University of Southern California Dissertations and Theses

Elements of next-generation wireless video systems: millimeter-wave and device-to-device algorithms

(USC Thesis Other)

Elements of next-generation wireless video systems: millimeter-wave and device-to-device algorithms

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Elements of Next-Generation Wireless Video Systems:
Millimeter-Wave and Device-to-Device Algorithms
by
Joongheon Kim
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulllment of the
Requirements of the Degree
DOCTOR OF PHILOSOPHY
(Computer Science)
August 2014
Copyright 2014 Joongheon Kim
Dedicated to My Loving Family
ii
Acknowledgements
First of all, I would like to express my deepest gratitude to my Ph.D. research advisor
and mentor Dr. Andreas F. Molisch. Without his guidance, encouragement, patience,
and inspiration, the research for this dissertation never would have taken place. I am also
grateful to Dr. Ramesh Govindan for managing my Ph.D. program in the Department
of Computer Science, and also to Dr. Giuseppe Caire for jointly guiding me in device-to-
device video research. In addition, I am grateful to Dr. Aiichiro Nakano and Dr. Antonio
Ortega for serving on my dissertation committee; and to Dr. Wonjun Lee for guiding my
M.S. research at Korea University.
I am also lucky to have such wonderful friends in the Wireless Devices and Systems
(WiDeS) Group and the Departments of Computer Science and Electrical Engineering:
Dr. Yafei Tian, Dr. Somasundaram Niranjayan, Dr. Junyang Shen, Seun Sangodoyin,
Hao Feng, Zheda Li, Rui Wang, Pei-Lan Hsu, Sundar Aditya, Vinod Kristem, Daoud
Burghal, Celalettin Umit Bas, Dr. Kun Yang, Dr. Ruisi He, Song-Nam Hong, Mingyue
Ji, Dilip Bethanabhotla, Hilmi Enes Egilmez, Vivek Sankaravadivel, Feiyu Meng, Peiyao
Chen, Pyojae Kim, Younghoon Lee, Hanjun Shin, YounKyu Lee, Dr. Ki-Young Jang,
Dr. Moo-Ryong Ra, and Dr. Jeongyeup Paek. They have provided lasting friendship,
enlightenment, encouragement, and entertainment.
In addition, I would like to express my thanks to Disney Research Z urich (especially
to Dr. Stefan Mangold), Intel Corporation (especially to Dr. Ali Sadri, Dr. Carlos
Cordeiro, Dr. Alexander Maltsev, Dr. Liang Xian, Kyle T. McCanta, Reza Are), and LG
iii
Electronics CTO Oce (especially to BeomJin Paul Jeon) for their valuable uncountable
supports. I am also indebted to the USC Annenberg Graduate Fellowship Program for
supporting my research.
Most importantly, I am grateful to my loving family (Sungeun Kim, Kate Seoyoung
Kim, and my son who will be born on June 2014), parents, and relatives for all of the
love and support they have given me over the years. Without their nourishment, I never
would have had the chance to succeed.
iv
List of Publications
Publications that Form Basis of the Dissertation
Magazine and Journal Publications
J. Kim, Y. Tian, S. Mangold, and A.F. Molisch, \Joint Scalable Coding and Routing for 60 GHz
Real-Time Live HD Video Streaming Applications," IEEE Transactions on Broadcasting, vol. 59,
no. 3, pp. 500 - 512, September 2013.
J. Kim and A.F. Molisch, \Fast Millimeter-Wave Beam Training with Receive Beamforming,"
IEEE/KICS Journal of Communications and Networks (Minor Revision, 3rd Round Review).
C. Cordeiro, A. Sadri, A. Maltsev, R. Are, J. Kim, and L. Xian, \Millimeter Wave Capable Small
Cells for Next Generation 5G Cellular Systems," IEEE Communications Magazine (1st Round
Review).
Conference Publications
J. Kim, F. Meng, P. Chen, H.E. Egilmez, D. Bethanabhotla, A.F. Molisch, M.J. Neely, G. Caire,
and A. Ortega, \Demo: Adaptive Video Streaming for Device-to-Device Mobile Platforms," in Pro-
ceedings of the ACM International Conference on Mobile Computing and Networking (MobiCom),
Miami, Florida, USA, 30 September - 4 October 2013.
J. Kim, Y. Tian, S. Mangold, and A.F. Molisch, \Quality-Aware Coding and Relaying for 60 GHz
Real-Time Wireless Video Broadcasting," in Proceedings of the IEEE International Conference on
Communications (ICC), Budapest, Hungary, 9 - 13 June 2013.
v
Publications not Treated in the Dissertation
Conference Publications
J. Kim, L. Xian, A. Maltsev, R. Are, and A. Sadri, \Required Frequency Rejection in 39 GHz
Millimeter-Wave Small Cell Systems: Intel's Preliminary Results," Under Review to be presented
at IEEE Global Communications Conference (GLOBECOM), Austin, Texas, USA, December 2014.
J. Kim and A.F. Molisch, \Quality-Aware Millimeter-Wave Device-to-Device Multi-Hop Routing for
5G Cellular Networks", in Proceedings of the IEEE International Conference on Communications
(ICC), Sydney, Australia, June 2014.
J. Kim and A.F. Molisch, \Enabling Gigabit Services for IEEE 802.11ad-Capable High-Speed Train
Networks, inProceedingsoftheIEEERadioandWirelessSymposium(RWS),APartofIEEERadio
and Wireless Week (RWW), Austin, Texas, USA, January 2013.
J. Kim, Y. Tian, A.F. Molisch, and S. Mangold, \Joint Optimization of HD Video Coding Rates
and Unicast Flow Control for IEEE 802.11ad Relaying," in Proceedings of the IEEE International
Symposium on Personal Indoor and Mobile Radio Communications (PIMRC), Toronto, Canada,
September 2011.
J. Paek, J. Kim, and R. Govindan, \Energy-Ecient Rate-Adaptive GPS-based Positioning for
Smartphones," in Proceedings of the ACM/USENIX International Conference on Mobile Systems,
Applications, and Services (MobiSys), San Francisco, California, USA, June 2010.
vi
Table of Contents
I Introduction 1
I.1 Next-Generation Wireless Video Systems . . . . . . . . . . . . . . . . . . 1
I.2 Millimeter-Wave Systems for High-Rate Data Transmission and Video . . 2
I.3 Device-to-Device Transmission Systems for High-Rate Data Transmission
and Video . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
I.4 Outline of The Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . 11
I.4.1 Millimeter-Wave Small Cell Video Networks . . . . . . . . . . . . . 11
I.4.2 Joint Scalable Coding and Relaying for 60 GHz High-Denition Video
Streaming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
I.4.3 Fast Millimeter-Wave Beam Training with Receive Beamforming . 12
I.4.4 Device-to-Device Adaptive Video Streaming for Distributed Pro-
grammable Wi-Fi Platforms . . . . . . . . . . . . . . . . . . . . . . 13
I.4.5 Joint Scheduling and Stochastic Streaming for Device-to-Device
Video Delivery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
II Millimeter-Wave Small Cell Video Networks 15
II.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
II.2 Small Cell Usages and Key Requirements . . . . . . . . . . . . . . . . . . 17
II.3 Proposed Millimeter-Wave Small Cells . . . . . . . . . . . . . . . . . . . . 18
II.3.1 Functionalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
II.3.2 Considering Frequency Bands . . . . . . . . . . . . . . . . . . . . . 20
II.4 Design Challenges and Considerations . . . . . . . . . . . . . . . . . . . . 21
II.4.1 RF, Antenna, and Physical Layer Design . . . . . . . . . . . . . . 21
II.4.2 Multiple Access and System Design . . . . . . . . . . . . . . . . . 25
II.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
IIIJoint Scalable Coding and Relaying for 60 GHz High-Denition Video
Streaming 29
III.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
III.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
III.3 A Reference System Model . . . . . . . . . . . . . . . . . . . . . . . . . . 34
III.3.1 Link Budget Analysis: Capacity Perspective . . . . . . . . . . . . . 34
III.3.2 Outdoor Broadcasting Systems with 60 GHz Wireless Links . . . . 36
vii
III.4 Joint Scalable Coding and Routing . . . . . . . . . . . . . . . . . . . . . . 40
III.4.1 Single-Beam Antennas at Sources and Relays . . . . . . . . . . . . 41
III.4.2 Single-Beam Antennas at Sources and Multiple-Beam Antennas at
Relays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
III.4.3 Multiple-Beam Antennas at Source and Relays . . . . . . . . . . . 45
III.4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
III.5 Re-Formulation: Convex Form . . . . . . . . . . . . . . . . . . . . . . . . 46
III.6 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
III.6.1 CDF of Aggregate Video Quality { Fixed Number of Relays and
Various Number of Sources . . . . . . . . . . . . . . . . . . . . . . 54
III.6.2 Impact of Lower Bound Setting . . . . . . . . . . . . . . . . . . . . 59
III.7 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
IV Fast Millimeter-Wave Beam Training with Receive Beamforming 63
IV.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
IV.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
IV.2.1 Mm-wave Wireless Packets . . . . . . . . . . . . . . . . . . . . . . 66
IV.2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
IV.3 Interactive Beam Training for Fixed Modulation . . . . . . . . . . . . . . 69
IV.3.1 Interactive 2D Beam Training . . . . . . . . . . . . . . . . . . . . . 69
IV.3.2 Interactive 3D Beam Training . . . . . . . . . . . . . . . . . . . . . 74
IV.3.3 Discussion: Tradeo { IFS Overheads . . . . . . . . . . . . . . . . 75
IV.3.4 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
IV.4 Iterative Beam Training for Adaptive Modulation . . . . . . . . . . . . . . 82
IV.4.1 Iterative 2D Beam Training . . . . . . . . . . . . . . . . . . . . . . 84
IV.4.2 Pseudo-Code and Computational Complexity . . . . . . . . . . . . 86
IV.5 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
IV.5.1 Interactive Beam Training for Fixed Modulation . . . . . . . . . . 87
IV.5.2 Iterative Beam Training for Adaptive Modulation . . . . . . . . . . 91
IV.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
V Device-to-Device Adaptive Video Streaming for Distributed Programmable
WiFi Platforms 97
V.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
V.2 Push-Strategic D2D Video Streaming . . . . . . . . . . . . . . . . . . . . . 99
V.2.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
V.2.2 Video Streaming Principles . . . . . . . . . . . . . . . . . . . . . . 101
V.2.3 Algorithm Details . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
V.3 D2D Video Streaming with Greedy Heuristic . . . . . . . . . . . . . . . . 104
V.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
V.3.2 Heuristic { Fetching Bits based on a Playback Order . . . . . . . . 105
V.3.3 An Operational Framework . . . . . . . . . . . . . . . . . . . . . . 106
V.4 Software Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
viii
V.4.1 A Protocol Operation Overview . . . . . . . . . . . . . . . . . . . . 108
V.4.2 Helper Implementation . . . . . . . . . . . . . . . . . . . . . . . . 111
V.4.3 User Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 111
V.5 Used Video for Demonstration . . . . . . . . . . . . . . . . . . . . . . . . 111
V.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
VI Joint Scheduling and Stochastic Streaming for Device-to-Device Video
Delivery 113
VI.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
VI.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
VI.2.1 A Reference Network Model: Macro View . . . . . . . . . . . . . . 118
VI.2.2 A Reference Link Model: Micro View . . . . . . . . . . . . . . . . 119
VI.2.3 Video Streaming System Model . . . . . . . . . . . . . . . . . . . . 120
VI.3 Quality-Aware Scheduling and Streaming for Device-to-Device Video De-
livery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
VI.3.1 Design Rationale . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
VI.3.2 Device-to-Device Scheduling . . . . . . . . . . . . . . . . . . . . . . 123
VI.3.3 Streaming with Quality-Aware Stochastic Control . . . . . . . . . 126
VI.4 Simulation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
VI.4.1 Simulation Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
VI.4.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
VI.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
VII Conclusions 145
BIBLIOGRAPHY 147
ix
List of Figures
I.1 Frequency Plan of 60 GHz Wireless Communications . . . . . . . . . . . 4
I.2 Research Issues in Millimeter-Wave Wireless Systems . . . . . . . . . . . 7
I.3 Research Issues in Device-to-Device Wireless Video Systems . . . . . . . 9
II.1 A HetNet scenario with MWSC . . . . . . . . . . . . . . . . . . . . . . . 18
II.2 High Level Block Diagram of the Proposed Large Antenna Array (left)
and Example of Layout for the Case of Planar Sub-Array Modules (right) 22
II.3 Implementing MIMO with MAA . . . . . . . . . . . . . . . . . . . . . . 24
II.4 Transmitter and Receiver Architectures for a Two Stream Antenna Module 25
II.5 Multiple Access in MWSC . . . . . . . . . . . . . . . . . . . . . . . . . . 26
III.1 Link Budget Analysis: Capacity (Unit: bit/s) vs. Log-Scale Distance
(Unit: meters) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
III.2 System Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
III.3 A System Model: SRC and RDC stand for the \Source-Relay Combina-
tion" and \Relay-Destination Combination", respectively. . . . . . . . . 39
III.4 Generalized relationship between the quality index of transmitted HD
video signals and data rates. Based on dierent kinds of HD video sources,
the curve can be varied, but the general form is given as logarithmically
and monotonically increasing as proved in [133, 85, 124]. . . . . . . . . . 42
III.5 Performance Evaluation Simulation Setup: Wireless HD video cameras
are uniformly distributed on top of the stadium. There is one broadcast-
ing center at bottom. Between wireless video cameras on top of stadium
and broadcasting center, relays are uniformly and linearly deployed. To
vary simulation setting, the deployment of relays has three dierent types:
the relays are distributed near cameras (Scenario A), in the middle of
cameras (on top of stadium) and broadcasting center (Scenario B), and
near broadcasting center (Scenario C). . . . . . . . . . . . . . . . . . . . 53
III.6 Simulation Result for Single-Beam Antennas at Sources and Relays: Var-
ious Number of Sources (N
S
= 5; 10; 15) and Fixed Number of Relays
(N
R
= 10) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
III.7 Simulation Result for Single-Beam Antennas at Sources and Multiple-
Beam Antennas at Relays: Various Number of Sources (N
S
= 5; 10; 15)
and Fixed Number of Relays (N
R
= 10) . . . . . . . . . . . . . . . . . . 59
x
III.8 Simulation Result for Multiple-Beam Antennas at Sources and Relays:
Various Number of Sources (N
S
= 5; 10; 15) and Fixed Number of Relays
(N
R
= 10) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
III.9 Simulation Result for Various Lower Bound Setting: Single-Beam Anten-
nas at Sources and Multiple-Beam Antennas at Relays, N
S
= 10;N
R
= 15 61
IV.1 A General Beamforming Procedure for Link Conguration . . . . . . . . 67
IV.2 A Fundamental Procedure of Interactive Beam Training . . . . . . . . . 71
IV.3 A Basic Concept of Interactive Beam Training . . . . . . . . . . . . . . . 71
IV.4 An Example for 3D Beam Search Process within One Segmented Space:
Gray and black rectangles stand for the median SNR beams and the
highest SNR beam, respectively. In the gure,
j
and
i
stand for azimuth
planes index j (180

j
180

) and elevation plane index i (90

i
90

), respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
IV.5 IFS Durations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
IV.6 CMS Frame Structure [16] . . . . . . . . . . . . . . . . . . . . . . . . . . 77
IV.7 Fundamental Concepts of Iterative Subpace Partitioning . . . . . . . . . 82
IV.8 A Procedure of Iterative Beam Training . . . . . . . . . . . . . . . . . . 83
IV.9 Pseudo-code to ndN
2D
sector
and ' . . . . . . . . . . . . . . . . . . . . . . 86
IV.10 Simulation Results for Mobile Wireless Services . . . . . . . . . . . . . . 87
IV.11 60 GHz Channel Realization . . . . . . . . . . . . . . . . . . . . . . . . . 95
IV.12 T
exhaustive
=T
iterative
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
V.1 A Reference Example Network Model: There exists 2 helpers, i.e.,H =
fh
1
;h
2
g, and 3 users, i.e.,U =fu
1
;u
2
;u
3
g. Suppose that all helper-user
pairs are connected, i.e., bipartite graph. Then, each helper has 3 queues
which are for serving dedicated/associated users. In each time slot, users
place chunks to their selected helpers and determine the quality mode
of the chunk. In addition, helpers are doing transmission scheduling in
terms of max-weight scheduling in each time slot. . . . . . . . . . . . . . 102
V.2 Motivation: In the case of multiple helper and single user, each helper
selects the single helper all the time but the user cannot be served by the
multiple helpers at the same time because of the single channel constraint
in WiFi-centric wireless systems. Then, the user selects one helper in a
random manner which is equivalent to random scheduling in the worst
case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
V.3 Device Software Architectures . . . . . . . . . . . . . . . . . . . . . . . . 109
VI.1 An Example of a Con
ict Graph . . . . . . . . . . . . . . . . . . . . . . 118
VI.2 A D2D Link Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
VI.3 Two Dierentiated Unit Time Scales . . . . . . . . . . . . . . . . . . . . 122
VI.4 A Device Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
xi
VI.5 Algorithm: MWIS-based scheduling with message-passing in each l
i
2
L;8i2f1; ;jLjg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
VI.6 An example of stall events: If all bits of next chunk are arrived at the
playback buer of D2D receiver, there is no stall event since the receiver
can immediately play the next chunk when the playing of current chunk is
completed. Otherwise, the stall event will be occurred when the playing
of current chunk is completed. . . . . . . . . . . . . . . . . . . . . . . . . 133
VI.7 Pre-buering: With the denition of pre-buering, the number of stall
events can be reduced, i.e., user satisfaction can be increased. . . . . . . 134
VI.8 Performance Comparison between mpMWIS-QP and FlashLinQ-QP/FlashLinQ
in terms of the Expected Number of Stall Events . . . . . . . . . . . . . 143
VI.9 Average Quality vs. Expected Number of Stall Events . . . . . . . . . . 144
xii
Abstract
This dissertation explores the possible issues and proposes promising solutions in next
generation wireless video systems.
For next generation wireless systems, one of the main research contributions is ded-
icated to multi-Gbps system design and implementation. To achieve multi-Gbps data
rates, using millimeter-wave wireless channels is one of the most promising topics since
the millimeter-wave systems can easily achieve multi-Gbps data rates according to ultra-
wide bandwidth that is 2.16 Gbps in 60 GHz. Therefore, millimeter-wave technologies
are actively discussing in next generation 5G cellular research in the bands of 28 GHz and
39 GHz. Even though the millimeter-wave wireless systems have this multi-Gbps benet,
research challenges also exist. According to the higher carrier frequencies, the attenuation
of signals is a major factor that should be handled. To deal with this issue, relaying and
beam training algorithms are mainly used and discussed.
For relaying in millimeter-wave wireless systems, we investigated a joint compression
and relaying algorithm for outdoor video applications. Transmission of high-denition
(HD) video is a promising application for millimeter-wave wireless links, since very high
transmission rates are possible. In particular we consider a sports stadium broadcasting
system where signals from multiple cameras are transmitted to a central location. Due
to the high path-loss of 60 GHz radiation over the large distances encountered in this
scenario, the use of relays might be required. The proposed algorithm analyzes the joint
selection of the routes and the compression rates from the various sources for maximiza-
tion of the overall video quality. We consider three dierent scenarios: (i) each source
transmits only to one relay and the relay can receive only one data stream, and (ii) each
xiii
source can transmit only to a single relay, but relays can aggregate streams from dierent
sources and forward to the destination, and (iii) the source can split its data stream into
parallel streams, which can be transmitted via dierent relays to the destination. For
each scenario, we derive the mathematical formulations of the optimization problem and
re-formulate them as convex mixed-integer programming, which can guarantee optimal
solutions. Extensive simulations demonstrate that high-quality transmission is possible
for at least ten cameras over distances of 300 m. Furthermore, optimization of the video
quality gives results that can signicantly outperform algorithms that maximize data
rates.
For beam training in millimeter-wave wireless systems, we investigated a fast beam
training algorithm with receive beamforming. Both IEEE standards and the academic
literature have generally considered beam training protocols involving exhaustive search
over all possible beam directions for both the beamforming initiator and responder. How-
ever, this operation requires a long time (and thus overhead) when the beamwidth is
quite narrow such as for mm-wave beams (1 degree in the worst case). To alleviate this
problem, we propose two types of adaptive beam training protocols for xed and adaptive
modulation, respectively, which take into account the unique propagation characteristics
of millimeter waves. For xed modulation, the proposed protocol allows for interactive
beam training, stopping the search when a local maximum of the power angular spectrum
is found that is sucient to support the chosen modulation/coding scheme. We further-
more suggest approaches to prioritize certain directions determined from the propagation
geometry, long-term statistics, etc. For adaptive modulation, the proposed protocol uses
iterative multi-level beam training concepts for fast link conguration that provide an
exhaustive search with signicantly lower complexity. Our simulation results verify that
xiv
the proposed protocol performs better than traditional exhaustive search in terms of the
link conguration speed for mobile wireless service applications.
For next generation mobile systems, direct communication between mobile stations,
i.e., called device-to-device communications, is actively discussed in next generation 3GPP
cellular mobile systems. In addition, one of major applications of device-to-device mo-
bile systems is adaptive video streaming. Video streaming is becoming the dominant
application for wireless data transmission. Its spectral eciency can be greatly enhanced
by caching and device-to-device (D2D) communications. In this dissertation, we con-
sider a system where each device pre-caches a subset of video les from a library, and
users requesting a le that is not in their own library gets it delivered through D2D
communication. We develop and analyze centralized and distributed algorithms for the
delivery phase, encompassing a link scheduling and a streaming component. The cen-
tralized scheduling is based on the max-weighted independent set (MWIS) principle, and
uses message-passing to determine max-independent sets. The distributed scheduling is
based on the FlashLinQ link scheduling algorithm enhanced by max-weight scheduling.
The streaming component for both centralized and distributed algorithms makes use of a
quality-aware stochastic optimization approach for which users place sequential requests
for video \chunks" into a request queue, and adapt the video quality according to the state
of such a queue. The streaming and the scheduling components are coupled by the length
of the users' request queues, which provide the weights of the objective function in the
max-weighted independent set problem that the scheduler must solve at each scheduling
time slot. The per-chunk quality adaptation is reminiscent of current DASH technology
(Dynamic Adaptive Streaming over HTTP), considered in 3GPP for such applications.
We evaluate the performance (PSNR and stall probability) of the proposed algorithms by
xv
extensive simulation and compare with a baseline scheme formed by the FlashLinQ D2D
protocol for scheduling and DASH for streaming. We nd that our proposed algorithms
provide sizeable performance gains in terms of user satisfaction and queue stability.
xvi
Chapter I
Introduction
I.1 Next-Generation Wireless Video Systems
According to the estimation by Cisco [181], Cisco visual networking index (VNI) antici-
pated the sum of all forms of video will be in the range of 80% to 90% of global consumer
trac by 2017, and the trac from wireless and mobile devices will exceed the trac from
wired devices by 2016. Therefore, video-aware algorithms are desired for next generation
communication and network architectures and systems.
To support high denition video data over wireless channels, two types of wireless
communication technologies have been getting a lot of attention both from industry and
academia, i.e., (i) millimeter-wave wireless systems and (ii) device-to-device proximal
wireless systems.
With millimeter-wave wireless technologies, we can achieve multi-gigabit/s data rates
with ultra-wideband benets in higher carrier frequencies such as 28 GHz, 39 GHz, 60
GHz, and so forth. However, it has research issues in terms of high signal attenuation
due to high carrier frequencies. To deal with issue, advanced communication architectures
(such as small cells) and algorithms (such as relaying and beam training) are desired.
1
With device-to-device wireless technologies, direction communication without central-
ized operation is allowed. In conventional mobile cellular networks, if a central base station
(BS) is not working properly due to unexpected breakdown or abnormally temporal sup-
porting for huge number of users over the limit, there is no way to communicate each other.
Thus, allowing device-to-device proximal networking is important for reliable communi-
cation services. However, it has also research issues in terms of interference management.
In traditional cellular systems, BS controls every single wireless transmission between BS
and its associated mobile users, i.e., interference due to the wireless transmission will
be handled and managed by the BS; however device-to-device wireless networking will
generate unexpected interferences for nearby neighbor device-to-device receivers. To deal
with this issue, advanced scheduling algorithms are required. Moreover, consideration of
streaming is also desired since one of main applications of device-to-device networking is
video streaming and sharing.
In this dissertation, we explore the issues and potential solutions for the two main
wireless video communication technologies.
I.2 Millimeter-Wave Systems for High-Rate Data Trans-
mission and Video
As stated, millimeter-wave wireless technologies are getting a lot of attention due to
its ultra-wideband benets. In indoor scenarios, a 60 GHz wireless channel is generally
used for millimeter-wave wireless communications, i.e., WirelessHD [1], Wireless Gigabit
Alliance (WiGig) [2], and IEEE 802.11ad [20]. In outdoor scenarios, 28 GHz, 38 GHz, and
2
39 GHz wireless channels are generally used for fth generation (5G) cellular systems [11,
97, 103].
60 GHz Millimeter-Wave Wireless Systems
In IEEE 802.11 wireless networking academia and industry societies, traditionally suc-
cessful IEEE 802.11a/b/g technologies were able to achieve several tens Mbit/s data rates;
however this amount of data rates is not enough for high-speed wireless video transmis-
sion. In addition, current multimedia resources require more data rates to support video
or rich-media streaming. By introducing multiple input and multiple output (MIMO)
technologies to IEEE 802.11, several hundreds Mbit/s data rates became available in
IEEE 802.11n systems. This is more benecial in terms of video or rich-media wireless
transmission. However, due to the increased demand for higher data rates in consumer
electronics applications (e.g., uncompressed high-denition (HD) video streaming from
set-top boxes to wireless HDTV displays), a new technology which can support multi-
gigabit/s data rates will be required because approximately 1.5 gigabit/s data rates are
needed in uncompressed 1080p HD video streaming applications.
As presented in Fig. I.1, even if only one sub-channel is used, the 60 GHz millimeter-
wave wireless communication can support more than the 1.5 gigabit/s data rate (i.e., basic
requirement of uncompressed 1080p HD video streaming)

. Not only 60 GHz, the other
millimeter-wave bands (from near 30 GHz up to 90 GHz) can also achieve multi-gigabit/s
data rates because of their ultrawideband natures.

If the used carrier frequency is bigger than 30 GHz, it named as millimeter-wave (millimeter-wave)
wireless communications because of =c=fc where , c, fc stand for wavelength, speed of light (3 10
8
meter per second), carrier frequency, respectively.
3
3 4 2 1
240
MHz
120
MHz
2160 MHz
57 58 59 60 61 62 63 64 65 66 f
GHz
1728 MHz
Figure I.1: Frequency Plan of 60 GHz Wireless Communications
Because of the benet of millimeter-wave gigabit/s technologies, there are a lot of
attention in industry and IEEE standards. I
2
R and Intel stated that "According to ABI
Research, in 2016 two thirds of the WLAN integrated circuits will include 60 GHz." [18].
In addition, Qualcomm-Atheros and Wilocity announced Tri-band Wi-Fi (i.e., 2.4 GHz,
5 GHz, and 60 GHz) wireless chipset (Model Name: AR90004TB) for exploiting multi-
gigabit wireless data rates [7].
Outdoor 28 GHz and 39 GHz Millimeter-Wave Cellular Systems
Not only 60 GHz wireless channels, two main licensed frequency bands under consideration
by the industry for next-generation 5G cellular systems are the 28 GHz [37] and 39
GHz [38] bands. In the United States, there is about 850 MHz of contiguous spectrum
in the 28 GHz band. While a strong candidate, the main challenge with the use of this
band for 5G systems is its coexistence with incumbent systems such as satellite. Moving
higher in the frequency spectrum, the 39 GHz band oers a larger amount of contiguous
spectrum than the 28 GHz band (1.4 GHz vs. 850 MHz). The 28 GHz cellular systems
have been designed and implemented by Samsung Electronics [37]; and the proposed
systems theoretically and practically achieve 1 gigabit/s data rates anywhere to provide
a uniform gigabit/s experience to all users. Eventually, the 28 GHz 5G cellular systems
4
are expected to support up to 5 and 50 gigabit/s rates for high-mobility and pedestrian
users [19, 30]. The 39 GHz cellular systems have been designed and implemented by Intel
Corporation [182]. This system is fundamentally based on the modular antenna arrays
(MAAs) to achieve improved signal-to-noise (SNR) ratio; and will be mainly used for
small cell network scenarios to improve capacity in particular service areas [38].
Research Challenges in Millimeter-Wave Wireless Systems
In millimeter-wave wireless communication research, there are a lot of research issues and
results in various directions.
Realizing the potential for advances using millimeter-wave frequency bands poses a
host of fundamental questions in RF front-end hardware design [57, 49, 58, 27, 80, 41, 116,
43, 73, 39, 25, 46, 79, 69, 84, 59, 56, 47] and advanced network architecture design [11, 30,
37, 38, 97] that are intimately correlated to the unique characteristics of millimeter-wave
radio propagation [151, 170, 182].
The multi-Gbps rates at millimeter-wave put severe power and processing require-
ments when conventional transceiver designs for say 2.4/5 GHz bands are used [50, 48,
72, 74, 40, 42, 64, 62, 66]. For example, high precision analog-to-digital converters (ADCs)
become a major bottleneck for achieving desired data rates. Thus, the millimeter-wave
wireless communication requires innovation for a new radio transceiver hardware and
signal processing designs [65, 54, 55, 45, 81, 51, 75, 78, 83, 63].
Millimeter-wave links are fundamentally dierent from those at lower carrier frequen-
cies. Propagation loss for omnidirectional transmission is worse at higher carrier fre-
quencies [113], while directional antennas are much easier to implement at these frequen-
cies [38]. Given the small wavelength in the order of millimeters, millimeter-wave wireless
5
links have a limited ability to diract around obstacles and are therefore susceptible to
blockage [151, 170, 113, 108]. These peculiar traits of millimeter-wave wireless links have
a profound impact on millimeter-wave network protocol design [11, 30, 37, 38, 97]
Among these possible research issues, the main challenge for millimeter-wave gigabit
technologies is the severe attenuation experienced by millimeter-wave wireless signals due
to their relatively high frequency. According to the well-known Friis path-loss model:
P
Rx
=G
Tx
G
Rx

c=f
c
4d

2
P
Tx
(I.1)
whereP
Tx
,P
Rx
,G
Tx
,G
Rx
,c,d,f
c
stand for the transmit power, receive power, transmit
antenna gain, receive antenna gain, speed of light, distance, and carrier frequency, the
strength of millimeter-wave wireless packets is extremely weaker than traditional 2.4 GHz
or 5 GHz (below 6 GHz) IEEE 802.11 WLAN or IEEE 802.15 WPAN wireless packets.
There are two possible approaches, i.e., (i) relaying and (ii) directional transmission, i.e.,
beamforming and training, which can be used to compensate for the disadvantages.
With relaying functionalities, the short communication range due to the weakness of
millimeter-wave wireless signals can be compensated. With beamforming and training, the
signal powers can be focused to specic directions. The beamwidths of current millimeter-
wave commercial antenna solutions are extremely narrow such as 1

(the narrowest case)
or 10

(the widest case) [3, 4].
Organization of Millimeter-Wave Research Results
In Chapter II, advanced millimeter-wave small cell network architectures and systems are
proposed and implemented.
6
Millimeter-Wave
Video Transmission
Millimeter-Wave
Small Cell
Video Networks
Joint Compression and
Relaying Algorithm in
60 GHz Outdoor Stadiums
Fast
Millimeter-Wave
Beam Training
Figure I.2: Research Issues in Millimeter-Wave Wireless Systems
In Chapter III, relaying operations are considered for an outdoor sports broadcasting
stadium to send recorded real-time video signals from wireless cameras towards a central
broadcasting center. Chapter III formulated to obtain (i) optimum set of links between
wireless cameras and relays and (ii) corresponding data rates for all optimum links.
In Chapter IV, fast millimeter-wave beam training schemes are proposed. Due to the
extremely narrow beamwidth of millimeter-wave links, operating traditional beam training
procedures are not adequate. Therefore, Chapter IV proposed fast beam training schemes
using (i) interactive beam training and (ii) prioritized sector search ordering.
These three topics in millimeter-wave wireless systems are illustrated in Fig. I.2.
I.3 Device-to-Device Transmission Systems for High-Rate
Data Transmission and Video
Nowadays, dening next generation mobile cellular systems (i.e., next-generation LTE) is
one of big issues in mobile cellular research and LTE-Advanced standardization societies.
One widely discussed promising topic is enabling the direct link conguration between
7
LTE-capable mobile phones (i.e., user entities (UE)). This technique is named device-to-
device (D2D) communications. This kind of peer-to-peer communications is quite useful
because it can reduce communication overheads on base stations (BS or eNB).
Nowadays, two kinds of D2D communications are developing actively, i.e., (i) pure
peer-to-peer distributed D2D systems [132, 135, 91] and (ii) BS-controlled D2D sys-
tems [26]. For the pure peer-to-peer distributed D2D systems, FlashLinQ is one of the
most well-known D2D wireless schedulers; and it has been also actively discussed in LTE
standard groups [132]. The medium access control (MAC) features of FlashLinQ are
as follows [135, 91]: FlashLinQ establishes a priority of the users, and lower-priority
users are allowed to transmit if they do not create signicant interference to the higher-
priority links; this is why measurements of channel strengths in both directions have to
be done [132]. Fundamentally, the priority is determined by randomizing the priority
of links over time. Then a basic level of fairness across links can be maintained [132].
Theoretically, it can guarantee the maximum number of activated D2D links as analyzed
by the theory of stochastic geometry [91]. Even if the FlashLinQ has quite well-designed
scheduling architectures, it does not contain quality-aware video streaming algorithms.
In device-to-device research, various topics are under discussion for designing advanced
device-to-device network architectures and algorithms including FemtoCell-based device-
to-device networking [109], resource allocation [164, 104, 171, 166, 105, 127, 106, 168, 161],
interference management [149, 167, 90, 126, 153], system architecture designs [122, 104,
130, 34, 125, 131, 32], WiFi-based device-to-device networking [123, 36, 154], spectrum
sparing [87], caching systems [157, 152, 142, 110, 165], location-aware schemes [173].
As well-studied in [35], video delivery is one of the main applications of D2D proximal
wireless networking. Our considering D2D network architecture is the networks where
8
Device-to-Device
Video Transmission
Implementation of
Device-to-Device
Adaptive Video Streaming
Quality-Aware Scheduling
and Streaming for
Device-to-Device Video
Figure I.3: Research Issues in Device-to-Device Wireless Video Systems
the devices themselves are caching video les, and transmitting them, upon request,
to other devices via high-spectral-eciency D2D links. For this type of network, we
only consider the transmission of video les, not streaming, and also neglect the issue
of video rate adaptation. We rst outline the principle and mathematical model we
consider. We consider a network where each device can cache a xed number M video
les, and send them - upon request - to other devices nearby. The BS also keeps track
of which devices can communicate with each other, and which les are cached on each
device. Such BS-controlled D2D communication is more ecient (and more acceptable
to spectrum owners if the communications occur in a licensed band) than traditional
uncoordinated peer-to-peer communications. Specically, consider a network that consists
of macrocells (for simplicity each macrocell is assumed to have the shape of a square,
with the side of the square having unit length), with n users per cell. We assumed that
inter-cell interference of the device-BS links is kept to a small level through appropriate
cell/frequency planning [14].
One of the well-designed D2D architecture is a Femtocaching architecture former by
a set of helper nodes (a.k.a., small cell base stations, with large storage capacity and
9
no or very weak backhaul) serving a set of user nodes, which request on-demand video
streaming sessions. The two main problems to be addressed for such architecture are: 1)
the generalization of dynamic adaptive streaming currently used in wireless video stream-
ing applications-layer protocols such as Microsoft Smooth Streaming (Silverlight) [176],
Apple HTTP Live Streaming [29], and 3GPP Dynamic Adaptive Streaming over HTTP
(DASH) [145], to the case of a network of multiple users and multiple helpers; 2) the cache
placement problem, i.e., how to optimally place the video les into the helper caches.
Organization of Device-to-Device Research Results
In Chapter V, we implemented pre-proposed joint transmission scheduling and admission
control D2D adaptive video streaming algorithm [114, 162]. For the Android-based im-
plementation, we gured our the fundamental limitations on the WiFi-based platforms;
and proposed a heuristic-based method that can handle the issue.
In Chapter VI, we proposed a new joint scheduling and quality-aware stochastic
streaming algorithm for the xed source and destination D2D video streaming pairs.
In our pre-proposed algorithm in [114, 162], the end users select its serving video storages
in every unit time; however, the operation introduces hando delays in WiFi-based im-
plementation. Thus, we consider xed source and destination scenarios, that is a general
usage cases in D2D; and proposed a joint scheduling (with max-weight independent set
formulation) and streaming (with stochastic network optimization) algorithm.
These two research topics in D2D wireless video systems are illustrated in Fig. I.3.
10
I.4 Outline of The Dissertation
I.4.1 Millimeter-Wave Small Cell Video Networks
Increasing the capacity of cellular networks is becoming one of the most challenging tasks
of the mobile industry this decade. As traditional mechanisms to increase spectral ef-
ciency approach their theoretical limits, new and disruptive approaches are needed to
satisfy the growing demand of mobile data trac. The fth generation (5G) cellular
system is expected to make extensive use of small cells to increase the density and capac-
ity by several hundred times in comparison with 4G systems. While considerable focus
has been rightfully put into exploiting licensed frequency bands below 6 GHz, the vast
amount of licensed frequency spectrum in millimeter-wave bands has seen little use by
cellular systems despite holding far greater potential for enhancing capacity. This chapter
introduces a novel architecture for the next generation cellular systems with millimeter
wave capable small cells. Details are in Chapter II.
I.4.2 Joint Scalable Coding and Relaying for 60 GHz High-Denition
Video Streaming
Transmission of HD video is a promising application for 60 GHz wireless links, since very
high transmission rates are possible. In particular we consider a sports stadium broad-
casting system where signals from multiple cameras are transmitted to a central location.
Due to the high pathloss of 60 GHz radiation over the large distances encountered in this
scenario, the use of relays might be required. This chapter analyzes the joint selection
of the routes (relays) and the compression rates from the various sources for maximiza-
tion of the overall video quality. We consider three dierent scenarios: (i) each source
11
transmits only to one relay and the relay can receive only one data stream, and (ii) each
source can transmit only to a single relay, but relays can aggregate streams from dierent
sources and forward to the destination, and (iii) the source can split its data stream into
parallel streams, which can be transmitted via dierent relays to the destination. For
each scenario, we derive the mathematical formulations of the optimization problem and
re-formulate them as convex mixed-integer programming, which can guarantee optimal
solutions. Details are in Chapter III.
I.4.3 Fast Millimeter-Wave Beam Training with Receive Beamforming
This chapter proposes fast millimeter-wave (millimeter-wave) beam training protocols
with receive beamforming. Both IEEE standards and the academic literature have gen-
erally considered beam training protocols involving exhaustive search over all possible
beam directions for both the beamforming initiator and responder. However, this op-
eration requires a long time (and thus overhead) when the beamwidth is quite narrow
such as for millimeter-wave beams (1

in the worst case). To alleviate this problem, we
propose two types of adaptive beam training protocols for xed and adaptive modulation,
respectively, which take into account the unique propagation characteristics of millimeter
waves. For xed modulation, the proposed protocol allows for interactive beam training,
stopping the search when a local maximum of the power angular spectrum is found that
is sucient to support the chosen modulation/coding scheme. We furthermore suggest
approaches to prioritize certain directions determined from the propagation geometry,
long-term statistics, etc. For adaptive modulation, the proposed protocol uses iterative
multi-level beam training concepts for fast link conguration that provide an exhaustive
search with signicantly lower complexity. Details are in Chapter IV.
12
I.4.4 Device-to-Device Adaptive Video Streaming for Distributed Pro-
grammable Wi-Fi Platforms
This chapter proposes a joint optimization framework of transmission scheduling and
admission control for adaptive video streaming in WiFi-centric mobile platforms. In our
considering network model, two types of devices are dened as helpers and users where
helpers have desired video chunks within their own storage and users are candidates which
desire wireless video services. In our previous scheme, we designed a stochastic network
utility maximization algorithm for device-to-device adaptive video streaming which can be
decomposed into two sub-problems, i.e., (i) admission control at users and (ii) transmission
scheduling at helpers. As a transmission scheduling algorithm, max-weight scheduling was
considered. However, it is not working properly in WiFi-centric single channel mobile
platforms due to the fact that multiple helpers cannot simultaneously transmit video
chunks toward a single same user. Therefore, this chapter proposes a heuristic algorithm
where users can fetch video chunks from helpers in terms of playback order instead of
doing transmission scheduling at helpers. Our proposed scheme is simulated to show the
performance enhancement and also implemented on top of Android mobile programmable
platforms. Details are in Chapter V.
I.4.5 Joint Scheduling and Stochastic Streaming for Device-to-Device
Video Delivery
We propose two kinds of joint link scheduling and quality-aware streaming algorithms for
device-to-device (D2D) wireless networks. We assume that video les from a library are
pre-cached into the user devices. Each device cache contains a subset of the les, such
13
that users must stream the requested video les that are not in their own cache from
other users' devices. The proposed two centralized or distributed algorithms involves
a link scheduling and a streaming component. The centralized scheduling is based on
the max-weighted independent set principle, and uses message-passing to determine max-
independent sets. On the other hand, the distributed scheduling is the improvement of the
popular FlashLinQ link scheduling algorithm with the principle of max-weight scheduling.
The streaming component for both centralized and distributed algorithms makes use of a
quality-aware stochastic optimization approach for which users places sequential requests
to video \chunks" into a request queue, and adapt the video quality according to the state
of such queue. The streaming and the scheduling components are coupled by the length
of the users' request queues, which provide the weights of the objective function in the
max-weighted independent set problem that the scheduler must solve at each scheduling
time slot. The per-chunk quality adaptation is reminiscent of current DASH technology
(Dynamic Adaptive Streaming over HTTP), considered in 3GPP for such applications.
We evaluate the performance of the proposed algorithm by extensive simulation and
compare with a baseline scheme formed by FlashLinQ at the MAC layer and DASH
at the application layer. In particular, we study the system performance in terms of
the number of video streaming stalls at receivers. According to our simulation results,
the proposed algorithm presents clear and sizeable performance gains in terms of user
satisfaction (video stall probability) and queue stability. Details are in Chapter VI.
14
Chapter II
Millimeter-Wave Small Cell Video Networks
II.1 Introduction
The remarkable growth in data trac led by the use of wireless video content is creating
a signicant challenge for future cellular networks [181]. Availability of licensed spectrum
in lower bands, continues to be scarce and quite expensive worldwide. Furthermore,
the traditional methods for improving spectrum usage are approaching their theoretical
limits [181]. As such, when coupled with the expected ten-fold increase in the density
of connected mobile devices [178], a rethinking of todays cellular network architecture is
required in order to meet the future demand of network capacity, device density, and user
experience.
The distribution of trac demand is unbalanced, both in terms of number of users
and the areas served within the mobile cells [8]. If the demand is not adequately met, user
experience across the multiple cells can signicantly degrade. Therefore, mobile operators
have been contemplating a combination of few approaches to serve the increased user
demand:
15
Using more spectrally ecient technologies such as 4G cellular systems, which can
provide up to two times performance boost when compared to 3G technologies [9];
Deploying a denser cell topology, which allows the same capacity to be served to a
smaller region or a fewer number of users. These cells are commonly known as small
cells, since they have a much smaller cell coverage than the existing macro cells.
When these approaches are used jointly, they are expected to increase network capacity by
more than one thousand times. Among these, however, several studies have shown that the
technique oering the greatest benet involves deployment of small cells [10, 179]. With
small cells, mobile operators can rely on macro cells to oer mobility while using small
cells for targeted inll of capacity. To meet the growing data trac demand, however,
mobile operators might have to deploy small cells in several orders of magnitude more
compared to existing macro cell numbers. Such large scale and dense deployment require
that the technology used within the small cells allows their seamless integration with
existing macrocell architecture, without causing undue interference to other cells [180].
The use of millimetre-wave (mm-wave) frequency bands for small cells is expected to
provide the necessary scalability, capacity and density required for a seamless integration
of these cells into the cellular network infrastructure [30, 11, 97].
The mm-wave frequency bands oer larger spectrum availability, increased network
capacity, and greater potential for spatial network densication while reducing the need
for high spectral eciencies. All these benets come at the expense of potentially greater
system complexity particularly in terms of radio frequency (RF) front end and antenna
design, but the recent advancements around unlicensed 60 GHz wireless systems develop-
ment have produced cost eective solutions that can be leveraged to overcome many of
16
these challenges [20]. In this chapter, we introduce the novel concept of mm-wave small
cells (MWSC) as a key ingredient of the next generation 5G cellular system. We demon-
strate that MWSC can oer a multi-fold increase in cellular system capacity through
densication and directional communication. MWSC can be deployed as an extension to
existing 4G network architecture, thus paving the way for an incremental and progressive
migration from 4G towards 5G cellular systems.
This rest of this chapter is organized as follows. Section II.2 gives an overview of
typical small cell scenarios and key requirements. The novel concept of MWSC, core
functionalities and target frequency bands are presented in section II.3. Section II.4
addresses the design challenges and considerations for the development of MWSC at the
RF/antenna, physical layer and medium access. The chapter is concluded in section II.5.
II.2 Small Cell Usages and Key Requirements
The main characteristic of a small cell is its range. For small cells developed using cel-
lular technologies, the range is expected to be typically around 10 to 200 meters under
NLOS conditions [180]. Small cells can be deployed indoor (e.g., femto cells) or out-
door. Small cells can be managed or unmanaged. Managed small cells are those that
are deployed under the control of the cellular operator. Unmanaged small cells are those
deployed/installed by users.
Small cells are deployed with one of two primary target usages:
Coverage extension/enhancement: deployment of small cells at the edge of a macro
cell to extend the coverage of the cellular system. Coverage of the small cell and
coverage of the macro cell may partially overlap.
17
Anchor BS
Booster BS
MS/UE
Backhaul Link
Backhaul Link
Backhaul Link
Anchor
Access
Link
Booster
Access
Link
Macro Cell
Macro Cell
Small Cell
Small Cell
Small Cell
Figure II.1: A HetNet scenario with MWSC
Capacity improvement: deployment of small cells within the coverage of a macro
cell to improve throughput of the cellular system. Usually, the coverage of the small
cell and the coverage of the macro cell overlap to a large extent.
Depending on the specic usage and the available spectrum, the number of small cells per
macro cell can vary considerably. In one outdoor metropolitan network deployment, it
has been found that between six and seven outdoor small cells would need to be deployed
per macrocell to satisfy the projected trac demand by 2019 [180].
II.3 Proposed Millimeter-Wave Small Cells
Given the high capacity, the target usage of MWSC is expected to be initially in hotspots
and for Quality-of-Experience enhancement. Over time and as the technology evolves,
MWSC can nd applications in many usages in both outdoor and indoor scenarios.
Fig. II.1 depicts a heterogeneous network (HetNet) scenario utilizing MWSC. We denote
the uplink base station (i.e., eNB) that serves as the point of attachment for a downlink
18
small cell eNB as an anchor eNB. A small cell eNB is denoted as a booster eNB. Note that
a booster eNB may not always have an anchor eNB. We believe that next generation 5G
cellular systems can use mm-wave capability in any of the following scenarios, as shown
in Fig. II.1:
In the access link between a booster eNBs and a user equipment (UE), i.e., the
booster access link;
If the booster eNB has a built-in Wi-Fi interface, in the ooad link between a
booster eNB and a UE using a 60 GHz mm-wave link [20];
In the backhaul link between a booster eNB and an anchor eNB, between two booster
eNB(s), and between two anchor eNBs. For managed small cells, the booster eNB
can be assumed to have a link with an anchor eNB either directly or via other
booster eNBs. This is not the case with unmanaged small cells, which typically
connect to the core network via traditional wired last mile technologies such as
digital subscriber line (DSL), ber, cable, etc.
In the HetNet scenario of Fig. II.1, an MWSC is operated by mm-wave capable booster
eNB that provides high-speed and low latency access to UEs. In addition, the use of
mm-wave frequencies can also be used to support high-speed backhaul links among the
MWSC, thereby providing the needed capacity to sustain the combined trac load of the
underlying MWSC.
II.3.1 Functionalities
The three main functions provided by a mm-wave capable booster eNB are:
19
Cellular access: mobile access is provided by LTE as well as a new licensed mm-
wave access. LTE can be used to support legacy UEs that either do not support
mm-wave access or for which mm-wave access is not the most suitable access given
the channel and/or trac conditions. MmWave access is used to provide the high-
capacity access for UEs that need and support mm-wave access capability.
Cellular ooad: cellular ooad is provided by unlicensed technologies, which in this
case constitute of both Wi-Fi (2.4/5GHz) and WiGig (60 GHz).
Cellular backhaul (and fronthaul): the mass deployment of small cells requires new

exible and cost acceptable backhaul (and fronthaul) solutions. For MWSC that
have high-capacity mobile access, this implies the wireless backhaul has to have a
capacity that is large enough to serve the trac demand needed for cellular access
and ooad.
II.3.2 Considering Frequency Bands
Millimeter wave bands generally refer to frequencies between 30 GHz and 300 GHz, which
correspond to wavelengths between approximately 10 and 1 millimeter. General propa-
gation characteristics of these bands in comparison to lower frequencies include higher
free-space transmission loss, higher diraction loss, higher penetration loss, higher atten-
uation due to rain and lower impact from atmospheric refraction. On the
ip side, smaller
wavelength in these frequencies compared to lower bands enables utilization of high gain
antenna arrays with very large number of elements in relatively smaller dimensions.
The two main licensed frequency bands under consideration by the industry and reg-
ulators for future MWSC deployment are the 28 GHz and 39 GHz bands. In the US
20
there is about 850 MHz of contiguous spectrum in the 28 GHz band (also known as lo-
cal multi-point distribution systems (LMDS) band) that can be harvested for MWSC.
While a strong candidate, the main challenge with the use of this band for 5G systems is
its coexistence with incumbent systems such as satellite. Moving higher in the frequency
spectrum, the 39 GHz frequency band oers a larger amount of contiguous spectrum than
the 28 GHz band (1.4 GHz vs. 850 MHz). Similar to the 28 GHz band, there are also
other incumbents in the 39 GHz band with whom coexistence needs to be investigated.
The industry is currently undertaking a series of studies and measurements to validate
the use of these frequency bands for MWSC.
II.4 Design Challenges and Considerations
II.4.1 RF, Antenna, and Physical Layer Design
The major technical challenge at the physical layer for the implementation of MWSC
is the availability of a cost eective, low power and small form factor RF and antenna
solution.
A viable MWSC requires an antenna with high directivity. This is in contrast with
modern communication systems, which require a station to be capable of covering a
relatively wide sector around it to communicate with other stations regardless of their
locations. Traditional antenna architectures used in mm-wave band are, generally, not ca-
pable of combining wide angle coverage with high directivity. Existing re
ective, parabolic
dishes and lens antennas can create narrow beam, thus delivering the needed 30-40 dB
antenna gain, but they lack the
exibility to cover wide angle coverage and are relatively
bulky. Phased patch antenna arrays allows steering the beam to a desired direction.
21
Figure II.2: High Level Block Diagram of the Proposed Large Antenna Array (left) and
Example of Layout for the Case of Planar Sub-Array Modules (right)
However, to achieve the necessary directivity, the array must consist of a large number of
elements (several hundred to thousands).
Antenna array architectures currently used for mass production and intended for per-
sonal devices employ a single module, containing an RFIC chip that includes controlled
analogue phase shifters capable of providing several discrete phase shifting levels. The
antenna elements are connected to the RFIC chip via feed lines. However, due to the loss
inherent in the feed lines, this approach reduces antenna gain and eciency, and becomes
a severe problem when the number of antenna elements and radio frequency increase [97]
as in MWSC.
To overcome this limitation, we propose a novel modular antenna array (MAA) ar-
chitecture for MWSC that provides
exibility in form factor choice, beam steering, and
array gain in a cost eective manner. The architecture is shown in Fig. II.2 and is essen-
tially a type of massive multiple-input multiple-output (MIMO) system. However, instead
of using an individual antenna module, MAA is constructed using modular, composite
mm-wave antenna arrays. Each module is implemented with a dedicated RFIC chip serv-
ing several antenna elements and an RF beamforming (RF-BF) unit with discrete phase
22
shifters. Given its modularity, the length of the feed lines in the MAA architecture can
be kept much shorter hence incurring much lower feed line loss. This makes the MAA
architecture much more
exible and ecient than existing approaches.
The aperture of the MAA and total transmitted power may exceed that of an indi-
vidual sub-array module, and is proportional to the number of sub-array modules used.
Therefore, much narrower beams may be created and, thus, MAA can provide much
greater antenna gains and transmit power in a cost ecient way compared to using an
individual array with the same number of antenna elements. This architecture also allows
dynamically conguring sectors covered by dierent sub-arrays in such a way as to vary
the coverage angle of the composite array, thereby creating several coverage angles (e.g.,
to communicate with several peer stations simultaneously).
In this section, a brief overview of the capabilities of MAA is presented for several
dierent scenarios. The use of mm-wave band makes it possible to arrange communication
with single or with multiple users simultaneously, because the short wavelength allows
creation of beamformed links in a very
exible manner. A mm-wave antenna of relatively
small size (e.g., 30 mm) is capable of creating narrow beams on the order of 1-2 degrees
that produce very little mutual interference between data streams directed to dierent
users. The propagation properties of the mm-wave channel also contribute to creating
narrow beams by greatly attenuating multi-path components, thus allowing the advantage
of directed transmissions over line-of-sight ray or over the best re
ected ray.
There are several challenges that aect applicability of traditional MIMO in the mm-
wave band. Traditional MIMO implementations in low frequency bands assume each
antenna element can make use with its own transceiver RF chain. In practice, this may
be too challenging for mm-wave owing to the sheer number of elements involved for any
23
Figure II.3: Implementing MIMO with MAA
degree of antenna array. Furthermore, such tightly located components will be prone to
cross-coupling which aect the MIMO performance. In the case, it is possible to use the
MAA architecture as depicted in Fig. II.3, where each phased array module (sub-array)
has a dedicated transceiver with coarse RF beamforming unit and, therefore, may create a
dedicated RF beam. Beams of individual sub-arrays may be steered in various directions
to achieve a number of goals.
In the case of Fig. II.3, each antenna sub-array module may be seen as a single
antenna port in the context of a MIMO system. The beamforming procedure of the
entire antenna system may be provided in hybrid manner by the coarse phasing of each
antenna element within a module and ne signal weighting in a MIMO Base Band (BB)
processor. For example, at the transmitter side, ne beamforming in BB may be used
to create spatially orthogonal signals for dierent antenna sub-arrays for both single user
MIMO (SU-MIMO) and multi-user MIMO (MU-MIMO) modes. In addition, dierent
24
Figure II.4: Transmitter and Receiver Architectures for a Two Stream Antenna Module
MIMO schemes may be implemented in the receiver, such as maximum ratio combining
and inter-stream interference cancellation.
An important aspect to note is that more complex antenna modules (sub-arrays)
may be used to create MAA derivatives capable of maintaining several spatial streams
(beams) in the RF, where each beam is formed using all available antenna elements. The
architecture of the transmitter (TX) and receiver (RX) for a two stream antenna module
is shown in Fig. II.4. As described in the following sections, dierent MAA congurations
may be implemented for dierent deployment scenarios, which demonstrate the
exibility
of this core architecture.
II.4.2 Multiple Access and System Design
Multiple Access
Current LTE adopts OFDMA as the DL access scheme and SC-FDMA as the uplink access
scheme. Each UE is allocated part of the frequency band and dierent precoders can be
applied for each individual resource block. In mm-wave, the coarse analog beamforming
is applied in RF, which implies that the entire frequency band will be beamformed in
25
• Single User & Multiple Modules
• Increased Directivity Gain (Range)
• Single User & Single Module
• Standard Directivity and Rate
• Multiple Users & Single Module
• Multiple Access (e.g., OFDMA) required
• Single User & Multiple Module
• 2 X 1 MIMO for TX diversity; or Increased Data Rate
Figure II.5: Multiple Access in MWSC
one or several directions at one time. So, within the direction of this analog beam, the
number of simultaneous served users may be reduced.
Fig. II.5 illustrates an example of a multiple access in a MWSC when the MAA
architecture is used by the booster eNB. By using MAA, the booster eNB forms multiple
beams to serve UEs in the MWSC. Depending on factors such as the number of antenna
modules at the booster eNB, geographical location of the UEs, channel propagation, etc.,
a beam from the booster eNB can cover more than one UE. Therefore, in addition to
providing a multiple access scheme that services UEs covered by dierent beams, the
multiple access mechanism used in a MWSC is also required to support multiple access
on a beam level.
26
Beamformed Link Management
Aspects such as environmental changes in the vicinity of a UE or booster eNB, UE mo-
bility, among others, can cause a beamformed link between a booster eNB and UE to
degrade and potentially break. Therefore, it is important to proactively maintain an al-
ready established beamformed link to deal with the consequences of environmental and
mobility changes. With respect to UE mobility, given the shorter range of small cells
when using mm-wave frequencies, it is reasonable to assume that serviced UEs move at
pedestrian speed and, in some cases, channel conditions can be primarily NLOS.
The nature of the change in the quality of a beamformed link between a UE and its
booster eNB impacts the approaches that should be considered to maintain this link:
Slow changes to beamformed link: in this case, there is no direct impact to schedul-
ing oered by the booster eNB. The beam tracking algorithm can be employed
during the actual data exchange to adapt to slow channel changes.
Full breakage of beamformed link: in this case, the beamformed link between the
UE and booster eNB is assumed to be lost and beam tracking cannot be used.
The mobility pattern of a UE may result in hando from one booster eNB to an-
other, which does not need to trigger re-beamforming. In this case, coordination
between nearby MWSC is desirable to facilitate the beam acquisition during hand-
o. Specically, the new booster eNB should be notied of the approaching UE,
its location and capabilities so that the mm-wave beamformed link can be quickly
established.
27
Spatial Sharing
Since the simultaneous communication between the booster eNB and the UEs can be
suciently spatially separated from each other to cause harmful mutual interference, a
booster eNB can schedule transmissions to and from dierent UEs simultaneously through
the use of spatial sharing mechanisms.
Spatial sharing may impact the design of various elements, including UE and booster
eNB measurements, beamforming and scheduling. Moreover, mutual interference is not
the only deciding factor. For example, depending on the type of trac being transmitted
or UE mobility pattern, the scheduler may decide not to perform spatial sharing allocation
between two links. Such aspects have to be considered in the overall system design.
II.5 Concluding Remarks
The use of mm-wave frequency bands as part of the next generation 5G cellular systems
is gaining increased momentum in the industry and academia. The mm-wave frequen-
cies oer larger spectrum availability and much faster data rates, thus being considered
an important component in enabling high capacity small cells. This chapter discusses
the exploitation of mm-wave bands through the introduction of a novel architecture for
next generation cellular systems with millimeter-wave small cells (MWSC). We show that
MWSC can signicantly increase capacity and density for 5G cellular systems, and also
discuss the basic functionalities of MWSC, target frequency bands, key radio access net-
work design challenges, antennas technologies, to name a few. Research on the use of
mm-wave bands for cellular communication is at its infancy, and this chapter provides a
number of promising technical directions in this area.
28
Chapter III
Joint Scalable Coding and Relaying for 60 GHz
High-Denition Video Streaming
III.1 Introduction
Recently, wireless data transmission and media streaming in the millimeter-wave (mm-
wave) frequency range have received a lot of attention by the wireless communications
and consumer electronics communities. In particular the 60 GHz frequency range is of
great interest: a 7 GHz wide band (58-65 GHz) has been made available for unlicensed
operation. This large bandwidth enables multi-Gbps wireless data transmission [21, 24],
which enables, in turn, high denition (HD) video streaming in an uncompressed, or less
compressed, manner. Therefore, several industry consortia such as WirelessHD [1] and
the Wireless Gigabit Alliance (WiGig) [2] have developed related technical specications.
Also within the IEEE, there are two 60 GHz mm-wave standardization activities, i.e.,
IEEE 802.15.3c Millimeter Wave Alternative PHY [16] and IEEE 802.11ad Very High
Throughput (VHT) [20]. First consumer electronics products for short-range transmission
(e.g., from Blue-Ray player to HDTV) have recently become available.
29
In this chapter, we analyze 60 GHz for longer-range outdoor applications, specially
an outdoor sports broadcasting system. In this system, there are multiple wireless HD
video cameras in a sports stadium for high-quality real-time broadcasting, all sending
their data to a single destination (called \broadcasting center", even though it is the
receiver of the data streams). To transmit uncompressed HD video streams in real time,
a data rate of approximately 1:5 Gbps is required

. Since the distance between wireless
HD video cameras and a broadcasting center is on the order of several hundred meters,
the high path-loss at 60 GHz is one of the main challenges that leads to a limitation of
coverage and/or reduction of the possible data rate. One promising way to deal with
this problem is using relays to extend the coverage range [128]. Increasing the number of
relays obviously improves performance, but also increases costs. We are thus interested
in nding the tradeo between performance and number of relays.
We furthermore take the complexity of the antennas into account. In order to com-
pensate for the high pathloss, as well as to reduce interference, high-gain antennas need
to be employed. We distinguish between the situations where the antenna can form only
a single beam, or multiple beams: (i) If both source and relay have only a single beam,
then each source has to select a suitable relay, and the relay can only receive from this
particular source. (ii) If sources have single beams, but relays have multiple beams, then
the source can transmit only to a single relay, but the relays can receive data from mul-
tiple sources and aggregate them before forwarding to the destination. (iii) If also the

In a single HD video frame, 1080 1920 pixels exist and each pixel has 24 bit information for RGB
format (8 bit for Red, Green and Blue color representation, respectively). In addition, for one second,
30 frames are required in a standard mode. Therefore, approximately 1:5 Gbps data rate is required for
uncompressed HD video streaming. In addition, for the enhanced mode, 60 frames are required [24]. This
chapter considers the standard mode but can be extended for the enhanced mode as well.
30
source has multiple beams, it can split its data stream into multiple parallel streams and
send them to the destination via parallel links. In case (i), some HD video streams from
sources cannot be served by the relays if the number of relays is less than the number of
sources. In cases (ii) and (iii), appropriate compression and routing of multiple streams
via the same relay can be used.
Relay selection for maximization of data throughput has been analyzed in many papers
(see Section III.2). However, for video transmission, we are not interested in maximizing
the data rate arriving at the destination, but rather the video quality, which is related to
the data rate in a nonlinear manner. To achieve this goal, the proposed mathematical
formulation will select the relays for every single HD video stream and decide the cod-
ing (compression) rates for each stream. With this formulation for the three cases, the
optimal solutions can be obtained by (i) the theory of unimodular matrices, (ii) a BBLP
algorithm [52], or (iii) standard convex optimization techniques.
The remainder of this chapter is organized as follows: Section III.2 gives an overview
of related work. Section III.3 explains the details of our reference system. Section III.4
presents the details of the joint scalable video coding and relaying algorithms for max-
imizing the delivered HD video qualities for the three cases. Section III.5 presents the
technique to convert the mathematical optimization framework of Section III.4 to a con-
vex form, which can guarantee optimal solutions. Section IV.5 evaluates the performance
and section V.6 concludes this chapter.
31
Table III.1: Related Work Comparison Table
Consideration Factors [118] [60] [71] [119] [92] [139] Proposed
Route Selection

Rate Allocation

Millimeter-Wave Channels

Multiple-Antenna Elements
Video Streaming

Insucient Number of Relays - - - - -
III.2 Related Work
The topic of wireless network technologies for outdoor sports stadium system was dis-
cussed in [136]; however the fundamental setup diers from ours in that [136] considers
content distribution to wireless devices of the audience in the stadium, while our inves-
tigations are for the real-time streaming service to a broadcasting center in the stadium
and from there to audiences at home. In terms of fundamental technology, our research is
related to both scalable video coding rate control and relay selection/routing for real-time
video transmission.
For the video relaying issue, example publications include [118, 60, 71, 119]. The pro-
posed scheme in [118] addresses opportunistic routing for video transmission over IEEE
802.11 wireless networks under given time constraints. The proposed scheme is ecient
in the given multi-hop IEEE 802.11 wireless networks, however, it does not consider the
route paths selection problem. Ref. [60] considers distributed video streaming in multi-
hop wireless networks. This chapter considers network architectures similar to ours (when
32
specialized to the two-hop case), but the proposed algorithm cannot consider the rate con-
trol mechanism. The formulation in [71] considers multipath selection for video streaming
in a mobile ad-hoc network (MANET) architecture. The main constraint for this algo-
rithm is the interference over the given wireless channel, a factor that does not play a role
in our 60 GHz mm-wave wireless channel, where the high directionality of the antennas
prevents inter-stream interference. The scheme in [119] considers a route selection mecha-
nism for video streaming, but using multipath video streaming with multicast techniques,
which diers from our setup where only a single destination is used. All of these papers
[118, 60, 71, 119] only consider the video multi-hop wireless networks but do not consider
the video coding rate control. For the same reason, the rich literature of relay selection
and routing of \conventional" data transmission is not applicable to our scenario. In
previous research on video streaming, schemes usually considered multipath video data
transmission to combat the limited wireless bandwidth [60, 71, 119]. In addition, some
of the research considered retransmission of video signals and tried to reduce transmis-
sion time [118]. However, thanks to the very large available bandwidth at 60 GHz, these
factors are not considered in this chapter.
A representative work which considers both rate control and route selection appeared
in [92]: the proposed algorithm selects the best relays for individual unicast data
ows and
it selects the corresponding data rates as well. However, the relays in [92] cannot aggregate
video streams, which is required when the number of relays is smaller than the number of
unicast
ows in real-time video streaming applications. In addition, the proposed frame-
work does not consider the properties of video, namely the nonlinear relationship between
data rate and video quality because it is for generalized cooperative multi-hop networking
systems. In addition, the algorithms in [118, 60, 71, 119, 92] do not consider the features
33
of mm-wave wireless channels; in particular, they do not consider beamforming for inter-
ference suppression, which is an essential part of our architecture. Ref. [88] considers the
main features of mm-wave wireless communications, i.e., high directionality. It designs
the medium access control mechanisms for 60 GHz wireless channels, however, it does not
consider the features related to video streaming. In [139], we considered the properties of
the 60 GHz channel as well as rate control and video quality, but we restricted ourselves
to the case that the number of relays exceeds the number of sources. This comparison
is summarized in Table III.1. In a conference version of our work [169], per-link quality
is considered instead of per-source quality consideration. Considering per-link quality is
meaningful when multiple streams emanate from one source location, each being trans-
mitted via one link. If, as assumed in this chapter, each source creates one video stream,
considering the quality of each source is the most meaningful consideration.
III.3 A Reference System Model
III.3.1 Link Budget Analysis: Capacity Perspective
A link budget analysis [14] provides the fundamental tradeo between data rate and range
that can be achieved. Using Shannon's equation for the capacity
C =B log
2
(1 + SNR) (III.1)
where SNR is equal to P
signal
=P
noise
as a linear scale, P
signal
and P
noise
stand for the
signal power and noise power, respectively. In addition, B stands for bandwidth and is
34
considered as 2:16 GHz following WiGig specication [20]. The signal power expressed in
dB, P
signal, dB
, can be computed as follows:
P
signal,dB
=E +G
r
WO(d) +F (d) (III.2)
where E stands for the equivalent isotropically radiated power (EIRP), which is limited
by frequency regulators to 40 dBm in the USA and 57 dBm in Europe. G
r
means the
receiver antenna gain; In our system it is set to 40 dB, which corresponds to commercial
high-gain 60 GHz outdoor scalar horn antennas [3, 4], which we propose to achieve large
communication range. W presents the shadowing margin and is set to 10 dB; while line-of-
sight (LOS) is anticipated for our deployment, obstacles such as passing-by people, raised
banners, etc., might attenuate the LOS. F (d) represents the path loss, which depends on
the distance (in meter) d between transmitter and receiver
F (d) = 10 log
10

4d

n
(III.3)
where n stands for the path loss coecient and is set as 2:5 [21]. In addition, the wave-
length () is 5 millimeter at 60 GHz. O(d) means the oxygen attenuation at d, which can
be computed as follows:
O(d) =
8
>
<
>
:
15
1000
d d 200 meter;
0 d< 200 meter [21]:
(III.4)
The noise power in dB, P
noise,dB
can be computed as follows:
P
noise,dB
= 10 log
10
(k
B
T
e
B) +F
N
(III.5)
35
Figure III.1: Link Budget Analysis: Capacity (Unit: bit/s) vs. Log-Scale Distance (Unit:
meters)
wherek
B
T
e
stands for the noise power spectral density which is174 dBm/Hz andF
N
is
the noise gure of the receiver and set as 6 dB.
The leads us to conclude that approximately 200 300 m is the maximum distance
for successful signal decoding when the maximum rate of 1:5 Gbit/s is used, as shown in
Fig. III.1.
III.3.2 Outdoor Broadcasting Systems with 60 GHz Wireless Links
It follows from the above link budget that the assistance of relays is required if the wireless
communication range between wireless HD video cameras and a single broadcasting center
is more than 200 300 m. In the general transport layer mechanisms such as TCP, the
congestion control mechanism encounters a number of new problems and suers from
36
Raw Video
Streams
Raw Video
Streams
Raw Video
Streams
Scalable
Video
Encoder
HD Video
Streams
Layer (N)
Layer (N-1)
Layer (N-2)
Layer (1)
Layer (2)
…
Bit-Stream Extractor
High-
Definition
(HD) Video
Camera
Transceiver
60 GHz Antenna
(Single-Beam
or Multiple-Beam)
Scalable Video Coding (SVC)
Optimal Coding Level Decision
Stream Handler
(a) Wireless HD Video Camera (Source) Structure
Layered
HD Video
Bit-Streams
Aggregation
Signal Processing
Transceiver
60 GHz Antenna
(Single-Beam
or Multiple-Beam)
Relay
Transceiver
Sub-Channel
Bit-Stream
Allocator
60 GHz
Single-Beam Antenna
(b) Relay Structure
SVC-Decoder
SVC Decoder Set
Broadcasting Center
Transceiver
Database
Broadcasting HD Video
Producing
Production
Sending to
Contents
Provider
Companies
SVC-Decoder
SVC-Decoder
…… 60 GHz Antenna
(Multiple-Beam)
(c) Broadcasting Center Structure
Figure III.2: System Components
37
poor performance in multi-hop wireless networks [61]. Thus, considering a small number
of hops is benecial. Furthermore, the size of sport stadium (i.e., from wireless HD video
cameras to the antennas of a broadcasting center) is not more than 500 m, e.g., the large
side of Los Angeles Memorial Coliseum is approximately 300 m. Thus, we can safely
restrict the number of relays to one, i.e., a two-hop network.
In our outdoor sports broadcasting system, mainly three components with 60 GHz
wireless communication capabilities are relevant, i.e., wireless HD video cameras, relays,
and a broadcasting center using multiple antennas. As presented in Fig. III.2(a), the
proposed wireless HD video cameras have scalable video coding (SVC) functionalities that
reproduce the real-world analog video signals as layered SVC-coded HD video bit streams.
If the achievable rate of a 60 GHz link is sucient for uncompressed HD video streaming
(i.e., 1:5 Gbit/s), then all SVC-coded layers can be transmitted, i.e., the optimal coding
level decision module selects all layers. Hence, this can preserve the maximum quality of
the delivered video streams. This achievable rate between x and y (A
x!y
) is computed
by (III.1). If the computed achievable rate is not enough for uncompressed HD video
streaming (i.e., less than 1:5 Gbit/s), the optimal coding level decision module has to
determine the maximum number of layers, reducing the overall video quality (see below)
as discussed in [67, 82].
Each wireless HD video camera can have one of two antenna types: single-beam an-
tennas, or multiple-beam antennas. Single-beam antennas usually are high-gain antenna
structures such as horn antennas or Cassegrain antennas. In the scenario with single-beam
antennas at the sources, all the multiple SVC-coded streams in each camera are assigned
to the single antenna, and thus transmitted to the same relay. If the antenna can form N
independent beams the multiple SVC-coded streams are divided into N parts and each
38
Broadcasting
Center, D
s
1
s
2
s
3
s
|S|
……

……

r
1
r
2
r
|R|
SRC

RDC

Figure III.3: A System Model: SRC and RDC stand for the \Source-Relay Combination"
and \Relay-Destination Combination", respectively.
part is assigned to a beam to be concurrently transmitted. Normally, the multiple beams
will be formed by phased-array type antennas, though the use of multiple horn antennas
pointing into dierent directions is possible as well.
If the number of sources exceeds the number of relays, the relays have to have multiple-
beam antennas for reception. If the number of sources is smaller than the number of relays,
single-beam antennas might be sucient. In either case, the number of beams for trans-
mission to the broadcast center need not exceed one, since there is only one destination.
The destination, however, always has to be able to receive multiple beams simultaneously.
Fig. III.2(b) shows the proposed architecture when the relays have multiple-beam anten-
nas. The relays use their built-in digital signal processing unit to aggregate the received
HD video signals and transmit them towards a broadcasting center via the single antenna.
We assume that the used relay in this system is an ideal decode-and-forward relay with
zero latency. As presented in Fig. III.2(c), the proposed broadcasting center has multiple
antennas which are facing the relays. We emphasize that due to the narrow beamwidth
(1.5-10 degree [3, 4]), of the antennas, multiple streams arriving at the broadcast center
39
do not interfere with each other (and similarly for the relays). This broadcasting center
selects important features of the current real-time sports game. For interactive TV where
users can select the camera/viewpoint they prefer, it is often desirable to maximize the
overall video quality, subject to constraints on the minimum acceptable quality for each
video stream.
III.4 Joint Scalable Coding and Routing
Fig. III.3 shows the system model with a set of sources S, a set of relays R, and a single
destination. In the relay-destination combination (RDC) part of Fig. III.3, all relays
(i.e., r
1
; ;r
N
R
) are connected to the single destination (i.e., the broadcasting center
D) where N
R
stands for the number of relays. Then the maximum achievable rates of
all possible relay-destination pairs are computed (denoted as a
RDC
r
1
!D
; ;a
RDC
r
N
R
!D
). We
assume that the destination can form a sucient number of independent beams so that
it has no limitations concerning the number of relays from which it can receive. Thus,
nding optimal combinations between sources and relays in SRC are considered for the
following three scenarios: (i) sources and relays have only a single beam (Section III.4.1),
(ii) sources have single beams and relays have multiple beams (Section III.4.2), and (iii)
both sources and relays can form multiple beams (Section III.4.3).
For all possible scenarios, our objective is the maximization of the sum of the overall
video qualities delivered to the destination. As a rst step, the relationship between the
video qualities and data rates should be dened. The quality of HD video signals is related
to the data rate in a nonlinear and monotonically increasing form, e.g., logarithmically.
One widely suggested model [133, 85, 124] which is applicable to H.264/MPEG4 AVC,
40
is presented in Fig. VI.4. In addition, there is no compression loss if the data rate is
more than 1:5 Gbit/s because we can exploit uncompressed HD video transmission in a
standard mode as shown in Fig. VI.4. We note, however, that this gure might depend
on the type of video source - e.g., will be dierent for fast-moving and slow-moving video.
The main feature we include in our modeling are a monotonic, but sublinear, increase of
video quality with data rate.
III.4.1 Single-Beam Antennas at Sources and Relays
Each source has a single-beam antenna and thus can send only to one relay, in addi-
tion, each relay also has a single-beam antenna for receiving data. Our objective is the
maximization of the sum of video qualities delivered from sources to the destination:
max :
X
N
S
i=1
f
q

X
N
R
j=1
1
2
a
SRC
s
i
!r
j
x
SRC
s
i
!r
j

(III.6)
subject to
X
N
S
i=1
a
SRC
s
i
!r
j
x
SRC
s
i
!r
j
A
RDC
r
j
!D
;8j; (III.7)
X
N
S
i=1
x
SRC
s
i
!r
j
1;8j; (III.8)
X
N
R
j=1
x
SRC
s
i
!r
j
1;8i; (III.9)
a
s
i

X
N
R
j=1
a
SRC
s
i
!r
j
x
SRC
s
i
!r
j
;8i; (III.10)
a
SRC
s
i
!r
j
A
SRC
s
i
!r
j
;8i;8j; (III.11)
x
SRC
s
i
!r
j
2 f0; 1g;8i;8j; (III.12)
x
RDC
r
j
!D
= 1;8j; (III.13)
41
Figure III.4: Generalized relationship between the quality index of transmitted HD video
signals and data rates. Based on dierent kinds of HD video sources, the curve can be
varied, but the general form is given as logarithmically and monotonically increasing as
proved in [133, 85, 124].
where N
S
stands for the number of sources and relays, respectively.
In this formulation,i andj are the indices of sources and relays wherei2f1; ;N
S
g
and j2f1; ;N
R
g where S and R stand for the sets of sources and relays, respectively.
If the connection between s
i
and r
j
is active (i.e., if source s
i
selects relay r
j
), then the
binary index variablex
SRC
s
i
!r
j
is 1 by (III.12); otherwise it is 0 by (III.12). The given relays
should be connected to the destination D, thus, x
RDC
r
j
!D
= 1 by (III.13). The A
SRC
s
i
!r
j
and
A
RDC
r
j
!D
are maximum achievable rates for the corresponding wireless links and computed
by (III.1). In addition, the desired data rates between s
i
and r
j
, i.e., a
SRC
s
i
!r
j
, should be
less than or equal to the computed A
SRC
s
i
!r
j
as shown in (III.11). As shown in (III.10), we
have to achieve the required minimum data rates for each
ow (i.e., a
s
i
,8s
i
) to guarantee
the required minimum video qualities for each
ow (i.e., f
q

a
s
i

,8s
i
) where f
q
() is a
function for the relationship between video quality and data rate (refer to Fig. VI.4). In
42
addition, A
SRC
s
i
!r
j
froms
i
tor
j
andA
RDC
r
j
!D
fromr
j
toD are xed values because the sources
and relays are not mobile and the channel is not time-varying.
Since each relay can receive one video stream, and these have to go to the destination
via the wireless link with a limited capacity A
RDC
r
j
!D
, (III.7) follows. For each individual
source, there is at most one outgoing
ow toward relays because the sources have single-
beam antennas, as formulated in (III.9). Similarly, each relay can form only one beam
in receiving mode, thus the number of incoming
ows from sources can be at most one
as formulated in (III.8). Finally, (III.6) describes the objective of nding the set of pairs
between sources and relays as well as nding the corresponding data rates for maximizing
the total video quality and the corresponding data rate value becomes half due to half-
duplex constraint.
The set of equations (III.6 - III.13) can be solved by the method of Section III.5. Al-
ternatively, we note that this setup is a special case of a scenario treated in our conference
paper [139]. In that paper, the proposed algorithm addresses the relay selection and coop-
erative communication scheme selection for a situation where several source-destination
unicast pairs exist, and furthermore transmission between the pairs can occur not only
using decode-and-forward, but also amplify-and-forward (AF), and non-cooperative com-
munications (non-CC)) direct transmission. In the system conguration of [139], an
algorithm selects the relay node and transmission mode for every single unicast pair in
terms of maximization of overall transferred video qualities. Thus, our mathematical
formulation, which shows that the connectivity matrix is totally unimodular, which in
turn guarantees polynomial-time solutions (i.e., a closed-form solution is possible), can be
applied also in this case. On the other hand, this framework cannot be easily generalized
to the multi-beam scenarios.
43
The proposed scheme in this section is meaningful under the assumption that the
number of relays is larger than or equal to the number of sources; otherwise, some video

ows from sources cannot reach to the destination. However, positioning many relays
obviously increases the cost of the network. To deal with this problem, more advanced
relay architectures, which allow multiple-beam antennas, are proposed and the schemes
for this case are discussed in the following two sections.
III.4.2 Single-Beam Antennas at Sources and Multiple-Beam Antennas
at Relays
Each source has a single-beam antenna and thus can send only to one relay, while relays
can aggregate streams from dierent sources. Therefore, the relays have multiple-beam
antennas, the constraint (III.8) is updated to allow multiple incoming
ows as follows:
X
N
S
i=1
x
SRC
s
i
!r
j
B
r
j
;8j; (III.14)
where B
r
j
stands for the number of antenna-beams at relay j;8j2f1; ;N
R
g; in the
following we will assume B
r
j
=jSj. Thus the corresponding formulation for maximizing
overall qualities of received HD video streams from sources to a destination is as follows:
max :
X
N
S
i=1
f
q

X
N
R
j=1
1
2
a
SRC
s
i
!r
j
x
SRC
s
i
!r
j

(III.15)
subject to (III.7), (III.9), (III.10), (III.11), (III.12), (III.13), and (III.14).
44
III.4.3 Multiple-Beam Antennas at Source and Relays
When the sources have multiple-beam antennas, the constraint (III.9) is updated to allow
multiple outgoing
ows for all sources as follows:
X
N
R
j=1
x
SRC
s
i
!r
j
B
s
i
;8i: (III.16)
where B
s
i
stands for the number of antenna-beams at source i;8i2f1; ;N
S
g; in the
following we will assume B
s
i
=jRj.
Summarizing, the mathematical formulation can be again written as follows:
max :
X
N
S
i=1
f
q

X
N
R
j=1
1
2
a
SRC
s
i
!r
j
x
SRC
s
i
!r
j

(III.17)
subject to (III.7), (III.10), (III.11), (III.12), (III.13), (III.14), and (III.16).
III.4.4 Discussion
In some cases direct transmission from source to a broadcasting center can guarantee bet-
ter video quality than transmission via relay. This can be incorporated in our framework
by placing a virtual relay (denoted as r
(v;j)
in this subsection) at a location very close to
the destination, and letting the capacity between the relay and a broadcasting center be
innity (i.e., A
RDC
r
(v;j)
!D
=1), while the achievable capacity between source and relay is
2A
SRC
s
i
!r
(v;j)
, where the factor 2 re
ects the fact that there is no half-duplex penalty in
direct transmission.
45
III.5 Re-Formulation: Convex Form
The proposed three mathematical formulations can be non-convex mixed-integer nonlinear
programs (MINLP) even though the quality function has a convex form (Fig. VI.4) as
shown in following theorem.
[Theorem1]: The three optimization formulations in Section III.4 can be a non-convex
MINLP.
[Proof 1]: If there exists a quality function which has logarithmically and mono-
tonically increasing property (Fig. VI.4) which can make our designed formulation be
non-convex MINLP, then the corresponding formulation is non-convex. Then, the follow-
ing equation is a possible quality index function:
f
q
(a) =
1
log

(a
max
+ 1)
log

(a + 1) (III.18)
is a base (1<), a
max
is a desired data rate for uncompressed video transmission (1:5
Gbit/s in a standard mode), anda is a given data rate. This proof considers the scenario
of one-source and one-relay, which is the simplest case. In this case the objective function
becomes
f

a
SRC
s
i
!r
j
;x
SRC
s
i
!r
j

,f
q

a
SRC
s
i
!r
j

x
SRC
s
i
!r
j
= K log

a
SRC
s
i
!r
j
+ 1

x
SRC
s
i
!r
j
(III.19)
where K =
1
log

(amax+1)
is a constant and 8i 2 f1 ;N
S
g;8j 2 f1; ;N
R
g. To
show that this given equation is non-convex, the second-order Hessian of this given
real function should be non positive denite [12]. The Hessianr
2
f

a
SRC
s
i
!r
j
;x
SRC
s
i
!r
j

is:
46
2
6
6
4
0
K
ln

1
a
SRC
s
i
!r
j
+1
K
ln

1
a
SRC
s
i
!r
j
+1
x
SRC
s
i
!r
j

K
ln

1
a
SRC
s
i
!r
j
+1

2
3
7
7
5
and then the corresponding two eigenval-
ues are
1
2
M
1
2
8
<
:
M
2
+ 4

K
ln

1
a
SRC
s
i
!r
j
+ 1
!
2
9
=
;
0:5
(III.20)
where M =x
SRC
s
i
!r
j

K
ln

1
a
SRC
s
i
!r
j
+1

2
, 0 a
SRC
s
i
!r
j
1:5, and 0 x
SRC
s
i
!r
j
1. These
values are not all positive, which shows that the Hessian is not positive denite, which
proves that the optimization function is non-convex. [End of Proof 1]
For non-convex MINLP, heuristic searches can nd approximate solutions but can-
not guarantee optimality. Among well-known approximation algorithms, branch-and-
rene based algorithms show relatively better performance for non-convex MINLP prob-
lems [175]. The detailed procedure of the branch-and-rene based algorithms is as fol-
lows: First, the integer terms are relaxed (relaxation), i.e., x
SRC
s
i
!r
j
2f0; 1g is converted
to 0x
SRC
s
i
!r
j
1. After that, the special ordered sets (SOS) approximation is used for a
linear approximation. This segments the multi-dimensional space of the given objective
function into multiple small triangular regions, each of which is plane; in other words,
the triangle regions approximate the multi-dimensional surface of the objective function.
For every single triangle region, optimum solutions can be obtained by running a simplex
based algorithm. Then, we run the branch-and-bound algorithm to nd integer solutions
for x
SRC
s
i
!r
j
for each single triangle region. Thus nally the optimum value is selected
among the solutions on the triangles. However, the branch-and-rene algorithm cannot
guarantee the optimum solutions in a non-convex MINLP. First of all, if the segmentation
into plane triangle regions is rough, then the approximated planes are not precise enough
to get the optimum solutions. If the segmentation into plane triangle regions is too ne,
47
the runtime becomes excessive, since the simplex algorithm should be operated for every
single triangular region. In conclusion, using branch-and-rene based algorithm provides
an approximation but cannot guarantee the optimum solutions and, moreover, this algo-
rithm cannot nd bounds on the approximation errors. We thus introduce the following
Theorem, with which our non-convex MINLP can be re-formulated as a convex program.
[Theorem 2]: For the given non-convex MINLP formulation in Section III.4, intro-
ducing
a
SRC
s
i
!r
j
A
SRC
s
i
!r
j
x
SRC
s
i
!r
j
;8i;8j (III.21)
instead of (III.11) makes the formulation convex.
[Proof 2]: For the non-convex MINLP formulation in Section III.4, x
SRC
s
i
!r
j
= 0 means
the link is disconnected. Thus the corresponding rate becomes 0 and (III.21) leads to the
same result when x
SRC
s
i
!r
j
= 0, i.e.,
a
SRC
s
i
!r
j
A
SRC
s
i
!r
j
0;8i;8j; (III.22)
thus,
a
SRC
s
i
!r
j
0;8i;8j: (III.23)
and then a
SRC
s
i
!r
j
is equal to 0 because a
SRC
s
i
!r
j
is non-negative. Otherwise, if x
SRC
s
i
!r
j
= 1,
then this term is equivalent to (III.11). Therefore, in turn, (III.6), (III.15), (III.17) are
also updated as
max :
X
N
S
i=1
f
q

X
N
R
j=1
1
2
a
SRC
s
i
!r
j

; (III.24)
48
(III.7) is updated as follows
X
N
S
i=1
a
SRC
s
i
!r
j
A
RDC
r
j
!D
;8j; (III.25)
and (III.10) is also updated as follows
a
s
i

X
N
R
j=1
a
SRC
s
i
!r
j
;8i: (III.26)
Then there are no non-convex terms in the proposed programs. [End of Proof 2]
Summarizing, the optimization problem can be reformulated as follows. For the single-
beam antennas at sources and relays
max :
X
N
S
i=1
f
q

X
N
R
j=1
1
2
a
SRC
s
i
!r
j

(III.27)
subject to
X
N
S
i=1
a
SRC
s
i
!r
j
A
RDC
r
j
!D
;8j; (III.28)
X
N
S
i=1
x
SRC
s
i
!r
j
1;8j; (III.29)
X
N
R
j=1
x
SRC
s
i
!r
j
1;8i; (III.30)
a
s
i

X
N
R
j=1
a
SRC
s
i
!r
j
;8i; (III.31)
a
SRC
s
i
!r
j
A
SRC
s
i
!r
j
x
SRC
s
i
!r
j
;8i;8j; (III.32)
x
SRC
s
i
!r
j
2 f0; 1g;8i;8j; (III.33)
x
RDC
r
j
!D
= 1;8j; (III.34)
49
where8i2f1; ;N
S
g;8j 2f1; ;N
R
g. In addition, for the single-beam antennas
at sources and multiple-beam antennas at relays case, the objective function is equiva-
lent to (III.27) and the corresponding constraints are (III.28), (III.30), (III.31), (III.32),
(III.33), (III.34), and (III.14) where8i2f1; ;N
S
g;8j2f1; ;N
R
g and nally for the
multiple-beam antennas at source and relays case, the objective function is equivalent to
(III.27) and the corresponding constraints are (III.28), (III.31), (III.32), (III.33), (III.34),
(III.14), and (III.16) where8i2f1; ;N
S
g;8j2f1; ;N
R
g.
With these given convex programs, the given integer constraints, i.e., x
SRC
s
i
!r
j
2f0; 1g
are relaxed as 0 x
SRC
s
i
!r
j
1,8i2f1; ;N
S
g;8j2f1; ;N
R
g. Then the convex
programs are solved using CVX which is the most popular matlab-based software for
solving convex optimization problems [177].
III.6 Performance Evaluation
To verify the superior performance of our proposed scheme, i.e., a joint HD video coding
rate decision and relay selection/routing scheme under the consideration of overall video
quality maximization (named VQM), we compare it with the following two schemes:
The joint HD video coding rate decision and relay selection/routing scheme under
the consideration of sum rate maximization. In this case, the proposed objective
function, i.e., (III.27), should be updated as follows:
max :
X
N
S
i=1

X
N
R
j=1
1
2
a
SRC
s
i
!r
j

(III.35)
50
due to the fact that the quality function (Fig. VI.4) is no longer considered. Note
that the half-duplex constraint is still existing. This method is named as SRM, i.e.,
sum rate maximization.
The scheme proposed in [92], which is an ecient algorithm that considers joint
relay selection/routing (called JRSR) and rate allocation at the same time. In order
to enable fair comparisons, we adapt the scheme to our outdoor-stadium architec-
ture (one-tier relay) and allow only decode-and-forward relaying. Lastly, Ref. [92]
has the same number of sources and destinations; however, to unify the setup for
performance comparison, all the given destinations are located at the same location
and operate as a single destination with multiple antenna elements.
With these given three schemes, i.e., VQM, SRM, JRSR, overall delivered video quality
values are simulated in the reference models shown in Fig. III.5. The sources (HD cam-
eras) are uniformly distributed on top of the stadium. Between stadium and broadcasting
center, multiple relays are uniformly deployed along a line. To vary the simulation sce-
narios, we consider this line to be near the sources (Scenario A), in the middle between
sources and broadcast center (Scenario B), and near the broadcast center (Scenario C).
Lastly, we also consider a scenario where the relay positions are uniformly randomly dis-
tributed. In Scenario A, there is a higher probability that the relay- to - destination
link might be the bottleneck, while Scenario C obviously has the source-relay link as its
bottleneck.
In addition, for the simulation studies with multiple antenna-beams at sources or
relays, the number of beams at relays and the number of beams at sources are set as N
S
and N
R
, respectively.
51
Table III.2: Expectation of Achieved Aggregated Video Quality for Single-Beam Antennas
at Sources and Relays Case (Values: the objective function results with optimal solutions)
N
S
N
R
Scenario VQM SRM JRSR
5 10 A 4.15 3.33 3.33
5 10 B 4.92 4.17 4.17
5 10 C 4.41 3.63 3.63
5 10 Random 4.43 3.65 3.65
10 5 A 4.15 3.35 3.35
10 5 B 4.92 4.18 4.18
10 5 C 4.41 3.65 3.65
10 5 Random 4.43 3.67 3.67
10 10 A 8.74 5.48 5.48
10 10 B 9.54 6.34 6.34
10 10 C 8.95 5.79 5.79
10 10 Random 9.17 5.20 5.20
10 15 A 8.78 5.63 5.63
10 15 B 9.79 6.46 6.46
10 15 C 9.18 5.93 5.93
10 15 Random 9.19 5.85 5.85
15 10 A 9.11 6.48 6.48
15 10 B 9.89 7.38 7.38
15 10 C 9.37 6.84 6.84
15 10 Random 9.39 6.85 6.85
52
Broadcasting Center
200 m
500 m 50 m 50 m
• Scenario A: d = 100 m
• Scenario B: d = 250 m
• Scenario C: d = 400 m
Relay
Deployment Range
(500 – d) m d m
……
Figure III.5: Performance Evaluation Simulation Setup: Wireless HD video cameras are
uniformly distributed on top of the stadium. There is one broadcasting center at bottom.
Between wireless video cameras on top of stadium and broadcasting center, relays are
uniformly and linearly deployed. To vary simulation setting, the deployment of relays
has three dierent types: the relays are distributed near cameras (Scenario A), in the
middle of cameras (on top of stadium) and broadcasting center (Scenario B), and near
broadcasting center (Scenario C).
As our performance measure we consider the cumulative probability distribution of the
aggregate video quality. We will show results for a xed number of relays (N
R
= 10) and
various number of sources (N
S
= 5; 10; 15) at rst (Section III.6.1). The cdf is obtained
as follows: we consider multiple realizations of the deployment of sources and relays (for a
xed scenario and number of relays, but random relay location according to the location
pdf of a given scenario). For each such realization, we optimize coding rates and relay
selection; thus each run gives us one realization of the aggregate video quality. We nally
plot the cdf of this quality. For the simulation of VQM, the lower bounds of each
ow
are set as 0:75, i.e., all a
s
i
where8i2f1; ;N
S
g are all set to 0:75 (Unit: Gbit/s) in
53
a standard mode (i.e., 50% of 1:5 Gbit/s). Later, Section III.6.2 shows the performance
behavior if various lower bound settings are observed.
III.6.1 CDF of Aggregate Video Quality { Fixed Number of Relays and
Various Number of Sources
Single-Beam Antennas at Sources and Relays
Fig. III.6 presents the cases that the number of sources is smaller, equal, or larger than
the number of relays (i.e., N
S
= 5; 10; 15, and N
R
= 10). We see that SRM and JRSR
show the same performance because they are equivalent for the given constraints of single-
beam antennas at sources and relays. For the case of 5 sources, we also see that with
the proposed VQM, the aggregate video quality is within 5% of its maximum for 83% of
simulation runs in Scenario A, 98% of simulation runs in Scenario B, 88% of simulation
runs in Scenario C and 89% of random deployment. Note that the given number of
sources is 5, thus, the maximum achievable video quality is 5 due to the fact that the
maximum video quality index in each
ow is normalized as 1 (refer to Fig. VI.4). We
also nd that deployment scenario Scenario B can obtain more quality than the others
deployment scenarios since it best balances capacity constraints on the source-relay and
relay-destination links. The mean achieved aggregate video qualities are also give in
Table III.2. For the case of 10 or 15 sources, the maximum achievable aggregated video
quality is 10 because the given number of relays is 10 which takes a role of threshold of
the delivered quality.
54
Table III.3: Expectation of Achieved Aggregated Video Quality for Single-Beam Antennas
at Sources and Multiple-Beam Antennas at Relays Case (Values: the objective function
results with optimal solutions)
N
S
N
R
Scenario VQM SRM JRSR
5 10 A 4.15 3.85 3.33
5 10 B 4.92 4.60 4.17
5 10 C 4.65 4.35 3.63
5 10 Random 4.57 4.27 3.65
10 5 A 4.16 3.86 3.35
10 5 B 4.95 4.61 4.18
10 5 C 4.66 4.36 3.65
10 5 Random 4.57 4.28 3.67
10 10 A 8.81 8.46 5.48
10 10 B 9.81 9.46 6.34
10 10 C 9.31 8.96 5.79
10 10 Random 9.29 8.96 5.20
10 15 A 8.86 8.55 5.63
10 15 B 9.86 9.55 6.46
10 15 C 9.36 9.05 5.93
10 15 Random 9.36 9.04 5.85
15 10 A 13.41 11.70 6.48
15 10 B 14.88 13.20 7.38
15 10 C 14.38 12.71 6.84
15 10 Random 14.22 12.53 6.85
55
(a) Scenario A (b) Scenario B
(c) Scenario C (d) Random
Figure III.6: Simulation Result for Single-Beam Antennas at Sources and Relays: Various
Number of Sources (N
S
= 5; 10; 15) and Fixed Number of Relays (N
R
= 10)
Single-Beam Antennas at Sources and Multiple-Beam Antennas at Relays
Fig. III.7 and Table III.3 present data similar to Fig. III.6 and Table III.2, but now
with multiple-beam antennas at the relays, so that the relays can aggregate and combine
video streams from dierent sources. In this case, there exists a performance dierence
between SRM and JRSR. The latter, by design, does not allow the exploitation of the
multiple-beam antennas at relays, and thus shows worse performance. As a matter of
56
fact, it has the same performance as with single-beam antennas at sources and relays
case if the network conguration is equivalent. This fact will even holds in the case of
multiple-beam antennas at sources and relays. We furthermore see that again SRM shows
lower performance than VQM due to the fact that SRM aims for maximization of sum
data rates, while VQM aims to maximize the overall delivered video quality. We also see
that for the case that the number of relays is sucient (larger than or equal to the number
of sources), the achieved video quality is not fundamentally dierent from the case with
single-beam antennas at the relays. However, this changes when the number of relays is
not sucient (i.e.,N
S
= 15,N
R
= 10). We can now achieve aggregate video quality larger
than 10 as some streams can be compressed with little video quality loss and forwarded
by the same relay. Nonideal performance is mainly caused by the capacity limitations
of the relay-destination links. These limitations have more impact on scenarios A and B
than in Scenario C.
Multiple-Beam Antennas at Sources and Relays
For multiple-beam antennas at both sources and relays, VQM and SRM have better per-
formance than the case of single-beam antennas at sources, while JRSR has the same
performance as in the case of single-beam antennas as described above. With the benet
of multiple-beam antennas, SRM and VQM have better performance than the other two
cases as shown in Fig. III.8 and Table III.4; however, the performance gain is minor. As
in the single-beam case at the source, the achieved aggregated video quality is limited by
the capacity between relays and a destination.
57
Table III.4: Expectation of Achieved Aggregated Video Quality for Multiple-Beam An-
tennas at Sources and Relays Case (Values: the objective function results with optimal
solutions)
N
S
N
R
Scenario VQM SRM JRSR
5 10 A 4.16 3.8734 3.3313
5 10 B 4.93 4.6469 4.1650
5 10 C 4.68 4.3969 3.6315
5 10 Random 4.59 4.3057 3.6534
10 5 A 4.16 3.87 3.35
10 5 B 4.95 4.62 4.18
10 5 C 4.66 4.37 3.65
10 5 Random 4.58 4.29 3.67
10 10 A 8.81 8.45 5.48
10 10 B 9.82 9.45 6.34
10 10 C 9.31 8.96 5.79
10 10 Random 9.31 8.85 5.20
10 15 A 8.88 8.57 5.63
10 15 B 9.87 9.56 6.46
10 15 C 9.38 9.07 5.93
10 15 Random 9.38 9.07 5.85
15 10 A 13.42 11.77 6.48
15 10 B 14.90 13.25 7.38
15 10 C 14.40 12.76 6.84
15 10 Random 14.24 12.59 6.85
58
(a) Scenario A (b) Scenario B
(c) Scenario C (d) Random
Figure III.7: Simulation Result for Single-Beam Antennas at Sources and Multiple-Beam
Antennas at Relays: Various Number of Sources (N
S
= 5; 10; 15) and Fixed Number of
Relays (N
R
= 10)
III.6.2 Impact of Lower Bound Setting
In the previous simulations, the lower bounds for the data rate per data stream (corre-
sponding to the desired lower bound on the video quality) are set as 0:75 Gbit/s. In this
section, we vary now this lower bound values from 0 Gbit/s (i.e., there is no lower bound)
up to 1:5 Gbit/s (i.e., we allow only uncompressed HD video transmission) in steps of 0:1
59
(a) Scenario A (b) Scenario B
(c) Scenario C (d) Random
Figure III.8: Simulation Result for Multiple-Beam Antennas at Sources and Relays: Var-
ious Number of Sources (N
S
= 5; 10; 15) and Fixed Number of Relays (N
R
= 10)
Gbit/s. As a performance quality measure we dene "stream outage", i.e., the probability
that at least one stream does not have the minimum required quality. Obviously, this
outage has to increase (and thus its complement, the probability of successful transmis-
sion, has to decrease) as the minimum video quality increases. We furthermore anticipate
that VQM will be better able to handle the increased requirements, as it is more
exible.
60
Figure III.9: Simulation Result for Various Lower Bound Setting: Single-Beam Antennas
at Sources and Multiple-Beam Antennas at Relays, N
S
= 10;N
R
= 15
This evaluation is performed for the case of single-beam antennas at the sources and
multiple-beam antennas at relays when N
S
= 10;N
R
= 15 in Fig. III.9.
As shown in Fig. III.9, the case of relay deployment with Scenario C suers signicantly
from the higher required per-stream quality. With Scenario C relay deployment, the data
rates between sources and relays are lower than in the other cases. Thus, when we set
the lower bound quite high, all
ows are disconnected. Thus, it achieves the lowest
performance. On the other hand, in Scenario A relay deployment, all
ows between
sources and relays have enough capacity to support uncompressed video transmission,
thus, a higher setting for minimum quality does not have such a strong impact. In
addition, Fig. III.9 shows that VQM has better performance than SRM for all possible
three types of relay deployment. For more details, the average normalized video quality
for VQM is 0:9531, 0:9138, 0:7288 for Scenario A, B, C, respectively. SRM cannot achieve
maximum aggregated video quality due to the fact that it maximizes the overall data
rates instead of overall delivered video qualities.
61
III.7 Concluding Remarks
This chapter addresses a joint scalable coding and routing scheme for a 60 GHz mm-wave
HD video streaming in an outdoor sports stadium broadcasting system. In the system,
there are multiple wireless HD video cameras distributed throughout the stadium. To
transmit the HD video data from the cameras over the wireless channels in a real-time
manner, 60 GHz wireless links are used because they can exploit multi-Gbit/s wireless data
transmission. However, according to the high path loss of 60 GHz links, relays are used to
extend communication coverage. We presented an algorithm for nding the combination
of wireless link pairs between wireless HD video cameras and relays that can maximize the
overall or per-
ow video qualities of delivered HD video streams to one single broadcasting
center. This chapter considers three kinds of cases, i.e., single-beam antennas at sources
and relays, single-beam antennas at sources and multiple-beam antennas at relays, and
nally, multiple-beam antennas at sources and relays. For each cases, the given problem
is initially formulated as non-convex MINLP and it is re-formulated as convex program,
which allows optimum solutions.
We demonstrated that our algorithm outperforms algorithms based on sum-rate maxi-
mization and other well-known methods in the literature, for a variety of relay deployment
scenarios, number of sources, and number of relays.
62
Chapter IV
Fast Millimeter-Wave Beam Training with
Receive Beamforming
IV.1 Introduction
Mm-wave data transmission has (re-) emerged as a highly promising approach to achieving
Gigabit/s throughput in wireless communications links. One application lies in cable re-
placement for in-home entertainment systems (Wireless HDMI or Wireless High-Denition
TV (HDTV) audio/video applications) and other personal-area networks that operate in-
doors over short distances (less than 20 meters) [1]. Mm-waves can also be used for
multi-Gigabit/s outdoor transmission, either as directional links for wireless backhaul, or
as access technology for transmissions between base stations (BSs) and mobile stations
(MSs) as a part of 5G research [108, 139, 151, 169, 170].
Due to the extremely high carrier frequencies, mm-wave signals experience high pathloss,
but are also highly directional [89]. High-gain antennas are thus required to obtain rea-
sonable signal-to-noise ratios (SNRs). For all applications where the propagation channel
(both indoor and outdoor) could change, an array antenna with adaptive beamforming is
63
required. Yet, determining the optimum beam direction is non-trivial. Mm-wave adaptive
arrays typically can (through RF-phase shifters) only \point" the beam in dierent di-
rections, so that nding the optimum direction essentially requires \sweeping" the beam
through all possible directions; the sweep has to be done in steps of the beamwidth.
Such an exhaustive search has been widely considered in the literature [53] and is fore-
seen in mm-wave standards such as IEEE 802.15.3c wireless personal area networks
(WPAN) [16, 31, 77] and IEEE 802.11ad wireless local area networks (WLAN) [20]. Since
the beamwidth for mm-waves can be as low as 1

, the overhead for the beam training is
signicant [3, 4].
To alleviate this problem, the current chapter proposes adaptive fast beam train-
ing protocols for both xed and adaptive modulation. For xed modulation, this chapter
addresses a fast link conguration protocol between beamforming initiator (BI) and beam-
forming responder (BR) using (i) interactive beam training and (ii) prioritized beam search
space ordering. For interactive beam training, the proposed protocol modies the beam
training procedure such that it immediately terminates the procedure when both BI and
BR nd their optimal beam directions. For this purpose, each BI and BR changes its
communication mode between transmitter (Tx) and receiver (Rx) in every single oper-
ation. Then, each BI and BR can give feedback to its opposite side when it nds an
"optimum" direction. We note that the optimum might be a local (not global) optimum,
but this has no impact on performance if the local optimum is \good enough" for the
considered modulation format. For prioritized beam search space ordering, we employ
long-term statistics, in particular the frequency of network association request/response
statistics. We know that due to the special characteristics of 60 GHz propagation (very
low diraction and through-wall transmission), certain angular ranges can be completely
64
blocked. Ordering of directions in terms of network association request/response statistics
gives an indication about such blocked regions, which consequently would not be searched
with high priority. For adaptive modulation, stopping before evaluating all beam direc-
tions is not benecial because it might miss the global optimal solution. Therefore, we
propose to perform an exhaustive search by starting out with a coarse-grained search that
is successively rened (multi-level beam training). We investigate the optimal numbers of
sectors for coarse-grained training and division orders for ne-grained training in terms
of fast operation for a given beamwidth.
The contributions of this chapter are as follows: (i) for xed modulation wireless
transmission, we design an interactive beam training algorithm that can dramatically
reduce the beam search space and thus the overhead and (ii) for adaptive modulation
wireless transmission, we additionally designed iterative beam training algorithm that
can also reduce overall beam training time. Simulations using standardized systems and
channel models conrm the eectiveness of our approach.
The remainder of this chapter is organized as follows: Sec. VI.2 presents the prelim-
inaries such as the basic properties of mm-wave wireless transmission, as well as related
work. For fast link conguration, Sec. IV.3 and Sec. IV.4 explain the details of our pro-
posed interactive and iterative beam training protocols for xed and adaptive modulation,
respectively. Sec. IV.5 simulates our proposed protocols in mobile wireless applications.
Sec. V.6 concludes this chapter and presents future research directions.
65
IV.2 Preliminaries
IV.2.1 Mm-wave Wireless Packets
According to the Friis path-loss model, the received power in free space is computed as
P
Rx
=G
Tx
G
Rx

c=f
c
4d

2
P
Tx
(IV.1)
whereP
Tx
,P
Rx
,G
Tx
,G
Rx
,c,d, andf
c
stand for the transmit power, receive power, trans-
mit antenna gain, receive antenna gain, speed of light, distance, and carrier frequency,
respectively. It can be seen that in order to compensate for the 1=f
2
c
term, a high antenna
gain is required. Many wireless propagation channel measurements have found that mm-
wave channels are highly directional, i.e., if transmission occurs in a particular direction,
signals are arriving at the receiver only from a few directions (and vice versa) [89, 21, 28].
This is in marked contrast to traditional microwave propagation, which is characterized
by a large angular dispersion [14]. Furthermore, the small wavelength allows to create
(adaptive or non-adaptive) antennas with high gain yet reasonable size: the beamwidths
of current mm-wave commercial antenna solutions range from 1

to 15

[3, 4]. We note
here that fully adaptive antennas that can perform digital beamforming allow very fast
beam training, since they can receive a training signal at all antenna elements, and thus
determine the optimum beam pattern in a single step. However, such antennas need a ra-
dio frequency (RF) chain for each antenna element, which is too costly for most consumer
applications. We are thus concentrating on antenna arrays that have only a single down-
conversion chain, plus RF phase shifters, and thus can shift the direction of a main beam
66
......
1
N
1
N
......
BI BR
FB
FB
(a) TxBF
BI BR
1
N
......
......
1
N
(b) RxBF
Figure IV.1: A General Beamforming Procedure for Link Conguration
in (adjustable) increments, but can only receive with a single beampattern/direction at a
time.
IV.2.2 Related Work
General Beam Training { Exhaustive Search
The general beamforming and training procedure using transmit beamforming (TxBF) is
illustrated in Fig. IV.1(a) [53]. In Fig. IV.1(a), each BI and BR has N beam directions.
To initiate the beam training procedure, the BI sweeps through all beam directions,
transmitting one training packet in each direction. During this time, BR receives the
packets with an omni-directional antenna pattern. At the end of this period, the BR
can identify which beam direction of the BI resulted in the highest SNR at the BR.
Subsequently, BI and BR exchange their roles and repeat the procedure, allowing the BI
to determine the direction of the BR leading to the highest SNR. A last step of exchanging
67
feedback packets allows both sides to learn their own optimal directions. A variant of this
approach uses receive beamforming (RxBF) instead of TxBF as illustrated in Fig. IV.1(b).
In Fig. IV.1(b), each node BI and BR has N beam directions. The BI transmits packets
in omni-directional mode, while the BR scans through all directions; then BI and BR
exchange roles. The two nodes then know their optimum beam directions without further
exchanging of feedback packets. Due to this advantage, and because Rx beamforming is
not impacted by constraints on EIRP (equivalent isotropically radiated power), but only
on transmitted absolute power, we henceforth consider RxBF in this chapter. As explained
before, the beamwidth of commercial mm-wave high-gain Cassegrain antennas is near 1

,
and similar values can be achieved with adaptive antennas of realistic size. Thus, in the
worst case, N should be 360 for 2D beam geometry and 360 180 6:48 10
4
for a
3D beam geometry. Consequently, the beam training procedures can require a signicant
overhead.
Beam Training Techniques in Mm-Wave IEEE Standards
This section reviews current standardized mm-wave beam training techniques. Specically
we consider the two standards for 60 GHz wireless communications, i.e., IEEE 802.15.3c
WPAN [31] and IEEE 802.11ad WLAN [20].
In the IEEE 802.11ad WLAN and IEEE 802.15.3c WPAN beamforming and train-
ing [16, 20, 31, 77], the protocols use a two-stage beamforming and training operation, i.e.,
coarse-grained beam training (named sector sweeping in IEEE 802.11ad and low-resolution
(L-Re) beam training in IEEE 802.15.3c) and ne-grained beam training (named beam re-
nement in IEEE 802.11ad and high-resolution (H-Re) beam training in IEEE 802.15.3c).
If the protocols consider TxBF, BF and BI determine the optimum coarse-grained beam
68
according to the exhaustive-search protocol described in Sec. IV.2.2. In the next stage,
ne-grained beam training, the same type of protocol is used to identify the best beam in
each coarse-grained beam. Similar principles hold when Rx beamforming is used; however,
the IEEE 802.15.3c WPAN standard does not consider RxBF. Even if both standards have
their own specic beamforming and training protocols, the protocols are fundamentally
based on two-level beam training. While this can accelerate the beamforming, it is still
slow, as demonstrated in Sec. IV.5.
In addition, we point out that nding the \best" direction at the Rx when the Tx
is omni-transmitting (i.e., the currently used algorithms for exhaustive search, two-stage
beam training, and our proposed scheme) might not be the same thing as nding the best
Rx-direction when a joint Tx-Rx optimization has been done. Note that current existing
mm-wave beam training protocols including IEEE 802.11ad and IEEE 802.15.3c have a
similar restriction as our algorithm.
IV.3 Interactive Beam Training for Fixed Modulation
IV.3.1 Interactive 2D Beam Training
Concepts
Our proposed beam training in a 2D geometry is designed based on two concepts, i.e., (i)
interactive beam training and (ii) prioritized sector search ordering. Fig. IV.2 shows the
fundamental process of the proposed two-level beam training. Prioritized sector search
ordering orders the sectors in terms of network association statistics at rst, in turn,
interactive beam training is used for nding desired beam directions within the sorted
sectors/beams.
69
Interactive Beam Training: An important source of ineciency in exhaustive
beam training lies in the fact that even when a BI or BR nds a good beam direction, it
cannot stop because it has to search all given beam directions. Of course, in order to nd
the globally optimum direction, a complete search is necessary. However, when a xed
modulation/coding scheme is used, it is often sucient to nd a \good enough" direction
that can sustain the communications with acceptable packet error rates. Thus, beam
training overhead can be reduced by letting the beam search stop when both BI and BR
nd acceptable beam directions. This principle requires modications in the transmission
protocol.
We thus propose that the protocol proceeds as follows: As illustrated in Fig. IV.3, BI
and BR change their communication mode between Tx and Rx after every training packet
transmission. Thus, after sending a training packet in an omni-directional Tx (Omni-Tx)
mode, the device, i.e., BI or BR, updates its communication mode as beamformed Rx
(BF-Rx) mode to receive the training packet from the given direction of the opposite side
via RxBF. When having identied a beam direction with \sucient quality", the Rx will
continue the search until it can be sure to have found a local optimum, i.e., determined
that the SNR is worse on both sides of the \suciently good" direction. This is done
in order to increase the robustness of the received scheme, and in light of the fact that
nding the local optimum does not impose a signicant increase in training overhead. If
either BI or BR nds an acceptable beam direction in a BF-Rx mode, it can \piggy-back"
this information on the next training packet. When both BI and BR have found their
acceptable beam directions, this beam training procedure immediately stops.
Prioritized Sector Search Ordering: To accelerate the average search speed, we
prioritize the order of Rx beam directions to be searched, For this purpose, our proposed
70
Interactive Beam Training
Sector Sector Sector
…… ……
……
……
Prioritized Sector Search Ordering
Figure IV.2: A Fundamental Procedure of Interactive Beam Training
BR

BI

Omni
-Tx
BF-Rx
BI and BR change their roles
Omni
-Tx
BF-Rx
BI and BR change their roles
Omni
-Tx
BF-Rx
Continues
Figure IV.3: A Basic Concept of Interactive Beam Training
protocol also considers multi-level beam training. The proposed protocol orders the seg-
mented spaces

in terms of network association request/response (NAR) statistics. This
prioritized sector search ordering is quite useful in mm-wave systems because physical ob-
stacles can constitute very strong attenuators, thus greatly restricting the angular range
from which useful signals can come in a given room (this is especially true for walls,

The term segmented spaces is equivalent to sectors in IEEE 802.11ad and low-resolution (L-Re) beams
in IEEE 802.15.3c.
71
which can be easily penetrated by microwaves, but are impervious to mm-waves, and
which might not be eective re
ectors for certain geometric congurations either). The
regions with the largest number of NAR statistics might thus constitute the angular re-
gions from which radiation can physically occur, or it might be a region that is preferred
by users. A detailed implementation is discussed in Sec. IV.3.1.
Detailed Designs of Algorithm Concepts
Interactive Beam Training: The piggy-backing of information on the training packet
structure needs to be compliant with the existing structures. Our proposed packet has
one bit Boolean information named beam notication (BN). The initial value of BN is
N which means that the current BI or BR does not know its sucient beam direction.
When that changes, the BN is set to Y from that time onwards. Let us assume for the
following discussion that BI rst nds a suitable beam direction. It then listens to the
incoming training packet from BR with its optimal Rx beam and will see that the BN
of the packet is the initial value, i.e., N. Recognizing that the BR has not yet found its
own optimal beam direction yet, it will send out training packets in an omni-directional
Tx mode, its BN eld is set to Y. When the BR also nds its optimal beam direction, it
will see the BN of the training packet and nd that the BN is Y. Then, BR will recognize
that BI already found its suitable beam direction. Now, the proposed interactive beam
training operation is terminated. Therefore, as expected, our proposed procedure can
be terminated when both sides nd their optimal beam directions. Thus, we can avoid
exhaustive search for all the given beam search directions. It is also noteworthy that the
training packets transmitted with a power/spreading factor that provides sucient SNR
72
at the receiver even though the Tx sends them omnidirectionally; this holds true also for
the currently standardized training packets.
Note that the search time of this algorithm is a random variable, strongly in
uenced
by the direction in which the search was started. The worst-case scenario occurs when
all possible directions have to be searched, i.e., the initially searched beam was close to
the only admissible beam direction, but the search started into the \wrong" direction. In
that case the overhead is as large as that of a conventional exhaustive search.
Note that even after the rst node nds a \promising" direction, it will continue to
search, in order to nd a possibly better direction.
Prioritized Sector Search Ordering: Our beam training protocol is based on
multi-level (i.e., two-level) beam training. Therefore, we have a given number of seg-
mented spaces and maintain the NAR statistics for each segmented space (i.e., the sum-
mation of NAR statistics of the beams within the segmented space), and prioritize in
order of aggregate number of NAR requests, where recent requests might be weighted
more heavily than earlier ones. For the given N
NAR
samples, the following equation
holds:
N
NAR
=
N
X
i=1
f
2D
NAR
(i) (IV.2)
where i stands for a beam index (1 i N), N stands for the number of beam search
spaces with the given beamwidth, andf
2D
NAR
(i) stands for the NAR statistics fori-th beam
(0f
2D
NAR
(i)N
NAR
). Then each given segmented space has NAR statistics as follows:
f
SS-2D
NAR
(k) =
N
N
SS
k
X
i=1+
N
N
SS
(k1)
f
2D
NAR
(i) (IV.3)
73
Segmented Space (k, l)
9
6
7
8
1
2
3
4
5
17 11
12
13
14 15 16
ϕ
j

θ
i

10
Figure IV.4: An Example for 3D Beam Search Process within One Segmented Space:
Gray and black rectangles stand for the median SNR beams and the highest SNR beam,
respectively. In the gure,
j
and
i
stand for azimuth planes indexj (180

j
180

)
and elevation plane index i (90

i
90

), respectively.
where i stands for a beam index (1 i N), k stands for a segmented space index
(1 k N
SS
) where N
SS
stands for the number of segmented spaces, and f
SS-2D
NAR
(k)
stands for the NAR statistics for segmented space k (0f
SS-2D
NAR
(k)N
NAR
). We search
the optimal Rx beam direction from the segmented space which has the highestf
SS-2D
NAR
(k)
value where 1 k N
SS
. Within the segmented space, the Rx beam direction search
order is sequential, e.g., if we consider the segmented space k, the beam search order is
from 1 +
N
N
SS
(k 1) up to
N
N
SS
k.
IV.3.2 Interactive 3D Beam Training
Even though most of beamforming and training research results are designed and im-
plemented in a 2D scenario, a 3D geometric computation is also considerable to design
more realistic beamforming and training protocols [141, 160]. The main dierence to the
2D scenario is that we have to search in a 3D geometry to make sure we have a local
maximum. Apart from the longer search time, the beam training mechanisms mentioned
above can be reused in a straightforward way. One key dierence is the search for the
74
local optimum within the segmented space. In 2D interactive beam training, the desired
beam direction can be found by a linear sequential search within every single segmented
space (i.e., sector). However, this sequential search is not applicable in 3D interactive
beam training due to the fact that the segmented space is formed as plane. The detailed
operation is as follows: As presented in Fig. IV.4, sequential search is doing in a vertical
manner (from 1 to 8) for every single column. While doing this sequential search, if we
start to listen to SNR (at 8-th beam in Fig. IV.4), then we look at the surrounding beam
directions to check whether a higher-SNR beam exists or not. If yes, then we move to the
beam and look at the neighbor beams, and continue this procedure. Then, nally, we are
able to reach the local maximum SNR beam (i.e., 11-th beam in Fig. IV.4).
In this section, we designed an interactive algorithm on top of a sequential search. Note
that the interactive algorithm concept can also be combined with other, more ecient
search algorithms (instead of the linear search), such as quadratic search, to provide a
faster algorithm.
IV.3.3 Discussion: Tradeo { IFS Overheads
Our proposed scheme changes Tx and Rx modes every single training packet transmission
for interactive operation. In general wireless systems, when devices change their commu-
nication modes, a time interval is dened for the mode change, i.e., inter-frame spacing
(denoted as IFS). Moreover, there is another inter-frame spacing between each beam
space evaluation (denoted as ifs). In general, IFS takes more time than ifs. For example,
ifs is 73 ns and IFS is 0:5s in IEEE 802.15.3c [16], i.e., IFS is 6:85 times longer than ifs.
As shown in Fig. IV.5(a) and Fig. IV.5(b), there are four and two IFS durations in the
exhaustive beam training procedure using TxBF and RxBF when each device changes
75
1 2 …… N 1 2 …… N
BI BR
IFS
IFS
IFS
IFS
Feedback
Feedback
BI BR
ifs
• IFS: 0.5 μs
• ifs: 73 ns
(a) IFS for the Exhaustive Search using TxBF
1 2 …… N 1 2 …… N
BI BR
IFS
ifs
• IFS: 0.5 μs
• ifs: 73 ns
IFS
(b) IFS for the Exhaustive Search using RxBF
1 1
IFS
IFS
2 2
IFS
IFS
3
IFS
N N
BI
IFS
IFS
……
BR BI BR BI BI BR
• IFS: 0.5 μs
(c) IFS for the Proposed Procedure
Figure IV.5: IFS Durations
its communication mode. In addition, 2 (N 1) ifs durations in the exhaustive beam
training procedure using both TxBF and RxBF. However, the proposed protocol requires
76
Table IV.1: IFS overhead
N BF mode
k
N
= 0:94
k
N
= 0:95
k
N
= 0:96
k
N
= 0:97
k
N
= 0:98
k
N
= 0:99
90 TxBF 0.9498 0.9599 0.9700 0.9801 0.9902 1.0003
90 RxBF 0.9606 0.9708 0.9810 0.9913 1.0015 1.0117
180 TxBF 0.9553 0.9655 0.9756 0.9858 0.9959 1.0061
180 RxBF 0.9607 0.9709 0.9812 0.9914 1.0016 1.0118
360 TxBF 0.9581 0.9683 0.9784 0.9886 0.9988 1.0090
360 RxBF 0.9608 0.9710 0.9812 0.9914 1.0017 1.0119
12.09 μs
CMS PHY
Payload
CMS Frame
Header
CMS PHY
Preamble
5.52 μs 3.26 μs
Figure IV.6: CMS Frame Structure [16]
much more IFS durations which can lead to time-overheads as shown in Fig. IV.5(c). To
observe the impact of IFS durations, the time-overheads are computed and simulated.
As a reference control frame structure in this chapter, the control frame dened in
the 60 GHz IEEE 802.15.3c standard is used. In IEEE 802.15.3c, common mode sig-
naling (CMS) is used for synchronization/beacon frame transmission including beam-
forming/training procedure (see Section 12.1.12 in [16]) and the frame structure is as
illustrated in Fig. IV.6 [16]. In Fig. IV.6, the CMS frame consists of CMS physical layer
(PHY) payload (42 octets and 47.8 Mpbs, thus, required time to transmit this part is
12.09s), CMS frame header (33 octets and 27.8 Mbps, i.e., required time to transmit
this part is 5.52 s), and CMS PHY preamble (3.26 s is required). Among them, the
77
CMS frame header consists of Reed-Solomon (RS) parity bits (16 octets)
y
, header check
sequence (2 octets), medium access control (MAC) header (10 octets), and PHY header
(5 octet). More details are in [16].
Finally, as can be seen in Fig. IV.6, the required time to transmit one training packet
is
t
pkt
= 12:09s + 5:52s + 3:26s = 20:87s (IV.4)
The results of IFS overhead simulations are as presented in Table IV.1 where k and
N stand for the search order of optimal beam direction and the total number of beam
directions, respectively. For example, if the optimal beam direction of a device is the
36-th beam direction and N is set to 360, then
k
N
=
36
360
= 0:1. The computed values
in Table IV.1 stand for
T
interactive
T
exhaustive
where T
interactive
and T
exhaustive
are the operation times
for the proposed interactive beam training and the general (exhaustive) beam training,
respectively. Thus, if this is bigger than 1, our proposed protocol is slower than the general
procedure because of IFS overheads. The simulation is performed for three dierent
search space sizes, i.e., N = 90; 180; 360. Note that there are no segmented spaces in this
simulation. As we can observe in Table IV.1, the performance is almost the same even
if the size of the search spaces are varying. If approximately k <
98
100
N, our proposed
protocol is faster than the traditional ones. Otherwise, the performance of the proposed
protocol is worse due to the IFS overheads. Therefore, it is true that the impact of IFS
y
The forward error correction scheme for CMS is RS coding and the coding rate is 239/255, i.e.,
RS(255; 239).
78
Table IV.2: Numerical Analysis of Training Speed (: Beamwidth)
= 1

= 4

= 7

= 10

E[T
exhaustive
]
E[T
interactive
]
10.511 10.519 10.527 10.534
E[T
multi-level
]
E[T
interactive
]
1.433 1.791 2.149 2.506
overheads is realized with approximately 2%, i.e., we cannot observe the IFS overhead
impacts, in general.
In conclusion, if the NAR statistics for all Rx beams are uniformly distributed (i.e.,
there are no preferred beam directions), we can achieve faster beam training results with
the probability of around 98% than the traditional exhaustive searches. By exploiting
beam prioritization, we can achieve an even higher overall reduction of the training time.
IV.3.4 Analysis
To analyse the expected link setup speed of the proposed protocol (i.e., E [T
interactive
]),
suppose that BI and BR start the training from the angles
BI
and
BR
where180

BI
180

and180

BR
180

. For simplicity, we only consider forward search,
thus, the training will stop when BI and BR meet the nearest global/local optimum
direction in forward search directions. S
B
means the set of global/local optima (e.g.,
S
B
=f152

;122

; 0

; 102

; 128

; 140

g and N
S
B
= 6 in the 60 GHz mm-wave IEEE
802.15.3c CM1.3 channel model as simulated in Fig. IV.11,s is an element of S
B
which is
located in the forward search as seen from the current angle of arrival (AoA). According
to the fact that forward search is used, e.g., if current direction is 130

, only 140

can be
s in the IEEE 802.15.3c CM1.3 channel model.
79
kxsk
min
S
B
stands for the AoA distance between x and the nearest s. Then, the time
that BI nds its global/local optimum direction (denoted as t

BI
) is as follows:
t

BI
= (t
pkt
+t
IFS
)

k
BI
sk
min
S
B

!
(IV.5)
where stands for beamwidth. Similarly, the time that BR nds its global/local optimum
direction (denoted as t

BR
) is as follows:
t

BR
= (t
pkt
+t
IFS
)

k
BR
sk
min
S
B

!
(IV.6)
and our algorithm stops when both nd their \good enough" global/local optima, i.e.,
2 maxft

BI
;t

BR
g: (IV.7)
To deal with possible cases with arbitrary
BI
and
BR
, the two parameters are dened
as follows:

BI
= +
2

max
BI

BI
(IV.8)

BR
= +
2

max
BR

BR
(IV.9)
where
max
BI
and
max
BR
mean the sampling sizes for the evaluation on BI and BR,
BI
and

BR
are indices for the sampling where 0
BI

max
BI
and 0
BR

max
BR
.
Finally, the expected link setup speed of the proposed protocol (i.e., E [T
interactive
])
can be computed by (IV.11). We also assume that the NAS statistics for sectors are
uniformly distributed.
80
E [T
interactive
] =
2 (t
pkt
+t
IFS
)

max
BR
+ 1

max
BI
+ 1

max
BR
X

BR
=0

max
BI
X

BI
=0
max
2
6
4

+
2

max
BI

BI

s

min
S
B

; (IV.10)

+
2

max
BR

BR

s

min
S
B

3
7
5 (IV.11)
The expected link setup speeds of the traditional exhaustive protocol with RxBF (see
Sec. IV.2.2) and multi-level protocol with RxBF (see Sec. IV.2.2), i.e., E [T
exhaustive
] and
E [T
multi-level
], can be computed as follows:
E [T
exhaustive
] = 2

2

t
pkt
+

2

1

t
ifs
+t
IFS

(IV.12)
E [T
multi-level
] = 2

2

s

+

s

t
pkt
+

2

s

+

s

2

t
ifs
+t
IFS

(IV.13)
wheredxe stands for the function which is the smallest integer not less than x where
0

360

, stands for the mm-wave beamwidth, and
s
stands for the width of each
sector. As shown in Table IV.2, the proposed protocol is 10:51110:534 times faster than
brute-force search and 1:433 2:506 times faster than multi-level beam training with 8
sectors.
81
BI BI
Exhaustive
Search
Coarse-Grained Beam Training Fine-Grained Beam Training
(a) Traditional Multi-Level Beam Training
BI BI
Coarse-Grained
Beam Training
Fine-Grained
Beam Training
Iteratively dividing into k subspaces
BI
(b) Iterative Subspace Partitioning
Figure IV.7: Fundamental Concepts of Iterative Subpace Partitioning
IV.4 Iterative Beam Training for Adaptive Modulation
For adaptive modulation, we have to nd the global optimum so that we can choose the
highest-rate modulation and coding scheme supported by the channel. Thus, we have to
nd a way to get the optimal Rx beam directions for BI and BR while observing all beam
directions. Yet, we still wish to reduce the overhead compared to the exhaustive search
described in Sec. II.
82
BI BR
SSP to find the optimal sector of BR
(Exhaustive Search)
Omni-Tx of
training
packets
Finding the
optimal
sector via
RxBF
Omni-Tx of
training
packets
Finding the
optimal
beam via
RxBF
Finding the
optimal
sector via
RxBF
Omni-Tx of
training
packets
Finding the
optimal
beam via
RxBF
Omni-Tx of
training
packets
ISP to find the optimal beam of BR
(Iterative Subspace Partitioning)
SSP to find the optimal sector of BI
(Exhaustive Search)
ISP to find the optimal beam of BI
(Iterative Subspace Partitioning)
Figure IV.8: A Procedure of Iterative Beam Training
In this section, we maintain the basic structure of multi-level beam training, i.e., (i)
coarse-grained beam training (L-Re/sector sweeping) and (ii) ne-grained beam training
(exhaustive search within the computed optimal L-Re/sector) as presented in Fig. IV.7(a).
In order to reduce the operation time of the protocol, we suggest iterative subspace parti-
tioning with given division order instead of exhaustive search, as illustrated in Fig. IV.7(b).
Therefore, our objective is nding (i) optimal number of segmented spaces (i.e., sec-
tors) for coarse-grained beam training and (ii) optimal division order that can minimize
the number of evaluated search spaces in iterative subspace partitioning.
83
IV.4.1 Iterative 2D Beam Training
For the sector sweeping phase (SSP), the omni-directional spaces of BI and BR are divided
intoN
2D
sector
number of segmented spaces (i.e., sectors) at rst. The proposed protocol nds
the optimum sectors according to the exhaustive search procedure described in Sec. II.
For the next phase (i.e., iterative subspace partitioning (ISP)), BI transmits training
packets in an omni-directional Tx mode, again. At the same time, BR breaks down the
optimum sector which were found in SSP into' subspaces where' stands for the division
order (' 2). Then we have to evaluate each subspace. Thus, we receive ' number of
training packets from the BI for each subspace partitioning. This procedure is repeated
until the subspace becomes smaller than the beam width (termination condition). Now,
BR nds its optimal beam direction. Then BI and BR exchange roles, and the optimum
subspace at the BI is found. This procedure is illustrated in Fig. IV.8. Note that SSP
could be interpreted to be just one further step in the ISP. However, in order to retain the
backward compatibility to the procedure described in the IEEE standards, we identify
them here as two separate stages in the beamforming process.
Let us now analyze the total number of training packets that need to be transmitted
by the BI (an equivalent number has to be transmitted by the BR). In the rst phase, the
BI transmitsN
2D
'
training packets. In addition, during each iteration step of the beam
renement, ' training packets are required. LetN
2D
'
be the number of iteration steps,
i.e., the minimum positive integer that satises
360
N
2D
sector

1
'

N
2D
'
: (IV.14)
84
To get theN
2D
'
from (IV.14), take the logarithm (base: ') on the left-hand side (LHS)
and right-hand side (RHS) of (IV.14), then following (IV.15) is eventually obtained:
N
2D
'
=

log
'

360=N
2D
sector

(IV.15)
wheredxe stands for the function which is the smallest integer not less than x where
0

360

. Thus, in our two-level beam training, we have to evaluateN
2D
sector
number
of coarse-grained beams at rst, and iteratively we have to break-downN
2D
'
times. In
each breaking down, we have to evaluate ' search spaces. Therefore,
N
2D
sector
+'

log
'

360=N
2D
sector

(IV.16)
number of training packet transmissions are required. After the conrmation of the
optimum beam direction of BR, BI and BR swap their roles and nd the optimum beam
direction of BI. Then, the proposed protocol is terminated.
The proposed scheme has three parameters as follows:
N
2D
sector
: the number of sectors, a positive integer
': a division order, an integer where ' 2
: a beamwidth, a real number where 1 10
Among them, is given and our main optimization framework nds the optimalN
2D
sector
and ' to minimize the number of training packet transmission, i.e., minimize (IV.16).
85
Figure IV.9: Pseudo-code to ndN
2D
sector
and '
IV.4.2 Pseudo-Code and Computational Complexity
The pseudo-code to nd optimal number of sectors and optimal division order in the
proposed iterative beam training for a 2D scenario is as represented in Fig. IV.9. The
86
Figure IV.10: Simulation Results for Mobile Wireless Services
protocol for a 3D case is equivalent to Fig. IV.8 and Fig. IV.9 with some straightforward
modications of the parameters. In addition, the computational complexities are O

N
2

for both 2D and 3D scenarios (polynomial-time computation).
IV.5 Performance Evaluation
This section evaluates the performance of proposed protocols in terms of the wireless link
conguration speed because our objective is fast beam conguration.
IV.5.1 Interactive Beam Training for Fixed Modulation
This section presents the simulation results in terms of the fastness of beam training in
mobile wireless service applications.
For the performance evaluation of mobile wireless services, our considered scenario
includes one access point (AP) as a BI and one device as a BR in a room. In this setting,
BI and BR are located at the top corner and ground, respectively. Here, BI uses only
1
8
of
87
Table IV.3: Link Establishment Time Comparison (Unit: s)
Beamwidth Exhaustive Proposed Search Proposed Search Proposed Search
Search with Highest with Lowest with Average
Performance Performance Performance
1

3:01 10
6
6:41 10
1
2:77 10
6
1:38 10
6
2

7:52 10
5
6:41 10
1
6:92 10
5
3:46 10
5
3

3:34 10
5
6:41 10
1
3:08 10
5
1:54 10
5
4

1:88 10
5
6:41 10
1
1:73 10
5
8:66 10
4
5

1:20 10
5
6:41 10
1
1:11 10
5
5:54 10
4
6

8:36 10
4
6:41 10
1
7:69 10
4
3:85 10
4
7

6:28 10
4
6:41 10
1
5:78 10
4
2:89 10
4
8

4:70 10
4
6:41 10
1
4:33 10
4
2:17 10
4
9

3:72 10
4
6:41 10
1
3:42 10
4
1:71 10
4
10

3:01 10
4
6:41 10
1
2:77 10
4
1:39 10
4
the beam directions (i.e., only Southeast down-side directions are available) and BR uses
1
2
(i.e., only upper-side directions are available) of the beam directions among all possible
sphere beam directions. Thus, we consider that the given network has 8 segmented spaces
with a

2
= 90

central angles value.
Then the link establishment time comparison between the proposed protocol and the
exhaustive search using RxBF is simulated as presented in Fig. IV.10. The simulation is
performed with various beamwidths, i.e., from 1

to 10

(x-axis). As shown in Fig. IV.10, if
the proposed protocol nds the optimal beam directions for the rst beam directions every
time (i.e., the best case), the link establishment only takes 64:11s that is signicantly
88
less than for the exhaustive search with RxBF. Even if the proposed protocol nds the
optimal beam directions for the last tested beam directions for the segmented spaces with
positive NAR statistics (i.e., the worst case and the segmented spaces which have no NAR
statistics are ignored), the link establishment time is almost half of the link establishment
time of exhaustive search with RxBF (due to the fact that it does not due unnecessary
searches in segments that do not have positive NAR values). Therefore, the performance
is almost twice better than the widely used RxBF based exhaustive search operation in
terms of link establishment time.
As summarized in [17], if seamless mobile services are required, the following latency
requirements should be met:
Video services: 1:66 10
4
s
Voice over IP (VoIP) services: 5 10
4
s
Then, as can be seen in Fig. IV.10 and Table IV.3, it is obvious that the currently
existing exhaustive search using RxBF cannot support seamless mobile video and VoIP
services at all. However, if our proposed protocol can nd optimal Rx beam directions in
the rst search space (i.e., the best case), then all kinds of seamless mobile services can
be supported even if there is no prioritized segmentation. Even in the worst case, at least
seamless mobile VoIP services can be supportable if the beamwidth is equal to or larger
than 8

; in the average case, seamless mobile VoIP services can be supportable when the
beamwidth is equal to or larger than 6

(without search space prioritization). In addition,
seamless mobile video services can be also achieved if the beamwidth is equal to or wider
than 9

in the average case.
89
As an example, we investigate the performance in a channel model that is adopted from
IEEE 802.15.3c [15]. Among the various IEEE 802.15.3c TSV channel model scenarios, we
considered channel model CM1.3, i.e., a line-of-sight (LOS) scenario (at a 5 meter distance,
which leads to an LOS-component attenuation of 82 dB) and a residential environment.
For this channel model, the angular power spectrum (APS) is shown in Fig. IV.11(a).
From this we derive the distribution of the SNR over the AoA with omni-directional
transmitter and receiver in Fig. IV.11(b). As can be seen, the receiver is able to get an
appreciable SNR around 6 times based on the CM1 channel model because of the multi-
path components by re
ection. Note that the SNR in 0

is the global optimum. The
other SNR representations are local maxima.
Let th
d
and th
v
be the threshold SNR for decodability of the MAC header and the
threshold SNR for the successful actual data transmission (video signal decoding), re-
spectively. As presented in Sec. IV.3.3, the training packet is CMS and CMS uses mod-
ulation/coding scheme level 0 (MCS0) [16]. As presented in [129], th
d
is 2:9 dB if our
desired bit-error-rate (BER) performance is 10
4
. Forth
v
, the minimum bounds of SNR
for decoding the most signicant bits (MSB) and the least signicant bits (LSB) of video
signals with BER performance 10
4
are 7 dB and 10:5 dB, respectively, in IEEE 802.15.3c
as veried in [121]. Note that in the scalable coded video, the base layer is encoded as
MSBs, while additional layers exist. If the wireless channel bandwidth is enough to trans-
mit all the scalable video layers, we can transmit the scalable video coding layers up to
LSB (i.e., all layers are able to be transmitted). Moreover, let G
s
andG
b
be the antenna
gain of a sector antenna (i.e., coarse-grained beam) and the gain of the adaptive array
that is used for the transmission over nal beams (ne-grained beam, i.e., used for actual
data transmission), respectively. G
s
andG
b
are set to 10 dB and 23 dB, respectively [15].
90
Based on Fig. IV.11(b), if both transmitter and receiver have omni-directional beam
patterns, the training packets are not able to be decoded because the maximum SNR
(AoA: 0

) is lower than 7 dB. However, due to the fact that the maximum SNR with
G
s
is larger than th
d
, the training packets of our coarse-grained beam training can be
decoded. Depending on the distance, only the globally optimum AoA, or several of
the local maxima, can be decoded. Moreover, for actual video data transmission, the
transmitted video signals are able to be decoded if following equation holds:
fmax SNR (AoA: 0

) in Fig. IV.11(b)g + 2G
b
th
v
: (IV.17)
Then, it is true that all MSB and LSB information with all given 10
4
BER performance
are satisfying this equation. Therefore, all video signals are able to be decoded as well.
It is noteworthy that decodability of the packet header implies that video packets can
always be decoded (due to the fact that the antenna gain during video transmission is
much higher than the sector antenna gain). On the other hand, there can be situations
where a video decoding would be possible but the connection cannot be set up since the
transmitter and receiver cannot nd suitable sectors. This is a weak point in the 802
protocols whose resolution would be important for future applications, but which likely
cannot be alleviated without impact on backward compatibility.
IV.5.2 Iterative Beam Training for Adaptive Modulation
With the given beamwidths, i.e., from 1

to 10

with 0:001

step size (i.e., totally 9001
sample evaluation), the optimal numbers of sectors and optimal division orders are sim-
ulated for a 2D scenario. The simulation results with 9001 samples are presented in
91
Table IV.4: Optimal Number of Sectors and Optimal Division Orders in Interactive 2D
Beam Training
# of Sectors: Division Order Statistics Statistics
N
2D
sector
' (out of 9001) (%)
2 2 1637 18.19%
2 3 4036 44.84%
3 2 1041 11.57%
3 3 1575 17.50%
4 2 0 0%
4 3 712 7.91%
Table IV.4. As presented in this table, the possible numbers of sectors are 2,3, and 4 and
the possible division orders are 2 and 3. The other values are not optimal in terms of
operation time (i.e., number of search spaces).
In Table IV.5, our proposed protocol is faster than the exhaustive search using RxBF
for all evaluated 9001 samples. In addition, our link conguration speeds for all evaluated
samples are less than 1:6610
4
s and 510
4
s, i.e., the requirements for seamless video
and VoIP services, respectively. This means that our protocol can set up the wireless link
within the delay bounds of video and VoIP services as summarized in [17]. Thus, even if
the wireless link is disconnected while it serves video or VoIP streams, we can establish
the link before the disconnection of the session.
In Fig. IV.12, link conguration time comparison, i.e., the link conguration time
of exhaustive search with RxBF over the link conguration time of iterative subspace
partitioning, is plotted. For the given beamwidths, the link conguration time of iterative
92
Table IV.5: Iterative 2D Beam Training Performance (Unit: s)
Beamwidth Iterative Beam Training Exhaustive Search
1

1:42 10
3
1:5 10
4
2

1:26 10
3
7:52 10
3
3

1:17 10
3
5:01 10
3
4

1:09 10
3
3:76 10
3
5

1:01 10
3
3:01 10
3
6

1:01 10
3
2:51 10
3
7

9:23 10
2
2:18 10
3
8

9:23 10
2
1:88 10
3
9

9:23 10
2
1:67 10
3
10

9:23 10
2
1:51 10
3
search is always faster than the one of exhaustive search. In the best case, iterative search
is approximately 10 times faster than exhaustive search when the beamwidth is near 1

.
In the average case, iterative search is around 3:46 times faster than exhaustive search.
IV.6 Concluding Remarks
This chapter investigated fast 2D and 3D interactive beam training protocols for narrow-
beam mm-wave wireless systems. In traditional beam training systems, exhaustive search
is widely used, however, this protocol denitely takes a lot of beam training time in
narrow-beam mm-wave wireless applications. For xed modulation, our proposed proto-
cols are designed based on interactive beam training and prioritized search space ordering
93
for both 2D and 3D cases. The performance of our protocols are veried by simulation in
terms of mm-wave wireless link conguration speed in standardized propagation channel
models. For adaptive modulation, multi-level beam training protocols are proposed which
have two steps, i.e., L-Re/sector sweeping and iterative subspace partitioning. For each
step, we nd the optimal number of sectors and division orders to minimize the operation
speed. These protocols can serve as relevant improvements for the existing mm-wave
IEEE standards, i.e., IEEE 802.15.3c or IEEE 802.11ad. Importantly, all the suggested
procedures can be adapted into the standards with minimum changes (i.e., introduction
of one new bit in the MAC header).
94
(a) Channel APS
(b) SNR Distribution: Omni-Tx and Omni-Rx
Figure IV.11: 60 GHz Channel Realization
95
Figure IV.12: T
exhaustive
=T
iterative
96
Chapter V
Device-to-Device Adaptive Video Streaming for
Distributed Programmable WiFi Platforms
V.1 Introduction
Recently, there are active research contributions which aim on the design and implementa-
tion of next generation wireless network systems and protocols including device-to-device
wireless video networks [35]. As a part of these research contributions, a stochastic adap-
tive video streaming mechanism which can jointly optimize (i) transmission scheduling
and (ii) admission control in wireless device-to-device networks was proposed in [162, 114]
as results of our previous research.
Our considering wireless device-to-device network consists of two types of devices,
i.e., helpers and users. Helpers are pre-deployed wireless devices which contain video
chunks. When an user wants to download a video le, these pre-deployed helpers serve
the sequence of corresponding chunks to the user. For more details, the users place video
streaming requests for each chunk to one of the helpers in each time slot and wish to
download sequences of video chunks corresponding to the desired video les. With this
97
distributed operation, each chunk can be downloaded from dierent helpers in each time
slot. The helpers contain cached video les and serve the user requests over device-to-
device wireless links. Therefore, the fundamental functionalities of users are (i) placing
the chunks of the desired video le to the distributed helpers as well as (ii) determining
the quality level of each chunk in each time slot (named admission control). In addition,
the fundamental functionalities of helpers are downloading the chunks of requested video
to users in terms of max-weight scheduling principle (named transmission scheduling) in
each time slot.
Even though the proposed algorithm can achieve max-weight throughput in terms
of network utility maximization, the desired performance cannot be obtained when we
implement the algorithm on top of mobile programmable platforms, e.g., Android mo-
bile platforms. In Android ad-hoc network programming, mobile apps can be designed
and implemented which utilize WiFi, Bluetooth, and near eld communications (NFC)
functionalities via corresponding application programming interfaces (APIs). Among the
network functionalities in API, Bluetooth and NFC are not good enough to be used for
video streaming applications due to their low data rates

. Therefore, Android mobile
platforms are only able to use WiFi for video streaming and the Android WiFi API is not
able to support multi-channel programming interfaces. Thus, the WiFi-based software
can handle only single channel operations. Then, when our previous algorithm is designed
and implemented on Android mobile software platforms, it is not possible to program the
case that multiple helpers are transmitting video chunk bits toward a single same user
at the same time because of the usage of single channel, i.e., receiving the signals at a

Bluetooth and NFC are introduced in mobile smartphones for small size audio streaming and text
messaging among nearby mobile smartphones.
98
single user from multiple helpers at the same time is not possible which is fundamentally
required for max-weight scheduling implementation. In this case, even if max-weight
scheduling is desired in our previous scheme, random scheduling will be performed in
the worst case. Therefore, an alternative method which can deal with this problem is
desired for designing real-world implementable device-to-device video streaming wireless
networks.
In this chapter, a heuristic approach is introduced in user-side software where the
users can fetch desired chunks from helpers in terms of playback order to minimize de-
lays. By introducing this heuristic, the transmission scheduling can be eliminated from
helper functionalities. By performing simulations, we verify that our proposed algorithm
can achieve improved performance compared to our previous algorithm when the given
D2D wireless network operates over single channel WiFi wireless links. In addition, our
algorithm is implemented and demonstrated via Android mobile programmable platforms.
The remainder of this chapter is organized as follows: Section V.2 provides the de-
tailed operations of our previous joint optimization algorithm. Section V.3 describes the
motivation and algorithm details of currently proposing scheme for single channel mobile
systems. In addition, section V.4 presents the software design and implementation and
then section V.6 concludes this chapter.
V.2 Push-Strategic D2D Video Streaming
Our previous scheme in [162], so called push-strategic D2D video streaming in this chap-
ter, is fundamentally formulated for network utility maximization in D2D wireless video
99
systems where the utility of a particular user is measured by its time-averaged quality
which is downloading from helpers.
V.2.1 System Model
Our considering system model is as follows: Suppose a D2D wireless video network is
dened by a bipartite graphG,fD;Eg whereD stands for the set of devices and consists
of two types of devices, i.e.,D ,fH;Ug whereH andU stand for the sets of helpers
and users. In addition,E stands for the set that consists of edges for all helper-user pairs
(h
i
;u
j
) such that there exists a potential wireless link between h
i
2H;8i2f1; ;jHjg
and u
j
2U;8j2f1; ;jUjg. In addition,N (u
j
)H;8j2f1; ;jUjg andN (h
i
)
U;8i2f1; ;jHjg stand for the sets of neighbors of users and helpers, respectively, i.e.,
N (u
j
) , fh
i
2Hj (h
i
;u
j
)2Eg; (V.1)
N (h
i
) , fu
j
2Uj (h
i
;u
j
)2Eg; (V.2)
where8i2f1; ;jHjg and8j2f1; ;jUjg.
With the given wireless bipartite network, each user u
j
2U,j2f1; ;jUjg requests
its desired video le v
j
. The request of user u
j
2U, j 2f1; ;jUjg for a chunk at
a particular time t can be assigned to any one of the helpers inN (u
j
). The chunks
have same duration in terms of playback time and these chunks must be reproduced in
sequence at receiver users.
100
V.2.2 Video Streaming Principles
As explained in [162], video streaming is dierent from video downloading in wireless
network research due to the fact that the playback starts while the whole video le has
not been transferred. In video downloading, the video can be played back once entire
bits of the video content have been downloaded from transmitter (Tx) to receiver (Rx).
To deal with this issue, video streaming consists of several chunks which can be played
in a stand-alone unit. Suppose that a video v
k
consists ofjCj number chunks where
c
(k;1)
;c
(k;2)
; ;c
(k;jCj)
. When Tx is doing streamingv
k
, the chunks should be transmitted
in a playback order. While the Tx sends the rst chunk c
(k;1)
, the Rx receives it. While
the Tx sends the next chunk c
(k;2)
, the Rx has the entire bits of c
(k;1)
so that the Rx can
playback the chunk and note that the chunks can be independently played in a stand-
alone unit while the Rx receives c
(k;2)
from Tx. Thus, Rx can keep playing back the
chunks in a playback order while Tx sends chunks. For this reason, the estimated delay in
streaming is the time for transmittingc
(k;1)
amount of bits while the delay in downloading
is v
k
amount of bits. Therefore, streaming has benet in terms of the minimization of
networking delays.
V.2.3 Algorithm Details
The algorithm consists of two separable parts, i.e., admission control at users (refer to
Sec. V.2.3) and transmission scheduling at helpers (refer to Sec. V.2.3). For more clear
understanding, our considering network is illustrated in Fig. V.1. More theoretical details
are presented in [162, 114].
101
h
1
h
2
u
1
u
1
u
2
u
3
u
2
u
3
u
1
u
2
u
3
Figure V.1: A Reference Example Network Model: There exists 2 helpers, i.e., H =
fh
1
;h
2
g, and 3 users, i.e.,U =fu
1
;u
2
;u
3
g. Suppose that all helper-user pairs are con-
nected, i.e., bipartite graph. Then, each helper has 3 queues which are for serving ded-
icated/associated users. In each time slot, users place chunks to their selected helpers
and determine the quality mode of the chunk. In addition, helpers are doing transmission
scheduling in terms of max-weight scheduling in each time slot.
Admission Control at Users
When a user wants to download a video, it broadcasts the request message to all neighbor
helpers. When the neighbor helpers listen to the request message, they check that whether
they have the desired video or not. If they have the video, each helper generates a queue
for the user (initial backlog size: 0) and replies back a message which contains the queue
backlog size for the dedicated user. Then, the user performs two distributed operations in
each time slot, i.e., (i) distributed chunk placement for the desired video and (ii) quality
model selection for each chunk.
For distributed chunk placement, each user, e.g., u
j
;8j 2f1; ;jUjg, observes its
dedicated queue backlog size at each helper, i.e., Q
(i;j)[t]
;8i2f1; ;jHjg foru
j
, in each
time slot. When the user receives this backlog size information from all helpers, it selects
102
one helper which has the smallest queue backlog size for the user, and then the user places
the bits which consist of the current chunk to the helper. In the example from Fig. V.1,
u
1
compares the its own queue sizes withinh
1
andh
2
. The queue ofu
1
withinh
1
is longer
than the queue of u
1
within h
2
. Thus, u
1
places current chunk into the queue of h
2
. For
the same reason,u
2
andu
3
place their current chunk toh
1
andh
2
, respectively. We note
that the numbers of bits which forms individual chunks are all dierent even though they
have same playback time.
For quality mode selection, each user determines the quality model of each chunk
according to the draft-plus-penalty (DPP) principle to deal with the tradeo between the
qualities of chunks and delay in a stochastic network optimization manner. Therefore,
each user computes the quality mode of each chunk in a stochastic manner for joint
optimization of the two criteria when the user is placing the chunk to one of neighbor
helpers. More theoretical investigations in terms of stochastic network optimization are
well-studied in [13].
Transmission Scheduling at Helpers
Each helper has queues for the currently associated users. To transmit the bits within
the queues, the helper selects one queue among N
u
number of queues where N
h
i
u
stands
for the number of associated users at helper h
i
, and the helper transmits the bits within
the queue for the given unit time slot interval. To select one queue/user, which the helper
wants to serve, the helper measures the data rates with currently associated users and
nds one user which maximizes following max-weight scheduling criterion:
Q
(i;j)
[t]Rate
(i;j)
[t] (V.3)
103
h
1
h
N
u
1
u
1
u
1
h
2
u
1
h
3
u
1
……

Figure V.2: Motivation: In the case of multiple helper and single user, each helper selects
the single helper all the time but the user cannot be served by the multiple helpers at
the same time because of the single channel constraint in WiFi-centric wireless systems.
Then, the user selects one helper in a random manner which is equivalent to random
scheduling in the worst case.
where Q
(i;j)
[t] stands for the queue backlog size at helper h
i
with one of associated user
u
j
at timet and Rate
(i;j)
[t] stands for the data rate at the wireless link between helper h
i
and user u
j
at time t.
V.3 D2D Video Streaming with Greedy Heuristic
V.3.1 Motivation
Even if the proposed stochastic streaming in Sec. V.2 can achieve desired performance as
studied in [162, 114], implementing the scheme in programmable open mobile platforms
has signicant diculty and the reason is as follows. The state-of-the-arts of Android
networking API supports several wireless networking programming capabilities including
WiFi, WiFi-Direct, Bluetooth, and NFC. Among these wireless network protocols, only
104
WiFi/WiFi-Direct is able to support reasonable data rates for wireless video streaming.
In Android WiFi API, it is not allowed to utilize multiple channels which is essentially
required for implementing max-weight scheduling which is a major part of our previous
scheme in [162, 114].
Suppose that an example network consists of one user u
1
and N number of helpers,
i.e.,jHj =fh
1
; ;h
N
g as illustrated in Fig. V.2. According to the fact that only one
user exists in the wireless network, each helper maintains only one queue for the dedicated
user u
1
. Then, as a result of transmission scheduling decision in all helpers, u
1
will be
selected in all given time slots by all given helpers. However, u
1
cannot receive the bits of
scheduled video chunks from multiple helpers because u
1
has only one channel in WiFi-
centric systems. Thus, u
1
should be served by only one helper. According to the fact
that our previous algorithm does not have any algorithms which can deal with this case,
u
1
randomly selects one helper. Eventually, even if our previous scheme in [162, 114]
is designed to perform max-weight scheduling, it works as random scheduling when the
algorithm performs over a single channel wireless systems such as WiFi in the worst case.
Therefore, we need better methods which can improve the performance.
V.3.2 Heuristic { Fetching Bits based on a Playback Order
Our proposed scheme in this chapter is designed by utilizing the information on user side.
According to the fact that each user in the network performs distributed chunk placement
in each unit time over the queues within the distributed neighbor helpers, each user knows
the distribution of chunks within the helpers. In addition, by referring to its own local
storage, each user knows that which chunks are not completely arrived yet. Therefore,
each user checks its own local storage and checks the received chunks in terms of playback
105
order whether the bits of the chunks are arrived. While doing this checking, when the
user nds a chunk which is not fully received its bits, the user checks which helper has
the remaining bits. Then, the user fetches the bits from the helper.
We note that the transmission scheduling can be eliminated from the helper side
algorithms by performing this fetching in terms of playback order at user-side software.
V.3.3 An Operational Framework
In terms of video data structure, a video consists of a sequence of chunks. In [162], the
unit of service processes of transmission queue is in bits. However, bit-level transmission
control is not allowed in current Android API (SDK Version 4.1). To cope with this
problem, we dene the concept of sub-chunk where a single chunk consists of a sequence
of sub-chunks. With this concept, we can control the amount of transmitted bits in
terms of the number of transmitted sub-chunks depending on the channel conditions.
The dynamics of the transmission queue at each helper for each user is modeled as follows
where t2f0; 1;g:
Q
h
u
(t + 1) = max
n
Q
h
u
(t)
h
u
(m;t); 0
o
+c
h
u
(m;t) (V.4)
where8(h;u)2E and Q
h
u
(0) = 0;8h2H;8u2U. In addition, c
h
u
(m;t) and
h
u
(m;t)
stand for the number of bits of incoming chunk with quality mode m at time t and the
number of bits of processed one or more sub-chunk(s) with quality mode m at time t,
respectively. Note that if the channel is good, a helper can transmit more sub-chunks in a
given time interval. The users can employ three dierent algorithms: (i) selecting helpers
for each chunk placement (i.e., Sec. V.3.3), (ii) selecting helper which should transmit
106
one or more sub-chunk(s) to the user based on the GPMD strategy (i.e., Sec. V.3.3), and
(iii) selecting quality mode for each chunk (i.e., Sec. V.3.3). Therefore, users can handle
entire protocol behaviors and helpers just transmit chunks to requested users with desired
quality mode.
Chunk Placement
For each time slot, each user determines which helper should place/stream the next chunk.
When each time slot starts, each user sends packets requesting the next video chunk to all
neighbor helpers, and each helper replies back the current queue backlog size to the user.
Then, the user select the helper which has the smallest queue backlog size and the user
sends one packet to the helper for letting it know that the helper should send the next
chunk in its queue. If there are multiple helpers which has the smallest queue backlog
sizes, the user performs a random selection.
Greedy Pull for Minimum Delay
For each time slot, each user selects its helper which has to transmit sub-chunks to the
user. In our previous work [162], transmission scheduling is dened and implemented
at helpers and the objective of the function is sum rate maximization. Then, one user
can be selected by multiple helpers due to the fact that helpers are operating in a fully-
distributed manner. Moreover, if only one user exists in the system, then all helpers select
the single user on every time slot. Unfortunately, this kind of system cannot be supported
by WiFi-based mobile platforms, because having multiple APs in a single WiFi network
is not allowed. Thus, we design an alternative algorithm at user side, named greedy pull
for minimum delay (GPMD). Suppose that a user is receiving a sequence of chunks from
107
helpers, and the sub-chunks of the desired video exist within the local storage of the
user. Then, the user checks the playback order of its own list of sub-chunks. When the
user gures out that a sub-chunk is not yet received from helpers and urgent in terms
of playback order, then user requests that sub-chunk from the helper. If the channel
condition is good enough, the user is able to download not only this sub-chunk but also
the sub-chunks within the FIFO queue of the currently serving helper. The amount of
sub-chunks to be transmitted from helpers to users is determined by the channel state of
the WiFi link.
Quality Mode Selection
For each user, the quality mode m
u
(t) is computed by
m
u
(t) = arg min
m2fH;Lg
n
Q
h
u
(t)kB
v
(m;t)VD
v
(m;t)
o
(V.5)
as explained in [162] where H and L stand for high-quality and low-quality, respectively.
In addition, D
v
(m;t) and kB
v
(m;t) stand for the video quality measurement index and
the number of bits for video v with quality mode m at time t. Last, V > 0 means a
control parameter which aects an utility-delay tradeo [162].
V.4 Software Implementation
V.4.1 A Protocol Operation Overview
For the rst time slot, the implemented protocol starts when a user broadcasts a request
of desired video to neighbor helpers. Then, the neighbor helpers reply back packets to
108
Functional Architecture
Helper
WiFi-AP
Samsung Galaxy S4
Storage
Video 1
Video N
……
Chunk (N,1)
Chunk (N,2)
Chunk (N,M)
……
Sub-Chunk (N,M,1)
Sub-Chunk (N,M,2)
Sub-Chunk (N,M,K)
……
Computation
……
Queue for u
U

Queue for u
1

Multi-Queues
Queue Backlog Size
Computation
Channel Condition
Measurement
Chunk Placement
(a) Helper Software Architecture
Functional Architecture
Samsung Galaxy
Tablet 2
Storage
Video 1
Video N
……
Chunk (N,1)
Chunk (N,2)
Chunk (N,M)
……
Sub-Chunk (N,M,1)
Sub-Chunk (N,M,2)
Sub-Chunk (N,M,K)
……
Computation
Helper Selection for
Chunk Placement
Quality Mode Selection
GPMD Function
User
WiFi-Station
(b) User Software Architecture
Figure V.3: Device Software Architectures
the requested user with the information of their queue backlog sizes. In this initial stage,
there are no sub-chunks in the queues yet, i.e., all neighbor helpers will reply back to the
109
user with zero. Then, the user selects one helper which has the shortest queue backlog
size. According to the fact that the queue backlog sizes of all neighbor helpers are zero at
this moment, the user performs a random selection where the selected helper is denoted
as h
0
. Now, the user transmits one packet to inform h
0
about: (i) h
0
is selected as the
serving helper, (ii) desired chunk of the requested video (note that the user needs the
rst chunk at this initial stage), (iii) maximum link rate between current user and h
0
, and
(iv) desired quality mode for the requested chunk computed by (V.5). Then, h
0
sends the
sub-chunks of the rst chunk into the queue with desired quality. Note that there is no
video sub-chunk transmission from helper to user in this initial time slot, i.e., just the
sub-chunks are pushed into the queue. For the next time slot, the user requests the next
chunk (i.e., the second chunk)from neighbor helpers. Then, the neighbor helpers reply
back packets to the user with their queue backlog sizes. After that, the user selects the
helper which has the shortest queue backlog size and follows the same procedure discussed
before (chunk placement procedure). In terms of video chunk transmission, if the user
have not received any chunks , i.e., the user requests the rst sub-chunk of the rst chunk
along with computed quality mode located at h
0
. On the helper side, h
0
transmits the
sub-chunks with desired quality mode (H or L)within the given time slot interval. If the
channel condition is good, the h
0
may transmit more sub-chunks, but in the worst case,
it can transmit only one sub-chunk. If the helper, h
0
, cannot send the all sub-chunks of
the rst chunk, then it can be transmitted later when the helper is scheduled. In the
following time slots, the previous operations are repeated until helpers transmit the last
sub-chunk of the last chunk.
110
V.4.2 Helper Implementation
The software for helper side is implemented on top of Samsung Galaxy smartphones
with Android platforms. The helpers in this system are set to WiFi-APs, i.e., they work
as WiFi APs. Therefore, the users can also periodically measure achievable link rates
or RSSI measures from the helpers, because the corresponding functions are allowed in
WiFi AP-centric systems, such asgetRssi() dened by the classes in the Android API's
android.net.wifi package. For multi-user support, the helpers can create multiple
TCP sockets for the multiple users which want to download sub-chunks from the helpers.
The roles of helpers are as follows: (i) returning the queue backlog sizes when they
receive chunk requests from users, (ii) placing sub-chunks when they are selected by users
for specic chunk service, and (iii) transmitting a number of sub-chunks depending on the
channel conditions. The corresponding software architecture is illustrated in Fig. V.3(a).
V.4.3 User Implementation
The software for user side is implemented on top of Samsung Galaxy tablet with Android
platforms. The roles of users are (i) determining which helpers should be selected for
chunk placements, (ii) selecting helpers for the GPMD strategy in terms of sub-chunks
downloading, and (iii) selecting quality modes for each chunk. The corresponding software
architecture is illustrated in Fig. V.3(b).
V.5 Used Video for Demonstration
Throughout this demo, we use standard test sequences "Bridge" and "Highway" both of
which have 2000 frames with CIF (352 288) resolution. Each video sequence is divided
111
into 10 chunks and then encoded using mpeg encoder [5] at 25 fps so that each chunk
corresponds to 8 seconds. In order to generate two dierent quality levels per video, we
encode each video at rates 250 kbps and 50 kbps corresponding to high quality and low
quality streams, respectively. Our demonstration video clip is available in [6].
V.6 Concluding Remarks
This chapter presents a practical implementation of an adaptive video streaming pro-
tocol for device-to-device mobile software platforms. In the given network, users can
dynamically select helpers and video quality modes (two quality levels) for adaptive video
streaming. In addition, we proposed the GPMD strategy to deal with the implementation
issues or limitations in our previous work [162].
112
Chapter VI
Joint Scheduling and Stochastic Streaming for
Device-to-Device Video Delivery
VI.1 Introduction
According to recent predictions of the Cisco Visual Networking Index (VNI) [181], the
sum of all forms of video will constitute 80% to 90% of global consumer data trac
by 2017, and the trac from wireless and mobile devices will exceed the trac from
wired devices by 2016. Therefore, ecient video-aware network algorithms for wireless
networks are of highest importance [35, 94, 96, 159, 155, 147, 144, 140]. It has been shown
recently [35, 115, 148] that the throughput for delivery of wireless video les can be greatly
enhanced by device-to-device (D2D) communications, where direct links between pairs of
user devices can be set up without requiring to go through a central base station. In
particular, these references proposed a system where: each device caches, at random
(according to a certain probability distribution), a subset of popular video les. When
a user needs a le, it obtains it from one of its neighbors through a spectrally ecient,
short-range D2D link. As user density increases, the aggregate storage space of the
113
devices increases, while the average communication distance decreases (and the spatial
reuse increases). For these reasons, D2D networks for video delivery are scalable, such
that (under many conditions) demand and throughput both increase linearly with user
density.
However, previous work concentrated on the delivery of entire video les, such that
playback would only start once the le has been completely delivered. This is not how
most video delivery happens nowadays. Rather, delivery is by streaming, such that (after
a brief buering time) the video is divided into \chunks", and later chunks are transmitted
while earlier chunks are being played back. A transmission algorithm for such a system
consists of two components: (i) a scheduling algorithm that determines which D2D pairs
are allowed to transmit at a given time, and (ii) a streaming component that determines
for a scheduled pair the quality at which a chunk should be transmitted. These two
components are obviously coupled.
Currently, the most well-known D2D scheduling protocol in both industry and academia
is FlashLinQ [132, 107]. It is a distributed algorithm that schedules D2D links accord-
ing to their priorities, such that the higher-priority links do not suer from signicant
interference of possibly scheduled lower-priority links. Theoretically, it can guarantee
the maximum number of activated D2D links as analyzed by the theory of stochastic
geometry [91]. However, FlashLinQ does not incorporate naturally a video quality-aware
mechanism, and therefore its suitability for D2D on-demand video streaming remains
open.
For the question of video streaming quality adaptation, a number of protocols have
been suggested in the past. The proposed protocols consider various characteristics of
video streaming, e.g., cloud-based video streaming protocols are proposed in [98, 101],
114
channel-aware streaming algorithms are discussed in [68], new architectural concepts are
presented in [111], rate-distortion theory based (or quality-aware) streaming is addressed
in [44, 112], and resource-aware streaming algorithms are mentioned in [100].
The key postulate of the proposed schemes is that D2D scheduling and streaming
are coupled, and that therefore a joint algorithm has to be used for such a task. In this
chapter, we propose and analyze both centralized and distributed algorithms that combine
aspects of the above-mentioned methods.
For the proposed centralized quality-aware streaming and streaming algorithm, we
utilize the benets of cellular centralized resources, i.e., while the devices communicate
directly with each other, they are under control of an existing base station (BS); see [26] for
a discussion and further references of such systems. Thus, the scheduling decision can be
made by the cellular BS and we assume that the BS knows the channel state information
(CSI) for all the possible pairs of D2D links. The scheduling component of the proposed
scheme is based on a link con
ict graph (e.g., formed centrally by the BS) such that the
links scheduled to be active simultaneously at any time slot must form an independent
set of such con
ict graph. This con
ict graph based scheduling can be formulated as
a max-weight independent set (MWIS) problem. The MWIS problem is known to be
NP-hard; in this work we make use of a message-passing algorithm proposed in [120, 76],
to (approximately) solve the MWIS problem. Therefore, our scheduling component is
designed based on this message-passing concept with D2D-related modications, i.e., the
weight is not only based on the queue backlog size but also the CSI. Our distributed
scheduling scheme is based on FlashLinQ, but we improve it by incorporating weights
(similar to MWIS) and thus providing a connection to the streaming decisions for video
quality.
115
Our streaming component is based on stochastic network optimization with the con-
sideration of the quality of video streaming. Each video consists of a number of chunks.
Each chunk can be requested at dierent quality levels,

such that higher quality cor-
responds to more bits per chunk to be delivered. Therefore, our algorithm dynamically
controls the quality mode of each chunk to maximize total quality subject to all data be-
ing supportable over the network. Note that the streaming decisions impact the weights
(for MWIS or modied FlashLinQ) of the scheduling.
This work extends previous investigations of adaptive video streaming algorithms [162,
114]; and we retain the notation of those chapters for streaming-related aspects. The algo-
rithms in [162, 114] are for adaptive stochastic video streaming for cache/helper (equiva-
lent to access points plus video database in WiFi-based networks) and apply to a bipartite
network topology where a set of infrastructure nodes (small cell base stations with cached
video les) serve a set of wireless users.
Furthermore, our work diers from these papers as follows:
In [162, 114] it is assumed that each user can be served simultaneously by multiple
infrastructure nodes over each scheduling slot. Instead, here we explicitly consider
the constraint of the D2D link con
ict graph, such that at each scheduling slot a
user can only be served by another (peered) user device, if the corresponding link
belongs to the scheduled independent set.
The algorithms in [162, 114] dynamically match source and destination pairs in
every single unit time operation. However, the algorithms in this chapter work on

For example, this can be obtained by storing multiple copies of the same video encoded at dierent
rates, as in current video on-demand delivery systems such as Net
ix or Amazon Prime, or by using
scalable video coding and requesting more or fewer renement layers [93, 86, 134].
116
xed source-destination pairs, as this is the relevant case for D2D communications.
Let us note that FlashLinQ also considers this case [132, 107].
Since user devices are paired permanently over a whole streaming session in the
present work, there is no hand-over delay, while per-slot dynamic association used
in [162, 114] gives rise to such delay. Such delay must be taken into account or
made small through a special network architecture, e.g., single-channel single-IP
implementation [22, 117].
We evaluate the performance of our proposed algorithms by extensive simulations
and compare them with a baseline scheme formed by FlashLinQ at the MAC layer and
DASH at the application layer. In particular, we study the system performance in terms
of total throughput, video quality (PSNR), and number of video streaming stalls at the
receivers. According to our simulation results, the proposed algorithm provides sizeable
performance gains for these quality measures.
The remainder of this chapter is organized as follows: Section VI.2 gives preliminaries
and background information. Section VI.3 explains the details of our proposed quality-
aware streaming and scheduling algorithms both for the centralized and the distributed
cases. Section VI.4 shows the simulation results compared to FlashLinQ variants. Sec-
tion VI.5 concludes this chapter.
VI.2 Preliminaries
This section consists of the reference model descriptions including (i) network model (see
Section VI.2.1), (ii) link model (see Section VI.2.2), and (iii) video streaming model (see
Section VI.2.3).
117
TX
1
RX
1
RX
2
TX
2
Interference
Link
l
1
Conflict Graph
l
1
l
2
Link
l
2
Figure VI.1: An Example of a Con
ict Graph
VI.2.1 A Reference Network Model: Macro View
Suppose there are multiple one-hop D2D wireless links [26, 33]. To schedule the D2D
links, a con
ict graph is constructed where the D2D links constitute the nodes, i.e.,
l
i
2L;8i2f1; ;jLjg whereL stands for the set of D2D links. As shown in Fig. VI.1,
two nodes in the con
ict graph are connected via an edge if the D2D links corresponding
to the considered nodes are interfering with each other when they are activated at the
same time t where t2f0; 1;g. Note that this representation implies a protocol model
where two nodes are considered to interfere with each other completely if the signal-to-
interference ratio (SIR) is below a certain threshold, and are not interfering at all if it
exceeds the threshold. The choice of the threshold will be discussed in Section VI.4.2.
However, for the computation of the achievable rates of each link, we still need to take
into account the residual interference, caused by the transmission from simultaneously
scheduled links.
118
Transmitter Receiver
Queue Backlog Size, Q
i
(t)
D2D Link, l
i
(t)
Streaming – Arrival Process:
Placement of Chunks
Streaming – Departure Process:
Transmission of Bits
Scheduling:
Message-Passing for MWIS
Tx

Rx

Figure VI.2: A D2D Link Model
Thus, edge in the con
ict graph between nodel
j
and nodel
k
is denoted asE
(j;k)
where
l
j
2L, l
k
2L, and it is represented as follows:
E
(j;k)
=
8
>
>
>
>
>
<
>
>
>
>
>
:
1; if l
j
interferes with l
k
where
l
j
2L, l
k
2L, and j6=k,
0; otherwise.
(VI.1)
In addition, the set of neighbor nodes of each node is dened as follows:
N (i),

l
a
jE
(i;a)
= 1 where l
a
2L

;8l
i
2L: (VI.2)
VI.2.2 A Reference Link Model: Micro View
As can be seen in Fig. VI.2, each D2D wireless link consists of one transmitter and its
associated receiver. Each transmitter has a queue whose length evolves according to
119
Q
i
(t + 1) = max [0;Q
i
(t)
i
(t)] +
i
(t) (VI.3)
where Q
i
(t),
i
(t), and
i
(t) stand for the queue backlog size at the transmitter of l
i
where l
i
2L at time t, the number of bits leaving the queue of the transmitter of l
i
, and
the number of bits added to the queue of the transmitter of l
i
, respectively. As shown in
Fig. VI.2, the arrival process (bits added to the queue) is associated with the placement
of chunks of the currently served video in each unit time t2f0; 1;g. When the D2D
link is scheduled for transmission by the centralized or distributed controller, the queue
has a departure process which depends on the channel states. More details are provided
in Section VI.3.
VI.2.3 Video Streaming System Model
The receiver of linkl
i
requests a video lef
i
that is located in the cache of its transmitter.
A video le forms a sequence of chunks, i.e., group of pictures (GOPs), which are encoded
and decoded as stand-alone units. Chunks must be reproduced in sequence at the D2D
receivers. The streaming thus consists of the transmission of sequential chunks from the
transmitter to its associated receiver such that the playback buer at each transmitter
contains the required chunks at the beginning of each chunk playback time.
The time scale for the scheduling decision and departure process, i.e., t, is not equiv-
alent to the chunk placement unit time , as can be seen in Fig. VI.3.
A chunk containsN =N
fpc
N
ppf
pixels whereN
ppf
denotes the number of pixels per
frame and N
fpc
stands for the number of frames per chunk. Suppose that each chunk of
each lef is encoded at a number of dierent quality modesq2M whereM =fq
1
q
M
g.
120
According to the variable bit-rate nature of video coding, the quality-rate prole may vary
from chunk to chunk. We let P
f
(q;) andNB
f
(q;) denote the video quality measure
(e.g., peak-signal-to-noise-ratio (PSNR))
y
and the number of bits for lef at chunk time
with quality mode q, respectively.
The chunk placement procedure consists of choosing the quality mode q
i
() of the
chunks requested at chunk unit time by the scheduled D2D transmitters i. The choice
q
i
() renders the choice of the point (P
f
i
(q
i
(););NB
f
i
(q
i
();)) from the nite set of
quality-rate tradeo pointsf(P
f
i
(q;);NB
f
i
(q;))g
q=q
M
q=q
1
. The network controller chooses
the quality modeq
i
() for chunk time for all requesting receiversi, allocates the source
coding rate (bit per pixel)B
f
i
(q
i
();). The transmitter of linki places the corresponding
NB
f
i
(q
i
();) bits in its transmission queueQ
i
(), to be sent to the receiver whose length
evolves according to.
Summarizing, as can be seen in Fig. VI.4, in each chunk time, the transmitter fetches
chunks from its local cache in the order in which they are to be played back. Then the
chunks will be coded based on the the source coding rates B
f
i
(q
i
();) with determined
quality modeq
i
() that is computed based on stochastic network optimization algorithms
(details are in Section VI.3.3). Then, the coded bits are packetized to be transmitted over
the air-interfaces. The enqueued packets will be transmitted depending on the departure
process
i
(t) and are dependent on channel states as well as interference from activated
neighbor D2D links, i.e., signal-to-interference-plus-noise (SINR) ratio.
y
There is a rich literature on video quality metrics, e.g., [112, 23, 99]. Our framework works with any
video quality measure, but for the sake of simplicity (and because details of quality measures are outside
the scope of this chapter), we use PSNR henceforth.
121
Time
Unit Time for Scheduling (t)
Unit Time for the Transmission of Bits as a part of Streaming (t)
Unit Time for the Placement of Chunks
as a part of Streaming ( τ)
……
Figure VI.3: Two Dierentiated Unit Time Scales
Transmitter
Q
i
(t)
μ
i
(t)
Tx

Bitstream
Generator
Quality-
Aware
Compressor
c
1,1
c
1,2
c
i,1
c
i,1
λ
i
(t)
……

……

c
1,|C
1
|
c
i,|C
i
|
Video v
1
Video v
i
…

…

…

Chunk
Fetching
Video
Storage

Figure VI.4: A Device Model
VI.3 Quality-Aware Scheduling and Streaming for Device-
to-Device Video Delivery
This section presents the basic design rationale of our proposed two quality-aware stream-
ing and scheduling algorithms.
VI.3.1 Design Rationale
The proposed centralized algorithm consists of two separable but interconnected parts,
i.e., centralized scheduling with MWIS formulation (refer to Section VI.3.2) and quality-
aware streaming (refer to Section VI.3.3). For the scheduling decision, the given problem
122
is formulated as MWIS and message-passing is used to solve the MWIS problem. For
the streaming decision, the operations to control the arrival and departure processes in
the queue of each D2D transmitter are dened. The entire link model is as illustrated in
Fig. VI.2.
The proposed distributed algorithm improves FlashLinQ with the concepts of dis-
tributed max-weight scheduling before transmission (refer to Section VI.3.2) and quality-
aware streaming (refer to Section VI.3.3).
VI.3.2 Device-to-Device Scheduling
Centralized Scheduling with MWIS Formulation
For centralized D2D scheduling, the objective is to nd the set of nodes that can maximize
the sum of weights of the selected nodes w
i
;8i2f1; ;jLjg, under the constraint that
simultaneously scheduled links must not interfere with each other. This problem is known
to be a maximum weight independent set (MWIS) problem and can be formulated as
follows:
max : F(L;E),
X
8l
i
2L
w
i
I
i
; (VI.4)
s.t. I
j
+I
k
1; ifE
(j;k)
= 1;8l
j
2L;8l
k
2L; (VI.5)
whereI
i
is a boolean index of l
i
;8l
i
2L that is dened as
I
i
=
8
>
<
>
:
1; if l
i
is scheduled where l
i
2L,
0; otherwise
(VI.6)
123
and w
i
;8i2f1; ;jLjg is formulated as follows for max-weight scheduling [13]:
w
i
,r
i
(t)Q
i
(t) (VI.7)
where Q
i
(t) is the queue backlog size at the transmitter of D2D link l
i
, that can be
formulated as (VI.3); and r
i
(t) stands for the maximum achievable rates of D2D link l
i
.
The exact value ofr
i
(t) of D2D linkl
i
cannot be obtained before a scheduling decision
is made because it is unknown which links will be activated, so that the interference
components contributing to the SINR are unknown.
To circumvent this problem, an estimated r
i
(t) can be computed as follows:
r
i
(t) = log
2

1 +
P
s
i
!r
i
(t)kh
i!i
k
2

2
+
!
(VI.8)
whereP
s
i
!r
i
(t) stands for the transmit power from s
i
at unit time t, h
i!i
stands for
the (complex amplitude) channel gain from s
i
to r
i
, is the standard deviation of the
(Gaussian) background noise,
stands for the interference thresholds, i.e., the maximum
admissible interference level
from a single interferer scheduled at the same time as the
considered link. In FlashLinQ,
is set to 9 dB [132, 107]. (VI.8) implies the assumption
that the aggregate interference from other links is equal to the interference from a maxi-
mally strong single interferer. Since the overall interference levels tend to be dominated
by the strongest interferer [70], this is a reasonable approximation.
After solving this MWIS problem, a set of nodes will be obtained as a solution and the
corresponding D2D links can be activated and transmit data from the transmitter of the
D2D link to its receiver. For solving MWIS problems, various heuristic and approximation
124
Figure VI.5: Algorithm: MWIS-based scheduling with message-passing in each l
i
2
L;8i2f1; ;jLjg
algorithms have been proposed due to the fact that MWIS is a well-known NP-hard
problem. One of these methods is the computation with message-passing [120, 76] which
we will use henceforth. The corresponding pseudo-code is presented in Fig. VI.5 with the
weights which are computed by (VI.7).
Distributed Max-Weight Scheduling
For distributed scheduling, we improve FlashLinQ with the concept of max-weight schedul-
ing.
125
FlashLinQ establishes a priority of the D2D wireless links, and lower-priority D2D links
can transmit only if they do not create signicant interference to the higher-priority D2D
links. The decision is based on measurements of channel strengths in both directions [132,
107]. In the original FlashLinQ, priorities are randomized over time, to provide a basic
level of fairness. With the concept of max-weight scheduling, we set the priorities of D2D
links instead as follows:
U
i
,
1
r
i
(t)Q
i
(t)
(VI.9)
where we implement the prioritization by making each link wait a time U
i
before trans-
mission. This means, if the queue backlog size of the transmitter of a D2D link is long,
U
i
becomes short and the link accesses the channel faster than the others. Similarly, if a
D2D link has high achievable rate, i.e., high r
i
(t), it can access the channel faster, thus it
has higher priority.
VI.3.3 Streaming with Quality-Aware Stochastic Control
The streaming consists of two parts, i.e., (i) placement of chunks (i.e., arrival process
of the queue) and (ii) transmission of bits (i.e., departure process of the queue). Notice
that the streaming method investigated in this section is used for both centralized and
distributed algorithms.
Arrival Process (Placement of Chunks)
In each chunk time slot 2f0; 1;g, the transmitter of every given link should place
chunks within its queue. Dierent quality modes are available in each chunk, where
126
obviously a higher quality mode requires more bits. Since knowledge of future rates is
not available, a stochastic chunk placement is performed as follows.
We aim to maximize the total quality over all scheduled links subject to rate stability
of the scheduled transmitter queues. Let P(t) =
P
l
i
2L
P
f
i
(q
i
(t);t); and
P
f
i
(q
i
(t);t) =
8
>
<
>
:
P
f
i
(q
i
(t);t); mod t = 0;
0; mod t6= 0:
(VI.10)
Then the following stochastic optimization formulation is our main objective function:
max lim
t!1
1
t
t1
X
t

=0
E [P(t

)] (VI.11)
subject to lim
t!1
1
t
t1
X
t

=0
E [Q
i
(t

)]<1;8l
i
2L (VI.12)
where (VI.12) means all given queues should fulll mean rate stability. Let (t) denote
the column vector of all scheduled queues at time t, and dene the quadratic Lyapunov
function
L(t) =
1
2

T
(t)(t) =
1
2
X
8i2L
jQ
i
(t)j
2
(VI.13)
where
T
(t) stands for the transpose of (t). Then, let (t) be a conditional quadratic
Lyapunov function that can be formulated as E [L(t + 1)j(t)]L(t), i.e., the drift on
slot t. The drift-plus-penalty (DPP) policy is designed to solve the given optimization
formulation by observing only the current queue backlog sizes (t) and choose q (i.e.,
quality mode) to maximize a bound on
P (t)(t) (VI.14)
127
where is a positive constant control parameter of the DPP policy that aects the
quality-delay tradeos [13].
Now, the quality control decision involves choosing q
i
(t), the quality mode for all
scheduled receivers at chunk time t; and (t) the column vector of the number of source-
coded bitsNB
f
i
(q
i
(t);t) with quality modeq
i
(t) that each receiveri must download from
its transmitter for the chunk requested at time t. These choices are made as
arg max
q
i
(t)2M
[P
f
i
(q
i
(t);t)fNB
f
i
(q
i
(t);t)gQ
i
(t)]: (VI.15)
Since the placement of chunks constitutes the arrival process of the queue,
i
(t) can
be denoted as follows when the optimal q is determined in each l
i
2L:

i
(t) =
8
>
<
>
:
NB
f
i
(q
i
(t);t); mod t = 0;
0; mod t6= 0:
(VI.16)
Departure Process (Transmission of Bits)
Once a set of links is determined to be scheduled by the algorithm in Fig. VI.5 with the
weights of r
i
(t)Q
i
(t),8l
i
2L

in each time t whereL

stands for the set of scheduled
D2D links, the transmitters of scheduled links can transmit bits up to the amount of the
maximum achievable rates, i.e.,
i
(t) =r
i
(t);8l
i
2L

.
According to Shannon's capacity equation,
i
(t) in (VI.7) can be computed as fol-
lows [14]:

i
(t) =B log
2
"
1 +
P
s
i
!r
i
(t)kh
i!i
k
2

2
+
P
j6=i
P
s
j
!r
j
(t)kh
j!i
k
2
#
(VI.17)
128
Table VI.1: Video Trace Information
Basic Information Resolution Average Bitrates
(Names of Test Sequences)
1 highway 352 288 Pixels 631 Kbps (8 Layers Encoded)
2 city, crew, harbour, train 704 576 Pixels 3908 Kbps (4 Layers Encoded)
3 parkrun, stockholm 1024 576 Pixels 6679 Kbps (4 Layers Encoded)
4 bridge close, bridge far 352 288 Pixels 556 Kbps (8 Layers Encoded)
where8l
i
2L

;8l
j
2L

;i6=j,P
sa!r
b
(t) stands for the power transmitted bys
a
intended
forr
b
, andh
j!i
stands for the channel gain from the transmitter of link j to the receiver
of link i where8a;8b2f1; ;jLjg at time t,B stands for the bandwidth of considering
wireless systems.
While here we have used the SINR-based capacity equation in (VI.17) to evaluate
achievable rates, any suitable function of SINR can be included in our algorithms, for
example, if the physical layer (PHY) of the D2D system includes a family of modulation
and coding scheme (MCS) each of which has a certain operational range of SINR and a
given rate, we can substitute such a piece-wise constant function into our scheme and get
meaningful results that explicitly include the properties of the MCS set (e.g., the MCS
modes of 802.11-based standards, or 3GPP LTE).
129
VI.4 Simulation Study
The performance of our proposed joint scheduling and streaming algorithm is simu-
lated and evaluated in this section. The basic simulation settings are presented in Sec-
tion VI.4.1; and the simulation results are presented and analyzed in Section VI.4.2.
VI.4.1 Simulation Settings
Video Traces
For the simulation study, we use four dierent types of video traces. Each video consists
of 14400 chunks where the playback time of each chunk is 0:5 seconds. Thus, the overall
playback time of each video trace is 2 hours, corresponding to a typical movie playback
time. The video sequences are standard moving picture experts group (MPEG) test
sequences, commonly used in the literature. The original video sequences consist of 200
chunks; To create one 2-hour video, we concatenated the same sequence 72 times. Details
of the traces are summarized in Table VI.1.
The quality of each chunk can be numerically represented by the PSNR, i.e.,P
f
i
(q
i
(t);t).
A selected high quality level leads to high PSNR but might negatively impact queue sta-
bility.
We note that the video streams are not synchronized between the D2D links, i.e.,
starting times for the dierent links (streams) are independent of each other.
130
Compared Algorithms
To show the eectiveness of our proposed centralized or distributed quality-aware stream-
ing and scheduling algorithms, their performances are evaluated and compared with the
performances of FlashLinQ variants.
FlashLinQ: This is the standard FlashLinQ algorithm with random U
i
timer se-
lection in each D2D link, such that each link has at least
1
N
probability of being
scheduled where N stands for the number of D2D links. Since FlashLinQ does not
consider video quality at all, we x the quality level as 2 in video trace 1 (among
given 4 levels), 4 in video trace 2 (among given 8 levels), 4 in video trace 3 (among
given 8 levels), and 2 in video trace 4 (among given 4 levels).
FlashLinQ-P: This variant of FlashLinQ uses prioritized U
i
selection. The U
i
are
computed as
U
i
=
1
r
i
(t)Q
i
(t)
(VI.18)
corresponding to the max-weight scheduling concept. Also FlashLinQ-P does not
consider video-quality related aspects, and we again x the qualities of the streams
to the same values as above.
FlashLinQ-Q: This variant of FlashLinQ uses the video streaming quality decisions
as in Section VI.3.3, but the link scheduling is standard FlashLinQ, i.e., the priorities
U
i
are chosen at random.
131
Simulation Topology Construction
The considering simulation topology consists of uniformly random deployed ten D2D
transmitter and receiver pairs in a 600m 600m square layout. The path-loss is computed
according to the Winner II model (indoor in 2.4 GHz D2D communications) [115]:
PL(d) =a
1
log
10
(d) +a
2
+a
3
log
10

f
GHz
c
5

+X

(VI.19)
where f
GHz
c
is the carrier frequency in a GHz scale. a
1
includes the path loss exponent
and its value is 18:7 dBm in LOS and 36:8 dBm in NLOS.a
2
is the intercept, which is 46:8
dBm in LOS and 43:8 dBm in NLOS. a
3
describes the path loss frequency dependence
and it is set to 20 in both LOS and NLOS. X

is the shadowing assumed to be a normal
distribution (in dB) with mean 0 and standard deviation , where = 3 dB in LOS
and = 6 dB in NLOS. Notice that we assume that no communication is possible for a
distance longer than 100 m.
With the given topology, we simulated transmission for 24 hours, assuming each D2D
link streams 10 video traces. Then, ten simulations will be operated with dierent random
geometries.
The connectivity of con
ict graphs changes according to the setting of the interference
threshold. If the received powers from nearby D2D transmittersj to current D2D receiver
i are less than the interference threshold, they will be considered as noise. Otherwise,
they will be considered as interference. Thus, low
increases the number of edges in the
corresponding con
ict graph. Consequently, a relatively small number of D2D links will
be scheduled; that may also reduce the sum rate. On the other hand, high
decreases the
number of edges in the given corresponding con
ict graph. Thus, a relatively large number
132
……
c
1
c
2
Tx

Rx

All bits of
chunk k are arrived
Playback:
Chunk k
c
1
All bits of
chunk (k+1) are arrived
Playback:
Chunk (k+1)
c
1
All bits of
chunk (k+2) are arrived
Playback:
Chunk (k+2)
Chunk Gap > 0
Stall Events!
Chunk Gap = 0
No Stall Events
Figure VI.6: An example of stall events: If all bits of next chunk are arrived at the
playback buer of D2D receiver, there is no stall event since the receiver can immediately
play the next chunk when the playing of current chunk is completed. Otherwise, the stall
event will be occurred when the playing of current chunk is completed.
of D2D links can be scheduled; however it will reduce the signal-to-interference-plus-noise
ratio and thus rate, for each link. Therefore, we need to consider various interference
threshold settings in the simulation studies.
Performance Analysis Metrics
Whenever a receiver nishes playing back chunk i, all bits of chunk i + 1 should have
arrived. Otherwise, a pause (stall) occurs in the playback, see Fig. VI.6. Obviously, the
stall events introduce user dissatisfaction. Typically, 3 or more stalls during one movie
133
All bits of
Chunk 1
are arrived
Playback
Chunk 3
Time
Playback
Chunk 2
Playback
Chunk 1
……
All bits of
Chunk 2
are arrived
All bits of
Chunk 3
are arrived
Stall
Event
Stall
Event
(a) No pre-buering: If there is no pre-buering,
stall events may be occurred between chunks.
All bits of
Chunk 1
are arrived
Playback
Chunk 3
Time
Playback
Chunk 2
Playback
Chunk 1
……
All bits of
Chunk 2
are arrived
All bits of
Chunk 3
are arrived
Pre-
buffering
(b) Pre-buering: By setting certain amounts of
pre-buering time, we can reduce the number of
stall events between chunks.
Figure VI.7: Pre-buering: With the denition of pre-buering, the number of stall events
can be reduced, i.e., user satisfaction can be increased.
would be judged to be unacceptable quality. Thus the number of stall events can be an
important index for measuring user satisfaction of video streaming.
To avoid stall events, pre-buering is frequently used. As shown in Fig. VI.7, a num-
ber of chunks are transmitted before playback at the receiver starts; this introduces a
viewing delay for the user but reduces the number of stall events. Obviously a very long
pre-buering time leads to user dissatisfaction as well, and if we set extremely large pre-
buering time, there is no dierence between streaming and downloading/transmission.
134
Therefore, there exists a tradeo, and dening an appropriate pre-buering time is re-
quired. In addition, once a stall occurs, the receiver buers again for the same amount of
pre-buering time as for startup
z
.
VI.4.2 Simulation Results
With the given two performance metrics, we evaluates the performance of our proposed
algorithms and three various FlashLinQ variants as a function of the following parameters:
various pre-buering time settings (see Section VI.4.2),
various , i.e., quality-delay tradeos (see Section VI.4.2),
various interference thresholds
(see Section VI.4.2), and
average quality vs. the expected number of stall events (see Section VI.4.2).
Notice that our proposed centralized algorithm is denoted as mpMWIS-QP (i.e.,
message-passing for MWIS formulation with Quality-awareness and max-weight Prioritization);
and also our proposed distributed algorithm is denoted as FlashLinQ-QP (i.e., improved
FlashLinQ with Quality-awareness and max-weight Prioritization).
Various Pre-Buering Time Settings
We vary the pre-buering time from 1 second to 10 seconds with a step size of 1 second.
In addition, and the interference threshold
are set to 2 and 5 dB, respectively. The
z
Notice that the re-buering time could also be shorter than the initial buering time, i.e., they do not
need to be the same, but we set it as equal in order to reduce the number of variables in the simulation.
135
Table VI.2: The Expected Number of Stall Events, i.e., E [N
s
], in Each Video Streaming
in Each D2D Link with Various Pre-Buering Times
Pre-
buering mpMWIS-QP: FlashLinQ: FlashLinQ-P: FlashLinQ-Q: FlashLinQ-QP:
Time E [N
s
] E [N
s
] E [N
s
] E [N
s
] E [N
s
]
1 second 13.4 34.8 26.0 26.6 14.5
2 second 11.1 30.2 20.9 22.9 11.9
3 second 7.9 28.5 15.5 18.3 8.9
4 second 5.0 21.8 10.2 11.3 6.1
5 second 2.1 11.1 4.4 5.7 2.7
6 second 0.9 6.0 2.2 3.0 1.1
7 second 0.7 4.6 1.6 2.4 0.8
8 second 0 3.7 1.2 1.9 0.3
9 second 0 3.3 0.6 1.1 0
10 second 0 2.6 0 0.3 0
resulting expected stall probability is presented in Table VI.2 and Fig. VI.8. For quanti-
tative comparison, two indices, i.e.,T
s
andM
s
, are dened as formulated in (VI.20) and
(VI.21), respectively.
T
s
=
E [N
s
] of FlashLinQ VariantshE [N
s
]i
hE [N
s
]i
(VI.20)
M
s
= E [N
s
] of FlashLinQ VariantshE [N
s
]i (VI.21)
136
whereE [N
s
] stands for the expected number of stall events andhE [N
s
]i denotes theE [N
s
]
of mpMWIS-QP.
As anticipated, the expected number of stall events reduces as the pre-buering time
increases. If there is no pre-buering time, mpMWIS-QP has 13.4 stall events on average
whereas FlashLinQ has 34.8. For 8 second pre-buering time, mpMWIS-QP has no
stall events; FlashLinQ-QP has 0.3; and this performance is the best among the given
three FlashLinQ variants. Pure FlashLinQ shows the lowest performance. As shown in
Fig. VI.8(c), it has 6 times more stall events when the pre-buering time is 7 second.
We furthermore see that the performance advantage of mpMWIS-QP vs. FlashLinQ
stems from a variety of factors: max-weight scheduling, incorporation of the intercon-
nection between scheduling and streaming, and centralized control. We see that only
incorporating max-weight scheduling (i.e., going from FlashLinQ to FlashLinQ-P) pro-
vides a signicant advantage, while only incorporating video quality without max-weight
scheduling (i.e., going from FlashLinQ to FlashlinQ-P gives slightly lower improvement.
The advantage of centralized scheduling over distributed scheduling (mpMWIS-QP) is
very small, which is an important insight for actual deployment.
Various
stands for a quality-delay tradeo constant as formulated in (VI.14). If is small,
quality-awareness takes higher priority. On the other hand, larger considers queue
stability with higher priority, so that lower probability of stalls can be anticipated. The
simulation results in this section investigate results when takes on the values 0:05,
0:1, 2, 4, and 8. Notice that our considered pre-buering time is 8 seconds, which is
137
Table VI.3: The Expected Number of Stall Events, i.e., E [N
s
], in Each Video Streaming
in Each D2D Link with Various
mpMWIS-QP: FlashLinQ: FlashLinQ-P: FlashLinQ-Q: FlashLinQ-QP:
E [N
s
] E [N
s
] E [N
s
] E [N
s
] E [N
s
]
8 0 0.8 0.3 0.5 0
4 0 1.6 1.1 0.9 0.1
2 0 3.7 1.2 1.9 0.3
0:1 1.2 7.0 3.7 4.2 1.6
0:05 3.0 15.3 8.4 9.0 4.7
Table VI.4: The Expected Number of Stall Events, i.e., E [N
s
], in Each Video Streaming
in Each D2D Link with Various Interference Thresholds

mpMWIS-QP: FlashLinQ: FlashLinQ-P: FlashLinQ-Q: FlashLinQ-QP:
E [N
s
] E [N
s
] E [N
s
] E [N
s
] E [N
s
]
0 dB 5.4 26.9 20.8 22.0 17.9
5 dB 0 3.7 1.2 1.9 0.3
13 dB 3.1 23.8 13.0 16.2 10.2
an optimum value for mpMWIS-QP; and the interference threshold
is set to 5 dB. A
further discussion of video quality versus stall events will be given in Section VI.4.2.
The simulation results are summarized in Table VI.3. If is 2, 4, or 8, there are no
stall events in mpMWIS-QP. Pure FlashLinQ has between 3.7 and 0.8 stall events for
those values. FlashLinQ-QP will have no stalls when = 8. This performance is lower
than the performance of mpMWIS-QP; however the performance of FlashLinQ-QP is the
best among the given FlashLinQ variants.
138
Table VI.5: PSNR Table of Given Four Video Traces
minimum PSNR Maximum PSNR
Video Trace 1 29.4835 dB 37.8063 dB
Video Trace 2 25.7136 dB 36.2584 dB
Video Trace 3 24.3273 dB 35.0470 dB
Video Trace 4 28.8283 dB 37.1691 dB
Various Interference Thresholds
As discussed in Section VI.4.1, the interference threshold trades o the number of active
links with the rate per link that can be obtained. Thus, nding an appropriate interference
threshold is important for optimizing performance. In our given network geometry, our
minimum and maximum interference thresholds are 0 dB and 13 dB, respectively. With

= 0 dB, our geometry is extremely densely connected in its corresponding con
ict
graph, i.e., only one D2D link will be scheduled in each unit time. On the other hand,
our geometry has no edges in its corresponding con
ict graph when
= 13 dB, i.e., all
D2D links will be scheduled and will generate interference all together in each unit time.
We additionally performed the simulation with
= 5 dB.
We again set the pre-buering time to 8 second; and set = 2. Results with the
mentioned three interference thresholds are listed in Table VI.4. For all algorithms, per-
formance with
= 5 dB is the best; and the performance with
= 0 dB is lower than the
performance with
= 13 dB.
139
Table VI.6: The Expected Number of Stall Events E [N
s
] vs. Average Quality (Average
PSNR) when the System Bandwidth is 2 MHz
mpMWIS-QP FlashLinQ-QP FlashLinQ-Q
(E [N
s
], Average PSNR) (E [N
s
], Average PSNR) (E [N
s
], Average PSNR)
0.05 (3.02, 32.84) (4.75, 32.78) (9.07, 31.48)
0.1 (1.15, 32.37) (1.58, 32.10) (4.18, 31.04)
2 (0.00, 31.85) (0.29, 31.44) (1.87, 30.75)
4 - (0.14, 31.15) (0.86, 30.54)
8 - (0.00, 30.91) (0.58, 29.88)
Table VI.7: The Expected Number of Stall Events E [N
s
] vs. Average Quality (Average
PSNR) when the System Bandwidth is 1 MHz
mpMWIS-QP FlashLinQ-QP FlashLinQ-Q
(E [N
s
], Average PSNR) (E [N
s
], Average PSNR) (E [N
s
], Average PSNR)
0.05 (6.29, 29.10) (10.50, 27.89) (20.95, 27.28)
0.1 (2.39, 27.12) (5.41, 26.99) (9.65, 26.78)
2 (0.89, 26.82) (2.55, 26.78) (4.32, 26.68)
4 (0.39, 26.72) (0.95, 26.68) (2.00, 26.67)
8 (0.16, 26.68) (0.32, 26.68) (1.32, 26.67)
Average Quality vs. Expected Number of Stall Events
This section presents average quality values depending on the expected number of stall
events for mpMWIS-QP, FlashLinQ-Q, and FlashLinQ-QP. The other two FlashLinQ
variants, i.e., FlashLinQ and FlashLinQ-P, are not considered in this simulation since they
statically select their quality mode. Pre-buering time is 8 second, and the interference
140
threshold is 5 dB. To numerically represent video quality, PSNR is used; the minimum
and maximum PSNR values in each video trace are listed in Table VI.5.
We simulate the three algorithms with =f0:05; 0:1; 2; 4; 8g, and compute the ex-
pected numbers of stall events and the average PSNR values. Results are shown for a
system bandwidth of 2 MHz in Table VI.6. In mpMWIS-QP, there are no stall events
if 2. However, higher leads to the degradation of average PSNR to guarantee
more stability on the D2D transmitter queue. If there are no stall events, there is no
need to improve the stability of the D2D transmitter queue. Thus, simulation results of
mpMWIS-QP where = 4 and = 8 are not shown. In addition, FlashLinQ-QP has no
stall events when = 8.
Fig. VI.9(a) shows that our mpMWIS-QP provides the highest PSNR for a given stall
probability; and the performance of FlashLinQ-QP is approximately 0:4 dB lower than the
performance of mpMWIS-QP. However, FlashLinQ-Q shows around 1:6 dB lower PSNR
compared to the PSNR of mpMWIS-QP.
Results for a system bandwidth of 1 MHz are in Table VI.7. Due to the lower band-
width, all the three evaluated algorithms have stall events. Similar to the cases of 2 MHz
system bandwidth, higher leads to the degradation of average PSNR to guarantee more
stability on the D2D transmitter queue. Fig. VI.9(b) shows that mpMWIS-QP again
provides the highest PSNR; and the PSNR of FlashLinQ-QP is approximately 1:5 dB
lower than the performance of mpMWIS-QP when the expected number of stall events is
near 5. However, FlashLinQ-Q shows around 2:3 dB lower PSNR compared to the PSNR
of mpMWIS-QP. According to Table VI.7, the lowest PSNR is near 26:7, which is close
to the PSNR of the lowest-quality mode available.
141
As observed in Fig. VI.9, a higher expected number of stall events is associated with
a higher PSNR (i.e., video quality). For guaranteeing more video quality, lower is used
for putting more weights on PSNR and lower weights on queue stability. Therefore, there
are more possibilities to increase the queue backlog sizes at D2D transmitters, i.e., this
leads to higher expected number of stall events at D2D receivers.
VI.5 Concluding Remarks
This chapter proposed centralized or distributed quality-aware streaming and scheduling
algorithms which can be used for device-to-device video delivery applications. In terms
of scheduling, we proposed two approaches, i.e., centralized and distributed ones. For
centralized scheduling, a message-passing based algorithm is used to obtain the solutions
from a maximum independent set problem formulation. For distributed scheduling, we
improved a FlashLinQ D2D scheduler with the principle of max-weight scheduling. In
terms of streaming, a quality-aware stochastic chunk selection algorithm is introduced that
works based on the queue backlog sizes in each D2D transmitter queue. The stochastic
algorithm in the streaming part controls the quality of each video chunk to maximize
the qualities of streamed video subject to queue rate stability. We can draw several
important conclusions from the simulations: (i) it is essential to use a transmission scheme
that accounts for the inter-relationship between scheduling and quality selection, (ii) a
good distributed scheme performs only marginally worse than our centralized scheme, and
(iii) we can trade o average video quality with probability of stalls. These results give
important insight in the deployment of D2D-based video streaming.
142
(a) Eectiveness of mpMWIS-QP
compared to FlashLinQ-QP in
terms ofTs dened in (VI.20)
(b) Eectiveness of mpMWIS-QP
compared to FlashLinQ-QP in
terms ofMs dened in (VI.21)
(c) Eectiveness of mpMWIS-QP
compared to FlashLinQ in terms
ofTs dened in (VI.20)
(d) Eectiveness of mpMWIS-QP
compared to FlashLinQ in terms
ofMs dened in (VI.21)
Figure VI.8: Performance Comparison between mpMWIS-QP and FlashLinQ-
QP/FlashLinQ in terms of the Expected Number of Stall Events
143
(a) System Bandwidth: 2 MHz (The data
in this gure is listed in Table VI.6)
(b) System Bandwidth: 1 MHz (The data
in this gure is listed in Table VI.7)
Figure VI.9: Average Quality vs. Expected Number of Stall Events
144
Chapter VII
Conclusions
In this dissertation, we explore two major communication technologies for next-generation
wireless video systems, i.e., millimeter-wave wireless technologies and device-to-device
proximal wireless technologies.
For millimeter-wave systems research, we developed, implemented, and demonstrated
novel small cell architectures that can increase capacity for next generation 5G cellular
systems as presented in Chapter II. Following Chapter III and Chapter IV are separately
discussing about two major research challenges and proposing corresponding solutions
in millimeter-wave wireless systems, i.e., relaying and beam training. In Chapter III,
we proposed a joint video compression and relaying algorithm that can jointly optimize
compression ratios and relay selection in 60 GHz outdoor broadcasting stadiums. In
Chapter IV, we proposed a fast millimeter-wave beam training algorithm that can signi-
cantly speed up link conguration time compared to standardized two-level beam training
methods.
For device-to-device video streaming research, rst of all, we implemented our theo-
retical results on top of Android mobile platforms. The proposed algorithm in [162, 114]
had limitations on WiFi-based implementation, thus we resolved the issues by proposing
145
greedy heuristics; and demonstrated working prototypes as described in Chapter V. In
Chapter VI, we proposed a new quality-aware scheduling (with max-weight independent
set formulation) and streaming (with stochastic network optimization with quality-delay
tradeos) algorithm for the xed source and destination D2D pairs.
146
BIBLIOGRAPHY
[1] WirelessHD; http://wirelessshd.org
[2] Wireless Gbit Alliance (WiGig); http://wirelessgbitalliance.org
[3] Comotech Corporation; http://www.comotech.com/en/index.html
[4] Millitech; http://www.millitech.com
[5] http://ffmpeg.org/
[6] https://www.youtube.com/watch?v=uRqwnCWDCsw
[7] Qualcomm Press; Qualcomm Atheros and Wilocity Announce Tri-band Wi-Fi: In-
dustrys First Standards-compliant, Multi-gigabit Wireless Chipset, May 2011.
[8] http://www.androidpit.com/en/android/blog/401283/
Are-You-the-1-Top-Mobile-Users-Consume-Half-World-s-Bandwidth
[9] RealWireless on behalf of Ofcom, \4G Capacity Gains," http://stakeholders.
ofcom.org.uk/market-data-research/other/technology-research/
2011/4G-Capacity-Gains/
[10] RealWireless on behalf of Ofcom, \Techniques for Increasing the Capacity of Wire-
less Broadband Networks: UK, 2012-2030," http://www.realwireless.biz/
mobile-capacity-in-the-uk-major-study-published/
[11] T.S. Rappaport, \The Massively Broadband Future," Spectrum 20/20 Conference,
June 2012., http://www.spectrum2020.ca/presentations/Rappaport.
pdf
[12] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press,
2004.
[13] M.J. Neely, Stochastic Network Optimization with Application to Communication
and Queueing Systems, Morgan & Claypool, 2010.
[14] A.F. Molisch, Wireless Communications, 2nd Ed., Wiley-IEEE, 2011.
147
[15] S.-K. Yong, \TG3c Channel Modeling Sub-committee Final Report," IEEE 15-07-
0584-01-003c, March 2007.
[16] IEEE 802.15.3c Millimeter-wave-based Alternative Physical Layer Extension; http:
//www.ieee802.org/15/pub/TG3c.html, October 2009.
[17] L. Cariou, et. al., \Fast Session Transfer," IEEE 802.11-10/491r2, May 2010.
[18] IEEE802.11-12/0140r4, May 2012.
[19] Samsung, \Technologies for Rel-12 and Onwards," 3GPP Workshop, RWS-120021,
June 2012.
[20] IEEE 802.11ad VHT Specication Version 1.0, December 2012.
[21] P. Smulders, \Exploiting the 60 GHz Band for Local Wireless Multimedia Access:
Prospects and Future Directions," IEEE Communications Magazine, vol. 40, no. 1,
pp. 140 - 147, January 2002.
[22] A. Mishra, M. Shin, and W. Arbaugh, \An Empirical Analysis of the IEEE 802.11
MAC Layer Hando Process," ACM SIGCOMM Computer Communications Review,
vol. 33, no. 2, pp. 93 - 102, April 2003.
[23] S. Winkler and P. Mohandas, \The Evolution of Video Quality Measurement: From
PSNR to Hybrid Metrics," IEEE Transactions on Broadcasting, vol. 54, no. 3, 2008.
[24] H. Singh, J. Oh, C. Kweon, X. Qin, H.-R. Shao, and C. Ngo, \A 60 GHz Wireless
Network for Enabling Uncompressed Video Communication," IEEE Communications
Magazine, vol. 46, no. 12, pp. 71 - 78, December 2008.
[25] H. Wang, K.-Y. Lin, Z.-M. Tsai, L.-H. Lu, H.-C. Lu, C.-H. Wang, J.-H. Tsai, T.-W.
Huang, and Y.-C. Lin, \MMICs in the Millimeter-Wave Regime," IEEE Microwave
Magazine, vol. 10, no. 1, pp. 99 - 117, February 2009.
[26] K. Doppler, M. Rinne, C. Wijting, C.B. Ribeiro, and K. Hugl, \Device-to-Device
Communication as an Underlay to LTE-Advanced Networks," IEEE Communications
Magazine, vol. 47, no. 12, pp. 42 - 49, December 2009.
[27] A. Niknejad, \Siliconization of 60 GHz," IEEE Microwave Magazine, vol. 11, no. 1,
pp. 78 - 85, February 2010.
[28] R. Daniels, J.N. Murdock, T. Rappapport, and R.W. Heath, Jr., \State-of-the-Art
in 60 GHz," IEEE Microwave Magazine, vol. 11, no. 7, pp. 44 - 50, December 2010.
[29] A. Begen, T. Akgul, and M. Baugher, \Watching Video over the Web: Part
1: Streaming Protocols," IEEE Internet Computing, vol. 15, no. 2, pp. 54 63,
March/April 2011.
148
[30] J. Pi and F. Khan, \An Introduction to Millimeter-Wave Mobile Broadband Sys-
tems," IEEE Communications magazine, vol. 49, no. 6, pp. 101 - 107, June 2011.
[31] T. Baykas, C.-S. Sum, Z. Lan, J. Wang, M.A. Rahman, H. Harada, and S. Kato,
\IEEE 802.15.3c: The First IEEE Wireless Standard for Data Rates over 1 Gb/s,"
IEEE Communications Magazine, vol. 49, no. 7, pp. 114 - 121, July 2011.
[32] G. Fodor, E. Dahlman, G. Mildh, S. Parkvall, N. Reider, G. Miklos, and Z. Tu-
ranyi, \Design Aspects of Network Assisted Device-to-Device Communications,"
IEEE Communications Magazine, vol. 50, no. 3, pp. 170 - 177, March 2012.
[33] L. Lei, Z. Zhong, C. Lin, and X. Shen, \Operator Controlled Device-to-Device Com-
munications in LTE-Advanced Networks," IEEE Communications Magazine, vol. 50,
no. 6, pp. 96 - 104, June 2012.
[34] M.J. Yang, S.Y. Lim, H.J. Park, and N.H. Park, \Solving the Data Overload: Device-
to-Device Bearer Control Architecture for Cellular Data Ooading," IEEE Vehicular
Technology Magazine, vol. 8, no. 1, pp. 31 - 39, March 2013.
[35] N. Golrezaei, A.F. Molisch, A.G. Dimakis, and G. Caire, \Femtocaching and Device-
to-Device Collaboration: A New Architecture for Wireless Video Distribution," IEEE
Communications Magazine, vol. 51, no. 4, pp. 142 - 149, April 2013.
[36] D. Camps-Mur, A. Garcia-Saavedra, and P. Serrano, \Device-to-Device Communi-
cations with Wi-Fi Direct: Overview and Experimentation," IEEE Wireless Com-
munications, vol. 20, no. 3, pp. 96 - 104, June 2013.
[37] W. Roh, J.-Y. Seol, J. Park, B. Lee, J. Lee, Y. Kim, J. Cho, K. Cheun, and F. Aryan-
far, \Millimeter-Wave Beamforming as an Enabling Technology for 5G Cellular Com-
munications: Theoretical Feasibility and Prototype Results," IEEE Communications
Magazine, vol. 52, no. 2, pp. 106 - 113, February 2014.
[38] C. Cordeiro, A. Sadri, A. Maltsev, R. Are, J. Kim, and L. Xian, \Millimeter Wave
Capable Small Cells for Next Generation 5G Cellular Systems," IEEE Communica-
tions Magazine (Under Review), 2014.
[39] Y. Chang, H.J. Kuno, and D.L. English, \High Data-Rate Solid-State Millimeter-
Wave Transmitter Module," IEEE Transactions on Microwave Theory and Tech-
niques, vol. 23, no. 6, pp. 470 - 477, June 1975.
[40] A.A. Abidi, \Direct-Conversion Radio Transceivers for Digital Communications,"
IEEE Journal of Solid-State Circuits, vol. 30, no. 12, pp. 1399 - 1410, December
1995.
[41] P.F.M. Smulders and L.M. Correia, \Characterization of Propagation in 60 GHz
Radio Channels," Electronics Communication Engineering Journal, vol. 9, no. 2, pp.
73 - 80, April 1997.
149
[42] B. Razavi, \Design Considerations for Direct-Conversion Receivers," IEEE Transac-
tions on Circuits and Systems II: Analog and Digital Signal Processing, vol. 44, no.
6, pp. 428 - 435, June 1997.
[43] C.R. Anderson and T.S. Rappaport, \In-Building Wideband Partition Loss Measure-
ments at 2.5 and 60 GHz," IEEE Transactions on Wireless Communications, vol. 3,
no. 3, pp. 922 - 928, May 2004.
[44] J. Chakareski, J.G. Apostolopoulos, S. Wee, W.-T. Tan, and B. Girod, \Rate-
distortion hint tracks for adaptive video streaming," IEEE Transactions on Circuits
and Systems for Video Technology, vol. 15, no. 10, pp. 1257 - 1269, 2005.
[45] C.H. Doan, S. Emami, A.M. Niknejad, and R.W. Brodersen, \Millimeterwave CMOS
Design," IEEE Journal of Solid-State Circuits, vol. 40, no. 1, pp. 144 - 155, January
2005.
[46] B.A. Floyd, S.K. Reynolds, U.R. Pfeier, T. Zwick, T. Beukema, and B. Gaucher,
\SiGe Bipolar Transceiver Circuits Operating at 60 GHz," IEEE Journal of Solid-
State Circuits, vol. 40, no. 1, pp. 156 - 167, January 2005.
[47] T.O. Dickson, M.-A. LaCroix, S. Boret, D. Gloria, R. Beerkens, and S.P. Voinigescu,
\30-100-GHz Inductors and Transformers for Millimeter-Wave (Bi)CMOS Integrated
Circuits," IEEE Transactions on Microwave Theories and Techniques, vol. 53, no. 1,
pp. 123 - 133, January 2005.
[48] N. Deparis, A. Bendjabballah, A. Boe, M. Fryziel, C. Loyez, L. Clavier, N. Rolland,
and P.-A. Rolland, \Transposition of a Baseband UWB Signal at 60 GHz for High
Data Rate Indoor WLAN," IEEE Wireless and Microwave Component Letters, vol.
15, no. 10, pp. 609 611, October 2005.
[49] D. Cabric, M.S.W. Chen, D.A. Sobel, S. Wang, J. Yang, and R.W. Brodersen, \Novel
Radio Architectures for UWB, 60 GHz, and Cognitive Wireless Systems," EURASIP
Journal on Wireless Communications and Networking, Article ID: 17957, 2006.
[50] H.-Y. Chang, P.-S. Wu, T.-W. Huang, H. Wang, C.-L. Chang, and J.G.J. Chern,
\Design and Analysis of CMOS Broad-Band Compact High-Linearity Modulators for
Gigabit Microwave/Millimeterwave Applications," IEEE Transactions on Microwave
Theory and Techniques, vol. 54, no. 1, pp. 20 - 30, January 2006.
[51] B. Razavi, \A 60-GHz CMOS Receiver Front-End," IEEE Journal of Solid-State
Circuits, vol. 41, no. 1, pp. 17 - 22, January 2006.
[52] M.M. Akbar, M.S. Rahman, M. Kaykobad, E.G. Manning, and G.C. Shoja, \Solving
the Multidimensional Multiple-Choice Knapsack Problem by Constructing Convex
Hulls," Computers and Operations Research, vol. 33, no. 5, pp. 1259 - 1273, May
2006.
150
[53] F. Dai and J. Wu, \Ecient Broadcasting in Ad Hoc Wireless Networks using Direc-
tional Antennas," IEEE Transactions on Parallel and Distributed Systems, vol. 17,
no. 4, pp. 335 - 347, August 2006.
[54] A. Babakhani, X. Guan, A. Komijani, A. Natarajan, and A. Hajimiri, \A 77-GHz
Phased-Array Transceiver with On-Chip Antennas in Silicon: Receiver and Anten-
nas," IEEE Journal of Solid State Circuits, vol. 41, no. 12, pp. 2795 - 2807, December
2006.
[55] A. Natarajan, A. Komijani, X. Guan, A. Babakhani, and A. Hajimiri, \A 77-GHz
Phased-Array Transceiver with On-Chip Antennas in Silicon: Transmitter and Local
LO-Path Phase Shifting," IEEE Journal of Solid State Circuits, vol. 41, no. 12, pp.
2807 - 2819, December 2006.
[56] S.K. Reynolds, B.A. Floyd, U.R. Pfeier, T. Beukema, J. Grzyb, C. Haymes, B.
Gaucher, and M. Soyuer, \A Silicon 60-GHz Receiver and Transmitter Chipset for
Broadband Communications," IEEE Journal of Solid-State Circuits, vol. 41, no. 12,
pp. 2820 - 2831, December 2006.
[57] N. Guo, R.C. Qiu, S.S. Mo, and K. Takahashi, \60-GHz Millimeter-Wave Radio:
Principle, Technology, and New Results," EURASIP Journal on Wireless Commu-
nications and Networking, Article ID: 68253, 2007.
[58] S.-K. Yong and C.-C. Chong, \An Overview of Multigigabit Wireless through Mil-
limeter Wave Technology: Potentials and Technical Callenges," EURASIP Journal
on Wireless Communications and Networking, Article ID: 78907, 2007.
[59] B.B.M.W. Badalawa and M. Fujishima, \60 GHz CMOS Pulse Generator," IET
Electronics Letters, vol. 43, no. 2, pp. 100 - 102, January 2007.
[60] S. Mao, X. Cheng, Y.T. Hou, H.D. Sherali, and J. Reed, \On Joint Routing and
Server Selection for MD Video Streaming in Ad Hoc Networks," IEEE Transactions
on Wireless Communications, vol. 6, no. 1, pp. 338 - 347, January 2007.
[61] H. Zhai, X. Chen, and Y. Fang, \Improving Transport Layer Performance in Multihop
Ad Hoc Networks by Exploiting MAC Layer Information," IEEE Transactions on
Wireless Communications, vol. 6, no. 5, pp. 1692 - 1701, May 2007.
[62] B. Razavi, \A Millimeter-Wave CMOS Heterodyne Receiver with Onchip LO and
Divider," IEEE Journal of Solid State Circuits, vol. 43, no. 2, pp. 477 - 485, February
2008.
[63] N. Deparis, C. Loyez, N. Rolland, and P.-A. Rolland, \UWB in Millimeter Wave
Band with Pulsed ILO," IEEE Transactions on Circuits and Systems II: Express
Briefs, vol. 55, no. 4, pp. 339 - 343, April 2008.
151
[64] T. Mitomo, R. Fujimoto, N. Ono, R. Tachibana, H. Hoshino, Y. Yoshihara, Y. Tsut-
sumi, and I. Seto, \A 60-GHz CMOS Receiver Front-End with Frequency Synthe-
sizer," IEEE Journal of Solid-State Circuits, vol. 43, no. 4, pp. 1030 - 1037, April
2008.
[65] E. Laskin, P. Chevalier, A. Chantre, B. Sautreuil, and S. P. Voinigescu, \165-GHz
Transceiver in SiGe Technology," IEEE Journal of Solid State Circuits, vol. 43, no.
5, pp. 1087 - 1100, May 2008.
[66] E. Moldovan, S.O. Tatu, and S. Aes, \A 60 GHz Multiport Front-End Architecture
with Integrated Phased Antenna Array," Microwave and Optical Technology Letters,
vol. 50, no. 5, pp. 1371 - 1376, May 2008.
[67] M.-C. Hwang, J.-H. Kim, D.T. Duong, and S.-J. Ko, \Hybrid Temporal Error Con-
cealment Methods for Block-Based Compressed Video Transmission," IEEE Trans-
actions on Broadcasting, vol. 54, no. 2, pp. 198 - 207, June 2008.
[68] H. Mansour, V. Krishnamurthy, and P. Nasiopoulos, \Channel Aware Multiuser
Scalable Video Streaming Over Lossy Under-Provisioned Channels: Modeling and
Analysis," IEEE Trasactions on Multimedia, vol. 10, no. 7, November 2008.
[69] K. Scheir, S. Bronckers, J. Borremans, P. Wambacq, and Y. Rolain, \A 52 GHz
Phased-Array Receiver Front-End in 90nm Digital CMOS," IEEE Journal of Solid-
State Circuits, vol. 43, no. 12, pp. 2651 - 2659, December 2008.
[70] M.Z. Win, P.C. Pinto, and L.A. Shepp, \A Mathematical Theory of Network Inter-
ference and Its Applications," Proceedings of the IEEE, vol. 97, no. 2, pp. 205 - 230,
February 2009.
[71] W. Wei and A. Zakhor, \Interference Aware Multipath Selection for Video Streaming
in Wireless Ad Hoc Networks," IEEE Transactions on Circuits and Systems for Video
Technology, vol. 19, no. 2, pp. 165 - 178, February 2009.
[72] A. Parsa and B. Razavi, \A New Transceiver Architecture for the 60-GHz Band,"
IEEE Journal of Solid-State Circuits, vol. 44, no. 3, pp. 751 - 762, March 2009.
[73] D.A. Sobel and R.W. Brodersen, \A 1 Gb/s Mixed-Signal Baseband Analog Front-
End for a 60 GHz Wireless Receiver," IEEE Journal of Solid-State Circuits, vol. 44,
no. 4, pp. 1281 - 1289, April 2009.
[74] A. Oncu and M. Fujishima, \49mW 5 Gbit/s CMOS Receiver for 60 GHz Impulse
Radio," IET Electronics Letters, vol. 45, no. 17, pp. 889 - 890, August 2009.
[75] A. Tomkins, R.A. Aroca, T. Yamamoto, S.T. Nicolson, Y. Doi, and S.P. Voinigescu,
\A Zero-IF 60 GHz 65nm CMOS Transceiver with Direct BPSK Modulation Demon-
strating up to 6 Gb/s Data Rates over a 2m Wireless Link," IEEE Journal of Solid-
State Circuits, vol. 44, no. 8, pp. 2085 - 2099, August 2009.
152
[76] S. Sanghavi, D. Shah, and A.S. Willsky, \Message Passing for Maximum Weight
Independent Set," IEEE Transactions on Information Theory, vol. 55, no. 11, pp.
4822 - 4834, October 2009.
[77] J. Wang, Z. Lan, C.W. Pyo, T. Baykas, C.-S. Sum, M.A. Rahman, J. Gao, R. Funada,
F. Kojima, H. Harada, and S. Kato, \Beam Codebook Based Beamforming Protocol
for Multi-Gbps Millimeter-Wave WPAN Systems," IEEE Journal on Selected Areas
in Communications, vol. 27, no. 8, pp. 1390 - 1399, October 2009.
[78] C.-S. Wang, J.-W. Huang, K.-D. Chu, and C.-K. Wang, \A 60-GHz Phased Array
Receiver Front-End in 0.13-um CMOS Technology," IEEE Transactions on Circuits
and Systems I: Regular Papers, vol. 56, no. 10, pp. 2341 - 2352, October 2009.
[79] D. Dawn, P. Sen, S. Sarkar, B. Perumana, S. Pinel, and J. Laskar, \60-GHz Integrated
Transmitter Development in 90-nm CMOS," IEEE Transactions on Microwave The-
ory and Techniques, vol. 57, no. 10, pp. 2354 - 2367, October 2009.
[80] S. Geng, \Performance and Capacity Analysis of 60 GHz WPAN Channel," Mi-
crowave and Optical Technology Letters, vol. 51, no. 11, pp. 2671 - 2675, November
2009.
[81] C. Marcu, D. Chowdhury, C. Thakkar, J.-D. Park, L.-K. Kong, M. Tabesh, Y. Wang,
B. Afshar, A. Gupta, A. Arbabian, S. Gambini, R. Zamani, E. Alon, and A.M.
Niknejad, \A 90nm CMOS Low-Power 60 GHz Transceiver with Integrated Baseband
Circuitry," IEEE Journal of Solid-State Circuits, vol. 44, no. 12, pp. 3434 - 3447,
December 2009.
[82] L. Zhou, B. Geller, B. Zheng, A. Wei, and J. Cui, \System Scheduling for Multi-
Description Video Streaming Over Wireless Multi-Hop Networks," IEEE Transac-
tions on Broadcasting, vol. 55, no. 4, pp. 731 - 741, December 2009.
[83] T. Mitomo, N. Ono, H. Hoshino, Y. Yoshihara, O. Watanabe, and I. Seto, \A 77
GHz 90nm CMOS Transceiver for FMCW Radar Applications," IEEE Journal of
Solid State Circuits, vol. 45, no. 4, pp. 928 - 937, April 2010.
[84] E. Cohen, G. Jakobson, S. Ravid, and D. Ritter, \A Bidirectional TX/RX Four-
Element Phased Array at 60 GHz with RF-IF Conversion Block in 90-nm CMOS
Process," IEEE Transactions on Microwave Theory and Techniques, vol. 58, no. 5,
pp. 1438 - 1446, May 2010.
[85] M.-P. Kao and T.Q. Nguyen, \Rate-Distortion Optimized Bitstream Extractor for
Motion Scalability in Wavelet-Based Scalable Video Coding," IEEE Transactions on
Image Processing, vol. 19, no. 5, pp. 1214 - 1223, May 2010.
[86] T. Schierl, Y. S. de la Fuente, R. Globisch, C. Hellge, and T. Wiegand, \Priority-
based Media Delivery using SVC with RTP and HTTP Streaming," Multimedia Tools
and Applications, vol. 55, no. 2, pp. 227-246, 2011.
153
[87] C.-H. Yu, K. Doppler, C.B. Ribeiro, and O. Tirkkonen, \Resource Sharing Optimiza-
tion for Device-to-Device Communication Underlaying Cellular Networks," IEEE
Transactions on Wireless Communications, vol. 10, no. 8, pp. 2752 - 2763, August
2011.
[88] S. Singh, R. Mudumbai, and U. Madhow, \Interference Analysis for Highly Direc-
tional 60-GHz Mesh Networks: The Case for Rethinking Medium Access Control,"
IEEE/ACM Transactions on Networking, vol. 19, no. 5, pp. 1513 - 1527, October
2011.
[89] S. Wyne, K. Haneda, S. Ranvier, F. Tufvesson, and A.F. Molisch, \Beamforming Ef-
fects on Measured mm-Wave Channel Characteristics," IEEE Transactions on Wire-
less Communications, vol. 10, no. 11, pp. 3553 - 3559, November 2011.
[90] H. Min, J. Lee, S. Park, and D. Hong, \Capacity Enhancement Using an Interference
Limited Area for Device-to-Device Uplink Underlaying Cellular Networks," IEEE
Transactions on Wireless Communications, vol. 10, no. 12, pp. 3995 - 4000, December
2011.
[91] F. Baccelli, J. Li, T. Richardson, S. Shakkottai, S. Subramanian, and X. Wu, \On
Optimizing CSMA for Wide Area Ad-hoc Networks," Queueing Systems, vol. 72, pp.
31 - 68, 2012.
[92] S. Sharma, Y. Shi, Y.T. Hou, H.D. Sherali, S. Kompella, and S.F. Midki, \Joint
Flow Routing and Relay Node Assignment in Cooperative Multi-Hop Networks,"
IEEE Journal on Selected Areas in Communications, vol. 30, no. 2, pp. 254 - 262,
February 2012.
[93] Y. S. de la Fuente, T. Schierl, C. Hellge, T. Wiegand, D. Hong, D. De Vleeschauwer,
W. Van Leekwijck, and Y. Le Louedec, \Ecient HTTP-based streaming using Scal-
able Video Coding," Signal Processing: Image Communication, vol. 27, no. 4, pp.
329 - 342, April 2012.
[94] A. Abdel Khalek, C. Caramanis, and R.W. Heath, Jr., \A Cross-layer Design for Per-
ceptual Optimization of H.264/SVC with Unequal Error Protection," IEEE Journal
On Selected Areas in Communications, vol. 30, no. 7, pp. 1157 - 1171, August 2012.
[95] L. Toni, P.C. Cosman, and L.B. Milstein, \Channel Coding Optimization Based on
Slice Visibility for Transmission of Compressed Video over OFDM Channels," IEEE
Journal on Selected Areas in Communications, vol. 30, no. 7, pp. 1172 - 1183, August
2012.
[96] S. Singh, J.G. Andrews and G. de Veciana, \Interference Shaping for Improved Qual-
ity of Experience for Real-Time Video Streaming," IEEE Journal On Selected Areas
in Communications, vol 30., no. 7, pp. 1259 - 1269, August 2012.
154
[97] T.S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G.N. Wong, J.K.
Schulz, M. Samimi, and F. Gutierrez, \Millimeter Wave Mobile Communications for
5G Cellular: It Will Work!," IEEE Access, vol. 1, pp. 335 - 349, 2013.
[98] X. Wang, M. Chen, T.T. Kwon, L.T. Yang, and V.C.M. Leung, \AMES-Cloud: A
Framework of Adaptive Mobile Video Streaming and Ecient Social Video Sharing
in the Clouds," IEEE Transactions on Multimedia, Vol. 15, no. 14, pp. 811 - 820,
2013.
[99] N. Staelens, D. Deschrijver, E. Vladislavleva, B. Vermeulen, T. Dhaene, and P. De-
meester, \Constructing a No-Reference H.264/AVC Bitstream-Based Video Quality
Metric Using Genetic Programming-Based Symbolic Regression," IEEE Transactions
on Circuits and Systems for Video Technology, vol. 23, no. 8, pp. 1322 - 1333, 2013.
[100] T.C. Thang, H.T. Le, H.X. Nguyen, A.T. Pham, J.W. Kang, and Y.M. Ro, \Adap-
tive video streaming over HTTP with dynamic resource estimation," Journal of Com-
munications and Networks, vol. 15, no. 6, pp. 635 - 644, 2013.
[101] X. Wang, T.T. Kwon, Y. Choi, H. Wang, and J. Liu, \Cloud-assisted adaptive
video streaming and social-aware video prefetching for mobile users," IEEE Wireless
Communications, vol. 20, no. 3, pp. 72 - 79, 2013.
[102] B. Kaufman, J. Lilleberg, and B. Aazhang, \Spectrum Sharing Scheme Between
Cellular Users and Ad-hoc Device-to-Device Users," IEEE Transactions on Wireless
Communications, vol. 12, no. 3, pp. 1038 - 1049, March 2013.
[103] T.S. Rappaport, F. Gutierrez, E. Ben-Dor, J.N. Murdock, Qiao Yijun, and J.I.
Tamir, \Broadband Millimeter-Wave Propagation Measurements and Models Us-
ing Adaptive-Beam Antennas for Outdoor Urban Cellular Communications," IEEE
Transactions on Antennas and Propagation, vol. 61, no. 4, pp. 1850 - 1859, April
2013.
[104] B. Zhou, H. Hu, S.-Q. Huang, and H.-H. Chen, \Intracluster Device-to-Device Relay
Algorithm With Optimal Resource Utilization," IEEE Transactions on Vehicular
Technology, vol. 62, no. 5, pp. 2315 - 2326, June 2013.
[105] Y. Pei and Y.-C. Liang, \Resource Allocation for Device-to-Device Communica-
tions Overlaying Two-Way Cellular Networks," in IEEE Transactions on Wireless
Communications, vol. 12, no. 7, pp. 3611 - 3621, July 2013.
[106] P. Phunchongharn, E. Hossain, and D.I. Kim, \Resource Allocation for Device-
to-Device Communications underlaying LTE-Advanced Networks," IEEE Wireless
Communications, vol. 20, no. 4, pp. 91 - 100, August 2013.
155
[107] X. Wu, S. Tavildar, S. Shakkottai, T. Richardson, J. Li, R. Laroia, and A. Jovicic,
\FlashLinQ: A Synchronous Distributed Scheduler for Peer-to-Peer Ad Hoc Net-
works," IEEE/ACM Transactions on Networking, vol. 21, no. 4, pp. 1215 - 1228,
August 2013.
[108] J. Kim, Y. Tian, S. Mangold, and A.F. Molisch, \Joint Scalable Coding and Routing
for 60 GHz Real-Time Live HD Video Streaming Applications," IEEE Transactions
on Broadcasting, vol. 59, no. 3, pp. 500 - 512, September 2013.
[109] J. Li, J.B. Song, and Z. Han, \Network Connectivity Optimization for Device-to-
Device Wireless System With Femtocells," IEEE Transactions on Vehicular Tech-
nology, vol. 62, no. 7, pp. 3098 - 3109, September 2013.
[110] K. Shanmugam, N. Golrezaei, A.G. Dimakis, A.F. Molisch, and G. Caire, \Femto-
Caching: Wireless Content Delivery Through Distributed Caching Helpers," in IEEE
Transactions on Information Theory, vol. 59, no. 12, pp. 8402 - 8413, December 2013.
[111] L. de Cicco and S. Mascolo, \An Adaptive Video Streaming Control System: Mod-
eling, Validation, and Performance Evaluation," IEEE/ACM Transactions on Net-
working, vol. 22, no. 2, pp. 526 - 539, 2014.
[112] S. Tavakoli, J. Gutierrez, and N. Garcia, \Subjective Quality Study of Adaptive
Streaming of Monoscopic and Stereoscopic Video," IEEE Journal on Selected Areas
in Communications, vol. 32, no. 4, pp. 684 - 692, 2014.
[113] J. Kim and A.F. Molisch, \Fast Millimeter-Wave Beam Training with Receive Beam-
forming," IEEE/KICS Journal of Communications and Networks, (Accepted).
[114] D. Bethanabhotla, G. Caire, and M.J. Neely, \Joint Transmission Scheduling and
Congestion Control for Adaptive Video Streaming in Small-Cell Networks" IEEE
Transactions on Communications, (Under Revision). Available on ArXiv: http:
//arxiv.org/abs/1304.8083
[115] M. Ji, G. Caire, and A.F. Molisch, \Wireless Device-to-Device Caching Networks:
Basic Principles and System Performance," IEEE Journal on Selected Areas in Com-
munications (Under Revision). Available on ArXiv: http://arxiv.org/abs/
1305.5216
[116] H. Xu, V. Kukshya, and T.S. Rappaport, \Spatial and Temporal Characteriza-
tion of 60 GHz Indoor Channels," in Proceedings of the IEEE Vehicular Technology
Conference (VTC), Boston, Massachusetts, USA, September 2000.
[117] I. Ramani and S. Savage, \SyncScan: Practical Fast Hando for 802.11 Infrastruc-
ture Networks," in Proceedings of IEEE INFOCOM, Miami, Florida, USA, March
2005.
156
[118] M.-H. Lu, P. Steenkiste, and T. Chen, \Time-Aware Opportunistic Relay for Video
Streaming over WLANs," in Proceedings of IEEE International Conference on Mul-
timedia & Expo (ICME), Beijing, China, July 2007.
[119] S. Murthy, P. Hegde, V. Parameswaran, B. Li, and A. Sen, \Improved Path Se-
lection Algorithms for Multipath Video Streaming in Wireless Ad-Hoc Networks,"
in Proceedings of IEEE Military Communications Conference (MILCOM), Orlando,
Florida, USA, October 2007.
[120] S. Sanghavi, D. Shah, and A.S. Willsky, \Message Passing for Max-weight Indepen-
dent Set," in Proceedings of Neural Information Processing Systems (NIPS), Van-
couver, Canada, December 2007.
[121] Z. Lan, J. Wang, C.-S. Sum, T. Baykas, C. Pyo, F. Kojima, H. Harada, and S.
Kato, \Unequal Error Protection for Compressed Video Streaming on 60GHz WPAN
System," in Proceedings of IEEE International Wireless Communications and Mobile
Computing Conference (IWCMC), Crete Island, Greece, August 2008.
[122] B. Kaufman and B. Aazhang, \Cellular Networks with an Overlaid Device to De-
vice Network," in Proceedings of the Asilomar Conference on Signals, Systems and
Computers, Pacic Grove, California, USA, October 2008.
[123] X. Bo, K. Seada, and N. Venkatasubramanian, \An Experimental Study on Wi-Fi
Ad-Hoc Mode for Mobile Device-to-Device Video Delivery," in Proceedings of the
IEEE INFOCOM Workshops, Rio de Janeiro, Brazil, April 2009.
[124] C.-Y. Tsai and H.-M. Hang, \Rate-Distortion Model for Motion Prediction E-
ciency in Scalable Wavelet Video Coding," in Proceedings of IEEE Packet Video
Workshop, Seattle, Washington, USA, May 2009.
[125] K. Doppler, M.P. Rinne, P. Janis, C. Ribeiro, and K. Hugl, \Device-to-Device Com-
munications; Functional Prospects for LTE-Advanced Networks," in Proceedings of
the IEEE International Conference on Communications (ICC) Workshops, Dresden,
Germany, June 2009.
[126] C.-H. Yu, K. Doppler, C.B. Ribeiro, and O. Tirkkonen, \Performance Impact of
Fading Interference to Device-to-Device Communication underlaying Cellular Net-
works," in Proceedings of IEEE International Symposium on Personal, Indoor and
Mobile Radio Communications (PIMRC), Tokyo, Japan, September 2009.
[127] B. Kaufman, B. Aazhang, and J. Lilleberg, \Interference Aware Link Discovery for
Device to Device Communication," in Proceedings of the Asilomar Conference on
Signals, Systems and Computers, Pacic Grove, California, USA, November 2009.
[128] H. Su and X. Zhang, \Joint Link Scheduling and Routing for Directional-Antenna
Based 60 GHz Wireless Mesh Networks," in Proceedings of IEEE Global Communi-
cations Conference (GLOBECOM), Honolulu, Hawaii, USA, December 2009.
157
[129] O. Homann, R. Kays, and R. Reinhold, \Coded Performance of OFDM and SC
PHY of IEEE 802.15.3c for Dierent FEC Types," in Proceedings of IEEE Global
Communications Conference (GLOBECOM) Workshop, Honolulu, Hawaii, USA, De-
cember 2009.
[130] S. Hakola, T. Chen, J. Lehtomaki, and T. Koskela, \Device-to-Device (D2D) Com-
munication in Cellular Network-Performance Analysis of Optimum and Practical
Communication Mode Selection," in Proceedings of the IEEE Wireless Communica-
tions and Networking Conference (WCNC), Sydney, Australia, Arpil 2010.
[131] K. Doppler, Chia-Hao Yu, C.B. Ribeiro, and P. Janis, \Mode Selection for Device-
To-Device Communication Underlaying an LTE-Advanced Network," in Proceedings
of the Wireless Communications and Networking Conference (WCNC), Sydney, Aus-
tralia, Arpil 2010.
[132] X. Wu, S. Tavildar, S. Shakkottai, T. Richardson, J. Li, R. Laroia, and A. Jovi-
cic, \FlashLinQ: A Synchronous Distributed Scheduler for Peer-to-Peer Ad Hoc Net-
works," in Proceedings of Allerton Conference on Communication, Control, and Com-
puting, Monticello, IL, USA, October 2010.
[133] H.L. Cycon, V. George, G. Hege, D. Marpe, M. Palkow, T.C. Schmidt, and M.
Wahlisch, \Adaptive Temporal Scalability of H.264-compliant Video Conferencing in
Heterogeneous Mobile Environments," in Proceedings of IEEE Global Communica-
tions Conference (GLOBECOM), Miami, Florida, USA, December 2010.
[134] Y. S. de la Fuente, T. Schierl, C. Hellge, T. Wiegand, D. Hong, D. De Vleeschauwer,
W. Van Leekwijck, and Y. Le Loudec, \iDASH: Improved Dynamic Adaptive Stream-
ing over HTTP using Scalable Video Coding," in Proceedings of ACM Multimedia
Systems Conference (MMSys), San Jose, California, USA, February 2011.
[135] F. Baccelli, J. Li, T. Richardson, S. Shakkottai, S. Subramanian, and X. Wu, \On
Optimizing CSMA for Wide Area Ad-hoc Networks," in Proceedings of IEEE Inter-
national Symposium on Modeling and Optimization in Mobile, Ad Hoc and Wireless
Networks (WiOpt), Princeton, NJ, USA, May 2011.
[136] D. Giustiniano, V. Vukadinovic, and S. Mangold, \Wireless Networking for Auto-
mated Live Video Broadcasting: System Architecture and Research Challenges," in
Proceedings of IEEE International Symposium on a World of Wireless Mobile and
Multimedia Networks (WoWMoM), Lucca, Italy, June 2011.
[137] L. Toni, P.C. Cosman, and L.B. Milstein, \Unequal Error Protection based on Slice
Visibility for Transmission of Compressed Video over OFDM Channels," in Proceed-
ings of IEEE International Conference on Multimedia and Expo (ICME) Workshop
on Acoustics and Video Coding and Communication (AVCC), Barcelona, Spain, July
2011.
158
[138] C. Chen, R.W. Heath Jr., A. C. Bovik, and G. de Veciana, \Adaptive Policies for
Real-Time Video Transmission: a Markov Decision Process Framework," in Pro-
ceedings of IEEE International Conference on Image Processing (ICIP), Brussels,
Belgium, September 2011.
[139] J. Kim, Y. Tian, A.F. Molisch, and S. Mangold, \Joint Optimization of HD Video
Coding Rates and Unicast Flow Control for IEEE 802.11ad Relaying," in Proceedings
of IEEE International Symposium on Personal Indoor and Mobile Radio Communi-
cations (PIMRC), Toronto, Canada, September 2011.
[140] Z. Lu and G. de Veciana, \Opportunistic Transport for Stored Video Delivery over
Wireless Networks: Optimal Anticipative and Causal Approximations," in Proceed-
ings of Allerton Conference on Communication, Control and Computing, Allerton,
Illinois, USA, October 2011.
[141] W. Zhang, X. Zhou, L. Yang, Z. Zhang, B.Y. Zhao, and H. Zheng, \3D Beam-
forming for Wireless Data Centers," in Proceedings of ACM Hot Topics in Networks
(HotNets), Cambridge, Massachusetts, USA, November 2011.
[142] N. Golrezaei, A.G. Dimakis, A.F. Molisch, and G. Caire, \Wireless Video Content
Delivery through Distributed Caching and Peer-to-Peer Gossiping," in Proceedings
of the Asilomar Conference on Signals, Systems and Computers, Pacic Grove, Cal-
ifornia, USA, November 2011.
[143] D. Wang, P.C. Cosman, and L.B. Milstein, \Cross Layer Resource Allocation Design
for Uplink Video OFDMA Wireless Systems," in Proceedings of IEEE Global Com-
munications Conference (GLOBECOM), Houston, Texas, USA, December 2011.
[144] A. A. Khalek, C. Caramanis, and R. W. Heath Jr., \Joint Source-Channel Adap-
tation for Perceptually Optimized Scalable Video Transmission," in Proceedings of
IEEE Global Communications Conference (GLOBECOM), Houston, Texas, USA,
December 2011.
[145] T. Stockhammer, \Dynamic Adaptive Streaming over HTTP: Standards and Design
Principles," in Proceedings of ACM Conference on Multimedia Systems, 2011.
[146] S. Dey, \Cloud Mobile Media: Opportunities, Challenges, and Directions," in Pro-
ceedings of IEEE International Conference on Computing, Networking and Commu-
nications (ICNC), Maui, Hawaii, USA, Januaray 2012.
[147] V. Joseph and G. de Veciana, \Jointly Optimizing Multi-user Rate Adaptation
for Video Transport over Wireless Systems: Mean-Fairness-Variability Tradeos," in
Proceedings of IEEE INFOCOM, Orlando, Florida, USA, March 2012.
[148] N. Golrezai, K. Shanmugam, A. Dimakis, A. F. Molisch, G. Caire, \FemtoCaching:
Wireless Video Content Delivery through Distributed Caching Helpers," in Proceed-
ings of IEEE INFOCOM, Orlando, FL, USA, March 2012.
159
[149] H.E. Elkotby, K.M.F. Elsayed, and M.H. Ismail, \Exploiting Interference Alignment
for Sum Rate Enhancement in D2D-Enabled Cellular Networks," in Proceedings of
the IEEE Wireless Communications and Networking Conference (WCNC), Paris,
France, April 2012.
[150] H. Ahlehagh and S. Dey, \Video Caching in Radio Access Network: Impact on Delay
and Capacity," in Proceedings of IEEE Wireless Communications and Networking
Conference (WCNC), Paris, France, April 2012.
[151] T.S. Rappaport, E. Ben-Dor, J.N. Murdock, and Y. Qiao, \38 GHz and 60 GHz
Angle-Dependent Propagation for Cellular & Peer-to-Peer Wireless Communica-
tions," in Proceedings of IEEE International Conference on Communications (ICC),
Ottawa, Canada, June 2012.
[152] N. Golrezaei, K. Shanmugam, A.G. Dimakis, A.F. Molisch, and G. Caire, \Wireless
Video Content Delivery through Coded Distributed Caching," in Proceedings of the
IEEE International Conference on Communications (ICC), Ottawa, Canada, June
2012.
[153] C. Xu, L. Song, Z. Han, Q. Zhao, X. Wang, and B. Jiao, \Interference-Aware
Resource Allocation for Device-to-Device Communications as an underlay using Se-
quential Second Price Auction," in Proceedings of the IEEE International Conference
on Communications (ICC), Ottawa, Canada, June 2012.
[154] N. Golrezaei, A.F. Molisch, and A.G. Dimakis, \Base-Station Assisted Device-to-
Device Communications for High-Throughput Wireless Video Networks," in Proceed-
ings of the IEEE International Conference on Communications (ICC) Workshop on
Video-Aware Wireless Networks ViOpt), Ottawa, Canada, June 2012.
[155] S. Singh, O. Oyman, A. Papathanassiou, D. Chatterjee, and J.G. Andrews, \Video
Capacity and QoE Enhancements over LTE," in Proceedings of IEEE International
Conference on Communications (ICC) Workshop on Realizing Advanced Video Op-
timized Wireless Networks (ViOpt), Ottawa, Canada, June 2012.
[156] H. Ahlehagh and S. Dey, \Hierarchical Video Caching in Wireless Cloud: Ap-
proaches and Algorithms," in Proceedings of IEEE International Conference on Com-
munications (ICC) Workshop on Realizing Advanced Video Optimized Wireless Net-
works (ViOpt), Ottawa, Canada, June 2012.
[157] N. Golrezaei, A.G. Dimakis, and A.F. Molisch, \Wireless Device-to-Device Com-
munications with Distributed Caching," in Proceedings of the IEEE International
Symposium on Information Theory Proceedings (ISIT), Cambridge, Massachusetts,
USA, July 2012.
160
[158] K. Shanmugam and G. Caire, \Wireless Downloading Delay under Proportional Fair
Scheduling with Coupled Service and Requests: An Approximated Analysis," in Pro-
ceedings of the IEEE International Symposium on Information Theory Proceedings
(ISIT), Cambridge, Massachusetts, USA, July 2012.
[159] A. Abdel Khalek, C. Caramanis, and R.W. Heath, Jr., \Video-aware MIMO Precod-
ing with Packet Prioritization and Unequal Modulation," in Proceedings of European
Signal Processing Conference (EUSIPCO), Bucharest, Romania, August 2012.
[160] X. Zhou, Z. Zhang, Y. Zhu, Y. Li, S. Kumar, A. Vahdat, B.Y. Zhao, and H. Zheng,
\Mirror Mirror on the Ceiling: Flexible Wireless Links for Data Centers," in Pro-
ceedings of ACM Special Interest Group on Data Communication (SIGCOMM) Con-
ference, Helsinki, Finland, August 2012.
[161] X. Chen, L. Chen, M. Zeng, X. Zhang, and D. Yang, \Downlink Resource Allocation
for Device-to-Device Communication underlaying Cellular Networks," in Proceedings
of IEEE International Symposium on Personal, Indoor and Mobile Radio Commu-
nications (PIMRC), Sydney, Australia, September 2012.
[162] D. Bethanabhotla, G. Caire, and M.J. Neely, \Joint Transmission Scheduling and
Congestion Control for Adaptive Streaming in Wireless Device-to-Device Networks"
in Proceedings of Asilomar Conference on Signals, Systems, and Computers, Pacic
Grove, California, November 2012.
[163] L. Toni, P.C. Cosman, and L.B. Milstein, \Subcarrier Mapping based on Slice Vis-
ibility for Video Transmission over OFDM Channels," in Proceedings of Asilomar
Conference on Signals, Systems and Computers, Pacic Grove, California, USA,
November 2012.
[164] L.B. Le, \Fair Resource Allocation for Device-to-Device Communications in Wire-
less Cellular Networks," in Proceedings of the IEEE Global Communications Confer-
ence (GLOBECOM), Anaheim, California, USA, December 2012.
[165] N. Golrezaei, A.G. Dimakis, and A.F. Molisch, \Device-to-Device Collaboration
through Distributed Storage," in Proceedings of the IEEE Global Communications
Conference (GLOBECOM), Anaheim, California, USA, December 2012.
[166] L. Su, Y. Ji, P. Wang, and F. Liu, \Resource Allocation using Particle Swarm
Optimization for D2D Communication underlay of Cellular Networks," in Proceed-
ings of the IEEE Wireless Communications and Networking Conference (WCNC),
Shanghai, China, April 2013.
[167] R. Zhang, X. Cheng, L. Yang, and B. Jiao, \Interference-Aware Graph based
Resource Sharing for Device-to-Device Communications underlaying Cellular Net-
works," in Proceedings of the IEEE Wireless Communications and Networking Con-
ference (WCNC), Shanghai, China, April 2013.
161
[168] D.H. Lee, K.W. Choi, W.S. Jeon, and D.G. Jeong, \Resource Allocation Scheme for
Device-to-Device Communication for Maximizing Spatial Reuse," in Proceedings of
the IEEE Wireless Communications and Networking Conference (WCNC), Shanghai,
China, April 2013.
[169] J. Kim, Y. Tian, S. Mangold, and A.F. Molisch, \Quality-Aware Coding and Re-
laying for 60 GHz Real-Time Wireless Video Broadcasting," in Proceedings of IEEE
International Conference on Communications (ICC), Budapest, Hungary, June 2013.
[170] Y. Azar, G. Wong, K. Wang, R. Mayzus, J. Schulz, H. Zhao, F. Gutierrez, and T.S.
Rappaport, \28 GHz Propagation Measurements for Outdoor Cellular Communica-
tions Using Steerable Beam Antennas in New York City," in Proceedings of IEEE
International Conference on Communications (ICC), Budapest, Hungary, June 2013.
[171] S. Wen, X. Zhu, X. Zhang, and D. Yang, \QoS-Aware Mode Selection and Re-
source Allocation Scheme for Device-to-Device (D2D) Communication in Cellular
Networks," in Proceedings of the IEEE International Conference on Communications
Workshops (ICC), Budapest, Hungary, June 2013.
[172] J. Kim, F. Meng, P. Chen, H.E. Egilmez, D. Bethanabhotla, A.F. Molisch, M.J.
Neely, G. Caire, and A. Ortega, \Demo: Adaptive Video Streaming for Device-
to-Device Mobile Platforms," in Proceedings of ACM International Conference on
Mobile Computing and Networking (MobiCom), Miami, Florida, USA, September
2013.
[173] D. Burghal and A.F. Molisch, \Location Aware Training Scheme for D2D Net-
works," in Proceedings of the Asilomar Conference on Signals, Systems and Comput-
ers, Pacic Grove, California, USA, November 2013.
[174] J. Kim, A. Turci, G. Caire, and A.F. Molisch, \Joint Scheduling and Stochastic
Streaming for Device-to-Device Video Delivery," IEEE Information Theory and Ap-
plications Workshop (ITA), San Diego, California, USA, February 2014. (Graduation
Day Talk)
[175] S. Leyer, A. Sartenaer, and E. Wanufelle, \Branch-and-Rene for Mixed-Integer
Nonconvex Global Optimization," Technical Report, Mathematics and Computer Sci-
ence Division, Argonne National Laboratory, ANL/MCS-P1547-0908, 2008.
[176] A. Zambelli, \IIS Smooth Streaming Technical Overview," Microsoft Corporation,
vol. 3, March 2009.
[177] M. Grant and S. Boyd, CVX: Matlab Software for Disciplined Convex Programming,
Version 1.21, April 2011.
[178] Ericsson, \More than 50 Billion Connected Devices," Ericsson White Pa-
per, February 2011., http://www.ericsson.com/res/docs/whitepapers/
wp-50-billions.pdf
162
[179] ABI Research, \Small Cells: Outdoor Pico and Micro Markets," RR-METR-11,
2011.
[180] Small Cell Forum, \Backhaul Technologies for Small Cells Use Cases, Requirements
and Recommendations," Draft White Paper, October 2012.
[181] Cisco, \Cisco Visual Networking Index: Forecast and Methodology 2012{2017,"
Cisco White Paper, 2013.
[182] A. Sadri, \mmWave Technology Evolution: From WiGig to 5G Small Cells," IEEE
Interational Confernece on Communications (ICC) Tutorial, Budapest, Hungary,
June 2013.
163

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Cloud-enabled mobile sensing systems

PDF

Real-time channel sounder designs for millimeter-wave and ultra-wideband communications

PDF

mmWave dynamic channel measurements for localization and communications

PDF

Exploiting side information for link setup and maintenance in next generation wireless networks

PDF

Channel sounding for next-generation wireless communication systems

PDF

A protocol framework for attacker traceback in wireless multi-hop networks

PDF

Enabling virtual and augmented reality over dense wireless networks

PDF

Signal processing for channel sounding: parameter estimation and calibration

PDF

Architectures and algorithms of charge management and thermal control for energy storage systems and mobile devices

PDF

Fundamentals of two user-centric architectures for 5G: device-to-device communication and cache-aided interference management

PDF

Hybrid beamforming for massive MIMO

PDF

High-performance distributed computing techniques for wireless IoT and connected vehicle systems

PDF

Design, modeling, and analysis for cache-aided wireless device-to-device communications

PDF

Rate adaptation in networks of wireless sensors

PDF

Understanding and optimizing internet video delivery

PDF

Multichannel data collection for throughput maximization in wireless sensor networks

PDF

Robust routing and energy management in wireless sensor networks

PDF

Optimal distributed algorithms for scheduling and load balancing in wireless networks

PDF

Double-directional channel sounding for next generation wireless communications

PDF

Silicon-based wideband & mm-wave power amplifier architectures and implementations

Asset Metadata

Creator Kim, Joongheon (author)

Core Title Elements of next-generation wireless video systems: millimeter-wave and device-to-device algorithms

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Computer Science

Publication Date 06/10/2014

Defense Date 04/10/2014

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag 5G,device-to-device,millimeter-wave,OAI-PMH Harvest,wireless video systems

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Govindan, Ramesh (committee chair), Molisch, Andreas F. (committee chair), Nakano, Aiichiro (committee member), Ortega, Antonio K. (committee member)

Creator Email joonghek@usc.edu,joongheon@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-418628

Unique identifier UC11296187

Identifier etd-KimJoonghe-2544.pdf (filename),usctheses-c3-418628 (legacy record id)

Legacy Identifier etd-KimJoonghe-2544.pdf

Dmrecord 418628

Document Type Dissertation

Format application/pdf (imt)

Rights Kim, Joongheon

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA