeasurements were located.

This same method has been used to calculate network discon-nection probabilities in a tropical climate, using measurementdata from Malaysia. Preliminary results show that unlike theconditions of a temperate climate, rain fading is a major issuein a rainy, tropical climate that will pose serious challenges. Weplan to publish a paper on this topic in the near future.

112

113

6. Conclusion and Discussion

Every research project presented in this dissertation consid-ered the highest level of technological development at the time.Since the work described in the first few chapters was con-cluded, the technology has evolved and changed. In 2016, it canbe seen which technologies became popular and the mistakeswhich were made in planning the following generation system.Some technologies, which were originally planned to be ubiq-uitous by this year, are still in the process of being graduallydeployed.

In contrast to the first few chapters, the success of wirelesstechnology generations described in the later chapters is still anopen question. The demand for ever higher wireless data rateshas been continuously growing in the past two decades, it is as-sumed that it will continue. The argument behind this predic-tion is that, if in the past, users were given a certain bandwidth,they always found some way of consuming it all, though notnecessarily efficiently. As the average price that can be chargedfor the delivery of a gigabyte of data or charged for a certaindata rate continuously decreases, the industry is forced to sellin ever larger quantities to maintain its revenue. Developingnetworks with higher data rates is in fact a means of decreasingthe so-called cost per bit. Due to this, it is highly likely that thedrive for higher data rates will continue for many more years,enabling completely new applications.

Conclusion and Discussion

114

6.1 Conclusions from Deployed Technologies

Time has proven GSM to be an excellent standard. It has com-peted with many other standards, and has become the mostpopular telecommunications standard in the world. There areapproximately 8 billion [96] GSM compatible Subscriber Iden-tity Module (SIM) cards in use today; this figure is more thanthe population of the entire planet, as it includes Machine toMachine (M2M) connections. All mobile phones that use latergeneration 3GPP technologies also feature GSM. At the time,its backhaul architecture which featured a BSC was a logicalchoice, even if the concept has been superseded.

The UMTS backhaul architecture proved to be short-lived.Just three years after the standardization of UMTS, the HSDPAstandard had to move important functionalities from the RNCto the Node B. This functionality shift continued with HSUPA,and the Iub interface became far more complicated than origi-nally anticipated. The separate RNC node was abandoned in I-HSPA and LTE, greatly simplifying these systems. On the onehand, in the case of GSM, it was beneficial to design the BTS tobe as simple as possible, and move most functionalities to theBSC, because this decreased the cost of the BTS. In the 1990s,a BTS device itself was very expensive and many more BTSsthan BSCs had to be deployed. On the other hand, by the time3G was deployed in the 2000s, due to the advance in computingtechnology, digital baseband processing was much less of a costissue, which did not justify in retrospect a separate RNC nodeand the additional complexity of having to standardize and de-velop an Iub interface. In many countries, when UMTS was firstdeployed, it was already the upgraded HSDPA version.

Femtocell deployment has proved to be easier when startingdeployment with the I-HSPA architecture. While the femtocellconcept has been incorporated into the popular “small cell”umbrella term, it has not yet become ubiquitous, despitefemtocells being available in dozens of countries. One of themajor reasons behind this is not technical, instead thus farthere is no consensus on the business model; femtocells arecurrently marketed according to all the possibilities describedin Chapter 3. It now seems that femtocells can be useful only in


115

the specific scenario where there is a wired internet connection,but no outdoor to indoor coverage anywhere within the cus-tomer premises. The average data rate of fixed data connectionshas also increased over the years; therefore, the maximumnumber of voice calls that can be served is less of an issue.

VoIP has been spreading gradually. Many operators havephased out their GSM CS core networks, with only GPRS re-maining. Voice over LTE (VoLTE), which is all-IP, has spreadto only a handful of countries to date. In some LTE networks,voice calls are simply dropped back to 3G or 2G, this is clearlyagainst the original design principles. The primary focus of theindustry currently is to provide reliable voice call service, whilethe efficiency of the underlying technology is secondary, asvoice represents only a fraction of the total traffic, but generatesmuch revenue.

6.2 Discussion on the Adoption of Future Technol-ogies

While the academic world has somewhat moved beyond re-searching CoMP, and the technology principles are well de-fined, the deployment of its more efficient versions is still a planfor the future. Inter-eNB CoMP has not received much atten-tion due to its demanding backhaul requirements. The inven-tion described in Section 4.2 lessens precisely these strict re-quirements. The version of CoMP that is under implementationis inter-site intra-eNB CoMP. Inter-site intra-eNB CoMP is veryeasy to implement in a C-RAN architecture. It is also one of thedriving forces behind the development of the C-RAN architec-ture. C-RAN enables even the most efficient and complicatedform of CoMP, namely JT to be implemented without any syn-chronization issues. On the one hand, a C-RAN structure de-creases the total cost of the devices, as RRHs are far less expen-sive than eNBs, and only a few BBUs are necessary. On theother hand, the fronthaul link between the BBU and each RRHis more expensive than a backhaul link. Thus, the extent towhich C-RAN will become popular, or the time of its deploy-ment is still undetermined.


116

Cognitive radio systems have already been promised a verylong time ago. However, due to the difficulties of perfect incum-bent protection, the incumbent users and regulators are reluc-tant to grant secondary access to expensive spectrum. Cognitiveradio is still researched in academia, and probably will be forthe next several years.

Millimetre-wave mobile communication has recently grownto be a very hot research topic. Thus far, no technical obstaclehas emerged that could prevent deployment, but its overall costand efficiency remains a much investigated question. The ex-tremely high capacities and data rates promised for 5G seemspossible only with either centimetre-wave massive MIMO ormillimetre-waves; therefore, there is a very serious demand forthis technology. Millimetre-wave multi-hop backhaul seems tobe the most convenient way of providing backhaul connectivity.Without multi-hop backhaul, deploying access points at the re-quired density seems excessively expensive; therefore, thereare currently very few opponents of multi-hop backhaul tech-nology. In-band backhaul is one cost efficient option; it caneven be combined with the other option of dedicated devicesand spectrum, which can enable higher capacities.

Rain fading is currently not a major concern, even for milli-metre-wave networks. This is also because new technologiesare often first deployed in temperate climates; in contrast, rainfading is more of an issue in extremely rainy tropical climates.

In the next decade, the future of the mobile telecommunica-tions industry will be undoubtedly determined by the successof 5G. Hence, the shape of things to come after 5G remains purespeculation.

117

References

[1] Á. Drozdy, Cs. Vulkán, L. Kőrössy. Integrált otthoni internet ésmobiltelefon szolgáltatás femtocellákkal. Híradástechnika,65(5-6):2-8, May 2010.

[2] Cs. Vulkán, A. Rákos, Z. Vincze, Á. Drozdy. Reducing Overheadon Voice Traffic. U.S. Patent No. 8,824,304. Washington, DC:U.S. Patent and Trademark Office. Also published asWO2011134527A1, US2013003557A1, JP2013516875A,JP5620518B2, EP2564569A1, CN102870390A, 2014.

[3] Cs. Vulkán, Á. Drozdy, Z. Vincze. Method for backhaul efficientinter-eNB Coordinated Multipoint transmission. Patentpublished as US20160037511, EP2974490A1,WO2014139588A1, 2016.

[4] E. Metsälä, J. Salmelin, eds. Mobile Backhaul. John Wiley andSons, 2012.

[5] H. Holma, A. Toskala, eds. WCDMA for UMTS – Radio Accessfor Third Generation Mobile Communications (3rd ed.). JohnWiley & Sons, 2004.

[6] H. Holma, A. Toskala, eds. HSDPA/HSUPA for UMTS – HighSpeed Radio Access for Mobile Communications. John Wiley &Sons, 2006.

[7] Nokia Siemens Networks. Simplified network architecturedelivers superior mobile broadband. White paper,http://networks.nokia.com/file/17371/i-hspa-profitable-wireless-broadband, 2010.

[8] Nokia Siemens Networks. The broadband world is flat. Whitepaper,http://networks.nokia.com/system/files/document/Profitable_mobile_broadband_with_I-HSPA.pdf, 2007.

[9] E. Metsälä, J. Salmelin, eds. LTE Backhaul: Planning andOptimization. John Wiley and Sons, 2016.

[10] H. Holma, A. Toskala, eds. LTE for UMTS – OFDMA and SC-FDMA Based Radio Access. John Wiley & Sons, 2009.

[11] Nokia. Nokia AirScale Cloud RAN. White paper,https://networks.nokia.com/solutions/airscale-cloud-ran,2016.

[12] L. T. W. Ho, H. Claussen. Effect of user-deployed, co-channelfemtocells on the call drop probability in a residential scenario.In Proceedings of the IEEE International Symposium onPersonal, Indoor and Mobile Radio Communications, Athens,Greece, 3-7 Sept. 2007.

[13] 3rd Generation Partnership project, 3GPP TR 25.820, 3G HomeNodeB Study Item Technical Report, version 8.0.0.http://www.3gpp.org/, Sept. 2008.

References

118

[14] 3rd Generation Partnership project. 3GPP TR 25.967, HomeNode B (HNB) Radio Frequency (RF) Requirements (FDD),version 13.0.0. http://www.3gpp.org/, Jan. 2016.

[15] Ch. David. Where can I buy a femtocell. ThinkSmallcell,https://www.thinksmallcell.com/Examples/where-can-i-buy-a-femtocell.html

[16] IEEE Instrumentation and Measurement Society. 1588-2008IEEE Standard for a Precision Clock Synchronization Protocolfor Networked Measurement and Control Systems.http://ieeexplore.ieee.org/servlet/opac?punumber=4579757,2008.

[17] International Telecommunication Union, TelecommunicationStandardization Sector. ITU-T G.8271 07/2016, Time andphase synchronization aspects in packet networks.https://www.itu.int/rec/T-REC-G.8271, July, 2016.

[18] L. Sun, E. C. Ifeachor. Prediction of perceived conversationalspeech quality and effects of playout buffer algorithms. InProceedings of the IEEE International Conference onCommunications (ICC), Anchorage, USA, May 2003.

[19] International Telecommunication Union, TelecommunicationStandardization Sector. ITU-T G.107 06/2015, The E-model: acomputational model for use in transmission planning.https://www.itu.int/rec/T-REC-G.107, June, 2015.

[20] International Telecommunication Union, TelecommunicationStandardization Sector. ITU-T P.862 02/01, Perceptualevaluation of speech quality (PESQ): An objective method forend-to-end speech quality assessment of narrow-bandtelephone networks and speech codecs.http://www.itu.int/rec/T-REC-P.862, Feb. 2001.

[21] M. Degermark, B. Nordgren, S. Pink. IP Header Compression.Internet Engineering Task Force (IETF) RFC 2507,https://tools.ietf.org/html/rfc2507, Feb. 1999.

[22] L-E. Jonsson, G. Pelletier, K. Sandlund. The RObust HeaderCompression (ROHC) Framework. Internet Engineering TaskForce (IETF) RFC 4995, https://tools.ietf.org/html/rfc4995,July 2007.

[23] O. Fresan, T. Chen, K. Ranta-aho, T. Ristaniemi. Dynamicpacket bundling for VoIP transmission over Rel’7 HSUPA with10ms TTI length. In Proceedings of the IEEE ISWCS, pp. 508-512, Trondheim, Norway, Oct. 2007

[24] G. Rittenhouse, H. Zheng. Providing VoIP service in UMTS-HSDPA with frame aggregation. In Proceedings of the ICASSP,vol. 2, pp. 1157-1160, Philadelphia, USA, Mar. 2005.

[25] J. Sjoberg, M. Westerlund, A. Lakaniemi, Q. Xie. RTP payloadformat and file storage format for the Adaptive Multi-Rate(AMR) and Adaptive Multi-Rate Wideband (AMR-WB) audiocodecs. Internet Engineering Task Force (IETF) RFC 4867,https://tools.ietf.org/html/rfc4867, Apr. 2007

[26] H. Schulzrinne, S. Casner, R. Frederick, V. Jacobson. RTP: ATransport Protocol for Real-time Applications. InternetEngineering Task Force (IETF) RFC 1889,https://tools.ietf.org/html/rfc1889, Jan. 1996.

References

119

[27] 3rd Generation Partnership project. 3GPP TS 29.414, CoreNetwork Nb Data Transport and Transport Signalling, version13.1.0. http://www.3gpp.org/, Mar. 2016.

[28] 3rd Generation Partnership project. 3GPP TS 29.814,Feasibility Study on Bandwidth Savings at Nb Interface with IPtransport, version 7.1.0. http://www.3gpp.org/, June 2007.

[29] P. Cameron, D. Crocker, D. Cohen, J. Postel. TransportMultiplexing Protocol (TMux). Internet Engineering TaskForce (IETF) RFC 1692, https://tools.ietf.org/html/rfc1692,Aug. 1994.

[30] R. Pazhyannur, I. Ali, C. Fox. PPP Multiplexing. InternetEngineering Task Force (IETF) RFC 3153,https://tools.ietf.org/html/rfc1889, Aug. 2001.

[31] Cisco Systems. Cisco Visual Networking Index: Global MobileData Traffic Forecast Update, 2015–2020. White Paper,http://www.cisco.com/c/en/us/solutions/collateral/service-provider/visual-networking-index-vni/mobile-white-paper-c11-520862.html. Feb. 2016.

[32] 3rd Generation Partnership project. 3GPP TR 36.814, FurtherAdvancements for E-UTRA Physical Layer Aspects, version9.0.0. http://www.3gpp.org/, Mar. 2010.

[33] 3rd Generation Partnership project. 3GPP TR 36.819,Coordinated Multi-point Operation for LTE Physical LayerAspects, version 11.2.0. http://www.3gpp.org/, Sep. 2013.

[34] T. Biermann, L. Scalia, Ch. Choi, W. Kellerer, H. Karl. HowBackhaul Networks Influence the Feasibility of CoordinatedMultipoint in Cellular Networks. IEEE CommunicationsMagazine, 51(8): 168-176, Aug. 2013.

[35] J. Lee, Y. Kim, H. Lee, B. L. Ng ; D. Mazzarese, J. Liu, W. Xiao,Y. Zhou. Coordinated Multipoint Transmission and Receptionin LTE-Advanced Systems. IEEE Communications Magazine,50(11):44-50, Nov. 2012.

[36] R. Heath, S. Peters, Y. Wang, J. Zhang. A Current Perspectiveon Distributed Antenna Systems for the Downlink of CellularSystems. IEEE Communications Magazine, 51(4):161-167,Apr. 2013.

[37] Nokia Siemens Networks. Liquid Radio Let traffic waves flowmost efficiently. White paper, nsn.com/system/files/document/liquid_radio_white_paper_280611.pdf, June 2011.

[38] P. Marsh, G. Fettweis eds. Cooperative MultipointCommunications – From Theory to Practice. CambridgeUniversity Press, 2011.

[39] R. Irmer, H. Droste, P. Marsch, M. Grieger, G. Fettweis, S.Brueck, H.-P. Mayer, L. Thiele, V. Jungnickel. CoordinatedMultipoint: Concepts, Performance, and Field Trial Results.IEEE Communications Magazine, 49(2):102-111, Feb. 2011.

[40] 3rd Generation Partnership project. 3GPP TS 36.420, EvolvedUniversal Terrestrial Radio Access Network (E-UTRAN); X2General Aspects and Principles, version 11.0.0.http://www.3gpp.org/, Sep. 2012.

[41] F. Diehm, G. Fettweis. On the Impact of Signaling Delays onthe Performance of Centralized Scheduling for Joint DetectionCooperative Cellular Systems. In Proceedings of the IEEE

References

120

Wireless Communications and Networking Conference(WCNC), Cancun, Mexico, Mar. 2011.

[42] 3rd Generation Partnership project. 3GPP TS 36.213, EvolvedUniversal Terrestrial Radio Access (E-UTRA); Physical LayerProcedures, version 10.4.0. http://www.3gpp.org/, Dec. 2011.

[43] G. Cili, H. Yanikomeroglu, F. R. Yu. Energy Efficiency andCapacity Evaluation of LTE-Advanced Downlink CoMPSchemes Subject to Channel Estimation Errors and SystemDelay. In Proceedings of the IEEE 78th Vehicular TechnologyConference (VTC Fall), Las Vegas, USA, Sep. 2013.

[44] Ch. Choi, Q. Wei, Th. Biermann L. Scalia. Mobile WDMBackhaul Access Networks with Physical Inter-Base-StationLinks for Coordinated Multipoint Transmission/ReceptionSystems. In Proceedings of the IEEE GlobalTelecommunications Conference (GLOBECOM), Houston,USA, Dec. 2011.

[45] T. Biermann, L. Scalia, H. Karl. Designing Optical Metro andAccess Networks for Future Cooperative Cellular Systems. InProceedings of the 14th ACM international conference onModeling, Analysis and Simulation of Wireless and MobileSystems, Miami Beach, USA, Nov. 2011.

[46] M. Sawahashi, Y. Kishiyama, A. Morimoto, D. Nishikawa, M.Tanno. Coordinated multipoint transmission/receptiontechniques for LTE-advanced. IEEE WirelessCommunications, 17(3):26-34, June 2010.

[47] D. Lee, H. Seo, B. Clerckx, E. Hardouin, D. Mazzarese, S.Nagata, K. Sayana. Coordinated Multipoint Transmission andReception in LTE-Advanced: Deployment Scenarios andOperational Challenges. IEEE Communications Magazine,50(2):148-155, Feb. 2012.

[48] V. Jungnickel, T. Wirth, M. Schellmann, T. Haustein, W.Zirwas. Synchronization of Cooperative Base Stations. IEEEInternational Symposium on Wireless CommunicationSystems (ISWCS), Reykjavik, Iceland, Oct. 2008.

[49] D, Bladsjö, M. Hogan, S. Ruffini. Synchronization Aspects inLTE Small Cells. IEEE Communications Magazine, 51(9):70-77, Sep. 2013.

[50] K. Manolakis, L. Thiele, Ch. Oberli, Th. Haustein, V.Jungnickel. Impairment Modeling for Joint TransmissionCoMP. In Proceedings of the 2nd International Conference onWireless Communication, Vehicular Technology, InformationTheory and Aerospace & Electronic Systems Technology(Wireless VITAE), Chennai, India, Feb.-Mar. 2011.

[51] A. Papadogiannis, E. Hardouin, D. Gesbert. DecentralisingMulticell Cooperative Processing: A Novel Robust Framework.EURASIP Journal on Wireless Communications andNetworking, 2009(1):1-10, Feb. 2009.

[52] A. Papadogiannis, H. J. Bang, D. Gesbert, E. Hardouin.Efficient Selective Feedback Design for Multicell CooperativeNetworks, IEEE Transactions on Vehicular Technology,60(1):196-205, Jan. 2011.

[53] V. Jungnickel, L. Thiele, M. Schellmann, T. Wirth, A. Forck, W.Zirwas, T. Haustein, E. Schulz. Implementation Concepts for

References

121

Distributed Cooperative Transmission. In Proceedings of the42nd Asilomar Conference on Signals, Systems andComputers, Pacific Grove, USA, Oct. 2008.

[54] W. Zirwas, W. Mennerich. Main Enablers for AdvancedInterference Mitigation. In Proceedings of the 8thInternational Workshop on Multi-Carrier Systems &Solutions (MC-SS), Herrsching, Germany, May. 2011.

[55] D. Hui. Distributed Precoding with Local Power Negotiationfor Coordinated Multi-Point Transmission. In Proceedings ofthe IEEE 73rd Vehicular Technology Conference (VTCSpring), Budapest, Hungary, May, 2011.

[56] T. Biermann, L. Scalia, J. Widmery, H. Karl. Backhaul Designand Controller Placement for Cooperative Mobile AccessNetworks. In Proceedings of the 73rd IEEE VehicularTechnology Conference (VTC Spring), Budapest, Hungary,May 2011.

[57] M. Draxler, Th. Biermann, H. Karl, W. Kellerer. Cooperatingbase station set selection and network reconfiguration inlimited backhaul networks. In Proceedings of the 23rd IEEEInternational Symposium Personal Indoor and Mobile RadioCommunications (PIMRC), Sydney, Australia, Sep. 2012.

[58] Th. Biermann, L. Scalia, Ch. Choi, H. Karl, W. Kellerer.Improving CoMP Cluster Feasibility by Dynamic Serving BaseStation Reassignment. In Proceedings of the 22nd IEEEInternational Symposium Personal Indoor and Mobile RadioCommunications (PIMRC), Toronto, Canada, Sep. 2011.

[59] R. Weber, A. Garavaglia, M. Schulist, S. Brueck, A. Dekorsy.Self-Organizing Adaptive Clustering for CooperativeMultipoint Transmission. In Proceedings of the IEEE 73rdVehicular Technology Conference (VTC Spring), Budapest,Hungary, May, 2011.

[60] Ch. Choi, L. Scalia, Th. Biermann, Sh. Mizuta. CoordinatedMultipoint Multiuser-MIMO Transmissions over Backhaul-Constrained Mobile Access Networks. In Proceedings of the22nd IEEE International Symposium Personal Indoor andMobile Radio Communications (PIMRC), Toronto, Canada,Sep. 2011.

[61] L. Scalia, T. Biermann, Ch. Choi, K. Kozu, W. Kellerer. Power–Efficient Mobile Backhaul Design for CoMP Support in FutureWireless Access Systems. In Proceedings of the IEEEConference on Computer Communications Workshops(INFOCOM Workshops), Shanghai, Apr. 2011.

[62] Q. Zhang, Ch. Yang, A. F. Molisch. Cooperative downlinktransmission mode selection under limited-capacity backhaul.In Proceedings of the IEEE Wireless Communications andNetworking Conference (WCNC), Shanghai, Apr. 2012.

[63] Q. Zhang, Ch. Yang. Transmission Mode Selection forDownlink Coordinated Multipoint Systems. IEEETransactions on Vehicular Technology, 62(1):465-471, Jan.2013.

[64] Q. Zhang, Ch. Yang, A. Molisch. Downlink Base StationCooperative Transmission Under Limited-Capacity Backhaul.

References

122

IEEE Transactions on Wireless Communications, 12(8):3746-3759, Aug. 2013.

[65] N. Seifi, M. Viberg, R. W. Heath, J. Zhang, M. Coldrey.Coordinated Single-Cell vs Multi-Cell Transmission withLimited-Capacity Backhaul. In Proceedings of the 44th

Asilomar Conference on Signals, Systems and Computers(ASILOMAR), Pacific Grove, USA, Nov. 2010.

[66] R. Zakhour, D. Gesbert. Optimized Data Sharing in MulticellMIMO With Finite Backhaul Capacity. IEEE Transactions onSignal Processing, 59(12):6102-6111, Dec. 2011.

[67] G. Zhou, G. Bauch, J. Berkmann. Inter-cell InterferenceCoordination via Cooperative Rate Splitting and Scheduling. InProceedings of the 7th IEEE International Conference onWireless and Mobile Computing, Networking andCommunications (WiMob), Shanghai, China, Oct. 2011.

[68] Signals Research Group. Exploring the world of wireless LTE-Advanced and Coordinated Multi-Point (CoMP): What goesaround, comes around. Signals Ahead Newsletter, 8(5), April2012.

[69] X. You, D. Wang, B. Sheng, X. Gao, X. Zhao, M. Chen.Cooperative distributed antenna systems for mobilecommunications [Coordinated and Distributed MIMO]. IEEEWireless Communications, 17(3):35-43, June, 2010.

[70] J. Ni, J. Zhang, D. Chen, X. Bi, Y. Wang. Distributed AntennaSystems and Their Applications in 4G Wireless Systems. InProceedings of the IEEE International Conference onCommunications Workshops (ICC Workshops), Kyoto, Japan,June, 2011.

[71] F. Diehm, P. Marsch, G. Fettweis. The FUTON Prototype: Proofof Concept for Coordinated Multi-Point in Conjunction with aNovel Integrated Wireless/Optical Architecture. InProceedings of the IEEE Wireless Communications andNetworking Conference Workshops (WCNCW), Sydney,Australia, April, 2010.

[72] A. Nagate, K. Hoshino, M. Mikami, T. Fujii. A Field Trial ofMulti-cell Cooperative Transmission over LTE System. InProceedings of the IEEE International Conference onCommunications (ICC), Kyoto, Japan, June 2011.

[73] Ericsson, Huawei Technologies, NEC, Alcatel Lucent, NokiaSiemens Networks. Common Public Radio Interface (CPRI);Interface Specification. Version 6.0, http://www.cpri.info/,Aug. 2013.

[74] R. Irmer, H. Mayer, A. Weber, V. Braun, M. Schmidt, M. Ohm,N. Ahr, A. Zoch, C. Jandura, P. Marsch, G. Fettweis. MultisiteField Trial for LTE and Advanced Concepts. IEEECommunications Magazine, 47(2):92-98, Feb. 2009.

[75] J. Holfeld, I. Riedel, G. Fettweis. A CoMP downlinktransmission system verified by cellular field trials. InProceedings of the 19th European Signal ProcessingConference (EUSIPCO), Barcelona, Spain, Aug. 2011.

[76] J. Holfeld, E. Fischer, G. Fettweis. Field Trial Results for CoMPDownlink Transmissions in Cellular Systems. In Proceedings

References

123

of the International ITG Workshop on Smart Antennas(WSA), Aachen, Germany, Feb. 2011.

[77] V. Jungnickel, L. Thiele, T. Wirth, T. Haustein, S.Schiffermuller, A. Forck, S. Wahls, S. Jaeckel, S. Schubert, H.Gabler, C. Juchems, F. Luhn, R. Zavrtak, H. Droste, G. Kadel,W. Kreher, J. Mueller, W. Stoermer, G. Wannemacher.Coordinated Multipoint Trials in the Downlink. In Proceedingsof the IEEE Global Telecommunications ConferenceWorkshops (GLOBECOM Workshops), Honolulu, USA, Nov-Dec. 2009.

[78] V. Jungnickel, A. Forck, S. Jaeckel, F. Bauermeister, S.Schiffermueller, S. Schubert, S. Wahls, L. Thiele, T. Haustein,W. Kreher, J. Mueller, H. Droste, G. Kadel. Field Trials usingCoordinated Multi-Point Transmission in the Downlink. InProceedings of the 21st IEEE International Symposium onPersonal, Indoor and Mobile Radio CommunicationsWorkshops (PIMRC Workshops), Istanbul, Turkey, Sep. 2010.

[79] T. Okamawari, L. Zhang, A. Nagate, H. Hayashi, T. Fujii.Design of Control Architecture for Downlink CoMP JointTransmission with Inter-eNB Coordination in Next GenerationCellular Systems. In Proceedings of the 74th IEEE VehicularTechnology Conference (VTC Fall), San Francisco, USA, Sep.2011.

[80] S. Brueck, J. Giese, L. Zhao, A. Dekorsy. On MAC LayerThroughput Enhancements in LTE-A by Downlink MacroDiversity. In Proceedings of the IEEE InternationalConference on Communications Workshops (ICC Workshops),Dresden, Germany, Jun. 2009.

[81] S. Brueck, L. Zhao, J. Giese, M. A. Amin. CentralizedScheduling for Joint Transmission Coordinated Multi-Point inLTE-Advanced. In Proceedings of the International ITGWorkshop on Smart Antennas (WSA), Bremen, Germany, Feb.2010.

[82] 3rd Generation Partnership project. 3GPP TR 25.814, PhysicalLayer Aspects for Evolved Universal Terrestrial Radio Access(UTRA), version 7.1.0. http://www.3gpp.org/, Sep. 2006.

[83] Nokia. Looking Ahead to 5G – Building a virtual zero latencygigabit experience. White paper,http://networks.nokia.com/file/28771/5g-white-paper, 2014.

[84] The Quality of Service and MObility driven cognitive radioSystems (QoSMOS) project. http:// http://www.ict-qosmos.eu.

[85] D. Tandur, J. Duplicy, K. Arshad, D. Depierre, K. Moessner, J.Lehtomäki, K. Briggs, L. Gonçalves, Atílio Gameiro. Cognitiveradio systems evaluation – Measurement, Modeling andEmulation. IEEE Vehicular Technology Magazine, 7(2):77-84,June 2012.

[86] M. Cudak, A. Ghosh, Th. Kovarik, R. Ratasuk, T. A. Thomas, F.W. Vook, P. Moorut. Moving towards mmWave-based beyond-4G (B-4G) technology. In Proceeding of the 77th IEEE VTCSpring, Dresden, Germany, June 2013.

[87] T. S. Rappaport, Sh. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang,G. N. Wong, J. K. Schulz, M. Samimi, F. Gutierrez. Millimeter

References

124

wave mobile communications for 5G cellular: It will work!.IEEE access, 2013(1):335-349.

[88] T. S. Rappaport, R. W. Heath, R. C. Daniels, J. N. Murdock.Millimeter Wave Wireless Communications. Prentice Hall,2014.

[89] W. Hong, K-H. Baek, Y. Lee, Y. Kim, S-T. Ko. Study andPrototyping of Practically Large-Scale mmWave AntennaSystems for 5G Cellular Devices. IEEE CommunicationsMagazine, 52(9):63–69, Sep. 2014.

[90] A. I. Sulyman, A. T. Nassar, M. K. Samimi, G. R. MacCartney,T. S. Rappaport, A. Alsanie. Radio Propagation Path LossModels for 5G Cellular Networks in the 28 GHz and 38 GHzMillimeter-Wave Bands. IEEE Communications Magazine,52(9):78–86, Sep. 2014.

[91] G. R. MacCartney, M. K. Samimi, T. S. Rappaport. Exploitingdirectionality for millimeter-wave wireless systemimprovement. In Proceedings of the IEEE InternationalConference on Communications (ICC), London, UK, June2015.

[92] S. Hur, S. Baek, B. Kim, Y. Chang, A. F. Molisch, T. S.Rappaport, K. Haneda, J. Park. Proposal on Millimeter-WaveChannel Modeling for 5G Cellular System. IEEE Journal onSelected Topics in Signal Processing, 10(3):454–469, Apr.2016.

[93] T. S. Rappaport, S. Deng. 73 GHz Wideband Millimeter-WaveFoliage and Ground Reflection Measurements and Models. InProceedings of the IEEE International Conference onCommunications Workshop (ICC Workshops), London, UK,June 2015.

[94] Sh. Sun, T. S. Rappaport, T. A. Thomas, A. Ghosh. APreliminary 3D mmWave Indoor Office Channel Model. InProceedings of the International Conference on Computing,Networking and Communications (ICNC), Garden Grove,USA, Feb. 2015.

[95] G. Fettweis. The Tactile Internet. IEEE Vehicular TechnologyMagazine, 9(1):64-70, March 2014.

[96] GSM Association Intelligence (GSMA). The Mobile Economy2016. http://gsmaintelligence.com, Feb. 2016.

-otl

aA

DD

9

2/

710

2

+ddadh

a*GMFTSH

9 NBSI 3-3037-06-259-879 )detnirp( NBSI 6-2037-06-259-879 )fdp(


ytisrevinU otlaA

gnireenignE lacirtcelE fo loohcS gnikrowteN dna snoitacinummoC fo tnemtrapeD

if.otlaa.www

+ SSENISUB YMONOCE

+ TRA

+ NGISED ERUTCETIHCRA

+ ECNEICS

YGOLONHCET

REVOSSORC


ydz

orD

dápr

Á l

uahk

caB

elibo

M fo

noit

ulov

E y

tisr

evi

nU

otla

A

7102


eliboM fo noitulovE luahkcaB

ydzorD dáprÁ


Documents

E volution of Mobile DD B ackhaul -otlaA