Vous êtes sur la page 1sur 200

THESE DE DOCTORAT DE

L'UNIVERSITE DE RENNES 1
COMUE UNIVERSITE BRETAGNE LOIRE

ECOLE DOCTORALE N° 596


Matière, Molécules, Matériaux
Spécialité : photonique

Par
Kaoutar BENYAHYA

Mode-Group Division Multiplexing for Short Reach


Optical Communications

Rapporteurs avant soutenance :


Sophie Larochelle, Université Laval
Yves Jaouen, Télécom Paris

Composition du Jury :

Examinateurs :
Christelle Aupetit-Berthelemot, Université de Limoges
Nicolas Dubreuil, Institut d'Optique Graduate School

Yves Jaouen, Télécom Paris

Invitée
Marianne Bigot-Astruc, Prysmian Group

Encadrants
Christian Simonneau, Alcatel-Submarine Networks
Amirhossein Ghazisaeidi, Nokia Bell-Labs

Directeur de thèse :
Peucheret Christophe , Université de Rennes 1
Remerciements

Tout d’abord j’aimerais commencer par remercier tous les membres du jury qui m’ont accordé
l’honneur de lire et d’évaluer mon travail de thèse. Merci pour vos précieux conseils et com-
mentaires qui ont contribué à la production de ce manuscrit.

A travers ces lignes, je tiens à orienter mon immense remerciement et ma gratitude envers mes
encadrants de thèse, Christian Simonneau et Amirhossein Ghazisaeidi, qui m’ont accompagné
depuis le début de mon stage de fin d’études et aussi tout au long de ma thèse. Leur grand
soutien et leurs encouragements ainsi que leur savoir-faire et leur patience représentent un point
clé de la réussite de cette thèse. Chers encadrants, je tiens donc à vous remercier autant pour
votre côté humain que professionnel, la combinaison de ces deux qualités ont fait de ma thèse une
très riche expérience. Malgré son départ au début de ma troisième année de thèse, Christian
Simonneau a continué à me superviser et a toujours été un soutien jusqu’à la fin de cette
aventure. Je n’oublie pas de remercier Amirhossein Ghazisaeidi pour son partage généreux de
savoirs et ses discussions intéressantes incluant, histoire, sciences, religion, . . . tout au long de
nos trajets sur le RER ‘B’. Je tiens également à remercier Jérémie Renaudier et Gabriel Charlet
pour leurs contributions impactantes vis-à-vis de ce sujet, la mise en œuvre généreuse de leur
grande expertise a toujours été bénéfique pour les différents axes de recherches que j’ai abordé.
J’aimerais aussi remercier mes collègues un par un, merci à vous pour l’ambiance magnifique
dans laquelle s’est déroulée ma thèse. Nos moments de convivialité lors de nos pauses café, des
sorties, des conférences, . . . me donnaient de l’énergie et boostaient ma motivation.

Enfin, je voudrais remercier ma famille ainsi que mes chers amis. Un énorme merci à Laurent
Moureu qui m’a ouvert les premières portes d’Alcatel Lucent lors de mon stage de découverte
d’entreprise à Lannion. Il n’a cessé de me suivre et m’orienter professionnellement toujours
vers les bons choix. Merci à mes parents qui depuis ma naissance sont là pour filtrer toutes les
difficultés que me cache la vie et m’aider à les surmonter, cette thèse est pour vous remercier
et vous rendre fiers aujourd’hui. Merci à Mustapha Emrabdan, homme de grand cœur qui
m’a toujours soutenue et a veillé sur moi. Merci à ma sœur Ikram Benyahya pour tous les
repas qu’elle m’a préparés avec amour pendant les périodes stressantes de ma thèse, à vrai dire,
je voulais dire merci les livraisons à domicile!! Finalement je tiens à remercier Rahif Kassab
pour son soutien ainsi que le temps qu’il a consacré à relire ma thèse et à dénicher les fautes
d’orthographe.

i
Résumé

La demande croissante du trafic de données sera alimentée par des technologies révolutionnaires
telles que la réalité virtuelle (VR), la réalité augmentée (AR) et l’Internet des objets (IoT).
Par conséquent, les réseaux optiques devraient répondre aux besoins de ces services en termes
de débit, faible temps de réponse et grande fiabilité. En effet, les hauts débits représentent un
besoin critique pour les systèmes de communication à fibre optique déployés dans les réseaux
locaux ainsi que dans les centres de données. Pour ces deux applications, les systèmes basés sur
la modulation d’intensité et la détection directe de cette dernière sont très attractifs en raison
de leur faible coût et de leur compatibilité avec les applications à courte distance.

Dans le cadre de cette thèse, nous répondons à la nécessité d’augmenter les débits pour les
systèmes de communication optiques à courte distance basés sur le multiplexage de groupe de
modes et la détection directe. Tout d’abord, nous visons à augmenter la capacité des fibres mul-
timodes standard déjà déployées dans les réseaux locaux et à l’intérieur des centres de données
où la distance est inférieure à 5 km. Deuxièmement, nous étendons notre solution aux applica-
tions avec des distances de déploiement plus longues telles que les connexions entre les centres
de données. Dans les deux cas, les architectures des liens optiques, y compris les émetteurs, les
récepteurs et les fibres optiques, sont analysées. De plus, les formats de modulation adaptés
aux systèmes basés sur la détection directe tels que le format de modulation mono-porteuse
4-PAM et celui multi-porteuse DMT sont comparés dans le contexte de la transmission basée
sur le multiplexage spatial.

Nous avons démontré les avantages du multiplexage de groupes de modes combiné à la détection
directe pour augmenter le débit transmit sur une seule fibre. Premièrement, 5 Tb / s ont été
démontré sur 2,2 km de fibre multimode conventionnelle (OM2). Deuxièmement, un record
de transmission de 14,5 Tb / s sur fibre OM2 est démontré au moment correspondant à sa
réalisation. Enfin, 200 Gb / s sur 20 km de fibre faiblement multimode (FMF) a été démontré,
ce qui étend les avantages du multiplexage par groupes de modes aux applications à longue
distance par rapport aux réseaux LAN où la distance maximale est limitée à 5 km.

ii
List of Acronyms

A|B|C|D|E|F|I|L|M|N|O|P|Q|R|S|T|V|W

ADC analog to digital converter.

APD avalanche photodiode.

ASIC application-specific integrated circuits.

AWGN additive white gaussian noise.

B2B back-to-back.

BER bit error rate.

BPP binary phased plates.

BPSK binary phase shift keying.

CD Chromatic Dispersion.

CMOS complementary metal oxide semiconductor.

CP cyclic prefix.

CWDM coarse wavelength division multiplexing.

iii
D

DAC digital to analog converter.

DC direct current.

DCF dispersion-compensating fiber.

DCI datacenter interconnect.

DD direct detection.

DFB distributed feedback laser.

DFE decision-feedback equalizer.

DML directly modulated laser.

DMT discrete multitone modulation.

DSP Digital Signal Processing.

EAM electro absorption modulator.

EDFA Erbium Doped Fiber Amplifier.

EML externally modulated laser.

FEC forward-error corrector.

FFE feed-forward equalizer.

FFT fast Fourier Transform.

FIR finite impulse response.

FMF few-mode-fiber.

iv
I

ICI inter-carrier interference.

IFFT inverse fast Fourier transform.

IMDD intensity modulation direct detection.

InP indium phospide.

ISI inter-symbol-interference.

IXT inter-model crosstalk.

LANs local area networks.

LMS least mean squares.

LMSE linear mean square equalizer.

LR long reach.

MCF multi-core fibers.

MGD mode-group demultiplexer.

MGM mode-group multiplexer.

MIMO multi-input multi-output.

ML maximum likelihood.

MMF multimode Fiber.

MMFs multimode fibers.

MMSE minimum mean square error.

v
MPI multipath interference.

MPLC multi-plane light conversion.

MSE mean square error.

MZM mach-zehnder modulator.

NRZ non return to zero.

OAM orbital angular momentum.

OFDM orthogonal frequency division multiplexing.

PAM pulse amplitude modulation.

PAPR peak to average power ratio.

PDF probability density function.

QAM quadrature amplitude modulation.

QPSK quadrature phase shift keying.

rms root mean square.

RRC root raised cosine.

Rx receiver.

vi
S

SDM spatial-division-multiplexing.

SER symbol error rate.

SLM spatial light modulator.

SMF single mode fiber.

SNR signal to noise ratio.

SSBI signal to signal beating interference.

SSMF standard single-mode fiber.

SWDM short wavelength division multiplexing.

TIA trans impedance amplifier.

Tx transmitter.

VCSELs Vertical-cavity surface-emitting lasers.

WDM wave division multiplexing.

vii
viii
Contents

Remerciements i

Résumé ii

1 Introduction 1

1.1 A Brief Introduction to IMDD Systems . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Datacenter Traffic Growth and Topology . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Datacenter: Transceivers and Optical Fiber Connectivity . . . . . . . . . . . . . 6

1.3.1 State of the art : Product . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.3.2 Toward Tb/s Interconnect Technologies: a Critical Challenge . . . . . . . 8

1.3.3 SDM in Support of Large Scale Datacenter Interconnects . . . . . . . . . 9

1.4 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Intensity Modulation Direct Detection Systems 15

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2 IMDD optical characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.2.1 IMDD Transmitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

ix
x CONTENTS

2.2.2 IMDD Receiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.3 Pulse Amplitude Modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.3.1 Symbol and Bit Error Probability . . . . . . . . . . . . . . . . . . . . . . 24

2.3.2 Classical M-PAM DSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.4 Discrete Multitone Modulation (DMT) . . . . . . . . . . . . . . . . . . . . . . . 35

2.4.1 Basic Concept of Conventional DMT . . . . . . . . . . . . . . . . . . . . 35

2.4.2 Inverse Fourier Transform (IFFT) . . . . . . . . . . . . . . . . . . . . . . 36

2.4.3 Cyclic Prefix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.4.4 Peak to Average Power Ratio . . . . . . . . . . . . . . . . . . . . . . . . 40

2.4.5 Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.4.6 Loading Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3 Space Division Multiplexing 56

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.2 Optical fibers structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.3 Optical fiber dispersion characteristics . . . . . . . . . . . . . . . . . . . . . . . 61

3.4 Spatial Mode Division Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . 65

3.4.1 Inter-modal Crosstalk . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.4.2 Mode-division-Multiplexing Techniques . . . . . . . . . . . . . . . . . . . 71

3.4.3 Spatial Mode-Group Multiplexing based on MPLC . . . . . . . . . . . . 76

3.5 Crosstalk Investigation for IMDD Schemes . . . . . . . . . . . . . . . . . . . . . 78


CONTENTS xi

3.5.1 Crosstalk impact due to MGM, MGD and splices . . . . . . . . . . . . . 79

3.5.2 Crosstalk Impact Using the M-PAM Format . . . . . . . . . . . . . . . . 82

3.5.3 Crosstalk Impact Using the DMT Format . . . . . . . . . . . . . . . . . 86

3.5.4 Crosstalk’ s impact comparison between M-PAM and DMT formats . . . 89

3.5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

4 Transmission Experiments 94

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

4.2 Mode Group Multiplexing for LAN and Intra


-Datacenter Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

4.2.1 5 Tb/s Bidirectional Transmission Over OM2 Fiber . . . . . . . . . . . . 96

4.2.2 Beyond 250Gb/s Single Wavelength Transmission Based on Low Cost EML103

4.2.3 14.5 Tb/s Transmission over OM2 Fiber . . . . . . . . . . . . . . . . . . 109

4.3 Mode-Group-Multiplexing for DCI . . . . . . . . . . . . . . . . . . . . . . . . . 121

4.3.1 Transition from OM2 to few-mode-fiber (FMF) . . . . . . . . . . . . . . 122

4.3.2 20 km Transmission Experiment . . . . . . . . . . . . . . . . . . . . . . . 127

4.4 Experimental Investigation of Crosstalk Impact on 4-PAM and DMT Formats . 135

4.4.1 Polarization Maintaining setup . . . . . . . . . . . . . . . . . . . . . . . 135

4.4.2 Experimental Investigation . . . . . . . . . . . . . . . . . . . . . . . . . . 137

4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

Conclusions and Perspectives 145


List of Publications 150

Bibliography 152

xii
List of Tables

1.1 First four generation of fiber optic communication systems [1] . . . . . . . . . . 2

1.2 Example of standardized and proposed optical interfaces for datacenter inter-
connect for 25Gb/s, 40Gb/s, 100Gb/s, and 400Gb/s . . . . . . . . . . . . . . . . 7

2.1 Example of training pattern signs. . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.1 Generated loss and Crosstalk of the MGM followed by the MGD. . . . . . . . . 98

4.2 Accumulated losses over the link where OM2 represents 2.2 km of OM2 fiber. . . 100

4.3 Estimated fiber crosstalk. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

4.4 Achieved net bit rate per channel for each mode group . . . . . . . . . . . . . . 118

4.5 ∆nef f between used mode groups . . . . . . . . . . . . . . . . . . . . . . . . . . 124

4.6 Generated loss by MGM+MGD+20 m of fiber (in case of OM2 and FMF). . . . 127

4.7 Loss and crosstalk generated by MGM,MGD, 20km of FMF and 2 splices. . . . 129

4.8 CD parameter estimation for each mode groups and its comparison with simu-
lated value by Prysmian. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

4.9 Measured Loss and modal crosstalk generated by maintaining polarization MGM,
MGD and 20 m of FMF. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

xiii
xiv
List of Figures

1.1 Global data center (a) IP traffic growth and (b) traffic by destination in 2021
(source: Cisco Global Cloud index, 2016-2021) . . . . . . . . . . . . . . . . . . . 3

1.2 Intra-datacenter networks architecture: folded CLOS. . . . . . . . . . . . . . . . 4

1.3 Ethernet traffic growth published by Ethernet Alliance [2]. . . . . . . . . . . . . 5

1.4 Typical ranges of link distances categories for datacenter interconnect . . . . . . 6

1.5 Different architectures for 400G transceivers for datacenter interconnects, from
left to right: 400G-BASE-SR16, 400G-BASE-LR8, 400G-BASE-DR4 . . . . . . 8

1.6 Photo of massive number of optical cables installed in datacenter (source: IBM) 10

1.7 Cross sections of most common optical fibers driving research on SDM (a) SMF,
(b) FMF, (c) standard MMF, (d) elliptical core fiber, (e) multi-cores fiber . . . . 10

2.1 Generic model of IMDD link. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.2 Conceptual model of IMDD transmitter based on (a) DML and (b) EML. . . . . 17

2.3 Laser structure for (a) DFB and (b) VCSEL. . . . . . . . . . . . . . . . . . . . 19

2.4 Model of linear intensity modulator. . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.5 Characteristic transfer function for (a) EAM and (b) mach-zehnder modulator
(MZM). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

xv
xvi LIST OF FIGURES

2.6 Conceptual model for intensity modulation direct detection (IMDD) receiver. . 22

2.7 Example of a Gray bit-to-symbol mapping for: (a) BPSK, (b) 4-PAM and (c)
8-PAM. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.8 PDF of each level of 4-PAM signal in the presence of Gaussian noise only. . . . 25

2.9 BER versus SNR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.10 DSP flow at (a) the transmitter and (b) the receiver, for M −PAM implementa-
tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.11 Raised-cosine pulse spectrum with different roll-off factors. . . . . . . . . . . . 27

2.12 Principle of digital equalization. . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2.13 Principle of feed-forward equalizer (FFE) equalizer. . . . . . . . . . . . . . . . 31

2.14 Experimental setup for 4-PAM DSP evaluation. . . . . . . . . . . . . . . . . . . 33

2.15 Impact of applying pre-emphasis filter on the performance in terms of (a) elec-
trical spectrum ((b) BER and (c) frequency response of 51-taps FFE equalizer. 34

2.16 Impact of the number of FFE taps on the bit error rate (BER) . . . . . . . . . 34

2.17 Schematic block diagram of principle of DMT modulator and DMT demodulator. 36

2.18 Representation of the impact of cyclic prefix on inter-frame interference. . . . . 39

2.19 Normalized distribution of of two generated DMT signals, with 2N + NCP = 23


(blue curve) and 2N + NCP = 26 (red curve). . . . . . . . . . . . . . . . . . . . 40

2.20 PAPR evolution as function of the number of subcarriers. . . . . . . . . . . . . . 41

2.21 Example of DMT waveform conversion with clipping technique forAth = 0.045 . 42

2.22 Synchronization symbol insertion in the time domain. . . . . . . . . . . . . . . 43

2.23 Example of time domain training sequence with 4 repeated patterns with differ-
ent sign excluding cyclic prefix (L = 4 and M = 16). . . . . . . . . . . . . . . . 44
LIST OF FIGURES xvii

2.24 Time domain metric Λ for L = 4, 8 and 16. . . . . . . . . . . . . . . . . . . . . . 45

2.25 Achievable bit per symbol per subcarrier versus its SNR for different value of Γ. 46

2.26 Performance target (BER) as a function of SNR gap Γ for different M-QAM
constellation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

2.27 SNR threshold generation algorithm. . . . . . . . . . . . . . . . . . . . . . . . . 50

2.28 bn and ξn attribution algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

2.29 Example of SNR threshold and bit allocation rules for a B ? of 10−3 . . . . . . . 52

2.30 Estimated SNR response using quadrature phase shift keying (QPSK) modula-
tion for all the 1024 subcarriers impacted by the channel response H(f). . . . . . 52

2.31 Achievable bitrate as a function of the BER target in (a) and the computed
BER at each bitrate as a function of the BER target in (b). . . . . . . . . . . . 53

2.32 Averaged BER over all the subcarriers carrying the same number of bit per
symbol versus bit/symbol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3.1 Standard optical fibers structure. . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.2 Normalized propagation constant b versus the normalized frequency V . . . . . 60

3.3 Example of intensity profiles of different LPlm modes. . . . . . . . . . . . . . . . 61

3.4 Power fading impact illustrated by (a)the magnitude response of R{Hlm (f )}


as a function of frequency for different fiber lengths (L=2.2,4.4 and 20km) for
20ps/(nm.km) of chromatic dispersion (b) their corresponding channel impulse
response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.5 Power fading impact illustrated by (a) the magnitude response of R{Hlm (f )} as
a function of frequency for different dispersion coefficient (Dlm = 20, 25 and 30
ps./(nm.km)) at 20km (b) their corresponding channel impulse response . . . . . 64
xviii LIST OF FIGURES

3.6 Impact of ∆β(z) on coupled power between two modes . . . . . . . . . . . . . . 69

3.7 Typical designs of fiber refractive index profiles published in conferences and
journal papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.8 Different methods of launching technique: (a) lens combination, (b) small spot
size injection and (c) SMF launch. . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.9 Principle of modes profile generation based on spatial light modulation . . . . . 72

3.10 Principle of modes phase profile generation based on binary phase plates. . . . . 73

3.11 Example of phase transition patters for the first six linearly polarized modes . . 74

3.12 Principle of MPLC mode conversion technique . . . . . . . . . . . . . . . . . . 75

3.13 Principle of MPLC mode conversion technique in multipass cavity. . . . . . . . . 76

3.14 Example of groups of modes in OM2 fiber. . . . . . . . . . . . . . . . . . . . . 77

3.15 Principle of mode-group multiplexing and demultiplexing . . . . . . . . . . . . 78

3.16 Mathematical model for statistical crosstalk investigation in the case of M-PAM
signaling in (a) and DMT format in (b) . . . . . . . . . . . . . . . . . . . . . . 79

3.17 Considered electro-optic conversion model in this study . . . . . . . . . . . . . . 80

3.18 Example of the crosstalk impact on each level for 2-PAM and 4-PAM format at
different crosstalk’s intensity when theta variates uniformally from 0 to 2π . . . 85

3.19 a) Example of the crosstalk impact on each symbol al belonging to a QPSK


constellation according to different crosstalk intensity and where bl is a QPSK
symbol. b) The crosstalk is at −10 dB with different constellation bl and theta
is uniform from 0 to 2π . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

3.20 Comparison of simulated and theoretical crosstalk distribution for different crosstalk
intensities for 4-PAM and 16-QAM-DMT . . . . . . . . . . . . . . . . . . . . . 90
LIST OF FIGURES xix

3.21 BER as function of the crosstalk’s intensity ρ2 for 4-PAM and 16-QAM-DMT. . 91

3.22 Crosstalk impact on each level for 4-PAM and 16-QAM constellation for −15dB
of crosstalk in the presence of AWGN. . . . . . . . . . . . . . . . . . . . . . . . 92

4.1 Bandwidth improvements and maximum achievable distance evolution at 10


Gb/s OOK using multimode VCSEL for standard optical multimode fibers. . . 95

4.2 Intensity patterns of LP modes in OM2 fiber and mode groups patterns used in
this experiment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.3 MGM and MGD used in bidirectional configuration. . . . . . . . . . . . . . . . 98

4.4 Illustration of wave division multiplexing (WDM) and mode group multiplexing
technique used in experimental studies. . . . . . . . . . . . . . . . . . . . . . . 99

4.5 Illustration of wavelength comb interleaving for all the groups. . . . . . . . . . 99

4.6 (a) transmitter (Tx) per mode group, (b) receiver (Rx) per mode group (c)
transfer functions of single mode WDM multiplexer and multimode WDM de-
multiplexer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

4.7 Achievable gross line rate versus Rx input power obtained with Chow’s bit load-
ing algorithm at two different BER target :1.2 × 10−2 and 3.8 × 10−3 . . . . . . 101

4.8 (a) single channel SNR response after 2.2 km propagation over G4 , (b) its corre-
sponding bit allocation for BER target of 3.9 × 10−3 using rate adaptive Chow4S
algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4.9 BER measurement of 80 channels modulated at (a) 80.96Gb/s (b) 68.8Gb/s . . 102

4.10 Total capacity over EDFA’s ouput power. . . . . . . . . . . . . . . . . . . . . . . 103

4.11 externally modulated laser (EML)-based transceivers standards for intra-datacenter


communications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
xx LIST OF FIGURES

4.12 Proposed architecture based on mode group multiplexing and CWDM using
EML . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

4.13 photo of EML test−bench as used in the Lab . . . . . . . . . . . . . . . . . . . 105

4.14 (a) microscopic top view of the EML (b) the SIBH waveguid with semi- insulating-
InP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.15 (a) Static output power as a function of bias voltage, (b) electro-optic bandwidth
for Ilaser=70mA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.16 Impact of 88Gs/s and 65Gs/s DAC in single mode B2B transmission with MZM
on (a) the SNR response at 0dBm , (b) achievable bit rate before FEC for BER
target of 3.9 × 10−3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

4.17 Single mode back-to-back:(a) bit rate comparison between MZM and EML for
a BER target of 3.9 × 10−3 (b) SNR for MZM and EML for a Rx input power
of 0dBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

4.18 Gross line rate for the 4 mode groups for 2.2km and 4.4km for the MZM and
for the EML . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

4.19 Experimental setup for comparison of 30GHZ PIN+TIA and 28GHz PIN in
single channel single mode B2B configuration . . . . . . . . . . . . . . . . . . . 111

4.20 (a) comparison of achievable bit rates before FEC between different photore-
ceivers for BER target of 3.9 × 10−3 (b) SNR versus frequency at Rx input
power of -3dBm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

4.21 MGM+MGD configuration as used in the experiment, thin lines represents SMF
while thick lines represents OM2 fiber . . . . . . . . . . . . . . . . . . . . . . . . 112

4.22 The achievable throughput comparing 28GHz PIN and 30GHz PIN+TIA for the
4 groups of modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

4.23 Converted modes received by the photodiode after mode group demultiplexing. . 113
LIST OF FIGURES xxi

4.24 Impact of MGM+MGD insertion compared to single mode B2B. . . . . . . . . . 114

4.25 (a) transmitter including single mode WDM multiplexer (b) receiver including
multimode WDM demultiplexer . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

4.26 (a) Spectrum of co-propagating G1 and G4 in one direction where A corresponds


for 1 km experiment and B for 2.2 km at the output of MZM modulators (b) 2
different configurations used to measure the penalty due to wavelength crosstalk
in the WDM multimode demultiplexer induced by adjacent channels . . . . . . . 115

4.27 (a) SNR of the optical channel in several configurations (inset filtering function
of the single mode WDM multiplexer and the multimode WDM demultiplexer,
(b) Experimental configuration used for measurements in (a). . . . . . . . . . . . 117

4.28 SNR response after 2.2km propagation for all mode groups at -3dB of input power.118

4.29 Allocated bit per subcarrier for G1 and G4 corresponding to 2.2km of propagation
for a BER target of 5 × 10−3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

4.30 Achieved bit rate for BER target of 5 × 10−3 versus distance of propagation for
single channel for all mode groups . . . . . . . . . . . . . . . . . . . . . . . . . . 119

4.31 Measured BER for the all 160 channels after 2.2km of OM2 fiber. . . . . . . . . 120

4.32 FMF refractive index profile and neff for the six LP modes used in the experiment.123

4.33 Modal composition of propagating mode groups and detected mode group after
mode-group demultiplexer (MGD) by the photodetector. . . . . . . . . . . . . . 123

4.34 Comparison of mode-group multiplexer (MGM) and MGD designed for trans-
mission over FMF and OM2 in terms of generated crosstalk in the absence of
splices , and in the presence of two splices between the MGM and MGD with
20 m of FMF and OM2 fiber. . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
xxii LIST OF FIGURES

4.35 (a) Measured fiber crosstalk for each group using both systems: blanked mark-
ers with MGM and MGD designed for OM2 based transmission, filled markers
with MGM and MGD designed for FMF based transmission. (b) deduced fiber
crosstalk in dB/km per each group. . . . . . . . . . . . . . . . . . . . . . . . . . 126

4.36 Impact of fiber’s crosstalk in percentage versus distance (a) for OM2 fiber and
(b) for FMF. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

4.37 Measured losses generated by two systems:(a) MGM+MGD+OM2 and (b) MGM+MGD+FMF.12

4.38 Experimental setup for transmissionn over FMF. . . . . . . . . . . . . . . . . . . 128

4.39 Measured SNR response per mode group after 20km propagation over FMF at
1 dBm Rx input power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

4.40 Approximate model to estimate the chromatic dispersion value per mode group . 132

4.41 Measured BER after 20km of propagation using 4−PAM and DMT : (a) for G1
and G3 , (b) for G2 and G4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

4.42 BER stability due to variations in modal crosstalk for all mode groups modulated
at 55Gb/s DMT: a) for G2 and G4 b) for G1 and G3 . . . . . . . . . . . . . . . 134

4.43 Description of concept of polarization maintaining experimental setup. . . . . . 137

4.44 Refractive index profiles of 6-LP-mode FMF with inner core depressed-index (a)
with a step-index structure, (b) with a trapezoidal-index structure. . . . . . . . 138

4.45 Experimental setup used for crosstalk’s fluctuations investigation . . . . . . . . . 138

4.46 Fluctuation of BER for all mode groups in two different configurations in terms
of co-propagating mode group polarization state (cross and aligned polarization). 139

4.47 Histogram of 5 hours fluctuation of log10 (BER) for 100Gb/s 4-PAM and dis-
crete multitone modulation (DMT) for (a) cross polarization and (b) aligned
polarization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.48 5 hours BER fluctuation in the case of aligned polarization for 4-PAM and DMT.141

4.49 Measured (a) 4-PAM histograms and (b) SNR per subcarriers for DMT ,for the
lowest and highest BER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

4.50 (a)Allocated bit per symbol per subcarrier (b) allocated power per subcarrier,
for 100G DMT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

xxiii
xxiv
Chapter 1

Introduction

1.1 A Brief Introduction to IMDD Systems

In 1976, the first generation of optical fiber communication systems was deployed. It oper-
ated at 45 Mbit/s at a wavelength of 850 nm over 11 km of multimode fibers (MMFs) with
attenuation near to 3 dB/km [3]. This invention was possible due to two key technological
breakthroughs. The first was the birth of semiconductor lasers in 1962 [4]. These lasers were
made of gallium aluminium arsenide (GaAlAs) and operated at 850 nm. The second revolu-
tionary breakthrough was the drastic reduction of optical fiber attenuation. In fact, in 1960s,
the available optical fibers had enormous loss of 1000 dB/km. In 1966, Charles Kao suggested
to reduce the fiber loss by removing impurities from silica glass [5, 6]. Hence, he introduced the
optical fiber as a new form of communication medium with high potential. Four years later,
an American company called Corning took up the challenge and fabricated a single mode fiber
with losses around 20 dB/km at 630 nm [7]. In 1972, the same company reported a loss of only
4 dB/km at 850 nm using titanium with germanium as a dopant inside the multimode fiber
silica core [8]. In 1979, a Japanese company produced an optical fiber having loss of 0.2 dB/km
in the spectral window around 1.55 µm [9]. They won the race by achieving a value close to
the fundamental limit of Rayleigh scattering in single mode fiber (SMF).

1
2 Chapter 1. Introduction

The oldest optical communication systems were based on IMDD schemes, where the optical
power is generated by directly driving the semiconductor laser using a two-level electrical sig-
nal. The optical power is launched into the optical fiber. Due to the propagation impairments,
signal is degraded before being detected with a semiconductor device that converts the optical
power into electrical current.
IMDD systems dominated four generations of fiber-optic communication systems until the be-
ginning of 2000s. Details about generation characteristics is given in Table. 1.1 [1].

Table 1.1: First four generation of fiber optic communication systems [1]
Fiber-optic
1st 2nd 3rd 4th
generation
1st year commercial
1980 1987 1990 1995
availibility
Fiber type MMF SMF SMF SMF
Wavelength (nm) ∼850 ∼ 1310 1550 15
CD management   ✓ ✓
WDM & EDFA    ✓
Bit rates 45 Mb/s 1.7 Gb/s 10Gb/s 40Gb/s
Distance 10 km 50 km 80 km 1000’s km

Migration toward a second generation of fiber-optic communication system has been driven
by the need to increase the repeater spacing and reduce significantly the system cost. This
need has been met by operating in the infrared region of 1.3 µm thanks to development of new
semiconductors devices based on indium phospide (InP). Furthermore, the optical fiber Chro-
matic dispersion is minimal in this wavelength window. However, modal dispersion limits the
capacity to 100 Mbit/s. SMF deployment enabled to overcome this limitation and increased
the bit rates to 1.7 Gb/s for distances up to 50 km [8]. Chromatic dispersion management
appeared for systems > 10 Gb/s and enabled to increase both reach and the bit rates for the
third and second generation. Furthermore, deployment of optical amplification combined with
wavelength-division multiplexing (WDM) enhanced the throughput of optical communication
systems.
In the early 2000s, the advances in high-speed application-specific integrated circuits (ASIC)
and analog to digital converter (ADC), digital to analog converter (DAC), together with the
1.2. Datacenter Traffic Growth and Topology 3

evolution of the backbone network traffic generated by the appearance of the Internet, moti-
vated the growing interest for coherent detection technologies. From 2010, coherent detection
with advanced Digital Signal Processing (DSP) to compensate for system impairments took
over direct detection in many application, from metro to long haul networks. However, IMDD
systems remain the preferred technology for short reach applications since they have lower cost.
Furthermore, optical networks are moving towards the fourth industrial revolution (industry
4.0) which creates a critical need for high speed short range technologies for different applica-
tions such as machines interconnections, autonomous vehicle and advance robotic. However,
requirements in terms of high reliability, low latency and high capacity should be satisfied. To
achieve these requirements, computing resources should be placed closer to where data orig-
inates to reduce time of data transfer to the computing location such as the cloud. Hence,
datacenter architectures will move from big datacenters to small distributed datacenters to
provide low latency and high reliability. Consequently, to ensure high-speed connections be-
tween datacenters as well as intra-datacenters, short range systems with low cost, low power
consumption and high level of scalability are highly demanded.

datacenter to user 14.9%


25
intra-datacenter
datacenter to datacenter
20 datacenter to datacenter
13.6%
datacenter to user
Zettabit/year

15 intra-datacentre 71.5%

10

0
2016 2017 2018 2019 2020 2021

(a) (b)

Figure 1.1: Global data center (a) IP traffic growth and (b) traffic by destination in 2021
(source: Cisco Global Cloud index, 2016-2021)

1.2 Datacenter Traffic Growth and Topology


Overgrowing demand of internet services drives the need for a huge number of servers running
together inside and outside datacenters, which results in a considerable increase of datacenter
4 Chapter 1. Introduction

capacity. For example, the bisection bandwidth of Google’s datacenters has increased by a
factor of thousand over the past decade [10]. According to Fig. 1.1(a), the total datacenter
IP traffic will be tripled by 2021 compared to 2016 and this will be reaching 20.6 Zettabit.
The most important part of this traffic is inside the datacenters. Fig. 1.1(b) reveals that
by 2021, 71.5% of the total traffic will be within datacenters. The need of data transfer
between different clouds drives the fast growth of the traffic between datacenters. Inter data-
center traffic, will represent more than 14% of total datacenter traffic by 2021. 13.6 % of
total datacenter traffic flows from the datacenter to end customers through the internet. Each
portion of datacenter traffic is distributed inside or outside datacenter according to different
links. Therefore, academia and industry research are focused on relevant technologies to meet
this need for distances ranging from few meters to hundreds of meters.Cables as well as optical
components should be designed while satisfying cost requirements and allowing high capacity,
low latency, reliability and scalability. The most important part of the total traffic runs inside

Core Core Core Core


optical fibers
40G => 100G=>400G
Edge aggregation
Blocks
1G => 10G=>40G
Top o the rack
(TOR)

Servers

Figure 1.2: Intra-datacenter networks architecture: folded CLOS.

datacenters. A massive number of servers are deployed and connected to each others forming a
super computing structure. Today, folded CLOS type networks predominate in datacenters. A
typical intra-datacenter architecture is presented in Fig. 1.2. The first layer contains the racks
1.2. Datacenter Traffic Growth and Topology 5

Figure 1.3: Ethernet traffic growth published by Ethernet Alliance [2].

where the intra-rack connections are ensured using printed circuit boards. The top of each rack
(TOR) is directly connected to the edge aggregation switches using optical fibers. The switches
communicate together and with the core switches using optical interconnections. Assuming a
huge number of servers in a datacenter, the number of connections between switches is very
large and with a distance ranging from few meters to hundreds of meters with a maximum
distance of 2 km. Datacenter traffic is mostly Ethernet traffic and datacenter network speed
usually refers to the switching speed.
Fig. 1.3 gives a high-level view of Ethernet speed evolution published by Ethernet Alliance [2].
We distinguish three speed curves, the serial speed curves represent the servers speed while the
Quad Speed refers to the switch speed. For example, 40 Gb/s Quad speed is implemented using
four 10 Gb/s which is the server speed. Hence, 200 Gb/s datacenter network is realized with
four lanes 50 Gb/s. The recent 400 Gb/s network standards (IEEE 802.3) requires a highly
parallel speeds since 100 Gb/s server speeds will not be available before 2025. Therefore eight
lanes supporting 50 Gb/s or 16 lanes with 25 Gb/s are used for 400G datacenter networks.

Regarding the large number of optical connections in addition to different speeds between
servers and switches as well as switches and routers, different technologies based on different
types of fibers connectivity are used. Deployed optical fibers lengths depend on whether they
are deployed inside the same building or to connect different buildings in a datacenter. In the
first case, the reach ranges from few meters to hundreds of meters . In the second case, the
maximum distance is 2 km.
6 Chapter 1. Introduction

1.3 Datacenter: Transceivers and Optical Fiber Connec-

tivity

Transceivers represent a critical challenge for datacenter applications. In fact, it is strongly


desired to have a small form factors since they impact the rack size which defines the datacenter
structure size. Thus, the larger the size, the more energy is required for cooling. The most
common transceivers are the small form-factor pluggable (SFP) and quad small form-factor
pluggable (QSFP). The form factor have different optical connectors depending on the fiber
structure and the number of lanes.

1.3.1 State of the art : Product

short reach : SR
Datacenter 2 long reach : LR
Datacenter 1 Extended reach : ER

Datacenter 3
≤ 20 km Datacenter 4
LR, SMF
≤ 40 km
LR, SMF
≤ 80 km
300m ≤ 500 m
SR, MMF LR, SMF ER, SMF

Figure 1.4: Typical ranges of link distances categories for datacenter interconnect

Three main link distance type can be distinguished. Fig. 1.4 illustrates these links. The first
type is the short reach (SR) range. It represents the intra-datacenter connectivity including
connections within racks, switch-to-switch and between racks. Most SR interconnects are based
on multimode technology which combines MMF and 850 nm Vertical-cavity surface-emitting
lasers (VCSELs). Today, SR multimode optics are the lowest priced optical interconnects due
to the low manufacturing cost of VCSELs and the easy light coupling between the VCSELs
and the large fiber core. However modal dispersion limits the bit rates at 10 Gb/s at 1550 m
1.3. Datacenter: Transceivers and Optical Fiber Connectivity 7

with the most advance OM4 fiber (802.3 -Edition 2015, Clause 52). SR interconnects carry up
to 71.5% of total datacenter traffic. For distances higher than 500 m, technologies based on
SMF are used for long reach (LR) range and extended reach (ER). For each range, Ethernet
optical interface have a naming convention according to IEEE standards. They are given by
xx-GBASE-yyn where xx is the speed, yy is the reach range and n is the number of lanes.
For a bidirectional configuration, the naming is given by xx-GBASE-yyn.2 Table. 1.2 gives
an example of standardized and proposed optical interfaces for datacenter interconnect for
25 Gb/s, 40 Gb/s, 100 Gb/s, and 400 Gb/s. For each speed, there are two types of optical

Table 1.2: Example of standardized and proposed optical interfaces for datacenter interconnect
for 25Gb/s, 40Gb/s, 100Gb/s, and 400Gb/s
N° of
Speed Naming convension Format Baud rate Fiber Cable type N° of wavelength
fiber
25G BASE-SR NRZ 25.8G MMF Duplex 2 1
25Gb/s 25G BASE-FR
NRZ 25.8G SMF Duplex 2 1
25G BASE-LR

40G BASE-SR4 MPO parallel 8 1


NRZ 10.3G MMF
40G SWDM4 Duplex 2 4
40Gb/s 40G BASE-LR4 10.3G MPO parallel 2 4
40G BASE-ER4 NRZ 10.3G SMF Duplex 2 4
40G BASE-FR 41.25G Duplex 2 1
100GBASE-SR10 NRZ 10.3G MPO parallel 20 1
100GBASE-SR2 PAM4 26.6G MPO parallel 4 1
FMF
100GBASE-SR4 NRZ 25.8G MPO parallel 8 1
100Gb/s 100G SWDM4 NRZ 25.8G Duplex 2 4
100G CWDM4 NRZ
100GBASE-LR4 NRZ 25.8G SMF Duplex 2 4
100GBASE-ER4 NRZ
400GBASE-SR16 NRZ 32 1
26.6G MMF MPO parallel
400GBASE-SR4.2 4-PAM 8 2
400Gb/s
400GBASE-FR8
4-PAM 26.6G SMF Duplex 2 8
400GBASE-LR8

transceivers. One with multimode Fiber (MMF) and the other one with SMF. For MMF
based transceivers, scaling capacity from 40 Gb/s to 400 Gb/s is mainly based on parallel
architecture where multiple parallel optics (MPO) and connectors are required. With non
return to zero (NRZ) modulation format, the number of MMF is increased by a factor 4 for
400 Gb/s compared to 40 Gb/s where 32 parallel MMFs (16 in each direction) are required
as shown in table 1.2. Next generation of transceivers based on MMF will implement short
wavelength division multiplexing (SWDM) to reduce the number of fibers and increase the
transmitted bit rate. SWDM uses four channels centered at 850, 880, 910 and 940 nm (30 nm
8 Chapter 1. Introduction

spacing) to implement higher data rates in wideband MMF standardized as OM5 according to
IEEE. In fact, recent research, demonstrated up to 100 Gb/s using 4-PAM over 100 m of OM5
fiber [11].
For SMF based transceivers, the use of coarse wavelength division multiplexing (CWDM) with
wavelength in the infra-red spectrum reduces the fiber cost. Furthermore, CWDM enables the
use of uncooled lasers such as EMLs and directly modulated laser (DML) due to the large
wavelength spacing. For example 400-GBASE-LR8 uses 8 lanes with line speed of 50 Gb/s
4-PAM. Next generation of 400G transceivers interface will be 400G-BASE-DR4 which uses
parallel optic deploying 4 parallel SMF with 100G speed. Fig. 1.5 gives the transceivers
architecture for 400G-BASE-LR8 and 400G-BASE-DR4.

400GBASE-SR16 400GBASE-LR8 400GBASE-DR4


16x25G NRZ LR8: 8x50G 4-PAM DR4: 4x100G 4-PAM
Up to 100m Up to 10km Up to 500m
MMF λ1 SMF
Multimode mode MPO cable and connector (32 fiber)

Driver VCSEL Driver DML/EML λ1


Driver EML

Single mode MPO cable and connector (8 fiber)


MMF λ2 SMF
Driver VCSEL λ2
Driver DML/EML Driver EML
MMF λ3 SMF
Driver VCSEL Driver DML/EML λ3
Driver EML
...

...

λ4 SMF
Duplex LC connectors

Driver EML
MMF λ8
Driver VCSEL Driver DML/EML

MMF SMF
TIA PD TIA PD
TIA PD
MMF
TIA PD SMF
TIA PD TIA PD
MMF SMF
TIA PD TIA PD
TIA PD
...

...

SMF
TIA PD
MMF
TIA PD TIA PD

Figure 1.5: Different architectures for 400G transceivers for datacenter interconnects, from left
to right: 400G-BASE-SR16, 400G-BASE-LR8, 400G-BASE-DR4

1.3.2 Toward Tb/s Interconnect Technologies: a Critical Challenge

Section 1.3.1 reviewed the state of the art for data-center interconnect products where de-
ployment of 400G Ethernet is still in progress. However, datacenter traffic keeps growing and
1.3. Datacenter: Transceivers and Optical Fiber Connectivity 9

meeting its need seems to be very challenging with current technologies. For example, to scale
Ethernet traffic to 1.6 Tb/s, serial speed will require at least 200 Gb/s if the number of lanes
is kept at 8. In fact, there are three main axes to scale the capacity. First, increasing the
symbolrate enables to achieve higher capacity. However, the bandwidths of the optical and
electrical components represent strong limitations in terms of devices and cost. Furthermore,
chromatic dispersion sensitivity increases as the symbol rate increases. Secondly, the use of
high order modulation format such as eight levels PAM and sixteen levels PAM enables to
increase the bitrate while keeping lower component bandwidth. Nevertheless, this axis re-
quires a high signal to noise ratio which increases the total power consumption. Furthermore,
inter-symbol-interference (ISI) due to the propagation impairments and bandwidth limitation
requires complex DSP and channel coding to recover the signal. Finally, increasing the num-
ber of spatial channels is an attractive technique for scaling datacenters capacity. Currently,
CWDM and SWDM are proposed to increase single fibers speed. However, increasing the
number of lanes will reduce the wavelength spacing and limit the use of low cost uncooled lasers.
Dense wavelength division multiplexing (DWDM) requires power amplification and expensive
adapted optics. Currently, parallel fibers architectures are implemented. Fig. 1.6 depicts the
huge number of cables deployed inside a datacenter and published by IBM [12]. Up to 1536
cables are used per rack and thus increasing the number of fibers will be limited by the avail-
able space in the data centers. Otherwise, mega datacenters structures will be required. The
most relevant approach to deal with capacity scalability is spatial-division-multiplexing (SDM)
which allows the creation of independent spatial channels within the fiber. SDM gained the
attention of many researchers and several technologies have been proposed.

1.3.3 SDM in Support of Large Scale Datacenter Interconnects

From 2010, the availability of dedicated fibers for SDM resulted in an explosion of the number
of published papers on this topic [13]. At least 100 conference and journal papers are published
each year containing significant research on this topic. Fig. 1.7 illustrates different examples
of fiber cross section for the most common fibers used for SDM. These fibers are FMF in Fig.
10 Chapter 1. Introduction

Figure 1.6: Photo of massive number of optical cables installed in datacenter (source: IBM)

1.7(a)smaller than standard MMF depicted in Fig. 1.7(c). Elliptical core FMF (EC-FMF) in
Fig. 1.7(b) is a non-circular core fiber used for SDM. The main advantage of breaking core
circularity is to reduce mode coupling due to birefringence arising from core asymmetry. Fur-
thermore, multi-core fibers (MCF)s based on SDM are another family of SDM techniques [14].
Their cross section is depicted in Fig. 1.7(d). An MCF has a multiple cores for signal propa-
gation. Each core of an MCF can accommodate single mode or number of modes depending
on fiber design and the multiplexing technique.

(b)

(a) (b) (c) (d) (e)

Figure 1.7: Cross sections of most common optical fibers driving research on SDM (a) SMF,
(b) FMF, (c) standard MMF, (d) elliptical core fiber, (e) multi-cores fiber

SDM based coherent systems demonstrated significant transmission records. The latest record
demonstrated a transmission 10 Petabits/s over a 6-mode, 19-core fiber using both C + L bands
[15]. This work comprised 84,246 SDM and WDM channels. Such high capacity is achieved
with coherent detection thanks to multi-input multi-output (MIMO) DSP [16] which allows
to significantly compensate the crosstalk between cores (w.r.t modes). On the other hand,
1.4. Thesis outline 11

SDM based IMDD systems focus on carefully designing different components such as fibers
and (de)-multiplexers.

MCFs have been investigated for SDM based IMDD schemes. Recently, transmission of 1.176
Tb/s over 10 m of 7 cores single mode fiber has been reported using 8-PAM modulation format
[17]. Furthermore, 533 Gb/s transmission over 10 km of 7-core MCF has been achieved in
direct detection systems using discrete multitone modulation format [18]. Another approach
for SDM dedicated to IMDD systems is photonic lanterns. This technique arose from the
field of astrophotonics [19]. Photonic lanterns technique is based on adiabatically merging
several single-mode cores into one multimode core. In this context, EC-FMF mode-selective
photonic lanterns have been designed for MIMO free SDM Systems [20]. Furthermore 3x10
Gb/s have been reported over 20 km of FMF using photonic lanterns with IMDD [21]. Fiber
orbital angular momentum (OAM) modes have been investigated in [22, 23]. Generation of
OAM modes with different topological number l is traditionally based on free space systems
such as spiral phase plates (SPPs) and spatial light modulator (SLM) [24]. In fact, OAM
generation is usually affected by atmospheric turbulences and has significant transmission losses.
2x20Gb/s over 18.4 km of graded-index ring-core fiber has been achieved with this technique
[25]. Recently, six modes all-fiber OAM devices has been proposed and fabricated based on
conventional OM3 fiber [26] with less than 8 dB of insertion loss.

1.4 Thesis outline

In this work, we aim to investigate suitable technologies for future short range optical communi-
cations. Our work lies in the core of the spatial division multiplexing based IMDD area. Scaling
capacity over a single fiber while keeping low cost direct detection systems will be assessed.
This work will cover two types of distances. The first one covers short distances up to 4 km such
as intra-datacenter, inter-datacenter buildings in addition to local area networks (LANs) where
standard multimode fibers are already deployed. In fact, future technological revolutions will
cover many areas for the purpose of automation of everything. For example, factories, hospitals
12 Chapter 1. Introduction

and universities will implement revolutionary application such as augmented reality, internet
of thing and virtual reality. Person-to-machine and machine-to-machine communication will
enable to connect physical and digital worlds. However, optical networks should support the
revolution industry requirements in terms of low latency, high capacity and reliability. Cur-
rently, local area networks (LAN) have 10 Gb/s Ethernet data traffic, the optical infrastructure
is based on standard MMF for distances shorter than 300 m. To meet the need of future ap-
plications, the optical network should enable higher capacity. Nowadays, a French company
called CAIlabs proposed to upgrade the Ethernet traffic within LAN using an innovative so-
lution that is easily installed to upgrade existing multimode optical networks. Their solution
is based on space division multiplexing on MMFs based on the multi plane light conversion
(MPLC) technology. Hence, this approach delivers a considerable increase in bandwidth with-
out requiring complex and costly deployment of new generation optical fibers. This solution
enabled to upgrade the LAN data traffic from 10Gb/s to 40 Gb/s creating 4 spatial channel
within the MMF each carrying 10 Gb/s. However, with the ever-growing demand on data
traffic, new and disruptive technologies are required to continue upgrading their transmission
throughput while keeping such low cost solution. The second types of application we are in-
terested in are inter-datacenter interconnects where distances are longer compared to the first
applications.They may reach up to 20 km. This thesis is structured as follows:

In the first chapter, we review the main key breakthroughs that have resulted in the birth of fiber
optic communication systems. Furthermore, the urgent need of high speed and scalable short
link systems with low cost, low power consumption and low latency is discussed. After that, we
focused on traffic growth in datacenters and the current deployed interface connectivity to meet
the ever-growing demand for higher capacity. As a part of this chapter, datacenter systems
architecture is presented where Folded CLOS type networks predominates. Speed evolution
of different layers have been detailed. Furthermore, we provided an overview of optical fibers
structures and transceivers implemented in different types of link for distances ranging from
few meters to 80 km. This overview enabled to investigate the challenges of large sclabality to
meet future requirement of Ethernet traffic. Spatial division multiplexing have been identified
as the most promising technology for scaling data rates inside and outside datacenters.
1.4. Thesis outline 13

Chapter 2 provides a concise review on generic IMDD transmitters (Tx) and receivers (Rx).
Advantages as well as disadvantages of some optical and electronic devices are explained. This
chapter further discusses the main aspect of different modulation formats used for IMDD sys-
tems, mainly 4-PAM and DMT formats. For 4-PAM classical transmitter (Tx) and receiver
(Rx), DSP are described. Furthermore, their impact on the performance is experimentally
evaluated using 56 GBd 4-PAM signal. In this case, the main source of ISI is the electronic
bandwith limitation. Moreover, the basic concept of DMT modulation format is reviewed and
loading algorithms are discussed. To further optimize systems at higher BER, an alternative
method is proposed and numerically compared with classical Chow’s bit loading algorithm.

Chapter 3 focuses on the optical fiber structure and impairments. Chromatic dispersion impact
on systems relying on IMDD is addressed. Mode division multiplexing is introduced and
inter-modal crosstalk originating from (de)multiplexers and fiber imperfections is reviewed.
Mode group division multiplexing is proposed to reduce the emphasized crosstalk between
modes having close propagation constants. Mode group multiplexing based on MPLC technique
is explained. Finally, as IMDD schemes represents the main pillar of this work, crosstalk
robustness of M-PAM and DMT format will be analytically compared in the case of localized
crosstalk.

Chapter 4 , experimentally demonstrates the achievable benefit of mode group multiplexing


technique to increase the transported capacity over already deployed standard MMF in IMDD
schemes. High speed multiterabits transmission experiments over single fiber will be reported.
Furthermore, advantages of mode group multiplexing for low cost applications adapted to
intra-datacenter environment will be explored. Moreover, transmission record of 14.5 Tb/s net
throughput will be presented in a bidirectional transmission over 2.2 km of conventional OM2
fiber using DMT modulation scheme and direct detection. The second part of this chapter, will
focus on extending the IMDD mode group multiplexing technique to longer reach applications
up to 20 km using FMF instead of standard MMF. First, a comparison between standard
MMF and six-LP modes FMF in terms of components design, generated crosstalk an losses
will be studied. The second comparison in terms of fiber’s impact will be assessed. Finally
a demonstration of 200 Gb/s net bitrate over 20 km of 6−LP modes FMF will be presented.
14 Chapter 1. Introduction

Experimental investigation of the impact of crosstalk fluctuations on both formats, 4-PAM and
DMT is reported.

Finally, we will provide a concise summary of achieved results and discuss the perspective and
challenges of their implementation in real word.
Chapter 2

Intensity Modulation Direct Detection


Systems

2.1 Introduction

The advent of IMDD technology for short reach applications is motivated by its numerous ad-
vantages. In fact, it requires low-cost electro-optical (E-O) components for both the transmitter
and the receiver. At the transmitter side, the intensity of light could be directly modulated
with multiple devices then directly detected at the receiver side.
In fact, IMDD links aim to simplify both the transmitter and the receiver hence their cost
is lower than coherent transmission schemes. In such systems, only the intensity of light is
detected with a unique photodiode, and consequently, the information about the phase of the
signal cannot be recovered. Therefore, it is not possible to digitally compensate some impair-
ments such as chromatic dispersion, as the phase of the signal is a key element for compensation.
Therefore, IMDD links are attractive for short reach applications where chromatic dispersion
is not an issue. In contrast to coherent systems where data can be encoded in phase and
amplitude, in IMDD systems, information is modulated according to amplitude only.

In the first section of this chapter, we give an overview of IMDD transmitter and receiver.
In the second section, we will discuss two modulation formats adapted to IMDD schemes,

15
16 Chapter 2. Intensity Modulation Direct Detection Systems

namely, pulse amplitude modulation (PAM) and DMT. For PAM format, examples of DSP
algorithms deployed to mitigate transmission impairments such as ISI due to the channel delay
are discussed. In this context, linear as well as non-linear equalizers are reviewed. Moreover,
this chapter will explain how the signal could be adapted to the channel response thanks to
different bit loading algorithms using DMT format.

2.2 IMDD optical characteristics

DATA in DATA out

electrical electrical
modulator demodulator electrical
domain
DC u(t) r(t)
n(t)
component
y(t)

optical optical channel optical optical


intensity h(t) photodetector domain
modulator

Figure 2.1: Generic model of IMDD link.

To investigate IMDD systems and understand their different applications, it is important to


introduce the general model of an IMDD link and to detail each block of this model. This model
is illustrated in Fig. 2.1. First incoming data is modulated to the adequate modulation format
using an electrical modulator. For IMDD schemes, only the intensity of light is modulated and
detected. Accordingly, to drive the optical intensity modulator, a unipolar signal is required.
To ensure the unipolarity, a direct current (DC) bias is added to the generated electrical signal
u(t). After the electrical-to-optical conversion in the optical intensity modulator, the optical
signal propagates over an optical channel with the impulse response h(t). At the receiver side,
the intensity is detected and converted to an electrical signal with an optical photoreceiver.
A trans-impedance amplifier could be used in addition to the photodetector to amplify the
photodetector output current and improve the receiver sensitivity. The DC component of the
electrical signal is then removed using a DC-block. An additive white gaussian noise (AWGN)
2.2. IMDD optical characteristics 17

representing the detector thermal noise is added to y(t) before being demodulated with an
electrical demodulator. In what follows, we will give a generic model of IMDD transmitter and
receiver respectively, as well as the classical DSP conducted in both of them.

2.2.1 IMDD Transmitter

A conceptual model of IMDD transmitter is depicted in Fig. 2.2.As shown in this figure, the
transmitter can be implemented according to two configurations. The first one is depicted in
Fig. 2.2(a) for DML and the second one in Fig. 2.2(b) for externally modulated laser (EML).
For both configurations, data is generated with a low complexity DSP. The DSP depends on
the modulation format and the transmission link’s characteristics. Generally, it includes data
encoding, pulse shaping and pre-emphasis filtering. These concepts will be introduced in the
rest of this chapter. DACs based on the complementary metal oxide semiconductor (CMOS)
technology enable data modulation in the electrical domain [27, 28]. Their analog bandwidth is
a limitation facing optical communications to achieve high data rates. The swing of electrical
signals which drives directly the laser (Fig. 2.2 (a)) or the modulator (Fig. 2.2 (b)), is fixed by
a driver amplifier after the digital to analog conversion.

DSP DSP

DAC Driver DAC Driver

To the optical To the optical


channel channel
Laser Laser Modulator

(a) (b)

Figure 2.2: Conceptual model of IMDD transmitter based on (a) DML and (b) EML.

Transmitters based on DML are the oldest form of transmitters and they produce general
intensity modulated (IM) optical signals in a simplistic manner [29]. In this case, the injected
current into the laser drives its active section. It impacts the carriers in the laser cavity which
18 Chapter 2. Intensity Modulation Direct Detection Systems

generates the photons according to the stimulated emission process. This efficient conversion
explains the low-cost nature of DMLs. Furthermore, transmitters based on DML are compact
with low power-consumption. However, DMLs exhibit a significant disadvantage which is the
frequency shift or the chirp effect[30]. In fact, changes in the stored carriers in the laser cavity
modify the refractive index inside this cavity due to the applied current modulation to the
active region of the semiconductor laser. Hence, it results in a change of the optical length of
the cavity where the resonant mode is shifted back and forth in frequency. Thus, the modu-
lated spectrum of the laser output is broadened [31, 32]. Hence, the interaction of laser chirp
with fiber’s chromatic dispersion causes signal distortion, mainly in applications at 1.55 µm.
Transmitters based on EMLs use an external intensity modulator which modulates the light’s
intensity. This mechanism decouples the light generation process from the intensity modulation
process. Hence, the laser properties do not change. Therefore, EMLs are advantageous over
DMLs due to their reduced chirp which depends on the modulator’s characteristics. Indeed,
they are massively applied for high speed data transmissions over long distances. For trans-
mitters based on DML, the laser’s technology can be vertical cavity emitting laser (VCSEL)
or distributed feedback laser (DFB). For transmitters based on EML, the DFBs are com-
monly used. Moreover, for EMLs, the intensity modulation can be ensured with two types of
modulators. The electro-absorption modulator (EAM) or LiNbO3 Mach-Zehneder modulator
(MZM).

Laser Technology

Optical laser sources operate around 850nm, 1.3 µm and 1.55 µm which respectively correspond
to the first, second and third window of the optical fiber. The firstly deployed semiconductor
lasers were the edge-emitting laser diodes[33, 34]. In this case, the active region is confined
between a double heterojunction as represented in Fig. 2.3(a). The laser beam is then emitted
in the parallel surface of the semiconductor wafer chip. DFB laser is an edge-emitting laser
where a periodical diffraction grating is directly incorporated along the active section. The
diffraction grating acts as a reflector which enables a single frequency operation[35]. Hence,
DFB laser is adapted to dense WDM applications. However it remains costly for short reach
2.2. IMDD optical characteristics 19

applications.
VCSEL have been proposed as a cheaper technology compared to edge-emitting lasers . In
1970, Iga et al. [36] have demonstrated the possibility to emit the laser beam perpendicularly
to its epitaxial wafer as illustrated in Fig. 2.3 (b). This result allowed remarkable cost savings
by reducing manufacturing costs. In fact, in contrast DFB, VCSEL allows testing in a wafer
form without being mounted [37, 38]. Furthermore, VCSEL discards the optical components
used for fiber coupling. It can be designed with a circular output beam with reduced numerical
aperture to be efficiently coupled to the optical fiber.

Distributed
grating

Electrical contact
Electrical contact
p-dopped
p-dopped

n-dopped n-dopped
Active section
substrate substrate

(a) Electrical contact (b)

Figure 2.3: Laser structure for (a) DFB and (b) VCSEL.

Modulator Technology

The principle of intensity modulation process, is given in Fig.2.4 by assuming a linear intensity
modulator. The figure plots the output optical power Popt as a function of the input driving
current u(t). When the driving current is below the threshold current Ith , no optical power is
emitted. If the driving current is above Ith , the output optical power is linearly proportional
to the driving current. To avoid a non-linear distortion, the bias current Ibias should satisfy:

IBias − upeak ≥ Ith (2.1)

in which upeak is the maximal amplitude value of the driving current u(t), i.e.:

upeak = max(|u(t)|) (2.2)


20 Chapter 2. Intensity Modulation Direct Detection Systems

Output optical power Popt


P0

Ith Applied current I

u(t)
IBias

Figure 2.4: Model of linear intensity modulator.

Assuming that Ith = 0, the output optical power is given by :

Popt (t) = α[Ibias + u(t)] (2.3)

where α is the quantum efficiency of the electrical current-to-optical power conversion. In fact,
equation (2.3) gives the electro-optical conversion for an ideal intensity modulator. In practice,
the implemented modulators have different conversion functions which poses some limitations.
These modulators are divided into two categories, electro absorption modulator (EAM) and
electro-optic MZM.

Intensity modulation for EAM is based on the modification of the absorption spectrum of the
semiconductor material due to the applied driving current [39, 40]. The transfer function of
an EAM is depicted in Fig. 2.5(a). The EAM structure is similar to semiconductor lasers,
however, the active region is designed for absorption instead of emission. The EAM is very
compact and can be easily integrated in the same chip as the DFB laser. This made EAM very
attractive for short reach applications. Nevertheless, EAM deployment is limited compared
to MZM modulators due to their chirp parameter. In fact, even if the laser is operated in
continuous wave (CW), the residual optical feedback from the output modulator laser causes
laser chirp. Furthermore, EAM have a reduced wavelength operating range and low output
2.2. IMDD optical characteristics 21

power. In chapter 4, we demonstrate high capacity (> 250 Gb/s) transmission using an EML
transmitter based on DFB and EAM over short distances (< 4 Km). The EML has been
designed by III-V labs. As a part of this work, in chapter 4, we will compare transmitters based

EAM MZM
1

Transmittance
Transmittance

quadrature
point
0 null
point
2𝑉𝜋

−1
−2𝑉𝜋 0 2𝑉𝜋
Voltage Voltage
(a) (b)

Figure 2.5: Characteristic transfer function for (a) EAM and (b) MZM.

on EAM and MZM.

In contrast with EAM, LiNbo3 MZM enables chirp-free modulation, high output power and
wavelength insensitivity due to its dependency on Pockels effect only [41, 42, 43]. In fact, its
structure consist of a Mach-zehnder interferometer with Y-branches and electrodes as depicted
in Fig.2.5 (a) Exploiting pockels effect, the phase difference between the two arms of MZM is
proportional to the driving voltage [ref pockels]. Hence the output optical field is given by :

h π π i
Eout (t) ∝ cos Ibias + u(t) (2.4)
2Vπ 2Vπ

where Vπ is the half-wave voltage representing the voltage that achieves a π phase shift between
the two arms. The output intensity is they given as follows:

1 1 hπ π i
Iout (t) = |Eout (t)|2 ∝ + cos Ibias + u(t) . (2.5)
2 2 Vπ Vπ

The MZM transfer function is not linear for both field and intensity as it can be seen in Fig.
2.5 (b). Therefore, to operate in the linear regime, the bias point where the amplitude is
maximally linear is the null point satisfying Ibias = −Vπ . This point is the optimal bias for
22 Chapter 2. Intensity Modulation Direct Detection Systems

coherent transmission. However, for IMDD systems, the optimal bias maximizing the linear
intensity response is the quadrature point where Ibias = − V2π . Despite the advantages of MZM
which are mainly represented by the satisfaction of dense WDM requirements, it remains a
costly technology for short range applications. Furthermore, its large size represent a strong
drawback for such applications.

2.2.2 IMDD Receiver

From the optical


channel RF
photodetector
amplifier

ADC

DSP

Figure 2.6: Conceptual model for IMDD receiver.

In Fig. 2.6, a common architecture of an IMDD receiver is represented. Light intensity is


directly detected using a photodetector which provides the optical to electrical conversion. The
electrical signal is the amplified with an RF amplifier then converted to a digital signal using
an ADC. Finally, DSP is applied to recover the signal. This depends on the modulation format
and channel’s impairments. In this chapter, we discuss the adaptation of the DSP for different
modulation formats for IMDD systems.

One of the most critical part in optical communications systems is the deployed photore-
ceiver. In fact, most common photoreceivers uses the PIN photodiode and avalanche photodiode
(APD). The APD has the main advantage of having a greater level of sensitivity compared to
PIN due to the avalanche effect[44, 45]. Hence it enables to detect low signal power and reduces
the impact of accumulated fiber attenuation and devices loss. However, the APD requires a
high operating voltage, which induces high non-linearities and provides higher level noise com-
pared to PIN. In fact, PIN has the advantage of reduced level of capacitance due to its small
2.3. Pulse Amplitude Modulation 23

active section which enables to reach a very high bandwidth [46]. Its active area ranges of
few µm which is suitable for light coupling with fibers with small cored diameter. However,
when the fiber core diameter increases, the active area should be increased for efficient light
coupling which limit the photodiode bandwidth. To improve the PIN photodiode’s sensitivity,
a trans impedance amplifier (TIA) is implemented to amplify the output current. A PIN+TIA
receiver provides a large sensitivity compared to only PIN based receiver [47]. In the context
of this thesis, the performance of both photodetectors is compared and studied in chapter 4.

2.3 Pulse Amplitude Modulation

As its name indicates, PAM is a form of modulation where the amplitude of the signal is
encoded according to a series of pulses. In this case, the transmitted symbols belong to the
real domain. It is the most viable modulation format for short-range optical communications
due to its simplicity. It is the only standardized format for datacenter interconnect (DCI) for
distances up to 80 Km as discussed in chapter 1. Recently, IEEE standards are using 4-PAM
signaling, namely for 100 GbE and 400 GbE (IEEE 802.3).

0 1
(a)
-A A

(b) 01 00 10 11

-3A -A A 3A
000 001 011 010 110 111 101 100
(c)
-7A -5A -3A -A A 3A 5A 7A

Figure 2.7: Example of a Gray bit-to-symbol mapping for: (a) BPSK, (b) 4-PAM and (c)
8-PAM.

Fig. 2.7 shows symbols encoded in amplitude levels according to a Gray bit-to-symbol mapping
for two levels binary phase shift keying (BPSK), four levels (4-PAM) and eight levels (8-PAM).
More specifically, this mapping reduces the bit-error in noisy channels where errors are often
24 Chapter 2. Intensity Modulation Direct Detection Systems

produced between neighbouring symbols. In fact, only one single bit error occur for any symbol
error (between consecutive symbols). In general, M -PAM signal is an encoded signal with M
levels where the minimum Euclidean distance between points is given by:

dmin = 2A (2.6)

where A is the inverse of the square root of the average power of the M-PAM signal. By
increasing the M-PAM levels, for a fixed baudrate the bit rates is increased. However, high
DAC and ADC resolution is required in addition to higher power.

2.3.1 Symbol and Bit Error Probability

To characterize the quality of the transmission for high data rate M -PAM signaling, BER
and signal to noise ratio (SNR) are the key metrics. BER denotes the number of erroneous
bits over the total number of transmitted bits. The SNR, as its name indicates, is the ratio
between the power of the signal and the power of noise, i.e.:

Psig
SN R = (2.7)
σ02

where Psig is the signal power and σ02 is the noise power variance. The SNR is independent
of the modulation format. For any M -PAM signal, assuming all M -PAM levels are equally
likely, the relationship between the BER and the SNR has been derived in [48] assuming a
Gaussian noise, it is given by:

r !
2(M − 1) 6
BER = Q SNR (2.8)
M log2 M M2 − 1

where Q is the Q-function and is defined as:


!

u2
Z
1
Q(x) = √ exp − du. (2.9)
2π x 2
2.3. Pulse Amplitude Modulation 25

In optical communications, by convention, the BER is sometimes converted to another metric,


the Q2 -factor, using the following formula:
!

Q2f actor = 20.log10 2.erfc−1 (2BER) . (2.10)

Furthermore, assuming the Gray-mapping, symbol error rate (SER) can be approximated
using equation 2.3.1. It is given by :

SER = BER × log2 M. (2.11)

An example of the probability density function (PDF) for the four levels of 4-PAM in the
presence of Gaussian noise is plotted in Fig. 2.8. The overlap between each PDF defines the
error that could be quantified with SER (resp. BER assuming Gray coding). Given the PAM
levels and the SNR, the SER could be numerically estimated using a Monte-Carlo simulations
[48].
Probability distribution

S1 S2 S3 S4

errors

Levels

Figure 2.8: PDF of each level of 4-PAM signal in the presence of Gaussian noise only.

Fig. 2.9 illustrates the BER as a function of SNR per symbol for different M -PAM modu-
lation formats (M = {2, 4, 8, 16}) under the impact of AWGN. We observe that for the same
probability of bit error, when the order of M -PAM increases the required SNR to increases.
For example, the required SNR per symbol for BER = 10−4 increases by around 7 dB when
M is doubled.
26 Chapter 2. Intensity Modulation Direct Detection Systems

M=16

M=8

BER
M=4

M=2

SNR per symbol [dB]

Figure 2.9: BER versus SNR

2.3.2 Classical M-PAM DSP

Resampling and clock


recovery

Matched filter

Eequalization
Bit to symbol mapping (FFE,DFE,..)

Pulse shaping and


De-mapping
resampling

Pre-emphasise filter FEC decoding

Tx-DSP Rx-DSP
(a) (b)

Figure 2.10: DSP flow at (a) the transmitter and (b) the receiver, for M −PAM implementation.

Generic signal processing for digital M -PAM is depicted in Fig. 2.10 for the Tx in (a) and Rx
in (b). The DSP at the Tx has lower complexity compared to the Rx. At the Tx, the most
important operations are the signal encoding which was discussed in the previous section, pulse
shaping and digital pre-compensation of partial (for example: only the DAC or the ADC) or
total analog bandwith (for example: both DAC and ADC) [49].
2.3. Pulse Amplitude Modulation 27

Pulse-shaping

Generated sequence of symbols should be converted into an appropriate waveform which is


adapted to the transmission over a band-limited channel. This problem can be written as:


X
x(t) = Sk g(t − kT )
n=1
∞ (2.12)
X
= Sk gk (t)
n=1

where Sk denotes the generated discrete symbols to be transmitted over the channel, gk (t)
is the pulse shape and T is the symbol period. gk (t) should be adapted to a band-limited
channel, accordingly, it should be designed to have a band-limited frequency response Gk (f ).
In optical communications, raised-cosine pulse is the most common pulse shape due to its
desirable spectrum. The frequency response of a raised-cosine pulse is given by:


1−β
if 0 ≤ |f | ≤



 T 2T


  h i
HRC (f ) = T πT 1−β 1−β 1+β (2.13)
1 + cos (|f | − ) if ≤ |f | ≤
2
 β 2T 2T 2T


 1+β
0 if |f | >

2T

where 0 ≤ β ≤ 1 is the roll-off factor, i.e., it is a metric for the excess bandwidth of HRC . In

HRC (f)

β=0

β=0.5

β=1

1 1 1 1 f
− −
𝑇 2𝑇 2𝑇 𝑇

Figure 2.11: Raised-cosine pulse spectrum with different roll-off factors.

Fig. 2.11, an example of a raised-cosine pulse spectrum with different roll-off factor is plotted.
28 Chapter 2. Intensity Modulation Direct Detection Systems

It can be seen that the excess bandwidth increases when increasing the roll-off factor. More
specifically, for β = 0.5, the excess bandwidth is 50% and for β = 1 the bandwidth occupied by
the signal is 100%. The impulse response of raised-cosine pulse obtained with a simple inverse
Fourier transformation denoted F −1 is given by:

hRC (t) = F −1 (HRC (f )) (2.14)

Finally, the transmitted signal can be written as:


X
x(t) = Sk hRC (t − kT ). (2.15)
n=1

Hence, by applying the pulse-shaping, we generate a band-limited signal with a maximum


1+β
frequency of 2T
. Furthermore, to minimize the effect of ISI induced by the channel delay, it
is common to apply a matched filter at the receiver. Hence, to satisfy the zero ISI condition,
the pulse shape at the transmitter denoted g(t) and the matched filter at the receiver denoted
f (t) should satisfy the following [ref proakis]:


Z 0 if l 6= k

gk (t)fl (t)dt = (2.16)

1 if l = k.

where
fk (t) = g(t − kT ) (2.17)

and
fl (t) = f (t − lT ). (2.18)

The above conditions can be achieved by splitting the raised cosine pulse shape between the
transmitter and the receiver. In this case, we have:

gk (t) = fk (t) = hRRC (t − kT ) (2.19)


2.3. Pulse Amplitude Modulation 29

where hRRC is the root-raised-cosine pulse shape and its frequency response is given according
to:
p
HRRC (f ) = HRC (f ) (2.20)

where HRC is given by equation (2.3.2).

Pre-emphasise Filter
This operation aims to boost the SNR of higher frequency signals. In fact, the limited −3 dB
analog bandwidth of the DAC and ADC induces signal distortion and limits the performance
of high baudrate transmission. To address this limitation, digital pre-emphasis is applied at
the transmitter to mitigate the low pass analog bandwith limitation [50, 51]. The applied
digital pre-emphasis filter may mitigate only the DAC limitation bandwith or both DAC and
ADC depending on their impact on the high baudrate signal. When considering only the
limitation of the analog bandwith of the DAC, the pre-emphasis filter can be estimated from
the measured spectrum using an optical spectrum analyzer. Hence, the digital pre-emphasis
filter is the inverse of the measured spectrum in a specific frequency window. In the case where
both DAC and ADC are considered, the digital pre-emphasis filter can be estimated from
the Rx according to a minimum mean square criterion. This approach will be experimentally
discussed in the following section.

At the Rx, the DSP is more complex compared to the Tx. First, the signal is resampled
and synchronized using over the clock-recovery block, it consists of compensating the sampling
frequency drift between the Tx (DAC) and the Rx (ADC). Then the matched filter, which
is the same filter applied for the pulse shaping block at the Tx is applied to reduce the ISI
which is given according to equation (2.3.2). Furthermore, digital equalization is applied to
recover the signal and finally feed-forward-error corrector (FEC) enables to achieve the error
free transmission.

Equalization

Propagation over a dispersive channel in addition to the bandwidth limitation and non linear-
ities of optical devices such as EML and DML generate significant ISI. Unfortunately, pulse
30 Chapter 2. Intensity Modulation Direct Detection Systems

shaping is not sufficient for ISI mitigation. In fact, equalizing the channel allows to mitigate
these distortions and reduces the ISI effect on the received signal. The principle of channel
digital equalization is depicted in Fig. 2.12 where an estimated digital filter equalizer G(z)
is applied to the received symbol yk to compensate channel propagation impairments. Chan-
𝑛𝑘

𝑠𝑘 𝑦𝑘 𝑧𝑘 Decision 𝑠Ƹ𝑘
Dispersive channel Equalizer filter
𝐻(𝑧) 𝐺(𝑧)

Figure 2.12: Principle of digital equalization.

nel equalization can be achieved with a linear equalizer such as linear mean square equalizer
(LMSE) [48], or non-linear equalize such as FFE and decision-feedback equalizer (DFE) [48].
The equalizer filters are optimized according to different cost functions such as zero forcing cri-
terion where the equalizer filter is the inverse of the Fourier transform of the channel response,
1
i,e, G(z) = H̃(z)
. This criterion is non optimal in the presence of additive noise because the
LMSE equalizer compensates the ISI at the expense of enhancing additive noise [48]. In fact,
applying the inverse of the channel impulse response amplifies the noise at frequencies where
the channel has a high attenuation. Moreover, this operation results in coloured noise which
complicates the implementation of optimal detectors. An alternative criterion that reduces
these unwanted effects is the minimum mean square error (MMSE) estimator. In this case
G(z) is designed based on minimising the mean square error (MSE):

!
ck |2 ]
GM M SE = arg min E[|Sk − S (2.21)
G(z)

where Sk and S
ck are respectively the k-th transmitted symbol and the its corresponding esti-

mated symbol with the MMSE criterion. Since, linear equalization generates noise colouration
and residual ISI, another approach is to use non-linear equalization with feedback decision to
remove ISI of detected symbol. Most popular equalizers are FFE and DFE. In this work only
FFE is implemented.
FFE block diagram is given in Fig. 2.13 where W (z) is a finite impulse response (FIR) filter
representing the most common components of FFE. W is then a vector of tap weight and
2.3. Pulse Amplitude Modulation 31

𝑛𝑘

𝑠𝑘 𝑦𝑘 Decision
Dispersive channel 𝑧𝑘 𝑠Ƹ𝑘
W(z)
𝐻(𝑧)

𝑒(𝑘)

Figure 2.13: Principle of FFE equalizer.

W = [w0 , w1 , . . . , wL−1 ] where L is the number of taps. The tap weights can be updated using
different algorithms such as ZF algorithm and least mean squares (LMS) algorithm. In this
work we used the most common adaptive technique which is direct decision LMS (DD-LMS).
In this case, we define the error between the desired symbol sbk and the FIR filter output symbol
zk as:

ek = sbk − zk . (2.22)

Based on a gradient decent method we minimize the cost function given by e2k . Hence the
(n + 1)-th weight tap is given by :

wn+1 = wn + µek Yk (2.23)

in which µ is the step size and Yk = [yk , yk−1 , . . . , yk−L+1 ] is the vector of the input signal of
the FIR filter W . The complexity of FFE implementation increases by increasing the number
of tap weights. In what follows, the impact of taps on the performances is assessed.

Forward Error Correction Codes

Channel equalization presented previously allows a significant reduction of ISI but does not
enable to reach very low BER that are of the order of 10−15 in current optical communication
systems. Therefore, to meet these stringent requirements, channel coding or forward-error
corrector (FEC) was adopted.
The main idea in FEC is to add redundant information to a bit sequence. After propagation
32 Chapter 2. Intensity Modulation Direct Detection Systems

over noisy channel, the FEC uses this redundancy to recover the transmitted data. The FEC
codes are characterized according to their code rate or overhead. The code rate is given by

K
r= (2.24)
N

in which K and N − K are the number of information and redundant bits respectively. The
overhead in % is given as
1−r
OH = . (2.25)
r

In 2009, the International Telecommunication Union-Telecommunication standardization sec-


tor (ITU-T) introduced a large recommendation to specify requirement of optical transport
networks (OTN) including the hard-decision FEC (HD-FEC) codes G.975.1 [52]. HD-FEC
codes are based one Reed Solomon (RS) codes [53]. The most common HD-FEC for short
reach application are those with 20% and 7% OH rate. Their pre-FEC BER limit are re-
spectively 1.2 × 10−2 and 3.8 × 10−3 . A new class of FEC codes which are the staircase codes
in ITU-T G.709 [54]. These codes enable to slightly relax the constraint on the pre-BER. For
example, with 7% OH, BER limit is 5 × 10−3 using staircase codes compared to 3.8 × 10−3
with HD-FEC. In this thesis, the FEC codes are not implemented. However, performance is
evaluated according to different BER limit ( FEC threshold).

112Gb/s 4-PAM DSP in back-to-back (B2B) Experiment

In this section, we highlight the advantages of the presented DSP blocks at the transmitter
(pulse-shaping and pre-emphasise filtering) and the receiver (FFE) in an experimental context
using 56 GBd 4-PAM. The experimental setup is depicted in Fig. 2.14. At the transmitter,
the digital 56 GBd 4-PAM signal is converted into the analog domain using 88 Gs/s CMOS
DAC with 22 GHz analog bandwidth. The transmitter uses an EML configuration where a
DFB laser is externally modulated using a LiNO3 -MZM. After few meters of propagation over
SMF which corresponds to B2B configuration, signal is detected with PIN+TIA photodetector
with 41 GHz bandwidth. 92 Gs/s ADC converts the electrical signal into the digital domain.
2.3. Pulse Amplitude Modulation 33

Offline-processing is then used to process received data.

DSP TX DAC 88Gs/s

Driver
SMF

DFB PIN+TIA
MZM ADC 92Gs/s DSP RX
LASER 42GHz
VOA

Figure 2.14: Experimental setup for 4-PAM DSP evaluation.

Firstly, 56 GBd 4-PAM is generated by applying only the root raised cosine (RRC) pulse-
shaping filter with a rolloff of 0.4. At the receiver, we estimate the channel according to
Y HX
maximum likelihood (ML) [48] criterion where H
b =
XH X
is the estimated channel, X the
transmitted training sequence and Y its corresponding received data. In this experiment, the
channel does not vary over time. Furthermore, it includes only the Tx and Rx impairments,
more specifically, the DAC and ADC analog bandwidth limitations. Hence, the channel is esti-
mated only once using all the 215 transmitted training sequences. The inverse of the estimated
channel is given in the inset of Fig. 2.14. Secondly, the inverse of the estimated channel is ap-
plied at the transmitter for pre-emphasising. The impact of applying pre-emphasis filter on the
performance is given in terms of electrical spectrum in Fig. 2.15(a), and BER in Fig. 2.15(b)
and frequency response of 51-taps FFE equalizer in Fig. 2.15(c). We observe in Fig. 2.15(a)
that the electrical spectrum, after applying the pre-emphasis filter, is flattened compared with
the absence of pre-emphasis filter. The impact of the pre-emphasis filter on the performance
can be clearly observed in Fig. 2.15(b) where BER is plotted as a function of Rx input power
assuming FFE with 51 taps. For a Rx input power lower than −5 dBm, the performances
are approximately the same, in this case, we are mainly limited by the photodetector’s noise
(more specifically, thermal noise). At −4 dBm, the pre-emphasised signal achieves the same
performance as the one at −1 dBm, hence, pre-emphasis gives a total of 3 dB more power
34 Chapter 2. Intensity Modulation Direct Detection Systems

margin. Furthermore, the best achieved BER at 0 dBm is 10−6 for pre-emphasised signal
while it reaches only 10−4 when pre-emphasis filter is not applied. For Rx input power higher
than 0 dBm, the BER increases, in this case, we are mainly limited by the non-linearity of
the PIN+TIA receiver. Fig. 2.15(c) plots the response of 51-taps FFE filter in the frequency
domain at 0 dBm, with and without pre-emphasis filter. For the latter, we observe that the
FFE equalizer enhances by a bigger factor higher frequencies which are highly impacted by the
analog bandwidth limitation. Increasing the number of taps beyond 51 taps does not improve
the performance which is saturated as shown in Fig. 2.16 where the BER is plotted as a func-
tion of FFE taps at 0 dBm assuming pre-emphasise filter. FFE complexity could be reduced
depending on the deployed FEC. For example, using KP4 FEC, 15 taps are sufficient.

(a) (b) (c)

Figure 2.15: Impact of applying pre-emphasis filter on the performance in terms of (a) electrical
spectrum ((b) BER and (c) frequency response of 51-taps FFE equalizer.

-2.5
Log10 (BER)

-4.5

-6.5
5 15 25 35 45 55 65

Number of FEE taps

Figure 2.16: Impact of the number of FFE taps on the BER


2.4. Discrete Multitone Modulation (DMT) 35

2.4 Discrete Multitone Modulation (DMT)

DMT is an important scheme for IMDD systems. It is a baseband version of the well-known
orthogonal frequency division multiplexing (OFDM) which is widely deployed for wireless ap-
plications such as wireless local area networks (LAN) and Wi-Fi [55]. DMT is commonly
adopted for high speed digital subscriber line (DSL) applications. In 1958, Collins Kineplex
modem was the first voiceband modem that uses multi-channel modulation format [56, 57].
Nowadays DMT is a very attractive format for short range optical communication systems
thanks to its high tolerance to transmission impairments by adapting the signal to the channel
response. A detailed study of the DMT format can be found in [58, 59, 60]. In this section,
the basic concept of conventional DMT is briefly reminded. Furthermore, advantages of DMT
such as loading algorithms are discussed and a new version of Chow’s rate adaptive bit loading
algorithm [58] is proposed. The synchronization technique used in this work is also described.

2.4.1 Basic Concept of Conventional DMT

In this paragraph, the basic building blocks of a DMT transmitter and receiver are presented.
The key difference with single carrier modulation formats like M-PAM, is that a high-speed
serial data sequence is divided into several parallel low-speed data streams. Each low-speed
data stream is modulated to one subcarrier, hence, the generated signal will carry multiple
subcarriers at different frequencies.
Fig. 2.17 illustrates the typical block diagram of a DMT modulator and demodulator. To gener-
ate a DMT signal, a high-speed serial bit sequence is converted to parallel binary sequences. For
each sequence, a number M of bits are grouped together and mapped to a dedicated complex
value n according to a specific M-array quadrature amplitude modulation (QAM) constellation.
The mapped symbols Cn with n = {0, . . . , N − 1} will be supported by N subcarriers. To have
a signal adapted to a low cost IMDD transmission, the generated DMT waveform should be
real-valued. Therefore, a Hermitian conjugate operation is required before the 2N -point inverse
fast Fourier transform (IFFT). Accordingly, 2N subcarriers are constructed where the first half
36 Chapter 2. Intensity Modulation Direct Detection Systems

supports the symbols Cn and the second half carries the complex conjugates of Cn according
to a Hermitian symmetry operation that converts the symbols C2N −n to Cn∗ . Consequently,
only half of the spectrum is utilized compared to a conventional orthogonal frequency division
multiplexing (OFDM) modulation. Moreover, we commonly put C0 = CN = 0 to remove the
DC component of the DMT signal, hence, the generated signal is a zero-mean signal.

DMT modulator

Serial bit Serial to QAM Hermitian Adding Parallel to


IFFT Resampling
sequence parallel mapping conjugate cyclic prefix serial

Bit and power


To the DAC
loading

DMT demodulator
Loading
algorithm From the ADC

Parallel to QAM Remove Serial o


Equalization FFT synchronization Resampling
serial demapping cyclic prefix parallel

Figure 2.17: Schematic block diagram of principle of DMT modulator and DMT demodulator.

2.4.2 Inverse Fourier Transform (IFFT)

The output sequence of the IFFT operation is real valued. In fact, the vectors resulting from
the IFFT are written as:
2N −1
1 X nk
v(k) = √ Cn W2N (2.26)
2N n=0
2.4. Discrete Multitone Modulation (DMT) 37

with

nk
nk 
W2N = exp j2π (2.27a)
2N
 N −1 2N −1 
1 X
nk
X
nk
v(k) = √ C0 + Xn W2N + CN + Xn W2N (2.27b)
2N n=1 n=N +1

or, for C0 = CN = 0, the vector v(k) is given as

N −1 2N −1 
1 X
nk
X
nk
v(k) = √ Cn W2N + Cn W2N . (2.28)
2N n=1 n=N +1

However, by doing the change of variables m = 2N − n, we obtain

2N −1 N −1
(2N −n)k
X X
nk
Cn W2N = Cn? W2N (2.29a)
n=N +1 n=1

(2N −n)k nk ?
and WN = (W2N ). (2.29b)

Hence,
N −1
1 X nk
+ Cn? (w2N
nk ?

v(k) = √ Cn W2N )
2N n=1
N −1
(2.30)
1 X nk
=√ 2<{Cn W2N }
2N n=1

with <{.} denotes the real part.


Finally, after the IFFT operation, v(k) is real-valued and given by:

2N −1
1 X  πnk 
v(k) = √ Cn exp j (2.31)
2N n=1 N

where k = {0, . . . , 2N − 1}.

2.4.3 Cyclic Prefix

Among the advantages of deploying the DMT modulation format is the use of cyclic prefix that
enables a better channel delay spread tolerance. As it can be seen in Fig. 2.17, the addition of
38 Chapter 2. Intensity Modulation Direct Detection Systems

a cyclic prefix (CP) takes place before the digital to analog conversion (DAC). The last NCP -
points of v(k) are copied to the beginning of v(k). The addition of the CP could be considered
as a phase shift in the Fourier domain. v(k) can be written as:

2N −1+N
X CP
1  k  NCP 
v(k) = √ Cn exp jπn exp − jπn
2N n=1
N N
2N −1+N
(2.32)
1 X CP k − NCP 

=√ Cn exp jπn
2N n=1
N

with k = {0, . . . , 2N − 1 + NCP }.


The (2N +NCP )-point sequence v(k) is called the DMT frame. When considering the sampling
rate of the digital to analog converter (DAC), v(k) can be written as:

2N −1+N
X CP
1  k − NCP 
v(k) = √ Cn exp j2πn ∆ts (2.33)
2N n=1
T

1
where ∆ts = fs
represents the sampling period and fs is the sampling frequency. The re-
sampled signal to be loaded in the DAC is given by:

2N −1+N
X CP
1
x(t) = √ v(k)δ(t − k∆ts ) (2.34)
2N n=1

where the operator δ is the Dirac operator defined as




1 if n = 0

δn = (2.35)

0 if n 6= 0.

The addition of the CP removes the frames interference due to the channel dispersion as il-
lustrated in Fig. 2.18. In the absence of the CP, i.e. NCP = 0, the propagating DMT signal
is dispersed due to the accumulated propagation delays. The insertion of NCP -points at the
beginning of each frame makes the signal less vulnerable to inter-frames interference and absorb
the effects of channel delay [60]. Therefore, the length of the cyclic prefix NCP can be calculated
when the channel delay value is known. The cyclic prefix should be chosen to be longer than
the accumulated delays during the propagation to eliminate the ISI [61]. Its TCP duration
2.4. Discrete Multitone Modulation (DMT) 39

Channel with
delay spread
t
DMT frame distorted DMT frame

NCP=0
frame 1 frame 2 frame 3 frame 4
t

NCP >> 0 NCP NCP NCP

frame 1 frame 2 frame 3 frame 4


t

Figure 2.18: Representation of the impact of cyclic prefix on inter-frame interference.

should be larger than the channel impulse duration and satisfies the following inequality:

c
TCP >> 2 |Dacc |.N.∆fs (2.36)
f02

where TCP is the cyclic prefix duration, Dacc is the accumulated dispersion in ps/(nm.km),
1
c = 299792458 m/s being the light velocity in vacuum , f0 Hz is the frequency, and ∆fs = Ts

is the subcarriers spacing, i.e., the inverse of the symbol rate. NCP is computed from (2.36) as
follows:
TCP
NCP = . (2.37)
∆ts

Finally, according to the central limit theorem, when N + Ncp is high, the generated DMT
signal to be loaded in the DAC is normally distributed with zero mean. Fig. 2.19 gives the
example of the normalized distribution of two generated DMT signals. The blue curve with
2N + NCP = 23 and the red curve with 2N + NCP = 26 . We observe that for small numbers
of subcarriers, the PDF of the negated signal is chaotic while 26 subcarriers are enough to
approach a Gaussian distribution, which is helpful for analytical investigations as it will be
discussed later in chapter 3.
40 Chapter 2. Intensity Modulation Direct Detection Systems

2N+NCP=23

PDF
2N+NCP=26

Time [UI]

Figure 2.19: Normalized distribution of of two generated DMT signals, with 2N + NCP = 23
(blue curve) and 2N + NCP = 26 (red curve).

2.4.4 Peak to Average Power Ratio

The generated Gaussian-distributed DMT signal may suffer from a major drawback which is
its high peak to average power ratio (PAPR). It is defined as the ratio between the maximum
amplitude of x(t) and its root mean square (rms) [62, 63]:

max|x(t)|
P AP R = p (2.38)
hx(t)2 i

In multi-carrier transmission systems such as OFDM and DMT, the sum of random M-array
QAM variables in the frequency domain causes a high peak apparition in the time domain
generates very large amplitude values. First, a high PAPR limits both the DAC at the
transmitter and the ADC at the receiver by increasing the quantization noise. Furthermore,
in practice, higher PAPRs induce non-linear distortions caused by optical devices such as the
optical intensity modulator. Fig. 2.20 plots the simulation curve of the PAPR in dB as a
function of 2N + Ncp points when each subcarrier carries independent random QAM symbols.
This figure illustrates the impact of subcarriers number on the PAPR. It is observed that the
PAPR is very sensitive to the 2N +NCP value and increases with the number of subcarriers until
a certain value after which it saturates as shown in Fig. 2.20. From 22 to 26 , the PAPR increases
by a factor of 3.4 dB, then saturates. It is important to control the PAPR for optical systems to
2.4. Discrete Multitone Modulation (DMT) 41

Figure 2.20: PAPR evolution as function of the number of subcarriers.

improve their performance. There are several methods to mitigate the PAPR like the selected
mapping method [64], gradient based optimization algorithms [65], Trellis-assisted constellation
selection [66], and chaotic active constellation extension[67]. However, these methods require
high complex implementation. The hard-clipping method is adopted for practical uses thanks to
its low complexity [68]. As its name indicates, this method suppresses the maximum amplitude
and convert it to a given threshold. The new signal envelop becomes:





 −Ath if x(t) < −Ath


w(t) = x(t) if − Ath ≤ x(t) ≤ Ath (2.39)





A

th if x(t) > Ath

where Ath is the amplitude threshold. Fig. 2.21 compares the generated DMT waveform before
and after applying such clipping with Ath = 0.045. After the clipping the amplitudes beyond
0.045 are suppressed. In this work, the clipping is applied by setting a specific percentage where
the parameter to be optimized is the clipping percent denoted as Pclipp . This method reuces
the PAPR by a percentage of Pclipp where:

max(|v(t)|) − Ath
Pclipp = × 100. (2.40)
max(|v(t)|)

Fig. 2.21 represents the case where Pclipp = 26.3%.


42 Chapter 2. Intensity Modulation Direct Detection Systems

before clipping after clipping

Signal amplitude
Signal amplitude

samples samples

Figure 2.21: Example of DMT waveform conversion with clipping technique forAth = 0.045

2.4.5 Synchronization

The received and transmit signal synchronization is crucial for signal recovery. It is important
to detect the beginning of each received symbol to avoid ISI as well as inter-carrier interference
(ICI) [69]. Various methods have been proposed for timing synchronization. The main differ-
ence between these methods is the design of symbol synchronization. Most of these algorithms
use a vector with repetitive parts for synchronization. In some cases, the patterns are repeated
with different signs. At the receiver, the detection of the beginning of the transmitted DMT
signal is based on the auto-correlation operation using these vectors. In practice, these vectors
can be implemented in the frequency domain as well as in the time domain. A well known
example of these algorithms is the Schmidt and Cox algorithm[70] which have been proposed
for OFDM systems. It uses two synchronization symbols with repetitive half. Another design
of preamble synchronization composed of Golay complementary pair is discussed in [71]. In this
thesis, Minn and Barghava algorithm [72] is implemented and the synchronization symbol is
added in the time domain as showed in Fig. 2.22. Training sequences for time synchronization
are inserted at the start of the DMT symbol, hence indicating the beginning of each detected
DMT symbol at the receiver. The training sequences are characterized by having L repeated
parts with different signs. To achieve synchronization, we maximize the timing metric Λ given
2.4. Discrete Multitone Modulation (DMT) 43

Symbol
synchronization
frame 1 frame 2 frame 3 frame 4 ….
t

DMT symbol to be
loaded in the ADC

Figure 2.22: Synchronization symbol insertion in the time domain.

by: !2
L |p(d)|
Λ = , (2.41)
L − 1 E(d)

with
L−2
X M
X −1
p(d) = b(k) y ? (d + kM + m)y(d + (k + 1)M + m), (2.42)
k=0 m=0

and
M
X −1 X
L−1
E(d) = |y(d + i + kM )|2 , (2.43)
i=0 m=0

where{p(k) : k = 0, 1, . . . .L − 2} denotes the sign of the repeated parts of the training se-
quence and M is the number of samples in each repeated part. The training sequence should
be designed to have a timing metric Λ with a steep rolloff to increase the precision of syn-
chronization and low PAPR to avoid non linear distortion. Golay complementary sequence
are widely deployed in this context due to their robust correlation properties and low PAPR.
Moreover, they enable to preserve the cyclic prefix structure. They could be inserted in both
time domain (TD) and frequency domain (FD). The synchronization symbol is then composed
of L repeated Golay sequences with a specific pattern with L signs. For example, Fig. 2.23
depicts a pattern with [+, +, +, −]. Minn & Barghava [72] give an example of training pattern
signs which gives the steeper rolloff obtained by computer search. Table 2.1 illustrates these
patterns for L = 4, 8 and 16. To show that the number of repeated parts L has an impact on
the precision of synchronization, we generate three different training sequences corresponding
to L = 4, 8 and 16. For each value of L, we use the first pattern presented in Table 2.1 to
generate the sequences. Then, these sequences are applied in the time domain and inserted
at the beginning of a DMT signal with N = 1024 and NCP = 0. The subcarriers carry the
same QAM constellation, in this case, QPSK (4-QAM). The length of the training sequences
44 Chapter 2. Intensity Modulation Direct Detection Systems

A A A -A

Figure 2.23: Example of time domain training sequence with 4 repeated patterns with different
sign excluding cyclic prefix (L = 4 and M = 16).

L Pattern
4 [−, +, −, −]
[+, +, +, −]
8 [+, +, −, −, +, −, −, −]
[−, +, +, −, −, −, +, −]
16 [+, −, −, +, +, +, −, −, +, −, +, +, −, +, −, −]
[−, −, +, +, −, +, +, −, −, −, −, +, +, +, +, −]

Table 2.1: Example of training pattern signs.

is the same as the number of subcarriers N . Assuming a noiseless ideal channel, Fig. 2.24
plots the time metric when using each training sequence. We notice that when L increases,
the peak is narrower and the rolloff is steeper which gives a higher precision in terms of time
synchronization.

2.4.6 Loading Algorithms

The main advantage of DMT modulation is bit loading. It provides the possibility to adapt
the signal to the channel response. There are two types of loading algorithms. The first type
is the rate adaptive bit loading that maximizes the data rate for a given fixed performance
(BER or Q2 -factor). The second type is the margin-adaptive bit loading that maximizes the
performance for a fixed data rate. As far as this work is concerned, only the rate-adaptive bit
loading algorithm has been investigated. The algorithm is based on the well-known Shannon
capacity formula:
C = 2M B log2 (1 + SNR) (2.44)
2.4. Discrete Multitone Modulation (DMT) 45

Time [UI]

Figure 2.24: Time domain metric Λ for L = 4, 8 and 16.

where M is the number of spatial channels, B is the bandwidth of the signal and the factor 2
reflects the two polarizations. Assuming only one polarization and a single spatial channel, the
Shannon capacity C is:
C = B log2 (1 + SNR). (2.45)

For multitone signals, the total capacity is the sum of each tone capacity cn :

N −1 N −1
X B X
C= cn = log2 (1 + SNRn ). (2.46)
n=0
N n=0

where SNRn is the SNR of the n-th subcarrier, given by:

SNRn = ξn .gn (2.47)

where ξn is the subcarrier power and gn , which depends on the channel response, represents
the subcarrier SNR when unit energy is applied:

|Hn |2
gn = (2.48)
σn2

with Hn and σn2 being the channel response and the AWGN variance at the n-th tone re-
spectively. In practice, the Shannon capacity is not reached due to transmission limitations
and impairments. These limitations could be represented by the parameter Γ. Hence, the real
46 Chapter 2. Intensity Modulation Direct Detection Systems

reached capacity R is below the ideal Shannon capacity and is given by:

N −1
!
B X SNRn
R= log2 1 + < C. (2.49)
N n=0 Γ

The constant Γ is the SNR gap that depends on the number bn of bits per symbol (i.e the
modulation format) and the SNRn via the ”gap approximation” formula:

!
SNRn
bn = log2 1+ . (2.50)
Γ

The expression of Γ is deduced from (2.50) and given by:

SNRn
Γ= − 1. (2.51)
2bn − 1

Fig. 2.25 shows bn as a function of SNRn for different values of Γ. We notice that for the
bn

SNRn[dB]

Figure 2.25: Achievable bit per symbol per subcarrier versus its SNR for different value of Γ.

same SNRn , value, the achievable bit per symbol bn increases when Γ is decreases. The case
where Γ = 0dB corresponds to the Shannon limit. In practice, for a given SNRn value, the bit
allocation operation is ensured by fixing the SNR gap Γ. Rate-adaptive bit loading algorithms
maximise R given in equation (2.49) by fixing the performance target included in the parameter
Γ. In fact, Γ can be expressed as a function of BER based on the analytical error probability
2.4. Discrete Multitone Modulation (DMT) 47

formula of M-QAM format given by [48]:

" r !!2 #
2 1 3
BERM −QAM = 1− 1 − 2(1 − √ )Q SNR . (2.52)
log2 (M ) M M −1

The QAM order is related to the number of bits per symbol via the equation:

bn = log2 M. (2.53)

Finally, using equations (2.51) and (2.52), the relation between the performance target per
subcarrier BER and the SNR gap Γ for square M-QAM modulation format is given by:

" !2 #
2 1  p 
BERM −QAM = 1− 1 − 2(1 − √ )Q 3(Γ + 1) . (2.54)
log2 (M ) M

Fig 2.26 illustrates the performance target BER as a function of Γ for different M -QAM
BER

Figure 2.26: Performance target (BER) as a function of SNR gap Γ for different M-QAM
constellation.

modulation formats. Γ is approximately the same for all the M -QAM formats when the BER
is lower than 10−5 . When the BER increases, the variance of the values of Γ between different
M-QAM formats starts to increase. More specifically, the lower order M-QAM (M = 2) is more
sensitive to Γ compared to higher order M-QAM (for e.g., M = 128). Most of the rate-adaptive
algorithms suppose that for a given BER target all the M-QAM constellations correspond to
48 Chapter 2. Intensity Modulation Direct Detection Systems

the same Γ. In fact, the error probability target value is set by fixing one corresponding Γ and
maximizing the data rate R subject to the energy constraint. The optimization problem can
be formulated as follows

N −1
!!
B X ξn .gn
max(R) = max log2 1+
ξn ξn N n=0 Γ
N −1
(2.55)
X
subject to ξn = ξtot
n=0

where ξtot is the total available energy in the system. The problem can be solved using Lagrange
multipliers. Consequently, we take the gradient of the Lagragian function corresponding to
(2.55) to zero as
∇ξn f (ξn ) + λ∇ξn h(ξn ) = 0 (2.56)

with
N −1
!
1 X ξn .gn
f (ξn ) = log2 1 + (2.57)
N n=0 Γ

and
N
X −1
h(ξn ) = ξn − ξtot . (2.58)
n=0

By differentiating with respect to ξn , equation (2.56) can be written as:

1 1
· Γ
+ λ = 0. (2.59)
ln(2) gn
+ ξn

The bit rate maximization problem is achieved with the well-known water-filling solution given
by:
Γ 1
+ ξn = − = constant. (2.60)
gn λ ln(2)
Γ
This solution is an illustration of the function gn
being a bowl where water which is the en-
ergy ξn is added to achieve a constant level of water. Some subcarriers may have a very low
Γ
signal to noise ratio, in this case gn
exceeds the constant level and thus these subcarriers are
eliminated and turned off. Computation of bit and power allocation based on water-filling
solution is detailed in [58]. This method can result in a non-integer values of bn which limits
2.4. Discrete Multitone Modulation (DMT) 49

its implementation. Therefore, different rate-adaptive bit loading algorithms are proposed to
achieve optimum bit and power allocation with fine granularity and approximate water-filling
solution. The most common one is Chow’s rate-adaptive bit loading [73]. It approximates the
water-filling solution with finite granularity for bn . This algorithm is based on a first decision
of the subcarriers that will be used for transmission and those eliminated according to their
SNR. The energy of discarded subcarriers is redistributed over the subcarriers with higher
SNR. Compared to water-filling approximation, the non-integer bn is rounded to the nearest
integer of bn and the corresponding ξn is respectively increased or decreased if bn ≤ [bn ] or
bn ≥ [bn ] respectively to achieve the target error probability. Chow’s algorithm uses a single
value of Γ for all the M -QAM formats, this gap-approximation is more accurate when the BER
is low, as shown in Fig 2.26. However, when the BER is increases, Γ has different values for
different modulation formats [73, 58]. In this case, it is not possible to reach exactly the same
target performance BER. In this thesis, we propose a more accurate method to deal with this
problem [74]. Instead of using equation (2.50) to determine the number of bits per symbol that
are allocated to each subcarrier according to their SNR, the required SNR for a given target
BER (denoted as B ? ) for any M -QAM format is computed using Monte Carlo simulations and
denoted as S(B ? , k) where k is related to M-QAM order by:

M = 2k + 1 (2.61)

with k = {0, 1, . . . , km ax} , where kmax is the maximum number of bits per symbol which can
be allocated to the subcarrier with the highest SNR for a fixed B ? . For each B ? , we use the
relationship between the BER and the SNR to determine the SNR thresholds denoted as Sk
for the bit and power allocation. The flow chart in Fig. 2.27 gives the process of thresholds
generation. Once the Sk are determined, the number of bits per symbol and the energy for each
subcarrier are allocated according to its SNR. If the SNR is lower than the first threshold S0 ,
the subcarrier is turned off and not used for transmission. Each SNRn is classified in a range
of SNR limited by two successive thresholds. bn is allocated according to these ranges. The
subcarrier energy is increased or decreased depending on whether the subcarrier’s SNR is above
or below the threshold. The flow chart in Fig. 2.28 illustrates bit and power allocation steps.
50 Chapter 2. Intensity Modulation Direct Detection Systems

Start

k=0

𝑆 𝐵∗ , 𝑘 + 1
𝑆𝑘 =
2

k=k+1

𝑆 𝐵∗ , 𝑘 + 𝑆 𝐵∗ , 𝑘 + 1
𝑆𝑘 =
2

k = k+1
yes No
k > kmax − 1

end

Figure 2.27: SNR threshold generation algorithm.

Fig. 2.29 depicts an example of bit and power allocation process for a target error probability
of 10−3 . Monte-Carlo simulations enables us to have the BER versus SNR curve even for non-
square M-QAM such as 8-QAM and 32-QAM, where a generalized analytical error probability
formula does not exist. In practice the BER versus SNR curves for all M-QAM formats are
first generated using Monte Carlo simulations then numerically interpolated. The interpolation
coefficients are then used to generate automatically the thresholds for any target BER. This bit
and power loading method ensures the same performance target over all the used subcarriers
for any attributed QAM format. Hence the average BER will be the same as the target.
2.4. Discrete Multitone Modulation (DMT) 51

Start

𝑁𝑜𝑛 = 𝑁

n =1

k=0

No yes
𝑆𝑁𝑅𝑛 ≥ S0

𝑁𝑜𝑛 = 𝑁 − 1
No yes
S𝑘+1 ≤ 𝑆𝑁𝑅𝑛 < S𝑘+2
𝑏𝑛 = 0

𝑏𝑛 = 𝑘 + 1
ξ𝑛 = 0
k = k+1

ξ𝑘 = 𝑆 𝐵∗ , 𝑘 + 1 − 𝑆𝑘
n= n+1

Compute the energy scaling factor:


𝑁𝑜𝑛
No yes
β = ෍ ξ𝑛 𝑛≥𝑁
𝑛=1
Scale all subcarrier energies to
ξ𝑛 ← β. ξ𝑛

end

Figure 2.28: bn and ξn attribution algorithm.

Numerical Comparison

In this section, Chow’s bit loading algorithm [ref] and the optimized method are compared by
simulation. The IFFT/FFT size is 1024, cyclic prefix length is 0 and clipping is neglected. The
occupied bandwidth signal is 56GHz. Only AWGN is considered by setting the total SNR to
18 dB. The propagation channel is assumed to have a frequency response H given by:

H(f ) = 2 cos(2π 2 Kf 2 ) (2.62)


52 Chapter 2. Intensity Modulation Direct Detection Systems

BER

S0 S1 S2 S3 S4 S5 S6

bn =1 bn =2 bn =3 bn =4 bn =5 bn =6

ξn is increased
ξn is decreased

SNR [dB]

Figure 2.29: Example of SNR threshold and bit allocation rules for a B ? of 10−3

where K is a constant representing the channel dispersion coefficient and it is set to 1.110− 21.
We use QPSK format for all subcarriers to estimate the SNR. The SNR response before ap-
plying the bit and power loading is given in Fig. 2.30.
Bit and power allocation are computed based on the SNR response. First, we use Chow’

22

16
SNR[dB]

10

4
0 8 16 24 32
Frequency[GHz]

Figure 2.30: Estimated SNR response using QPSK modulation for all the 1024 subcarriers
impacted by the channel response H(f).

s algorithm with SNR gap Γ corresponding to QPSK modulation format (see the curve cor-
responding to M = 4 in Fig. 2.26) to compute the allocated bit and power per subcarrier
for different BER target. Second, we apply the proposed method to compute the allocated
2.4. Discrete Multitone Modulation (DMT) 53

bit and power per subcarrier for the same BER target. In both cases, bit and power loading
are applied and the BER is computed over more than 219 samples. Fig. 2.31(a) plots the
achievable bitrate as a function of the BER target. Fig. 2.31(b) plots the computed BER
at each bitrate as a function of the BER target in for both methods. It is observed that the
proposed optimized method results in a higher bitrate while achieving the BER target. On the
contrary, Chows algorithm results in lower achievable bitrate as the target BER is not achieved
due to the non-optimized value of Γ. For BER target lower than 10−3 , both methods converge
toward the same performance, in this case, the gap approximation is valid and the fixed Γ is
approximately the same for all QAM formats.
To evaluate the case where Γ corresponds to the higher order QAM, we set the BER target to
1.7 × 10−2 and Γ corresponding to 64-QAM. When applying Chow’s algorithm using the value
of Γ corresponding to QPSK and 64-QAM, the achievable bit rate is respectively 84.8 Gb/s
and 78.4 Gb/s. Using the proposed method, the achievable bitrate is 81.5 Gb/s. For all cases,
we apply the computed bit and power allocation. Fig. 2.32 plots the averaged BER over all the
subcarriers carrying the same number of bits per symbol. We observe that only the BER of
subcarriers carrying the modulation format for which we fixed the parameter Γ matches with
the BER target. In contrast, the BER for all subcarriers matches with the BER target when
using the proposed method. In fact, for optical communications where the FEC limit should
be respected, the proposed method will give a good trade-off between the transmitted bitrate
and the targeted BER.

140

135
5.0E-03
130
Computed BER
Bitrate [Gb/s]

125

120

115 Optimized method Optimized method


Chow : SNR garp corresponding
Γ corresponding to QPSKto QPSK
Γ corresponding
Chow : SNR garp corresponding
to QPSKto QPSK target
110 5.0E-04
1.0E-03 1.0E-02 1.0E-03 1.0E-02
BER target BER target
(a) (b)

Figure 2.31: Achievable bitrate as a function of the BER target in (a) and the computed BER
at each bitrate as a function of the BER target in (b).
54 Chapter 2. Intensity Modulation Direct Detection Systems

ΓSNR gap corresponding


corresponding to QPSKto QPSK
Γ corresponding to 64-QAM
SNR gap corresponding to 64QAM
Optimized method
5.0E-02 Target

BER

5.0E-03
1 2 3 4 5 6
bit/symbol

Figure 2.32: Averaged BER over all the subcarriers carrying the same number of bit per symbol
versus bit/symbol

2.5 Summary

In the first part of this chapter, we have detailed the general model of IMDD transmission
links. Furthermore, the generic transmitter architecture was discussed. More specifically, the
advantages and disadvantages of DML based transmitters were reviewed. The DML advantages
include low cost and power consumption. As for the disadvantages, we mention the chirp
phenomena which limits the transmission distance. The latter can be avoided using EML
based transmitters.
In the second part, we have discussed the main aspects of different modulation format deployed
for IMDD schemes. First, the generation of multilevel signaling Pulse Amplitude Modulation
(PAM) was described. For the case of a linear intensity modulator, the performance in the
presence of additive Gaussian noise was assessed were different M-PAM formats are compared.
Pulse shaping technique has been introduced to determine the spectral width occupied by the
signal. Moreover, the main equalization block used to mitigate the Inter-symbol-interference
that impairs the optical signal has been reviewed. Finally, experiments were conducted to
evaluate the impact of classical DSP, namely, pre-emphasise filtering and FFE equalizer on the
BER improvements.
2.5. Summary 55

The last part of this chapter discussed the basic concept of DMT modulation format and the
main advantages of using this format. These include cyclic prefix to reduce the inter-carrier-
interference generated by the channel delay, and the rate-adaptive bit-loading algorithm that
allows the adaptation of the signal to the channel response. An alternative approach to ”gap
approximation” has been introduced in order to optimize the bit and power allocation for high
values of the BER. Moreover, the main disadvantages of DMT where presented. These include
high PAPR and clipping operation to reduce this effect. A critical point in DMT, which is time
synchronization, has been also discussed in this Chapter.

After introducing the details of both the receiver and the transmitter in this chapter, in addition
to the most common modulation formats as well as the most important DSP blocks in chapter
2, Chapter 3 will focus on the propagation channel in addition to different methods to increase
the data rate by investigating the spatial diversity within the channels.
Chapter 3

Space Division Multiplexing

3.1 Introduction

The optical fiber is designed to support single mode or multiple modes. When optical fibers
were first deployed, MMFs have been used for short distances. In this case, modal dispersion
was the main source of ISI. The introduction of single mode fibers eliminated the modal dis-
persion [8]. However, chromatic dispersion has been identified as a source of ISI. Each type of
fibers has a specific use case. For example, for datacenters, MMFs are deployed for distances
less than 500 m. In fact, MMFs are attractive for such applications due to their high tolerance
for connectors misalignment. Furthermore, they enable the use of directly modulated 850 nm
VCSELs which could be easily coupled to MMFs and has low power consumption and man-
ufacturing cost. However, the inter-modal dispersion limits the reach at high data rates in
MMFs which made single mode fiber the preferred solution for reaches longer than 500 m.

This chapter is organized as follows: In section 3.2, we will review the fundamentals of optical
fibers. Section 3.3 will discuss the optical fiber dispersion impairment in the context of IMDD
systems. Section 3.4 will focus on SDM as a good technique to reduce modal dispersion while
increasing the transported capacity over MMFs. In this section, we will discuss the inter-modal
crosstalk as a main limitation of SDM systems. Furthermore, advantages of mode-group-
division-multiplexing over mode-division-multiplexing will be explained. Finally, in section 3.5,

56
3.2. Optical fibers structure 57

modulation formats used for direct detection systems will be analytically compared in terms of
robustness to the localized crosstalk.

3.2 Optical fibers structure

A general optical fiber structure is represented in Fig. 3.1. The fiber is usually protected from
the environment degradation by many layers. The external one is the cable which contains
several fibers. Each fiber is surrounded by a minimum of two non-optical layers of polymer
called the coating. The first optical layer that contributes to light propagation is the cladding
made of silica and permits to confine the light in the core due to the total internal reflections
that occur at the cladding-core interface. These reflections are ensured due to the difference of
the refractive indices n1 and n2 of the core and the cladding respectively where n1 > n2 .

1st coating
cable
core
cladding

2nd coating

Figure 3.1: Standard optical fibers structure.

The core diameter as well as the fiber refractive profile are the main parameters to be opti-
mized depending on the target application. Assuming a weakly guiding approximation, the
propagation of light over the optical fiber is described according to the following equation of
propagation:
n2 ∂ 2 Ψ
∆2 Ψ − =0 (3.1)
c2 ∂t2
c
where Ψ is the electric field,, n is the refraction index in the fiber with n = ν
and ν is the phase
velocity in the optical fiber. The solution of equation (3.1) has the following form:

Ψ(r, φ, z, t) = ψ(r, φ)ei(ωt−βz) (3.2)


58 Chapter 3. Space Division Multiplexing

where (r, φ, z) are the cylindrical coordinates, β is the propagation constant and ω is the angular
frequency. Given the circular symmetry of the optical fiber’s cross section, the propagation
equation (3.1) can be expressed in the cylindrical coordinates system as follows:

∂ 2 Ψ(r, φ) 1 ∂Ψ(r, φ) 1 ∂ 2 Ψ(r, φ)


2
+ + 2 2
+ [k02 n2 (r) − β 2 ]Ψ(r, φ) = 0 (3.3)
∂r r ∂r r ∂φ

ω
where k0 = c
. The desired solution is obtained by performing a separation of variable and
modeling the solution as a product of two functions as follows:

Ψ(r, Φ) = R(r)Φ(φ). (3.4)

Thus, equation (3.3) can be rewritten as:

r 2 d2 R r dR 2 2 2 2 1 d2 Φ
+ + r [k0 n (r) − β ] = − . (3.5)
R dr2 R dr φ dφ2

The optical fiber is a circular waveguide which implies that Φ(φ) is a periodic function with a
period of 2π. Hence:
1 d2 Φ
= constant = −l2 (3.6)
φ dφ2

hence, the solution, is required to be 2π-periodic, therefore l needs to be positive integer. Using
equations (3.5) and (3.6) we have:

r2 d2 R r dR
2
+ + r2 [k02 n2 (r) − β 2 ] = l2 (3.7)
R dr R dr

Equation (3.6) is easily solved by taking the double integral with respect to φ. The solution of
(3.6) is constant when l = 0 and has two independent solutions when l > 0. These solutions
are given by:
Φ(φ) = {cos(lφ), sin(lφ)}. (3.8)

The solutions of equation (3.7) depend on the ranges of values of r. The first range is defined by
0 < r < a (i.e., in the core) and the second range is defined for r > a (i.e., in the cladding) where
a is the core diameter. The solutions are given by investigating the condition of continuity at
3.2. Optical fibers structure 59

the core-cladding interface (r = a) [75] in cylindrical coordinates:



Ψ0
· Jl ( U.r



Jl (U ) a
) if |r| ≤ a
R(r) = (3.9)
Ψ0
Kl ( W.r



Kl (W )
· a
) if |r| ≥ a

where Jl is the Bessel function of the first kind and Kl is the modified Bessel function of the
second kind. The parameters U and W are defined as

q
U = a n21 k02 − β 2 (3.10)

q
W = a β 2 − n22 k02 . (3.11)

Finally, the solution of (3.7) is given by:


 
 


 cos(lφ)

Ψ U.r
0
.J ( ) if|r| ≤ a

Jl (U ) l a 





 sin(lφ)

Ψl (r, φ) =  (3.12)

 
cos(lφ)

 
Ψ0 W.r
.Kl ( a ) if |r| ≥ a




 Kl (W ) 

 sin(lφ)

In equation (3.10) and (3.11), new parameters which are the normalized frequency V , and the
normalized propagation constant b are introduced, they are defined as follows:

√ q
V = U + W = k0 a n21 − n22
2 2 (3.13)

W2
b= (3.14)
V2

Using equations (3.12), (3.13) and (3.14), in addition to the condition of continuity of R(r) and
dR
dr
at r = a, the resulting fiber characteristic function is given by[75]:

√ √
√ Jl−1 ( V2 1 − b) √ Kl−1 ( V b)
2
1 − b. √ =− b √ if l ≤ 0. (3.15)
Jl ( V2 1 − b) Kl ( V2 b)

For each value of l, there exist a finite number m of values of b verifying equation (3.15).
60 Chapter 3. Space Division Multiplexing

N= 1 N= 2 N=4 N= 6 N= 9 N= 12 lm
1
01
0.9 11
0.8 21
02
0.7
31
0.6 12
41
0.5
b

22
0.4 03
51
0.3
32
0.2
13
0.1
42
0
0 1 2 3 4 5 6 7 8 9 10
V

Figure 3.2: Normalized propagation constant b versus the normalized frequency V

Each solution defines a non-degenerated mode of the optical fiber. These modes are commonly
referred by LPlm where LP stands for linearly polarized, l and m are respectively the number
of radial and azimuthal zeros. Furthermore, due to polarization, each LPlm mode is two times
degenerated for l = 0 and four times degenerated for l ≥ 1. In fact, the number of solutions
is equal to the number of values of b, which depends on the fiber structure, mainly, the core
diameter a and the refractive indices of the core and the cladding.
Fig. 3.2 depicts the normalized constant b as function of normalized frequency V . When V is
lower than 2.405, which is the cut-off frequency, only the LP01 mode propagates in the fiber.
This mode is known as the fundamental mode. In the this case, the fiber is known as SMF.
When V ≥ 2.405, the number of propagating modes N is higher than 1, in this case, the optical
fiber is a MMF. Some examples of intensity distribution of modes are illustrated in Fig. 3.3.

The fiber geometry, in addition to the signal wavelengths, define whether the fiber is a single-
mode fiber or a multi-mode fiber. The standard single-mode fiber (SSMF) are defined by
ISO/IEC 11801 according to OS1, OS2 and OS3 classes. SSMF are widely deployed in many
optical network architectures. The main limiting factors in the SSMFs are the attenuation,
3.3. Optical fiber dispersion characteristics 61

l=0, m=1 l=1, m=1 l=2, m=1 l=0, m=2

l=3, m=1 l=1, m=2 l=4, m=1 l=2, m=2

l=0, m=3 l=5, m=1 l=3, m=2 l=, m=3

Figure 3.3: Example of intensity profiles of different LPlm modes.

chromatic dispersion and nonlinearities. Modal dispersion is another impairment which is


present in the MMFs and limits their deployment. Nevertheless, MMFs allow to increase the
number of propagation channels by investigating its spatial diversity which is possible due to
the orthogonality of the LPlm modes [75, 76] in terms of inner product. This can be represented
as: 
1 if l = l0 & m = m0

2π +∞
Z Z 
Ψlm Ψ?l0 m0 drdφ = (3.16)
0 0  0 0
0 if
 l 6= l & m 6= m

3.3 Optical fiber dispersion characteristics

Chromatic dispersion as well as modal dispersion are an important impairment in optical fiber
communications. In fact, an optical pulse transmitted into the fiber will spread in the time
domain while propagating over the fiber. This leads to signal degradation and limits high data
rate transmissions over long distances. The dispersion phenomena is due to the difference of
propagation velocities among different components of radiations having different frequencies.
In fact, the refractive index is wavelength sensitive and depends on the dopants within the
materials according to Sellmeir dispersion formula [68, 77]. Hence, the propagation constant β
62 Chapter 3. Space Division Multiplexing

is also wavelength-dependent which results in the frequency dependence of the group delay:

1 dβ
τg = = . (3.17)
νg dω

Chromatic Dispersion (CD)) represents the first derivative of the group delay given in equation
(3.17). It is a combined effect of the waveguide dispersion and material dispersion [78] denoted
as β2 . Since different spatial modes have different propagation constants, the CD parameter
per mode is given by :
d2 βlm (ω) λ2
β2,lm = = −Dlm (3.18)
dω 2 2πc
 
where Dlm = d 1/νg /dλ is the dispersion coefficient of the LPlm mode. For SSMF, typical
value of the dispersion coefficient is D01 = 20ps/(nm.km). Considering only the impact of
chromatic dispersion, and all the spatial modes are independent, each LP mode propagates
over independent dispersive channel having the following transfer function [ref]:

2β 2L
Hlm (f ) = e−i2π 2,lm f
(3.19)

where L is the fiber length. In the case of coherent transmission, pulse broadening due to βlm
can be fully compensated with DSP thanks to phase information availability. In contrast with
coherent systems, for low cost IMDD systems, chromatic dispersion is not managed which
limits transmission of high data rates due to power fading effect. In fact, assuming a noiseless
dispersive channel, the received signal can be modeled by :

y(t) = |(C + x(t)) ∗ hlm (t)|2


(3.20)
= C 2 + 2Cx(t) ∗ R{hlm (t)} + |x(t) ∗ hlm (t)|2

where ∗ denotes the convolution operator, C being the optical carrier, x(t) the real-valued
transmitted data and hlm (t) the inverse Fourier transform of Hlm given in equation (3.19). The
first term is the DC component, the second term is the desired signal while the third term
represents the signal to signal beating interference (SSBI). By neglecting the SSBI as well as
3.3. Optical fiber dispersion characteristics 63

the DC component, the received signal can be approximated by:

y(t) ≈ 2Cx(t) ∗ R{hlm (t)}. (3.21)

In the frequency domain we have :

Y (f ) ≈ 2CX(f ) × R{Hlm (f )}. (3.22)

Hence, in IMDD systems, the impact of dispersive channel is given by its real part which is
written as:
  πD λ2 f 2 L 

2 2 lm
R{Hlm (f )} = cos 2π β2,lm f L = cos . (3.23)
c

Power fading effect arises when the cosine term is equal to 0 which leads to spectral zeros in
the signal spectrum. Operating at a fixed wavelength λ, power fading strongly depends on the
dispersion coefficient Dlm , the propagation distance and the signal bandwidth.
ℱ −1 (ℛ 𝐻𝑙𝑚 (𝑓) )
ℛ 𝐻𝑙𝑚 (𝑓)

(a) (b)

Figure 3.4: Power fading impact illustrated by (a)the magnitude response of R{Hlm (f )} as
a function of frequency for different fiber lengths (L=2.2,4.4 and 20km) for 20ps/(nm.km) of
chromatic dispersion (b) their corresponding channel impulse response

Assuming a signal baudrate of 56 GBd, operating at 1550 nm and a chromatic dispersion


coefficient of 20 ps/(nm.km). Fig. 3.4 (a) gives the magnitude response of R{Hlm (f )} in dB as
a function of the frequency for different fiber lengths namely 2.2, 4.4, and 20 km. The impulse
response given by the inverse Fourier transform (denoted as F −1 ) of R{Hlm (f )} is plotted in
Fig. 3.4 (b). We notice that, for 2.2km the impact of CD is neglected due to the absence of
64 Chapter 3. Space Division Multiplexing

power fading. Furthermore, we notice that the numbers of spectral zero frequencies increases
as the fiber length increases. The first spectral zero frequency is 26.6 GHz at 4.4 km while
it is 12.46GHz at 20 km. In this case, the use of a single carrier modulation format such as
M-PAM will be strongly limited for baudrate higher than 53.2 GBd at 4.4km and 25GBd at
20km. dispersion-compensating fiber (DCF) have been proposed in [79] for CD compensation,
however , their cost and high attenuation limit their deployment for short reach distances.
Recently, research on reducing the impact of CD with direct detection have been proposed in
the case of 4-PAM format. For example, a first trail of 60 Gb/s net bitrate over 30km of SSMF
have been demonstrated based on Tomlinson-Harashima Precoding (THP) [80, 81]. However,
DSP to combat power fading increases the system complexity. Multi-tones modulation format
as DMT are less sensitive to the power fading. In fact, adapting the signal to the channel
response allows the avoidance of spectral zero frequencies by turning off the subcarriers at
these frequencies. Hence, the error induced by CD can be reduced.
Furthermore, in the case where the optical fiber supports different spatial modes having
ℱ −1 (ℛ 𝐻𝑙𝑚 (𝑓) )
ℛ 𝐻𝑙𝑚 (𝑓)

(a) (b)

Figure 3.5: Power fading impact illustrated by (a) the magnitude response of R{Hlm (f )} as a
function of frequency for different dispersion coefficient (Dlm = 20, 25 and 30 ps./(nm.km)) at
20km (b) their corresponding channel impulse response

different Dlm , the impact of CD is not the same on all modes assuming that each mode propagate
over an independent channel within the optical fiber where each channel has the frequency
response Hlm (f ). Fig. 3.5 illustrates the impact of intra-modal dispersion or CD on three
spatial modes having a DLM of 20, 25 and 30 ps/(nm.km) at 20 km for a signal baudrate of 56
GBd operating at 1550 nm. We observe from Fig. 3.5 (a) that the power fading is not the same
for the three modes. Furthermore, the impulse response in Fig. 3.5 (b) is also different. In
3.4. Spatial Mode Division Multiplexing 65

the case where all modes are used for transmission, each mode sees its own dispersive channel.
This difference of chromatic dispersion is a difference of the slope of the group delay given
in equation (3.17). In fact, this difference of group delay is known as inter-modal dispersion
and has the unit of ps/m. It describes the signal degradation due to the different propagation
paths of the modes. Inter-modal dispersion is the main limiting factor for MMF deployment.
Differential mode group Delay or DMD has been proposed to quantify the modal dispersion
[82, 83]. The DMD measurement is based on short laser pulse launching into a SMF which
is scanned across the face of the MMF. Information about the difference between the slowest
and the fastest pulses is exploited to compute the DMD. Optimization of MMF parameters
such as the refractive profile namely graded-index MMF (GI-MMF) to reduce the DMD and
reach longer distance. A second approach for combating the modal dispersion is mode-division
multiplexing which consists of transmitting different data over different spatial modes. Hence,
the mode dependency is reduced. However, mode coupling limits the performances . Moreover,
given the difference in CD impact on each mode, the transported capacity over each group will
be also different. In the following section, spatial mode multiplexing is introduced.

3.4 Spatial Mode Division Multiplexing

Spatial diversity over optical fibers is an attractive solution to increase the transmission capacity
of the fiber [84]. Theoretically, each LPlm mode could be considered as an independent spatial
channel [75, 85] due to its orthogonality as expressed in equation (3.16). In fact, the optical
field at the input of the MMF will be decomposed according to an orthonormal basis formed
by the orthogonal eigenmodes as follows:

X
Ein (r, φ) = alm Ψlm (r, φ) (3.24)
l,m

where Ein is the launched optical field at the input of the MMF and alm is the weight of the
LPlm mode. alm is a complex coefficient because the propagating modes could have different
66 Chapter 3. Space Division Multiplexing

phases. The total optical power is given as:

X
P= |alm |2 . (3.25)
l,m

To investigate the modes orthogonality, we multiply equation (3.24) by Ψ?l0 m0 (r, φ) and integrate
it over the MMF cross section Ω:

ZZ X ZZ
Ein (r, φ)Ψ?l0 m0 (r, φ)drdφ = alm Ψlm (r, φ)Ψ?l0 m0 (r, φ)drdφ (3.26)
Ω l,m Ω

In the above equation, only the term alm remains when l = l0 and m = m0 due to modes
orthogonality and can be written as:

RR
E (r, φ)Ψ?lm (r, φ)drdφ
alm = RR in

. (3.27)

|Ψlm (r, φ)|2 drdφ

alm is the excitation coefficient of the LPlm mode. In the absence of mode coupling, after
propagation, the optical field is a linear combination of the guided modes and it can be expressed
as:
X
Eout (r, φ) = A0lm (ω, z)Ψlm (r, φ) (3.28)
l,m

where Alm (z) is a complex valued coefficient describing the amplitude and phase propagation.
It is given by :
Alm (z) = a0lm (z)e−iβLM (ω)z (3.29)

at z = 0, a00 is the excitation coefficient alm Hence spatial-division-multiplexing is an approach


that shapes the optical field distribution at the input of the MMF in order to excite only the
target mode. The excited mode is an independent channel within the fiber. The number of
excited modes creates an equivalent number of spatial paths in the fiber which increases the
data rate. Furthermore, the mode-division multiplexing technique involves the conversion of the
transmitted mode to independent output optical fields. However, in real world implementations,
this technique is limited by the inter-modal coupling.
3.4. Spatial Mode Division Multiplexing 67

3.4.1 Inter-modal Crosstalk

The inter-model crosstalk (IXT) describes the coupling between signals in different modes [86].
In fact, unintended crosstalk may arises from multiplexers [87] which generate a light field
which does not match exactly with the desired mode in the fiber, thus, this light field may
excite other undesired modes. Moreover, the optical fiber could be a source of crosstalk [88]due
to its manufacturing imperfections. For example, non-circularity of the core and variation
in the index profile. Furthermore, mechanical constraints on the fiber, such as splicing or
bending cause modes coupling. The latter impacts the quality of transmission as it will be
experimentally detailed later in chapter 4. For systems using coherent detection, MIMO DSP
is deployed to fully compensate linear crosstalk between modes [89, ?]. However, MIMO DSP
complexity increases as the number of multiplexed modes increases. For IMDD systems, mode
coupling is mainly reduced by carefully designing the optical components.
IXT can be studied according to a field coupling model [86, 90, 91] which describes the changes
of eigenmodes caused by mode coupling or power coupling model [92, 90, 93, 88], which is a
simplified model giving the coupling energy between modes. Assuming a power coupling model,
the crosstalk induced in mode x is defined as [94, 88] :

P
x6=y Px−y
XTx = (3.30)
Px−x

where XTx is the modal crosstalk on the x-th mode , Px−y is the transferred power from mode
y to mode x and Px−x is the power of x-th mode. To have a more general idea about the power
transfer process between different modes, we assume a field coupling model. In the absence of
crosstalk, the propagating field in the fiber is given by:

  
A (ω, z) 0 0 ... 0 Ψ (r, φ)
 01   01 
  

 0 A11a (ω, z) ... ... 0   Ψ11a (r, φ) 
 
 ..  
E(z, r, φ, ω) = 
 0 0 A11b (ω, z) ... .   Ψ
  11b (r, φ) 
 (3.31)
.. .. .. .. . .
  

 . . . . .. 
 .. 

  
0 0 0 . . . ALM (ω, z) ΨLM (r, φ)
68 Chapter 3. Space Division Multiplexing

where the indices a and b indicate the degenerate modes, L and M are the usual LPLM nomen-
clature and Alm (z) is given by equation (3.29). Assuming the field coupling model, when the
refractive index is perturbed, the propagation and coupling are described according to the
following equation [90]:

dAlm X
= −iβlm Alm + clm−l0 m0 Al0 m0 (3.32)
dz 0
l6=l m6=m 0

where clm−l0 m0 is the coupling coefficient from LPlm modes and LPl0 m0 . Considering only two
modes. For simplicity, we denote Ax the complex coefficient of desired mode and Ay the
complex coefficient of coupled mode. Assuming that only mode i, at z = 0 we have Ax (0) = 1
and Ax (1) = 0. Assuming that the perturbation does not depend on z, according to equation
(3.32) we have:
dAx
= −xβx Ax + cx−y Ay
dz (3.33)
dAy
= −iβy Ay + cy−x Ax
dz
Assuming the initial conditions, the solutions of equation 3.33 are given by [86]:

h ∆β i βx +βy
Ax (z) = cos(αz) − i sin(αz) e−i 2 z (3.34)
2

and
Cx−y βx +βy
Ay (z) = sin(αz)e−i 2 z (3.35)
α

where
∆β 2
α2 = + |Cx−y |2 . (3.36)
2

The power of modes x and y are given by:

Px (z) = h|Ax (z)|2 i


(3.37)
Py (z) = h|Ay (z)|2 i
3.4. Spatial Mode Division Multiplexing 69

where hi denotes an ensemble average. Hence, the amount of coupled power in mode x is
described according to:
1
|Ay (z)|2 = ∆β
2 sin(αz)2 . (3.38)
2|Cx−y |
+1

We notice that the quantity of coupled power is proportional to the ratio ∆β/|Cx−y |. In fact,
when ∆β  |Cx−y |, the maximum coupled power tend to unity and when ∆β  |Cx−y |, the
coupled power becomes small. Fig. 3.6 plots |Ay (z)|2 as a function of z for ∆β/|Cx−y | equal
to 0.25, 1 and 1.5 . We notice that power is coupled back and forth due to the sin(αz)2 factor.
Furthermore, for high ∆β, the coupled energy is lower compared to lower ∆β value. In fact,
even for weak perturbation, when mode propagation constants have close values, these modes
are highly coupled. The most common technique to reduce mode coupling is to increase the

Figure 3.6: Impact of ∆β(z) on coupled power between two modes

∆β by modifying the design of the fiber refractive index. In fact, optimizing special forms of
refractive index profile will increase ∆nef f given by:

λ∆β
∆nef f = . (3.39)

In this context, FMF are the most attractive fibers for refractive index profile manipulation
due to their simplicity in terms of design and manufacturing. Fig. 3.7 illustrates examples of
typical fiber’s refractive index profile published on conferences and journal papers. Introduction
70 Chapter 3. Space Division Multiplexing

of array an air-holes in a FMF has been proposed [95], where increasing the size of holes
increases the ∆nef f . Fiber’s profile referring to raised cosine function have been reported
[96]. Introduction of depressed zones in the center of the fiber index profile has been proposed
to reduce the ne f f of the cylindrically-symmetric spatial modes [97, 98]. Furthermore, ring-
assisted enabled to reduce the coupling by coinciding the ring structure with the trough of
mode power distribution of LP02 mode [99]. Moreover, inter-core crosstalk reduction has been
also approached based optimization of cores refractive profiles [100]. This techniques showed
promising results for SDM based IMDD systems, however, we have to find the optimum trade-
off between the ∆nef f LP modes and modes effective area Aef f that have to be as large as
possible to limit intra-mode non-linearity [97, 101, 102]. As far as this work is concerned,
crosstalk reduction using a weakly coupled 6-LP modes FMF introducing a depressed-index
zone in the center of the step-index core (having min(∆nef f ) = 1.6 × 10−3 ) will be compared
with induced crosstalk of a standard MMF.

A.Gulistan OSA S.Jiang, JLT, 2018 L.Wang Optics express 2014 M.Bigot, OFC,2019
continuum,2018

S.Jiang, JLT, 2018 S.Jiang, JLT, 2018


P.Sillard, JLT, 2014

Figure 3.7: Typical designs of fiber refractive index profiles published in conferences and journal
papers

Finally, inter-modal crosstalk depend on different factors. The fidelity of desired mode exci-
tation is the main factor inducing mode coupling. In this case, the crosstalk between modes
strongly depends on two different factors. First, it depends on the imperfection of mode exci-
3.4. Spatial Mode Division Multiplexing 71

tation and its enhancement for modes having close propagation constants. In fact, the modes
with the nearest propagation constant are highly coupled compared to those with large dif-
ference of propagation constants. As shown in section 3.2, the propagation constant depends
on the fiber characteristics (core diameter, refraction index..), hence, some research has been
conducted to reduce crosstalk. For example, MMFs with reduced number of modes known as
FMF has been proposed with different refractive index profiles to achieve low mode coupling.
Another approach for mode coupling reduction is to multiplex group of modes having nearly
propagation constants instead of multiplexing each mode independently. It consist of using
group of modes having close propagation constant as independent channels within the MMF
to reduce the crosstalk. This approach will be detailed in section 3.4.3. Moreover, to ensure
mode multiplexing and demultiplexing operations with minimum crosstalk, many technologies
have been proposed. In section 3.4.2, we will discuss some of them.

3.4.2 Mode-division-Multiplexing Techniques

Mode division multiplexing techniques aim to create a limited number of spatial channels within
the multimode fibre to boost the data rate per single fiber [103, 13].Therefore, two mandatory
operations which are mode multiplexing and demultiplexing are needed to carry out the trans-
mission. The mode multiplexer is a key component that generates the desired spatial mode
profiles, combines and injects them into the transmission fiber. In practice, it converts the
fundamental mode of a single mode fiber into one given mode profile in the multimode fiber.
This mode conversion could be done for N input single mode fibers to create N different spatial
modes. At the end of the transmission link, a mode demultiplexer is needed to separate the
spatial channels at the output of the multimode fiber. These two operations can be imple-
mented based on different techniques. Mode selectivity is a very challenging process. In fact,
the quality of the transmission as well as its degradation due to mode coupling depend strongly
on the reliability of the mode selectivity. Many techniques have been proposed in the literature,
each with its advantages and drawbacks in terms of cost, complexity, generated crosstalk and
loss.
72 Chapter 3. Space Division Multiplexing

Launching technique: The offset launching technique has been discussed in [104, 105]. This
technique consists of projecting a small circular spot onto the MMF cross section with different
radial positions. Reference [104] presented three different methods to implement the launching

SMF MMF

(a)
MMF
SMF

(b)

SMF MMF

(c)

Figure 3.8: Different methods of launching technique: (a) lens combination, (b) small spot size
injection and (c) SMF launch.

technique, as shown in Fig. 3.8. The first one represented in Fig. 3.8(a) employs a combina-
tion of two lenses to collimates the light beam while the second technique, represented in Fig.
3.8(b), focuses it as a small spot onto the MMF using a fiber lense positioned closer to the
MMF. Finally, the third technique in Fig. 3.8(c) is based on SMF lunch. This method is very
limited since only the lower order modes, typically the fundamental modes, are excited.
Spatial light modulator (SLM)
The principle of modes profile generation based on spatial light modulation is represented in

𝑺𝑳𝑴
𝐿𝑃01 𝐿2
𝐿1 𝐸𝑖𝑛_𝑆𝐿𝑀 𝐸𝑜𝑢𝑡_𝑆𝐿𝑀

𝐸𝑖𝑛

𝑓 𝑓 𝑓

Figure 3.9: Principle of modes profile generation based on spatial light modulation

Fig. 3.9 where the fundamental mode of an SMF is converted to the desired mode profile thanks
to the use of an spatial light modulator (SLM) [106, 3]. As explained in the previous section,
3.4. Spatial Mode Division Multiplexing 73

the incident field onto the MMF should have the same spatial distribution as the guided mode
LPlm to achieve proper mode excitation. Therefore the SLM transfer function T (fx , fy ) should
be : !
T (fx , fy ) = T F −1 Ψl,m (x, y) (3.40)

where T F −1 denotes the inverse Fourier transform symbol. fx and fy are the spatial frequencies
given by :
x
fx = (3.41)
λf

and
y
fy = (3.42)
λf

where λ is the signal wavelength and f the focal length. In fact, the SLM transfer function
should be optimized to increase the excitation coefficient alm given in equation (3.27). Moreover,
the main signal information is carried by its phase, consequently, in most cases, the amplitude
of T (fx , fy ) is set to 1. Hence, T (fx , fy ) is a phase transformation. Both, phase only SLM
and amplitude SLM could be based on liquid crystal on silicon display (LCOS) [?] which is a
reflection type SLM to get the mode profile matching the selected mode phase. The presented
method remains a costly technology in addition to its dependence on polarization.

Binary Phase Plates


Mode conversion based on binary phased plates (BPP) is a low cost technique compared to

𝑩𝑷𝑷

𝐿𝑃01 𝐿2
𝐿1 𝐸𝑖𝑛_𝑆𝐿𝑀 𝐸𝑜𝑢𝑡_𝑆𝐿𝑀

𝐸𝑖𝑛

𝑓 𝑓 𝑓

Figure 3.10: Principle of modes phase profile generation based on binary phase plates.

SLMS. In fact, SLM represented in Fig. 3.9 is replaced by a phase plate as shown in Fig.
3.10. The SMF fundamental mode is converted to higher order mode using the same Fourier
74 Chapter 3. Space Division Multiplexing

transform process as for SLM [107]. It is a special case of SLM where only the phase of the field
is modulated. Fig. 3.11 displays an example of these patterns for the first six linearly polarized
modes, LP01 , LP02 , LP11a , LP11b ,LP21a , and LP21b . As LP01 corresponds to the originate op-
tical beam, no mode conversion is needed. While the other phase masks are a combination of
multiple 0 and π transitions [108]. In fact, loss is is one of the main drawbacks of phase plates
employment. Loss increases as the number of multiplexed mode increases due to the use of
3dB beam combiners. Furthermore, one mask phase operates at only one wavelength. In fact,
the thickness of the phase determines the phase shift for a given wavelength. The phase mask
fabrication process is often based on substrate overcoated photoresist [109] or using holography
[110]. In both cases, in the presence of fabrication defects, the phase transition is not accurate.
Hence, the generated mode profile is different from the target mode. It is a linear combination
of all the spatial modes in the fiber which results in crosstalk.

𝐿𝑃 𝐿𝑃 𝐿𝑃 𝐿𝑃 𝐿𝑃 𝐿𝑃

𝐿𝑃 mode

Its corresponding π π
phase mask π π π π
π

Figure 3.11: Example of phase transition patters for the first six linearly polarized modes

Multi Plan light conversion


The multi-plane light conversion (MPLC) technique was invented in 2010 by the founders
of Cailabs [111] who demonstrated a programmable unitary spatial mode converter (UPMC).
Theoretically, any desired unitary transformation of the light field can be achieved using a suc-
cession of reflections on deformable mirror with Fourier transform separation. This succession
of transverse phase profiles is achieved using phase plates or SLMs and it can be used in trans-
mission or in reflection [112, 113]. In this thesis, we worked in collaboration with Cailabs, to
meet the requirements of new generation of multiplexers and demultiplexers suitable for IMDD
optical communications.
Fig. 3.12 illustrates the principle of mode conversion based on the MPLC technique where
3.4. Spatial Mode Division Multiplexing 75

three independent beams in blue, red, and green pass through seven optical phases plate and
optical Fourier transform to form three spatial mode profiles. In fact, [111] showed that with a
number K of input beams, the equivalent number of modes is generated using typically 2K + 1
phase profiles. Hence 3 input fundamental modes are converted into 3 modes using 7 phase
profiles. The MPLC process is a low loss process because it is adiabatic and does not imply
any intrinsic loss. The losses are generated due to the imperfections of optical element. A
key point is that the loss does not scale with number of modes unlike other schemes requiring
beam splitters. Moreover, the MPLC process is bidirectional and could be used in the reverse
direction to demultiplex the same modes. The realization of the MPLC technique presented

OFT OFT OFT OFT OFT OFT

Mode Converter
Phase plates &
Optical Fourier
Transform (OFT)

Transverse
intensity

propagation

Figure 3.12: Principle of MPLC mode conversion technique

in Fig. 3.12 requires 2K + 1 phase profiles which occupy a large space. To overcome the use
of such expensive and complex system, the MPLC process is implemented using a multipass
cavity [113, 114] as described in Fig. 3.13. In fact all the required transverse phase profiles
are printed in one reflective phase plate and located in different areas of the plate. The optical
Fourier transform is obtained with multiple reflections in the cavity formed by the spherical
mirror and the phase plate. To generate K spatial modes, 2K + 1 reflections are needed. Fig.
3.13 shows an example of 3 reflections on the phase plate. The fidelity of the mode conversion is
computed according to the overlap integral between the k-th generated field distribution Ein,k
and the desired theoretical mode LPlm of the multimode fiber Ψlm :

Z Z 2

?
ρlm = Ein,k (r, φ).Ψlm (r, φ)drdφ (3.43)

The overlap between the produced mode and the theoretical non desired mode LPl0 m0 with
{l0 , m0 } 6= {l, m} is the crosstalk of the produced mode generated by the multiplexer. The
76 Chapter 3. Space Division Multiplexing

Fiber array +
microlens array
D-shaped mirror
SMF
SMF
SMF
Input beams
Spherical Phase plate or
mirror SLM
Output beam

Figure 3.13: Principle of MPLC mode conversion technique in multipass cavity.

metric ρk corresponds to the squared absolute value of the excitation coefficient al,m given
earlier in equation (3.27). This technique demonstrated a high mode selectivity [115, 116, 15,
117, 114, 118]. However, inter modal crosstalk is still challenging due to fiber imperfections such
as micro-bending, macro-bending and splices. In this thesis, Mode group division multiplexing
based on MPLC process is proposed to deal with this challenge.

3.4.3 Spatial Mode-Group Multiplexing based on MPLC

The Mode-Group multiplexing technique is another approach of spatial multiplexing in MMF.


It consists of using groups of modes as independent spatial channels instead of each mode
independently. In fact, the groups of modes with the same propagation constant propagate at
nearly the same velocity while those with different propagation constant propagate at different
velocity. Hence, the modes in the fiber can be classified into different groups according to
their constant of propagation. For example, standard OM2 fiber supports 55 modes at 1550
nm which can be divided into 9 groups. Fig. 3.14 illustrates some of these groups in which
each white box indicates a specific group of modes. The principle of mode-group multiplexing
and demultiplexing is illustrated in Fig. 3.15. Standard single mode fiber are used at the
input of the mode group multiplexer (MGM), which converts the fundamental mode of these
fibers into one mode of the corresponding mode group. The MGM is based on the MPLC
technology explained in the previous section. In fact, even if the goal is to use groups of modes
instead of independent modes, it is not necessary to generate all the modes profiles to excite all
3.4. Spatial Mode Division Multiplexing 77

Figure 3.14: Example of groups of modes in OM2 fiber.

the modes of each group. Since all the modes in the same group are highly coupled, exciting
only one of them allows us to excite all the modes within the same group which reduces the
complexity. After propagation in the fiber, due to the large intra-mode group coupling, all
the modes of the same mode group need to be detected simultaneously in order to avoid large
power fluctuation at the receiver. The MGD is designed to selectively demultiplexe each mode
group. The MGD is based on the MPLC technology. Its design is more complex than the MDM
design due to the conversion of a higher number of mode profiles. The output of the MGD is
connected to multiple MMFs to detect the mode composition of each group. Furthermore one
SMF is connected to the output of the MGD and carries the first group which contains only
the fundamental mode. Different MGM and MGD configurations will be discussed in chapter
4. The design of these components depends on the target application in terms of distance
and deployed fiber. Indeed, the mode-group multiplexing technique tolerates the crosstalk by
avoiding the large coupling of modes sharing the same propagation constant. Unfortunately,
other sources of crosstalk between mode groups exist such as splices and fiber bending which
should be carefully optimized. However, modulation formats could represent some resilience as
well as vulnerability to the crosstalk. In the case of coherent detection, QAM formats have been
assessed in [119] where authors showed that the robustness to in-band crosstalk of QAM signals
is reduced as the number of levels in the QAM format increases. In section 3.5, we will assess
78 Chapter 3. Space Division Multiplexing

input 𝐺1 output 𝐺1
SMF SMF
input 𝐺2 Mode Group Mode Group output 𝐺2
SMF MMF
Multiplexer Demultiplexer



(MGM) (MGD)
MMF
converts N converts N spatial
input 𝐺𝑁 fundamental modes channels to N desired output 𝐺𝑁
SMF to N desired modes groups of modes MMF

Figure 3.15: Principle of mode-group multiplexing and demultiplexing

the impact of crosstalk on modulation format for direct detection, namely 4-PAM and DMT.
Therefore, we conduct a simplified theoretical investigation assuming a localized crosstalk and
that coupling occurs only between two modes. We will consider a coherent crosstalk, where the
crosstalk’s phase which is the relative phase between the signal of interest and the crosstalk
signal evolves randomly.

3.5 Crosstalk Investigation for IMDD Schemes

The crosstalk arises from multiplexers, demultiplexers, fibers and mostly non-optimized splices
[120] . Furthermore, as the relative phase between coupled modes variates randomly over time,
it causes fluctuations in the performance and errors at the receiver which are mainly given by the
BER metric [121]. Their mitigation for IMDD schemes is very challenging even with complex
digital signal processing (DSP). In case of coherent detection, generated in-band crosstalk due to
the effect of interfering connections at the same wavelength has been investigated for advanced
modulation formats in [119, 122]. Recently, inter-core crosstalk in multicore fiber systems has
been theoretically and experimentally characterized for on-off keying signaling in [123]. In this
thesis, we use mode-group multiplexing technique to reduce mode coupling impairment without
complex MIMO processing. However, the design of multiplexers and demultiplexers depends
on the modal composition and the number of excited mode groups. Furthermore, the mode
conversion is not perfect and crosstalk is generated by these components. In fact, the design of
mode group multiplexers and de-multiplexers as well as MMFs and splices optimization can
reduce the crosstalk impact on the transmission quality. Nevertheless, some modulation formats
3.5. Crosstalk Investigation for IMDD Schemes 79

are less vulnerable to crosstalk compared to other formats. In this work, the modulation format
can be either single carrier multilevel amplitude modulation (M-PAM) or DMT, as discussed
in chapter 2.

3.5.1 Crosstalk impact due to MGM, MGD and splices

In this section, we consider the main components of the crosstalk generated by multiplexers and
splices. Hence, we deal with localized crosstalk, and neglect the fiber crosstalk. Furthermore,
we do not consider the propagation effects such as chromatic dispersion. We conduct this study
to evaluate the impact of the crosstalk on 4-PAM and DM T formats using a simplified model.
More specifically, our model assume crosstalk between two signals where the relative phase
between these signals is random and uniformly distributed from 0 to 2π. Firstly, we derive
the analytical distribution of crosstalk impact in the case of 4-PAM and DMT modulation
formats. Secondly, for both format, we compare the impact of crosstalk on the BER in the
presence of AWGN using simulation and theory. Fig. 3.16 shows the model we adopt for the

𝑋 𝐸 (𝑡) RX1
. 𝐼 RX2
Analytical
𝐶1 investigation
n
𝑋 𝐸 (𝑡)

RX1
DMT 𝑋 𝐸 (𝑡) 𝐼 DMT 𝐼
Modulator . Demodulator RX2
(IFFT,..) (FFT,..)
𝐶1
n Analytical investigation
DMT 𝐸 (𝑡)
𝑋
Modulator
(IFFT,..)

Figure 3.16: Mathematical model for statistical crosstalk investigation in the case of M-PAM
signaling in (a) and DMT format in (b)

theoretical investigations in the case of M-PAM signaling in (a) and DMT in (b) neglecting the
80 Chapter 3. Space Division Multiplexing

propagation over the optical fiber. At the transmitter, we generate two independent signals
denoted as Xi , where i = 1 for the desired signal and i = 2 for the crosstalk signal. To model

𝐼𝑖
𝐸𝑖

DC component:
𝐶𝑖
0 Voltage

Data: 𝑋𝑖

Figure 3.17: Considered electro-optic conversion model in this study

the electrical to optical conversion, as shown in Fig. 3.17, a DC component denoted as Ci is


added to the generated signal Xi . We assume that the quantum efficiency of the electrical-
to-optical conversion is equal to 1. Furthermore, we assume to be in the linear region of the
transfer function of both intensity and field. The output optical field can be modeled as a linear
combination of Ci and Xi and it is given by:

Ei (t) = (Xi (t) + Ci )e(jωc t) for i = {1, 2} (3.44)

with ωc being the carrier frequency. We assume that both the optical field of the desired signal
and of the crosstalk signal have the same ωc . The coupling coefficient from mode 2 to mode 1
is given by :
c2−1 = ρeiθ (3.45)

where ρ is the crosstalk’s amplitude and θ is the crosstalk’s phase and which is uniformly
distributed in the interval [0, 2π]. The direct detection receiver detects the optical signal and
converts it into an electrical signal as:

I(t) = |Etot (t)|2 + n(t) (3.46)


3.5. Crosstalk Investigation for IMDD Schemes 81

where Etot (t) is detected optical field and n is the AWGN. Etot (t) is given by :

Etot (t) = E1 (t) + ρeiθ E2 (t). (3.47)

The photocurrent I, is given by :

I(t) = |E1 (t) + ρeiθ .E2 (t)| + n(t). (3.48)

Since X1 and X2 are real-valued signals, I can be written as:

 
I(t) = 2C1 X1 (t) + N , (3.49)

where N (t) is the term representing the total impairments impacting the signal of interest
X1 (t) including AWGN, signal-to-signal beating interference (SSBI) and the crosstalk. It can
be written as:
N (t) = C12 + X1 (t)2 + z(t) + n(t), (3.50)

where z(t) is the crosstalk contribution and given by:

    2
z(t) = 2ρ cos(θ) C1 + X1 (t) C2 + X2 (t) + C2 + X2 (t) ρ2 (3.51)

When |θ| = 0, z(t) is maximized which corresponds to the worst case scenario. For |θ| = π2 ,
 2
we have z(t) = C2 + X2 (t) ρ2 which corresponds to the best case scenario where the impact
of crosstalk is negligible. Hence, according to the phase variation, the penalties will fluctuate
π
between the worst case where |θ| = 0 and the best-case scenario |θ| = 2
. To characterize the
distribution of these fluctuations, in the following, we will derive the PDF of the crosstalk term
when X1 and X2 are M-PAM signals. In this case the investigation is done in the time domain
as shown inf Fig. 3.16(a). We also derive this PDF for X1 and X2 being DMT signals. In this
case, the investigation is done in Frequency domain as shown inf Fig. 3.16(b). The PDF of
the crosstalk term will be compared for both modulation format.
82 Chapter 3. Space Division Multiplexing

3.5.2 Crosstalk Impact Using the M-PAM Format

In this section, the impact of modal crosstalk in the case of multilevel pulse amplitude mod-
ulation is analytically approached. We consider {X1 , X2 } ∈ {A0 , . . . , AM −1 } where M is the
number of signal amplitude levels. The probability density function of the crosstalk term z is
given by :

PZ (k) = P (z = k|X2 = A0 )P (X2 = A0 ) + · · · + P (z = k|X2 = AM −1 )P (X2 = AM −1 ) (3.52)

where {A0 , . . . , AM −1 } are equiprobable :

1
P (X2 = A0 ) = · · · = P (X2 = AM −1 ) = (3.53)
M

hence
1
PZ (k) = [P (z = k|X2 = A0 ) + · · · + P (z = k|X2 = AM −1 )]. (3.54)
M

where z could be expressed according to :

   2
z = Y C2 + X2 + C2 + X2 ρ2 (3.55)

in which Y (t) is given as:


 
Y = 2ρcos(θ) C1 + X1 . (3.56)

First, the distribution of Y (t) is denoted as PY and is given by:

1h i
PY (y) = P (Y = y|X1 = A0 ) + · · · + P (Y = y|X1 = AM −1 ) . (3.57)
M

and 
  



 2ρcos(θ) C1 + A0 if X1 = A0


Y = ... (3.58)



  

2ρcos(θ) C1 + AM −1
 if X1 = AM −1
3.5. Crosstalk Investigation for IMDD Schemes 83

Given that the crosstalk phase θ is uniformly distributed from 0 to 2π, the distribution of
v = 2ρ cos(θ) is given by:
1 1
fV (v) = r  2 . (3.59)
4|ρ|π v
1 − 2ρ

Hence, for m ∈ [0, M − 1], PY is given by :

!
1 Y y 1
P (Y = y/X1 = Am ) = r (3.60)
4|ρ(C1 + Am )|π 4|ρ(C1 + Am )|  2
1 − 2ρ(C1y+Am )

Q
where is the function defined by:

1

1 if |t| ≤

Y 2
(t) = (3.61)
0 if |t| > 1 .


2

From equation (3.57) and (3.60), the PDF of Y is given by:

" M −1 ! #
1 X 1 Y y 1
PY (y) = r (3.62)
M 4|ρ(C1 + Am )|π 4|ρ(C1 + Am )| 2
m=0
1 − ( 2ρ(C1y+Am )

where U is the product of two independent random variables Y and C2 + X2 and is given by:

U = Y (C2 + X2 ), (3.63)

its pdf is given by :

" M −1 M −1 !
1 XX 1 Y u
PU (u) = 2
M i=0 m=0
4|ρ(C1 + Am )(C2 + Ai )|π 4|ρ(C1 + Am )(C2 + Ai )|
(3.64)
#
1
×r  2 .
u
1− 2ρ(C1 +Am )(C2 +Ai )
84 Chapter 3. Space Division Multiplexing

Finally the PDF of the crosstalk term z denoted as PZ is given by:

" M −1 M −1 !
1 XX 1 Y z − (C2 + Ai )2 ρ2
PZ (z) = 2
M i=0 m=0
4|ρ(C1 + Am )(C2 + Ai )|π 4|ρ(C1 + Am )(C2 + Ai )|
(3.65)
#
1
×r  2
z−(C2 +Ai )2 ρ2
1− 2ρ(C1 +Am )(C2 +Ai )

Fig. 3.18 plots an example of the crosstalk conditional probability denoted as P (z/X1 = Am )
for 2-PAM and 4-PAM signals having respectively M = 2 and M = 4. C1 and C2 have been
set to C1 = C2 = C and C > max(X1 ). For 2-PAM, X1 and X2 take up value in the set
of 2-PAM constellations defined by {−1, 1}. The conditional probabilities P (z/X1 = A0 ) and
P (z/X1 = A1 ) are plotted for different value of crosstalk’s intensity ρ2 , for 10 log10 (ρ2 ) equal to
-50dB, -25dB and -10dB. At -50 dB, the crosstalk’s impact is negligible, hence the conditional
probabilities P (z/X1 = A0 ) and P (z/X1 = A1 ) have approximately the same distribution while
they have different shape when the crosstalk increases. We observe that for higher amplitude
level the crosstalk spread is wider. In fact, the squared root term in equation (3.68) verify:

−2ρ(C1 + Am )(C2 + Ai ) + (C2 + Ai )2 ρ2 < z < 2ρ(C1 + Am )(C2 + Ai ) + (C2 + Ai )2 ρ2 (3.66)

Hence, increasing the amplitude level Am will increase the window of variation of z. Fur-
thermore, we observe the apparition of a defined number of peaks, each peak represents a
given combination between the transmitted symbol and its interference with each level of the
crosstalk’s signal X2 . Moreover, Fig. 3.18 depicts also an example of conditional distribution
4-PAM signal. In this case, we have the same observation as in the 2-PAM case. In fact, for
any M-PAM signal, a large part of errors is due to the high level amplitudes where the crosstalk
spread is more important. Including the SSBI, the PDF of w = z + X12 is given by :

" M −1 M −1 !
1 XX 1 Y w − ((C2 + Ai )2 ρ2 + A2m )
PW (w) = 2
M i=0 m=0
4|ρ(C1 + Am )(C2 + Ai )|π 4|ρ(C1 + Am )(C2 + Ai )|
#
1
×r  2
w−((C2 +Ai )2 +A2m )ρ2
1− 2ρ(C1 +Am )(C2 +Ai )
(3.67)
3.5. Crosstalk Investigation for IMDD Schemes 85

2-PAM 2-PAM 2-PAM

4-PAM 4-PAM 4-PAM

Figure 3.18: Example of the crosstalk impact on each level for 2-PAM and 4-PAM format at
different crosstalk’s intensity when theta variates uniformally from 0 to 2π

The PDF of total impairments N (t) including crosstalk, SSBI and AWGN given in equation
(3.50) is obtained according to a convolution operation of the PDF of w given in equation (3.67)
and the PDF of n denoted Pn . In fact, n ∼ N (0, σ02 ), hence Pn is the Gaussian distribution
with zero mean and variance σ02 :

(y−z)2
M −1 M −1 Z βi,m +γi,m −
1 1 XX 1 e N0
PN (y) = √ v !2 dz
2πM 2 πN0 i=0 m=0 γi,m βi,m −γi,m uu
t1 − z−βi,m
γi,m (3.68)
M −1 M −1
1 1 XX
= √ Wi,m (y)
2πM 2 πN0 i=0 m=0

in which the two functions βi,m and γi,m are given by :

βi,m = (C2 + Ai )2 ρ2 + C12 + A2m (3.69)

γi,m = |2ρ(C1 + Am )(C2 + Ai )|. (3.70)


86 Chapter 3. Space Division Multiplexing

The PDF of N will allow us to plot the conditional PDF PN (y|X1 = Am ) for each transmitted
symbol. Accordingly, using a semi-analytical method to compute the surface of the overlapped
area limited by the PDF PN (y|X1 = Am ) and PN (y|X1 = Am ) for two consecutive symbols.

3.5.3 Crosstalk Impact Using the DMT Format

In the presence of DMT signals, the derivation of the crosstalk PDF PZ is different from the
case where X1 and X2 are multi-levels signals. We assume that both X1 and X2 have the same
number of subcarriers denoted Nsc and the cyclic prefix of both signals is set to 0. In chapter
2, section 2.4, we showed that when the number of subcarriers is high enough, the generated
DMT signals have zero-mean Gaussian distributions, consequently, we have:

X1 (t) ∼ N (0, σ12 ) (3.71)

X2 (t) ∼ N (0, σ22 ) (3.72)

where σ12 and σ22 are respectively the variance of X1 and X2 . Normalizing the power of X1 and
X2 , we take σ12 = σ22 = 1. We consider θ to be constant during a DMT symbol and derive the
PDF of V (t) which is defined as:

  
V (t) = 2ρ cos(θ) 2ρC1 C2 + C1 + X1 C2 + X2
  (3.73)
= cos(θ) 2ρC1 X2 + 2ρC2 X1 + 2ρX1 X2 .

As shown in Fig. 3.16 (b), at the receiver side, the DMT signal is demodulated using FFT
operation, hence the FFT of V is denoted as Vbl and given by:

2N sc −1
2ρ X kl
Vbl = √ V (k)e−j2π 2Nsc
2Nsc k=0
(3.74)
2Nsc −1 
h 1 X  kl
i
= cos(θl ) µ1,l + √ C1 C2 + X1 (k)X2 (k) e−j2π 2Nsc
2Nsc k=0
3.5. Crosstalk Investigation for IMDD Schemes 87

where

µ1,l = 2ρ(C2 al + C1 bl ) (3.75)

and al and bl are respectively the allocated complex QAM symbols at the l−th subcarrier of
X1 and X2 . Hence, Vb can be expressed as:

Vbl = cos(θl )Yl (3.76)

in which Yl is:
2Nsc −1 
2ρ X  kl
Yl = µ1,l + √ C1 C2 + X1 (k)X2 (k) e−j2π 2Nsc . (3.77)
2Nsc k=0

According to the central limit theorem, Yl has a normal distribution with mean µ1 and variance
4ρ2 as follows:

Yl ∼ N (µ1,l , 4ρ2 ). (3.78)

Hence, we derive the PDF of Vbl . It is denoted as PVb and given by :

2µ2
1,m,k +v
2
M −1 X
M1 − Z +∞
X e 4ρ v2 vµ1,m,k
PVb (v) = √ e− cosh(2x) 4ρ cosh[cosh(x) ]dx (3.79)
m=0 k=0 4ρ 2π 3 0 2ρ

in which

µ1,m,k = 2ρ(C2 am + C1 bk ) (3.80)

and M is the QAM order format assigned to each subcarrier. Finally, the FFT of the total
crosstalk impact Z given in equation (3.51) is denoted Zb and can be expressed as :

Zbl = Vbl + Wl (3.81)

where according to the central limit theorem Wl is a normally distributed function as:

Wl ∼ N (µ2,l , 4ρ4 ), (3.82)


88 Chapter 3. Space Division Multiplexing

where the average µ2 is given by

µ2,l = 2ρ2 C2 bl . (3.83)

The term Wl being proportional to ρ2 could be neglected. In this case, the pdf of Zb is equal to
PVb given in equation (3.84). However, in the case where Wl is considered non-negligible, PZb is
given by:
2µ2
1,m,k +(z−µ2,k )
2
M −1 X
M1 − Z +∞ (z−µ2,k )2
X e 4ρ(ρ+1)
− cosh(2x)
PZb (z) = √ e 4ρ(ρ+1)

m=0 k=0 4ρ(ρ + 1) 2π 3 0


! (3.84)
(z − µ2,k )µ1,m,k
× cosh cosh(x) dx.
2ρ(ρ + 1)

Fig. 3.19 plots the real part of the normalized conditional PDF denoted as PZb (z|al ∈ Srk )
where Srk is the set of QAM symbols having equal real part. First, PZb (z|al ∈ Srk ) is plotted
when both al and bl are QPSK symbols at different crosstalk intensities (-50 dB,-25dB and -
10dB). We notice that in contrast with M-PAM signal, the impact of crosstalk is approximately
the same regardless the constellation level. Furthermore the crosstalk spread depends strongly
on the crosstalk level as its distribution is wider when its intensity increases. At a crosstalk
intensity of -10 dB, the PDF PZb (z|al ∈ Srk ) is plotted for al being a QPSK symbol and bl
being 8-QAM, 16-QAM or higher order 64-QAM. We observe that the crosstalk’s spread is
approximately the same for all cases and does not depend on the constellation of the crosstalk
signal.
The probability density function of the total noise N (t) including SSBI, crosstalk and AWGN
impacts is derived by assuming that the fast Fourier Transform (FFT) of the SSBI term X1 (t)2
has a normal distribution. According to the central limit theorem and since the power of X1 is
normalized, we have:
c2 ∼ N (0, 2).
X (3.85)
1

In the frequency domain, the PDF of N (t) is denoted PNb and given according to :

s
(y−z)2
Z
1 −
PNb (y) = e N0
PZb (z)dz (3.86)
4πN0 ρ(ρ + 1) R
3.5. Crosstalk Investigation for IMDD Schemes 89

𝑎𝑙 :QPSK 𝑎𝑙 :QPSK 𝑎𝑙 :QPSK


𝑏𝑙 :QPSK 𝑏𝑙 :QPSK 𝑏𝑙 :QPSK

𝑎𝑙 :QPSK 𝑎𝑙 :QPSK 𝑎𝑙 :QPSK


𝑏𝑙 :8QAM 𝑏𝑙 :16QAM 𝑏𝑙 : 64QAM

Figure 3.19: a) Example of the crosstalk impact on each symbol al belonging to a QPSK
constellation according to different crosstalk intensity and where bl is a QPSK symbol. b) The
crosstalk is at −10 dB with different constellation bl and theta is uniform from 0 to 2π

in which N0 is the constant AWGN noise density. In what follows, we’re going to plot the
conditional probability of N for each recieved QAM symbol am given by PN (y|am ∈ Sri ) where
Sri is the i-th set of transmitted QAM symbols having the same real part. For example, when
16-QAM is assigned to all subcarriers, the number and set is 4 and i = 1, 2, 3, 4. The numerical
computation of the area limited by the overlap PN (y|am ∈ Sri ) for consecutive symbols am
allows us to deduce the BER.

3.5.4 Crosstalk’ s impact comparison between M-PAM and DMT

formats

In this section, we compare the performance in terms of error probability in the presence
of crosstalk based on the theoretical studies in sections 3.5.2 and 3.5.3. We use numerical
simulations to validate the derived equations in both previous sections. An ideal optical channel
having an impulse response h(t) = 1 is assumed. First, two signals X1 and X2 are generated
90 Chapter 3. Space Division Multiplexing

independently. The DC components C1 and C2 are added respectively to X1 and X2 . Each


signal drives an optical intensity modulator where the emitted power is linearly proportional to
the driving current. The quantum efficiency of the electrical current-to-optical power conversion
of both modulators are set to 1 and the maximum output power from the modulators is Pmax
and is equal to 4 mW . In this model, the amplitude of X1 and X2 varies between 0 and 4 V

and C1 = C2 = 2 2. At the output of the intensity modulators, the optical field E2 is added
to E1 according to ρeiθ coefficient. The photodetector responsivity is equal to 1. We assume
only thermal noise. We will consider the 4-PAM signaling where the generation of 4-PAM is
assumed by mapping a 221 -length bits sequence into 4-PAM symbols. In the case of DMT , the
signal is generated with an FFT length of 512 subcarriers, and no cyclic prefix is added due
to the absence of dispersive channels. All subcarriers carry 16-QAM symbols for both signals.
No equalization is applied for 4-PAM and 16-QAM-DMT. The crosstalk phase θ takes value
uniformly distributed within [0,2π] and is assumed to be constant during a DMT symbol.

𝜌² = −25𝑑𝐵 𝜌² = −15𝑑𝐵 𝜌² = −10𝑑𝐵


normalized 𝑃𝑍

normalized 𝑃𝑍
normalized 𝑃𝑍
4-PAM

Crosstalk’s levels Crosstalk’s levels Crosstalk’s levels

𝜌² = −25𝑑𝐵 𝜌² = −15𝑑𝐵 𝜌² = −10𝑑𝐵


16QAM-DMT

normalized 𝑃𝑧Ƹ
normalized 𝑃𝑧Ƹ

normalized 𝑃𝑧Ƹ

inphase inphase inphase

Figure 3.20: Comparison of simulated and theoretical crosstalk distribution for different
crosstalk intensities for 4-PAM and 16-QAM-DMT

Fig. 3.20 displays the distribution of the crosstalk impact PZ for 4-PAM and PZb for 16-QAM-
DMT using simulation and the derived analytical expressions at −25 dB,−15 dB and −10 dB
of crosstalk. For both formats, we observe that the simulations match the theory which shows
the correctness of the analysis. Furthermore, we notice the strong dependency of the crosstalk
3.5. Crosstalk Investigation for IMDD Schemes 91

spread on the 4-PAM amplitude levels.

𝜌2 [dB]

Figure 3.21: BER as function of the crosstalk’s intensity ρ2 for 4-PAM and 16-QAM-DMT.

To compare the impact of crosstalk on each format, the BER metric is used. We normalized
the power of the noiseless received signal given by:

y = |E1 + ρeiθ E2 |2 (3.87)

and we added an AWGN and setting its variance to have a BER around 10−3 for both modu-
lation formats. Fig. 3.21 depicts the BER as a function of the crosstalk intensity when θ varies
uniformally between 0 and 2π, for both simulation and theory, using 4-PAM and 16-QAM-
DMT. When the impact of crosstalk is negligible (i.e., crosstalk intensity between −50 dB and
−40 dB), the BER for both 4-PAM and 16-QAM-DMT is approximately the same. When the
crosstalk increases, 16-QAM-DMT outperforms 4-PAM because the levels of 4-PAM are more
vulnerable to the crosstalk as explained earlier. Moreover, this result can be further seen in
Fig. 3.22 which shows the relative probability of total received 4-PAM signal and the real part
of 16-QAM-DMT after the DMT demodulation at −15 dB of crosstalk.
In summary, statistical characterization of inter-modal crosstalk in direct detection regime has
been approached theoretically for M-PAM and DMT modulation formats. Theoretical results
have been validated numerically with 4-PAM and 16-QAM-DMT which revealed the resilience
of DMT to crosstalk fluctuation compared to 4-PAM. This simplified study revealed a prelimi-
92 Chapter 3. Space Division Multiplexing

𝜌² = −15𝑑𝐵
4-PAM 16QAM-DMT
simulation simulation
theory

Normalized probability
theory
Normalized probability

levels inpase

Figure 3.22: Crosstalk impact on each level for 4-PAM and 16-QAM constellation for −15dB
of crosstalk in the presence of AWGN.

nary understanding of modulation formats robustness to the crosstalk impact. We expect that
DMT format could be the best candidate for mode group multiplexed transmission. In the
next chapter, DMT modulation will be adopted for experimental realizations. Furthermore it
will be experimentally compared to 4-PAM format to confirm this results in a concrete case.

3.5.5 Summary

This chapter discussed the basics of modes propagation in multimode optical fibers. First,
the theory of linearly polarized modes have been reviewed. Second, spatial mode multiplexing
has been introduced based on incident optical field decomposition in the basis of LPlm modes.
Mode coupling limitations have been also presented.
Example of mode multiplexing techniques have been presented such as, off-set launching tech-
nique, mode selectivity based on SLM and BPP, and finally the MPLC process which will be
used in this work. Furthermore, the mode group multiplexing approach is introduced to deal
with the limitation of strong mode coupling between the modes having close propagation con-
stants. In this work, this technique is implemented based on the MPLC technology.
Finally, statistical investigation of inter-modal crosstalk has been performed by comparing two
different modulation formats for IMDD applications. Theoretical derivation of crosstalk prob-
3.5. Crosstalk Investigation for IMDD Schemes 93

ability density function has been validated using simulation. 4-PAM and 16-QAM-DMT have
been compared in terms of robustness to localized crosstalk. 4-PAM showed high vulnerability
to crosstalk phase fluctuation compared to DMT format. To further confirm this result, both
modulation formats will be experimentally compared in the presence of inter-modal croststalk.
In the next chapter, mode group multiplexed transmission will be experimentally approached
for different target short reach applications. For each application, the design of the multiplexers,
demultiplexers, as well as the optical fiber will be detailed. This technique will be implemented
for both DWDM and low cost transmission using EML with EAM.
Chapter 4

Transmission Experiments

4.1 Introduction

In this chapter, we focus on high data rate transmissions over different types of MMFs for
short reach applications. In fact, increasing capacity over MMF is the subject of numerous
research due to the ever-growing demand of broadband services. Regardless of the application,
inter-modal dispersion is the main limiting factor for the reach at high data rates in MMF
as explained in chapter 3. However, MMF links are widely deployed and optimized for short
reach direct detection (DD) transmission, namely, local area networks (LANs), where the
distance is shorter than 5 km and intra-datacenter, where the distance is shorter than 2 km.
According to the ISO/IEC 11801 standard [68], several classes of optical multimode fibers have
been defined. They are classified as OM1, OM2, OM3, OM4 and OM5. Fig. 4.1 gives the
evolution of MMF standards over time and compares theses fibers in terms of maximum reach
at 10 Gb/s using directly modulated multimode 850 nm VCSELs and OOK format in addition
to their bandwidth widening over time. The OM1 fiber is the first standardized one, its core
diameter is 62.5 µm with a limited reach of 33 m supporting 10 Gb/s. It was deployed for
100 Mb/s Ethernet applications. By the deployment of 1 Gb/s Ethernet, the OM2 fiber has
been standardized with reduced core diameter compared to OM1 (from 62.5 µm to 50 µm) and
higher bandwidth from (200 MHz·km to 500 MHz·km). The maximum reach at 10 Gb/s has

94
4.1. Introduction 95

been increased to 83 m. The OM3 fiber has been adapted to 40 Gb/s and 100 Gb/s Ethernet
for 100 m distance. The use of 850 nm multimode VCSELs increased the reach to 300 m over
OM3 and drove the standardization of OM4 fiber with higher bandwidth compared to OM3.
Hence, With the more advanced OM4 fiber, the 10 Gb/s system reach can be upgraded to
550 m (802.3 -Edition 2015, Clause 52). Furthermore, the introduction and standardization of
a new wideband MMF (OM5) goes in the same direction of the current standard 200 GBE
(SR4.2) and 400 GBE (SR8.2) based on SWDM, thus, to keep a low-cost transmission, the
use of multiwavelength over OM5 fiber is feasible with multiple VCSELs [11].

OM1 OM2 OM3 OM4 OM5

550m

300m 28000
82m 4700
33m
2000
500
200
1989 1998 2002 year
2009 2014

Figure 4.1: Bandwidth improvements and maximum achievable distance evolution at 10 Gb/s
OOK using multimode VCSEL for standard optical multimode fibers.

The creation of parallel spatial channels within MMFs is another approach to increase the
throughput transmitted over short length with the DD scheme. On this topic, many technolo-
gies have been introduced as discussed in chapter 3.

In this chapter, we will explain how the data rates for transmission over MMFs can be ex-
perimentally increased based on the mode group multiplexing technique for different target
applications. Firstly, 5 Tb/s transmission over standard OM2 fiber is discussed for a maximum
distance of 2.2 km. Secondly, we will explain how the mode group multiplexing technique
could also be extended to applications where conventional OMx fibers are not appropriate,
for example, it could be used for LR datacenter InterConnect communications where longer
reach is required. Comparison between systems using OM2 fibers and those based on FMF
96 Chapter 4. Transmission Experiments

are compared in terms of MGM and MGD design, crosstalk and losses. In this work, specially
designed FMFs adapted for longer distances are provided by Prysmian. Mode group multi-
plexers and demultiplexers have been also adapted to FMF based transmission. For each link,
we worked with Cailabs to improve and propose new designs of mode group multiplexers and
demultiplexers to achieve higher transmission performance.

4.2 Mode Group Multiplexing for LAN and Intra

-Datacenter Applications

In this section, spatial mode group multiplexing is investigated over standard multimode fibers,
mainly OM2 fibers. In fact, this fiber being already standardized and installed, it is not possible
to modify its parameters such as refractive index profile and core diameter to improve the
performance. This work consists in improving the transported capacity over this type of fiber
by proposing an adapted architecture of transmission link. First, a demonstration of 5 Tb/s
bidirectional transmission (2.5 Tb/s in each direction) over 2.2 km of OM2 fiber using selective
excitation of 4 mode groups and wavelength multiplexing with DMT modulation format and
direct detection is detailed. Furthermore, compatibility of low cost component with mode group
multiplexed transmission will be proved. In a second part, a record transmission of 14.5 Tb/s
over 2.2 km of OM2 fiber will be presented.

4.2.1 5 Tb/s Bidirectional Transmission Over OM2 Fiber

In this part, we present the first demonstration of the mode group multiplexing technique
allowing to significantly increase the throughput over conventional OM2 fibers. The OM2 fiber
supports 55 modes operating at 1.55 µm which can be divided into 9 groups of modes. In
this experiment, we will use the first 4 groups presented in Fig. 4.2. The first group denoted
as G1 supports only the fundamental mode LP01 , G2 supports the degenerated modes LP11a
and LP11b , G3 with LP02 and the degenerated modes LP21a and LP21b . Finally the 4-th group
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 97

𝐺1
𝐺2
𝐺3
𝐺4

Figure 4.2: Intensity patterns of LP modes in OM2 fiber and mode groups patterns used in
this experiment.

G4 contains the higher number of modes which are LP31a , LP31b ,LP12a and LP12b . Through a
collaboration with Cailabs, specially designed MGM and MGD have been leveraged for this
experiment. The MGM generates the mode profile of only one mode within each group. It
generates the LP01 , for G1 , LP11a for G2 , LP21a for G3 and LP31a for G4 . As explained in
chapter 3, excitating a single mode is sufficient since all the modes within a group are highly
coupled. However, due to this large coupling, it is important to demultiplex and detect all the
modes within the same group at the receiver to avoid large power fluctuations. Hence, SMFs
are used at the inputs of the MGM while OM2 fibers are used at the output of the MGD,
except for G1 which carries only the fundamental mode, and can therefore be demultiplexed
to SMF. In fact, the MGD design is more complex compared to MGM design. To reduce
the system (MGM+MGD) complexity, both MGM and MGD will operate in bidirectional
configuration, as shown in Fig. 4.3. G1 and G4 will propagate in the same direction while
G2 and G3 will propagate in the opposite direction. Therefore, the phase plates of the MGM
and MGD will manage 5 modes in each direction. Table4.1 shows the measured losses and
the encountered crosstalk for each group due to the MGM and the MGD. The measurements
are done using an amplified spontaneous emission (ASE) source. To measure the losses, the
power is launched into the input of each group and measured at its output.The crosstalk of each
group is measured depending on its co-propagating group, for example, for G1 , the crosstalk is
given by the measured optical power at the output of the MGD corresponding to G4 , which
is the amount of transferred power from G1 to G4 . The average loss the MGM and MGD
98 Chapter 4. Transmission Experiments

SMF
input G1 SMF output G1
MGM OM2 MGD OM2
SMF
input G4 output G4

OM2
SMF
output G2 input G2
OM2 MGD MGM
output G3 SMF input G3

EDFA
Figureinput
4.3:G1 MGM and MGD used in bidirectional configuration.

WDM WDM
output G1
WDM WDM

20 l 20 l

SMF EDFA MGM OM2 MGD OM2


input G4 loss [dB] crosstalk[dB] output G4 20 l
20 l
G1 -5.6 -28 EDFA
WDM WDM

output G2 G2 -5.2 -14.5 input G2

WDM WDM
20 l 20 l
SMF G3
MGD -5.6splices -12.7
MGM EDFA SMF
output G3 input G3
20 l G4 -5.6 -17.2 20 l

Table 4.1: Generated loss and Crosstalk of the MGM followed by the MGD.

is −5.54 dB while the average crosstalk is −15.6 dB. As to this crosstalk value, we notice
that the interference between G2 and G3 is more important compared to that between G1 and
G4 . In fact, G1 contains only the LP01 mode which is the least complex mode. Moreover, its
propagation velocity is larger compared to other modes. The use of the MGM and MGD,
enables the creation of 4 spatial channels within the OM2 fiber, i.e., with 2 channels per
direction. To further increase the transported capacity over this single fiber, 20 wavelength
channels are multiplexed per mode group. Fig. 4.4 illustrates the experimental combination
of WDM and mode group multiplexing techniques that will be used in our experiment. We
multiplexed 20 wavelengths per group, which results in the creation of 80 channels within the
MMF. In fact, to avoid coherent crosstalk, each mode group uses a different set of wavelengths,
where the wavelength spacing is 200 GHz. The wavelength combinations of all mode groups
is represented in Fig. 4.5. Wavelengths are interleaved in such way that in each direction a
100 GHz spaced is propagating. At the MGM side, we use conventional single mode WDM
multiplexers. To conserve the modal composition per group, we use multimode WDM de-
multiplexers with OM2 fiber pigtail with FC-PC connectors at both ends. The fiber length at
the output of the WDM multiplexer is shorter than 0.5 m to avoid modal dispersion impact.
Moreover, to compensate for the loss of the WDM (de)multiplexer and of the MGM and
MGD, we use single mode single stage Erbium Doped Fiber Amplifier (EDFA) at the outputs
MGM OM2 MGD OM2
SMF
input G4 output G4

4.2. Mode Group Multiplexing


OM2
for LAN and Intra-Datacenter Applications 99
SMF
output G2 input G2
OM2 MGD MGM
output G3 SMF input G3
of the WDM multiplexers.

EDFA

WDM WDM
input G1 output G1

WDM WDM
20 l 20 l

SMF EDFA MGM OM2 MGD OM2


input G4 output G4
20 l
20 l
EDFA

WDM WDM
output G2 input G2

WDM WDM
20 l 20 l
SMF MGD splices MGM EDFA SMF
output G3 input G3
20 l 20 l

Figure 4.4: Illustration of WDM and mode group multiplexing technique used in experimental
studies.

G1: comb1 OM2


G2: comb2
G4: comb4 G3: comb3

200 GHz
100GHz
50GHz

G1 G2 G4 G3 G1 G2 G4 G3
l

Figure 4.5: Illustration of wavelength comb interleaving for all the groups.

First, single wavelength experiments are conducted for each group of modes. Independent Tx
and Rx are used for each group. Tx and Rx are depicted in Fig. 4.6(a) and 4.6(b) respectively.
At the transmitter side, the signal is generated by a 65 Gs/s CMOS Digital to Analog Converter
with an effective number of bit (ENOB) equal to 5. An MZM intensity modulator driven by a
32 GHz linear driver is used for the electro-optical conversion process. The 20-wavelengths comb
operates in the C band. Single mode EDFAs are added at the output of the WDM multiplexer.
For each group of mode, the signal is generated using independent Txs (i.e., different combs,
different MZMs,..) such that independent data is propagating over each group. At the Rx
side, the signal is demultiplexed using the multimode WDM demultiplexer and single channel
is detected per modegroup with an 28 GHz multimode PIN photodiode. In fact, as explained
in chapter 1, it is challenging to benefit from large bandwidth multimode photodiode because
of its large effective area. After the opto-electrical conversion, the signal is then amplified by an
RF driver and resampled with a 50 Gs/s oscilloscope with 16 GHz analog bandwidth. Off-line
100 Chapter 4. Transmission Experiments

processing is used to process data and compute the errors. 2.2 km of standard OM2 fiber is
spliced between the MGM and MGD. The loss of link including the MGM and MGD with 2.2
km of fiber and 2 splices are given in Table 4.2. The loss of the multimode WDM demultiplexer

MGM+MGD [dB] MGM+MGD+OM2[dB] MGM+MGD+OM2+ WDM demux [dB]


G1 -5.6 -6 -10
G2 -5.2 -5.7 -9.7
G3 -5.6 -6.4 -10.4
G4 -5.6 -6.9 -10.9

Table 4.2: Accumulated losses over the link where OM2 represents 2.2 km of OM2 fiber.

for the worst channel is -4dB, which gives a maximum loss of ∼ −11dB for the worst channel of
the worst mode group. Furthermore, Fig. 4.6 (c) compares the normalized transfer function of
both single mode WDM multiplexer and multimode WDM demultiplexer. We notice that the
bandwidth of the multimode WDM demultiplexer is wider than the single mode multiplexer.
Hence, the multimode WDM demultiplexer will not limit the signal bandwidth. The main
bandwidth limitation will be due to the 16 GHz bandwidth oscilloscope.

Tx Rx Single mode WDM mutliplexer


Multimode WDM demutliplexer
PIN RF 0
Transmission loss [dB]

DAC Amplifier
WDM

Driver BW=28GHz
65Gs/s
-5

multimode
-10
EDFA
WDM

20 l MZM Oscilloscope
Off-line
50Gs/s -15
processing 192.78 192.98
single mode Frequency [THz]

(a) (b) (c)

Figure 4.6: (a) transmitter (Tx) per mode group, (b) receiver (Rx) per mode group (c) transfer
functions of single mode WDM multiplexer and multimode WDM demultiplexer.

The DMT format is used for this experiment. The signal is generated with 1024 useful subcar-
riers, which means 2048-FFT length. A cyclic prefix of 4 samples is added after the FFT. The
clipping ratio has been optimized to 25%. Rate adaptive Chow’s algorithm [73] is applied to
perform bit and power loading. Off-line signal processing is used to synchronize, demodulate
the DMT signal and count the errors. The first step before applying the bit loading algorithm
is to assign the same QAM constellation for the 1024 subcarriers, then compute the SNR per
subcarrier using off-line processing at the Rx. In this work, we modulate all the subcarriers
with the QPSK format. Fig. 4.7 displays the gross line rate (without taking into account the
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 101

FEC overhead) obtained in single channel transmission for each mode group as a function of
Rx input power. For each input power and for each mode group, we apply optimum bit loading
given by Chow algorithm for two BER targets which are 1.2 × 10−2 corresponding to a hard
decision FEC with 20% overhead [53] and 3.8 × 10−3 corresponding to a hard decision FEC
with 7% overhead. G1 , G2 and G3 show approximately similar bit rates for both BER target

100
6
90
5.5
BER ≈1.2 x 10-2
Gross line rate [Gbits/s]

Total troughput [Tb/s]


80
5

70
BER ≈ 3.9x 10-3 4.5

60 4
G1
50 G2
3.5
G3
G4
40 3
-8 -6 -4 -2 0 2 4 16 18 20
Input Power [dBm] EDFAs Output p

Figure 4.7: Achievable gross line rate versus Rx input power obtained with Chow’s bit loading
algorithm at two different BER target :1.2 × 10−2 and 3.8 × 10−3

whereas the bit rate obtained with G4 is lower. This penalty could be originated from multipath
interference (MPI) generated by the splices in addition to inter-modal coupling with higher
order modes which are not used for transmission. For each group, the EDFA total output
power is set to 21 dBm which gives 8 dBm per channel, therefore the optical power at the Rx
for the worst channel is −3 dBm. We choose to apply the same bit loading for all mode groups.
In order to have some margin (to take into account gain excursion of the EDFAs for instance),
we considered that the worst channel input power on the Rx would be −4 dBm. For all mode
groups, we apply the computed bit loading of G4 for an input power of −4 dBm which results
a bitrate of 80.86 Gb/s at BER of 1.2 × 10−2 before FEC and 68.8 Gb/s at BER of 3.9 × 10−3
before FEC. Fig. 4.8 (a) displays the measured SNR versus frequency for G4 for an input
power of −4 dBm. We observe that after 16 GHz the SNR decreases abruptly which is due to
the oscilloscope bandwidth limitation. Furthermore, we remark the spurious behavior of the
102 Chapter 4. Transmission Experiments

DAC/ADC at certain frequencies. Bit allocation will adapt the signal to this SNR response.
Fig. 4.8(b) displays the bit per symbol versus frequency for a BER target of 3.9 × 10−3 . We
observe that the subcarriers at 8.15 GHz and 16.2 GHz are turned off due to the spurs location.
We now move to WDM experiments. Fig. 4.9 shows the BER measurements of the 80 channels.

6
20 G1 G4 G4
5

15 4

bits/symbol
SNR [dB]

3
10
2
5
1

0 0
0 5 10 15 20 25 0 5 10 15 20 25
Frequency [GHz] Frequency [GHz]

(a) (b)

Figure 4.8: (a) single channel SNR response after 2.2 km propagation over G4 , (b) its corre-
sponding bit allocation for BER target of 3.9 × 10−3 using rate adaptive Chow4S algorithm

In Fig. 4.9(a) for BER target of 1.2 × 10−2 and in (b) for BER target of 3.9 × 10−3 . In both
configurations all the channels are above the FEC limit. Each BER value is averaged over 20
measured waveforms with 219 samples each. The total transmitted capacity before FEC is 6.47
Tb/s in (a) and 5.5 Tb/s in (b). Assuming a 20% FEC overhead, a total net throughput of 5
Tb/s is achieved in (a) and assuming 7% FEC overhead, a total net throughput of 5.1 Tb/s is
achieved in (b). Fig. 4.10 shows the estimated (using measurement of Fig. 4.7) total capacity

1.E-01 1.E-01

FEC threshold
1.E-02 1.E-02
log10(BER)

log10(BER)

FEC threshold

1.E-03 1.E-03

G1 G2 G1 G2
G3 G4 G3 G4
1.E-04 1.E-04
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Channel index Channel index
(a) (b)

Figure 4.9: BER measurement of 80 channels modulated at (a) 80.96Gb/s (b) 68.8Gb/s
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 103

as a function of the EDFAs output power when all the groups have the same bit loading as G4
(blue dot) and when we use optimum bit loading for each mode group (red square). Applying
optimal bit loading for each mode group enable 5% increased capacity. Increasing the output
power of the EDFA by 3 dB (from 21 dBm to 24 dBm) increases the capacity by only 8%,
whereas decreasing the EDFA output power by 3 dB (from 21 dBm to 18 dBm) decreases the
capacity by 35%. We found that 21 dBm output power is the best compromise in our set-up
between transported capacity and EDFA cost and power consumption.

5.5
BER ≈1.2 x 10-2
Total troughput [Tb/s]

BER ≈ 3.9x 10-3 4.5


same bit loading (G4)
4 for all mode groups
G1 each mode group with
G2 optimized bit loading
3.5
G3
G4
3
-6 -4 -2 0 2 4 16 18 20 22 24 26
Input Power [dBm] EDFAs Output power [dBm]

Figure 4.10: Total capacity over EDFA’s ouput power.

4.2.2 Beyond 250Gb/s Single Wavelength Transmission Based on

Low Cost EML

EMLs are a widely deployed technology for intra-data centres communications due to their
advantages over other technologies such as DMLs. More specifically, they have a reduced
frequency chirp and can achieve a higher extinction ratio. Furthermore, they present an eco-
nomical advantage due to their low cost, especially for coarse wavelength division multiplexing
(CWDM) with no need to (temperature-) stabilize or optical filters.
Fig. 4.11 provides an example of current standards of single mode 100G LR4 and 400G LR8
transceivers. In the case of 100G LR4, four EML modulated at 25 G NRZ using CWDM
are utilized. As for 400G LR8, eight wavelength lanes are used with 50Gb/s 4-PAM. Scaling
beyond 800 Gb/s is very challenging with current path because it requires a high modulation
104 Chapter 4. Transmission Experiments

CFP 8

DML/EML

WDM
DML/EML

DML/EML

Band WDM
DML/EML
SMF
QSFP TOSA

DML/EML DML/EML
WDM

WDM
DML/EML SMF DML/EML
DML/EML DML/EML
DML/EML DML/EML
TOSA
TOSA

100GBASE-LR4 : 4x lx25Gb/s (OOK) 400GBASE-LR4 : 4x lx100Gb/s (PAM4)


400GBASE-LR8 : 8x lx50Gb/s (PAM4)

Figure 4.11: EML-based transceivers standards for intra-datacenter communications.

baudrate which needs a wide electro-optical bandwidth. Furthermore, increasing the number
of wavelengths is not adapted to the limited wavenlength stability of un-cooled lasers sources
and remains very expensive for short reach applications. In fact, to keep low cost direct detec-
tion, a combination of CWDM and mode group division multiplexing will allow for maximum
scaling while keeping moderate channel spacing and reducing laser constraint. This method
is illustrated in Fig. 4.12 with an example of selective excitation of 2 mode groups. In this

QSFP QSFP

EML Rx

MMF
WDM

SMF
WDM

EML Rx

EML Rx

EML MMF Rx
MGM

MGD

(one direction)
QSFP
QSFP
Rx
EML
MMF
WDM

SMF
WDM

Rx
EML
Rx
EML
EML Rx

Figure 4.12: Proposed architecture based on mode group multiplexing and CWDM using EML

section, we will present a single wavelength transmission based on EML and mode group mul-
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 105

tiplexing technique. For this purpose, we compare channel bit rate as a function of OM2 fiber
length for two different transmitter configurations. The first configuration is the same as the
one used in section 4.2.1 using a DFB laser and an MZM. In a second configuration, we use
an EML with 18 GHz 3-dB bandwidth [124]. The EML has been deployed and optimized in
collaboration with III-V Labs. We used it in a test−bench compatible with chips on carrier
enabling short-time loop validation of EML designs without expensive and time-consuming
packaging, as illustrated in Fig. 4.13. This EML is specifically designed for very high output

Figure 4.13: photo of EML test−bench as used in the Lab

power and semi-cooled operation and operates with +6 dBm modulated power at 45◦ C. The
EML results from the monolithic integration on an InP substrate of a 500 µm-long DFB laser
with a 150µm-long electro-absorption modulator. The laser has a continuous grating and high
back-facet reflectivity. The EAM is separated from the DFB section by an H+-implanted sec-
tion for electrical isolation. The EAM-DFB integration is done by butt−joint technology. Both
EAM and DFB layer structures are grown by Molecular Beam Epitaxy (MBE) and are based
on InGaAsP/InGaAsP multi-quantum wells structures. The waveguide structure is a semi-
insulated buried heterostructure (SIBH), with semi- insulating-InP (SI-InP) grown along the
waveguide and p-doped InP on top. The EML is operated at 35◦ C and its emission wavelength
is 1537.4 mn (195 THz). Fig. 4.14 depicts the microscopic top view of the EML in (a) and
the SIBH technology in (b). The laser section is operated at 70 mA and the EAM section bias
106 Chapter 4. Transmission Experiments

MQW p-InP

250 µm
Laser Modulator SI-InP

n-InP
900 µm

(a) (b)

Figure 4.14: (a) microscopic top view of the EML (b) the SIBH waveguid with semi- insulating-
InP

is −1.1 V. Fig. 4.15 (a) shows the normalized output power as a function of the bias voltage
when the current of the laser section is 70 mA. Fig. 4.15 (b) shows the electro-optic response of
the EML from which electro-optical 3dB bandwidth of typically 18 GHz can be infered. The
peak-to-peak voltage of the signal applied to the EAM is 1.4 V. The modulated output power is
−1 dBm. In this section, we use the same Rx and Tx introduced previously in Fig. 4.6 except

0 3

-4
Normalized power [dB]

0
S12 [dB]

-8
-3
-12

-6
-16

-20 -9
-2 -1.5 -1 -0.5 0 0 10 20 30
EAM Bias [V] Frequency [GHz]
(a) (b)

Figure 4.15: (a) Static output power as a function of bias voltage, (b) electro-optic bandwidth
for Ilaser=70mA

that we use a 88 Gs/s CMOS DAC instead of 65 Gs/s DAC for practical reasons. However,
the same 50 Gs/s oscilloscope with 16 GHz analog bandwidth is still used. The impact of
the DACs on the performance is compared in single mode single channel experiment using an
MZM modulator. Fig. 4.16 illustrates the impact of 88 Gs/s and 65 Gs/s DAC on the SNR
response at 0 dBm Rx input power in (a) and the achievable bit rate before FEC for BER
target of 3.8 × 10−3 in (b). We observe that the 88Gs/s DAC enables higher bitrates, up to 10
Gbit/s more capacity compared to the 65 Gs/s DAC at high Rx input power (2 dBm). This
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 107

is due to the difference in terms of ENOB which is 6 for the 88 Gs/s DAC compared to 5 for
the 65Gs/s DAC in addition to the wider bandwidth of the 88 Gs/s DAC. However, for low
Rx input powers, both DACs show the same performance. In this case the main limitation
comes from the thermal noise and not the quantization noise. Moreover, we observe from Fig.
4.16(a) that we are still limited by the 16 GHz oscilloscope bandwidth.

25 110

100
20

90

Rate[Gbit/s]
15
SNR[dB]

80
10
70
B2B with 65Gs/s DAC
5 60
B2B with 88Gs/s DAC B2B with 88Gs/s DAC
B2B with 65Gs/s DAC
0 50
0 5 10 15 20 -6 -4 -2 0 2
Frequency[GHz] Rx input power[dBm]

(a) (b)

Figure 4.16: Impact of 88Gs/s and 65Gs/s DAC in single mode B2B transmission with MZM
on (a) the SNR response at 0dBm , (b) achievable bit rate before FEC for BER target of
3.9 × 10−3

The EAM and MZM performance will be compared with the 88 Gs/s DAC. Hence Fig.4.17(a)
displays the bit rate (excluding 7% FEC overhead) as a function of the receiver input power for
two different configurations using the MZM and the EML for the reference back to back over
less than 10 m of single mode fiber used between the transmitter and the receiver including
the EDFA. We measure a reduction of approximately 10 Gb/s for all input power when using
the EML instead of the MZM. Fig. 4.17(b) shows the SNR for an input power of 0 dBm
when using the EML or the MZM. The bit rate reduction could be attributed to the EML
bandwidth which reduces the SNR of some subcarriers compared to the MZM, as can be seen
in 4.17(b) for an input power below -1 dBm. To reach the same bit rate with the EML the
power of the modulated channel needs to be increased by 1dB compared to the MZM.

We have then compared MZM and EML in a single channel transmission experiment over
2.2 km and 4.4 km of OM2 fiber. As for the previous experimental setup represented in Fig.
4.6, the channel power at the output of the EDFA is set to 8 dBm. Fig. 4.18 displays the
108 Chapter 4. Transmission Experiments

110 25

100
Gross line rate[Gb/s] 20
90

80 15

SNR[dB]
70 10
60 EML
EML
5
50
MZM MZM
40 0
-6 -4 -2 0 2 0 5 10 15 20
Rx input power [dBm] Frequency[GHz]
(a) (b)

Figure 4.17: Single mode back-to-back:(a) bit rate comparison between MZM and EML for a
BER target of 3.9 × 10−3 (b) SNR for MZM and EML for a Rx input power of 0dBm

bit rate for the 4 mode groups for 2.2 km and 4.4 km for both MZM and EML. For each
measurement, the bit loading with 1024 subcarriers has been optimized to reach the highest
bit rate for a BER target of 3.8 × 10−3 . For 2.2 km, using the MZM we can reach more than
79 Gbit/s per mode group, which permits an allowable total transmission throughput of 5.87
Tb/s after applying FEC comparing to 5 Tb/s obtained in section 4.2.1. This difference is due
to the use of a 88 Gs/s DAC instead of a 65 Gs/s DAC. For the same distance, using the
EML, more than 68.5 Gbit/s is achieved per mode group. This demonstrate that 5Tb/s total
throughput would be reachable with the use of low cost EML instead of a DFB laser + lithium
niobate MZM. When the OM2 fiber length is increased from 2.2 km to 4.4 km, the channel
bit rate is decreased firstly due the increased loss of the link and secondly due to the increase
of accumulated chromatic dispersion. When using a MZM the channel bit rate of the worst
mode group is decreased from 79 Gb/s to 65.5 Gb/s: this is the same bit rate obtained with
the 65 Gs/s DAC for 2.2 km. Using the 88 Gs/s DAC instead of the 65 Gs/s DAC enables us
to double the reach while keeping the same throughput. Finally, when using an EML for 4.4
km reach, the channel bit rate for mode G4 is 57 Gb/s, therefore the bit rate is reduced by only
13% due to use of an EML instead of a MZM when all other experimental parameters are the
same. In this section, we demonstrated a total throughput of more than 5 Tbs/s bidirectional
transmission (2.5 Tb/s in each direction) over 2.2 km of conventional OM2 fiber using DMT
modulation and direct detection using a MZM. We also demonstrated that the use of low cost
EMLs is compatible with the mode group multiplexing technique in the case of a single channel
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 109

100
MZM 2.2km EML 2.2km

MZM 4.4km EML 4.4km


90

Gross line rate[Gb/s]


80

70

60

50
1 2 3 4
Mode Group index

Figure 4.18: Gross line rate for the 4 mode groups for 2.2km and 4.4km for the MZM and for
the EML

experiment.

4.2.3 14.5 Tb/s Transmission over OM2 Fiber

In this section, demonstration of 14.5Tb/s bidirectional transmission over 2.2 km of OM2


fiber using selective excitation of 4 mode groups based on MPLC technology, WDM and direct
detection is detailed. The same MGM and MGD as in section 4.2.1 are used in this experiment.
However the key elements enabling to increase the transported bitrate are the opto-electronic
devices. The MGM and MGD did not degrade over time and we verified that the crosstalk value
given in Table 4.1 are approximately the same as in Table 4.1. Table 4.3 shows the estimated
modal crosstalk of the OM2 fiber. To estimate the fiber crosstalk, first the total crosstalk of
the MGM, MGD and 20 m of fiber with 2 splices is measured. Second, we replace the 20 m
of fiber by 2.2 km of fiber. Finally, we subtract the results from the second measurement from
the first one, to extract the modal crosstalk of the fiber alone. The modal crosstalk generated
by 20m of fiber is negligible. As the modal crosstalk induced by fiber splice could be different
from one splice to another, we made 3 different splices for each measurement and averaged the
value. Moreover, due to the use of bidirectional transmission, in addition to the uses of different
wavelengths per mpde group, modal crosstalk is generated by the copropagating mode group
only. For instance, the crosstalk for mode G4 is created only by G1 and the crosstalk for G1
110 Chapter 4. Transmission Experiments

is created only by the co-propagated G4 . There is no crosstalk due to the propagated mode
groups in the opposite direction and the return loss is lower than −28 dB for all mode groups.

crosstalk of MGM+MGD [dB] crosstalk of OM2 [dB/km]


G1 -28 -41
G2 -14.5 -21
G3 -12.7 -20
G4 -17.2 -26

Table 4.3: Estimated fiber crosstalk.

B2B single mode single channel experiment

Fig. 4.19 shows the experimental setup in B2B configuration where Tx and Rx are linked
by 10 m of SMF. At the transmitter, a tunable laser source is modulated using an MZM
which is driven by 32 GHz linear driver. The modulated signal is generated using 88 Gs/s
DAC with 22 GHz analog bandwidth used in the previous section. At the receiver side, two
multimode photodetectors are tested. At switch 1 output, the 28 GHz PIN used previously and
at switch 2 output a new 30 GHz PIN and transimpedance amplifier (TIA) are used to enable
better sensitivity. Both photodetector systems are connected to two independent inputs of an
oscilloscope with 80 GHz sampling rate and 33 GHz analog bandwidth. The digitized signal
is off-line processed. To compare the gain in terms of achievable bit rate and power budget
using PIN+TIA instead of PIN, we generate a DMT signal with an IFFT/FFT length of 1024,
a cyclic prefix length of 2 samples, and a subcarrier separation frequency of 52.7 MHz. The
typical PAPR value before clipping is 7.2 dB. A 25% maximum amplitude clipping is applied
to the generated signal. Moreover, bit and power allocation per subcarrier are applied in order
to compensate for the bandwidth limitation and channel impairment without a complex signal
processing. The bit loading procedure is described in section 2.4

Fig. 4.20 (a) compares the achievable gross line rate for a BER target of 3.9 × 10−3 versus Rx
input power when signal is detected with either the 28 GHz PIN or the 30 GHz PIN+TIA.
We notice that the receiver based on PIN+TIA enables to transmit approximately 25 Gbit/s
more compared to the one with PIN. In fact this difference comes mainly from the benefit of
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 111

DAC, 88Gs/s Driver


BW = 22GHz BW=32GHz
1 PIN RF
BW=28GHz Amplifier

Laser MZM 2 PIN+TIA


source
BW=30GHz

Off-line Oscilloscope, 80Gs/s


processing BW=33GHz

Figure 4.19: Experimental setup for comparison of 30GHZ PIN+TIA and 28GHz PIN in single
channel single mode B2B configuration

150 25

140
20
Gross mine rate[Gbit/s]

130
SNR [dB]

120 15

110 10
100
28GHz PIN 5 28GHZ PIN
90
30GHz PIN+TIA 30GHz PIN+TIA
80 0
-6 -5 -4 -3 -2 -1 0 8 16 24 32
Rx input power [dBm] Frequency [GHz]
(a) (b)

Figure 4.20: (a) comparison of achievable bit rates before FEC between different photoreceivers
for BER target of 3.9 × 10−3 (b) SNR versus frequency at Rx input power of -3dBm

the TIA which gives 3dB more power budget. while the difference in terms of bandwidth has
a minimal impact as it can be seen in Fig. 4.20 (b) where the SNR versus frequency is plotted
for both receivers. We notice that both SNR responses decrease abruptly from 22 GHz. The
bandwidth limitation is mainly originating from the 22 GHz DAC. Furthermore, we see in Fig.
4.20 (a) that at -1dBm of Rx input power, the achievable bit rate decreases from 141 Gbit/s
to 132 Gbit/s for PIN+TIA detector while it saturates using the PIN detector. This is due to
the non-linearity effects, which are more important in the case of PIN+TIA.
112 Chapter 4. Transmission Experiments

B2B multimode mode single channel experiment

In this section the MGM and MGD are inserted instead of 10 m of SMF as shown in Fig. 4.21.
Each mode group carries independent data using 4 independent Tx and Rx. In practice we
used only two Tx and Rx for the copropagating groups of modes (because the return loss of
-28dB has non impact and could be neglected). The Tx and Rx have the same design as in
Fig.4.19.

G1 G1
TX1 SMF
RX1
SMF
MGM OM2 MGD
G4 G4
TX4 RX4
SMF OM2

G2 G2
RX2 SMF
TX2
OM2
MGD MGM
G3 G3
RX3 OM2 TX3
SMF

Figure 4.21: MGM+MGD configuration as used in the experiment, thin lines represents SMF
while thick lines represents OM2 fiber

For each group we compared the achievable bit rate using PIN and PIN+TIA receiver at dif-
ferent Rx input power and plotted the gain in bit/s as a function of mode group index in
Fig. 4.22. The gain is computed by subtracting the achievable bit rate for a BER target of
3.8 × 10−3 when the signal is detected using the 28 GHz PIN from the one when the signal is
detected using 30GHz PIN+TIA. We notice that at high Rx input power (-3 dBm) the gain is
higher than 25 Gbit/s for all the groups. However when decreasing the Rx input power, the
gain decreases as a function of mode group index. G4 decreases much faster while the gain of
G1 represents the lowest penalty. In this case, we notice that the gain depends on the modal
composition of each group. In fact the PIN+TIA bandwidth is 2 GHz larger, which means that
its effective area is smaller compared to 28 GHz PIN. The effective area reduction impacts the
detection of modes with large spot size. Therefore, the MGD has been designed to convert the
higher order modes profiles to low order modes profiles to adapt the signal to the small effective
area of the photoreceiver. As indicated in Fig. 4.23 for G3 and G4 , all the modes of each mode
group are converted to lower order modes and sent to an output OM2 fiber. The modes of
G3 (LP02 , LP21a , LP21b ) are converted to LP01 , LP11a , LP11b modes, and the modes of mode
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 113

group 4 (LP31a , LP31b , LP12a , LP12b ) are converted to LP01 , LP11a , LP11b and LP02 modes.
However the gain penalty of G4 could be explained by the imperfect detection of this group by
the photodetector since it contains the LP02 mode which could be non adapted to the effective
area of the 30GHz PIN+TIA. Concerning the penalty due to the insertion of the MGM and
MGD compared to single mode back-to-back, we verified that MGM and MGD insertion has
a negligible impact by compared the achievable bit rate for G1 and single mode B2B obtained
in the previous section. The results are displayed in Fig.4.24.

35

30

25
Gain[Gbit/s]

20

15

10

5 -3dBm -4dBm -6dBm

0
1 1.5 2 2.5 3 3.5 4
mode group index

Figure 4.22: The achievable throughput comparing 28GHz PIN and 30GHz PIN+TIA for the
4 groups of modes

Propagating Modes detected by


modes the photodiode

G1:
LP01 LP01

G2:
LP11a LP1b LP11a + LP11b

G3:
LP02+LP21a+ LP21b LP01+ LP11b+ LP11a

G4:
LP12a +LP12b+LP31a+ LP31b LP01+ LP02 + LP11a + LP11b

Figure 4.23: Converted modes received by the photodiode after mode group demultiplexing.
114 Chapter 4. Transmission Experiments

150
B2B single mode using PIN
B2B single mode using PIN+TIA
140
B2B G1 using PIN
B2B G1 mode using PIN+TIA
130

Gross mine rate[Gbit/s]


120

110

100

90

80
-6 -5 -4 -3
Rx input power [dBm]

Figure 4.24: Impact of MGM+MGD insertion compared to single mode B2B.

Multimode WDM Transmission Experiment

In this section, a WDM experiment is achieved based on the same link as the previous section.
We will use the 30GHz PIN+TIA photodetector. Schematics of the transmitter and the receiver
are displayed in Fig. 4.25(a) and Fig. 4.25(b) respectively. To compensate for the insertion

TXi, i=1,2,3,4 RXi, i=1,2,3,4


OM2
MUX
WDM

20l, 200GHz PIN+TIA


40l, 100GHz

50/50 EDFA OM2 BW=30GHz


WDM
Dmux

MZM 40l
WDM

SMF
MUX

20l, 200GHz
Driver
BW=32GHz Oscilloscope, 80Gs/s
Off-line
processing BW=33GHz
Odd 40λ-comb for i=1,3 DAC, 88Gs/s
Even 40λ-comb for i=2,4 BW = 22GHz

(a) (b)

Figure 4.25: (a) transmitter including single mode WDM multiplexer (b) receiver including
multimode WDM demultiplexer

loss of the MGM and MGD as well as that of WDM Mux/Demux, we use single-mode single
stage EDFAs at the output of the WDM multiplexers at the transmitter side. The total output
power of each EDFA is 22.5 dBm. A standard OM2 fiber is spliced between the MGM and
the MGD. We used DMT modulation with an IFFT/FFT length of 1024, a cyclic prefix length
of 2 samples, and a subcarrier separation frequency of 52.7 MHz. A 25% maximum amplitude
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 115

clipping is applied to the generated signal. The 40 wavelength channels of each mode group
are modulated by a separate modulator. Off-line signal processing is used to demodulate the
DMT signal and count the errors. The detected signal is normalized and resampled. Time
synchronization is required to determine the beginning of DMT frames. For synchronization,
a vector of 256 samples with repeated part is inserted at the beginning of each transmitted
frame with 217 samples. The synchronization vector is detected through an auto-correlation
operation over a sliding window [72] as explained earlier in chapter 2. After synchronization,
cyclic prefixes are removed, and an FFT operation enables to obtain the data symbols in each
subcarrier. We are considering a hard decision FEC limit corresponding to a BER of 5 × 10−3
and assume a 7% overhead. The achieved net bit rate assumes the transmission overhead due to
the cyclic prefix, time synchronization and the FEC. At the transmitter side, conventional single
mode WDM multiplexers are used. At the receiver side, a multimode WDM de-multiplexer is
used with multimode OM2 fiber pigtails having FC-PC connectors at both ends. The measured
modal crosstalk for both MGD, MGM and OM2 fiber displayed in Table 4.3 are high and the
use of the same wavelength comb for various co-propagating modes groups would generate
high BER penalties. Therefore, to avoid coherent crosstalk penalty, each co-propagating mode
group uses a different wavelength comb with a wavelength spacing of 100 GHz.

A:1km 50/50
Even l24
comb 50/50
l26
MZM
l25
PSD [5dB/Div]

Configuration.1

Odd B:2.2km
comb l24 50/50
MZM 1 50/50
l26

l25 MZM 2

192900 192950 193000 Configuration.2


Frequency [GHz]
(a) (b)

Figure 4.26: (a) Spectrum of co-propagating G1 and G4 in one direction where A corresponds
for 1 km experiment and B for 2.2 km at the output of MZM modulators (b) 2 different
configurations used to measure the penalty due to wavelength crosstalk in the WDM multimode
demultiplexer induced by adjacent channels
116 Chapter 4. Transmission Experiments

In our experiment, we used an array of 80 DFB lasers with 50 GHz spacing, partitioned into
two interleaved combs of “even” an “odd” sources, each consisting of 40 laser sources of 100
GHz spacing. The odd (even) comb was assigned to the mode groups 1 and 3 (2 and 4), leading
to a total number of 160 transmitted channels. The wavelength combination for G1 and G4
are indicated in Fig. 4.26(a) where modulated signals for group 1 and 4 are displayed for two
distances 1km and 2.2 km. Wavelengths are interleaved in such a way that in each direction the
same 50GHz-spaced comb is propagating. The same odd and even wavelength combs are used
for G3 and G2 in the opposite direction. Even if the measured modal crosstalk value does not
allow for the use of the same wavelengths for co-propagating mode groups, it is nevertheless
possible to have a partial spectral overlap as it can be seen in Fig. 4.26(a). In case of 1 km
transmission, the signal bandwidth is less impacted by chromatic dispersion compared to 2.2
km transmission. This overlap would not be possible in case of single mode fiber transmission
as it would lead to high inter symbol interference (ISI) penalties due to the impossibility to
filter out adjacent channels at the RX. Furthermore, we can clearly observe the impact of power
loading on the shape of the power spectral density for both groups. For each mode group, the
40 wavelengths with 100 GHz spacing are modulated with the same MZM modulator. Consid-
ering the channel power and the transmission distance in our experiments, there are no cross
nonlinear interactions during propagation among WDM channels within a group of modes.
Channel propagation is linear and the correlation of data between those channels has no im-
pact on the measured channels BER. However, the filtering function of the multimode WDM
multiplexer (displayed in the inset of Fig. 4.27 (a)) generates a small ISI penalty because it
is not wide enough considering the spectrum width of the modulated signal. The penalty due
to crosstalk induced by adjacent channels is 0.45 dB in terms of Q2 factor in our experimen-
tal configuration. This case corresponds to the configuration 1 displayed in Fig.4.26(b). To
compare our experimental configuration with a more realistic system configuration where each
channel would be modulated with independent data, we also measure the penalty caused by
adjacent channels in the WDM demultiplexer in the configuration 2 displayed in Fig.4.26(b)
where adjacent channels are modulated with different data than the measured channel. In this
configuration 2, it has been measured the same 0.45 dB penalty than in the configuration 1
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 117

which validate that our experimental set-up could be representative of an actual system.

20m of SMF
24

A: TX RX
20
20m of OM2
A In out
𝐺1 𝐺1
16 TX MGM RX
B: MGD
B
SNR [dB]

12 0 20m of OM2 out


In
𝐺1 𝐺1
-3
C TX RX
C: WDM
PSD (5dB/Div)

MGM MGD
8 -6
-9
MM WDM Demux D 2.2km of OM2 out
4 -12 In
SM WDM Mux
𝐺1 𝐺1
-15 Frequency (200 GHz/Div) TX WDM RX
D: MGM MGD
192,84 192,88 192,92 192,96
0
0 4 8 12 16 20 24 28 32
Frequency [GHz]
(a) (b)

Figure 4.27: (a) SNR of the optical channel in several configurations (inset filtering function of
the single mode WDM multiplexer and the multimode WDM demultiplexer, (b) Experimental
configuration used for measurements in (a).

In order to assess the impact of components on the performance, we measured the SNR re-
sponse reduction when inserting different devices of the transmission link described in Fig.4.25.
Fig.4.27(a) displays the SNR versus frequency for a single channel for G1 for an input power
of -3dBm and for several configurations A, B, C and D. Each configuration is represented in
Fig. 4.27(b). Case A is the reference bacl-to-back in single mode. The SNR displayed for
Case B when only the MGM and MGD are inserted showing they have very small impact
on the channel. The SNR displayed for Case C after the insertion of the multimode WDM
demultiplexer. Case C shows the impact of the filtering function of the multimode WDM de-
multiplexer. The filtering function of the multimode WDM demultiplexer is not as wide as a
single mode demultiplexer and therefore it reduces the channel bandwidth. The inset of Fig.
4.27(a) displays the bandwidth of a typical single mode (continuous line) and the multimode
(dashed line) WDM (de)multiplexer. It has been verified that the flat-top and wide filtering
of the typical single mode WDM mux would have no impact on the channel bandwidth in
this system. Finally, Case D displays the SNR of the end-to-end transmission with 2.2 km
of OM2 fiber. This curve shows further reduction of the channel bandwidth induced by the
chromatic dispersion. After 2.2 km propagation, the SNR response is different for each group
of modes as it can be seen in Fig. 4.28. In fact, the chromatic dispersion impact is different
118 Chapter 4. Transmission Experiments

24

20

16

SNR [dB]
12

𝐺1 Group 1
Mode
8
𝐺2 Group 2
Mode

4 𝐺3 Group 3
Mode
𝐺4 Group 4
Mode
0
0 4 8 12 16 20 24 28
Frequency [GHz]

Figure 4.28: SNR response after 2.2km propagation for all mode groups at -3dB of input power.

per group. Hence the bandwidth reduction is also different. Therefore, to optimize the spectral
efficiency, We applied different bit loading for each of the mode groups. For all mode groups
the highest number of bits per subcarrier was 6 (64QAM) in the low frequency region, and the
lowest number of bit was 1 (BPSK) in high frequency region. As displayed in Table 4.4. We
computed a net bit rate of 100.8 Gb/s for the 40 channels of G1 and 89.1 Gb/s per channel for
mode Group G2 and G3 and 86.2 Gb/s for G4 . The bit allocation for groups G1 and G4 are
displayed in Fig. 4.29.

Modulation format Number of subcarriers Number of subcarriers Number of subcarriers


G1 G2 &G3 G4
64QAM 69 44 32
32QAM 162 177 183
16-QAM 112 91 93
8QAM 91 60 58
QPSK 55 63 42
BPSK 5 11 25
Gross bitrate [Gb/s] 108.3 96.2 92.6
Net bitrate [Gb/s] 100.8 89.1 86.2

Table 4.4: Achieved net bit rate per channel for each mode group

Fig. 4.30 gives the achievable bit rate before assuming the 7% FEC overhead as a function of
reach for each group. Up to 4.4 km, G1 has the best performance, G2 and G3 allow to transmit
approximately the same bit rate while G4 has the lowest bit rate. Even at 1 km where CD is
not an issue, G4 has the worst performance. So the assumption that this group is the most
4.2. Mode Group Multiplexing for LAN and Intra-Datacenter Applications 119

6
G1
𝐺1
5 𝐺4
G4

bit/symbol
3

0
0 4 8 12 16 20 24 28
Frequency [GHz]

Figure 4.29: Allocated bit per subcarrier for G1 and G4 corresponding to 2.2km of propagation
for a BER target of 5 × 10−3 .

120
𝐺1
G1 𝐺3
G2

100 𝐺2 𝐺4
Gross bit rate [Gb/s]

G3 G4

80

60

40
1 2.2 4.4 6.6 8.8
Reach [km]

Figure 4.30: Achieved bit rate for BER target of 5 × 10−3 versus distance of propagation for
single channel for all mode groups .

impacted by CD is not valid for distance shorter than 4.4 km. In this case, the penalty could
originate from MPI generated by splices. For distances beyond 4.4 km the performance of G2
diverges from that of G3 . It is reduced by the increased chromatic dispersion. Thus, G2 is most
vulnerable to CD. In our experiment, the target distances are below 4.4km. At a distance of 2.2
km, a minimum net bit rate per channel of 86.2 Gb/s for all mode groups is achieved and the
total throughput assuming the 40 channels per group is 14.5 Tb/s. At 4.4 km, the minimum
achievable net bit rate per group is 84 Gb/s which leads to a total throughput of 13.5 Tb/s.
To verify that we have successfully transmitted all the 160 channels at 2.2 km. The BER is
measured for all the channel. The result is given in Fig. 4.30. Each value of BER represents
120 Chapter 4. Transmission Experiments

5.0E-02
𝐺1
Group 1 𝐺2
Group 2

𝐺3
Group 3 𝐺4
Group 4

BER FEC threshold


5.0E-03

5.0E-04
0 20 40 60 80 100 120 140 160
Channel index

Figure 4.31: Measured BER for the all 160 channels after 2.2km of OM2 fiber.

the averaged BER over 30 waveforms. we observe that all the channels operate below the FEC
limit of a BER of 5 × 10−3 . The total net throughput of 14.5 Tb/s is successfully transmitted.

This section has been dedicated to transmission over standard OM2 fiber for LAN and intra-
datacenter applications where the reach is shorter than 5 km. The main driving factor for this
work is that the market already exists and the multimode fibers are already standardized and
deployed. The fiber installation cost is then neglected. We demonstrated that mode group
multiplexing enables to increase the capacity over these fibers while keeping low cost direct
detection schemes without any complex DSP. First 5 Tb/s have been demonstrated using
selective excitation of 4 mode groups and multiplexing 20 wavelengths per group. Second,
demonstration of the same concept using low cost integrated EMLs with EAM has been
achieved for single wavelength transmission. Finally, a record transmission of 14.5 Tb/s has
been discussed in this section. Nevertheless, cost reduction of the presented setups should be
further reduced. For example, removing EDFAs will offer a good cost and power reduction
consumption. However multimode photodetectors with high sensitivity should be used. As a
part of this thesis, we worked in collaboration with III-V Labs to develop APDs adapted to
our application. The tested APDs have a narrow bandwidth which limits their use. However
research on this topic is still in progress. Moreover, operating at 1.3 µm instead of 1.55 µm will
reduce the impact of chromatic dispersion hence more capacity could be transported when the
4.3. Mode-Group-Multiplexing for DCI 121

fiber length is increased. However, fiber attenuation is more important in this case. Finally,
mode group multiplexing could be a good candidate to scale capacity over already installed
MMF. In the following section, we will explain how mode group multiplexing technique could be
extended to other application areas such as inter-datacenter or campus datacenter application
where standard MMFs are not appropriate.

4.3 Mode-Group-Multiplexing for DCI

Section 4.2 detailed over achievements of multi-terabit transmission over conventional OM2
fibers using selective excitation of 4 groups of modes. However, we are still limited in terms
of reach. The technology based on OM2 fiber could not be deployed for large distances due
to the increased fiber’s crosstalk at higher reach in addition to chromatic dispersion impact.
The accumulated losses did not limit the performance because they have been compensated
using an EDFA. However in a real short reach system implementation, the cost of EDFA
is unwanted and their suppression is desirable. Therefore, the large bandwidth multimode
avalanche photoreceiver could be a low cost alternative. Nevertheless, mode group multiplexing
technique enabled a high boost in terms of transported capacity over standard MMFs and could
find large application areas such as LANs and short reach intra-data centers communications.

This technique could be extended to applications where conventional OMx fibers are not appro-
priate, for example, in datacenter interconnect (DCI) where longer reach is required. Currently,
SMF fibers are deployed for DCI. The current IEEE standard is 400 Gb/s employing 4-PAM
at 25 Gbaud using 8 optical lanes (LR8) and doubling the symbol rate to 50Gbaud using 4
optical lanes (LR4). Moreover, FMFs could be a good alternative instead of OMx fibers. Since
standard multimode fibers are already deployed and standardized, to respect the low cost crite-
rion, mode group multiplexing using these fibers does not imply modifying any fiber parameter.
Innovation and optimizations concerned only the DSP, opto-electronic devices and the MGM
and MGD design. In contrast with OM2 fiber, research on FMFs enables to modify their
properties in terms of number of propagating modes and refractive index profile. As detailed
122 Chapter 4. Transmission Experiments

in chapter 3, the use of FMFs enables to modify their fiber properties to reduce the mode
coupling and attenuation. Proper design of the refractive index profile enables to maximize
the effective index difference ∆neff between the LP modes hence reduce the modes coupling.
However, the fiber refractive index profile should be carefully designed to find a good trade
off between ∆neff , macro-bending loss, the effective area Aef f that has to be large to avoid
non-linearities and the fiber attenuation per mode .

In this section, systems based on OM2 fibers and FMFs are compared in terms of MGM and
MGD and fiber design. Furthermore, generated crosstalk and losses due to MGM and MGD
in addition to the fiber are compared. This comparison enables to understand the transition
toward longer reach applications using mode group multiplexing.

4.3.1 Transition from OM2 to FMF

The fiber used in this work is a weakly-coupled FMF [ref] presented earlier in chapter 3, it
supports 10 spatial modes: 2 non-degenerate LP modes (LP01 &LP02 ) and 4 two-time-degenerate
LP modes (LP11a,b , LP21a,b , LP31a,b &LP12a,b ). This FMF has a standard glass diameter of 125
µm and has been fabricated with standard manufacturing processes that allow for large-scale
production. The minimum effective index difference between the LP modes (min|∆neff |) is as
high as 1.5 × 10−3 , which reduces mode coupling. The effective areas are between 84 µm2 and
100 µm2 . An improved trade-off between high min|∆nef f | and effective areas and attenuation
compared to step-index profiles with the same number of LP modes has been obtained by
introducing an optimized depressed zone in the center of the step-index core as shown in
Fig.4.32.
With this FMF we use a lower number of modes compared to our previous work where we
used ten spatial modes supported by 4 different groups of modes. Hence, both MGM and
MGD will handle a lower number of modes and have reduced modal crosstalk due to their
lower complexity. CAILabs designed the MGM and MGD based on the MPLC technology.
We used the first four LP modes for transmission. We used a bidirectional configuration. Fig.
4.33 shows the propagating modes in each direction. In one direction two groups of modes are
4.3. Mode-Group-Multiplexing for DCI 123

Index difference (10-3)

Radius(µm)

Figure 4.32: FMF refractive index profile and neff for the six LP modes used in the experiment.

used: G1 with LP01 mode and G3 with LP21ab modes. In the reverse direction G2 with LP11a,b
modes and G4 with LP02 mode are used. Moreover, since the large bandwidth photodiode has
a small effective area, it is important that the spot size on the receiver is as small as possible.
Therefore, the MGD converts the high order mode profiles into lower order mode profiles.
Fig. 4.33 illustrates the modal composition of the detected mode groups. We observe that the
higher order modes profiles are the LP11ab supported by G2 and G3 while G1 and G4 carry the
fundamental mode. Therefore the MGD has been designed to selectively demultiplex G1 and
G4 into SMF while it demultiplex G2 and G3 into OM2 fiber. The copropagating mode groups

Propagating Modes detected by


modes the photodiode

𝐺1 :
Mode group1
LP01 LP01
𝐺2 :
Mode group2
LP11b+ LP11a LP11b+ LP11a
𝐺3 :
Mode group3
LP21a+LP21b LP11b+ LP11a
𝐺4 :
Mode group4

LP02 LP01

Figure 4.33: Modal composition of propagating mode groups and detected mode group after
MGD by the photodetector.

have been selected so that they have the maximum ∆neff between them. The ∆neff between
the groups is given in Table 4.5. We notice that the LP02 mode which is supported by G4 has
the minimum ∆neff . Hence, if we suppose that the crosstalk strongly depend on the ∆neff ,
124 Chapter 4. Transmission Experiments

this group will represent the higher impairments in transmission. To evaluate the crosstalk

LP01 − LP11 LP11 − LP21 LP21 − LP02


difference of refractive 2.8 3.8 1.5
index (10−3 )

Table 4.5: ∆nef f between used mode groups

reduction enabled by the low complex design system based on FMF compared to the one based
on OM2, the generated crosstalk by the couple MGM and MGD is compared for both systems
in two configurations, tHe first configuration has 20 m of fiber in the absence of splices (as
it has been received from CAIlabs). While in the second configuration we cut the fiber and
made two splices. In fact, the crosstalk is vulnerable to the quality of splices, therefore, we
did several splices and averaged the value of crosstalk. The crosstalk was measured using an
ASE source. Fig. 4.34 plots the crosstalk as a function of mode group index. Blank markers
represents the case with no splices while filled markers represents the case with two splices.
For system with OM2, in the absence of splices, the worst crosstalk was measured for G3 and
is equal to -19.6 dB, in the presence of splices, this value is increased by 4.2 dB. In fact, this
degradation is principally due to the MGM and MGD. Even with no splices this group has the
high measured crosstalk. Moreover, when splicing the fiber, this group is more degraded. For
system with FMF, in the absence of splices, the worst crosstalk is attributed to G2 and G3
and equal to -19.6 dB. The splices increase the crosstalk by approximately 2dB for all groups
except G3 which is the less impacted one with only 0.8 dB more crosstalk. Hence, OM2 fiber
is more vulnerable to splices compared to FMF. Furtheremore, comparing both systems, we
notice that using FMF the worst case is improved by 5.3 dB. This could be explained by the
low complexity of the phase plates of the MGM and MGD. In fact, each plate manages 3 modes
profiles compared to 5 modes for the system based on OM2.

The newly designed MGM and MGD will enable longer reach transmission over FMF. However
fiber crosstalk scaling with fiber length should be considered. The fiber crosstalk is measured
for OM2 and FMF and compared for both fibers. First, the generated crosstalk by the MGM
and MGD is measured when few meters of fiber (20m) is spliced between the MGM and MGD
(two splices) as represented in Fig. 4.34. Second, the crosstalk is measured when the fiber
4.3. Mode-Group-Multiplexing for DCI 125

-5

-10

10.4dB
-15
5.3dB
crosstalk [dB 6.6 dB
-20

-25
MGM+MGD no splices (OM2)
0.4dB MGM+MGD + 2 splices (OM2)
-30
MGM+MGD no splices (FMF)
MGM+MGD +2 splices (FMF)
-35
1 2 3 4
mode group index

Figure 4.34: Comparison of MGM and MGD designed for transmission over FMF and OM2
in terms of generated crosstalk in the absence of splices , and in the presence of two splices
between the MGM and MGD with 20 m of FMF and OM2 fiber.

spool is spliced between the MGM and the MGD. Finally, the measured crosstalk with 20 m
of fiber is subtracted from the measured crosstalk with multiple length of fiber. To have a
good estimation of fiber crosstalk, we did several measurements with different splices. The
measured crosstalk values are then averaged over all these measurements. Fig. 4.35(a) depicts
the crosstalk as a function of distance for each group of mode considering the MGM, MGD ,
fiber and 2 splices. For both systems, G1 has the lowest crosstalk. G3 represents the worst
group with OM2 while G2 represents the worst with FMF. In fact, this is mainly due to MGM,
MGD and splices as explained previously. Fig. 4.35(b) gives the fiber crosstalk in dB/km for
each group in the case of FMF and OM2. We observe that both fibers have the same impact
on G1 . Furthermore FMF outperforms OM2, where the worst case is improved by 6dB. Fig.
4.36 illustrates the evolution of fiber’s crosstalk contribution in percentage compared to the
total crosstalk (MGM + MGD +2 splices +fiber) as a function of distance for all groups, in
(a) for FMF and in (b) for OM2. For FMF, the crosstalk evolution is more important for G4 .
This may be explained by the low ∆neff of the mode LP02 supported by this group. Moreover,
the percentage of fiber crosstalk for G3 is lower. In fact, this later has the higher ∆nef f . For
OM2 fiber, G2 is the most impacted group. In fact, even at 20 m of fiber this group has the
highest crosstalk when the fiber length increases, this group is coupled with higher order modes
126 Chapter 4. Transmission Experiments

supported by the fiber so it suffers from higher crosstalk.


-5 -10.00

-10
-20.00

Crosstalk [dB/km]
-15
10.5dB 13.4dB
Crosstalk [dB]

-30.00 6dB
-20

-25
-40.00

-30 OM2 FMF


G1 (OM2) G2 (OM2) G3 (OM2) G4 (OM2)
G1 (FMF) G2 (FMF) G3 (FMF) G4 (FMF) -50.00
-35 1 2 3 4
0 2 4 6 8 10 mode group index
Reach [km]

(a) (b)

Figure 4.35: (a) Measured fiber crosstalk for each group using both systems: blanked markers
with MGM and MGD designed for OM2 based transmission, filled markers with MGM and
MGD designed for FMF based transmission. (b) deduced fiber crosstalk in dB/km per each
group.

50 50

40
Percentage of fiber crosstalk
Percentage of fiber crosstalk

40

30 30

20 20

10 10 G1 (OM2) G2 (OM2)
G1 (FMF) G2 (FMF)
G3 (FMF) G4 (FMF) G3 (OM2) G4 (OM2)
0 0
1 4 7 10 1 4 7 10
Reach [km] Reach [km]
(a) (b)

Figure 4.36: Impact of fiber’s crosstalk in percentage versus distance (a) for OM2 fiber and (b)
for FMF.

Systems with FMF and OM2 are also compared in terms of generated losses. Table .4.6 gives
the generated losses by the MGM and MGD with few meter of spliced fiber between the MGM
and the MGD in the case of OM2 and FMF configurations. For OM2 the losses are degraded
compared to those given in section 4.2 by 0.6 dB for G3 and 1.1 dB for G4 . In fact, the values
4.3. Mode-Group-Multiplexing for DCI 127

presented previously correspond to the best optimized splices while those in Table 4.6 are the
averaged value over many splices (including the best and the worst ones). Furthermore, we
notice that systems with FMF improves the power budget by approximately 1.3 dB for the
worst group which is due to the system complexity reduction. System with FMFs result in
higher power budget. However, for transmission over long distances amplification using EDFAs
may be necessary in the absence of photoreceiver with great level of sensitivity such as APDs.

MGM+MGD+20m loss of G1 loss of G2 loss of G3 loss of G4


of fiber (with two splices) in dB in dB in dB in dB
OM2 -5.7 -5.25 -6.25 -6.73
FMF -5 -4.58 -5 -5.5

Table 4.6: Generated loss by MGM+MGD+20 m of fiber (in case of OM2 and FMF).

Fig. 4.37 presents the measured loss for each group for systems with OM2 (a) and FMF in
(b) as a function of distance. The losses include the losses of MGM, MGD, splices and fiber.
We notice that the losses are different for each group, This difference is due to the MGM and
MGD for both systems in addition to the different losses generated by the fiber. Moreover, we
observe that G1 has more loss compared to G2 . In the case of FMF this could be explained
by the depressed zone in the center of the index profile. Furthermore, standard OM2 fiber has
been designed to transmit all the supported modes (55 modes at 1.5 µm) at the same time.
Hence its refractive index profile could have some manufacturing defects which is negligible
because its impact on all the modes together is negligible. The comparison between systems
based on FMFs and those based on OM2 showed a good reduction of crosstalk and losses in the
case of FMF, which will enable to transmit data over longer distances compared to OM2 fiber.
It enables mode group multiplexing to find applications out of intra datacenters and LAN.
However, when operating at 1.5 µm, chromatic dispersion remains an important impairment.

4.3.2 20 km Transmission Experiment

In this section, we demonstrate bidirectional transmission over 20 km using the system with
FMF, MGM and MGD presented previously with low modal crosstalk. Depending on the losses,
128 Chapter 4. Transmission Experiments

-4 -4
G1 (OM2) G2 (OM2) G1 (FMF) G2 (FMF)
-5 G3 (OM2) G4 (OM2) G3 (FMF) G4 (FMF)

-6 -6
Loss [dB]

Loss [dB]
-7

-8 -8

-9

-10 -10
0 2.2 4.4 6.6 8.8 0 2.2 4.4 6.6 8.8
Reach [km] Reach [km]

(a) (b)

Figure 4.37: Measured losses generated by two systems:(a) MGM+MGD+OM2 and (b)
MGM+MGD+FMF.

crosstalk and chromatic dispersion, each group will be used to transmit different capacities.
Achievable bitrates using DMT and 4−PAM are experimentally compared. Fig. 4.38 depicts the

RX3
DAC1 Driver1
EDFA1 RX1
G1 G1 Multimode RF
MZM1 PIN amplifier
LASER
(lco) G3 MGM MGD
G3
MZM3 20km of FMF Offline ADC
processing
Driver3 EDFA3
DAC3
Driver2 DAC2
EDFA2
Offline G2 splices G2
ADC MZM2 LASER
processing MGD MGM
G4 G4 (lcontra)
RF Multimode MZM4
amplifier PIN
RX2 EDFA4
Driver4 DAC4
RX4

Figure 4.38: Experimental setup for transmissionn over FMF.

experimental set-up. At the transmitter, light from single laser source is split by a polarization
maintaining 3 dB coupler. Each output is then modulated with a dedicated MZM driven
by an 88 Gs/s CMOS DAC with 22 GHz analog bandwidth, such that independent data are
used for co-propagating mode groups. Laser sharing is possible due to the low crosstalk of the
system. In section 4.2, we used different wavelengths for the co-propagating groups to avoid
coherent crosstalk due to the high system crosstalk. In the case of this experiment the low
modal crosstalk allows for the use of the same laser source and reduces the system cost. In the
4.3. Mode-Group-Multiplexing for DCI 129

reverse direction, a different wavelength is used due to non-optimized return loss in the MGD
and MGM. Crosstalk and losses corresponding to MGM, MGD, 20 km of FMF and 2 splices
are given in Table. 4.7. As expected from the previous section, G4 is the most attenuated
group with 12.4 dB of loss at 20 km. Moreover, G2 has the largest crosstalk value which is
compatible with the previous measurements.

Mode group Losses [dB] crosstalk [dB]


G1 -10 -23
G2 -10.5 -10.1
G3 -11.7 -17
G4 -12.4 -16.6

Table 4.7: Loss and crosstalk generated by MGM,MGD, 20km of FMF and 2 splices.

To partially compensate for the losses of the MGM, MGD and the FMF, a single-mode EDFA
is used after each modulator. The input power of each photodiode is set to +1dBm by tuning
the output power of each EDFA. The output power of EDFA1,2,3,4 are respectively 10.9, 11.7
,12.6 and 13.2 dBm since the ouput power of EDFAs are not high, they can be removed using
PIN TIA or APDs. At the receiver, the signal is detected with a multimode PIN photodiode
with 28 GHz bandwidth followed by a 32 GHz linear amplifier. The signal is then sampled by
a 92 Gs/s ADC, followed by off-line signal processing. We investigate two different modulation
formats, i.e., 4-PAM and DMT. In the case of the 4-PAM format, at the transmitter, the levels
are optimized to compensate the non-linear response of the MZM modulator then a root raised
cosine filter with roll-off factor of 0.2 is applied. At the Rx side, the signal is resampled, time
recovery is used to synchronize the signal, we apply the matched filter and a T/2-spaced feed
forward equalizer with 51 taps for equalization. In the case of the DMT format, the FFT length
is set to 512, with 30% maximum amplitude clipping and a cyclic prefix composed of 16 samples.
Optimal bit loading adapted to the channel response is applied to maximize the rate for a BER
target of 5 × 10−3 . The latter, corresponds to the threshold of a hard decision FEC with 7%
overhead. Firstly, the performance in the absence of the co-propagating modes is measured. It is
measured by launching power into one group and turning off its corresponding co-propagating
group (for example: G1 on and G3 off). Using the DMT format, the SNR response for all
groups is measured as a function of subcarrier frequency by assigning the QPSK format to all
130 Chapter 4. Transmission Experiments

subcarriers. Fig. 4.39 depicts the SNR response for all groups after 20 km of propagation. We

20
G1 G3
15
SNR [dB]

10

0
0 4 8 12 16 20 24 28 32
Frequency [GHz]

20
G2 G4
15
SNR [dB]

10

0
0 4 8 12 16 20 24 28 32
Frequency [GHz]

20
G1 G4
15
SNR [dB]

10

0
0 4 8 12 16 20 24 28 32
Frequency [GHz]

Figure 4.39: Measured SNR response per mode group after 20km propagation over FMF at 1
dBm Rx input power

observe power fading arising from the accumulated fiber chromatic dispersion (CD). However
this fading is different for all groups which means that CD value is different per group. In
chapter 3, it has been explained that CD depends on the constant of propagation of LP modes.
Hence, its value will be different for each LP mode. This phenomenon can be clearly seen in
Fig. 4.39. Moreover the shape of the power fading of G1 indicates that G1 has the lowest
CD while G3 is the worst group in terms of CD. To validate our observation, we estimated
the dispersion value of each group based on their SNR response. In fact the SNR response is
proportional to the channel’s intensity responses according to :

SNRGi (f ) ∝ |HGi (f )|2 (4.1)

in which SN RGi is the SNR of group with i index Gi and HGi is the propagation channel
4.3. Mode-Group-Multiplexing for DCI 131

response seen by Gi . In direct detection, the information about the propagation channel is
given by its real part. Hence, in the presence of CD we have:

|HGi |2 ∝ cos2 2π 2 β2,i f 2 L



(4.2)

and
λ2
β2,i = −Di (4.3)
2πc

where Di is the CD parameter of Gi in unit of ps/(nm.km). In Fig. 4.39 the frequencies of


power fading represent the solutions of:

cos2 2π 2 β2,i f 2 L = 0.

(4.4)

and are therefore given by :

1  c 
f0 = +k ,k ∈ Z (4.5)
2 λ2 .Di

Hence, Di is estimated from the experimental SNR response of Gi . We take the value of the
frequency located at the power fading and compute Di from equation (4.5). The values of these
frequencies are given in Table 4.8. Furthermore, estimated Di and the simulated values Di by

Mode group f0 [GHz] estimated Di simulated Di equivalent length


[ps/(nm.km)] [ps/(nm.km)] (Prysmian) of SMF [km]
G1 21.62 20.01 20.5 23.5
G2 24.62 25.73 25.35 29.4
G3 22.74 30.15 31.25 36.4
G4 23.7 27.66 26.95 31.6

Table 4.8: CD parameter estimation for each mode groups and its comparison with simulated
value by Prysmian.

Prysmian are also presented in table 4.8. As expected, we remark that G1 is less impacted by
CD. Considering only CD impact, its propagation is equivalent to 23.5 km over SMF with 17
ps/nm.km of CD. Moreover, we observe that G3 has the highest dispersion and its propagation
over 20 km of FMF is equivalent to 36.4 km propagation of SMF in terms of CD impact.
Consequently G3 will transport the lowest capacity compared to the other groups. Also, the
132 Chapter 4. Transmission Experiments

estimated value of Di are in a good match with those simulated by Prysmian. Fig. 4.40
illustrates the measured SNR response and our approximated model of |HGi |2 . To consider the
total bandwidth attenuation, a triangular filter with 3 dB bandwidth of 22 GHz (corresponding
to the analog bandwidth of the DAC) is considered. We notice that for all the mode groups,
the adopted model approximates the SNR, response which validates the estimated value of Di .

Figure 4.40: Approximate model to estimate the chromatic dispersion value per mode group

The difference of CD impact on each mode groups explains the results in Fig. 4.41 (a) and (b)
where the achievable rates for each group will also be different. Fig. 4.41 (a) and (b) display
respectively the BER versus achievable bitrate (without assuming 7% overhead) for G1 and
G3 in the first direction and G2 and G4 in the reverse direction considering 4-PAM and DMT
format. For each value, 20 waveforms have been processed. First, we evaluate the performance
in the absence of crosstalk (blanked markers). In this case, it is observed that DMT achieves
4.3. Mode-Group-Multiplexing for DCI 133

higher bitrates compared to 4-PAM for all mode groups. Moreover, for G3 as the bit rate
increases, the BER of 4-PAM increases much faster than the BER of DMT. In fact this group
is the most impacted group due to its highest dispersion compared to the other groups ( ∼ 30.15
ps/(nm.km)), as given in Table 4.8. This shows the resilience of DMT to chromatic dispersion
achieved by adapting the bit loading to the channel response. Furthermore, using DMT, the
achievable bitrate per mode group is different, and the BER degradation when the bitrate
increases is different for thess two mode groups. Secondly, the BER is measured considering
the propagation of the co-propagating mode group with the same wavelength (filled markers).
Fig. 4.41(b) shows that 4-PAM is highly impacted by modal crosstalk. This observation has
been expected from the analytical investigations presented in chapter 3. In this case, using the
4-PAM format, the net bitrate is limited to 38 Gb/s considering a BER threshold of 5 × 10−3 .
Contrary to 4-PAM, DMT modulation scheme shows good resilience to modal crosstalk fo all
mode groups. In fact, G1 and G3 present only a small penalty due to their low crosstalk values
displayed in the Table 4.7. Achievable bitrates for G1 and G3 , are respectively 80 Gb/s and 70
Gb/s. Moreover, for G2 with the highest crosstalk, 4−PAM does not allow for any transmission
over this group. Therefore, we consider only DMT modulation. Fig. 4.41(b) shows the BER
as a function of the transmitted bitrate for G2 and G4 . It is observed that G2 has the highest
BER penalty due to its high modal crosstalk value. Achievable bitrates for G2 and G4 are 65
Gb/s and 70 Gb/s respectively. This results in a total of 285Gb/s gross bit rate (265 Gb/s
net) using the four mode groups. However, signals can encounter time fluctuation due to the
random crosstalk phase variation. Hence, depending on the speed of its variation, the BER will
fluctuates between a worst and best case. Furtheremore, [121] has shown that for multicore fiber
the modal crosstalk of carrier-supported signals fluctuate over time. Hence, to ensure sufficient
margin in our, system we choose to modulate all the mode groups at 55 Gb/s, resulting in a
total of 220 Gb/s (205 Gb/s net bitrate) and measured BER fluctuation over 150 minutes.

Stability results are displayed in Fig. 4.42 (a) and (b) showing the BER as a function of time
for all mode groups. The BER performance of all mode groups stays below the FEC threshold,
with G1 giving the largest margin such that its corresponding rate could be set higher than 55
Gbit/s due to its lowest modal crosstalk and dispersion value. G2 gives the lowest margin, due
134 Chapter 4. Transmission Experiments

10-1 10-1 G2 With G4


G1 Without G3
G1 With G3 G2 without G4
G4 with G2
G3 Without G1 G4 without G2
G3 With G1

10-2 10-2
FEC threshold FEC threshold
BER

BER
10-3 10-3

10-4 10-4

4-PAM DMT
4-PAM DMT
10-5 10-5
20 30 40 50 60 70 80 90 20 30 40 50 60 70 80 90
Rate [Gbit/s] Rate [Gbit/s]
(a) (b)

Figure 4.41: Measured BER after 20km of propagation using 4−PAM and DMT : (a) for G1
and G3 , (b) for G2 and G4

10-1
10-1
G2 G4 G1 G3

10-2
FEC threshold 10-2 FEC threshold

10-3 10-3
BER

BER

10-4 10-4

10-5 10-5
0 30 60 90 120 150
0 30 60 90 120 150
Time [mn]
Time [mn]
(a)
(b)

Figure 4.42: BER stability due to variations in modal crosstalk for all mode groups modulated
at 55Gb/s DMT: a) for G2 and G4 b) for G1 and G3

to its high crosstalk. The performance of all groups fluctuates between a worst and best case.
This observation could give an interesting information about the variation speed of crosstalk’s
phase. The latter variates between two acquisition and does not fluctuate rapidly during one
acquisition. To better understand and confirm our observations as well as the conclusion of the
analytical investigation in case of 4-PAM and DMT presented in chapter 3, we did additional
4.4. Experimental Investigation of Crosstalk Impact on 4-PAM and DMT Formats 135

experiment with new MGM, MGD and FMF designs. This experiment will be detailed in the
next section.

4.4 Experimental Investigation of Crosstalk Impact on

4-PAM and DMT Formats

Inter-modal crosstalk is a great challenge for MMF deployment. The main difficulty is to
increase optical transmission capacity over MMF while keeping low cost and low power con-
sumption. Therefore, inter-modal crosstalk compensation using complex DSP is not a suitable
solution. Significant progress in optical components have been made. For instance, low com-
plexity MGM and MGD in addition to FMF characteristic presented in section 4.2 and 4.3
demonstrated a transmission throughput up to 14.5 Tb/s over 2.2 km of standard OM2 fiber
and 200 Gb/s over 20 km of FMF. Despite these progress, some modulation format such as
4-PAM showed a poor performance compared to DMT implementing mode group multiplex-
ing technology. First, we operate at 1.55 µm and CD limits the use of single carrier 4-PAM
due to the power fading impact. Second, its vulnerability to modal crosstalk has been clearly
observed compared to DMT. In fact, preliminary analytical study of localized crosstalk in
chapter 3 section 3.5.1 revealed the resilience of DMT to inter-modal crosstalk compared to
4-PAM. Moreover, experimental results in section 4.3 confirmed our expectations in the case of
distributed crosstalk. In this section we conduct a deep experimental assessment to compare
both formats, 4-PAM and DMT. New designed MGM, MGD and FMF were used in this
work.

4.4.1 Polarization Maintaining setup

The crosstalk is divided into coherent crosstalk where the phase of the desired signal and that of
the crosstalk’s signal are correlated and incoherent crosstalk where the phase of the interfered
signals are not correlated. The coherent crosstalk causes power fluctuation, hence the signal
136 Chapter 4. Transmission Experiments

performance will fluctuates between a best and worst case scenarios depending on the speed
of crosstalk phase variation. Incoherent crosstalk has much lower impact. Indeed, incoherent
crosstalk causes power noise since the crosstalk is only added in intensity. To better compare
crosstalk fluctuations using 4-PAM and DMT format, we focus on the case of coherent crosstalk.
We use bidirectional transmission where two groups of modes are co-propagating in the same
direction. We maximize the range of crosstalk phase variation by using the same wavelength and
aligning the polarizations of the co-propagating mode groups. Therefore, Cailabs designed new
MGM and MGD thar allow for polarization maintaining. Fig. 4.43 illustrates the concept
of this component. We keep the bidirectional configuration with two groups of modes per
direction: G1 and G3 co propagate in the forward direction and G2 and G4 co propagate in the
opposite direction. The modal composition per group is the same as for the MGM and MGD
used in section 4.3. Single mode polarization maintaining fiber is connected at the input of
the MGM. The added functions to these components compared to those presented in previous
section is the polarization maintaining fiber (PMF) used at the input of the MGM that allows
to create mode profile having the same polarization state at its input. This MGM enables
the investigation of two configurations as illustrated in Fig. 4.43. Configuration 1, where the
co-propagating mode groups will have cross polarizations. This configuration increases the
incoherent crosstalk while configuration 2 enables the co-propagating mode groups to have
aligned polarization, hence, it emphasizes the coherent crosstalk due to the use of the same
wavelength for the copropagating mode groups. The latest configuration represents the best
scenario to compare the crosstalk’s fluctuation using different modulation formats. Table. 4.9
gives the crosstalk and losses per mode group generated by MGM, MGD and 20 m of FMF
without splices.
Mode group Losses [dB] crosstalk [dB]
G1 -5 -28
G2 -6 -10.4
G3 -4.7 -12.4
G4 -5.2 -17.2

Table 4.9: Measured Loss and modal crosstalk generated by maintaining polarization MGM,
MGD and 20 m of FMF.

We used a weakly coupled 6-LP mode fiber designed by Prysmian [98]. Its refractive index
4.4. Experimental Investigation of Crosstalk Impact on 4-PAM and DMT Formats 137

config1 config2
config1 config2
𝐺1 𝐺1 𝐺2 𝐺2
𝐺1 PMF SMF 𝐺1
MGM MGD
FMF PMF 𝐺2
𝐺2 FMF
𝐺3 𝐺3 PMF FMF 𝐺4
𝐺4
𝐺3 𝐺3
SMF MGD MGM PMF
𝐺4 𝐺4

Input polarization states Input polarization states


: polarization fast axis

: polarization slow axis

Figure 4.43: Description of concept of polarization maintaining experimental setup.

profile has been designed to reduce higher order mode attenuation due to Rayleigh and small
angle light scattering (SALS). In fact, the scattering effect on the LP modes is more impacting
when the mode energy is important at the core-cladding interface and this interface is steep
in terms of difference of refractive index. Hence, the attenuation increases significantly with
the order of LP modes. Therefore, Prysmian proposed a smooth interface by introducing a
trapezoidal-index profile and keeping the core depressed-index zone in the center of the step
index profile to reduce mode coupling that takes place between different modes of cylindrical
symmetric spatial modes namely the LP02 mode. Fig. 4.44 compares the refractive index
profiles of thew FMFs used in section 4.3 in (a) and the one that will be used in this experiment
in (b). Both fibers have the same min|∆neff | which is equal to 1.6 × 10−3 however the fiber in
(b) has approximately 4 times lower differential mode attenuation (0.025 dB/km) compared to
the fiber in (b)(0.11dB/km). More details about the fibers characteristics could be found in
reference [98].

4.4.2 Experimental Investigation

Fig. 4.45 depicts the experimental setup used to investigate the crosstalk’s phase fluctuation
and its effects on BER fluctuation for 4−PAM and DMT. We use the same wavelength for the
co-propagating mode groups. Thus, at the transmitter side, a single laser source is split by a
polarization maintaining 3 dB coupler. Each output is then modulated with a dedicated MZM
138 Chapter 4. Transmission Experiments

(a) (b)

Figure 4.44: Refractive index profiles of 6-LP-mode FMF with inner core depressed-index (a)
with a step-index structure, (b) with a trapezoidal-index structure.

DAC1 Driver1 RX3


RX1
PMF G1 G1
MZM1 Multimode RF
LASER PIN amplifier
PMF G3 MGM
(lco) MGD G3
MZM3 Few meters of FMF Offline ADC
processing
DAC3 Driver3
Driver2 DAC2

Offline G2 G2 PMF
ADC MZM2 LASER
processing MGD MGM G4 PMF (lcontra)
G4
RF Multimode MZM4
amplifier PIN
RX2
RX4 Driver4 DAC4

Figure 4.45: Experimental setup used for crosstalk’s fluctuations investigation

driven by an 88 Gs/s CMOS DAC we use PMFs to connect each MZM and the input of the
MGM. At the receiver, the signal is detected with a 28 GHz multimode PIN photodiode. The
input power is set to −1 dBm. The signal is then sampled by a 92 Gs/s ADC followed by
off-line signal processing. We investigate two different modulation formats, particularly, 4-PAM
and DMT. In case of 4-PAM, a root raised cosine filter with roll-off factor of 0.16 is used, and
a T/2-spaced FFE equalizer with 51 taps is applied for equalization. For the DMT format, the
FFT length is set to 512 with 25% maximum amplitude clipping. Assuming the use of FEC
with 7% overhead, 100 Gb/s net bitrate is transmitted using 56 GBd for 4-PAM and adapted
bit loading algorithm for DMT where the mapped bit per symbol ranges from 1 to 6.

We assessed crosstalk fluctuations for all mode groups via the BER metric in the case of 56
4.4. Experimental Investigation of Crosstalk Impact on 4-PAM and DMT Formats 139

4-PAM: G1 Aligned polarization 4-PAM: G3 Aligned polarization


4.E-02 Cross polarization 4.E-02 Cross polarization

BER
BER

4.E-03 4.E-03

4.E-04 4.E-04
0 25 50 75 100 125 150 0 25 50 75 100 125 150
Waveform index Waveform index

Aligned polarization 4-PAM: G4 Aligned polarization


4-PAM: G2
4.E-02 4.E-02 Cross polarization
Cross polarization

BER
BER

4.E-03 4.E-03

4.E-04 4.E-04
0 25 50 75 100 125 150 0 25 50 75 100 125 150
Waveform index Waveform index

Figure 4.46: Fluctuation of BER for all mode groups in two different configurations in terms
of co-propagating mode group polarization state (cross and aligned polarization).

GBd 4-PAM in B2B for 150 measurements and for Rx input power of −1 dBm. Results are
shown in Fig. 4.46. We evaluate the BER fluctuation in the case of cross polarization (blanked
markers) and for aligned polarization (filled markers). Measurements are done consecutively
in the same environmental condition. In the case of cross polarization, the fluctuations of the
crosstalk are minimal. In the latter, the BER of all mode groups is around 1.3 × 10−3 . As
for aligned polarization, two observations can be made. For the first direction (where the co-
propagating mode groups are G1 and G3 ), the fluctuation’s range is approximately the same
as in the case of cross polarization. This is because, as indicated in Table. 4.9, the crosstalk
between these groups is low and thus the measurement of large fluctuations is not possible. For
the second direction (where the co-propagating mode groups are G2 and G4 ), and in contrast
to the first direction, large fluctuations can be measured and the BER exceeds the FEC limit
of 5 × 10−3 . We note that G2 has the highest fluctuations for aligned polarizations due to its
high crosstalk value given in Table 4.9 (-10.4 dB), thus, we turn our focus to this group in the
remainder of this section.
In what follows, we compare 100 Gb/s DMT and 4-PAM formats using G2 . More specifically, we
140 Chapter 4. Transmission Experiments

measured the BER fluctuation of these formats for 5 hours. Fig. 4.47(a) shows the distribution

(a) (b)

Figure 4.47: Histogram of 5 hours fluctuation of log10 (BER) for 100Gb/s 4-PAM and DMT
for (a) cross polarization and (b) aligned polarization.

of BER fluctuations when the polarization are crossed and Fig. 4.47(b) gives the distribution
of BER fluctuation in the case of aligned polarization. In the case of cross polarization,
100Gb/s 4-PAM shows better performance compared to 100 Gb/s DMT. The average BER
is 1.2 × 10−3 for 4-PAM and 2 × 10−3 for DMT. In fact, the 4-PAM spectrum occupies 28
GHz electrical bandwidth while DMT signal has 32 GHz useful bandwidth to optimize bit
loading and achieve 100 Gbit/s net bitrate. Hence the slight difference of performance could be
explained by bandwidth limitation in case of DMT in addition to its other impairments such as
its sensitivity to PAPR and quantization noise. In the presence of coherent crosstalk, where the
polarizations are aligned, in case of 4-PAM modulation, fluctuations are more important than
for DMT modulation. The measured worst BER for 4-PAM is 2 × 10−2 while it is 6.7 × 10−3
for DMT. Furthermore, Fig. 4.48 plots the temporal changes of BER at each moment when
one waveform is processed. For both formats, the fluctuations are uniform and their envelop
is constant. BER fluctuation occurs randomly and with a short time scale duration. In fact
as explained in section 3.5.1, desired signal and the signal of crosstalk couple constructively or
destructively depending on the crosstalk phase θ. Hence, the lowest BER corresponds to the
case where the crosstalk phase is close to π/2 (cos(θ) ∼ 0) while the highest BER corresponds to
the case where the crosstalk phase is close to 0 (cos(θ) ∼ 1). Fig. 4.50 (a) depicts the measured
4.4. Experimental Investigation of Crosstalk Impact on 4-PAM and DMT Formats 141

4-PAM aligned polarization


4.E-02
DMT aligned polarization
BER

4.E-03

4.E-04
0 100 200 300 400 500 600 700 800 900 1000
Waveforme index

Figure 4.48: 5 hours BER fluctuation in the case of aligned polarization for 4-PAM and DMT.

histograms of 100 Gb/s PAM-4 for the best case (lowest BER) and the worst case (highest BER).
Fig. 4.50 (b) illustrates the measured SNR per subcarrier for 100 Gb/s DMT in the worst and
best cases. For the best case 4-PAM, in the absence of crosstalk’s impact, the four levels are
equally spaced and clearly distinguished. All levels have the same Gaussian distribution due to
the impact of AWGN. For the worst case 4-PAM, in the presence of crosstalk’s impact, 4-PAM
levels do not have the same distribution. The distribution spread increases as the amplitude
level increases. We notice that the large part of errors is due to the highest order levels. For
the best case DMT, the stair steps in the SNR are due to the applied power allocation. In
fact, subcarriers carrying the same QAM format have approximately the same SNR. Applied
bit and power allocation are respectively given in Fig .4.49(a) and Fig .4.49(b). For the worst
case DMT, we notice that the SNR reduction is not the same for all subcarriers. Subcarriers
carrying higher order QAM format are more impacted. The SNR is reduced by approximately
2.2 dB for subcarriers having 5 and 4 bit/symbol, 1.2 dB for subcarriers having 3 bit/symbol
and 0.3 dB for subcarriers with 2 bit/symbol. The impact of crosstalk on subcarriers carrying
1 bit/symbol is negligible.
Finally, the presented results confirm our expectation regarding the vulnerability of 4-PAM to
the crosstalk phase variation and the resilience of DMT to these fluctuations.
142 Chapter 4. Transmission Experiments

(a) (b)

Figure 4.49: Measured (a) 4-PAM histograms and (b) SNR per subcarriers for DMT ,for the
lowest and highest BER

(a) (b)

Figure 4.50: (a)Allocated bit per symbol per subcarrier (b) allocated power per subcarrier, for
100G DMT

4.5 Summary

In the first part of this chapter, we demonstrated the achievable benefits of the mode group
multiplexing technique to increase the transported capacity over already deployed standard
MMFs in IMDD schemes. We employed selective excitation of 4 modes groups to create 4
spatial channel in single OM2 fiber. We used the DMT format and multiplexed 20 wavelength
channels per group to achieve 5 Tb/s throughput over 2.2 km of OM2 fiber without complex
DSP. WDM technique was made possible due to the availability of a 200 GHz grid multimode
WDM demultiplexer, which allowed keeping the same modal composition per mode group
4.5. Summary 143

and hence avoided large power fluctuations at the photoreceiver. Furthermore, we exploited
the advantages of mode group multiplexing for low cost applications by demonstrating a net
bitrate of 262 Gbit/s over 2.2 km of OM2 using an EML with an electro-optical 3 dB bandwith
of 18 GHz . The use of higher bandwidth EMLs could allow the net bit rate per mode group
by 34.5 Gbit/s and be in line with 400 Gb/s Ethernet standardization [IEEE P802.3bs].
Moreover, in this chapter we demonstrated the highest throughput transmitted over standard
MMF using direct detection at the time of its realization. In fact, we reported a transmission
record of 14.5 Tb/s net throughput in a bidirectional configuration over 2.2 km of conventional
OM2 fiber using DMT modulation scheme and direct detection. It was not only the highest
throughput transmitted but also the highest throughput ever transmitted over an optical fiber
for IMDD regime. We achieved this result by using selective excitation of 4 mode groups and
WDM multiplexing of 40 wavelengths for each mode group using a 100 GHz grid multimode
WDM demultiplexer. Furthermore, comparison and evaluation of different type of multimode
photodiodes, DACs and ADCs and their impact on transmitted bitrate have been analyzed.

In the second part of this chapter, we extended the IMDD mode group multiplexing technique
to longer reach applications up to 20 km using FMFs instead of OM2 fiber. Transition from
systems based on OM2 fibers to those based on FMFs has been detailed. First, a comparison
of the two systems is made in terms of MGM, MGD design, generated crosstalk and losses
per mode group. Second, a comparison in terms of fiber’s crosstalk impact has been evalu-
ated. Modal crosstalk has been remarkably reduced for systems with FMF which allowed for
increasing the transmission distance. The concept of mode group multiplexing using FMFs for
short reach has been experimentally demonstrated with 200 Gb/s net bitrate achieved over 20
km of 6-LP modes FMF. As part of this experiment, we investigated two modulation formats,
namely 4-PAM and DMT. DMT outperformed 4-PAM not only in terms of resilience to CD
but also in terms of resilience to modal crosstalk which was in line with our expectation from
the analytical study presented in section 3.5.1.

The last section of this chapter aimed to experimentally investigate the impact of crosstalk’s
fluctuation on both 4-PAM and DMT formats. To this end, we maximized the coherent crosstalk
by aligning the polarization of the desired signal and the crosstalk contribution. This was
144 Chapter 4. Transmission Experiments

achieved using a specially designed MGM and MGD which enable to maintain the polarization
state. The stability measurements of BER of 100Gb/s 4-PAM and 100 Gb/s DMT was
discussed. We have found that DMT is more resilient to crosstalk and thus it could be the
best candidate for SDM application using IMDD systems over short distances.
Conclusions and Perspectives

Optical network requirements are driven by the next technological revolution where the main
goal is the automation of everything. Machines would be asked to do creative tasks instead of
repetitive tasks. Physical and digital worlds will be connected thanks to revolutionary appli-
cations such as the Internet of things, augmented reality and virtual reality. Future networks
should provide high capacity to ensure person-to-person, person-to-machine and machine-to-
machine communications. Hence, low cost IMDD based short reach systems are very attractive
to support future optical networks requirements. Furthermore, optical architectures in local
area networks such as factory, hospitals and universities should be adapted to the ever-growing
demand on data traffic to implement the future revolutionary technologies.
Moreover, future optical networks should provide sub millisecond control with high reliability.
For example, critical applications such as autonomous cars and robots require less than 1 ms
response time. Hence, due to the finite speed of light, assuming an acceptable response time
of 1 ms, the maximum allowed round trip distance in an optical fiber is 100 km. Furthermore,
in practice, optical fibers are not deployed straightly which represents a maximum attainable
distance of 40 to 50 km between the users and the servers. Therefore, optical networks archi-
tecture will move from big datacenters to multiple distributed cloud processing centers close
to the users. Low cost IMDD based short reach systems are very attractive to support future
optical networks requirements.
In this thesis, we address the need of high capacity for short reach applications using mode
group multiplexed transmission over multimode fibers based on low cost IMDD schemes. First,
we focused on increasing the throughput over standard multimode fibers for two target mar-
kets. Namely, LAN for short term and intra-datacenter interconnects for long term. Regarding

145
146 Chapter 4. Transmission Experiments

LAN, the market already exists. Multimode fibers are already deployed and the mode group
(de)multiplexer are sold in the whole world by Cailabs. However, the current solution is 4 × 10
Gb/s using serial speed plugable and the reach is much lower than 5 km. For intra-datacenter,
MMF are mainly deployed for distances less than 500 m. Second, we turned our focus toward
a more ambitious and challenging application which is inter-datacenter interconnects.

The second chapter of this thesis was devoted on reviewing the classical architecture of IMDD
transmitter and receiver. The most common devices have been discussed. DML and EML
based transmitters were reviewed and compared in terms of cost advantages and disadvantages.
Furthermore we described the main aspects of two attractive modulation formats for IMDD
schemes which are 4-PAM and DMT. For 4-PAM, classical Tx and Rx DSP have been detailed.
It includes the pulse shaping to limit the signal bandwidth, the pre-compensation at the trans-
mitter side by applying the pre-emphasise filter and the Rx equalization block used to mitigate
the Inter-symbol-interference that impairs the optical. 56 GBd 4-PAM has been used to eval-
uate experimentally the impact of classical DSP functionalities on the BER improvements.
The experiment has been conducted in back-to-back were the main limitation arises from the
DAC and ADC analog bandwidth. The FFE enabled to significantly mitigate the ISI and the
pre-emphasize filtering increased the power margin by 2 dB. In the second part of chapter 2,
the conventional operational principle of DMT modulation has been explained. In contrast to
single carrier M-PAM formats, DMT has the advantage of better dispersion tolerance achieved
by cyclic prefix to reduce the inter-carrier-interference generated by the channel delay. Fur-
thermore, DMT has better spectral efficiency thanks to its loading algorithms. In this thesis,
rate-adaptive bit-loading has been discussed. Furthermore, we proposed to adapt the loading
algorithm for high BER target. In fact, an alternative approach to ”gap approximation” has
been introduced in order to optimize the bit and power allocation for high values of the BER.
The proposed method is based on a simple SNR thresholds setting from the relationship be-
tween the BER and the SNR. Chow’s bit loading algorithm and our optimized method have
been numerically compared for a highest BER target namely 1.7 × 10−2 .

In the third chapter of this thesis, we reviewed the fundamentals of optical fibers to explain the
modal composition of multimode fiber and the parameters contributing on whether the fiber is
4.5. Summary 147

a single mode, a few mode or a multimode fiber. Power fading impairment due to chromatic
dispersion impact in IMDD regime have been discussed. In multimode fibers, chromatic disper-
sion coefficient is different for each mode. Hence, the impact of chromatic dispersion is different
for each mode which leads to a difference in terms of achievable bit rates for each mode in the
case of spatial division multiplexing transmission.
In this chapter, the difference between mode division multiplexing and mode group division
multiplexing has been introduced. In fact, due to high mode coupling between modes having
close propagation constants, mode group multiplexing is implemented instead of independent
mode multiplexing to reduce crosstalk between multiplexed channels within the MMF without
using complex MIMO DSP. The principle of mode group (de)multiplexers based on MPLC
technology has been detailed. This technique has been proposed by CAIlabs in 2010. CAIlabs
contributed to this work by designing different pairs of mode group (de)multiplexers adapted
to our target applications.
Moreover, statistical investigation of inter-modal crosstalk has been performed by comparing
two different modulation formats for IMDD applications. We derived the crosstalk probability
density function theoretically and validated it using simulations. 4-PAM showed high vulner-
ability to crosstalk phase fluctuation compared to DMT format. This simple investigation has
been conducted in the case of localized crosstalk neglecting the propagation over the fiber. It
enabled us to confirm our observation during the experiments where we noticed higher penalties
for 4-PAM compared to DMT in the same experimental conditions using the same devices and
for the same level of crosstalk.

The last chapter of this thesis was dedicated to experimental investigations. It is divided into
three main parts as follows:
The first part discussed mode-group multiplexed transmission over standard multimode fibers,
namely, OM2 fibers. In fact, the motivation behind using standard multimode fibers is their
presence in many areas such as LAN and intra-datacenters. Our goal is to upgrade the through-
put over these fibers without adding cost due to fiber installation. We focused on carefully de-
signing the link architecture, the adapted optical components and modulation formats. First,
We reported 5 Tb/s bidirectional transmission over 2.2 km of OM2 fiber using selective excita-
148 Chapter 4. Transmission Experiments

tion of 4 mode groups and 20 wavelength per mode group. This result has been obtained using
direct detection and DMT format. In fact, this result demonstrated the potential of increase
capacity over a single fiber using mode group multiplexing and WDM. However, its implemen-
tation in real world use cases remains challenging due to the effective cost of deployed devices.
Therefore, we investigated the advantages of mode group multiplexing using low cost EMLs
designed by III-V lab. We reported a net bit rate greater than 250 Gb/s over 2.2km and 200
Gb/s over 4.4km of OM2 fiber using 4 modes groups and single wavelength. These throughputs
represent more than 4 times the current deployed solution based on mode group multiplexing
in LAN ( 4 × 10 Gb/s). Furthermore, it could be a good candidate for scaling capacity in
intra-datacenter while maintaining a low number of deployed fibers. In fact, implementation
of CWDM4 based on 4 wavelength multiplexing using EMLs could increase the capacity to 1
Tb/s over 2.2 km and 800 Gb/s over 4.4 km. However deployment of mode group multiplexing
is challenging for the short term due to the presence of fiber optic patch panels which will
represent additional sources of crosstalk.
Furtheremore, we reported a capacity record of 14.5 Tb/s net throughput in bidirectional
transmission over 2.2 km of conventional OM2 fiber using DMT modulation scheme and direct
detection. We achieved this result by using selective excitation of 4 mode groups and WDM
multiplexing of 40 wavelengths for each mode group using a 100 GHz grid multimode WDM
demultiplexer. In this experiment, we achieved single wavelength net throughput of 450 Gb/s.
This capacity is ten times higher than current products’ capacity in LAN which is 4 × 10 Gb/s
SFP ( serial speed). The latter made mode group multiplexing a scalable technology for meeting
future requirements in terms of high data rates. For example, factories, companies, universities
and hospitals will benefit from high speed networks to implement bandwidth hungry applica-
tions such as smart sensors, internet of things, augmented reality and advanced robotics.
The second part of chapter 4 focused on extending the IMDD mode group multiplexing tech-
niques to longer reach applications up to 20 km using FMFs instead of OM2 fiber. As mentioned
at the beginning of this summary, optical network architecture will move from mega datacen-
ters to distributed cloud processing centers close to the users to respect sub-millisecond latency
requirement. Hence, high speed short reach systems are required to connect these distributed
4.5. Summary 149

datacenters. Therefore, we investigated mode group multiplexing for meeting this need. To
achieve longer distances, we reduced system crosstalk by using few mode fiber as a transmis-
sion medium instead of OM2 fiber. Hence, the mode group (de)multiplexers design handles
lower number of modes which reduces the system complexity and crosstalk. We reported single
wavelength 200 Gb/s net bitrate over 20 km of weakly coupled 6-LP modes FMF designed
by Prysmian. In this experiment, we were mainly limited by power fading that arising from
chromatic dispersion. Operating around 1.3 µm may enables to reduce the impact of chromatic
dispersion and reach higher distances. However, fiber attenuation should be compensated using
EDFAs or photoreceivers with improved sensitivity should be used.
The last part of chapter 4 reported an experimental comparison in terms of crosstalk robustness
for 4-PAM and DMT formats. This investigation has been driven by many observations and
measurements where we noticed the vulnerability of 4-PAM to modal crosstalk compared to
DMT format. Hence, we experimentally investigated the impact of crosstalk phase fluctuation
on BER fluctuation for both 4-PAM and DMT formats. Our investigation confirmed that DMT
is more resilient to crosstalk and thus it could be the best candidate for SDM application using
IMDD systems over short distances.
Finally, we believe that mode group multiplexing could have a promising potential for future
LAN applications. However, its implementation for intra-datacenters as well as inter-datacenter
remains challenging. In fact, coherent transceivers may be a strong competitors for IMDD sys-
tems due to their low cost. Furthermore, light coherent transceivers will reduce the complexity
of classical coherent transceivers namely in terms of managed chromatic dispersion.
List of Publications

First Author Journal Publications

1. K. Benyahya, C. Simonneau, A. Ghazisaeidi, P. Jian, J.F. Morizur, G. Labroille, M.


Bigot, P. Sillard, J. Renaudier, and G. Charlet, ”High-Speed Bi-Directional Transmission
Over Multimode Fiber Link in IM/DD Systems,” in Journal of Lightwave Technology,
vol. 36, no. 18, pp. 4174-4180, 15 Sept.15, 2018. (Invited paper)

2. K. Benyahya, C. Simonneau, A. Ghazisaeidi, N. Barré, P. Jian, J.F. Morizur, G.


Labroille, M. Bigot, P. Sillard, J. Provost, H. Debrégeas, J. Renaudier, and G. Charlet,
”Multiterabit Transmission Over OM2 Multimode Fiber With Wavelength and Mode
Group Multiplexing and Direct Detection,” in Journal of Lightwave Technology, vol. 36,
no. 2, pp. 355-360, 15 Jan.15, 2018. (Invited paper)

Co-authored Journal Publications

1. J.M. Estarán, S. Almonacil, R. Müller, H. Mardoyan, P. Jenneve, K. Benyahya, C.


Simonneau, S. Bigo, J. Renaudier, and G. Charlet,”Sub-Baudrate Sampling at DAC and
ADC: Toward 200G per Lane IM/DD Systems,” in Journal of Lightwave Technology, vol.
37, no. 6, pp. 1536-1542, 15 March15, 2019. (Invited paper)

150
4.5. Summary 151

First Author Conference Publications

1. K. Benyahya, C. Simonneau, A. Ghazisaeidi, N. Barré, P. Jian, J. Morizur, G. Labroille,


P. Sillard, J. Renaudier, and G. Charlet, ”5Tb/s transmission over 2.2 km of multimode
OM2 fiber with direct detection thanks to wavelength and mode group multiplexing,” in
Optical Fiber Communication Conference, OSA Technical Digest (online) (Optical Society
of America, 2017), paper M2D.2. (Finalist for the best student paper award)

2. K. Benyahya, C. Simonneau, A. Ghazisaeidi, N. Barré, P. Jian, J.F. Morizur, Guillaume


Labroille, Marianne Bigot, Pierre Sillard, Jérémie Renaudier and Gabriel Charlet,”14.5Tb/s
Mode-Group and Wavelength Multiplexed Direct Detection Transmission over 2.2 km
OM2 Fiber,” 2017 European Conference on Optical Communication (ECOC), Gothen-
burg, 2017, pp. 1-3. (Top scored and best student paper award)

3. K. Benyahya, C. Simonneau, A. Ghazisaeidi, R. Muller, M. Bigot, P. Sillard, P. Jian,


G. Labroille, J. Renaudier, and G. Charlet, ”200Gb/s Transmission Over 20km of FMF
Fiber Using Mode Group Multiplexing and Direct Detection,” 2018 European Conference
on Optical Communication (ECOC), Rome, 2018, pp. 1-3.

4. K. Benyahya, A. Ghazisaeidi and C. Simonneau, ”Mode Group Division Multiplexed


Transmission Over Multimode Fiber,” 2019 24th OptoElectronics and Communications
Conference (OECC) and 2019 International Conference on Photonics in Switching and
Computing (PSC), Fukuoka, Japan, 2019, pp. 1-3. (Invited paper)

5. K. Benyahya, A. Ghazisaeidi, C. Simonneau, G. Labroille, A. billaud, M. Bigot, P.


Sillard and J. Renaudier, ”Statistical characterization of intermodal crosstalk in direct
detection schemes: M-PAM versus DMT,” 2018 European Conference on Optical Com-
munication (ECOC), Dublin, 2019.
152 Chapter 4. Transmission Experiments

Co-authored Conference Publications

1. M. Trajkovic, K. Benyahya, C. Simonneau, F. Blache, H, Debregeas, J. Provost, K. A


Williams and X. Leijtens,”112 Gb/s PAM-4 Transmission over 1.5 km with an EAM in
Generic Integration Platform,” 2019 24th OptoElectronics and Communications Confer-
ence (OECC) and 2019 International Conference on Photonics in Switching and Com-
puting (PSC), Fukuoka, Japan, 2019, pp. 1-3.

2. M. Bigot-Astruc, J. B. Trinel, H. Maerten, M. van Stralen, I. Milicevic, L. Bigot, S.


Plus, A. Masselot, R. Habert, C. Simonneau, K. Benyahya, G. Labroille, and P. Sillard,
”Weakly-Coupled 6-LP-Mode Fiber with Low Differential Mode Attenuation,” in Optical
Fiber Communication Conference (OFC) 2019, OSA Technical Digest (Optical Society of
America, 2019), paper M1E.3.

3. J.M. Estaran, R. Müller, K. Benyahya, P. Jenneve, C. Simonneau, S. Bigo, J. Renaudier,


G. Charlet, ”Sub-Baudrate Sampling at DAC and ADC for Next Generation 200G Sys-
tems,” 2018 European Conference on Optical Communication (ECOC), Rome, 2018, pp.
1-3.
Bibliography

[1] M. Chagnon, “Optical communications for short reach,” in 2018 European Conference on
Optical Communication (ECOC), Sep. 2018, pp. 1–3.

[2] [Online]. Available: https://ethernetalliance.org

[3] G. P. Agrawal, “Optical communication: its history and recent progress,” in Optics in
Our Time. Springer, Cham, 2016, pp. 177–199.

[4] L. Lukasiak and A. Jakubowski, “History of semiconductors,” Journal of Telecommuni-


cations and information technology, pp. 3–9, 2010.

[5] K. C. Kao and G. A. Hockham, “Dielectric-Fibre Surface Waveguides for Optical Frequen-
cies,” Proceedings of the Institution of Electrical Engineers, vol. 113, no. 7, pp. 1151–1158,
July 1966.

[6] C. K. Kao, “Nobel lecture: Sand from centuries past: Send future voices fast,” Reviews
of Modern Physics, vol. 82, no. 3, p. 2299, 2010.

[7] R. D. Maurer and P. C. Schultz, “Fused silica optical waveguide,” May 2 1972, uS Patent
3,659,915.

[8] S. E. Miller, E. A. Marcatili, and T. Li, “Research toward optical-fiber transmission


systems,” Proceedings of the IEEE, vol. 61, no. 12, pp. 1703–1704, 1973.

[9] T. Miya, Y. Terunuma, T. Hosaka, and T. Miyashita, “Ultimate low-loss single-mode


fibre at 1.55 m,” Electronics Letters, vol. 15, no. 4, pp. 106–108, February 1979.

153
154 BIBLIOGRAPHY

[10] A. Singh, J. Ong, A. Agarwal, G. Anderson, A. Armistead, R. Bannon, S. Boving, G. De-


sai, B. Felderman, P. Germano et al., “Jupiter rising: A decade of clos topologies and
centralized control in google’s datacenter network,” ACM SIGCOMM computer commu-
nication review, vol. 45, no. 4, pp. 183–197, 2015.

[11] J. Lavrencik, S. Varughese, V. A. Thomas, and S. E. Ralph, “Scaling vcsel-mmf links to


1 tb/s using short wavelength division multiplexing,” Journal of Lightwave Technology,
vol. 36, no. 18, pp. 4138–4145, Sep. 2018.

[12] D. M. Kuchta, “High capacity vcsel-based links,” in 2017 Optical Fiber Communications
Conference and Exhibition (OFC), March 2017, pp. 1–94.

[13] P. J. Winzer, D. T. Neilson, and A. R. Chraplyvy, “Fiber-optic transmission and net-


working: the previous 20 and the next 20 years,” Optics express, vol. 26, no. 18, pp.
24 190–24 239, 2018.

[14] K. Singh and G. Kaur, “Multi-core fibers: An overview,” 12 2013.

[15] D. Soma, Y. Wakayama, S. Beppu, S. Sumita, T. Tsuritani, T. Hayashi, T. Nagashima,


M. Suzuki, H. Takahashi, K. Igarashi, I. Morita, and M. Suzuki, “10.16 peta-bit/s dense
sdm/wdm transmission over low-dmd 6-mode 19-core fibre across c+l band,” in 2017
European Conference on Optical Communication (ECOC), Sep. 2017, pp. 1–3.

[16] R. Ryf and N. K. Fontaine, “Space-division multiplexing and mimo processing,” Wiley
series in microwave and optical engineering, 2016.

[17] C. Yang, M. Luo, and X. Li, “Direct detection of 1.176-tb/s (168-gb/s× 7) nyquist-
pam-8 over 7-core fiber using digital pre-distortion,” in 2018 Asia Communications and
Photonics Conference (ACP). IEEE, 2018, pp. 1–4.

[18] L. Zhang, J. Van Kerrebrouck, O. Ozolins, R. Lin, X. Pang, A. Udalcovs, S. Spiga, M. C.


Amann, L. Gan, M. Tang, S. Fu, R. Schatz, G. Jacobsen, S. Popov, D. Liu, W. Tong,
G. Torfs, J. Bauwelinck, X. Yin, S. Xiao, and J. Chen, “Experimental demonstration of
BIBLIOGRAPHY 155

503.61-gbit/s dmt over 10-km 7-core fiber with 1.5 − µm sm-vcsel for optical intercon-
nects,” in 2018 European Conference on Optical Communication (ECOC), Sep. 2018, pp.
1–3.

[19] J. Bland-Hawthorn and P. Kern, “Molding the flow of light: Photonics in astronomy,”
Physics Today, vol. 65, no. 5, p. 31, 2012.

[20] X. Sai, Y. Li, X. Zeng, L. Feng, W. Li, J. Qiu, X. Hong, Y. Zuo, H. Guo, and J. Wu,
“Elliptical-core mode-selective photonic lanterns for mimo-free mode division multiplex-
ing systems,” in 2017 Conference on Lasers and Electro-Optics (CLEO), May 2017, pp.
1–2.

[21] H. Liu, H. Wen, J. C. A. Zacarias, J. E. Antonio-Lopez, N. Wang, P. Sillard, A. A. Cor-


rea, R. Amezcua-Correa, and G. Li, “3× 10 gb/s mode group-multiplexed transmission
over a 20 km few-mode fiber using photonic lanterns,” in Optical Fiber Communication
Conference. Optical Society of America, 2017, pp. M2D–5.

[22] S. H. Murshid, Optical Fiber Multiplexing and Emerging Techniques: SDM and OAM.
Morgan & Claypool Publishers, 2018.

[23] N. Bozinovic, Y. Yue, Y. Ren, M. Tur, P. Kristensen, H. Huang, A. E. Willner, and


S. Ramachandran, “Terabit-scale orbital angular momentum mode division multiplexing
in fibers,” science, vol. 340, no. 6140, pp. 1545–1548, 2013.

[24] H. Huang, Y. Ren, G. Xie, Y. Yan, Y. Yue, N. Ahmed, M. P. Lavery, M. J. Padgett,


S. Dolinar, M. Tur et al., “Tunable orbital angular momentum mode filter based on
optical geometric transformation,” Optics letters, vol. 39, no. 6, pp. 1689–1692, 2014.

[25] J. Zhang, G. Zhu, J. Liu, X. Wu, J. Zhu, C. Du, W. Luo, Y. Chen, and S. Yu, “Orbital-
angular-momentum mode-group multiplexed transmission over a graded-index ring-core
fiber based on receive diversity and maximal ratio combining,” Optics express, vol. 26,
no. 4, pp. 4243–4257, 2018.

[26] K. Yang, Y. ge Liu, Z. Wang, H. wei Zhang, Y. Han, B. wei Mao, and R. jing He,
“All-fiber orbital angular momentum laser generated with titled fiber bragg grating pair
156 BIBLIOGRAPHY

written in few-mode ring-core fiber,” in Optical Fiber Communication Conference (OFC)


2019. Optical Society of America, 2019, p. W3C.4.

[27] T. Alpert, F. Lang, D. Ferenci, M. Grözing, and M. Berroth, “A 28gs/s 6b pseudo


segmented current steering dac in 90nm cmos,” in 2011 IEEE MTT-S International Mi-
crowave Symposium, June 2011, pp. 1–4.

[28] J. Savoj, A. Abbasfar, A. Amirkhany, M. Jeeradit, and B. W. Garlepp, “A 12-gs/s


phase-calibrated cmos digital-to-analog converter for backplane communications,” IEEE
Journal of Solid-State Circuits, vol. 43, no. 5, pp. 1207–1216, May 2008.

[29] S. L. Chuang, Physics of photonic devices. John Wiley & Sons, 2012, vol. 80.

[30] L. A. Coldren, S. W. Corzine, and M. L. Mashanovitch, Diode lasers and photonic inte-
grated circuits. John Wiley & Sons, 2012, vol. 218.

[31] T. L. Koch and J. E. Bowers, “Nature of wavelength chirping in directly modulated


semiconductor lasers,” Electronics Letters, vol. 20, no. 25, pp. 1038–1040, December
1984.

[32] T. L. Koch and R. Linke, “Effect of nonlinear gain reduction on semiconductor laser
wavelength chirping,” Applied Physics Letters, vol. 48, no. 10, pp. 613–615, 1986.

[33] R. N. Hall, G. E. Fenner, J. Kingsley, T. Soltys, and R. Carlson, “Coherent light emission
from gaas junctions,” Physical Review Letters, vol. 9, no. 9, p. 366, 1962.

[34] M. I. Nathan, W. P. Dumke, G. Burns, F. H. Dill Jr, and G. Lasher, “Stimulated emission
of radiation from gaas p-n junctions,” Applied Physics Letters, vol. 1, no. 3, pp. 62–64,
1962.

[35] H. Kawanishi, Y. Yamauchi, N. Mineo, Y. Shibuya, H. Mural, K. Yamada, and H. Wada,


“Eam-integrated dfb laser modules with more than 40-ghz bandwidth,” IEEE Photonics
Technology Letters, vol. 13, no. 9, pp. 954–956, Sep. 2001.

[36] H. Soda, K.-i. Iga, C. Kitahara, and Y. Suematsu, “Gainasp/inp surface emitting injection
lasers,” Japanese Journal of Applied Physics, vol. 18, no. 12, p. 2329, 1979.
BIBLIOGRAPHY 157

[37] M. S. Lebby, C. A. Gaw, W. Jiang, P. A. Kiely, C. L. Shieh, P. R. Claisse, J. Ramdani,


D. H. Hartman, D. B. Schwartz, and J. Grula, “Use of vcsel arrays for parallel optical
interconnects,” in Fabrication, Testing, and Reliability of Semiconductor Lasers, vol. 2683.
International Society for Optics and Photonics, 1996, pp. 81–91.

[38] K. Iga, F. Koyama, and S. Kinoshita, “Surface emitting semiconductor lasers,” IEEE
Journal of Quantum Electronics, vol. 24, no. 9, pp. 1845–1855, 1988.

[39] K. K. Ng, Complete guide to semiconductor devices. IEEE Press, 2002.

[40] C. Madarash-Hill and J. Hill, “Enhancing access to ieee conference proceedings: a case
study in the application of ieee xplore full text and table of contents enhancements,”
Science & Technology Libraries, vol. 24, no. 3-4, pp. 389–399, 2004.

[41] M. Izutsu, Y. Yamane, and T. Sueta, “Broad-band traveling-wave modulator using a


linbo3optical waveguide,” IEEE Journal of Quantum Electronics, vol. 13, no. 4, pp. 287–
290, April 1977.

[42] F. Leonberger, “High-speed operation of linbo 3 electro-optic interferometric waveguide


modulators,” Optics letters, vol. 5, no. 7, pp. 312–314, 1980.

[43] R. C. Alferness, “Waveguide electrooptic modulators,” IEEE Transactions on Microwave


Theory Techniques, vol. 30, pp. 1121–1137, 1982.

[44] T. Muoi, “Receiver design for high-speed optical-fiber systems,” Journal of Lightwave
Technology, vol. 2, no. 3, pp. 243–267, June 1984.

[45] K. Kato, “Ultrawide-band/high-frequency photodetectors,” IEEE Transactions on Mi-


crowave Theory and Techniques, vol. 47, no. 7, pp. 1265–1281, July 1999.

[46] J. Bowers and C. Burrus, “Ultrawide-band long-wavelength p-i-n photodetectors,” Jour-


nal of Lightwave Technology, vol. 5, no. 10, pp. 1339–1350, October 1987.

[47] S. Nishihara, M. Nakamura, K. Nishimura, K. Kishine, S. Kimura, and K. Kato, “A fast-


response and high-sensitivity pin-tia module with wide dynamic range for 10g burst-mode
158 BIBLIOGRAPHY

transmissions,” in 33rd European Conference and Exhibition of Optical Communication,


Sep. 2007, pp. 1–2.

[48] M. S. John G Proakis, Digital communications. McGraw-hill, 2001.

[49] S. Gorshe, “Beyond 100g otn interface standardization,” in 2017 Optical Fiber Commu-
nications Conference and Exhibition (OFC), March 2017, pp. 1–62.

[50] J. Huo, X. Zhou, K. P. Zhong, J. Tu, J. Yuan, C. Guo, K. Long, C. Yu, A. P. T. Lau, and
C. Lu, “Transmitter and receiver dsp for 112 gbit/s pam-4 amplifier-less transmissions
using 25g-class eml and apd,” Opt. Express, vol. 26, no. 18, pp. 22 673–22 686, Sep 2018.

[51] N. Eiselt, D. Muench, A. Dochhan, H. Griesser, M. Eiselt, J. J. V. Olmos, I. T. Monroy,


and J. Elbers, “Performance comparison of 112-gb/s dmt, nyquist pam4, and partial-
response pam4 for future 5g ethernet-based fronthaul architecture,” Journal of Lightwave
Technology, vol. 36, no. 10, pp. 1807–1814, May 2018.

[52] I. Union, “Itu-t recommendation g. 975.1,” 2004.

[53] I. S. Reed and G. Solomon, “Polynomial codes over certain finite fields,” Journal of the
society for industrial and applied mathematics, vol. 8, no. 2, pp. 300–304, 1960.

[54] B. P. Smith, A. Farhood, A. Hunt, F. R. Kschischang, and J. Lodge, “Staircase codes:


Fec for 100 gb/s otn,” Journal of Lightwave Technology, vol. 30, no. 1, pp. 110–117, 2011.

[55] J. M. Cioffi, “A multicarrier primer,” ANSI T1E1, vol. 4, pp. 91–157, 1991.

[56] J. A. Bingham et al., “Multicarrier modulation for data transmission: An idea whose
time has come,” IEEE Communications magazine, vol. 28, no. 5, pp. 5–14, 1990.

[57] M. L. Doelz, E. T. Heald, and D. L. Martin, “Binary data transmission techniques for
linear systems,” Proceedings of the IRE, vol. 45, no. 5, pp. 656–661, May 1957.

[58] [Online]. Available: https://web.stanford.edu/group/cioffi/book/

[59] J. Lee, F. Breyer, S. Randel, J. Zeng, F. Huijskens, H. Van den Boom, A. Koonen, and
N. Hanik, “24-gb/s transmission over 730 m of multimode fiber by direct modulation of
BIBLIOGRAPHY 159

an 850-nm vcsel using discrete multi-tone modulation,” in National Fiber Optic Engineers
Conference. Optical Society of America, 2007, p. PDP6.

[60] J. Lee, “Discrete multitone modulation for short-range optical communications,” Eind-
hoven: Technische Universiteit Eindhoven, 2009.

[61] W. Shieh, X. Yi, Y. Ma, and Y. Tang, “Theoretical and experimental study on pmd-
supported transmission using polarization diversity in coherent optical ofdm systems,”
Opt. Express, vol. 15, no. 16, pp. 9936–9947, Aug 2007.

[62] T. Jiang, D. Chen, C. Ni, and D. Qu, “Chapter 5 - peak-to-average power


ratio,” in OQAM/FBMC for Future Wireless Communications, T. Jiang, D. Chen,
C. Ni, and D. Qu, Eds. Academic Press, 2018, pp. 95 – 129. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/B978012813557000005X

[63] D. Lim, S. Heo, and J. No, “An overview of peak-to-average power ratio reduction schemes
for ofdm signals,” Journal of Communications and Networks, vol. 11, no. 3, pp. 229–239,
June 2009.

[64] C. Wang, S. Ku, and C. Yang, “A low-complexity papr estimation scheme for ofdm signals
and its application to slm-based papr reduction,” IEEE Journal of Selected Topics in
Signal Processing, vol. 4, no. 3, pp. 637–645, June 2010.

[65] B. Koussa, S. Bachir, C. Perrine, C. Duvanaud, and R. Vauzelle, “A comparison of several


gradient based optimization algorithms for papr reduction in ofdm systems,” in CCCA12,
Dec 2012, pp. 1–6.

[66] R. Yoshizawa and H. Ochiai, “Trellis-assisted constellation subset selection for papr re-
duction of ofdm signals,” IEEE Transactions on Vehicular Technology, vol. 66, no. 3, pp.
2183–2198, March 2017.

[67] J. Zhong, X. Yang, and W. Hu, “Performance-improved secure ofdm transmission us-
ing chaotic active constellation extension,” IEEE Photonics Technology Letters, vol. 29,
no. 12, pp. 991–994, June 2017.
160 BIBLIOGRAPHY

[68] F. Chang, Datacenter Connectivity Technologies: Principles and Practice. River pub-
lishers, 2018.

[69] M. Speth, S. A. Fechtel, G. Fock, and H. Meyr, “Optimum receiver design for wireless
broad-band systems using ofdm. i,” IEEE Transactions on Communications, vol. 47,
no. 11, pp. 1668–1677, Nov 1999.

[70] T. M. Schmidl and D. C. Cox, “Robust frequency and timing synchronization for ofdm,”
IEEE Transactions on Communications, vol. 45, no. 12, pp. 1613–1621, Dec 1997.

[71] Luo Zhi-nian, Zheng Jin-jin, Zhu Jiang, and Zhang Er-yang, “Golay complementary pair
aided time synchronization method for ofdm systems,” in 2012 IEEE 14th International
Conference on Communication Technology, Nov 2012, pp. 166–170.

[72] Hlaing Minn, V. K. Bhargava, and K. B. Letaief, “A robust timing and frequency syn-
chronization for ofdm systems,” IEEE Transactions on Wireless Communications, vol. 2,
no. 4, pp. 822–839, July 2003.

[73] P. S. Chow, J. M. Cioffi, and J. A. C. Bingham, “A practical discrete multitone transceiver


loading algorithm for data transmission over spectrally shaped channels,” IEEE Trans-
actions on Communications, vol. 43, no. 2/3/4, pp. 773–775, Feb 1995.

[74] K. Benyahya, C. Simonneau, A. Ghazisaeidi, P. Jian, J. Morizur, G. Labroille, M. Bigot,


P. Sillard, J. Renaudier, and G. Charlet, “High-speed bi-directional transmission over
multimode fiber link in im/dd systems,” Journal of Lightwave Technology, vol. 36, no. 18,
pp. 4174–4180, Sep. 2018.

[75] B. E. Saleh and M. C. Teich, Fundamentals of photonics. John Wiley & Sons, 2019.

[76] S. Addanki, I. Amiri, and P. Yupapin, “Review of optical fibers-introduction and appli-
cations in fiber lasers,” Results in Physics, vol. 10, pp. 743–750, 2018.

[77] R. Lingle, Jr, D. Peckham, A. McCurdy, and J. Kim, Light-Guiding Fundamentals and
Fiber Design, 12 2007, pp. 19–68.
BIBLIOGRAPHY 161

[78] D. Large and J. Farmer, “Chapter 4 - Linear Fiber-Optic Signal Transportation,” in


Broadband Cable Access Networks, ser. The Morgan Kaufmann Series in Networking,
D. Large and J. Farmer, Eds. Boston: Morgan Kaufmann, 2009, pp. 81 – 126. [Online].
Available: http://www.sciencedirect.com/science/article/pii/B9780123744012000048

[79] T. A. Birks, D. Mogilevtsev, J. C. Knight, and P. St. J. Russell, “Dispersion compensation


using single-material fibers,” IEEE Photonics Technology Letters, vol. 11, no. 6, pp. 674–
676, June 1999.

[80] R. D. Wesel and J. M. Cioffi, “Achievable rates for tomlinson-harashima precoding,”


IEEE Transactions on Information Theory, vol. 44, no. 2, pp. 824–831, March 1998.

[81] R. Rath and W. Rosenkranz, “Tomlinson-harashima precoding for fiber-optic communi-


cation systems,” in 39th European Conference and Exhibition on Optical Communication
(ECOC 2013), Sep. 2013, pp. 1–3.

[82] K. Petermann, “Simple relationship between differential mode delay in optical fibres and
the deviation from optimum profile,” Electronics Letters, vol. 14, no. 24, pp. 793–794,
November 1978.

[83] T. I. Association et al., “Fotp-220 differential mode delay measurement of multimode fiber
in the time domain,” Standard document TIA-455-220-A (Telecommunication Industry
Association, 2003), 2003.

[84] D. Richardson, J. Fini, and L. E. Nelson, “Space-division multiplexing in optical fibres,”


Nature Photonics, vol. 7, no. 5, p. 354, 2013.

[85] M. H. Weik, “space-division multiplexing,” in Computer Science and Communications


Dictionary. Boston, MA: Springer US, 2001, pp. 1623–1623. [Online]. Available:
https://doi.org/10.1007/1-4020-0613-61 7780

[86] K.-P. Ho and J. M. Kahn, “Chapter 11 - Mode Coupling and its Impact on
Spatially Multiplexed Systems,” in Optical Fiber Telecommunications (Sixth Edition),
sixth edition ed., ser. Optics and Photonics, I. P. Kaminow, T. Li, and A. E.
162 BIBLIOGRAPHY

Willner, Eds. Boston: Academic Press, 2013, pp. 491 – 568. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/B9780123969606000110

[87] R. Ryf, M. A. Mestre, A. Gnauck, S. Randel, C. Schmidt, R. Essiambre, P. Winzer, R. Delbue,


P. Pupalaikis, A. Sureka et al., “Low-loss mode coupler for mode-multiplexed transmission in
few-mode fiber,” in National Fiber Optic Engineers Conference. Optical Society of America,
2012, pp. PDP5B–5.

[88] R. Olshansky, “Mode coupling effects in graded-index optical fibers,” Applied optics, vol. 14,
no. 4, pp. 935–945, 1975.

[89] R. Ryf, S. Randel, A. H. Gnauck, C. Bolle, A. Sierra, S. Mumtaz, M. Esmaeelpour, E. C.


Burrows, R. Essiambre, P. J. Winzer, D. W. Peckham, A. H. McCurdy, and R. Lingle, “Mode-
division multiplexing over 96 km of few-mode fiber using coherent 6 × 6 mimo processing,”
Journal of Lightwave Technology, vol. 30, no. 4, pp. 521–531, Feb 2012.

[90] D. Marcuse, Theory of dielectric optical waveguides. Elsevier, 2013.

[91] K. Ho and J. M. Kahn, “Frequency diversity in mode-division multiplexing systems,” Journal


of Lightwave Technology, vol. 29, no. 24, pp. 3719–3726, Dec 2011.

[92] D. Gloge, “Optical power flow in multimode fibers,” The Bell System Technical Journal, vol. 51,
no. 8, pp. 1767–1783, Oct 1972.

[93] D. Marcuse, “Losses and impulse response of a parabolic index fiber with random bends,” The
Bell System Technical Journal, vol. 52, no. 8, pp. 1423–1437, Oct 1973.

[94] P. Genevaux, C. Simonneau, G. Le Cocq, Y. Quiquempois, L. Bigot, A. Boutin, and G. Charlet,


“A five-mode erbium-doped fiber amplifier for mode-division multiplexing transmission,” Jour-
nal of Lightwave Technology, vol. 34, no. 2, pp. 456–462, 2015.

[95] A. Gulistan, S. Ghosh, and B. Rahman, “Enhancement of modal stability through reduced
mode coupling in a few-mode fiber for mode division multiplexing,” OSA Continuum, vol. 1,
no. 2, pp. 309–319, 2018.
BIBLIOGRAPHY 163

[96] S. Chebaane, H. Fathallah, H. Seleem, and M. Machhout, “Trenched raised cosine fmf for differ-
ential mode delay management in next generation optical networks,” Optics Communications,
vol. 408, pp. 15–20, 2018.

[97] P. Sillard, M. Bigot-Astruc, and D. Molin, “Few-mode fibers for mode-division-multiplexed


systems,” Journal of Lightwave Technology, vol. 32, no. 16, pp. 2824–2829, Aug 2014.

[98] M. Bigot-Astruc, J. Trinel, H. Maerten, M. van Stralen, I. Milicevic, L. Bigot, S. Plus, A. Mas-
selot, R. Habert, C. Simonneau et al., “Weakly-coupled 6-lp-mode fiber with low differential
mode attenuation,” in Optical Fiber Communication Conference. Optical Society of America,
2019, pp. M1E–3.

[99] S. Jiang, L. Ma, Z. Zhang, X. Xu, S. Wang, J. Du, C. Yang, W. Tong, and Z. He, “Design and
characterization of ring-assisted few-mode fibers for weakly coupled mode-division multiplexing
transmission,” Journal of Lightwave Technology, vol. 36, no. 23, pp. 5547–5555, Dec 2018.

[100] J. H. Chang, S. Bae, H. Kim, and Y. C. Chung, “Heterogeneous 12-core 4-lp-mode fiber based
on trench-assisted graded-index profile,” IEEE Photonics Journal, vol. 9, no. 2, pp. 1–10, 2017.

[101] S. Chebaane, H. Seleem, H. Fathallah, and M. Machhout, “Design tradeoffs of few-mode step
index fiber for next generation mode division multiplexing optical networks,” in 2015 Inter-
national Conference on Information and Communication Technology Research (ICTRC), May
2015, pp. 262–265.

[102]

[103] G. M. Saridis, D. Alexandropoulos, G. Zervas, and D. Simeonidou, “Survey and evaluation of


space division multiplexing: From technologies to optical networks,” IEEE Communications
Surveys Tutorials, vol. 17, no. 4, pp. 2136–2156, Fourthquarter 2015.

[104] L. Raddatz, I. White, D. Cunningham, and M. Nowell, “An experimental and theoretical study
of the offset launch technique for the enhancement of the bandwidth of multimode fiber links,”
Journal of Lightwave Technology, vol. 16, no. 3, p. 324, 1998.
164 BIBLIOGRAPHY

[105] D. H. Sim, Y. Takushima, and Y. C. Chung, “High-speed multimode fiber transmission by using
mode-field matched center-launching technique,” Journal of Lightwave Technology, vol. 27,
no. 8, pp. 1018–1026, April 2009.

[106] C. Pinho, I. Alimi, M. Lima, P. Monteiro, and A. Teixeira, “Spatial light modulation as a
flexible platform for optical systems,” in Telecommunication Systems. IntechOpen, 2019.

[107] G. Stepniak, L. Maksymiuk, and J. Siuzdak, “Binary-phase spatial light filters for mode-
selective excitation of multimode fibers,” Journal of Lightwave Technology, vol. 29, no. 13,
pp. 1980–1987, July 2011.

[108] K. Igarashi, D. Souma, K. Takeshima, and T. Tsuritani, “Selective mode multiplexer based on
phase plates and mach-zehnder interferometer with image inversion function,” Optics express,
vol. 23, no. 1, pp. 183–194, 2015.

[109] G. A. Cirino, R. D. Mansano, P. Verdonck, L. Cescato, and L. G. Neto, “Diffractive phase-shift


lithography photomask operating in proximity printing mode,” Optics express, vol. 18, no. 16,
pp. 16 387–16 405, 2010.

[110] A. Bañas and J. Glückstad, “Light shaping with holography, gpc and holo-gpc,” Optical Data
Processing and Storage, vol. 3, no. 1, pp. 20–40, 2017.

[111] J.-F. Morizur, L. Nicholls, P. Jian, S. Armstrong, N. Treps, B. Hage, M. Hsu, W. Bowen,
J. Janousek, and H.-A. Bachor, “Programmable unitary spatial mode manipulation,” JOSA A,
vol. 27, no. 11, pp. 2524–2531, 2010.

[112] J. Morizur, Pu Jian, B. Denolle, O. Pinel, N. Barré, and G. Labroille, “Efficient and mode-
selective spatial multiplexer based on multi-plane light conversion,” in 2015 Optical Fiber Com-
munications Conference and Exhibition (OFC), March 2015, pp. 1–3.

[113] G. Labroille, B. Denolle, P. Jian, P. Genevaux, N. Treps, and J.-F. Morizur, “Efficient and
mode selective spatial mode multiplexer based on multi-plane light conversion,” Optics express,
vol. 22, no. 13, pp. 15 599–15 607, 2014.
BIBLIOGRAPHY 165

[114] G. Labroille, P. Jian, N. Barre, B. Denolle, and J. Morizur, “Mode selective 10-mode multiplexer
based on multi-plane light conversion,” in 2016 Optical Fiber Communications Conference and
Exhibition (OFC), March 2016, pp. 1–3.

[115] K. Lenglé, X. Insou, P. Jian, N. Barré, B. Denolle, L. Bramerie, and G. Labroille, “4× 10 gbit/s
bidirectional transmission over 2 km of conventional graded-index om1 multimode fiber using
mode group division multiplexing,” Optics express, vol. 24, no. 25, pp. 28 594–28 605, 2016.

[116] M. Arikawa, Y. Ono, and T. Ito, “Mode diversity coherent receiver with few-mode fiber-coupling
for high-speed free-space optical communication under atmospheric turbulence,” in Free-Space
Laser Communication and Atmospheric Propagation XXX, vol. 10524. International Society
for Optics and Photonics, 2018, p. 1052412.

[117] G. Labroille, P. Jian, L. Garcia, J. Trinel, R. Kassi, L. Bigot, and J. Morizur, “30 gbit/s
transmission over 1 km of conventional multi-mode fiber using mode group multiplexing with
ook modulation and direct detection,” in 2015 European Conference on Optical Communication
(ECOC), Sep. 2015, pp. 1–3.

[118] P. Genevaux, C. Simonneau, M. Bigot-Astruc, P. Sillard, and G. Charlet, “Real time transmis-
sion of 2 200 gb/s pdm-16qam using two modes over 20km of step-index few mode fibre,” in
2016 Optical Fiber Communications Conference and Exhibition (OFC), March 2016, pp. 1–3.

[119] P. J. Winzer, A. H. Gnauck, A. Konczykowska, F. Jorge, and J. . Dupuy, “Penalties from


in-band crosstalk for advanced optical modulation formats,” in 2011 37th European Conference
and Exhibition on Optical Communication, Sep. 2011, pp. 1–3.

[120] J. Vuong, P. Ramantanis, Y. Frignac, M. Salsi, P. Genevaux, D. F. Bendimerad, and


G. Charlet, “Mode coupling at connectors in mode-division multiplexed transmission over
few-mode fiber,” Opt. Express, vol. 23, no. 2, pp. 1438–1455, Jan 2015. [Online]. Available:
http://www.opticsexpress.org/abstract.cfm?URI=oe-23-2-1438

[121] G. Rademacher, R. S. Luı́s, B. J. Puttnam, Y. Awaji, and N. Wada, “Crosstalk dynamics


in multi-core fibers,” Opt. Express, vol. 25, no. 10, pp. 12 020–12 028, May 2017. [Online].
Available: http://www.opticsexpress.org/abstract.cfm?URI=oe-25-10-12020
166 BIBLIOGRAPHY

[122] M. Mohamed and M. Ab-Rahman, “Analytical analysis of in-band crosstalk, out-of-band


crosstalk and gvd-based power penalties in dwdm and tdm/dwdm-pons,” Journal of Computer
Science, vol. 11, pp. 573–589, 03 2015.

[123] T. M. F. Alves, J. L. Rebola, and A. V. T. Cartaxo, “Outage probability due to intercore


crosstalk in weakly-coupled mcf systems with ook signaling,” in 2019 Optical Fiber Communi-
cations Conference and Exhibition (OFC), March 2019, pp. 1–3.

[124] H. Debrégeas, F. Lelarge, R. Brenot, C. Caillaud, J.-G. Provost, F. Pommereau, T. Nguyen,


D. Lanteri, K. Mekhazni, S. Barbet, A. Garreau, J.-F. Paret, G. Glastre, C. Fortin, E. Der-
ouin, O. Drisse, M. Achouche, J.-L. Gentner, B. Saturnin, F. Martin, Y. Moustapha-Rabault,
F. Blache, M. Goix, B. Duval, and P. Charbonnier, “Record 6dbm electroabsorption modulated
laser for 10gb/s and 25gb/s high power budget access networks,” Optical Fiber Communication
Conference, 2017.
Résumé de Thèse de doctorat

Multiplexage de Groupe de Modes pour les Communications


Optiques à Courte Distance

Les systèmes de communication optiques basés sur la modulation de l’intensité


lumineuse et la détection directe constitué les premières générations de systèmes de
communication par fibre optique. En effet, le premier système a été mis en œuvre en 1976 et
ne permettait de transmettre que 45 Mbit/s sur une distance allant jusqu’à 11 km.
L’implémentation de ce système fut possible grâce à deux grandes innovations. La première est
l’introduction des lasers à semiconducteurs en 1962. La deuxième innovation est la réduction
draconienne de l’atténuation des fibres optiques qui aujourd’hui, est de l’ordre de 0.2 dB par
kilomètre pour une fibre monomode standard. Dans ces premiers systèmes, la puissance optique
était générée en modulant directement un laser à semi-conducteur à l’aide d’un signal électrique
à deux niveaux avant d’être injectée dans la fibre optique. En raison des imperfections de
propagation telles que la dispersion chromatique et la dispersion modale, le signal est dégradé
avant d'être détecté avec un dispositif à semi-conducteur qui convertit la puissance optique en
courant électrique. Ce principe de modulation et détection de l’intensité lumineuse a dominé
pendant les quatre premières générations de systèmes de communication par fibre optique
jusqu’aux années 2000 où la détection cohérente associée au traitement numérique du signal
ont emporté l’intérêt de la communauté scientifique.

Les techniques de multiplexage en longueur d’onde (WDM) renforcées par la détection


cohérente ont permis d’accompagner l’explosion du trafic Internet dans le monde. De plus, les
techniques de transmission les plus avancées s’appuyant sur le traitement du signal avancé,
l’utilisation des convertisseurs numérique-analogique et analogique-numérique à haute
résolution et large bande ont permis de se rapprocher des limites théoriques de capacité que
l’on peut transmettre sur une fibre optique monomode. Néanmoins, ce type de système reste
couteux et complexe pour des transmissions sur courtes distances. Dans ce cas, la technique de
modulation et détection de l’intensité lumineuse est très attractive et fortement demandée du
fait de son faible coût.
En outre, le monde est face à une nouvelle révolution technologique, connue sous le nom de
future industrie 4.0, qui vise à instaurer des innovations liées à l’automatisation des robots, la
réalité augmentée, l’intelligence artificielle, l’internet des objets et autres applications qui
exigent une grande fiabilité du réseau optique, un grand débit pour opérer et un temps de
réponse qui soit très faible, estimé à moins de 1 ms. Pour ce faire, les systèmes de
communication optiques installés dans les réseaux locaux tels que les usines, les entreprises et
les hôpitaux doivent délivrer un grand débit. Cependant, les systèmes actuels ne permettent pas
de transmettre plus de 40 Gb/s sur quelques mètres (moins de 500 m) de fibre multimode
conventionnelle (type OM1-4). De plus, afin de respecter une faible latence pour ces
applications, les centres de données doivent être placés à proximité des utilisateurs, ce qui réduit
la distance de déploiement des fibres et accentue la demande des systèmes de communication
optiques à faible coût pour des courte portées.

Les travaux réalisés dans cette thèse s’inscrivent dans le cadre ’de l’augmentation de la capacité
transmise sur des fibres multimodes, tout en gardant un faible coût grâce à l’utilisation de
techniques de modulation compatibles avec la détection directe de l’intensité lumineuse, afin
d’accompagner la croissance de capacité des réseaux locaux optiques ainsi qu’au sein, ou entre
les centres de données dans la prochaine décennie. Cette thèse vise donc à répondre aux besoins
de deux différents types de marchés, premièrement les réseaux locaux et deuxièmement les
centres de données.

Les réseaux locaux représentent un terrain très attractif pour augmenter le débit de transmission
sur des fibres multimodes. En effet, les fibres multimodes conventionnelles sont déjà installées
dans ces infrastructures, ce qui épargne le coût considérable de l’installation de nouvelles fibres.
Ces fibres multimodes standard de type OM1-4 offrent un avantage d’utilisation de composants
à faible coût. Cependant, la dispersion intermodale limite la portée à des débits de données
élevés dans ces fibres, conduisant à une portée maximum de 80 m à un débit de 10 Gb/s contre
5 km à 100 Mb/s avec une fibre de type OM2 (satisfaisant à la norme IEEE 802.3). Néanmoins,
une compagnie française nommée CAIlabs a mis en œuvre une technique innovante qui a
permis de multiplier le débit transmis sur des fibres multimodes par un facteur 4 et de passer de
10 Gb/s à 40 Gb/s. Leur technique consiste à créer différents canaux de transmission dans une
seule fibre grâce au multiplexage spatial, en l’occurrence 4 canaux transmettant 10 Gb/s chacun.
Cette solution est basée sur le multiplexage de groupes de modes. La fibre multimode standard
(OM2) contient 55 modes à 1550 nm, qui peuvent être répartis en 9 groupes de modes. Chaque
groupe de modes représente un canal de transmission possible dans la fibre multimode. Cette
technique est désormais utilisée dans plusieurs réseaux locaux dans le monde. Toutefois,
l’évolutivité de cette solution vers des débits plus élevés représente un point critique à étudier.
Dans cette thèse, nous nous basons sur le multiplexage de groupe de modes, en collaboration
avec CAIlabs, afin de démontrer des capacités de l’ordre de plusieurs Tb/s sur des distances
inférieures à 5 km en utilisant la détection directe.

Le deuxième champ d’intérêt est constitué par les centres de données (data centers). En effet,
la demande croissante des services Internet entraîne le besoin d'un grand nombre de serveurs
fonctionnant à l'intérieur de ces centres de données, ce qui entraîne une augmentation
considérable de la capacité de transmission nécessaire au sein de ces centres. Par exemple, le
débit utilisé dans les centres de données de Google a été multipliée par mille au cours de la
dernière décennie. De plus, le trafic Internet total des centres de données atteindra les 20,6
zettabit en 2021 et ne cessera de croître. La fraction la plus importante de ce trafic aura lieu à
l'intérieur des centres de données où la majorité des fibres déployées sont des fibres multimodes.
La distance de déploiement de ces fibres est inférieure à 300 m en moyenne. La nouvelle norme
proposée par l’IEEE pour les réseaux opérant à 400 Gb/s implique l’utilisation de 16 fibres
multimodes en parallèle, chacune supportant 25 Gb/s en utilisant la modulation d’intensité à
deux niveaux avec codage non-retour à zéro (NRZ). De ce fait, on remarque que l’évolution
vers des nouveaux standards opérant à des débits de l’ordre du Tb/s est contraignant, vu la
nécessité d’installer un nombre important de fibres optiques. Le multiplexage spatial dans les
fibres multimodes offre la possibilité d’augmenter le débit tout en gardant un nombre réduit de
fibres. Dans ce cadre de cette thèse, l’exploitation de la technique de multiplexage de groupes
de modes nous a permis de répondre à ce besoin. De plus, la transition entre d’énormes centres
de données vers des centres de données de dimensions plus modestes distribués plus proche des
utilisateurs, nous a motivé à étendre nos recherches afin de proposer de nouvelles architectures
de systèmes assurant les communications optiques entre ces centres de données. Dans ce cas,
nous nous somme basés sur le même principe de multiplexage de groupes de modes sur de
nouvelles fibres expérimentales conçues par Prysmian. Ces fibres supportent, à longueur d’onde
donnée, un nombre de modes réduit par rapports aux fibres multimodes conventionnelles, ce
qui permet de réduire le couplage entre les modes et donc d’atteindre des distances plus longues,
de l’ordre d’une dizaine de kilomètres. Dans ce cas, le design des multiplexeurs,
démultiplexeurs, fibres et autres composants sont étudiés.

Dans le premier chapitre de cette thèse, les systèmes de communications optiques sont
introduits, en particulier sous une perspective historique. De plus la problématique qui consiste
à répondre à la demande croissante de ces systèmes pour des courtes distances en offrant des
très grands débits est explicitée dans des cadres concrets. L’état de l’art actuel en termes de
produits utilisés dans les réseaux locaux ainsi que les centres de données est détaillé dans ce
chapitre. Etant donné que l’approche du multiplexage spatial est prometteuse afin d’augmenter
le débit des prochaines générations de réseaux optiques, différents aspects de cette dernière ont
été brièvement introduits dans le cadre des différents travaux de recherches dans ce domaine.

Le chapitre 2 a été axé sur, les principaux aspects d'un système de communication optique basé
sur la modulation de l’intensité lumineuse et la détection directe de cette dernière. D’abord le
principe général de ce type de transmission est détaillé. Par la suite, les différents blocs
constituant un émetteur ainsi qu’un récepteur classique ont été abordés. L’émission des
données peut être assurée soit en modulant l’intensité émise par un laser à semiconducteur
directement à l’aide d’un signal électrique, ou bien en utilisant un modulateur externe. Ces deux
types d’émetteurs représentent des avantages ainsi que des inconvénients vis-à-vis de leur
complexité, consommation énergétique, taille ainsi que des phénomènes qui peuvent être
parfois limitants tels que la dérive en fréquence (« chirp »). En ce qui concerne la réception des
données, différents types de photodétecteurs se distinguent en termes de sensibilité. Ce
chapitre, étudie aussi les formats de modulation numériques adaptés aux systèmes de
communication optiques avec modulation et détection d’intensité. En effet, étant donné que
seule l’intensité du signal est détectée, le signal transmis est à fortiori réel. Dans le cadre de
cette thèse, on utilise deux formats de modulation qui répondent à ce critère. Premièrement, la
modulation mono-porteuse d’amplitude à plusieurs niveaux (PAM) ainsi que ses différents
aspects liés au traitement du signal tel que la mise en forme d’impulsion, la pré-compensation
numérique ainsi que l’égalisation numérique au niveau du récepteur sont introduits. L’impact
du traitement du signal a été expérimentalement démontré dans le cas d’un signal à 4 niveaux
occupant 28 GHz de bande qui résulte en un débit de 112 Gb/s. Dans ce cas, le facteur limitant
est dû à la bande passante étroite du convertisseur numérique-analogique ainsi que du
convertisseur analogique-numérique utilisé dans le récepteur. Deuxièmement, le principe d’un
format de modulation multi-porteuses « DMT » permettant la modulation numérique de
l’information dans le cadre de la détection directe à été expliqué. Ce type de modulation permet
d’adapter le signal à la réponse du canal de transmission ce qui offre un représente un grand
avantage pour les systèmes de transmission optique dont le canal présente une sélectivité en
fréquence, ce qui est le cas d’un canal dispersif à détection directe. Pour ce faire, différents
algorithmes ont été proposés dans la littérature. Dans ce travail, nous appliquons celui proposé
par Chow. Il permet de maximiser le débit transmis pour un taux d’erreur cible. Cet algorithme
est basé sur une approximation appelé « gap approximation », qui a été proposée pour des taux
d’erreur très faibles de l’ordre de 10-6. Nous avons proposé un algorithme alternatif permettant
d’obtenir le même taux d’erreur par sous porteuse quand ce dernier est plus élevé, de l’ordre de
10-3. Cet algorithme a été comparé à celui de Chow à l’aide d’un outil de simulation.

Le troisième chapitre s’est focalisé sur le canal de transmission qui est la fibre optique. Tout
d’abord, nous avons rappelé les équations fondamentales de la propagation de la lumière dans
la fibre, ce qui permet de distinguer diffèrent types de fibre dont la fibre monomode « SMF »,
la fibre multimode « MMF ». La fibre multimode est celle utilisée dans ce travail compte tenu
de la possibilité qu’elle offre d’accroître le débit grâce à l’introduction du multiplexage spatial.
Dans un premier temps nous avons évoqué l’impact de la dispersion chromatique qui, associée
au mécanisme de détection directe quadratique, induit des creux d’évanouissement dans la
réponse en fréquence du canal (incluant la propagation dans la fibre optique et la détection
directe). Les fréquences auxquelles surviennent ces creux d’évanouissement dépendent du
coefficient de dispersion de la fibre et de sa longueur, ainsi que de la longueur d’onde. Par la
suite, la dispersion modale, qui limite la bande passante de transmission sur des fibres
multimodes est introduite. Afin de contourner cette limitation et augmenter le débit sur ces
fibres, le multiplexage spatial peut être appliqué. On peut distinguer deux types de multiplexage
spatial : le multiplexage de modes et le multiplexage de groupes de modes. Le multiplexage de
modes utilise chaque mode individuellement comme un canal de transmission indépendant, ce
qui est rendu théoriquement possible par l’orthogonalité des modes. Néanmoins, cette
indépendance n’est pas vérifiée en pratique à cause des imperfections des (de)multiplexeurs,
des défauts de fabrication de la fibre ainsi que des contraintes supplémentaires que subit la fibre
comme le soudage et les courbures. Ces imperfections créent de la diaphonie entre les modes
qui est plus accentuée entre les modes ayant des constantes de propagation très proches. Le
multiplexage de groupes de modes vise à réduire cette limitation en utilisant l’ensemble des
modes ayant des constantes de propagation très proches comme étant un seul canal de
transmission. Ce type de multiplexage est celui utilisé dans cette thèse. Ce chapitre donne
l’exemple de plusieurs techniques de multiplexage. La technologie des (de)multiplexeurs
utilisés dans nos travaux repose sur la conversion de lumière multi-plans. L’impact de la
diaphonie entre les groupes de modes peut être diffèrent selon le format de modulation que l’on
utilise. Afin d’analyser la sensibilité des différents formats de modulation à cette diaphonie, des
études théoriques ainsi que numérique ont été menées. Nous nous sommes plus particulièrement
intéressés aux modulations de type mono-porteuse à plusieurs niveaux d’amplitude (PAM), et
multi-porteuses (DMT). L’analyse a été conduite dans le cas de diaphonie entre deux signaux,
dont la phase relative est une variable aléatoire distribuée uniformément dans l’intervalle [0
2π]. On a donné une expression analytique du terme représentant l’impact de la diaphonie pour
les deux formats de modulation considérés dans le cas de la détection directe. Ce résultat nous
a permis par la suite de quantifier l’impact de la diaphonie sur le taux d’erreur en la présence
de bruit blanc gaussien. A l’issue de cette investigation, nous avons conclu que les signaux émis
par modulation d’amplitude à plusieurs niveaux sont plus sensibles à la diaphonie par rapport
à ceux émis par modulation multi-porteuse. Cette sensibilité est due à la nature de la modulation
où les niveaux les plus élevés sont plus impactés et donc engendrent plus d’erreurs de
transmission. Cependant, les formats de modulation multi-porteuses ont une distribution
d’intensité gaussienne qui procure une certaine robustesse envers la diaphonie.

Le chapitre 4 est dédié aux études expérimentales et est divisé en trois parties. La première
partie détaille les expérimentations réalisées sur des fibres multimodes conventionnelles pour
des distances inférieures à 5 km. La deuxième partie décrit la transition entre des distances de
transmission inférieure à 5 km et des distances allant jusqu’à 20 km en termes de design des
fibres et des (de)multiplexeurs. La dernière partie compare expérimentalement l’impact de la
diaphonie entre les groupes de modes dans le cas de la modulation d’amplitude à 4 niveaux et
de la modulation multi-porteuse.

Dans la première partie, nous avons utilisé les quatre premiers groupes de modes : le premier
groupe comprend un seul mode tandis que le quatrième groupe comprend 4 modes. Lors du
multiplexage, un mode unique est excité pour chaque mode. Après la propagation dans la fibre
multimode, en raison du couplage important entre les modes appartenant à un même groupe,
tous les modes du groupe doivent être détectés simultanément afin d'éviter de larges fluctuations
de puissance au niveau du récepteur. Le démultiplexeur de modes a donc été conçu pour
démultiplexer sélectivement chaque groupe de mode dans une fibre multimode en sortie. Cette
technique permet de multiplier le débit transmis par 4 (le nombre de groupes de modes
multiplexés). Tout d’abord on a démontré une transmission de 5 Tb/s sur 2,2 km de fibre OM2 .
Ce résultat a été obtenu grâce à la combinaison de la technique de multiplexage de longueur
d’onde (WDM) et celle du multiplexage de groupes de modes. 20 longueurs d’ondes ont été
multiplexées par groupe de modes, chacune transportant 68,8 Gb/s, avec un taux d’erreur de
3,810-3. L’utilisation de la technique WDM a été rendue possible grâce à un démultiplexeur de
longueurs d’onde multimode, qui a permis de conserver la composition modale de chaque
groupe de mode et donc d’éviter de larges fluctuations de puissance au niveau du
photodétecteur. Ce résultat a démontré le potentiel d'augmentation de la capacité sur une seule
fibre en utilisant simultanément le multiplexage de groupes de modes et le multiplexage en
longueur d'onde. Cependant, sa mise en œuvre dans des cas d'utilisation concrets reste difficile
en raison du coût effectif des composants utilisés. Par conséquent, nous avons étudié les
avantages du multiplexage de groupes de modes en utilisant des composants à bas coût tel qu’un
laser à modulateur intégré (EML). Ce dernier a été conçu par III-V Lab. Nous avons comparé
l’utilisation de ce composant à celle d’une source laser continue indépendante suivie d’un
modulateur électro-optique externe employant une structure d’interféromètre de Mach-
Zehnder. La comparaison a été basée sur le débit transmissible par les deux technologies. Nous
avons pu démontrer un débit net supérieur à 250 Gb/s sur 2,2 km et à 200 Gb/s sur 4,4 km de
fibre OM2 en utilisant 4 groupes de modes et une seule longueur d'onde. Ces débits représentent
plus de 4 fois celui de la solution déployée actuellement dans les réseaux locaux qui est de 410
Gb/s. En outre, nous avons également démontré un record de capacité de 14,5 Tb/s de débit net
en transmission bidirectionnelle sur 2,2 km de fibre OM2 en utilisant la modulation multi-
porteuse et la détection directe. Nous avons obtenu ce résultat en utilisant l’excitation sélective
de 4 groupes de modes et le multiplexage WDM de 40 longueurs d’onde espacées de 100 GHz
pour chaque groupe de modes. Dans cette expérience, nous avons atteint un débit net de 450
Gb/s sur une seule longueur d'onde. Ce résultat important nous a permis de démontrer que le
multiplexage de groupes de mode est une technologie évolutive qui est en mesure de répondre
aux futures exigences des systèmes de communications optiques pour des courtes distances.

La deuxième partie du chapitre démontre comment la technique de multiplexage de groupes de


mode en détection directe peut être étendue à des applications avec des distances bien
supérieures à 5 km. Par exemple, cette technologie peut être proposée pour assurer la
communication à haut débit entre les centres de données où la distances est de l’ordre de
quelques dizaines de kilomètres. Il était donc crucial de réduire la diaphonie entre les groupes
de modes afin de pouvoir appliquer le multiplexage de groupes de modes pour des longues
distances. Pour ce faire, nous avons utilisé un nouveau type de fibre ayant un nombre réduit de
modes (12 modes) par rapport à l’OM2 (55 modes). Cette fibre est dite faiblement multimode
(FMF). Par conséquent la complexité des (de)multiplexeurs a été réduite, ce qui a réduit la
diaphonie entre les groupes de modes. Une comparaison en termes de pertes et de diaphonie
entre les systèmes utilisant la fibre OM2 et les nouveaux systèmes avec la fibre FMF est
détaillée dans ce chapitre. Cette configuration nous a permis de démontrer une transmission de
200 Gb/s sur 20 km de FMF en utilisant la modulation multi-porteuse DMT. Cette dernière a
été comparée avec le format de modulation 4-PAM dans le cadre de cette expérience.

La troisième partie du chapitre 4 a permis de confirmer notre étude préliminaire faite au chapitre
2 et qui consistait à comparer les deux formats de modulation mono-porteuse 4-PAM et multi-
porteuse DMT en termes de résilience à la diaphonie. Nous avons donc comparé
expérimentalement des transmissions à 100 Gb/s en utilisant les formats 4-PAM et DMT. Nous
avons en particulier étudié l'impact de la fluctuation de phase de diaphonie sur la fluctuation du
taux d’erreur pour ces deux formats. Les mesures des statistiques de taux d’erreur sur une durée
de 5 heures ont confirmé que le DMT résiste mieux à la diaphonie et qu'il pourrait donc être le
meilleur candidat pour les applications basées sur le multiplexage spatial et la détection directe
pour de courtes distances.

Le multiplexage de groupe de mode appliqué dans le cadre de la détection directe est une
solution attractive pour accompagner le croissement du Traffic dans les réseaux locaux.
Cependant son implémentation pour des application liées aux datacenters reste compliqué à
court terme.