Cloud Radio Access Network Architecture Towards 5G Network
Cloud Radio Access Network Architecture Towards 5G Network
Cloud Radio Access Network Architecture Towards 5G Network
Checko, Aleksandra; Berger, Michael Stbert; Kardaras, Georgios; Dittmann, Lars; Christiansen, Henrik
Lehrmann
Publication date:
2016
Document Version
Final published version
Link to publication
Citation (APA):
Checko, A., Berger, M. S., Kardaras, G., Dittmann, L., & Christiansen, H. L. (2016). Cloud Radio Access
Network architecture. Towards 5G mobile networks. Technical University of Denmark.
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
You may not further distribute the material or use it for any profit-making activity or commercial gain
You may freely distribute the URL identifying the publication in the public portal
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
PhD Thesis
Aleksandra Checko
Abstract v
Resum vii
Preface ix
Acknowledgements xi
Acronyms xxi
1 Introduction 1
1.1 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 A Note on contributions . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Publications prepared during the course of the PhD project . . . . 4
2 C-RAN overview 7
2.1 What is C-RAN? Base Station architecture evolution . . . . . . . . 9
2.2 Advantages of C-RAN . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Challenges of C-RAN . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Transport network techniques . . . . . . . . . . . . . . . . . . . . . 23
2.5 RRH development . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.6 Synchronized BBU Implementation . . . . . . . . . . . . . . . . . . 33
2.7 Virtualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.8 Likely deployment Scenarios . . . . . . . . . . . . . . . . . . . . . . 36
2.9 Ongoing work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.10 Future directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
iii
iv Contents
Appendices 119
Bibliography 127
Abstract
Cloud Radio Access Network (C-RAN) is a novel mobile network architecture which
can address a number of challenges that mobile operators face while trying to support
ever-growing end-users needs towards 5th generation of mobile networks (5G). The main
idea behind C-RAN is to split the base stations into radio and baseband parts, and pool
the Baseband Units (BBUs) from multiple base stations into a centralized and virtualized
BBU Pool. This gives a number of benefits in terms of cost and capacity. However, the
challenge is then to find an optimal functionality splitting point as well as to design the so-
called fronthaul network, interconnecting those parts. This thesis focuses on quantifying
those benefits and proposing a flexible and capacity-optimized fronthaul network.
It is shown that a C-RAN with a functional split resulting in a variable bit rate on the
fronthaul links brings cost savings due to the multiplexing gains in the BBU pool and
the fronthaul network. The cost of a fronthaul network deployment and operation can be
further reduced by sharing infrastructure between fronthaul and other services.
The origins of multiplexing gains in terms of traffic burstiness, the tidal effect and
various possible functional splits are analyzed and quantified. Sharing baseband resources
between many cells is possible for traditional C-RANs. However, in order to further
benefit from multiplexing gains on fronthaul, it is recommended to implement a functional
split yielding variable bit rate in the fronthaul. For the analyzed data sets, in deployments
where diverse traffic types are mixed (bursty, e.g., web browsing and constant bit rate, e.g.,
video streaming) and cells from various geographical areas (e.g., office and residential) are
connected to the BBU pool, the multiplexing gain value reaches six. Using packet-based
fronthaul has the potential to utilize fronthaul resources efficiently. However, meeting
synchronization and delay requirements is a challenge. As a possible solution, the use
of IEEE Precision Time Protocol (PTP) (also known as 1588v2) has been evaluated,
and for the analyzed scenario it can assure synchronization on the nanosecond level,
fulfilling mobile network requirements. Furthermore, mechanisms to lower delay and
jitter have been identified, namely: source scheduling and preemption. An innovative
source scheduling scheme which can minimize jitter has been proposed. The scheme is
optimized for symmetric downlink and uplink traffic, but can also be used when downlink
traffic exceeds uplink. Moreover, a demonstrator of a Software Defined Networking
(SDN) controlled Ethernet fronthaul has been built.
v
vi Abstract
Resum
(Summary in Danish)
vii
viii Resum
men kan ogs finde anvendelse nr downlink overskrider uplink. Slutteligt er der opbygget
en demonstrator af et Software Defined Network (SDN) styret, Ethernet baseret fronthaul.
Preface
This dissertation presents a selection of the research work conducted during my PhD study
from January 1, 2013 until February 15, 2016 under supervision of Associate Professor
Michael Stbert Berger, Associate Professor Henrik Lehrmann Christiansen, Professor
Lars Dittmann, Dr Georgios Kardaras (January 2013-December 2013), Dr Bjarne Skak
Bossen (January 2014-May 2015), and Dr Morten Hgdal (June 2015-February 2016). It
is submitted to the Department of Photonics Engineering at the Technical University of
Denmark in a partial fulfillment of the requirements for the Doctor of Philosophy (PhD)
degree.
This Industrial PhD project was done in the Networks Technologies and Service
Platforms group at the Department of Photonics Engineering at the Technical University
of Denmark (DTU), Kgs. Lyngby, Denmark and in MTI Radiocomp, Hillerd, Denmark,
where I was employed. The work was co-financed by the Ministry of Higher Education
and Science of Denmark within the framework of Industrial PhD Program. The work
benefited from collaboration within the EU HARP project, especially from a six-month
external stay at Alcatel-Lucent Bell Labs France under the guidance of Mr. Laurent
Roullet.
Aleksandra Checko
Kgs. Lyngby, February 2016
ix
x
Acknowledgements
Being a PhD student is not an easy job. Not only does one need to define a topic that
is challenging, but also advance the state-of-the-art within it. Whatever one finds, he or
she is welcome to re-search. On top of that, various formal aspects of the PhD project
need to be taken care of. Fortunately, I was not alone in this battle. I would like to take
the opportunity to thank some of the people that made the accomplishment of this thesis
possible.
First and foremost, I would like to thank my supervisors: Michael Berger, Bjarne
Bossen, Henrik Christiansen, Lars Dittmann, Georgios Kardaras, and Morten Hgdal for
their mentoring and support. I am particularly grateful to Michael Berger and to MTI
Radiocomp for giving me this opportunity. Special thanks to Henrik Christiansen for
making things as simple as possible, but not simpler. Moreover, I am grateful to Lara
Scolari and Thomas Nrgaard for their help in starting this project, and to Laurent Roullet,
Thomas Nrgaard, and Christian Lanzani for asking interesting questions and helping to
shape the directions this thesis took.
I would like to acknowledge the members of the Networks group at DTU Fotonik
and colleagues from MTI for all the fruitful discussions, and for maintaining a friendly
atmosphere at work. Lunch and seminar discussions were one-of-a-kind.
Merci beaucoup Laurent Roullet for welcoming me in his group at Alcatel Lucent
Bell Labs for a six-month external stay, and to Aravinthan Gopalasingham for our fruitful
collaboration. I am also grateful to the colleagues from Bell Labs for a warm welcome,
guidance and making my time in Paris more enjoyable, especially for our trips to Machu
Picchu.
I would like to show my gratitude to Henrik Christiansen, Michael Berger, Morten
Hgdal, Andrea Marcano, Matteo Artuso, Jakub Jelonek, Magorzata Checko and Geor-
gios Kardaras for reviewing this thesis. Mange tak til Henrik Christiansen for translating
the abstract.
This project would be a different experience if it was not for my cooperation with
members of the HARP project, the Ijoin project, and the IEEE Student Branch at DTU.
I am thankful for this multicultural experience.
And last but not least, I am grateful to my family and friends for their continuous
support. A most special thanks to Jakub Jelonek who stood by me in all joyful but also
xi
xii Acknowledgements
tough moments.
I had a great opportunity to work with so many bright people, from whom I have
learned so much. Dziekuje!
List of Figures
xiii
xiv List of Figures
4.11 Maximum phase error observed during stable operation for various scenar-
ios with offset averaging applied. . . . . . . . . . . . . . . . . . . . . . 97
4.12 Maximum frequency error observed during stable operation for various
scenarios with drift averaging applied. . . . . . . . . . . . . . . . . . . 98
4.13 Clock recovery scheme inside an RRH combined with CPRI2Eth gateway 99
4.14 Delays in Ethernet-based fronthaul . . . . . . . . . . . . . . . . . . . . 100
4.15 RRH-BBU distance assuming no queuing . . . . . . . . . . . . . . . . . 101
4.16 RRH-BBU distance for various queuing . . . . . . . . . . . . . . . . . . 101
4.17 Protected window, here for the fronthaul traffic . . . . . . . . . . . . . 102
4.18 Source scheduling used to reduce jitter. Here an example for UL . . . 103
4.19 Preemption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.20 Source scheduling algorithm . . . . . . . . . . . . . . . . . . . . . . . . 106
4.21 Ethernet L1 and L2 as well as 1904.3 overhead comparing to Ethernet
frame payload size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.22 Demonstrator of Ethernet fronthaul network . . . . . . . . . . . . . . . 108
4.23 Ethernet-based C-RAN fronthaul - laboratory setup . . . . . . . . . . . 109
4.24 Ping RTT over 1 - 3 switches . . . . . . . . . . . . . . . . . . . . . . . 110
4.25 Fronthaul, backhaul and midhaul . . . . . . . . . . . . . . . . . . . . . 112
4.26 Proposed architecture for Fronthaul over Ethernet . . . . . . . . . . . . 114
A.1 C-RAN architecture where OTN is used to transport fronthaul streams 121
A.2 Reference setup for CPRI over OTN testing . . . . . . . . . . . . . . . 122
A.3 CPRI over OTN mapping measurement setup . . . . . . . . . . . . . . 122
A.4 Detailed measurement setup [86] . . . . . . . . . . . . . . . . . . . . . 122
A.5 Results 64 QAM with OBSAI Using TPO125 Device . . . . . . . . . . 126
xvi
List of Tables
2.1 Comparison between traditional base station, base station with RRH and
C-RAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 IQ bit rates between a cell site and centralized BBU Pool . . . . . . . 21
2.3 Requirements for cloud computing and C-RAN applications [54] . . . . 23
2.4 Comparison of IQ compression methods. Compression ratio 33% corre-
sponds to 3:1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 DSP and GPP processors . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.6 Research directions for C-RAN . . . . . . . . . . . . . . . . . . . . . . 42
2.7 Requirements for different functional splits [148] for the LTE protocol
stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1 Assumed pooling gains on different layers of the LTE protocol stack. . 54
3.2 Estimations of baseband complexity in GOPS of cell- and user-processing
for UL and DL and different cell sizes. Numbers are taken from [155]. 56
3.3 Multiplexing gains (MG) looking at traffic-dependent resources. . . . . 64
3.4 Traffic generation parameters for network modeling; C - Constant, E -
Exponential, L - log-normal, U - uniform, UI - uniform integer . . . . . 66
3.5 BBU save for various office/residential cell mixes, measured using different
methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6 Traffic generation parameters for network modeling; C - Constant, E -
Exponential, L - log-normal, G - gamma, U - uniform . . . . . . . . . . 69
3.7 Simulation parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.8 Multiplexing gains calculated in different projects. MG - multiplexing gain. 76
xvii
xviii List of Tables
ACK acknowledgment
AM Acknowledged Mode
BB Baseband
CA Carrier Aggregation
CP cyclic-prefix
xxi
xxii Acronyms
CPRI2Eth CPRI-to-Ethernet
DL downlink
IP Internet Protocol
IQ In-phase/Quadrature
LTE-A LTE-Advanced
MG Multiplexing gain
NACK non-acknowledgement
PA Power Amplifier
PG Pooling gain
RANaaS RAN-as-a-Service
RF Radio Frequency
RLC Radio Link Control
RNC Radio Network Controller
ROHC RObust Header Compression
RRH Remote Radio Head
RRM Radio Resource Management
RTT Round Trip Time
UE User Equipment
UL uplink
xxvi Acronyms
UM Unacknowledged Mode
UMTS Universal Mobile Telecommunications System
More and more of our communication becomes mobile. In 1991, the first digital
mobile call was made through the Global System for Mobile Communications (GSM)
(2G) network by a Finnish prime minister. By 2001, the number of network subscribers
exceeded 500 million [1]. That same year, the first 3G network, the Universal Mobile
Telecommunications System (UMTS), was introduced to increase the speed of data
transmission. The mobile Internet connectivity has gained a wide spread popularity with
1
2 Chapter 1. Introduction
60
50
File sharing
40
Audio
EB/month
0 Video
2010 2015 2020
Year
Figure 1.2: Traffic growth in mobile networks. Source: Ericsson, November 2015.
Long Term Evolution (LTE) networks, which have been commercially deployed since
2011. By the end of 2012, the number of LTE subscribers had exceeded 60 million,
while the total number of mobile devices amounted to 7 billion [1], exceeding the worlds
population. The number of subscribers is forecast to grow further, reaching 9 billion
in 2021 [2], especially with the increasing popularity of Machine to Machine (M2M)
communication.
By 2021, a new generation of mobile networks will be standardized [3] - 5G - to
satisfy ever-growing traffic demand, offering increased speed and shorter delays, the latter
enabling a tactile Internet. It will be also able to offer ultra-high reliability, and connect
a vast number of various devices, such as cars and sensors, to the Internet, forming an
Internet of Things (IoT) [4]. The actual applications of the future mobile networks are
probably beyond our current imagination.
Historical data as well as traffic forecast shows that this increase in number of sub-
scribers is accompanied by an exponential growth in traffic, with video occupying the
most bandwidth [5], [2], as shown on Figure 1.2. In order to support such a growth more
cells or higher capacity per cell needs to be provided, which results in increased cost. At
the same time, users would like to pay less and use more data. Therefore the increase
of cost cannot follow the same trend line as traffic growth. The same applies to power
consumption both in terms of cost as well as respect towards the environment. Disruptive,
affordable solutions are needed to deliver more capacity, shorter delays, and improved
reliability at the same time not increasing the power consumption.
In order to address the capacity requirements more cells can be deployed, or higher
capacity needs to be provided per each cell. Small cells can be deployed in places with
high user activity. Self-Organizing Networks (SON) techniques are important to ease man-
agement and optimization of networks with many cells. In dense deployments interference
is a challenge, therefore techniques like enhanced Inter-cell Interference Coordination
(eICIC) and Coordinated Multi-Point (CoMP) are essential. To increase the capacity of
each cell more antennas can be used using a Massive Multiple Input Multiple Output
(MIMO). As spectrum is scarce in currently explored bands, higher frequency bands,
1.1. Thesis Outline 3
C-RAN is a novel RAN architecture that can address a number of challenges that the
operators face while trying to support the growing needs of end-users, therefore it is seen
as a major technological foundation for 5G networks [[4], [6], [7], [8], [9]]. Chapter 2
provides an extensive introduction to C-RAN, outlining its fundamental aspects, advan-
tages, and the technical solutions that can address the challenges it faces. In addition,
future research directions are outlined.
However, for the deployments to be widespread, they need to be economically feasible.
Chapter 3 evaluates energy and cost savings coming from the statistical multiplexing gain
in C-RAN. Three studies are covered, the first one not including protocol processing, the
second exploring a tidal effect, and the third looking at different method of measuring
multiplexing gains.
One of the major challenges to address to deploy C-RAN is a flexible fronthaul with
optimized capacity requirements. Chapter 4 puts special focus on fronthaul networks,
outlining their requirements, evaluating more traditional, circuit-switched transport so-
lution, as well as innovative packet-based transport techniques. Emphasis is placed on
Ethernet, packet-based fronthaul, exploring options to ensure clock delivery and to meet
delay requirements.
Finally, Chapter 5 concludes the dissertation and explores future research directions.
4 Chapter 1. Introduction
1.3.1 Journals
1. A. Checko1st , A.P. Avramova1st , M.S. Berger and H.L. Christiansen, Evaluating C-
RAN fronthaul functional splits in terms of network level energy and cost savings,
1.3. Publications prepared during the course of the PhD project 5
1.3.2 Conferences
1. A. Checko, A. Juul, H. Christiansen and M.S. Berger, "Synchronization Challenges
in Packet-based Cloud-RAN Fronthaul for Mobile Networks", In proceedings
of IEEE ICC 2015 - Workshop on Cloud-Processing in Heterogeneous Mobile
Communication Networks (IWCPM)
7. A. Checko, H.L. Christiansen and M.S. Berger, "Evaluation of energy and cost sav-
ings in mobile Cloud Radio Access Networks", In proceedings of OPNETWORK
2013 conference, Washington D.C., USA, August 26-30 2013
6 Chapter 1. Introduction
1.3.4 Others
1. A. Checko, G. Kardaras, C. Lanzani, D. Temple, C. Mathiasen, L. Pedersen,
B. Klaps, OTN Transport of Baseband Radio Serial Protocols in C-RAN Architec-
ture for Mobile Network Applications. Industry white paper, March 2014
2. Contribution to the deliverable of the EU project HARP D6.4 Eth2CPRI prototype
implementation, presented to European Commission in November 2015
3. Contribution to the deliverable of the EU project HARP D6.3 Protocol Extensions
Design and Implementation, presented to European Commission in August 2015
4. Coordination and contribution to Deliverable 6.2 Aggregation network optimiza-
tion of the EU project HARP, presented to European Commission in January
2015
5. Contribution to Y2 report of the EU project HARP, presented to European Com-
mission in November 2014
6. Contribution to Y1 report of the EU project HARP, presented to European Com-
mission in November 2013
7. Contribution to the deliverable of the EU project HARP D6.1 Aggregation Network
Definition, presented to European Commission in May 2014
8. Contribution to the deliverable of the EU project HARP D3.1 Requirements, metrics
and network definition, presented to European Commission in July 2013
9. Contribution to the deliverable of the EU project HARP D1.5 Final plan for the
use and dissemination of Foreground, presented to European Commission partners
in May 2014
CHAPTER 2
C-RAN overview
Base Transceiver Station (BTS), NodeB, eNodeB. Those are the names used to describe a
base station in GSM, UMTS and LTE standards, respectively. As a concept, and logical
node, a base station is responsible for receiving signal from/sending to user, prepare it to
be send up to/received from the core network and organize transmission. Physically, this
node can be deployed as a standalone base station, base station with RRH or Cloud RAN
(C-RAN).
As spectral efficiency for the LTE standard is approaching the Shannon limit, the
most prominent way to increase network capacity is by either adding more cells, creating
a complex structure of Heterogeneous and Small cell Networks (HetSNets) [16] or by
implementing techniques such as multiuser MIMO [17] as well as Massive MIMO [18],
where numerous antennas simultaneously serve a number of users in the same time-
frequency resource. However, this results in growing inter-cell interference levels and
high costs.
The Total Cost of Ownership (TCO) in mobile networks includes CAPital EXpenditure
(CAPEX) and OPerating EXpenditure (OPEX). CAPEX mainly refers to expenditure
relevant to network construction which may span from network planning to site acquisition,
RF hardware, baseband hardware, software licenses, leased line connections, installation,
civil cost and site support, like power and cooling. OPEX covers the cost needed to
operate the network, i.e. site rental, leased line, electricity, operation and maintenance as
well as upgrade [19]. CAPEX and OPEX are increasing significantly when more base
stations are deployed. More specifically, CAPEX increases as base stations are the most
expensive components of a wireless network infrastructure, while OPEX increases as,
among others, cell sites demand a considerable amount of power to operate, e.g., China
Mobile estimates 72% of total power consumption originates from the cell sites [20].
Mobile network operators need to cover the expenses for network construction, operation,
maintenance and upgrade; meanwhile, the Average Revenue Per User (ARPU) stays flat
or even decreases over time, as the typical user is more and more data-hungry but expects
to pay less for data usage.
C-RAN is a novel mobile network architecture, which has the potential to answer
the above mentioned challenges. The concept was first proposed in [21] and described
in detail in [20]. In C-RAN, baseband processing is centralized and shared among sites
in a virtualized BBU Pool. This means that it is well prepared to adapt to non-uniform
traffic and utilizes resources, i.e., base stations, more efficiently. Due to that fact that
fewer BBUs are needed in C-RAN compared to the traditional architecture, C-RAN also
7
8 Chapter 2. C-RAN overview
has the potential to decrease the cost of network operation, because power and energy
consumption are reduced compared to the traditional RAN architecture. New BBUs
can be added and upgraded easily, thereby improving scalability and easing network
maintenance. A virtualized BBU Pool can be shared by different network operators,
allowing them to rent Radio Access Network (RAN) as a cloud service. As BBUs from
many sites are co-located in one pool, they can interact with lower delays therefore
mechanisms introduced for LTE-Advanced (LTE-A) to increase spectral efficiency and
throughput, such as eICIC and CoMP are greatly facilitated. Methods for implementing
load balancing between cells are also facilitated. Furthermore, network performance is
improved, e.g., by reducing delay during intra-BBU Pool handover.
C-RAN architecture is targeted by mobile network operators, as envisioned by China
Mobile Research Institute [20], IBM [21], Alcatel-Lucent [22], Huawei [23], ZTE [24],
Nokia Siemens Networks [19], Intel [25] and Texas Instruments [26]. Moreover, C-RAN
is seen as a typical realization of mobile network supporting soft and green technologies
in fifth generation (5G) mobile networks [27]. However, C-RAN is not the only candidate
architecture that can answer the challenges faced by mobile network operators. Other
solutions include small cells, being part of HetSNets and Massive MIMO. Small cells
deployments are the main competitors for outdoor hot spot as well as indoor coverage
scenarios. All-in-one small footprint solutions like Alcatel-Lucents LightRadio [28] can
host all base station functionalities in a several-liter box. They can be placed outdoors
reducing cost of operation associated to cooling and cell site rental. However, they will be
underutilized during low-activity periods and cannot employ collaborative functionalities
as well as C-RAN. Moreover, they are more difficult to upgrade and repair than C-RAN.
A brief comparison between C-RAN, Massive MIMO and HetSNets is outlined in [16].
Liu et al. in [29] prove that energy efficiency of large scale Small Cell Networks is higher
compared with Massive MIMO. Furthermore, cost evaluation on different options needs
to be performed in order for a mobile network operator to choose an optimal solution.
Comparison of TCO including CAPEX and OPEX over 8 years of a traditional LTE
macro base station, LTE C-RAN and LTE small cell shows that the total transport cost
per Mbps is highest for macro cell deployments - 2200$, medium for C-RAN - 1800$
and 3 times smaller for small cell - 600$ [30]. Therefore the author concludes that
C-RAN needs to achieve significant benefits to overcome such a high transportation cost.
Collaborative techniques such as CoMP and eICIC can be implemented in small cells
giving higher benefits in HetNet configuration instead of C-RAN. The author envisions
that C-RAN might be considered for special cases like stadium coverage. However,
C-RAN is attractive for operators that have free/cheap fiber resources available.
This chapter surveys the state-of-the-art literature published on C-RAN and its im-
plementation until 2014. Such input helps mobile network operators to make an optimal
choice on deployment strategies. The chapter is organized as follows. In Section 2.1 the
fundamental aspects of C-RAN architecture are introduced. Moreover, in Section 2.2
the advantages of this architecture are discussed in detail along with the challenges that
need to be overcome before fully exploiting its benefits in Section 2.3. In Section 2.4 a
2.1. What is C-RAN? Base Station architecture evolution 9
RRH 2
RRH 2
...
RRH 1 ...
RRH n RRH 1
RRH n
BBU 2
Aggregated
BBU 1 BBU n Traffic (h)
Cloud
24 h BBU Pool
Mobile Mobile
Backhaul Network Backhaul Network
Figure 2.1: Statistical multiplexing gain in C-RAN architecture for mobile networks.
The area which a mobile network covers is divided into cells, therefore mobile
networks are often called cellular networks. Traditionally, in cellular networks, users
communicate with a base station that serves the cell under coverage of which they are
10 Chapter 2. C-RAN overview
located. The main functions of a base station can be divided into baseband processing and
radio functionalities. The main sub-functions of baseband processing module are shown
in left side of Figure 2.2. Among those one can find coding, modulation, Fast Fourier
Transform (FFT), etc. The radio module is responsible for digital processing, frequency
filtering and power amplification.
L3 L2 L1
S D CFR
IQ DL R U DAC
De-/Quantization
CPRI/OBSAI/ORI
De-/Modulation
Mapping-MIMO
Transport-MAC
Resource-block
De-/Sampling
CPRI/ Frequ
Channel de-/
Control-RRC
Mapping C C DPD
IFFT/FFT
Antenna
OBSAI ency
coding
CoMP
eICIC
BBU RRH
RRC Radio Resource Control SRC Sampling Rate Conversion DAC Digital-to-Analog Converter
MAC Media Access Control DUC/DDC Digital Up/Downconversion ADC Analog-to-Digital Converter
FFT Fast Fourier Transform CFR Crest Factor Reduction Power Amplifier
DPD Digital Predistortion
Figure 2.2: Base station functionalities. Exemplary baseband processing functionalities inside
BBU are presented for LTE implementation. Connection to RF part and sub modules of
RRH are shown.
Antenna
BaseBand
Transport
Control
Synch
PA
RF
BS
cell
RF RRH
BaseBand
Transport
Control
Synch
S1/X2 Ir
BBU
cell
RF RRH
RF RRH
BaseBand
Transport
BaseBand
Transport
Control
BaseBand
Transport
Control
Synch
BaseBand
Transport
Control
Synch
Control
Synch
Synch
S1/X2 X2 RF RRH
Ir
Virtual BBU Pool
where a BBU needs to be placed close to the antenna. RRHs can be placed up on poles
or rooftops, leveraging efficient cooling and saving on air-conditioning in BBU housing.
RRHs are statically assigned to BBUs similarly to the traditional RAN. One BBU can
serve many RRHs. RRHs can be connected to each other in a so called daisy-chained
architecture. An Ir interface is defined, which connects RRH and BBU.
Common Public Radio Interface (CPRI) [33] is the radio interface protocol widely
used for In-phase/Quadrature (IQ) data transmission between RRHs and BBUs - on Ir
interface. It is a constant bit rate, bidirectional protocol that requires accurate synchro-
nization and strict latency control. Other protocols that can be used are Open Base Station
Architecture Initiative (OBSAI) [34] and Open Radio equipment Interface (ORI) [35],
[36]. For LTE base stations the X2 interface is defined between base stations, the S1
interface connects a base station with mobile core network.
Table 2.1: Comparison between traditional base station, base station with RRH and C-RAN
RRH X2
RRH MME
Aggregation
network
EPC
S1 PGW
BBU
Base
RRH Base
Base
Base
pool
band
band
band
band SGW
40
35
30
25
Load
20
15
10
0
0 6 12 18 24
Time (h)
Figure 2.5: Daily load on base stations varies depending on base station location. Data
source: [20].
aggregated into one single cloud base station. Aggregation of 57 sectors in a single BBU
Pool saves more than 25% of the compute resources. Moreover, the user distribution has
a strong influence on the utilization of the compute resources. The results of last three
works converge giving around 25% of potential savings on baseband resources. In [42]
Bhaumik et al. show that the centralized architecture can potentially result in savings of at
least 22% in compute resources by exploiting the variations in the processing load across
base stations. In [40] Namba et al. analyze statistical multiplexing gain as a function
of cell layout. The analysis shows that for the metropolitan area, the number of BBUs
can be reduced by 75% compared to the traditional base station. In [43] Madhavan et
al. quantify the multiplexing gain of consolidating WiMAX base stations in different
traffic conditions. The gain increases linearly with network size and it is higher when
base stations are experiencing higher traffic intensity. On the contrary, in [44] Liu et
al. analyzed that lighter load can increase the statistical multiplexing gain in virtual
base station pool. Moreover, multiplexing gain reaches significant level even for the
medium-size pools and the increase in gain for larger pools in negligible.
In authors previous work [11] initial evaluation of statistical multiplexing gain of
BBUs in C-RAN was presented. The paper concludes that 4 times less BBUs are needed
for user data processing in a C-RAN compared to a traditional RAN for specific traffic
patterns, making assumptions of the number of base stations serving different types
of areas. The model does not include mobile standard protocols processing. After
including protocol processing in [12] the statistical multiplexing gain varied between 1.2
and 1.6 depending on traffic mix, reaching the peak for 30% of office and thereby 70%
of residential base stations, thereby enabling saving of 17% - 38%. Those results are
presented in Sections 3.2 and 3.3, respectively. Secondly, results obtained via simulations
[12] have been compared to the ones achieved with teletraffic theory [45].
All those works referred to the traditional - Baseband (BB)/RF - functional split of
16 Chapter 2. C-RAN overview
C-RAN. In [46] authors estimate what they define as statistical multiplexing convergence
ratio on fronthaul links by averaging observed daily throughput. Calculated ratio equals
to three. However, the analysis took only average network load into account and therefore
can be interpreted mostly as an average daily fronthaul utilization. In authors most recent
work [13] different functional splits (as presented in Section 2.10.1) and different, precisely
defined application mixes are investigated. A numerical evaluation was given supporting
the intuitive conclusions that the maximum multiplexing gain on BBU resources can be
achieved for a fully centralized C-RAN. The more functionality is moved from the BBU
pool to the cell site, the lower the multiplexing gain on the BBU pool. However, when
traffic starts to be variable bit rate, a multiplexing gain on fronthaul links can be achieved,
lowering the required capacity. Those results are presented in Section 3.4.
Statistical multiplexing gain can be maximized by employing a flexible, reconfigurable
mapping between RRH and BBU adjusting to different traffic profiles [47]. Statistical
multiplexing gain depends on the traffic, therefore it can be maximized by connecting
RRHs with particular traffic profiles to different BBU Pools [12]. More on multiplexing
gain evaluation can be found in Chapter 3.
Coverage upgrades simply require the connection of new RRHs to the already existing
BBU Pool. To enhance network capacity, existing cells can then be split, or additional
RRHs can be added to the BBU Pool, which increases network flexibility. Deployment of
new cells is in general more easily accepted by local communities, as only a small device
needs to be installed on site (RRH) and not a bulky base station. If the overall network
capacity shall be increased, this can be easily achieved by upgrading the BBU Pool, either
by adding more hardware or exchanging existing BBUs with more powerful ones.
As BBUs from a large area will be co-located in the same BBU Pool, load balancing
features can be enabled with advanced algorithms on both the BBU side and the cells
side. On the BBU side, BBUs already form one entity, therefore load balancing is a
matter of assigning proper BBU resources within a pool. On the cells side, users can be
switched between cells without constraints if the BBU Pool has capacity to support them,
as capacity can be assigned dynamically from the pool.
takes 46% of cell site power consumption [20]. Due to the usage of RRHs air conditioning
of radio module can be decreased as RRHs are naturally cooled by air hanging on masts
or building walls, as depicted in Figure 2.3. ZTE estimates that C-RAN enables 67%-80%
power savings compared with traditional RAN architecture, depending on how many cells
one BBU Pool covers [24], which stays in line with China Mobile research claiming 71%
power savings [49].
Civil work on remote sites can be reduced by gathering equipment in a central room,
what contributes to additional OPEX savings.
In total, 15% CAPEX and 50% OPEX savings are envisioned comparing to RAN with
RRH [49] or traditional RAN architecture [50]. However, the cost of leasing the fiber
connection to the site may increase CAPEX. IQ signal transported between RRHs and
BBUs brings up a significant overhead. Consequently, the installation and operation of
transport network causes considerable costs for operators.
Moreover, virtualization helps to reduce cost of network deployment and operation, at
the same time enabling operators to offer additional services, not only serve as pipelines
for carrying user date.
and multiple-user joint scheduling has been analyzed for a non-ideal channel with carrier
frequency offset [57]. When carrier frequency offset does not exceed 3 5ppb, C-RAN
can achieve remarkable performance gain on both capacity and coverage even in non-ideal
channel, i.e., 20%/52% for cell average/cell edge.
With the introduction of the BBU Pool cooperative techniques, as Multi-Cell MIMO
[58] can be enhanced. This can be achieved due to tighter cooperation between base station
within a pool. In [59], Liu et al. present a downlink Antenna Selection Optimization
scheme for MIMO based on C-RAN that showed advantages over traditional antenna
selection schemes.
technology, C-RAN BBU Pool can support multi-standard multi-system radio communi-
cations configured in software. Upgrades to new frequencies and new standards can be
done through software updates rather than hardware upgrades as it is often done today on
non-compatible vertical solutions. Multi-mode base station is therefore expected to allevi-
ate the cost of network development and Operations, Administration and Maintenance
(OAM).
Figure 2.6: Results of the survey [63] on operators drives for deploying C-RAN.
BBU
Base
Base
Base
poolBase
band
band
band
band
RRH BBU
Implementation
Section 2.7
2.3.1 A need for high bandwidth, strict latency and jitter as well as
low cost transport network
The C-RAN architecture brings a huge overhead on the optical links between RRH and
BBU Pool. Comparing with backhaul requirements, the one on fronthaul are envisioned
to be 50 times higher [54].
IQ data is sent between BBU and RRH as shown in Figure 2.2. The main contributors
to the size of IQ data are: turbocoding (e.g., in UMTS and LTE 1:3 turobocode is used
resulting in three times overhead), chosen radio interface (e.g., CPRI) IQ sample width and
oversampling of LTE signal. For example, 30.72 MHz sampling frequency is standardized
for 20 MHz LTE, which is more than 20 MHz needed according to Nyquist - Shannon
sampling theorem. Total bandwidth depends also on number of sectors and MIMO
configuration. Equation 2.1 summarizes factors that influence IQ bandwidth. Scenario of
20 MHz LTE, 15+1 CPRI IQ Sample width, 10/8 line coding, 2x2 MIMO transmission
resulting in 2.4576 Gbps bit rate in fronthal link is often treated as a baseline scenario.
Consequently, for 20 MHz 4x4 MIMO, 3 sector base station, the expected IQ throughput
exceeds 10 Gbps. Examples on expected IQ bit rate between cell site and BBU in LTE-A,
LTE, Time Division Synchronous Code Division Multiple Access (TD-SCDMA) and
GSM networks can be found in Table 2.2. The centralized BBU Pool should support 10 -
1000 base station sites [20], therefore a vast amount of data needs to be carried towards it.
2.3. Challenges of C-RAN 21
Table 2.2: IQ bit rates between a cell site and centralized BBU Pool
The transport network not only needs to support high bandwidth and be cost efficient,
but also needs to support strict latency and jitter requirements. Below different constraints
on delay and jitter are summarized:
1. The most advanced CoMP scheme, JT, introduced in Section 2.2.3 requires 0.5s
timing accuracy in collaboration between base stations, which is the tightest con-
straint. However, it is easier to cope with synchronization challenges in C-RAN
compared to traditional RAN due to the fact that BBUs are co-located in the BBU
Pool.
2. According to [20], regardless of the delay caused by the cable length, round trip
delay of user data may not exceed 5 s, measured with the accuracy of 16.276ns
on each link or hop [33].
3. The sub-frame processing delay on a link between RRHs and BBU should be kept
below 1 ms, in order to meet HARQ requirements. Due to the delay requirements
of HARQ mechanism, generally maximum distance between RRH and BBU must
not exceed 20-40 km [20].
Table 2.3: Requirements for cloud computing and C-RAN applications [54]
This section presents the technical solutions enabling C-RAN by discussing on transport
network, covering physical layer architecture, physical medium, possible transport net-
work standards and devices needed to support or facilitate deployments. Moreover, IQ
compression techniques are listed and compared.
The main focus of this section is on fronthaul transport network, as this is characteristic
for C-RAN. Considerations on backhaul network can be found in, e.g., [67]. The choice
of the solution for the particular mobile network operator depends on whether C-RAN
is deployed from scratch as green field deployment or introduced on top of existing
infrastructure. More on deployment scenarios can be found in Section 2.8.
24 Chapter 2. C-RAN overview
RF RRH RF RRH
O&M
L3 RF RRH
L2 RF
L1 RRH
RF RRH/L1 RF RRH/L1
L1 L1
O&M
L3
RF RRH/L1
RF L1
L2
L1 RRH/L1
Fiber Microwave
Figure 2.8: C-RAN architecture can be either fully or partially centralized depending on L1
baseband processing module location.
2.4. Transport network techniques 25
CPRI / CPRI /
OBSAI
Microvawe
OBSAI
Compression Compression
RRH BBU
Pool
Point to point
CPRI / OBSAI
Technology is ready
CPRI /
OBSAI
WDM CPRI /
OBSAI
(capacity not necessarily)
WDM-PON
Internet
OTN
CPRI/ CPRI/
OBSAI OBSAI
Ethernet
Eth/CPRI/ Eth/CPRI/
OBSAI OBSAI
2.4.2.2 WDM/OTN
Wavelength-division multiplexing (WDM)/Optical Transport Network (OTN) solutions
are suitable for macro cellular base station systems with limited fiber resources, especially
in the access ring. The solution improves the bandwidth on BBU-RRH link, as 40-80
optical wavelength can be transmitted in a single optical fiber, therefore with 10 Gbps
large number of cascading RRH can be supported, reducing the demand on dark fiber. On
the other hand, high cost of upgrade to WDM/OTN need to be covered. However, as the
span on fronthaul network does not exceed tens of kilometers, equipment can be cheaper
than in long distance backbone networks. Usage of plain WDM CPRI transceivers was
discussed and their performance was evaluated in [73]. [23] applies WDM in their vision
of C-RAN transport network.
In [74] Ponzini et al. describe the concept of non-hierarchical WDM-based access for
2.4. Transport network techniques 27
C-RAN. The authors have proven that WDM technologies can more efficiently support
clustered base station deployments offering improved flexibility in term of network
transparency and costs. Using that concept already deployed fibers, such as Passive
Optical Networks (PONs) or metro rings, can be reused to carry any type of traffic,
including CPRI, on a common fiber infrastructure. By establishing virtual P2P WDM
links up to 48 bidirectional CPRI links per fiber can be supported.
For scarce fiber availability ZTE proposes enhanced fiber connection or xWDM/OTN
[65]. Coarse WDM is suitable to be used for TD-SCDMA, while Dense WDM for LTE,
due to capacity requirements.
OTN is a standard proposed to provide a way of supervising clients signals, assure
reliability compared with Synchronous Optical NETworking (SONET)/SDH network as
well as achieve carrier grade of service. It efficiently supports SONET/SDH as well as
Ethernet and CPRI. CPRI can be transported over OTN over low level Optical channel
Data Unit (ODU)k containers as described in ITU-T G.709/Y.1331 [75], [76].
IEEE 802.1Q Virtual LAN (VLAN) standard through IEEE 802.1ad Provider Bridges (PB)
and IEEE 802.1ah Provider Backbone Bridges (PBB). To achieve Quality of Service (QoS)
of Ethernet transport service, traffic engineering is enabled in Carrier Ethernet. PBB-TE
uses the set of VLAN IDs to identify specific paths to given MAC address. Therefore
a connection-oriented forwarding mode can be introduced. Forwarding information is
provided by management plane and therefore predictable behavior on predefined paths
can be assured. Carrier Ethernet ensures 99.999% service availability. Up to 16 million
customers can be supported which removes scalability problem of PBB-TE predecessor
[80]. Carrier Ethernet grade of service can also be assured by using MPLS Transport
Profile (MPLS-TP). Technologies are very similar, although PBB-TE is based on Ehternet
and MPLS-TP on Multiprotocol Label Switching (MPLS).
The main challenge in using packet passed Ethernet in the fronthaul is to meet the
strict requirements on synchronization, syntonization and delay. Synchronization refers
to phase and syntonization to the frequency alignment, respectively. Base stations need
to be phase and frequency aligned in order to, e.g., switch between uplink and downlink
in the right moment and to stay within their allocated spectrum. For LTE-A frequency
accuracy needs to stay within 50ppb (for a wide area base station) [6.5 in [81]] while
phase accuracy of 1.5s is required for cell with radius 3km [82]. Time-Sensitive
Networking (TSN) features help to achieve delay and synchronization requirements. More
information about them can be found in Section 4.7, while the whole Chapter 4 gives a
deep overview of challenges and solutions for Carrier Ethernet-based fronthaul.
Altiostar is an exemplary company providing Ethernet fronthaul [83].
Power Design
consumption complexity
Latency
Figure 2.10: Factors between which a trade off needs to be reached choosing an IQ
compression scheme.
Reducing signal sampling rate is a low complex solution having minimal impact on
protocols, improves compression up to 66% with some performance degradation [20].
30 Chapter 2. C-RAN overview
By applying non-linear quantization, more quantization levels are specified for the
region in magnitude where more values are likely to be present. This solution improves
Quantization SNR (QSNR). Mature, logarithmic encoding algorithms, like -Law or
A-law are available to specify the step size. Compression efficiency up to 53% can be
achieved. This method creates additional Ir interface complexity (interface between RRH
and BBU) [20].
IQ data compression can be done using e.g., Digital Automatic Gain Control (DAGC)
[20], [87]. This technique is based on reducing the signals dynamic range by normal-
izing the power of each symbol to the average power reference, therefore reducing the
signal dynamic range. This method affects Signal-to-noise ratio (SNR) and Error Vector
Magnitude (EVM) deteriorates in DL. Potential high compression rate can be achieved,
however the method has a high complexity and no mature algorithms are available.
One example of a frequency domain scheme is to perform sub carrier compression.
Implementing the FFT/Inverse FFT (IFFT) blocks in the RRH allows 40% reduction
of Ir interface load. It can be easily performed in DL, however RACH processing is a
big challenge. This frequency domain compression increases IQ mapping and system
complexity. It also requires costly devices, more storage and larger FPGA processing
capacity [20]. On top of that, it limits the benefits of sharing the equipment in C-RAN, as
L1 processing needs to be assigned to one RRH. Several patents have been filed for this
type of compression schemes.
In [88] Grieger et al. present design criteria for frequency domain compression
algorithms for LTE-A systems which are then evaluated in large scale urban filed trials.
Performance of JD under limited backhaul rates was observed. The authors proved
that a Gaussian compression codebook achieves good performance for the compression
of OFDM signals. The performance can be improved using Frequency Domain AGC
(FDAGC) or decorrelation of antenna signals. However, field tests showed very limited
gains for the observed setups.
Samardzija et al. from Bell Laboratories propose an algorithm [89] which reduces
transmission data rates. It removes redundancies in the spectral domain, performs block
scaling, and uses a non-uniform quantizer. It keeps EVM below 8% (3GPP requirement
for 64 QAM, as stated in [81]) for 17% of relative transmission data rate (compression
ratio defined as transmission rate achieved after compression to the original one). The
algorithm presented by Guo et al. [90], which authors are also associated with Alcatel-
Lucent Bell Labs removes redundancies in spectral domain, preforms block scaling, and
uses non-uniform quantizer. EVM stays within 3GPP requirements in simulations for
30% compression ratio. TD-LTE demo test results showed no performance loss for 50%
compression ratio.
Alcatel-Lucent Bell Labs compression algorithm reduces LTE traffic carried over
CPRI interface from 18 Gbps to 8 Gbps [22], achieving a 44% compression ratio.
The solution discussed in [91] adapts to the dynamic range of the signal, removes
frequency redundancy and performs IQ compression creating 10.5 effective bits out of 12
2.5. RRH development 31
bits of data. This method allows 50% to 25% of compression ratio introducing 0.5% 1 to
8% of EVM and latency below 1s for LTE signal.
Lorca et al. from Telefonica I + D in [92] propose a lossless compression technique
where actual compression ratios depend upon the network load. For downlink direction,
the algorithm removes redundancies in the frequency domain. Secondly, the amount of
control data is reduced to minimum sending only the necessary information to reconstruct
control signals at RRH. Moreover, a special constellation coding is used to reduce number
of bits needed to represent constellation symbols for QPSK, 16QAM and 64QAM mod-
ulations. For uplink direction user detection is used to transmit only occupied carriers.
Compression ratio of 33% is achieved at full cell load. Compression ratio up to 6.6% are
achieved for 20% cell load.
Park et al. [93] propose a robust, distributed compression scheme applicable for
UL transmission, which they combine with an efficient base station selection algorithm.
Their current work focuses on implementing layered compression strategy as well as joint
decompression and decoding. Results in terms of compression ratio and EVM are not
available.
Table 2.4 summarizes and compares various compression methods discussed in this
section. Compression of 33% is achieved by all the algorithms for which the ratio was
available. The best result, where the algorithm is known, is achieved by [89] and by [92]
under small network load.
To conclude, in order not to lose the cost benefit of BBU Pooling for renting a transport
network, mobile network operator needs to either own substantial amount of fiber or
use an IQ compression scheme. Moreover, the cost of the optical high speed module
must stay comparable to traditional SDH transport equipment in order to make C-RAN
economically attractive.
sector 20 MHz LTE macro cell with 4x2 MIMO [22], see Table 2.2. Existing standards -
CPRI and OBSAI can support connections between the BBU Pool and RRHs in C-RAN.
Moreover, NGMN in [94] envisions ORI as a future candidate protocol. However, as
the nature of the interface between RRH and BBU is changing with an introduction of
C-RAN, the existing protocols may need to be redefined in order to be optimized for high
volume transmission over long distances.
Alcatel-Lucent is offering a lightRadio solution for C-RAN [22]. It uses a multiband,
multistandard active antenna array, with MIMO and passive antenna array support. Alcatel-
Lucent is working towards two multiband radio heads (one for high and one for low bands).
Built-in digital modules are used for baseband processing. For C-RAN L1, L2 and L3 are
separated from radio functions.
In 2012, Ericsson announced the first CPRI over microwave connection implemen-
tation [95], which is interesting for operators considering the deployment of a partially
centralized C-RAN architecture.
2.6.2 Processors
Nowadays, Field-Programmable Gate Arrays (FPGAs) and embedded Digital Signal Pro-
cessor (DSP) are used for wireless systems. However, the improvement in the processing
power of General Purpose Processor (GPP) used in IT is giving the possibility to bring IT
and telecom worlds together and use flexible GPP-based signal processors.
DSPs are developed to be specially optimized for real-time signal processing. They
are powerful and use multicore (3-6) technology with improved processing capacity.
What is important for C-RAN, a real time OS running on DSP facilitates virtualization of
processing resources in a real time manner. However, there is no guarantee of backwards
compatibility between solutions from different, or even from the same manufacturer, as
they are built on generally proprietary platforms.
Texas Instruments [26] favors the usage of specialized wireless System on a Chip
(SoC), providing arguments that SoC consumes one-tenth of the power consumed by a
typical server chip, and has wireless accelerators and signal processing specialization.
Considerations about power consumption of signal processors are essential to achieve
reduction in power consumption for C-RAN architecture compared to the traditional
RAN. In addition, for the same processing power, a DSP solution will also have a lower
price compared to GPP. In [96] Wei et al. present an implementation of SDR system on
an ARM Cortex-A9 processor that meets the real-time requirements of communication
system. As SDR technology further enables to benefit from C-RAN this is an important
proof of concept.
GPPs are getting more and more popular for wireless signal processing applications.
The usage of GPP is facilitated by muli-core processing, single-instruction multiple data,
low latency off-chip system memory and large on-chip caches. They also ensure backward
compatibility, which makes it possible to smoothly upgrade the BBU. Multiple OSs with
real-time capability allow virtualization of base station signal processing.
China Mobile Research Institute proved that commercial IT servers are capable of
performing signal processing in a timely manner. Intel is providing the processors for
both C-RAN and traditional RAN [25]. More on Intel GPP solutions for DSP can be
found in [97]. In [98], Kai et al. present a prototype of a TD-LTE eNB using a GPP.
2.7. Virtualization 35
It did not meet real-time requirements of LTE system, which is of great concern when
using general processors for telecommunication applications. It used 6.587 ms for UL
processing, with turbo decoding and FFT taking most of it and 1.225 ms for downlink
(DL) processing, with IFFT and turbo coding being again the most time consuming.
However, this system was based on a single core, and multi-core implementation with 4
cores should make the latency fall within the required limits. Another approach to reach
the requirements is to optimize the turbo decoder as described in [99], where Zhang et al.
prove that using multiple threads and a smart implementation, 3GPP requirements can be
met. De-Rate Matching and demodulation have been optimized for GPP used for LTE
in [100]. In [101] Kaitz et al. propose to introduce a dedicated co-processor optimized
for wireless and responsible for critical and computation intensive tasks. This optimizes
power consumption at the cost of decreased flexibility. They have considered different
CPU partitioning approaches for LTE-A case.
The issue of real-time timing control and synchronization for SDR has been addressed
in [102]. A real-time and high precision clock source is designed on a GPP-based SDR
platforms and users are synchronized utilizing Round-Trip Delay (RTD) algorithm. The
mechanism is experimentally validated.
Table 2.5 summarizes the characteristics of DSP and GPP.
DSP GPP
Flexibility dedicated solution general purpose
Vendor compatibility vendor specific, higher compatibil-
proprietary ity between ven-
dors
Backward compatibility limited assured
Power consumption lower higher
Real-time processing optimized, only possible with
achieved high power hard-
ware
Virtualization of BBU possible possible
2.7 Virtualization
In order to implement Centralized RAN, BBU hardware needs to be co-located. However,
in order to deploy RAN in the cloud - Cloud RAN - virtualization is needed. Although
virtualization is not the focus point of this thesis, it is briefly presented in this section for
the completeness of the C-RAN introduction, as it is an important foundation of C-RAN.
36 Chapter 2. C-RAN overview
lishing the network from scratch, as well as deployment of additional cells for boosting
the capacity of an existing network. Moreover, different stages of C-RAN deployment to
leverage its full potential are listed.
It is advised to deploy C-RAN in metropolitan area to benefit from statistical multi-
plexing gain, as users are moving through the day, but still remain within the maximum
distance (resulting from propagation and processing delay, up to 40 km) between RRH
and BBU. However, a metropolitan area might be served by a few BBU Pools.
of capacity improvement deployment scenarios are listed below [68]. The combination of
mentioned solutions is also possible.
RRH f2 f1
RRH RRH/ RRH/ RRH/ RRH/
RRH RRH RRH Repeater Repeater Repeater Repeater
RRH RRH
RRH RRH RRH Railway/Highway
RRH RRH
Operator 2, Operator 1,
RRH
RRH RRH ... RANaaS RANaaS
RRH RRH
RRH RRH RRH
RRH
O&M
L3
RRH
g L2
ss rin rnet
Acce r Ethe L1
Overlay C arrie BBU Pool 1
TN/ Mux/Demux
O
RRH
L1
RRH
L1
1, 2, ... , n
RRH
RRH
L1 L1
RRH WDM-UniPON
RRH
L1
RRH
Operator 3
L1 L1
RRH RRH RRH O&M Mux/Demux
L1 L1 L1
L3
RRH
L1 L2 CPRI/OBSAI/ORI
RRH RRH RRH
L1 L1 L1
RRH
RRH RRH
L1 L1 BBU Pool 2
RRH RRH RRH
L1 L1 L1
RRH RRH
L1
RRH
L1 RRH
RRH
Cell split, small cell deployment RRH
RRH
with L1 integrated into RRH RRH
RRH RRH RRH
a) HetNets. Existing BBUs of macro Base Stations can be replaced by BBU Pool and
additional RRHs can be deployed to form small cells.
b) Cell split. Existing macro cells can be split into smaller ones increasing the system
capacity. Interference management techniques are needed as all the cells will
operate at the same frequency. As explained in Section 2.2.3, C-RAN can enhance
cooperative techniques like CoMP and eICIC. This scenario can also be used to
provide indoor coverage by deploying RRHs on each floor of the building or group
of offices offering high capacity. However, in this scenario Wi-Fi can be a cheaper
solution, if users will have Wi-Fi connection in their devices switched on, enabling
offload from cellular network to Wi-Fi.
c) Overlay. Additional frequency band or a new cellular standard can be introduced to
boost system capacity. In Figure 2.11 one RRH provides coverage in frequency f1 .
2.9. Ongoing work 39
d) Super hot spots, e.g., stadium, transportation hub. It is a scenario where many users
are present in one location. Small cells are needed to assure the capacity, as well
as provide the coverage in complex scenery, e.g., with balconies, ramps, etc. The
density of users is high, therefore it is crucial to efficiently support interference
management schemes like CoMP and eICIC.
e) Railway/highway. Users are moving with a fast speed in this scenario, therefore
BBU Pool shall handle frequent handovers faster than traditional RAN.
Figure 2.11 summarizes C-RAN transport solutions and physical layer architecture
discussed in this chapter. Moreover, a possibility of sharing BBU Pool and rent RANaaS
is emphasized. For a particular network operator the choice of physical medium and
transport network depends on whether an existing infrastructure is already deployed.
1. Centralized RAN, where baseband units are deployed centrally supporting many
RRHs. However, resources are not pooled, nor virtualized.
2. Cloud RAN
- Phase 1, where baseband resources are pooled. Baseband processing is done using
specialized baseband chip - DSPs,
- Phase 2, where resources are virtualized, using GPP, thereby leveraging full benefits
of C-RAN. Sometimes this deployment is referred to as V-RAN standing for
Virtualized-RAN.
leading the effort. More operators, like Vodafone Hutchison Australia are planning the
deployments in 2018-2020 time frame.
- Constant bit rate, independent on user traffic. CPRI line bit rate is constant,
independent on whether there is any user activity within a cell. A solution scaling
the bit rate depending on user activity is desired.
- One-to-one mapping RRH-BBU. With CPRI each RRH is assigned to one BBU.
For load balancing it is desired to be able to move RRH affiliation between pools.
- FH bandwidth scales with cell configuration. Current CPRI line rate depend on
cell configuration: carrier bandwidth, number of sectors and number of antennas.
Already for LTE-A up to 100 MHz can be used, and for Massive MIMO 100+
antennas can be expected. That yields 100+Gbps of data per cell, which is not
feasible. Solution depending on user activity, and operating on lower bit rates
is desired. Compression techniques presented in Section 2.4.4 received a lot of
attention, however, none of the solutions are economically viable or flexible to meet
5G requirements. Therefore more disruptive solutions are needed.
44 Chapter 2. C-RAN overview
Before 2013, to address the challenge of optimizing the fronthaul bit rate and flexibility
various compression techniques were considered. As more disruptive methods were
needed to achieve higher data rate reduction, nowadays (from 2013 onwards) a new split
between RRH and BBU functionality is under extensive analysis by Alcatel Lucent Bell
Labs [146], [147], NGMN [68], Small Cell Forum [148], [149] and many others. A
new working group NGFI is under preparation (2015) under the sponsorship of an
IEEE-SA Standards Sponsor with founding members of AT&T, Huawei, CMCC and
Broadcom (more to join) [46]. Its goal is to encourage discussion on optimal functional
splits between the pool and the cell site to address the above mentioned shortcomings of
traditional fronthaul. The considered functional splits are marked with arrows in Figure
2.12. Functions to the left of each arrow will be executed centrally, while functions to
the right of the arrow will be executed by the device at the cell site. Split can be done
per cell [148], or even per bearer [150]. It may also be beneficial to implement different
split for DL and UL. Most likely a few functional splits will be implemented per BBU
pool, as one solution does not fit all the deployment scenarios. That requires a variety
of devices at the cell site that will have the remaining functions implemented. With the
current functional split (BB-RF), RRH could be standard-agnostic, especially when high
frequency range was supported. Moving parts of L1 (and higher layers) to the devices at
the cell site makes them dependent on mobile network standards.
L3 L2 L1
Encoding
Resource
mapping
mapping
Antenna
QAM +
S D
CP in
CFR
IFFT
FEC
CPRI/OBSAI/ORI
CPRI/OBSAI/ORI
IQ DL R U / DAC
Control-RRC
Applicaions
C C DPD Frequ
PDCP
... ency
MAC
MAC
RLC
S D filter
IQ UL
demapping
processing
Decoding
Resource
R D
Antenna
ADC
QAM-1 +
CP out
FEC-1
FFT
C C
Executed centrally
Various functional splits pose different throughput and delay requirements, as pre-
sented in Table 2.7. To benefit the most from centralization, the lowest split is rec-
2.10. Future directions 45
ommended, i.e. the closest to the current BB-RF split. However, in order to save on
bandwidth higher splits can be more applicable. User Equipment (UE)-Cell split (sep-
arating user and cell specific functionalities) is the lowest one for which data will have
a variable bit rate, dependent on user traffic. Moreover, higher splits allow for higher
fronthaul latency, e.g., 30 ms for PDCP-RLC split.
Table 2.7: Requirements for different functional splits [148] for the LTE protocol stack
NodeB
RRH
Mobile core
3G
network
NodeB RNC
eNodeB
Mobile core
4G
network
eNodeB
2.11 Summary
This chapter presents a detailed overview of the novel mobile network architecture
called C-RAN and discusses the advantages and challenges that need to be solved before
its benefits can be fully exploited. C-RAN has the potential to reduce the network
deployment and operation cost and, at the same time, improve system, mobility and
coverage performance as well as energy efficiency.
The work towards resolving C-RAN challenges has been presented. Critical aspects
such as the need for increased capacity in the fronthaul, virtualization techniques for the
BBU pool and hardware implementation have been discussed. First prototypes and field
trials of networks based on C-RAN have also been presented, together with most likely
deployment scenarios.
While the concept of C-RAN has been clearly defined, more research is needed to
find an optimal architecture that maximizes the benefits behind C-RAN. Mobile network
operators as well as the telecommunication industry show a very high interest in C-RAN
due to the fact that it offers potential cost savings, improved network performance and
possibility to offer IaaS. However, the implementation of C-RAN needs to be justified by
particular network operators taking into account available fronthaul network capacity as
well as cost of virtualization of BBU resources. As the required fronthaul capacity is one
of the main challenges for C-RAN deployments, the work on defining the new functional
split is of utter importance.
2.11. Summary 47
Mobile core
2G/3G/4G
network
RRH
BBU
RRH
Mobile core
3G/4G
network
BBU
RRH
RRH
Mobile core
4G/5G?
network
BBU pool
c) C-RAN
RRH++
RRH++
Mobile core
5G?
network
BBU pool --
49
50 Chapter 3. Multiplexing gains in Cloud RAN
comprehensive analysis of the results obtained in various projects with different models
used and exposing traffic burstiness and tidal effect. Section 3.7 summarizes this chapter.
A part of this section, Sections 3.2.3 - 3.1 and 3.4 were previously published in
[13]. Section 3.2.2 was previously published in [11], while Section 3.3 was previously
published in [12]. All of them are tailored and updated to fit this chapter.
3.1 Methodology
This chapter addresses the topic of multiplexing opportunities that are possible in C-RAN
inside a BBU pool and on a fronthaul network. The conclusions are drawn based on
the observation of the traffic aggregation properties in the simulated models and are an
approximation of actual multiplexing gains on BBUs and fronthaul. This section aims at:
1) defining terms: multiplexing gain and pooling gain, 2) detailing elements of the LTE
protocol stack and their impact on traffic shape, 3) explaining the relationship between
the traffic aggregation properties and the gain in power consumption, processing resource
and transport.
3.1.1 Definitions
This section defines Multiplexing gain (MG) and Pooling gain (PG), both in terms of
processing resources and power.
MG - ratio between the sum of single resources and aggregated resources, showing
how many times fewer resources are needed to satisfy users needs when they are aggre-
gated. Equation 3.1 shows an example for fronthaul resources, e.g. throughput. Link
resources specify sufficient bandwidth for a given deployment. They can be defined in
several ways, as providing peak throughput requested by users is costly: 1) the 95th per-
centile of requested throughput (used in Section 3.4.4), 2) mean plus standard deviation of
requested resources (used in the teletraffic approach in [13]), 3) peak requested throughput
averaged over given time (used in Sections 3.3, 3.2 and 3.4.2), and 4) link data rate for
which the application layer delay is acceptable (used in Section 3.4.3). MG can refer to
any resources, however, in the thesis it refers to savings on the transport resources. They
are possible when several links are aggregated.
cells
P
SingleCellLinkResources
MG = (3.1)
AggregatedLinkResources
PG - savings on BBUs comparing C-RAN with RAN (base station architecture with
or without RRH, but not centralized). We can distinguish pooling gains in terms of
processing, computational resources P Gprocessing and in terms of power savings
P Gpower . Definition of P Gprocessing is similar as for multiplexing gain but instead of
Link Resources, BBU resources are considered, as presented in Equation 3.2. P Gpower
is defined in 3.3 in terms of power consumed by both architectures.
3.1. Methodology 51
cells
P
BBResourcesRAN
P Gprocessing = (3.2)
BBResourcesBBU pool
cells
P
BBP owerRAN
P Gpower = (3.3)
BBP owerCRAN
The multiplexing gain value obtained in the thesis refers to M G. Section 3.1.3
describes the relation between M G, P Gprocessing and P Gpower .
PDCP PGPDCP
User processing
RLC PGRLC
Variable
bit rate MAC PGMAC
Bit-level
PGBLP
processing
QAM P PGQAM
Antenna mapping
Resource mapping
H
processing
IFFT Y PGRM_IFFT
Cell
Constant
bit rate
CP PGCP
Figure 3.1: Layer 2 (green) and Layer 1 (yellow) of user-plane processing in DL in a LTE
base station towards air interface. Possible P Gs are indicated. Based on [151], [148], [147],
[46].
layer. Still, it is assumed that the P G on this layer is equal to M G due to variable
bit rate traffic.
- Media Access Control (MAC) - this layer is responsible for data transfer and radio
resource allocation. It maps logical channels and transport channels into which
information is organized, as well as multiplexes different logical channels into
transport blocks to be delivered to the physical channel on transport channels, and
demultiplexes information from different logical channels from transport blocks
delivered from the physical layer on transport channels. It reports scheduling
information, does error correction by Hybrid ARQ (HARQ), performs priority
handling between UEs and between logical channels of one UE, and it selects the
transport format [154]. Variable bit rate traffic is expected, following the users
activity. Above mentioned processing will change data pattern comparing to the
traffic present at the RLC layer. Still, it is assumed that the P G on this layer is
equal to M G due to variable bit rate traffic.
3.1. Methodology 53
- various bit level processing takes place including Cyclic Redundancy Check (CRC)
insertion, channel coding, HARQ and scrambling, increasing the traffic volume
[151]. Still, a variable bit rate traffic is expected, following the users activity.
This functionalities can be put under the umbrella of Forward Error Correction
(FEC). Above mentioned processing will change data pattern comparing to the
traffic present at the MAC layer. Still, it is assumed that the P G on this layer is
equal to M G due to variable bit rate traffic.
- Quadrature Amplitude Modulation (QAM) - downlink data modulation transforms
a block of scrambled bits to a corresponding block of complex modulation symbols
[151]. A variable bit rate traffic is expected in a form of IQ data, therefore it is
assumed that the P G on this layer is equal to M G.
- the antenna mapping processes the modulation symbols corresponding to one or
two transport blocks and maps the result to different antenna ports [151]. A variable
bit rate traffic is expected in a form of IQ data, therefore it is assumed that the P G
on this layer is equal to M G.
- the resource-block mapping takes the symbols to be transmitted on each antenna
port and maps them to the resource elements of the set of resource blocks assigned
by the MAC scheduler for the transmission. In order to generate an Orthogonal
Frequency-Division Multiple (OFDM) signal IFFT is performed [151] resulting in
constant bit rate stream of IQ data, therefore it is assumed that the P G on this layer
is equal to one.
- cyclic-prefix (CP) insertion - the last part of the OFDM symbol is copied and
inserted at the beginning of the OFDM symbol [151]. The data remains constant
bit rate, therefore it is assumed that the P G on this layer is equal to one.
For each of these steps computational resources are required. The BBU pool needs to
be dimensioned according to the planned user activity in order to meet strict real-time
requirements of LTE processing. Depending on the functional split, for each of the layers,
when it is included in the BBU pool, pooling gain is possible, both in terms of process-
ing resources they require and power those resources consume: P GP DCP , P GRLC ,
P GM AC , P GBLP , P GQAM , P GRM _IF F T and P GCP . Section 3.1.3 elaborates on
such gains, splitting functionalities in user-processing and cell-processing. Assumed pool-
ing gain on user-processing resources equals to M G while pooling gain on cell-processing
resources will be equal to one, as listed in Table 3.1. In fact P GP DCP will be closest to
M G as the data is closest to the backhaul data. Processing by lower layers will change
traffic properties, therefore pooling gain can be approximated to M G.
Table 3.1: Assumed pooling gains on different layers of the LTE protocol stack.
PG Value
P GP DCP MG
P GRLC MG
P GM AC MG
P GBLP MG
P GQAM MG
P GRM _IF F T 1
P GCP 1
Subsections below explain sources and nature of gains in processing resources, power
consumption and transport network resources. The relation between M G, P Gprocessing
and P Gpower is described.
The BBU pool needs to be dimensioned according to the planned user activity. The
more traffic is expected, the more computing resources will need to be available (e.g.
rented from a cloud computing provider). Modules for control plane BBctrl will need to
be provided, for cell-processing traffic-independent functionalities BBc , like FFT/IFFT,
decoding/encoding as well as for user-processing traffic dependent resources BBu , like
Bit-level processing (BLP). Equation 3.2 can be further expanded into equation 3.4 by
inserting these components. Control resources, if dimensioned precisely, will be similar in
RAN and C-RAN. In fact, in case of C-RAN, more control information may be needed to
coordinate cells for e.g. CoMP. Cell-processing resources, also if dimensioned precisely
will be similar in RAN and C-RAN. However, user-processing resources could be reduced
by the factor of M G as listed in equation 3.5.
3.1. Methodology 55
cells
P
BBResourcesRAN
P Gprocessing = =
BBResourcesBBU pool
cells
P
(BBRANctrl + BBRANc + BBRANu )
= (3.4)
BBBBU poolctrl + BBBBU poolc + BBBBU poolu
cells
P cells
P cells
P
BBRANctrl + BBRANc + BBRANu
cells
P
(3.5)
cells
P cells
P BBRANu
BBRANctrl + BBRANc + MG
Therefore it can be seen that multiplexing gain calculated in the PhD project will
affect only the amount of processing resources required for user-processing modules,
leaving the fixed part aside. The factor of M G cannot be applied directly, as real life
implementation needs to respond in real time to changing traffic, while averaged values
were evaluated in the project, hence the almost equal to sign between equations 3.4 and
3.5. However, referring to the equation 3.2, it is worth noticing that, in practice, BBUs
will always be dimensioned with a margin comparing to planned consumption, allowing
to accommodate higher traffic peaks if they occur and to account for forecasted overall
traffic growth. When BBUs are aggregated in a pool, such a margin could be shared,
allowing additional pooling gain in C-RAN comparing to RAN. Such an additional margin
is especially applicable to resources required to support traffic peaks in different cells. In
other words, capacity can be scaled based on average utilization across all cells, rather
than all cells peak utilization. Moreover, the processing power can be dynamically shifted
to heavily loaded cells as the need arises.
In [41], Werthmann et al. show results on pooling gain, analyzing compute resource
utilization, in Giga Operations Per Second (GOPS), for different number of sectors
aggregated into a BBU pool. In [155], Desset et al. elaborate on how many GOPS are
needed for different base station modules, including:
Table 3.2: Estimations of baseband complexity in GOPS of cell- and user-processing for UL
and DL and different cell sizes. Numbers are taken from [155].
DL UL
macro femto macro femto
GOP Sc 830 180 700 240
GOP Su 30 25 140 120
%GOP Su 3% 12% 17% 33%
- FEC
For the reference case, assuming 20 MHz bandwidth, single-antenna, 64-QAM, rate-1
encoding and a load of 100%, and treating DPD, filtering, SERDES, OFDM, and linear
part of FD as cell-processing, and non-linear part of FD and FEC as user-processing,
Table 3.2 summarizes estimations on GOPS for UL and DL for macro and femto cell
for cell-processing GOP Sc and for user-processing GOP Su . It would require further
investigation on how to split platform control processor resources. It can be concluded
that especially for downlink, share of user-processing resources is small comparing to
cell-processing resources (3-12%). For uplink processing relatively more can be saved on
user-processing (17-33%).
To conclude, M G calculated in the project applies only to user-processing resources,
which are a fraction of total resources. Moreover, as mentioned above there are factors
implying that M G calculated in the project is actually both lower (protocol processing)
and higher (dimensioning margin) in reality. Given the complexity of the system and
amount of factors influencing the final result, obtained M G gives only a contribution to
the approximation of the real-life P Gprocessing .
- powering the main equipment (51% of the cell site power consumption for tradi-
tional base station in China Mobile [20], 83% for Vodafone [156] found in [157]
and [158])
- air conditioning (46% of the cell site power consumption for traditional base station
in China Mobile [20], 10-25% for Vodafone [156])
Power consumption of the main equipment can be further broken down into:
3.1. Methodology 57
. Based on this data P owerctrl + P owerc + P oweru account for 5-10% of total base
station power consumption. In [159] Auer et al. reports that 2-24 % of total base station
consumption are spent on baseband processing.
Similar as for processing resources, equation 3.3 can be further expanded into equation
3.6 by inserting these components. Also, only user-processing resources could be reduced
by the factor of M G as listed in equation 3.7.
cells
P
BBP owerRAN
P Gpower = =
BBP owerCRAN
cells
P
(P owerRANctrl + P owerRANc + P owerRANu )
= (3.6)
P owerBBU poolctrl + P owerBBU poolc + P owerBBU poolu
cells
P cells
P cells
P
P owerRANctrl + P owerRANc + P owerRANu
cells
P
(3.7)
cells
P cells
P P owerRANu
P owerRANctrl + P owerRANc + MG
In [155], Desset et al. propose a more detailed power model of a base station,
listing factors that impact power consumption. Power consumption is proportional to
GOPS: the number of operations that can be performed per second and per Watt is 40
58 Chapter 3. Multiplexing gains in Cloud RAN
GOPS/Ws for large base stations and default technology, here i.e., 65 nm General Purpose
Complementary metal-oxide-semiconductor (CMOS), and it is three times larger for pico
and femto cells. Therefore P Gpower on user-processing resources can be reduced by the
same factor as P Gprocessing on user-processing resources.
To conclude, P Gpower , similarly to P Gprocessing , accounting for 2-24% of base sta-
tion power consumption, is affected by M G by reducing power spent by user-processing
resources. Model including the whole LTE processing stack would be needed to determine
overall savings more precisely, as well as detailed information on how user traffic impacts
power consumption.
3.1.4 Approach
Mathematical, teletraffic theories have been used to calculate an overbooking factor [160]
that dimensions the link, based on the definition of an effective bandwidth [161]. They
provide an important indication when the fundamentals of the networks are studied.
However, teletraffic theories focus on well-defined traffic models, such as ON - OFF
source traffic e.g., Interrupted Poisson Process, Interrupted Bernoulli Process [161]. As
such, they do not capture all the aspects of real-life networks. In current and future mobile
networks there is a big variety of applications and the traffic varies throughout the day
depending on cell location (office or residential). In order to capture this heterogeneity the
analysis is done in a discrete event-based simulator - OPNET [162]. Such scenario with
detailed and heterogeneous traffic definition is especially important to evaluate the UE-
Cell split (introduced in Figure 2.12. Simulations allow to capture protocol interactions
and thereby observe implications of different network architectures on end-to-end delay
seen at the application layer. On the other hand, a mathematical approach allows to create
simpler models that can run with lower simulation time, thereby enabling testing more
extended scenarios, e.g., with more cells. Both approaches are important to compare
and validate the results. This chapter reports simulation results that have been validated
analytically by collaborators, as presented in Section 3.3.2.
Quantifying multiplexing gains has been addressed by research and industry commu-
nities. Related work is presented in Section 2.2.1 and 3.4.4.
3.2. Origins of multiplexing gain 59
40 2
DL data
1.5
Load
30
20 1
10 0.5
0 0
0 6 12 18 24 0 6 12 18 24
Traffic in New York, from MIT/Ericsson Traffic in New York, from MIT/Ericsson
Lower Manhattan Ridgewood Business Residential Commercial
2.5 2.5
2 2
DL data
DL data
1.5 1.5
1 1
0.5 0.5
0 0
0 6 12 18 24 0 6 12 18 24
Time (h) Time (h)
(c) Example of traffic in office and (d) Overall traffic in New York
residential cell in New York
and residential areas, only peaking in the office area in business hours, hence lower values
of multiplexing gain. Data for Los Angeles: Downtown and Florence, resembles data
from London. Values for New York, Figure 3.2(c) lie in between values for London and
data from China Mobile, as the curves of traffic.
Typically in the cities there are not only residential and office areas but also commercial
areas, like shopping malls, movie theaters as well as parks and mixed areas. They will
all have different daily and weekly traffic distributions that will affect the value of
multiplexing gain. Based on the available data impact of commercial sites is estimated
as shown in Figure 3.4. It is based on the overall data for New York from MIT/Ericsson
presented in Figure 3.2(d). Analyzed deployments consist of office, residential and
commercial cells. Values vary between 1.0, for 100% of one cell type, to 1.21 for 30% of
office cells, 70% residential and 0% commercial. As shown in Figure 3.2(d) commercial
areas in New York have a uniform traffic distribution in the daytime therefore they lower
the multiplexing gain.
To conclude, in order to maximize multiplexing gains it is best to combine in a BBU
pool cells whose daily traffic distributions are complementary i.e. traffic is low in some
3.2. Origins of multiplexing gain 61
1.3 London/Los
Multiplexing gain
Angeles
1.2 New York
1.1
Hong Kong
1.0
China Mobile
0.9
0.8
0 10 20 30 40 50 60 70 80 90 100
Office cells (%)
Figure 3.3: Multiplexing gains for different locations based on daily traffic distributions
between office and residential cells.
Figure 3.4: Multiplexing gains for different distributions between office, residential and
commercial cells.
by office workers on one floor. In order to capture traffic burstiness real-life application
characteristics need to be taken into account.
Paper [11] presents an initial evaluation of the multiplexing gain including both the
tidal effect and traffic burstiness. A real case scenario is modeled with the mobile traffic
forecast for year 2017, a number of recommendations on traffic models, including a
daily traffic variations between office and residential cells as presented in [20] and a
proposed RAN and C-RAN implementation. RAN and C-RAN are modeled by nodes
performing simple traffic aggregation. Multiplexing gain is evaluated as in equation 3.1,
where LinkResources are represented by the peak requested throughput.
The results show that the statistical multiplexing gain for user traffic in a C-RAN
architecture, when traffic burstiness is taken into account, is 4.34, = 1.42, compared
to a traditional RAN architecture. Please refer to [11] for more details on the model and
results.
Residential cell
Aggregated Traffic traffic (over 24h)
(over 24h)
a)
Aggregation link
b)
PDCP-RLC split
c)
UE-Cell split
CPRI
CPRI
d)
CPRI
BB-RF split
PDCP
RLC/MAC Present layer
backaul
PHY Absent layer BBU
BBU pool Switch RRH
Layers present fronthaul
at BBU side Legend
Figure 3.5: Possible multiplexing gains on BBU pool and fronthaul links depending on base
station architecture.
As stated before, multiplexing gain comes from traffic burstiness and from the tidal effect.
This section explores the tidal effect and summarizes efforts to find the most optimal mix
of residential and office cells in order to maximize the multiplexing gain. The results can
be applied to the BBU pool for BB-RF split.
Application Application
TCP TCP
IP IP
Office cells
Aggregated
.
Traffic (h)
Traffic (h)
24 h
24 h Switch
Residential
BBU Pool
cells
.
Traffic (h)
24 h
10
1
0 4 8 12 16 20 24
Time (h)
Figure 3.7: Modeled traffic from residential, office and aggregated traffic from 5 office and
5 residential cells.
66 Chapter 3. Multiplexing gains in Cloud RAN
Table 3.4: Traffic generation parameters for network modeling; C - Constant, E - Exponential,
L - log-normal, U - uniform, UI - uniform integer
MG 1 1.6 1
BBUS aving = = (3.8)
MG 1.6
Multiplexing gain for all the cases is above 1.2, corresponding to 17% of BBU savings.
Different traffic profiles will affect the results, however the same model can be used to
process different input data.
3.4. Exploring different resources measurement methods and application mixes 67
1.8
1.6
1.4
Multiplexing gain
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0 10 20 30 40 50 60 70 80 90 100
Office cells (%)
Figure 3.8: Optimal distribution of office and residential cells - simulation results. Confidence
intervals for 95% level are shown.
The previous section investigates multiplexing gains resulting from the tidal effect. In this
section the most optimal office/residential cells mix is considered and the different meth-
ods of evaluating LinkResources from equation 3.1 are used and compared. Moreover,
the investigation focuses on the application mix impact, therefore the percentage of web
and video traffic is varied, while the total offered traffic follows the daily load for each
simulation run. Those results study the impact of traffic burstiness, therefore they can be
applied to both BBU pool and fronthaul for UE-Cell split.
68 Chapter 3. Multiplexing gains in Cloud RAN
Table 3.5: BBU save for various office/residential cell mixes, measured using different
methods.
Table 3.6: Traffic generation parameters for network modeling; C - Constant, E - Exponential,
L - log-normal, G - gamma, U - uniform
Parameter Value
Modeler and simula- OPNET 17.5.A PL3
tion software
Simulated time 16 h
Seeds 24, random
Values per statistic For throughput measurements every
10 ms, for delay measurements ev-
ery 1 s
70 Chapter 3. Multiplexing gains in Cloud RAN
Application Application
TCP TCP
IP IP
3 office cells
Aggregated
.
Traffic (h)
Traffic (h)
24 h
24 h Switch
7 residential
BBU Pool
cells
.
Traffic (h)
24 h
1 ms. For practical reasons, in order to be able to process the results efficiently, the
data is collected every 10 ms. Operators will most likely not dimension their networks
for peak user data measured over 1 ms, but allow some buffering, thereby saving the
costs, although lowering the user data rate. Therefore different averaging is applied over
simulated throughput as follows.
For each cell c and for the aggregated link a the data set resulting from simulations
consists of 16hours
10ms = 5760000 throughput measurements x measured at time t. An
averaging window (bucket) is defined with width W such that for the samples (ti , xi )
where i = 0, 1, ..., n and tn t0 = W . The averaging window size represents the networks
ability to smoothen the traffic and has a similar function to a buffer. 16 hours simulated
time is divided into such windows W and for each of them an average throughput is
n
P
xi
i=0
calculated y = n . Out of all the y values, a maximum value of all the averages is
found ymax for each cell ymax,c and for an aggregated link ymax,a . Based on equation
(3.1) M GAV G is calculated as presented in equation (3.9).
cells
P
ymax,c
c=1
M GAV G = (3.9)
ymax,a
Values of M GAV G coming from simulations for different web and video traffic
3.4. Exploring different resources measurement methods and application mixes 71
5
Multiplexing gain
10 ms
4 100 ms
3 1s
10 s
2
57.6 s
1
100 s
0
0 10 20 30 40 50 60 70 80 90 100
Web traffic (%)
Figure 3.10: Multiplexing gain for for different percentage of web traffic in the system
and different throughput averaging windows: M GF HU ECell (10 ms, no averaging) and
M GF HU ECellAV G (100 ms, 1 s, 10 s, 57 s and 100 s).
mixes are presented in Figure 3.10. Confidence intervals for 95% level are calculated
using the Students t distribution. Different series present data averaged over 100 ms,
1 s, 10 s, 57 s and 100 s (averaging window W ). For 10 ms series throughput is not
averaged, only the maximum values are taken for each cell and aggregated link to compute
M G. Values vary for different mixes of web traffic. For no web traffic present in the
network, the multiplexing gain has similar value within our averaging intervals, as video
conferencing sessions have constant bit rates. As soon as web traffic is present (17-100%)
multiplexing gain varies from 1.5 to 6 depending on the averaging window. It can be
seen that multiplexing gain is very sensitive to the measurement interval. There is a clear
dependence of the averaging period on the multiplexing gain.
In principle, if we take longer, up to infinite, averaging periods the multiplexing gain
should be getting lower and reaching one, as the average bandwidth of an aggregation
link will need to match the sum of average bandwidths of single links. Therefore it is not
straightforward why the value is low for every 10 ms, then increases for 100 ms and 1 s and
then lowers again. The possible cause could be that the Cumulative Distribution Function
(CDF) of the throughput looks in a way that for 90% of the time the throughput to base
stations is below 100 Mbps and aggregated throughput is below 1 Gbps, as presented in
Figure 3.11. This indicates that by adequate dimensioning the multiplexing gain value
can be different. Moreover, if the dimensioning is done according to the results from
averaging over longer periods, the risk of dropped packets and connections will increase,
as buffer sizes may be exceeded and packets may be dropped or users may not be satisfied
72 Chapter 3. Multiplexing gains in Cloud RAN
with the delay. In this study none of the packets were dropped. The averaging is done
only in post processing of the data, so actually it is not verified what will be the impact
of providing only the data rates as averaged on the application level delays. For video
conferencing and web browsing averaging only up to 10 - 100 ms is safe, as application
layer delays should not exceed the order of magnitude of 150 ms and 1 s, respectively.
Delays on the application level will give an ultimate criterion for network dimensioning.
The following section elaborates on them.
1.2
1.0
0.8
Office (1)
CDF
0.6
Residential (1)
0.4 Total (10)
0.2
0.0
0 1 2 3 4 5 6
Throughput (Gbps)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0 0.1 0.2 0.3
Figure 3.11: CDFs of throughput for an sample office and residential cell as well as total
throughput for all ten cells for 50% web traffic mix.
Mbps. For video conferencing 90th percentile of packet end-to-end delays are below 150
ms for link 200 Mbps.
As expected, the delays are lower when the offered link data rates are higher. The
impact on delay is higher for the cases with less web traffic. It is due to the fact, that the
more video traffic is present in the network, the delays are more sensitive to the aggregated
link bandwidth. Small change of data rate affects delay considerably, even by a factor of
10 (for 17% of web traffic - 83% of video traffic). The reason could be that video occupies
a link at a constant bit rate for at least a minute, so if the links are under-dimensioned
queuing occurs. The conclusion is, that the more bursty the traffic is, the less sensitive is
it to under-dimensioning. The more video traffic present in the network the dimensioning
becomes more relevant for achieving quality of service. Traffic forecast [2] predicts that
in 2020 60% of the mobile traffic will be video; it will, however, vary between light and
heavy users.
100.0
90th percentile of web page
response time, s
10.0
100 Mbps
150 Mbps
1.0 200 Mbps
250 Mbps
0.1
0 10 20 30 40 50 60 70 80 90 100
Web traffic (%)
Figure 3.12: 90th percentile of web page response time for different percentage of web traffic
in the system and for different aggregated link data rate.
100.000
conferencing Packet End-to-End
90th percentile of v ideo
10.000
1.000
100 Mbps
Delay (s)
0.100
150 Mbps
0.010 200 Mbps
0.001 250 Mbps
0.000
0 10 20 30 40 50 60 70 80 90 100
Web traffic (%)
Figure 3.13: 90th percentile of video conferencing packet End-to-End delay for different
percentage of web traffic in the system and for different aggregated link data rate.
sum of 90th percentile of throughputs from fronthaul links and the 80th percentile of
aggregated link need to be provided (here 200 Mbps). In case of under-dimensioning, for
higher percentages of web traffic the delay increase will be lower, as is the sum of the
80th and the 90th percentiles.
Figure 3.14: 80th, 90th and 95th percentile of base stations throughput for different
percentage of web traffic in the system.
These results can be used not only for quantifying multiplexing gains but also for
network dimensioning providing traffic distribution with a CDF similar to the one studied
here. For upgrading existing networks, operators could measure single cell throughput
3.5. Discussion and verification of the results 75
and calculate the 90th percentile of it and measure 80th percentile of aggregated traffic.
Then depending on how many cells should belong to one BBU pool/be aggregated on a
single fronthaul link, the higher number will assure the necessary capacity. If it is not
met, it means that links should be upgraded to the next available data rate. For green
field deployments based on traffic forecast, the operators will need to estimate what is the
90th percentile of throughput they would like to offer to the users. The challenge then
would be to add those throughputs taking into account forecasted user activity. Having the
sum of the 90th percentile of such aggregated traffic for each base station, and the 80th
percentile of aggregated traffic, the capacities need to be summed considering how many
cells will be aggregated on a link/BBU pool. The sum will give the desired link/BBU
pool resources.
The results for the sum of the 95th percentiles can be applied to the equation (3.1),
where sufficient AggregatedLinkResources are 200 Mbps, based on delay measure-
ments. Using this method computed M G95th is in the range of 1.27 - 2.66 which con-
verges with the results for M G for throughput measured every 10 ms (1.5-2.2). Moreover,
2.7 is the result of M GBBU BBRF concluded with the teletraffic method published in
[13], which confirms the thesis that M GF HU ECell = M GBBU BBRF stated at the
beginning of this chapter. These results serve as validation of the two approaches.
Projects 1, 3 and 5 take into account different application definitions. Still the results
remain similar (3 - 6).
Figure 3.15: Multiplexing gains for different locations based on daily traffic distributions
between office and residential cells. Data from China Mobile and Ericsson.
Project 5 explores the most optimal cell mix from project 4. Similar values (1.5 -
1.7) can be observed for project 4, with 57.6 s and 100 s averaging. For longer averag-
ing in project 5, the multiplexing gain value 1.6 matches the value 1.6 from project 4,
despite different application definitions and mixes. Therefore it can be concluded that
multiplexing gain coming from the tidal effect equals to 1.0 - 1.33. The value is rather
small, therefore enabling multiplexing gains should not be treated as the main reason to
introduce C-RAN, rather an additional benefit on top of advantages mentioned in Section
2.2. However, there are cells with occasional traffic load, like stadiums, that will greatly
benefit from multiplexing gains. Therefore C-RAN is beneficial for such deployments on
a local scale, smaller than metropolitan scale.
78 Chapter 3. Multiplexing gains in Cloud RAN
3.5.6.2 Methodology
In modeling work the traffic aggregation properties were observed. In fact, they refer
to the multiplexing gain on transport resources. In order to truly evaluate pooling gain
actual computational resources will need to be modeled. Each of the LTE protocol stack
layers requires different amount of GOPS, that will scale differently depending on the
user activity. Moreover, in C-RAN an additional functionality will need to be in place, e.g.
orchestrator, to enable multiplexing gains on BBUs. Full LTE BBU pool implementation
will enable a more detailed evaluation of those resources. Lastly, modeling of actual
power consumption will give a better insight on possible power savings.
80 Chapter 3. Multiplexing gains in Cloud RAN
3.5.7 Conclusions
Based on the analyzed data set, in a typical, real-life mobile network, with a mix of
constant bit rate and bursty traffic, multiplexing gain on traffic-dependent resources in
range of 3 - 6 can be expected.
The contribution to the multiplexing gain that C-RAN as such enables on a metropoli-
tan scale is between 1.0 and 1.33, depending on the cell mix. Such a multiplexing gain
can be obtained on a BBU pool for all functional splits. To enable higher gains, cells with
occasional traffic, like stadiums, should be included in the BBU pool.
However, for the functional splits above UE-Cell, where fronthaul carries variable bit
rate data, multiplexing gain 3 - 6 can be achieved on the fronthaul links. This is a strong
motivation to investigate and deploy other functional splits than the traditional one, where
all the baseband processing is done in the pool.
3.7 Summary
C-RAN is seen as a candidate architecture for 5G mobile networks because of the cost and
performance benefits it offers. Due to the fact that securing fronthaul capacity to deploy
3.7. Summary 81
fully centralized C-RAN can be costly, a careful analysis of cost and energy savings for
different functional splits is of great importance.
Multiplexing gain observed from traffic aggregation approximates possible gains on
aggregation transport links. Such a multiplexing gain is proportional to power savings
and processing resources savings only on parts of the BBU pool responsible for user-
processing functionalities of data-plane processing. However, as exact power and cost
models are complex in the analyzed scenarios, the results present only an approximation.
Gains in power consumption and pooling gains will be lower than the multiplexing gain.
Pooling gains on processing resources can be achieved only on user-processing
resources, which are a fraction of overall signal processing resources - 3-12% on downlink,
17-33% on uplink.
Concerning power savings, 2-24% of total base station power consumption is spend
on baseband signal processing. Multiplexing gain can be achieved on user-processing
modules, which constitute a fraction of these resources - 3-12% on downlink, 17-33% on
uplink.
As only a fraction of resources are impacted by pooling gains, those gains should not
be a priority in designing new functional splits. However, a variable bit rate functional
split reduces needed bandwidth and enables multiplexing gains on the fronthaul network,
which is an important motivation.
Gains in transport are the closest to derive from multiplexing gain value, however, ac-
tual traffic patterns in fronthaul network for UC-Cell split will be different than simulated
when the LTE protocol stack is included.
Such a lack of the full LTE protocol stack is a major simplification in the modeling
work. Others include: methodology based on traffic throughput measurements, not small
enough time scale of measurements, and lack of control plane considerations. A full
LTE protocol stack BBU pool implementation will enable more precise measurements on
pooling gains on baseband resources as well as multiplexing gain on fronthaul link for
variable bit rate splits. As energy and cost savings are related to the multiplexing gain, in
this study the multiplexing gain is evaluated for different functional splits. Multiplexing
gains on throughput-dependent functionalities of a base station are calculated for different
C-RAN functional splits: BB-RF and separating user and cell specific functions using four
different approaches. For given traffic definitions, a quantitative analysis of multiplexing
gains is given. However, the results are sensitive to the traffic profiles as well as to the
multiplexing gain calculation method. Therefore the main outcome of this study is to
provide the trend lines that will facilitate finding an optimal trade off when fronthaul or
BBU resources are more costly for an operator.
For fully centralized C-RAN - with BB-RF split - maximum multiplexing gain on
BBU resources can be achieved. However, the required fronthaul capacity is the highest.
Therefore this split is vital for operators with access to a cheap fronthaul network. Addi-
tionally, if the traffic load is high, the operator will mostly benefit from the multiplexing
gain at the BBU pool.
82 Chapter 3. Multiplexing gains in Cloud RAN
The more functionality is moved from the BBU pool to the cell site, the lower the
multiplexing gain on the BBU pool. However, when traffic starts to be variable bit rate, a
multiplexing gain on the fronthaul links can be achieved, lowering the required capacity.
Hence, for low traffic load, and even more for bursty traffic, the BBU pool should only
have higher layer processing and then the requirements on the fronthaul link can be
relaxed.
An initial evaluation concludes that up to 1.3 times fewer BBUs are needed for
user data processing in C-RAN compared to a traditional RAN looking at daily traffic
distributions between office, residential and commercial areas. This number grows to 4
when taking into account specific traffic patterns, making assumptions about the number
of base stations serving different types of areas. The latter model does not include
mobile standard protocols processing. After including protocols processing the statistical
multiplexing gain varied between 1.2 and 1.6 depending on traffic mix, reaching the peak
for 30% of office and thereby 70% of residential base stations, enabling savings of 17% -
38%.
The application level delays are verified for different aggregated link bit rates and
thereby it is concluded what is the practical value of the multiplexing gain that can be
achieved. Rules of thumb for network/BBU dimensioning are proposed based on the
CDFs of throughput. The more video traffic is present in the network, the delays are more
sensitive to the aggregated link bandwidth, what influences achievable multiplexing gain.
A high impact on the multiplexing gain value is observed depending on the multi-
plexing gain calculation method, i.e. using different throughput averaging windows. The
results vary between 1.0 and 6. The multiplexing gain that C-RAN in traditional, BB-RF
split solely enables due to the tidal effect is between 1.0 and 1.33 depending on the cell
mix, thereby enabling savings of up to 25%. The major contribution to the numbers higher
than 3 comes from traffic burstiness, and in order to benefit from it on the fronthaul links,
functional splits that result in variable bit rate on the fronthaul need to be in place.
CHAPTER 4
Towards packet-based fronthaul
networks
While the connection between RF and BB parts in a traditional base station was just an
interface in one device, in a base station with RRH it often spanned few meters between
RRH and BBU in a point-to-point connection, up to the distance from the basement to
the rooftop. For C-RAN it is beneficial to connect sites from a metropolitan area, which
requires a whole network to support those connections. This network is called fronthaul
and spans between cell sites, traditionally equipped with RRH up to the centralized
processing location, traditionally equipped with BBU pool.
High capacity requirements on fronthaul are seen as one of the major deal-breakers
for introducing C-RAN. Therefore, a thorough analysis of fronthaul requirements as well
as solutions to optimally transport data are of high importance, which is reflected in
numerous standardization activities, referenced throughout the chapter.
As introduced in Section 2.4, several solutions can be used to organize transport in
the fronthaul network, such as: point to point, WDM, WDM over PON, microwave often
combined with compression, OTN and Ethernet. This chapter elaborates on the fronthaul
requirements (Section 4.1) as well as presents a proof of concept of CPRI/OBSAI over
OTN transport (Section 4.2). Moreover, it motivates using Ethernet for fronthaul (Section
4.3) as well as explores challenges and possible solutions to deliver synchronization
(Section 4.5 and 4.6) and fulfill delay requirements for a packet-based fronthaul network
(Section 4.7 and 4.8). Furthermore, it presents a demonstration of Ethernet fronthaul built
in this project (Section 4.9) as well as future directions (Section 4.10). Finally, Section
4.11 summarizes relevant standardization activities, while Section 4.12 summarizes the
chapter.
The material presented in Sections 4.3-4.6 was originally published in [15]. It was
expended and updated for this dissertation.
4.1 Requirements
This section elaborates on throughput, EVM, delay, jitter and synchronization require-
ments that fronthaul network needs to support in order for a mobile network to work
according to 3GPP specifications.
83
84 Chapter 4. Towards packet-based fronthaul networks
4.1.1 Throughput
As introduced in Section 2.3.1 a popular LTE cell configuration (2x2 MIMO, 20MHz) will
require 2.5 Gbps capacity on the fronthaul link, provided that IQ samples are transmitted
using CPRI. For higher bandwidth and more antennas this requirement will scale almost
linearly 1 . With the UE-Cell split this throughput can be reduced to 498 Mbps, 700 Mbps
[46] or 933 Mbps [148], depending whether 7, 10 or 16 bit sample width is used. [46]
assumes 7 bits for DL, 10 bits for UL, while [148] reserves 16 bits. With MAC-PHY
bandwidth can be further reduced to 152 Mbps for DL [148].
4.1.2 EVM
Fronthaul network may introduce errors in data transmission. EVM defines how much the
received signal is distorted from the ideal constellation points. No matter the fronthaul
architecture, general requirements for LTE-A need to be observed for different modulation
schemes on the main data bearing channel (Physical Downlink Shared Channel (PDSCH)):
Quadrature Phase Shift Keying (QPSK), 16 QAM and 64 QAM, as listed in Table 4.1.
4.1.3 Delay
A HARQ is a process that poses the most stringent delay requirement for LTE-A. As
a retransmission mechanism, it takes part in error control and correction. According
to the LTE-A standard, for FDD the HARQ RTT Timer is set to 8 subframes (8ms)
[170], which means that the user using subframe n needs to know whether retransmission
or transmission of new data should occur at subframe n + 8, as illustrated in Figure
4.1. Due to the timing advance, the user sends data ahead of time compensating for the
propagation delay, in order to fit into the subframe structure at the base station. It appears
to be an industry standard that a base station needs to prepare a HARQ acknowledgment
(ACK)/non-acknowledgement (NACK) within 3ms [20], [171], [172]: decode UL data,
prepare ACK/NACK and create a DL frame with ACK/NACK. Only then will the user
receive a ACK/NACK in the 4th subframe after sending the data, 3ms processing time at
UE can be accommodated, and a possible retransmission can occur during 8th subframe.
If a user will not get ACK/NACK it will retransmit the data in 8th subframe. This 3ms
delay budget is spent on BBU and RRH processing as well as UL and DL fronthaul delay,
1 For CPRI line bit rate option 7A (8110.08 Mbps) and above a more optimal 64B/66B line coding can be
leaving 100-200 s [20]/220 s [46]/ 250 s [148], [173] for fronthaul one way delay or
in other words 200-500 s Round Trip Time (RTT). Otherwise throughput will be lowered
[148]. Section 4.10 discusses possible future directions for the HARQ requirement.
UE RRH BBU
Rx Tx Rx Tx
n subframe
1 subframe
1ms
PropDel Data
Data FHULDel
BS BBU
processing processing
3ms 2.5ms
UE
processing
3ms
n+8
1 st Rtx/Data
Figure 4.1: Delays associated with the UL HARQ process. PropDel - propagation delay, Rtx
- retransmission, FHULDel - Fronthaul UL Delay
Looking at Metro Ethernet Forum Classes of Service for mobile backhaul, Class
"High" backhaul could be reused for fronthaul (frame delay 6 1ms) [174] or a new class
could be specified looking at lower delays, in the order of 100 s.
4.1.4 Jitter
Nowadays BBUs and RRHs read the delay at the bootup time, therefore the delay needs
to be constant. This requirement can be relaxed by buffering the data, however, this will
be done at the cost of a higher delay.
86 Chapter 4. Towards packet-based fronthaul networks
4.1.5 Synchronization
A proper synchronization is essential for mobile network operation. In order for a RRH
to modulate the data to a particular frequency it needs to know the precise definition of
1 Hz. It is important to keep the carrier frequency sharp in order for the signal coming
from the base stations operating in a different frequency band not to overlap and for the
UE to be able to receive it. For successful TDD network operation, the RRH needs to
follow time frames precisely in order for DL and UL frames not to overlap. Two types
of synchronization can be differentiated: frequency and phase (time) synchronization.
Clocks are synchronized in frequency if the time between two rising edges of the clock
matches. For phase synchronization, rising edges must happen in the same time, as shown
in Figure 4.2.
In Table 4.2 the requirements that need to be observed for various network features,
like Time-Division Duplex (TDD), Frequency-Division Duplex (FDD), MIMO, Carrier
Aggregation (CA), eICIC and CoMP are summarized. For latter three, the requirements are
expressed relatively to a common reference between the cells/streams involved, otherwise
the maximum deviation from an ideal source is listed. Moreover, for Enhanced 9-1-1
services, FCC requires the localization accuracy to be within 50 meters [175], which
requires synchronization in the 100 ns range [176].
- OTN is a promising solution for optical transport network of C-RAN when existing
OTN legacy network can be reused for C-RAN fronthaul connecting RRHs to the
BBU Pool.
- OTN provides FEC allowing the transport of client signals like CPRI in noisy
environments or over longer distances.
Partnership Project (3GPP) specifications for LTE-Advanced. It was a very useful exercise
towards understanding frequency error and EVM. More details about those measurements
can be found in Appendix A.
CPRI / CPRI /
OBSAI OBSAI
(a) Traditional C-RAN architecture (b) C-RAN with packet based fron-
with a dedicated line to each RRH. thaul.
CPRI
Variable delay BBU Pool++
BBU
RRH RRH++
Figure 4.3: Traditional and discussed C-RAN architecture together with principles of deriving
synchronization for them
The use cases of IEEE 1904.3 cover both legacy RRH and BBU as well as Ethernet-
based RRH and BBU. The first version of the standard is planned for May 2017.
IEEE 1904.3 will be referred to as 1904.3 later on in the text.
CCSA founded a project to study the requirements, scenarios and the key technolo-
gies for next-generation fronthaul [180].
3ms Round Trip Time latency allowed, 200-500us RTT left for fronthaul link
Switch
RRH CPRI2Eth BBU
Note: switch may not Pool
GW support 1588, leading to
Radio, LTE-A TDD CPRI
bigger timing recovery
errors
LTE LTE CPRI CPRI IQ IQ
Figure 4.4: Model of the requirements and factor introducing uncertainties in LTE, CPRI,
1588 and Ethernet layers.
CPRI2Eth gateways or the RRHs++ with GPS. This solution assures both frequency and
phase delivery. However, it increases the cost, spent not only on GPS equipment but also
on an oven-controlled oscillator. Moreover, the coverage indoors and in metropolitan
valleys (small cell on a lamp post between high buildings) will be problematic. For some
operators it is also important not to depend on a third-party solution for their network.
Another solution could be to implement a 1588 slave in the CPRI2Eth gateways or in
the RRH++. This solution assures lower equipment cost, however, it will be affected by
variable network delay present in Ethernet networks. Ashwood [186] shows that such a
jitter, when the background traffic is present, can be in the order of s per Ethernet switch.
The architecture considered is presented in Figure 4.4. Ethernet packets are sent
from the BBU through the network of switches to reach the CPRI2Eth gateway. They
are repacked there to CPRI stream to reach legacy RRHs. Alternatively, Ethernet RRHs
can be used, omitting CPRI. A packet-based solution - IEEE 1588 - is used to assure
synchronization with a 1588 master present in BBU Pool and a 1588 slave present in
CPRI2Eth gateway. Figure 4.4 summarizes requirements on different layers: LTE, CPRI
and Ethernet as well as factors introducing inaccuracies in 1588 and Ethernet layers.
In the section below a feasibility study of using 1588 for timing delivery is performed.
The modeling work takes into account the factors influencing the performance of 1588
92 Chapter 4. Towards packet-based fronthaul networks
mentioned in Figure 4.4. A dedicated Ethernet network was taken into account, leaving
for future work the case of sharing Ethernet infrastructure with other types of traffic.
It is worth noticing, that the requirements need to be fulfilled on two levels:
- delay requirements,
where the delay of synchronization packets lies between them, as 1588 helps to recover
the clock, but is affected by network delays, as emphasized in Figure 4.5.
Delay
Delay of 1588
packets
Clock
recovery
message leaves the node - timestamp t3 . The master node notes the time when it receives
the message - t4 and sends it to the slave via the DelayResp message. Based on those
timestamps, the delay and offset between the clocks can be computed as shown in the
Equation (4.2) and (4.3), respectively. It can be observed that the messaging of the time
stamps from the 1588 master to the slave is a feed forward messaging algorithm. More
information on the 1588 operation can be found in [187].
BBU
Pool
CPRI2Eth Eth
gateway switches
Applica
tion
IP
Relay
PDCP
PDCP GTP-U
UDP
L2
RLC RLC
IP
MAC MAC
L2
Eth Eth Eth
Radio Radio
1588
1588
1588
1588
Eth switch
1588
Eth
MAC
Eth
PHY
1588
messages Ingress Egress
timestamp timestamp
Figure 4.8: Ingress and egress timestamps should be taken as soon as Sync or DelayReq
packets enter and leave the node, respectively.
4.6. Feasibility study IEEE 1588v2 for assuring synchronization 95
The 1588 standard defines the establishment of a clock hierarchy and format of
messages, so that the equipment from different vendors can communicate in the network.
The standard gives an example on how to calculate the clock drift as presented in the
Equation (4.4). However, the implementation of an actual synchronization algorithm
lies outside of its scope and this can possibly give a competitive advantage to certain
vendors over others in performance of their solutions. Various works have been published
presenting synchronization algorithms. Xie et al. in [188] propose to maintain two time
scales in the slave: syntonous (frequency aligned) and synchronous (time and frequency
aligned). For delay calculations they use t2 and t3 measured in a syntonous time scale.
They concluded that it is the most optimal to apply drift correction every third time the
exchange of 1588 messages is completed. This way of drift calculation was implemented,
as in Equation (4.5).
Drif tStd =
t2(N ) t2(0)
= (4.4)
(t1 + Delay + CFS )N (t1 + Delay + CFS )(0)
Drif tImplemented =
t2(N ) t2(N 3)
= (4.5)
(t1 + Delay + CFS )N (t1 + Delay + CFS )(N 3)
A network model in OPNET modeler was built checking the performance of this
algorithm. OPNET is an event-driven simulation software, where a user can build his
scenario from self-defined nodes and processes. A network consisting of a 1588 Master,
a 1588 Slave and a variable number of Ethernet switches working as 1588 transparent
clocks was built. Sync and DelayReq packet rate is 64 packets per second (pps). The
novelty of this work is a network view where the protocol was tested against various
errors that can occur in the network. The slave node has an initial frequency drift of 1
ppm or 100 ppm (maximum that an Ethernet switch can have). Each of the Ethernet
switches has a frequency error of 1 ppm or 100 ppm and timestamping error of 1 ns or 4
ns. As it is not possible to measure the exact arrival time of a packet using the internal
clock reference in the switch, a random timestamping error is introduced up to 1 ns and 4
ns, respectively. The values mentioned above represent parameters of newer and older
generation equipment from the industry. The models follow the protocol stack from
Figure 4.7 on the link between the BBU and the CPRI2Eth gateways. 2.5 Gbps CPRI
traffic was sent over a 10 Gbps Ethernet network. In between master and slave node 0-21
Ethernet switches were put. The simulations were run for 10 minutes, while CPRI traffic
was sent after 30 s. 30 seconds were required to get the network operational with Ethernet
switching topology getting established.
For each exchange of 1588 messages, after all timestamps are gathered, delay, drift
and offset are computed. Drift correction is applied to slave clock frequency fS every
96 Chapter 4. Towards packet-based fronthaul networks
third exchange of timestamps. That affects both synchronous and syntonous time scales
at slave having impact on t2(Syntonous) and t3(Syntonous) . The synchronous time scale
of the slave is updated for the offset after each exchange of timestamps. That affects local
time at the slave tS . A relative frequency error between master clock frequency fM and
slave clock frequency fS was measured, as presented in Equation (4.6) and an absolute
phase error between time in master tM and time at slave tS as presented in Equation (4.7).
fM fS
F requencyError = (4.6)
fM
P haseError = tM tS (4.7)
Phase error
for different # of Eth switches (x) and different
timestamping errors (2 series)
40.0
35.0
Maximum error (ns)
30.0
25.0
1 ns
20.0
4 ns
15.0
10.0
5.0
0.0
0 5 10 15 20 25
#Switches
Figure 4.9: Maximum phase error measured for various scenarios during stable operation.
Figure 4.9 presents the maximum observed phase (time) error during stable operation
(after initial time discovery) for different numbers of the Ethernet switches present in the
network and two different timestamping error values. The phase error stays in the order
of nanoseconds and is highly dependent on timestamping errors. The dependency is close
to linear. The results are shown for the worst-case scenario of 100 ppm drift for both the
Ethernet switches and slave, as drift value had marginal effect whether it was 1 ppm or
100 ppm in both cases.
Figure 4.10 shows the frequency error for the afore mentioned scenarios. The fre-
quency error falls way above required values (16 ppm or below, depending on implemen-
tation). It is also highly dependent on timestamping errors. That is the reason why the
improvements to this method were applied.
Frequency error
for different # of Eth switches (x) and different
timestamping errors (2 series)
800.0
700.0
Maximum error (ppb)
600.0
1 ns
500.0
400.0 4 ns
300.0
200.0
100.0
0.0
0 5 10 15 20
#Switches
Figure 4.10: Maximum frequency error measured for various scenarios during stable operation.
the offset. The frequency is adjusted by averaged drift computed taking only a fractional
value of currently computed drift - drif t - and previously computed drift - drif tP rev.
The performance of the system for different values of alpha was checked, and it was
concluded that the higher it gets, the lower frequency error is observed. However, for
higher values of alpha it takes more time for the system to converge to stable operation.
For simulations alpha = 0.99 was used. After 180 s the system reached stable operation.
The frequency error got significantly smaller for all the cases (20 times smaller), while
the phase error got slightly smaller (2 times) as presented in Figures 4.11 and 4.12.
Phase error
for different # of Eth switches (x) and different
timestamping errors (2 series)
when drift averaging was applied
18.0
16.0
Maximum error (ns)
14.0
12.0
1 ns
10.0
8.0 4 ns
6.0
4.0
2.0
0.0
0 5 10 15 20 25
#Switches
Figure 4.11: Maximum phase error observed during stable operation for various scenarios
with offset averaging applied.
98 Chapter 4. Towards packet-based fronthaul networks
Frequency error
for different # of Eth switches (x) and different
timestamping errors (2 series)
when drift averaging was applied
35.0
Maximum error (ppb)
30.0
25.0 1 ns
20.0
4 ns
15.0
10.0
5.0
0.0
0 5 10 15 20
#Switches
Figure 4.12: Maximum frequency error observed during stable operation for various scenarios
with drift averaging applied.
IQ Baseband/CPRI
Ethernet2CPRI gateway
combined with RRH
LTE-A
requirements
Figure 4.13: Clock recovery scheme inside an RRH combined with CPRI2Eth gateway. LO -
local oscillator
Table 4.3: Delays in an Ethernet switch
Maximum Transmission Delay Switching
10 Gbps link, s Delay, s
Number of MTU, B E.g.
switches 1500 3000 6000 9000
1 1.2 2.4 4.8 7.2 3.0
2 2.4 4.8 9.6 14.4 6.0
5 6.0 12.0 24.0 36.0 15.0
10 12.0 24.0 48.0 72.0 30.0
3ms Round Trip Time latency allowed, 200-500us RTT left for fronthaul link
Switch
RRH CPRI2Eth BBU
GW Pool
010101
011110
101001
001101
001010
size. It assumes no queuing delay. Figure 4.16 takes the best case (cut through switches,
MTU 1500) and evaluates it for different queuing delay per switch.
If a dedicated Ethernet link is used for a fronthaul link, there will not be any queuing
delay. However, for multiplexing gains on links it is desired that many fronthaul streams
will share the link, possibly also with other types of traffic. Even if various fronthaul
streams will be given the highest priority, it needs to be assured they will not collide.
Moreover, the lower priority packets should not slow down fronthaul packets if they
happen to be under processing by a switch. The following two methods are aiming to
address these problems and are currently under the standardization in IEEE:
40
CT MTU 6000
30
CT MTU 9000
20
SAF MTU 1500
10 SAF MTU 3000
0 SAF MTU 6000
0 5 10 15 20
SAF MTU 9000
Number of switches
Figure 4.15: Allowed distance between RRH and BBU for a total delay budget of 250 s
depending on the number of switches in the network, MTU size and type of switch (SAF -
Store and forward, CT - Cut through). Assumed queuing delay is zero.
50
Distance, km
40 0 s
30 20 s
20 40 s
10 60 s
80 s
0
0 1 2 3 4 5 6 7 8
Number of switches
Figure 4.16: Allowed distance between RRH and BBU for a total delay budget of 250 s
depending on the number of switches in the network, MTU size and queuing delay per switch.
CT switch, MTU 1500.
102 Chapter 4. Towards packet-based fronthaul networks
Both solutions belong to the TSN set of standards, which can be applied to any
time sensitive application. In April 2015 the Project Authorization Request (PAR) of a
new standard was prepared - 802.1CM Time-Sensitive Networking for Fronthaul that
aims at defining profiles that select features, options, configurations, defaults, protocols
and procedures of bridges, stations and LANs for the fronthaul applications [191]. The
sections below describe the source scheduling and preemption.
Case 1: one fronthaul stream (protected traffic), one or more other traffic streams
(unprotected traffic) One approach to achieve that can be to assure that only one
stream has access to the network at specific times (protected window, from T1 to T2 in
Figure 4.17), in other words, the transmission for protected traffic is scheduled between
time T1 and T2. However, in order to make sure that the unprotected traffic is not
under transmission while the protected one arrives, a certain guard band needs to be
in place (T0-T1). The simplest solution would be to have a guard band equal to the
transmission time of the largest packet size supported by the network. That however, leads
to resource inefficiency, as the network could be unnecessarily idle. More optimally, the
implementation could check if there are any packets in the queues, whose transmission
could end before the transmission of protected traffic. Unprotected traffic will be sent on
the best effort basis.
Start of a Start of a End of a
guard band protected protected
(T0) window (T1) window (T2)
B A A B
B Unprotected traffic
Case 2: many fronthaul streams In this case it is not sufficient to schedule a protected
window when the fronthaul stream requests it, as many streams may collide, creating
variable (and non-deterministic) delay due to queuing in the switches. In the fronthaul
application it is important that the delay would be as stable as possible (low jitter) while
maintaining the delay within the requirements. Traffic scheduling can be implemented
already in the sources and it is illustrated in Figure 4.18. Packets A and B initially would
arrive at the switch at the same time and one of them will need to wait in the queue,
therefore experiencing a non-deterministic delay. When one of the packets here B - is
initially delayed its delay would be larger, but deterministic-predictable. It is especially
important when packets are going through many switches and a non-deterministic delay
would create a big jitter.
A A
B A B A
B No scheduling B delay non BBU pool B Scheduling B delayed BBU pool
deterministic initially, delay deterministic
Figure 4.18: Source scheduling used to reduce jitter. Here an example for UL
Case 3: many fronthaul streams and one or more other traffic streams (unprotected
traffic) Traffic scheduling as described in the section above can be used between many
fronthaul streams. Unprotected traffic will be sent on the best effort basis.
Using gate opening to create protected windows The 802.1Qbv standard allows the
implementation of traffic scheduling by means of a gate opening. A sequence of gate
open/close operations can be scheduled for each port and traffic class, allowing each
traffic class to be sent at a given time but be blocked at others. The schedule can be
changed periodically. A frame of a given class can be transmitted only if the gate is open
and it will remain open long enough for the packet to be transmitted.
preempted packets are then merged. The technical details of preemption are described
in 802.3br Draft Standard for Ethernet Amendment: Specification and Management
Parameters for Interspersing Express Traffic [192].
This approach is especially useful when a fronthaul stream will share the network with
other applications. It can optimize guard band size for scheduled traffic, as bigger parts of
the packets can fit before gate closing. The smallest size of a preempted packet needs to be
not smaller than 64 B, and thereby packets up to 124 B can be preempted. Alternatively,
only with source scheduling the whole packets will need to wait for transmission, even if
the switch would be free for some time. Therefore the capacity would be wasted. With
preemption it is safe to allow a portion of the packet to be sent.
4.7.4 Discussion
As can be seen from the delay analysis, it is possible to enable Ethernet-based fronthaul
without TSN functionalities for dedicated links. However, as soon as more than one
fronthaul stream will use the link, it is important to observe potential queuing delays.
Source scheduling is a beneficial technique that requires intelligence only at the network
edges (CPRI2Eth gateway or Ethernet RRH), while the legacy switches can be used. If
other services are supposed to share the network with the fronthaul streams, then the delay
may get even bigger. Preemption is a technique that assures timely delivery of higher
priority packets at the same time allowing other traffic to use the link whenever it is not
colliding with fronthaul packets.
If only preemption is enabled, than the delay can be kept deterministic for only one
stream. For more streams algorithms minimizing the overall delay on the network level
would need to be in place.
For fast networks (10+ Gbps) with low load TSN features are less significant.
Wan et al. in [193] presents discrete-event based simulation results on transmitting
CPRI over Ethernet. The authors measured that for a tree topology network where each
traffic stream went through 2-4 switches the delay was about 90 s and the jitter was up to
400 s. When background traffic was inserted, the Ethernet with preemption performed
similarly - delay was 91 s and the jitter was 410 s. However, with the scheduling
algorithm the authors proposed the performance was not consistent - in the majority of the
4.8. Source scheduling design 105
cases the jitter was removed, but in some grew up to 1000 s. The authors recommend to
use Ethernet with preemption and buffering at the edge.
Farkas et al. in [194] present other simulation results of CPRI over Ethernet transport
with TSN features. Without the TSN features in a tree topology comprised of 10 Gbps
links, where each stream went through 1-3 switches, the switching delay was 1500 ns
5 ns variation. For the scenario with background traffic, the jitter raised up to 3 s, but
when the preemption was enabled jitter stabilized in 100 ns level. With source scheduling
the jitter amounted to up to 50 ns. When source scheduling with 70 ns guard band was
added on top of preemption the jitter was reduced to 0 ns. The delay for all the cases
was below 26 s. Therefore it was shown in simulations that usage TSN standards and a
proper network configuration the delay and jitter requirements for fronthaul can be met.
1. 4G - CPRI traffic packetized into Ethernet frames. Many streams could share one
link. The bit rate per each stream would be 2 Gbps instead of the original 2.5 Gbps
when 8B/10B line coding is removed (Ethernet already implements error detecting
coding by means of Frame check sequence (FCS)),
2. 5G - Fronthaul data from one of the new functional splits sent in the Ethernet
packets. The bit rate per each stream would be between 150 Mbps and 2 Gbps,
instead of a 2.5 Gbps CPRI. Optionally, for small cells the CPRI split can be used,
as bit rates would be lower for small cells.
The BBU schedules the DL traffic to all the RRHs. The packets leave the output port
of the BBU pool in order, which can be compared to virtual timeslots. For each packet
received in DL, the RRH will send one UL packet. Therefore the UL packets will not
collide if they are sent with correct timing advance fitting the virtual timeslots prepared for
106 Chapter 4. Towards packet-based fronthaul networks
DL. At the RRH bootup, control packets will need to be exchanged to measure the delay,
in a similar way in which timing advance is implemented in GSM. Figure 4.20 describes
the initial DL scheduling (left) and UL and DL packets fitting the virtual timeslots (right).
2 Gbps traffic DL
A 4 Gbps traffic
ADL DL
AUL
2 Gbps traffic A B ADL BDL
B BDL BUL AUL
10 Gbps link BUL UL
BBU
Switch BBU Switch
RRHs RRHs
A2 B1 B2 A1 time
rcx rcx sent rcx
In case more BBU pools are present in the network, a more generic solution will be
needed. The methods used for wavelength assignment in wavelength routed networks
[195] can be exploited. In case the network will be shared by other, non-fronthaul services
it is recommended to add the preemption functionality.
1. Stream setup
Overhead
40%
HeadersSize/PayloadSize 35%
30%
25%
20%
15%
10%
5%
0%
0 2000 4000 6000 8000 10000
Payload size, B
Figure 4.21: Ethernet L1 and L2 as well as 1904.3 overhead comparing to Ethernet frame
payload size
b) Upon packet arrival RRH notes the time (T2), sends it back to BBU
Low priority,
RoE ctrl, subtype=Access_2
c) BBU computes delay Delay=T2-T1 informs RRH
Low priority,
RoE ctrl, subtype=Access_3
d) RRH acknowledges
Low priority,
RoE ctrl, subtype=Access_4
2. Data transmission
a) BBU schedules traffic to the cell, sends it
RoE pkt_type =data(000001b - 000100b or 100100b),
b) After receiving at least two packets m and m+1 at times Rcx(m) and Rcx(m+1),
respectively, RRH measures receiving interval RcxInt = Rcx(m + 1)
Rcx(m)
c) Cell receives the traffic in time DLrcx, cell can send back the traffic in the
earliest time in the future such that time DLrcx 2 Delay + n RcxInt
RoE data
For the dependency between required RRH-BBU distance and allowed number of
switches, the delay budget analysis presented in Section 4.7.1 applies.
108 Chapter 4. Towards packet-based fronthaul networks
10 G Eth
switches
Loop connection
A for DPDK B
Data, 10 G cables
Control, 1 G cables
Controller
1588 1588
Slave
Eth Master
RRH
BBU
Pool
T imeInP Cs is the one needed for the kernel to process packets in each of the PCs.
In order to calculate the switching delay (SwitchDel), the values where the minimum
RTT crosses the y axes were noted down, corresponding to the delay where transmission
delay depending on packet size (T ransDel(B)) is zero and queuing delay is also zero -
e.g. P ingRT T 1switch(0) for one switch. That leads to:
By subtracting the P ingRT T N switch(0) values (and dividing the difference by 2),
as shown in Table 4.5, an average switch delay is 2.3 s.
250
200
RTT, us
Min
150
y = 0.0075x + 53.472 Avg
100
Max
50 Linear (Min)
0
0 2000 4000 6000 8000 10000
Ping size, B
(a)
700 700
600 600
Min Min,
Avg Avg,
500 500
Max Max,
Linear (Min)
Linear (Min,)
RTT, us
RTT, us
400 400
300 300
200
200
y = 0.0106x + 62.463
y = 0.0089x + 58.19
100
100
0
0 0 2000 4000 6000 8000 10000
0 2000 4000 6000 8000 10000
Ping size, B
Ping size, B
(b) (c)
BBU
Base
RRH Base
Base
Base
pool
band
band
band
band
S1
RRH X2
RRH MME
EPC
S1 PGW
BBU
Base
RRH Base
Base
Base
pool
band
band
band
band SGW
RRH RRH
BBU BBU
Midhaul
- IEEE 1904.3 Standard for Radio Over Ethernet Encapsulations and Mappings
that aims to define encapsulation of fronthaul data, independent on the functional
split, to Ethernet frames.
- IEEE 1588 Standard for a Precision Clock Synchronization Protocol for Net-
worked Measurement and Control Systems v3 is under preparation which aims,
among others to improve accuracy to 100s ps and enhance security. The expected
completion date of v3 of the standard is 31 December 2017 [7].
IEEE 802.1Qbv, IEEE 802.1Qbu and IEEE 802.1CM are part of TSN Task Group.
4.12. Summary 113
4.12 Summary
In order to lower the costs and improve the flexibility of C-RAN deployments, existing
packet-based, e.g. Ethernet, networks can be reused for fronthaul.
Sections above analyze requirements on bit rates, delay, phase and frequency synchro-
nization for fronthaul with current and new functional splits. Architecture sufficient for
fulfilling them was derived, namely: CPRI/OBSAI over OTN and CPRI/new functional
split data over Ethernet. It has been proved that OTN can support existing deployments
with CPRI/OBSAI. Factors that are challenging for achieving synchronization in packet-
based C-RAN fronthaul were analyzed. A feasibility study was presented showing the
performance for frequency and phase synchronization using 1588 in Ethernet networks
under various inaccuracies that can be present in the network. Apart from possible queu-
ing delays, the one that has the highest impact is the timestamping error associated with
the way timestamps are generated in Ethernet switches. Whether this performance will
meet the requirements of future mobile network depends on PLL and local oscillator
implementation, based on 1588 feedback to clock offset and drift. Moreover, an Ethernet
network that is ready to be integrated in the fronthaul part of a C-RAN demo and which
fulfills the delay requirements was built. The delay measurements were performed, which
allowed to obtain a better understanding of the delays encountered in an Ethernet network.
To address the delay requirements in shared networks source scheduling and preemption
are investigated. A source scheduling algorithm is proposed, which is optimized for
symmetrical fronthaul traffic, however can also be applied in cases when downlink traffic
exceeds uplink traffic.
Figure 4.26 presents a protocol stack enabling various functionalities for fronthaul
over Ethernet transport. Devices at the cell site can either be legacy RRHs running on
CPRI and connected to a CPRI2Eth gateway or native Ethernet RRHs. 1588 assures
synchronization between a master clock located in a BBU pool and a slave clock located
at the cell site. It is beneficial to allow on path support of 1588 for better compensation
of queuing related delays, however, in principle it is not mandatory. It is recommended
to use TCs for a better network delay compensation. If an Ethernet network consists
of unknown types of switches, non-TSN enabled, it is recommended to enable source
scheduling at the network edges to control the delay. Moreover, if it is desired to share the
network with other services, preemption is a recommended feature for Ethernet switches.
114 Chapter 4. Towards packet-based fronthaul networks
CPRI2Eth
Devices Gateway,
1588 Slave BBU
Pool
EthRRH,
Eth
1588 slave
switches
Applica
tion
IP
Relay
PDCP
Data plane PDCP GTP-U
UDP
L2
RLC RLC
IP
MAC MAC
L2
Radio Radio 1904.3
L1
Figure 4.26: Proposed architecture for Fronthaul over Ethernet. Dashed lines highlight
optional functionality.
CHAPTER 5
Conclusions and outlook
In order to satisfy the ever growing need for capacity in mobile networks, and, at the same
time, to create cost and energy efficient solutions, new and disruptive ideas are needed.
With its performance gains and cost benefits, C-RAN proves to be a major technological
foundation for 5G mobile networks. By applying the concept of NFV, C-RAN mobile
networks are following the IT paradigm towards virtualization and cloudification. De-
coupling hardware from software enables software to orchestrate various components,
including BBU pool and network resources. Such a flexible, automated, and self organized
network allows for various optimizations.
C-RAN builds on the base station architecture evolution from a traditional base station,
via a base station with RRH to a centralized and virtualized one. In C-RAN base station
functionalities are split between cell locations and the centralized pool. It is a challenge
to find an optimal splitting point as well as assuring efficient interconnectivity between
the parts. This dissertation summarizes the efforts to analyze this architecture, evaluate
its benefits towards energy and cost savings as well as investigates a flexible fronthaul
architecture. To conclude, C-RAN is a candidate architecture worth considering for 5G
deployments to address its performance needs and optimize deployment costs. Central-
ization and virtualization indeed offer cost savings. The point where the functionality
needs to be split between the cell site and the centralized location needs to be chosen for a
particular deployment, and a variable bit rate splits are foreseen. Ethernet-based fronthaul
is not straightforward to implement. However, with the discussed techniques it has the
potential to meet mobile networks requirements.
A comprehensive overview of C-RAN is presented. Details of this architecture are
provided, along with its benefits and technical challenges. Due to sharing of baseband
resources, C-RAN adapts well to traffic fluctuations among cells. Capacity can be scaled
more easily, energy and cost of baseband units pool deployment and operation can be
lowered. Moreover, cooperative techniques are enhanced, increasing the overall cell
throughput, while dealing efficiently with interference. However, requirements of high
capacity and low delay are put on the fronthaul network, and development of virtualization
techniques is needed. Answers to those challenges follow in terms of: analysis of possible
transport network mediums and techniques, the needed RRH and BBU development
as well as an overview of virtualization techniques. Likely deployment scenarios are
presented together with a broad overview of industry and academic work on developing
and boosting C-RAN up to the beginning of 2014. Last, but not least, an overview of
future directions for C-RAN is provided with a strong focus on defining a new functional
115
116 Chapter 5. Conclusions and outlook
split between BBU and RRH in order to enable flexible fronthaul and lower data bit rate
in this network segment.
One of the main advantages of C-RAN, multiplexing gains, is thoroughly analyzed.
Various sources of multiplexing gains have been identified and quantified for traffic-
dependent resources, namely: users changing location daily between e.g. work and home
- the so called tidal effect, traffic burstiness, as well as different functional splits. For
traditional C-RAN deployments, with a functional split between baseband and radio
functionalities, a multiplexing gain can be achieved on baseband units. However, when
the functional split allows for variable bit rate, multiplexing gains can also be exploited
on fronthaul links. The latter is an important motivation and guideline for designing the
functional split not only to lower the bit rate, but also costs of the fronthaul network. For
the analyzed data sets, the multiplexing gain value reaches six, in deployments where
various traffic types are mixed (bursty, e.g. web browsing and constant bit rate, e.g. video
streaming) and cells from various areas (e.g. office and residential) are connected to the
same BBU pool.
In order to further optimize the cost of C-RAN deployments, the possibility of reusing
existing Ethernet networks has been exploited. Such an architecture is especially optimal
for functional splits resulting in variable bit rate traffic in the fronthaul. Packet-based
networks enable multiplexing gains and flexible, multipoint-to-multipoint connectivity be-
tween cell sites and centralized locations. However, assuring synchronization and meeting
stringent delay requirements is a challenge. Mechanisms for delivering a reference clock
to cell sites have been analyzed, and an architecture employing IEEE 1588, also known as
PTP, has been evaluated as a candidate technology for C-RAN. For the tested CPRI-like
scenario, the proposed filtering gave sufficient accuracy to fulfill the requirements of
mobile networks - in the order of nanoseconds. Regarding the delay requirements, the
sources of delays have been identified and quantified. The non-deterministic queuing
delay is challenging because of its possibly variable value, however, it can be addressed
with: 1) source scheduling, especially at the edge of the network, and 2) preemption in
the switches. A source scheduling algorithm has been proposed to address the jitter and
delay constraints. It is optimized for cases where downlink traffic equals uplink traffic,
but can also be used when downlink traffic exceeds uplink. Moreover, a demonstrator of
an SDN controlled Ethernet-based fronthaul has been prepared.
This thesis identifies several possible directions for future work, both to further
improve its findings and to explore a wider perspective.
The traditional base station functional split between baseband and radio functionali-
ties was fine for short scale deployments, e.g., between rooftop and basement. However,
with C-RAN, bringing fronthaul into metropolitan scale with ever-growing capacity needs,
more disruptive solutions are needed. When this project was started, the main focus to
address this issue was on compression. Now, towards the end of the project, new func-
tional spits are being discussed: load-dependent and multipoint-to-multipoint. The user
centric/cell centric split is a very promising one, as it brings the fronthaul data rate almost
to the level offered for users, creating backhaul-like traffic, still leaving many functions
centralized. Fronthaul networks are receiving high interest from the standardization
bodies, in terms of both defining a functional split and on meeting synchronization and
delay requirements. The main examples are IEEE NGFI, IEEE 1904.3, and 802.1CM -
TSN profile for fronthaul.
This project concentrated on studying multiplexing gains on resources dependent on
user traffic. As there are parts of the base station that need to be on to provide coverage,
even when no users are active, it would be beneficial to study overall multiplexing gains
including those modules, too. Moreover, another method could be used to quantify the
multiplexing gains, not applying traffic averaging, but possibly quantifying computational
resources in terms of e.g., the number of operations per second.
In this project, a CPRI-like scenario, including one data stream was analyzed for
synchronization accuracy. Future work could investigate how the varying network load
for several data streams from a variable-bit rate splits affects synchronization accuracy.
The proposed source scheduling algorithm does not cover scenarios with uplink traffic
exceeding downlink. Moreover, it is assumed that only one BBU pool is present in the
network. A more generalized algorithm for source scheduling, with multiple BBU
pools and possibly other services present in the network, is of interest for 5G.
One of the main constraints for fronthaul is the delay requirement coming from
the HARQ process. With the current design, not meeting this requirement results in
retransmission, thereby lowering application data throughput. Therefore more studies are
needed on whether HARQ is needed in 5G, if the timer could be extended, or what could
be the delay budget share between base station and UE.
Moreover, substituting fibers with a wireless fronthaul is an important research and
development topic. Exploring new frequency bands, together with increasing spectral
efficiency of transport via this medium can address growing capacity needs.
Last but not least, solutions for hardware sharing are of high interest. Virtualization
enables sharing e.g. baseband resources, however methods for sharing other elements,
like fronthaul or cell site equipment, are important, too. The topic of network sharing is
connected with hybrid fronthaul and backhaul optimization. Future mobile networks
will most likely consist of standalone base stations as well as base stations aggregated in
a C-RAN. Joint capacity and control plane optimization for midhaul networks will enable
more efficient usage of resources, thereby lowering network deployment and operation
costs.
118
Appendices
119
APPENDIX A
OTN-based fronthaul
The following sections on OTN demonstrator are published in [86].
AI
/ OBS
CPRI OTN
O ve r
OTN
CPRI / OTN
Mapper
OBSAI Mapper
Figure A.1: C-RAN architecture where OTN is used to transport fronthaul streams
A.2 Overview
We benchmarked the CPRI/OBSAI over OTN transport performance against a reference
setup shown in Figure A.2. BSE sends IQ data, here over CPRI protocol to RRH. A signal
analyzer is used to measure EVM and Frequency Error of the transmitted signal from the
RRH antenna port.
The actual measurement setup is shown in Figure A.3. We introduced the Altera
TPO124/125 OTN multiplexer that maps CPRI client signals to OTN containers and back
from OTN containers to CPRI. We measured data EVM and Frequency Error and compare
it to the one achieved with the setup presented in Figure A.2. The detailed overview of
the system is presented in Figure A.4.
121
122 Appendix A. OTN-based fronthaul
RF CPRI
RRH
Spectrum analyzer
BSE
OTN OTN
Tamdhu
ref. board
CPRI CPRI
OTU2
10 MHz
CLKIN
CPRI-ODUk mapping
options supported: Agilent MXA
N9020A
Four separate sets of measurements were conducted, focusing on CPRI and OB-
SAI, using TPO124 and TPO125 mappers as summed up in Table A.1. Parameters are
summarized in Table A.2.
A.3 Results
Table A.3 summarizes the results for CPRI and OBSAI protocols. For three different
modulations: QPSK, 16 QAM and 64 QAM measurements were taken without OTN
device, as presented in Figure A.2, and then with TPO124 and TPO125 as OTN mappers,
as presented in Figure A.3. We present maximum observed EVM and frequency error for
each modulation, for each scenario. Table A.3 presents worst case observations over a
1-minute interval.
Looking at the performance of transmission of CPRI over OTN, we can see that OTN
transmission caused negligible EVM comparing to reference scenario. The frequency
error increased; however, it stays within the requirements. The performance of OBSAI
over OTN transmission should be compared to its reference scenario. The conclusions are
similar: negligible EVM increase, frequency error stays within requirements.
Figure A.5 shows an example of results for 64 QAM modulated signal transmitted with
OBSAI protocol over OTN with TPO125 device. Upper left figure shows the modulation
constellation. In the upper right figure the frequency error is displayed for each of the 20
slots of a 10 ms LTE frame. Lower right figure shows summary of measurements with
EVM and averaged Frequency Error.
A.4 Conclusion
OTN is a posible optical transport solution for IQ transport between RRH and BBU Pool
when mobile network operator has a cost-efficient access to legacy OTN network, which
can be reused for C-RAN. We present a proof of concept of transmitting radio interface
protocols over OTN enabling to exploit benefits of C-RAN. Tested solution introduces
negligible EVM increase and small frequency error. It is fully compliant with 3GPP
requirements for LTE-Advanced. Future work could include integration of setup with
higher CPRI/OBSAI bit rates up to 10 Gbps and introducing verification of deterministic
delay measurements.
124 Appendix A. OTN-based fronthaul
127
128 Bibliography
[21] Y. Lin, L. Shao, Z. Zhu, Q. Wang, and R. K. Sabhikhi. Wireless network cloud:
Architecture and system requirements. In: IBM Journal of Research and De-
velopment 54.1 (Jan. 2010), 4:14:12. ISSN: 0018-8646. DOI: 10.1147/JRD.
2009.2037680 (cit. on pp. 7, 8, 12, 42).
[22] J. Segel. lightRadio Portfolio: White Paper 3. Tech. rep. Alcatel-Lucent Bell Labs,
2011 (cit. on pp. 8, 21, 25, 29, 30, 32, 33, 42).
[23] Huawei. Cloud RAN Introduction. The 4th CJK International Workshop Technol-
ogy Evolution and Spectrum. September 2011 (cit. on pp. 8, 26, 42).
[24] ZTE Green Technology Innovations White Paper. Tech. rep. ZTE, 2011 (cit. on
pp. 8, 17, 42).
[25] Intel Heterogenous Network Solution Brief. Tech. rep. 2011 (cit. on pp. 8, 33, 34,
41, 42).
[26] T. Flanagan. Creating cloud base stations with TIs KeyStone multicore architec-
ture. Tech. rep. Texas Instruments, October 2011 (cit. on pp. 8, 22, 34, 42).
[27] C.-L. I, C. Rowell, S. Han, Z. Xu, G. Li, and Z. Pan. Toward Green and Soft: a 5G
Perspective. In: Communications Magazine, IEEE 52.2 (Feb. 2014), pp. 6673.
ISSN : 0163-6804. DOI : 10.1109/MCOM.2014.6736745 (cit. on p. 8).
[28] lightRadio Network. Alcatel-Lucent. [cited: January 2016]. URL: https://
www.alcatel-lucent.com/solutions/lightradio (cit. on p. 8).
[29] W. Liu, S. Han, C. Yang, and C. Sun. Massive MIMO or Small Cell Network:
Who is More Energy Efficient? In: Wireless Communications and Networking
Conference Workshops (WCNCW), 2013 IEEE. 2013, pp. 2429. DOI: 10.1109/
WCNCW.2013.6533309 (cit. on p. 8).
[30] J. Madden. Cloud RAN or Small Cells? Tech. rep. Mobile Experts, April 2013
(cit. on pp. 8, 42).
[31] G. Kardaras and C. Lanzani. Advanced Multimode Radio for Wireless and Mo-
bile Broadband Communication. In: European Wireless Technology Conference,
2009. EuWIT 2009, pp. 132135 (cit. on p. 10).
[32] E. Dahlman, S. Parkvall, J. Skold, and P. Beming. 3G Evolution: HSPA and LTE
for Mobile Broadband. 3G Evolution. Elsevier Science, 2010. ISBN: 9780080923192
(cit. on p. 10).
[33] Common Public Radio Interface (CPRI); Interface Specification V6.0. Aug. 2013
(cit. on pp. 12, 21, 31, 87).
[34] Open Base Station Architecture Initiative (OBSAI) BTS System Reference Docu-
ment Version 2.0. 2006 (cit. on p. 12).
[35] ETSI GS ORI 002-1 V1.1.1 (2011-10). Open Radio equipment Interface (ORI);
ORI Interface Specification; Part 1: Low Layers (Release 1) (cit. on p. 12).
130 Bibliography
[36] ETSI GS ORI 002-2 V1.1.1 (2012-08). Open Radio equipment Interface (ORI);
ORI Interface Specification; Part 2: Control and Management (Release 1) (cit. on
p. 12).
[37] S. Zhou, M. Zhao, X. Xu, J. Wang, and Y. Yao. Distributed wireless communica-
tion system: a new architecture for future public wireless access. In: Communi-
cations Magazine, IEEE 41.3 (Mar. 2003), pp. 108113. ISSN: 0163-6804. DOI:
10.1109/MCOM.2003.1186553 (cit. on p. 12).
[38] F. Anger. Smart Mobile Broadband. In proceedings of RAN Evolution to the Cloud
Workshop. June 2013 (cit. on p. 14).
[39] G. Brown. Converging Telecom & IT in the LTE RAN. Tech. rep. Samsung, Feb
2013 (cit. on p. 14).
[40] S. Namba, T. Matsunaka, T. Warabino, S. Kaneko, and Y. Kishi. Colony-RAN
architecture for future cellular network. In: Future Network Mobile Summit
(FutureNetw), 2012. July 2012, pp. 18 (cit. on pp. 14, 15, 22, 42, 78).
[41] T. Werthmann, H. Grob-Lipski, and M. Proebster. Multiplexing gains achieved in
pools of baseband computation units in 4G cellular networks. In: Personal Indoor
and Mobile Radio Communications (PIMRC), 2013 IEEE 24th International
Symposium on. Sept. 2013, pp. 33283333. DOI: 10.1109/PIMRC.2013.
6666722 (cit. on pp. 14, 42, 55, 78).
[42] S. Bhaumik, S. P. Chandrabose, M. K. Jataprolu, A. Muralidhar, V. Srinivasan,
G. Kumar, P. Polakos, and T. Woo. CloudIQ: A framework for processing base
stations in a data center. In: Proceedings of the Annual International Conference
on Mobile Computing and Networking, MOBICOM (2012), pp. 125136 (cit. on
pp. 15, 33, 42).
[43] M. Madhavan, P. Gupta, and M. Chetlur. Quantifying multiplexing gains in a
Wireless Network Cloud. In: Communications (ICC), 2012 IEEE International
Conference on. 2012, pp. 32123216. DOI: 10.1109/ICC.2012.6364658
(cit. on pp. 15, 42).
[44] J. Liu, S. Zhou, J. Gong, Z. Niu, and S. Xu. On the statistical multiplexing gain
of virtual base station pools. In: Global Communications Conference (GLOBE-
COM), 2014 IEEE. Dec. 2014, pp. 22832288. DOI: 10.1109/GLOCOM.2014.
7037148 (cit. on p. 15).
[45] A. Avramova, H. Christiansen, and V. Iversen. Cell Deployment Optimization
for Cloud Radio Access Networks using Teletraffic Theory. In: The Eleventh
Advanced International Conference on Telecommunications, AICT 2015. June
2015 (cit. on pp. 15, 67, 68).
Bibliography 131
[46] White Paper of Next Generation Fronthaul Interface. Tech. rep. China Mobile
Research Institute, Alcatel-Lucent, Nokia Networks, ZTE Corporation, Broadcom
Corporation, Intel China Research Center, October 2015 (cit. on pp. 16, 44, 51,
52, 84, 85, 88).
[47] C. Liu, K. Sundaresan, M. Jiang, S. Rangarajan, and G.-K. Chang. The case for re-
configurable backhaul in Cloud-RAN based small cell networks. In: INFOCOM,
2013 Proceedings IEEE. Apr. 2013, pp. 11241132. DOI: 10.1109/INFCOM.
2013.6566903 (cit. on pp. 16, 41, 42).
[48] H. Jinling. TD-SCDMA/TD-LTE evolution - Go Green. In: Communication
Systems (ICCS), 2010 IEEE International Conference on. 2010, pp. 301305.
DOI : 10.1109/ICCS.2010.5686439 (cit. on pp. 16, 17, 42).
[49] C. Chen. C-RAN: the Road Towards Green Radio Access Network. Presentation.
August 2012 (cit. on pp. 17, 41, 42).
[50] C-RAN - Road Towards Green Radio Access Network. Centralized baseband,
Collaborative radio, and real-time Cloud computing RAN. Presentation. EXPO
2010 (cit. on pp. 17, 42).
[51] J. Acharya, L. Gao, and S. Gaur. Heterogeneous Networks in LTE-Advanced.
Wiley, 2014. ISBN: 9781118693957. URL: https://books.google.dk/
books?id=RiDnAgAAQBAJ (cit. on p. 17).
[52] P. Marsch and G. Fettweis. "Coordinated Multi-Point in Mobile Communications:
From Theory to Practice". Cambridge University Press, 2011. ISBN: 9781107004115
(cit. on p. 17).
[53] R. Irmer, H. Droste, P. Marsch, M. Grieger, G. Fettweis, S. Brueck, H.-P. Mayer,
L. Thiele, and V. Jungnickel. Coordinated multipoint: Concepts, performance,
and field trial results. In: Communications Magazine, IEEE 49.2 (2011), pp. 102
111. ISSN: 0163-6804. DOI: 10.1109/MCOM.2011.5706317 (cit. on pp. 17,
22, 42).
[54] H. Holma and A. Toskala. "LTE-Advanced: 3GPP Solution for IMT-Advanced".
John Wiley and Sons, Ltd, 2012. ISBN: 9781119974055 (cit. on pp. 17, 20, 23,
42).
[55] Y. Huiyu, Z. Naizheng, Y. Yuyu, and P. Skov. Performance Evaluation of Co-
ordinated Multipoint Reception in CRAN Under LTE-Advanced uplink. In:
Communications and Networking in China (CHINACOM), 2012 7th International
ICST Conference on. 2012, pp. 778783. DOI: 10.1109/ChinaCom.2012.
6417589 (cit. on pp. 17, 42).
[56] L. Li, J. Liu, K. Xiong, and P. Butovitsch. Field test of uplink CoMP joint
processing with C-RAN testbed. In: Communications and Networking in China
(CHINACOM), 2012 7th International ICST Conference on. 2012, pp. 753757.
DOI : 10.1109/ChinaCom.2012.6417584 (cit. on pp. 17, 41, 42).
132 Bibliography
[57] J. Li, D. Chen, Y. Wang, and J. Wu. Performance Evaluation of Cloud-RAN Sys-
tem with Carrier Frequency Offset. In: Globecom Workshops (GC Wkshps), 2012
IEEE. Dec. 2012, pp. 222226. DOI: 10.1109/GLOCOMW.2012.6477573
(cit. on pp. 18, 42).
[58] D. Gesbert, S. Hanly, H. Huang, S. Shamai Shitz, O. Simeone, and W. Yu. Multi-
Cell MIMO Cooperative Networks: A New Look at Interference. In: Selected
Areas in Communications, IEEE Journal on 28.9 (2010), pp. 13801408. ISSN:
0733-8716. DOI: 10.1109/JSAC.2010.101202 (cit. on pp. 18, 42).
[59] A. Liu and V. Lau. Joint power and antenna selection optimization for energy-
efficient large distributed MIMO networks. In: Communication Systems (ICCS),
2012 IEEE International Conference on. 2012, pp. 230234. DOI: 10.1109/
ICCS.2012.6406144 (cit. on pp. 18, 42).
[60] L. Liu, F. Yang, R. Wang, Z. Shi, A. Stidwell, and D. Gu. Analysis of handover
performance improvement in cloud-RAN architecture. In: Communications and
Networking in China (CHINACOM), 2012 7th International ICST Conference
on. 2012, pp. 850855. DOI: 10.1109/ChinaCom.2012.6417603 (cit. on
p. 18).
[61] Wipro Technologies. Software-Defined Radio Technology Overview, White Paper.
August 2002 (cit. on p. 18).
[62] M. Bansal, J. Mehlman, S. Katti, and P. Levis. OpenRadio: A Programmable
Wireless Dataplane. In: ACM SIGCOMM Workshop on Hot Topics in Software
Defined Networking (HotSDN12), ACM SIGCOMM 2012, pp. 109114 (cit. on
p. 18).
[63] S. Thomas. Poll: Savings Drive CRAN Deployments. LightReading, July 2015
(cit. on pp. 19, 41).
[64] Front-haul Compression for Emerging C-RAN and Small Cell Networks. Tech. rep.
Integrated Device Technology, April 2013 (cit. on p. 21).
[65] W. Huitao and Z. Yong. C-RAN Bearer Network Solution. Tech. rep. ZTE, Novem-
ber 2011 (cit. on pp. 21, 27, 41, 42).
[66] S. Namba, T. Warabino, and S. Kaneko. BBU-RRH Switching Schemes for
Centralized RAN. In: Communications and Networking in China (CHINACOM),
2012 7th International ICST Conference on. 2012, pp. 762766. DOI: 10.1109/
ChinaCom.2012.6417586 (cit. on pp. 22, 42).
[67] H. Raza. A brief survey of radio access network backhaul evolution: part I. In:
Communications Magazine, IEEE 49.6 (2011), pp. 164171. ISSN: 0163-6804.
DOI : 10.1109/MCOM.2011.5784002 (cit. on pp. 23, 25, 42).
[68] C. Chen, J. Huang, W. Jueping, Y. Wu, and G. Li. Suggestions on Potential
Solutions to C-RAN. Tech. rep. NGMN Alliance, 2013 (cit. on pp. 24, 38, 40, 42,
44, 88).
Bibliography 133
[69] J. Segel and M. Weldon. lightRadio Portfolio: White Paper 1. Tech. rep. Alcatel-
Lucent Bell Labs, 2011 (cit. on p. 25).
[70] 3GPP TR 36.932 Scenarios and requirements for small cell enhancements for
E-UTRA and E-UTRAN V 12.1.0. Mar. 2013 (cit. on pp. 25, 37).
[71] Z. Ghebretensae, K. Laraqui, S. Dahlfort, J. Chen, Y. Li, J. Hansryd, F. Ponzini,
L. Giorgi, S. Stracca, and A. Pratt. Transmission solutions and architectures for
heterogeneous networks built as C-RANs. In: Communications and Network-
ing in China (CHINACOM), 2012 7th International ICST Conference on. 2012,
pp. 748752. DOI: 10.1109/ChinaCom.2012.6417583 (cit. on pp. 25, 27,
41, 42).
[72] E-BLINK. Wireless Fronthaul Technology. [cited: January 2016]. URL: http:
//e-blink.com/ (cit. on p. 25).
[73] J. H. Lee, S.-H. Cho, K. H. Doo, S.-I. Myong, J. H. Lee, and S. S. Lee. CPRI
transceiver for mobile front-haul based on wavelength division multiplexing. In:
ICT Convergence (ICTC), 2012 International Conference on. 2012, pp. 581582.
DOI : 10.1109/ICTC.2012.6387205 (cit. on pp. 26, 42).
[74] F. Ponzini, L. Giorgi, A. Bianchi, and R. Sabella. Centralized radio access
networks over wavelength-division multiplexing: a plug-and-play implementa-
tion. In: Communications Magazine, IEEE 51.9 (2013). ISSN: 0163-6804. DOI:
10.1109/MCOM.2013.6588656 (cit. on pp. 26, 42).
[75] ITU-T G.709/Y.1331, Interfaces for the optical transport network. Geneva, Feb.
2012 (cit. on p. 27).
[76] ODU0 and ODUflex A Future-Proof Solution for OTN Client Mapping. Tech. rep.
TPACK, February 2010 (cit. on p. 27).
[77] B. Liu, X. Xin, L. Zhang, and J. Yu. 109.92-Gb/s WDM-OFDMA Uni-PON with
dynamic resource allocation and variable rate access. In: OPTICS EXPRESS,
Optical Society of America 20.10 (May 2012) (cit. on pp. 27, 42).
[78] J. Fabrega, M. Svaluto Moreolo, M. Chochol, and G. Junyent. WDM overlay
of distributed base stations in deployed passive optical networks using coherent
optical OFDM transceivers. In: Transparent Optical Networks (ICTON), 2012
14th International Conference on. 2012, pp. 14. DOI: 10 . 1109 / ICTON .
2012.6253934 (cit. on pp. 27, 42).
[79] S. Chia, M. Gasparroni, and P. Brick. The Next Challenge for Cellular Networks:
Backhaul. In: IEEE Microwave (August 2009) (cit. on p. 27).
[80] R. Snchez, L. Raptis, and K. Vaxevanakis. Ethernet as a Carrier Grade Tech-
nology: Developments and Innovations. In: IEEE Communications Magazine
(September 2008) (cit. on p. 28).
134 Bibliography
[81] 3GPP TS 36.104 Evolved Universal Terrestrial Radio Access (E-UTRA); Base
Station (BS) radio transmission and reception V 12.0.0. July 2013 (cit. on pp. 28,
30, 84, 87, 124, 125).
[82] 3GPP TS 36.133 Evolved Universal Terrestrial Radio Access (E-UTRA); Require-
ments for Support of Radio Resource Management. V 12.0.0. July 2013 (cit. on
pp. 28, 87).
[83] NFV Cloud RAN with Ethernet Fronthaul. [cited: January 2016]. URL: http:
//www.altiostar.com/solution/ (cit. on pp. 28, 36).
[84] W. Liu, K. Chen, W. Ma, and S. Norberg. Remote radio data transmission over
ethernet. United States Patent Application 20120113972 A1. 2012 (cit. on p. 28).
[85] Altera. SoftSilicon OTN Processors. [cited: September 2013]. URL: http://
www.altera.com/end-markets/wireline/applications/otn/
softsilicon-processors/proc-index.html (cit. on pp. 29, 42).
[86] A. Checko, G. Kardaras, C. Lanzani, D. Temple, C. Mathiasen, L. A. Pedersen,
and B. Klaps. OTN Transport of Baseband Radio Serial Protocols in C-RAN
Architecture for Mobile Network Applications. Tech. rep. MTI Mobile and Altera,
March 2014 (cit. on pp. 29, 42, 121, 122).
[87] D. Holberg, Hughes Aircraft Company. An adaptive digital automatic gain
control for MTI radar Systems. 3781882. 1973 (cit. on p. 30).
[88] M. Grieger, S. Boob, and G. Fettweis. Large scale field trial results on frequency
domain compression for uplink joint detection. In: Globecom Workshops (GC
Wkshps), 2012 IEEE. 2012, pp. 11281133. DOI: 10.1109/GLOCOMW.2012.
6477737 (cit. on pp. 30, 42).
[89] D. Samardzija, J. Pastalan, M. MacDonald, S. Walker, and R. Valenzuela. Com-
pressed Transport of Baseband Signals in Radio Access Networks. In: Wireless
Communications, IEEE Transactions on 11.9 (September 2012), pp. 32163225.
ISSN : 1536-1276. DOI : 10.1109/TWC.2012.062012.111359 (cit. on
pp. 3032, 42).
[90] B. Guo, W. Cao, A. Tao, and D. Samardzija. CPRI compression transport for
LTE and LTE-A signal in C-RAN. In: Communications and Networking in China
(CHINACOM), 2012 7th International ICST Conference on. 2012, pp. 843849.
DOI : 10.1109/ChinaCom.2012.6417602 (cit. on pp. 30, 32, 42).
[91] I. D. Technology. Compression IP for Wireless Infrastructure Applications. Prod-
uct brief. [July 2013] (cit. on pp. 30, 32, 42).
[92] J. Lorca and L. Cucala. Lossless compression technique for the fronthaul of
LTE/LTE-advanced Cloud-RAN architectures. In: World of Wireless, Mobile and
Multimedia Networks (WoWMoM), 2013 IEEE 14th International Symposium and
Workshops on a. June 2013, pp. 19 (cit. on pp. 31, 32, 42).
Bibliography 135
[93] S.-H. Park, O. Simeone, O. Sahin, and S. Shamai (Shitz). Robust and Effi-
cient Distributed Compression for Cloud Radio Access Networks. In: Vehicular
Technology, IEEE Transactions on 62.2 (February 2013), pp. 692703. ISSN:
0018-9545. DOI: 10.1109/TVT.2012.2226945 (cit. on pp. 31, 32, 42).
[94] Philippe Sehier et al. Liaisons, Contributions to 3GPP ETSI on Collaborative
Radio/MIMO, ORI Interface, ect. Tech. rep. NGMN Alliance, 2013 (cit. on p. 33).
[95] Ericsson. Worlds first microwave connection between LTE main and remote
radio units. [February 2012]. URL: http://www.ericsson.com/news/
1588074 (cit. on pp. 33, 42).
[96] X. Wei, X. Qi, L. Xiao, Z. Shi, and L. Huang. Software-defined Radio based
On Cortex-A9. In: Communications and Networking in China (CHINACOM),
2012 7th International ICST Conference on. 2012, pp. 758761. DOI: 10.1109/
ChinaCom.2012.6417585 (cit. on p. 34).
[97] D. Martinez-Nieto, V. Santos, M. McDonnell, K. Reynolds, and P. Carlston. Dig-
ital Signal Processing on Intel
R Architecture. In: Intel
R Technology Journal
13 (1 2009) (cit. on p. 34).
[98] N. Kai, S. Jianxing, C. Kuilin, and K. K. Chai. TD-LTE eNodeB prototype
using general purpose processor. In: "Communications and Networking in China
(CHINACOM), 2012 7th International ICST Conference on". 2012, pp. 822827.
DOI : 10.1109/ChinaCom.2012.6417598 (cit. on p. 34).
[99] S. Zhang, R. Qian, T. Peng, R. Duan, and K. Chen. High throughput turbo
decoder design for GPP platform. In: Communications and Networking in China
(CHINACOM), 2012 7th International ICST Conference on. 2012, pp. 817821.
DOI : 10.1109/ChinaCom.2012.6417597 (cit. on p. 35).
[100] Z. Guanghui, N. Kai, H. Lifeng, and H. Jinri. A method of optimizing the de-Rate
Matching and demodulation in LTE based on GPP. In: Communications and
Networking in China (CHINACOM), 2012 7th International ICST Conference
on. 2012, pp. 828832. DOI: 10.1109/ChinaCom.2012.6417599 (cit. on
p. 35).
[101] T. Kaitz and G. Guri. CPU-MPU partitioning for C-RAN applications. In:
Communications and Networking in China (CHINACOM), 2012 7th International
ICST Conference on. 2012, pp. 767771. DOI: 10.1109/ChinaCom.2012.
6417587 (cit. on p. 35).
[102] H. Duan, D. Huang, Y. Huang, Y. Zhou, and J. Shi. A time synchronization
mechanism based on Software Defined Radio of general-purpose processor. In:
Communications and Networking in China (CHINACOM), 2012 7th International
ICST Conference on. 2012, pp. 772777. DOI: 10.1109/ChinaCom.2012.
6417588 (cit. on p. 35).
136 Bibliography
[146] U. Dotsch, M. Doll, H.-P. Mayer, F. Schaich, J. Segel, and P. Sehier. Quantita-
tive analysis of split base station processing and determination of advantageous
architectures for LTE. In: vol. 18. 1. June 2013, pp. 105128 (cit. on p. 44).
[147] B. Haberland. Smart Mobile Cloud. 2013 (cit. on pp. 44, 51, 52).
[148] Small cell virtualization functional splits and use cases. Tech. rep. Small Cell
Forum, June 2015 (cit. on pp. 44, 45, 51, 52, 84, 85).
[149] H. Lee, Y. O. Park, and S. S. Song. A traffic-efficient fronthaul for the cloud-
RAN. In: Information and Communication Technology Convergence (ICTC),
2014 International Conference on. Oct. 2014, pp. 675678. DOI: 10.1109/
ICTC.2014.6983252 (cit. on p. 44).
[150] B. Haberland. Cloud RAN architecture evolution from 4G to 5G Mobile Systems.
2014 (cit. on p. 44).
[151] E. Dahlman, S. Parkvall, and J. Skold. 4G: LTE/LTE-Advanced for Mobile Broad-
band. Elsevier Science, 2011. ISBN: 9780123854902. URL: https://books.
google.com.qa/books?id=DLbsq9GD0zMC (cit. on pp. 52, 53).
[152] 3GPP, Technical Specification Group Radio Access Network; Evolved Universal
Terrestrial Radio Access (E-UTRA). Packet Data Convergence Protocol (PDCP)
specification (Release 8). Tech. rep. 36.323, v. 8.6.0. 2009 (cit. on p. 51).
[153] 3GPP, Technical Specification Group Radio Access Network; Evolved Univer-
sal Terrestrial Radio Access (E-UTRA). Radio Link Control (RLC) protocol
specification (Release 8). Tech. rep. 36.322, v. 8.8.0. 2010 (cit. on p. 51).
[154] 3GPP, Technical Specification Group Radio Access Network; Evolved Universal
Terrestrial Radio Access (E-UTRA). Medium Access Control (MAC) protocol
specification (Release 8). Tech. rep. 36.321, v. 8.11.0. 2011 (cit. on p. 52).
[155] C. Desset, B. Debaillie, V. Giannini, A. Fehske, G. Auer, H. Holtkamp, W. Wajda,
D. Sabella, F. Richter, M. J. Gonzalez, H. Klessig, I. Gdor, M. Olsson, M. A. Im-
ran, A. Ambrosy, and O. Blume. Flexible power modeling of LTE base stations.
In: Wireless Communications and Networking Conference (WCNC), 2012 IEEE.
Apr. 2012, pp. 28582862. DOI: 10.1109/WCNC.2012.6214289 (cit. on
pp. 5557).
[156] Alcatel-Lucent and Vodafone Chair on Mobile Communication Systems. Study
on Energy Efficient Radio Access Network (EERAN) Technologies. Unpublished
Project Report, Technical University of Dresden, Dresden, Germany, 2009 (cit. on
pp. 56, 57).
[157] L. M. Correia, D. Zeller, O. Blume, D. Ferling, Y. Jading, I. Gdor, G. Auer,
and L. V. D. Perre. Challenges and enabling technologies for energy aware
mobile radio networks. In: IEEE Communications Magazine 48.11 (Nov. 2010),
pp. 6672. ISSN: 0163-6804. DOI: 10.1109/MCOM.2010.5621969 (cit. on
p. 56).
Bibliography 141
[171] B. Sadiq, R. Madan, and A. Sampath. Downlink Scheduling for Multiclass Traffic
in LTE. In: EURASIP Journal on Wireless Communications and Networking
2009.1 (2009), p. 510617. ISSN: 1687-1499. DOI: 10.1155/2009/510617.
URL : http : / / jwcn . eurasipjournals . com / content / 2009 / 1 /
510617 (cit. on p. 84).
[172] T. Pfeiffer and F. Schaich. Optical Architectures For Mobile Back- And Fronthaul-
ing. OFC/NFOEC wireless backhauling workshop. Los Angeles: Alcatel-Lucent
Bell Labs Stuttgart, Mar. 5, 2012 (cit. on p. 84).
[173] J. H. et al. Further Study on Critical C-RAN Technologies, Version 1.0, tech. rep.
NGMN Alliance, 2015 (cit. on p. 85).
[174] Metro Ethernet Forum. Mobile Backhaul Implementation Agreement Phase 2,
Amendment 1 Small Cells, Tech. Spec. MEF 22.1.1. July 2014 (cit. on pp. 85,
111).
[175] Federal Communications Commission FCC 15-9, PS Docket No. 07-114. Feb.
2015 (cit. on p. 86).
[176] M. Weiss. Telecom Requirements for Time and Frequency Synchronization. Pre-
sented at 52nd Meeting of the Civil GPS Service Interface Committee in Nashville,
September 17th-18th, 2012. Time and Frequency Division, NIST (cit. on pp. 86,
87).
[177] E. Metsl and J. Salmelin. Mobile Backhaul. John Wiley and Sons, Ltd., 2012.
ISBN : 978-1-119-97420-8 (cit. on p. 86).
[178] ETSI/GSM. GSM Recommendation 05.10, 3GPP TS 05.10 Radio subsystem
synchronization, Release Ph1. Oct. 1992 (cit. on p. 86).
[179] H. Li. Consideration on CRAN fronthaul architecture and interface. Presented at
1st NGFI workshop in Beijing, China, June 4th, 2015. China Mobile Research
Institute, June 4, 2015 (cit. on p. 86).
[180] C.-L. I. NGFI: Next Generation Fronthaul towards 4.5G and 5G. Presented at
1st NGFI workshop in Beijing, China, June 4th, 2015. China Mobile Research
Institute, June 4, 2015 (cit. on pp. 86, 88, 89).
[181] D. Bladsjo, M. Hogan, and S. Ruffini. Synchronization aspects in LTE small
cells. In: Communications Magazine, IEEE 51.9 (2013), pp. 7077. ISSN: 0163-
6804. DOI: 10.1109/MCOM.2013.6588653 (cit. on p. 87).
[182] C.-L. I, Y. Yuan, J. Huang, S. Ma, C. Cui, and R. Duan. Rethink fronthaul for
soft RAN. In: Communications Magazine, IEEE 53.9 (Sept. 2015), pp. 8288.
ISSN : 0163-6804. DOI : 10.1109/MCOM.2015.7263350 (cit. on p. 88).
[183] IEEE 1904.3 Task Force. URL: http://www.ieee1904.org/3 (cit. on
p. 88).
Bibliography 143
[184] IEEE Standard for a Precision Clock Synchronization Protocol for Networked
Measurement and Control Systems. In: IEEE Std 1588-2008 (Revision of IEEE
Std 1588-2002) (July 2008), pp. c1269. DOI: 10.1109/IEEESTD.2008.
4579760 (cit. on p. 90).
[185] IEEE Standard for Ethernet. In: IEEE Std 802.3-2012 (Dec. 2012), pp. 1. DOI:
10.1109/IEEESTD.2012.6419735 (cit. on p. 90).
[186] P. A. Smith. CPRI FrontHaul requirements discussion with TSN. San Diego,
CA: Huawei, July 1, 2014 (cit. on p. 91).
[187] Precise Timing for Base Stations in the Evolution to LTE. Tech. rep. Vitesse, 2013
(cit. on p. 93).
[188] L. Xie, Y. Wu, and J. Wang. Efficient time synchronization of 1588v2 technology
in packet network. In: Communication Software and Networks (ICCSN), 2011
IEEE 3rd International Conference on. May 2011, pp. 181185. DOI: 10.1109/
ICCSN.2011.6014030 (cit. on p. 95).
[189] Time Sensitive Networking Task Group of IEEE 802.1, IEEE 802.1Qbv/D2.3
Bridges and Bridged NetworksAmendment: Enhancements for Scheduled Traf-
fic, IEEE. In: (2015) (cit. on p. 100).
[190] Time Sensitive Networking Task Group of IEEE 802.1, IEEE 802.1Qbu/D2.2
Bridges and Bridged NetworksAmendment: Frame Preemption, IEEE. In:
(2015) (cit. on p. 100).
[191] IEEE Higher Layer LAN Protocols Working Group (C/LM/WG802.1), Time-
Sensitive Networking for Fronthaul, P802.1CM PAR. In: (2015) (cit. on p. 102).
[192] LAN/MAN Standards Committee, IEEE P802.3br/D1.0 Draft Standard for Eth-
ernet Amendment: Specification and Management Parameters for Interspersing
Express Traffic., IEEE. In: (2014) (cit. on p. 104).
[193] T. Wan and P. Ashwood. A Performance Study of CPRI over Ethernet. Presented
at IEEE 1904.3 Task Force meeting in Louisville, January 30th, 2015. Huawei
Canada Research Center, Jan. 30, 2015 (cit. on p. 104).
[194] J. Farkas and B. Varga. Applicability of Qbu and Qbv to Fronthaul. Ericsson,
Nov. 11, 2015 (cit. on p. 105).
[195] T. Stern and K. Bala. Multiwavelength Optical Networks: A Layered Approach.
Addison-Wesley professional computing series. Addison-Wesley, 1999. ISBN:
9780201309676. URL: https://books.google.dk/books?id=C%5C_
lSAAAAMAAJ (cit. on p. 106).
[196] A. De La Oliva, X. Costa Perez, A. Azcorra, A. Di Giglio, F. Cavaliere, D.
Tiegelbekkers, J. Lessmann, T. Haustein, A. Mourad, and P. Iovanna. Xhaul:
toward an integrated fronthaul/backhaul architecture in 5G networks. In: Wireless
Communications, IEEE 22.5 (Oct. 2015), pp. 3240. ISSN: 1536-1284. DOI:
10.1109/MWC.2015.7306535 (cit. on p. 111).
144 Bibliography