Harp d2 1 1 Proceedings of First Special Session

HARP deliverable: D2.1.
1 Proceedings of first workshop/ special session at an international conference

Project Title: High capacity network Architecture with Remote radio heads & Parasitic antenna arrays
Grant Agreement Number: 318489
D2.1.1 Proceedings of first workshop/ special session at international conference
D2.1.1 Proceedinngs of the first workshop/special session at

an international conference
Authors:
Laura Cottatellucci, Ralf Mller, Lars Dittmann, Henrik L. Christiansen, Aleksandra Checko,
Stephane Senecal, Jean-Marc Kelif.
-1-
HARP deliverable: D2.1.1 Proceedings of first workshop/ special session at an international conference
Table of Contents
1.
Introduction and summary ............................................................................................................... 3
2.
General Information about the Workshop ....................................................................................... 4
3.
Workshop Structure and Contributors ............................................................................................. 4
4.
Overview of the Technical Contributions ......................................................................................... 5
5.
Appendix ........................................................................................................................................... 8
-2-
1.
Introduction and summary
In line with its vision and philosophy for the development of the future 5G wireless network, HARP
promoted and organized a workshop titled Wireless optical network convergence in support of cloud
architectures, shortly
WONC, at IEEE Globecom 2014, a flagship conference of the IEEE
communications society.
As from the title, the workshop assigns a prominent role to the concept of radio access networks based
on cloud computing to enable disruptive technological and architectural changes required by the
challenging targeted Key Performance Indicators (KPI) of 5G wireless networks. For the technical and
economic success of this concept, the integration and convergence of radio access and optical networks
play a strategic role and the workshop focuses on this specific aspect of cloud radio access networks. In
the workshop vision, the cloud architecture is based on massively distributed antennas or remote radio
heads (RRHs) connected to a computing cloud through an optical fiber infrastructure that replaces the
traditional backhaul network.
It is worth noting that the original declared aim of HARP was to bring distributed multi-antenna wireless
access to reality by combining two powerful technologies: 1) radio remote heads (RRHs), which allow for
widely geographically distributed access via radio-over-fiber connections to a central processing unit;
and 2) electronically steerable passive array radiators, which provide multi-antenna-like functionality
with a single active RF chain only, and thus developing the RRH optical aggregation network, required
for efficient coordination between disparate access areas. Then, the core of HARP is completely aligned
to the scope of the workshop and HARP contributed to WONC by proposing solutions for high capacity,
low-cost, and environmentally friendly multi-antenna RRHs supported by a flexible optical aggregation
network.
The relevance of the topic proposed in WONC is corroborated by the fact that several European projects
are working on this subject from different and complementary perspectives as well as major players in
the worldwide communications arena. Then, primary objective of the workshop was to bring together
academic and industrial actors in the field to draw the current state of art, identify the technical
challenges, exchange results and strategic visions, foster discussions to identify the key driving concepts
and technologies, and, eventually, develop synergies. This objective was achieved by inviting paper
authors and keynote speakers from relevant flagship EU funded research projects such as TROPIC,
CONTENT, SODALES, iJOIN, and HARP and representatives of leading industries as Huawei (China) and
prominent American scientists very active in the field.
The workshop was structured into two sessions and covered a wide range of key issues on the wireless
optical network convergence in support of cloud architectures. In particular, they span from emerging
technologies for remote radio heads with low capital and operational costs to signal processing
techniques for distributed antenna systems taking into account the non-ideality of the front-haul and
the back-haul, to allocation and scheduling of radio, optical and computational resources. In addition,
-3-
emphasis was given in the architectural challenges associated with the converged optical/wireless
networks to support cloud and mobile cloud services.
The rest of the deliverable is structured as follows. In Section 2 we provide general information about
the workshop. Section 3 provides information about the structure of the workshop, keynote speakers
and the invited paper authors that contributed to the workshop. Section 4 gives an overview of the
technical contributions. Section 5 describes the actions taken to promote the workshop. Finally, Section
6 contains the full contribution to WONC by the invited authors.
2.
General Information about the Workshop
Title: International Workshop on Wireless Optical Network Convergence in Support of Cloud

Architectures
Website Link: http://www.fp7-harp.eu/globecom-2014wonc.html
Hosting Conference: IEEE Globecom 2014
Date and Place: December 12, 2014 at Austin, TEXAS
Workshop Organizer/Chair: Laura Cottatellucci (HARP member, EURECOM, France), Anna Tzanakaki
(University of Bristol, United Kingdom)
Number of Attendees: About 20 in average. Not official number is available because there was a unique
registration for all the workshops planned on December 12, 2014.
Number of Keynote Speeches: 2, one out of two was a HARP member.
Number of Invited papers: 7, two out of seven were HARP contributions.
Number of TPC members: 12, four out of 12 were HARP members.
3.
Workshop Structure and Contributors
The workshop was structured into two sessions. The first session titled Radio Access Technologies for
Wireless Optical Networks (RAT) collected all the contributions related to the physical layer, from
hardware design to algorithms for transmission and resource allocation. The second session titled
Wireless Optical Networks for Cloud Architectures (WON) was focused on architectural issues raised
by the cloud architectures. Each of the sessions was open by a keynote speech and followed by the
presentation of invited papers.
-4-
The RAT session was open by the keynote speech of Prof. Ralf Mueller, HARP member from University
Erlangen-Nrnberg, Germany. Prof. Mueller presented an innovative technologies for multiantenna
arrays with single active RF chain developed in the framework of HARP in a speech titled "Load
Modulated MIMO: A New Hardware Concept to Reduce Cost, Size, and Amplifier Back-off".
Followed the works of the invited authors:
1) Prof. Wei Yu from University of Toronto, Canada and holding a Canada Research Chair in
Information Theory and Digital Communications (directly presented).
2) Prof. Gerhard Fettweis, University of Dresden, Germany, holding a Vodafone chair and active
member of the European STREP project iJOIN (presented by the co-author Jens Bartelt).
3) Prof. Sergio Barbarossa from University of La Sapienza, Roma, Italy and active member of the
European STREP project TROPIC (due to health problems, presented by his colleague Adrian
Agustin, from UPC, Barcelona also member of TROPIC).
The WON session was open by an introductory keynote speech prepared by the invited speaker Dr.
Ganghua Yang, vice-director of the communication technology lab at Huawei, Shanghai and due to VISA
problem, the speech was given by his colleague Dr. Jian Li from the same lab. The keynote speech titled
Wireless Optical Network Convergence in 5G Access Network provides a broad vision at the cutting
edge of current technological developments from Huawei perspective.
The keynote speech was followed by the presentation of the works of the invited authors:
1) Prof. Lars Dittmann, Technical University of Denmark, Denmark, member of the European STREP
Project HARP (directly presented).
2) Dr. Markos Anastasopoulos University of Bristol, UK, member of the EU Project CONTENT
(directly presented).
3) Jordi Ferrer Riera from Fundaci i2CAT, Internet i Innovaci Digital a Catalunya, Spain, active
member of the EU Project SODALES (directly presented).
4) Dr. Jean Marc Kelif, Orange Labs, France and member of the European Project HARP.
4.
Overview of the Technical Contributions
The keynote speech opening the RAT session and titled "Load Modulated MIMO: A New Hardware
Concept to Reduce Cost, Size, and Amplifier Back-off" presented recent results of the HARP project on
the architecture of massive antenna arrays. The proposed architecture enables multi antenna array with
single power amplifier chain and allows significant front-end hardware savings resulting in an appealing
enabler for future small and powerful cell nodes able to carry massive MIMO technology. The talk
reviewed and compared radio frequency (RF) hardware architectures with a single power amplifier and
showed that they are superior to conventional MIMO implementations in terms of cost, dissipated heat,
and physical size. This progress on the RF-side have relevant impact also on the future network
architectures since it allows to merge the two competing cellular concepts of virtual and massive MIMO
into a hybrid approach of remote radio heads with massive MIMO arrays.
-5-
The keynote speech was followed by the presentation of two works that focus on the non-ideality of the
backhaul and the front-haul.
The article titled Optimized Beamforming and Backhaul Compression for Uplink MIMO Cloud Radio
Access Networks studied the optimization of transmit beamforming and backhaul compression
strategies for the uplink of cloud radio access networks (C-RAN), in which multi-antenna users
communicate with a cloud-computing based central processor (CP) through multi-antenna base-stations
(BSs). The BSs perform compress-and-forward strategy to quantize the received signals and send the
quantization bits to the CP via capacity-limited backhaul links for decoding. In contrast to the previous
work on the uplink C-RAN, which mostly focuses on the backhaul compression strategies only, in this
work the authors propose the joint optimization of the transmit beamformers and the quantization
noise covariance matrices at the BSs for maximizing the benefit brought by C-RAN. A novel weighted
minimum-mean-square-error successive convex approximation (WMMSE-SCA) algorithm was developed
for finding a local optimum solution to the problem along with suboptimal low-complexity
approximations. Simulations show that with optimized beamforming and backhaul compression, C-RAN
can significantly improve the overall performance of wireless cellular networks.
The impact of imperfect optical links in the front-haul is studied in the project iJoin and presented in the
work titled An Improved Decoder for Cloud-Based Mobile Networks under Imperfect Fronthaul. The
fronthaul of a cloud-RAN network forwards digitized samples to the central processing unit. However,
with increasingly dense deployments of small cells, the availability of high quality fiber fronthaul cannot
be expected at all sites. Longer fronthaul distances or the employment of wireless technologies renders
fronthaul links imperfect in the sense that they introduce bit-errors into the digital samples. The authors
build upon so-called hardware noise resilient decoders to investigate decoders for imperfect fronthaul
that can achieve very good performance even under high bit-error rates on the fronthaul.
The session was concluded by the presentation of results of the European project Tropic. The work titled
Distributed Mobile Cloud Computing: Joint Optimization of Radio and Computational Resources
focused on the challenging problem of jointly optimize the radio interface by beamforming design and
computational resource allocation to minimize the overall power consumption under latency constraints.
More specifically, the authors consider a scenario composed by multiple mobile users asking for
computation offloading of their applications to a set of cloud servers. They aim to find the optimal
assignment of each mobile user to a cloud server through the most convenient base station and,
contextually, the optimal MIMO precoding matrices and computational rates (virtual machines) to each
user, under latency constraints dictated by the users' Quality of Experience (QoE). The radio resources
assigned to users belonging to the same cell are orthogonal to each other, but users belonging to
different cells might interfere with each other. The latency constraint imposes a strict relationship
between the time spent for transferring the program execution from the mobile device to the fixed
server (and viceversa) and the time needed to execute the computation. The authors formulate the
computation offloading problem as a joint optimization of the radio and computational resources, with
the objective of minimizing the overall energy consumption, at the mobile terminal side, while meeting
the latency constraints. Numerical results show that the proposed joint optimization yields significant
-6-
energy savings with respect to more traditional schemes performing a separate optimization of the
radio and computational resources.
The WON session was focused on the network architecture aspects and started with a keynote speech
titled "Wireless Optical Network Convergence in 5G Access Network". Huawei presented its vision on
5G wireless systems in term of service requirements and network architecture. Some scenarios
for wireless optical network convergence at the access network and at base-stations were analyzed.
Finally, hot research topics on wireless optical network convergence were suggested and discussed.
The following work reflected contribution in the field of the European project Conect and was titled
Dependable Multi-Tenant Infrastructures Supporting Cloud and Mobile Cloud Services. The authors
focus on the design of dependable Mobile Optical Virtual Networks (MOVNOs) to enable multi-tenancy
over converged wireless, optical and computational infrastructures supporting cloud and mobile cloud
services. Dependability is defined as the MOVNO's ability to offer resilience in case of network or
computing resource failures and security guarantees in terms of isolation between MOVNOs that share
the common physical infrastructure. The VI planning optimization objective is to minimize the overall
energy consumption satisfying at the same time specific end-to-end delay constraints. To address these
issues, together with the uncertainty introduced by the mobility of the end devices the authors
proposed a MOVNO planning scheme based on non-linear programming. This modeling results in
interesting trade-offs between the overall energy consumption of the MOVNOs, the mobility of the enddevices, as well as the dependability impact on the energy consumption and the utilization of the
infrastructure resources.
The subsequent presentation reflected the HARP vision on the network architecture with special focus
on the design of the fronthaul and the flow aggregation at that level. The work was titled Meeting
fronthaul challenges of future mobile network deployments - the HARP approach. This contribution
addresses the need to design new architecture to be able to perform a flexible network aggregation at
different levels in future mobile networks. The authors presented the design approach for a C-RAN
based mobile aggregation network used in the EU HARP project. With the proposed architecture fronthaul aggregation is performed which might be an option for future generation of mobile networks.
The SODALES (SOftware-Defined Access using Low-Energy Subsystems) activity on the field and its vision
on wireless optical networks supporting cloud architecture is presented in the work titled Softwaredefined wired-wireless access network convergence: the SODALES approach. In the SODALES vision an
active remote node (ARN) at an intermediate location between the central office and end-user premises
is introduced to solve the issues associated to bandwidth and wired-wireless convergence. The ARN
represents the key architectural design innovation of the SODALES network. The purely passive
approaches (e.g. as exemplified by passive optical networking (PON) access architectures) are shown to
be too restrictive for future 5G networks when compared the modularity and flexibility offered by an
active remote node. A performance analysis of the ARN node demonstrates the potential benefits in
terms network performance, operational efficiency, and flexible functionality of the proposed
architecture. Future ARN capabilities can also be expected to include hierarchical caching, customer
-7-
premises equipment (CPE) virtualization, and nearer to the end-user location of software-defined
platforms supporting ubiquitous cloud services.
The WON session is concluded by an HARP contribution to facilitate simulations at a network level and is
titled Analytical Performance Model for Poisson Wireless Networks with Pathloss and Shadowing
Propagation. The authors propose a fluid network model to easily compute the SINR (signal to
interference plus noise ratio) for wireless networks analysis in scenarios including pathloss and
shadowing.
5.
Workshop Promotion Activities
The workshop was promoted by regularly announcing it in worldwide emailing lists of the
telecommunication society such as TCCC-announce (tccc-announce@COMSOC.ORG) and TCCN
(tccn@COMSOC.ORG). This allowed us to reach a very wide audience. A more targeted advertising was
achieved by announcing the workshop in European project mailing lists such as HARP CONTENT,
SODALES, iJOIN focused on the specific topic of our workshop.
Additionally, the workshop flier was shown on the EURECOM website and the websites of the above
mentioned European projects.
6.
Appendix
This appendix contains the following documents:
1) Presentation slides of the keynote speech: Load Modulated MIMO: A New Hardware
Concept to Reduce Cost, Size, and
Amplifier Back-off, Ralf Mller (University
Erlangen-Nrnberg, Germany).
2) Invited paper: Optimized Beamforming and Backhaul Compression for Uplink MIMO
Cloud Radio Access Networks, Yuhan Zhou and Wei Yu (University of Toronto, Canada).
3) Invited paper: An Improved Decoder for Cloud-Based Mobile Networks under Imperfect
Fronthaul, Jens Bartelt (Dresden University of Technology, Germany); Gerhard Fettweis
(Technische Universitt Dresden, Germany).
4) Invited paper: Distributed Mobile Cloud Computing: Joint Optimization of Radio and
Computational Resources, Sergio Barbarossa and Stefania Sardellitti (University of Rome
La Sapienza, Italy); Gesualdo Scutari (State University of New York at Buffalo, USA).
5) Invited paper: Dependable Multi-Tenant Infrastructures Supporting Cloud and Mobile
Cloud Services, Markos Anastasopoulos, Anna Tzanakaki and Dimitra Simeonidou
(University of Bristol, United Kingdom).
-8-
6) Invited paper: Meeting fronthaul challenges of future mobile network deployments - the
HARP approach, Lars Dittmann, Henrik L. Christiansen (Technical University of Denmark,
Denmark), and Aleksandra Checko (MTI Radiocomp).
7) Invited paper: Software-defined wired-wireless access network convergence: the
SODALES approach, Jordi Ferrer Riera, Carlos Bock and Eduard Escalona (Fundaci i2CAT,
Internet i Innovaci Digital a Catalunya, Spain); Volker Jungnickel (Fraunhofer Heinrich Hertz
Institute, Germany); Kai Habel (Fraunhofer HHI, Germany); Michael Parker, Stuart D Walker
and Terry Quinlan (University of Essex, United Kingdom); Victor Marques (Portugal Telecom
Inovacao, Portugal); David Levy (Ethernity Networks, Israel).
8) Invited paper: Analytical Performance Model for Poisson Wireless Networks with Pathloss
and Shadowing Propagation, Jean-Marc Kelif and Stephane Senecal (Orange Labs, France);
Marceau Coupechoux (Telecom ParisTech, France); Constant Bridon (ENS Cachan, France).
-9-
FAU, Germany
NTNU, Norway
Introduction
Load Modulated MIMO
1 / 18
Ralf Mller (FAU & NTNU)
Load Modulated MIMO
12-Dec-2014
2 / 18
Increase user density by increasing the number of antennas at the same site.
Massive MIMO philosophy:
Increase user density by partitioning of cells into subcells, i.e. increasing the
number of base stations.
Small Cell philosophy:
Massive MIMO vs. Small Cells
This work was supported in part by the EU-FP7 project
12-Dec-2014
Constantinos Papadias, Vlasis Barousis

AIT, Greece
12-Dec-2014
Georg Fischer
Mohammad A. Sedaghat
joint work with
Friedrich-Alexander Universitt Erlangen-Nrnberg (FAU)

Norwegian University of Science & Technology (NTNU)
Ralf R. Mller
Load Modulated MIMO:

A New Hardware Concept to Reduce
Cost, Size, and Amplifier Back-off
Introduction
Load Modulated MIMO
12-Dec-2014
2 / 18
Load Modulated MIMO
Path loss overcome by array gain.
Larger distance to the users.
12-Dec-2014
2 / 18
Remote radio heads.
More base station sites required.
Larger distance to the users.
More base station sites required.
Introduction
Massive use of bandwidth

Large processing gain
Introduction
Load Modulated MIMO
Massive use of bandwidth

Large processing gain
Massive use of antenna elements

Large array gain
Load Modulated MIMO
Both systems can operate in arbitrarily strong noise and interference.
Massive MIMO:
Spread spectrum:
Massive MIMO mimics the idea of spread spectrum.
Massive MIMO
Spread spectrum:
Massive MIMO mimics the idea of spread spectrum.
Massive MIMO
Introduction
12-Dec-2014
12-Dec-2014
3 / 18
3 / 18
Introduction
Load Modulated MIMO
Load Modulated MIMO
For cell densification transmit arrays should be small.
Remote Radio Heads (RHHs)
For cell densification transmit arrays should be small.
Remote Radio Heads (RHHs)
Introduction
12-Dec-2014
Can massive arrays

be compact and
cost-effective?
Is massive MIMO
compatible with
RRHs?
12-Dec-2014
4 / 18
4 / 18
Introduction
Load Modulated MIMO
12-Dec-2014
Size issue:
Load Modulated MIMO
12-Dec-2014
For a large number of antennas, the hardware cost becomes significant.

Standard approaches require one linear amplifier and mixer per transmit
antenna.
RF-costs limit the number of antenna elements and thus the overall system
performance.
Cost issue:
Open Issues
Size issue:
Cost issue:
Open Issues
Introduction
5 / 18
5 / 18
Promises
Load Modulated MIMO
Promises:
Load Modulated MIMO
Load Modulated Massive MIMO
Antenna coupling prohibits dense element spacing.
Size issue:
12-Dec-2014
12-Dec-2014
Standard approaches require one linear amplifier and mixer per transmit
antenna.
RF-costs limit the number of antenna elements and thus the overall system
performance.
For a large number of antennas, the hardware cost becomes significant.
Cost issue:
Open Issues
Introduction
6 / 18
5 / 18
Promises
Load Modulated MIMO
12-Dec-2014
6 / 18
Load Modulated MIMO
No D/A-converters required (digital RF).
12-Dec-2014
6 / 18
MIMO transmitters can be implemented with a single amplifier that serves all
antenna elements.
Promises:
antenna elements.
Promises:
Promises
Promises
Load Modulated MIMO
12-Dec-2014
6 / 18
Load Modulated MIMO
12-Dec-2014
The amplifier efficiency is the higher the more antenna elements are used.
6 / 18

The single amplifier can operate at constant envelope with high efficiency
despite arbitrarily (e.g. Gaussian) distributed signals on each antenna element.
antenna elements.
Promises:

antenna elements.
Promises:
Promises
General Concept
Load Modulated MIMO
9W
=W
Load Modulated MIMO
!"#$""%
LW
Voltage vs. Load Modulation
Sub-/2 spacing due to digital mitigation of mutual coupling.
12-Dec-2014
12-Dec-2014
7 / 18
6 / 18

The amplifier efficiency is the higher the more antenna elements are used.
antenna elements.
Promises:
Promises
General Concept
Load Modulated MIMO
!"#$""%
LW
+ Source perfectly matched

Source driven with back-off
V (t) s(t)
Z (t) = const.
Voltage modulation:
9W
=W
Load Modulated MIMO
!"#$""%
LW
9W
=W
General Concept
V (t)
Z (t) + R
V (t)
Z (t) + R
12-Dec-2014
must be proportional to the

data bearing signal s(t).
i(t) =
The current
12-Dec-2014

i(t) =
The current
7 / 18
7 / 18
V (t)
Z (t) + R

i(t) =
The current
General Concept
Load Modulated MIMO
V(t)
2
V(t)
1
iN(t)
i2(t)
i1(t)
Antenna
Antenna
Antenna
FWBMVBUJPO
Load Modulated MIMO
The hardware effort scales linearly with

the number of antennas.
Linear amplifiers with significant backoff areFWBMVBUJPO
required.
R
V(t)
1
Voltage modulation
12-Dec-2014
12-Dec-2014
8 / 18
7 / 18
Nonlinear relation between s(t) and Z (t)
Source in constant mismatch
+ Source with constant envelope
V (t) = const.
1
s(t)
Z (t) + R
Load modulation:
!"#$""%
LW
Voltage vs. Load Modulation in Massive MIMO
Source driven with back-off
+ Source perfectly matched
Z (t) = const.
V (t) s(t)
Voltage modulation:
9W
=W
General Concept
V(t)
2
V(t)
1
iN(t)
i2(t)
i1(t)
Antenna
Antenna
Antenna
Implementation
FWBMVBUJPO
V
Z1(t)
Z2(t)
Z1(t)
R
Antenna
Antenna
i1 (t) Antenna

R
i2(t)
i1(t)
Load modulation
FWBMVBUJPO
8 / 18
=(t)
Antenna
Load Modulated MIMO
12-Dec-2014
9 / 18
The load modulators shall use only reactive components, e.g. banks of PIN-diodes
acting as variable capacitors, to avoid power dissipation into heat.
12-Dec-2014
The impedance seen by the source converges to a constant as N due

to the law of large numbers. Arbitrary
good source matching is possible.
Load Modulated MIMO
Lossless Load Modulators
The hardware effort scales linearly with

the number of antennas.
Linear amplifiers with significant backoff areFWBMVBUJPO
required.
V(t)
1
Voltage modulation
Voltage vs. Load Modulation in Massive MIMO
General Concept
Antenna
L1
D1
D3
L2
L3
D3
Load
Modulator
io1
"
&
'
Implementation
Load Modulated MIMO
Load Modulated MIMO
The data signal does not pass through the power amplifier (PA).
Block Structure
(SAW drift can be considered part of the channel.)
12-Dec-2014
12-Dec-2014
Pulse shaping after the load modulator by surface acoustic wave (SAW) filters.
=(t)
10 / 18
9 / 18
The load modulators shall use only reactive components, e.g. banks of PIN-diodes
acting as variable capacitors, to avoid power dissipation into heat.
Lossless Load Modulators
Implementation
w

"
%
&
'
(
1
2
6
.
Implementation
Load Modulated MIMO
"
&
'
" "
!"
&
'
!"
!"
!"
8029/-992.
!"$
!"
!"%0+:;<<;9=0<,3>2>;:;/?
Load Modulated MIMO
12-Dec-2014
Let the current per antenna be complex Gaussian distributed. For N transmit
antennas, the transmitted power is 2 distributed with 2N degrees of freedom.
!"
12-Dec-2014
Example: OFDM Modulation, CB or ZF precoding
The data signal does not pass through the power amplifier (PA).
The PA provides kinds of RF-power sockets where the antennas are plugged.
Example: OFDM Modulation, CB or ZF precoding
Implementation
+,-./012+/3,04567
10 / 18
10 / 18
Cost Reduction
Load Modulated MIMO
12-Dec-2014
11 / 18
Load Modulated MIMO
12-Dec-2014
In the following example, we will compare load and voltage modulation at

identical radiated power for 100 antenna elements.
11 / 18
In 3G and LTE, the installed peak power is much higher than the average radiated
power.
installed peak power (hardware cost)
In 3G and LTE, most of the consumed power is dissipated into heat.
power consumption (radiated power and dissipated power)
Two cost factors:
Cost Reduction
In 3G and LTE, the installed peak power is much higher than the average radiated
power.
installed peak power (hardware cost)
In 3G and LTE, most of the consumed power is dissipated into heat.
power consumption (radiated power and dissipated power)
Two cost factors:
Cost Reduction
Cost Reduction
20
Average power
40
60
Output power (%)
0
0
10
20
30
40
50
60
70
80
40
60
Output power (%)
Crest factor=8dB
80
100
Load Modulated MIMO
Maximum power ~ eff.=78%
20
Average power
100
Load Modulated MIMO
Cost Reduction
Crest factor=8dB
80
Cost Reduction
Maximum power ~ eff.=78%
Amplifier Efficiency
0
0
10
20
30
40
50
60
70
80
Amplifier Efficiency
Power efficiency (%)

Power efficiency (%)
12 / 18
12-Dec-2014
F = 80%.
For class F amplifiers,

the efficiency is
12 / 18
For amplifiers of class A or B,

the efficiency is given by
Pmean
A/B =
4 Pmax
12-Dec-2014
For amplifiers of class A or B,

the efficiency is given by
Pmean
A/B =
4 Pmax
Size Reduction
Load Modulated MIMO
Multiple-PAs
100
Class A/B
8 dB
31%
1.9 W
4.2 W
6.1 W
12.0 W
Single-PA
100
Class F
1.2 dB
61%
1.9 W
1.2 W
3.1 W
2.5 W
12-Dec-2014
Load Modulated MIMO
12-Dec-2014
Are there theoretical limits for close element spacing (aperture)?
Array Densification
49% consumed power reduction
Load modulation (single PA) gives

79% peak power reduction
Transmitter
# antennas
Amplifier
Crest
Efficiency
Radiated power
Dissipated power
Consumed power
Peak power
Consider 19 mW average radiated power per antenna element.
Example
Cost Reduction
14 / 18
13 / 18
Size Reduction
Load Modulated MIMO
12-Dec-2014
Load Modulated MIMO
YES for antenna gain!
NO for directivity! (1946)
12-Dec-2014
Array Densification
Array Densification
Size Reduction
14 / 18
14 / 18
Size Reduction
Load Modulated MIMO
12-Dec-2014
14 / 18
Load Modulated MIMO
12-Dec-2014
14 / 18
These two theoretical limits are neither related to frequency nor wavelength.
The maximum antenna gain is limited by the physical size of the antenna.
The maximum directivity does not depend on the physical size of the antenna.

Bad news for noise limited systems

Good news for interference limited systems
Array Densification


Array Densification
Size Reduction
Size Reduction
Load Modulated MIMO
12-Dec-2014
14 / 18
Load Modulated MIMO
YES!
(in particular, for voltage modulation)
Are there practical limits for close element spacing?
12-Dec-2014
14 / 18


Array Densification
Are there practical limits for close element spacing?


Array Densification
Size Reduction
=W

=W
1
9W

9W
1
LW
1
LW

LW

antennas
coupled
FWBMVBUJPO
Size Reduction
=W

=W
1
9W

9W
1
LW
1
LW

antennas
coupled
FWBMVBUJPO
N N matching network required.

FWBMVBUJPO
Infeasible
for large N.
= Antenna mismatch.
=W

9W

LW

LW

12-Dec-2014
=W
1
=W

=W

LW
1
LW

antennas
coupled
15 / 18
Load Modulated MIMO
12-Dec-2014
15 / 18
N 1 matching network required.

Load modulators as matching network.
Digital compensation of mutual coupling.
9W
MC in load modulated MIMO:
Load Modulated MIMO
Mutual coupling in standard MIMO:
Close Element Spacing

FWBMVBUJPO
Infeasible
for large N.
= Antenna mismatch.
=W

9W

Size Reduction
=W

=W
1
9W

9W
1
LW
1
LW

antennas
coupled
=W
1
=W

=W

LW

LW
1
LW

antennas
coupled
N 1 matching network required.

Load modulators as matching network.
Digital compensation of mutual coupling.
9W
MC in load modulated MIMO:
FWBMVBUJPO
Conclusions
15 / 18
12-Dec-2014
16 / 18
+ Class F amplifier sufficient

Nonlinear relation between signals
and loads
Load Modulated MIMO
Many linear amplifiers required
12-Dec-2014
+ Close element spacing possible

+ Low backoff
Load modulation:
Load Modulated MIMO
Matching network required (for

closely spaced antennas)
Source driven with high back-off
Voltage modulation:
Conclusions
Mitigation of mutual coupling by baseband processing.
Load modulated antennas can be spaced more closely without much effort.

FWBMVBUJPO
Infeasible
for large N.
= Antenna mismatch.
=W

9W

LW

Size Reduction
+ Low backoff
+ Class F amplifier sufficient
Nonlinear relation between signals
and loads
+ Close element spacing possible
Load modulation:
Related Work
Load Modulated MIMO
12-Dec-2014
Parasitic element
"
Active element

Load Modulated MIMO
Varactor

12-Dec-2014
Goal: Try to reduce the number of RF-chains and the size of the array:
ESPAR MIMO
17 / 18
16 / 18
Load modulation moves the border between digital and analog circuitry one step
further towards the antenna.
Load modulation is well-suited for massive MIMO and allows for dense arrays.
Many linear amplifiers required
Source driven with high back-off
Matching network required (for

closely spaced antennas)
Voltage modulation:
Conclusions
Conclusions
=W
1
=W

=W

LW

LW

LW
1
LW

Related Work
FWBMVBUJPO
Load Modulated MIMO
antennas
coupled
12-Dec-2014
=W
1
=W

=W

FWBMVBUJPO
9W

LW
1
LW

Load Modulated MIMO
certain
antennas
coupled
17 / 18
12-Dec-2014
17 / 18
A tested prototype exists. blabla
ESPAR MIMO
FWBMVBUJPO
9W

ESPAR MIMO
Related Work
=W
1
=W

=W

LW

LW
1
LW

certain
antennas
coupled
Literature
FWBMVBUJPO
Load Modulated MIMO
12-Dec-2014
Load Modulated MIMO
12-Dec-2014
R. C. Hansen.
Electrically Small, Superdirective, and Superconducting Antennas.
John Wiley & Sons, 2006.
RM, M. A. Sedaghat, R. Fischer.

Load Modulated Massive MIMO.
Proc. IEEE GlobalSIP, Atlanta, GA, Dec 2014.
V. I. Barousis, M. A. Sedaghat, RM, C. Papadias.

Massive Antenna Arrays with Low Front-End Hardware Complexity: An
Enabling Technology for the Emerging Small Cell and Distributed Network
Architectures.
arXiv:1407.7733, Jul 2014.
18 / 18
17 / 18
A tested prototype exists. blabla

ThisFWBMVBUJPO
is a valuable alternative if the number of antenna elements is not very large.
9W

ESPAR MIMO
Related Work
Globecom 2014 Workshop - Wireless optical network convergence in support of cloud architectures
Optimized Beamforming and Backhaul Compression

for Uplink MIMO Cloud Radio Access Networks
Yuhan Zhou and Wei Yu
Department of Electrical and Computer Engineering
University of Toronto, Toronto, Ontario, Canada
Email: {yzhou, weiyu}@ece.utoronto.ca
AbstractThis paper studies the optimization of transmit

beamforming and backhaul compression strategies for the uplink
of cloud radio access networks (C-RAN), in which multi-antenna
user terminals communicate with a cloud-computing based central
processor (CP) through multi-antenna base-stations (BSs) serving
as relay nodes. The BSs perform compress-and-forward strategy
to quantize the received signals and send the quantization bits
to the CP via capacity-limited backhaul links for decoding. In
contrast to the previous works on the uplink C-RAN, which mostly
focus on the backhaul compression strategies only, this paper proposes the joint optimization of the transmit beamformers and the
quantization noise covariance matrices at the BSs for maximizing
the benefit brought by the C-RAN architecture. A weighted sumrate maximization problem is formulated under the user power
and backhaul capacity constraints. A novel weighted minimummean-square-error successive convex approximation (WMMSESCA) algorithm is developed for finding a local optimum solution
to the problem. This paper further proposes a low-complexity
approximation scheme consisting of beamformers matching to the
strongest channel vectors at the user side along with per-antenna
scalar quantizers with uniform quantization noise levels across
the antennas at each BS. This simple separate design strategy
is derived by exploring the structure of the optimal solution to
the sum-rate maximization problem under successive interference
cancellation (SIC) while assuming high signal-to-quantizationnoise ratio (SQNR). Simulation results show that with optimized
beamforming and backhaul compression, C-RAN can significantly
improve the overall performance of wireless cellular networks.
With SIC, the proposed separate design performs very close to the
optimized joint design in the SQNR regime of practical interest.
I. I NTRODUCTION
To meet the exponential increasing data demand in wireless communication driven by smartphones, tablets, and video
streaming, modern cellular systems are moving towards heterogenous networks consisting of base-stations (BSs) deployed
with high density, covering progressively smaller areas. As a
consequence, inter-cell interference management becomes a
challenging problem. Cloud radio access network (C-RAN)
is a promising network architecture for dealing with this
interference problem. In a C-RAN architecture, the encoding/decoding functionalities of the BSs are migrated to a cloudcomputing based central processor (CP). By taking advantage
of high backhaul links between the CP and the BSs, C-RAN
allows coordination and joint signal processing across multiple
cells, and the possibility for inter-cell interference cancellation,
which has the potential to significantly improve the overall
throughput of the wireless cellular network [1].
978-1-4799-7470-2/14/$31.00 2014 IEEE
One of the main problems arising in the practical implementation of C-RAN is how to optimally utilize the backhaul when
the backhaul link capacities are finite. Substantial research
works have made progress towards this direction, either from
information theoretic point of view [2][4] or algorithm design
point of view [5], [6]. In [7], a virtual multiple access channel
(VMAC) scheme, which is a two-stage compress-and-forward
strategy based on successive decoding, is proposed for the
single-input single-output (SISO) C-RAN architecture. It is
shown that the VMAC scheme can achieve the sum capacity of
C-RAN to within a constant gap under a sum backhaul capacity
constraint.
This paper studies the linear transceiver and backhaul compression design in the VMAC scheme for the uplink multipleinput-multiple-output (MIMO) C-RAN model as shown in
Fig. 1. As a generalization of [7] which considers the SISO
case only, this paper considers the MIMO case where both the
users and the BSs are equipped with multiple antennas. The
main difference between the SISO case and the MIMO case is
the impact of transmit optimization at the user terminals. In the
SISO case, since most of the intra-cluster interference has been
nulled by multicell decoding, it is near optimal for the users
to transmit with their maximum powers. In the MIMO case,
the users are capable of doing transmit beamforming, so the
optimal transmit beamforming design is more involved. This is
especially so when the backhaul links in C-RAN have limited
capacities.
The backhaul compression problem for the uplink C-RAN
model has been considered extensively in the literature. Various
algorithms such as the gradient projection method [5], the
robust backhaul compression approach [6], and the alternating
convex optimization approach [7] are developed for maximizing the (weighted) sum rate under the bachaul constraints.
All of these algorithms focus only on the optimization of
quantization noise covariance matrices across the BSs, while
keeping the transmit beamformers fixed. This paper goes one
step further by considering the joint transmit beamformer and
quantization noise covariance matrix optimization problem.
Accounting for both the transmit beamforming and the quantization design problem together in the optimization framework
is nontrivial because the two are coupled through the backhaul
constraints. To tackle this problem, this paper proposes a
weighted minimum-mean-square-error successive convex approximation (WMMSE-SCA) algorithm for efficiently finding
1472
..
x1
z1
H11
H21
z2
..
x2
1
y1 : y
H12
H22
is given by xk = Vk sk with E[sk sH

k ] = I. The transmit
beamformers
are
subjected
to
per-user
power constraints, i.e.,
(
)
Tr Vk VkH Pk for k = 1, 2, . . . , G. The received signal at
BS i, yi , can be expressed as
C1
2 C2
y2 : y
1
y
2
y
H1G
H2G
HLG
Hik Vk sk + zi
for
i = 1, 2, . . . , L,
where sk CN (0, I) is the intended signal vector for user k,

and zi CN (0, i2 I) represents the additive white Gaussian
noise at BS i. Based on rate-distortion theory, the quantized
received signal y
i for the ith BS is given by
L
yL : y
k=1
Central
Processor
L
y
HL1
HL2
..
xG
yi =
CL
zL
y
i = yi + qi
Fig. 1. Uplink of a cloud radio-access network with capacity-limited backhaul
a local optimum solution to the weighted sum rate maximization problem. Moreover, a simple approximation scheme
consisting of transmit beamformers matching to the strongest
channel signal vector and per-antenna scalar quantizers with
uniform quantization noise levels across the antennas at each
BS is also developed for the sum-rate maximization problem
when the signal-to-quantization-noise ratio (SQNR) is high and
when successive interference cancelation (SIC) is applied at
the receiver. This low-complexity scheme designs the transmit
beamforming and the backhaul compression separately. Numerical simulations show that with SIC implemented at the
receiver the proposed separate design is near optimal in the
SQNR regime of practical interest.
where qi CN (0, KQi ) represents the additive Gaussian

quantization noise.
Following the results in [7], [8], assuming that the linear
minimum-mean-squared-error (MMSE) receiver is applied at
the CP, the transmission rate Rk for user k for the VMAC
scheme is given by

1

Rk = log I + VkH HH
(2)
k Jk Hk Vk
where
Jk = JLE
k =
Hj Vj VjH HH
j + Kz + KQ ,
(3)
j=k
with Kz = diag(i2 I) and KQ = diag(KQi ). To achieve

higher throughput, the SIC scheme can also be applied. In this
SIC
case, the matrix JLE
expressed as
k is replaced by Jk
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION

We consider the uplink of a C-RAN, where G mobile users
communicate with a CP through L BSs, as shown in Fig. 1.
The BSs are connected to the CP via noiseless backhaul links
of finite capacities Ci . Each user terminal is equipped with M
antennas, and each BS is equipped with N antennas. In this
paper, we assume that the user scheduling is fixed, and perfect
channel state information (CSI) is made available to all the
BSs and to the CP.
The VMAC scheme with single-user compression is applied in the C-RAN system, where the ith BS quantizes the
received signal using single-user compression, then forwards
the compressed bits to the CP for decoding. By single-user
compression, we mean that the compression process only
involves a conventional vector quantizer, and it does not utilize
the correlation between the received signals across the BSs. At
the CP side, the quantization codewords are first decoded, then
the user messages are decoded sequentially. Define Hik as the
N M complex channel matrix between the kth user and the
ith BS, the channel matrix from user k to all the BSs is given
as the N L M matrix, i.e., Hk = [HT1k , HT2k , . . . , HTLk ]T .
It is assumed that each user intends to transmit d parallel
data streams to the CP. Let Vk CM d denote the transmit
beamfomer that user k utilizes to transmit signal sk Cd1
to the central receiver. Then the transmit signal at user k
(1)
Jk = JSIC
=
k
Hj Vj VjH HH
j + Kz + KQ .
(4)
j>k
with Kz = diag(i2 I) and KQ = diag(KQi ). The compression

rates at the BSs should also satisfy the backhaul link capacity constraints. Based on the vector quantization theory, the
backhaul constraints can be expressed as follows

G

2
+
I
+
K
k=1 Hik Vk VkH HH
Qi
i
ik
log
Ci
(5)
|KQi |
for i = 1, 2, . . . , L, where Hik is the complex channel gain
matrix between kth user and ith BS.
To characterize the tradeoff between the achievable rates for
the users and the system resources, we formulate the following
weighted sum-rate maximization problem:
1473
max
Vk ,KQi
s.t.

1

k log I + VkH HH
k Jk Hk Vk
(6)
k=1
Jk = JLE
or Jk = JSIC
k

k

G
2
+
I
+
K
k=1 Hik Vk VkH HH
Qi
i
ik
Ci ,
log
|KQi |
KQi 0, for i = 1, 2, . . . , L,
(
)
Tr Vk VkH Pk , for k = 1, 2, . . . , G,
where k s are the weights representing the priorities associated with the mobile users.
Due to the non-convexity of both the objective function and
the backhaul capacity constraints in problem (6), finding the
global optimum solution of (6) is challenging. We point out
that the present formulation (6) can be easily extended to a
more general case where the user scheduling strategy is also
included. More specifically, one can consider a weighted sum
rate over all the users in the network, where the beamformers
for the users are set to be the zero vector if they are not
scheduled. In this paper, we focuses on the active uses only
and assume that the user scheduling is done prior to solving
problem (6).
for i = 1, 2, . . . , L. It is not hard to see that the original

backhaul constraint (5) is always feasible when the convex
constraint (8) is feasible. The two constraints are equivalent
when
G
2
Hik Vk VkH HH
(9)
i =
ik + i I + KQi .
k=1
Now we approximate the objective function in (6) using

the WMMSE approximation. Let Uk CN Ld be the linear
receiver applied at the CP for recovering sk . The transmission
rate Rk in (2) can be expressed as the following [9] [10],
Rk = max log |E1
k |
where
III. P ROPOSED S OLUTIONS
H
H
Ek = (I UH
k Hk Vk )(I Uk Hk Vk )
Uk .
+ UH
Hj Vj VjH HH
k
j + Kz + KQ
A. The WMMSE-SCA Algorithm

This section presents a novel algorithm which is able to find
a stationary point of the problem (6). The main difficulties
in solving (6) come from the fact that the objective function
and backhaul capacity constraints are both nonconvex functions with respect to the optimization variables. Inspired by
the recent work on the WMMSE approach for beamforming
design [9], [10], we first reformulate the objective function
in problem (6) as a convex function with respect to the
MMSE-matrix given by the users target signal sk and decoded
signal sk assuming that the MMSE receiver is used. We then
linearize the convex objective function and the compression
rate expressions in the backhaul constraints of (6) to obtain
a convex approximation of the original problem. Finally we
successively approximate the optimal solution by optimizing
this convex approximation. The idea of convex approximation
roots from modern optimization techniques including block
successive minimization method and minorize-maximization
algorithm, which have been previously applied for solving
related problems in wireless communications [11], [12].
Before presenting the proposed algorithm, we first state the
following lemma, which is a direct consequence of concavity
of the log | | function.
Lemma 1: For positive definite Hermitian matrices ,
CN N ,
(
)
log || log || + Tr 1 N
(7)
j=k
By applying Lemma 1 again, we rewrite rate expression (10)

as
Rk = max (log |Wk | Tr(Wk Ek ) + d)
Wk ,Uk
2
Hik Vk VkH HH
ik + i I + KQi ,
k=1
we can approximate the backhaul constraint (5) with the

following convex constraint:
(
( G
))
1
H H
2
log |i | + Tr i
Hik Vk Vk Hik + i I + KQi
k=1
log |KQi | Ci + N
(11)
where Wk is the weight matrix introduced by the WMMSE

method. The optimal Wk is given by
H
Wk = E1
k = I + (Uk ) Hk Vk .
(12)
where Uk is the MMSE receive beamformer given by
H
H
Hj Vj Vj Hj + Kz + KQ Hk Vk .
Uk =
(13)
j=k
Using (11) and (8) to replace the objective function and

the backhaul constraints in problem (6), we reformulate the
weighted sum-rate maximization problem as follows
max
Vk ,KQi ,Uk ,
Wk ,i
k (log |Wk | Tr(Wk Ek ))
k=1
+
s.t.
with equality if and only if = .

By applying Lemma 1 to the first log-determinant term in
the backhaul constraint (5) and by setting
=
(10)
Uk
log |i | +
i i F
( i=1 )
Tr 1
i i
(14)
log |KQi | Ci ,
KQi 0, for i = 1, 2, . . . , L,
(
)
Tr Vk VkH Pk , for k = 1, 2, . . . , G,
where
is some positive constant1 , Ci = Ci + N , and
G
2
i = k=1 Hik Vk VkH HH
ik + i I + KQi . Note that the last
term in the objective function which involves a summation of
Frobenius norms is a quadratic regularization item. It makes
problem (14) a strongly convex optimization with respect to
each optimization variable.
It is easy to verify that problem (14) is convex with respect
to any one of the optimization variables when the other
(8)
1474
1 In
the simulation part of this paper, is set to be 1000.
optimization variables are fixed. Specifically, when the other

variables are fixed, the optimal values of i , Wk , and Uk are
given by equations (9), (12), and (13) respectively. When i ,
Uk , and Wk are fixed, the optimal values of Vk and KQi are
the solution of the following optimization problem:
min
Vk ,KQi
s.t.
k=1
k Tr(Wk Ek ) +
i i F
(15)
i=1
(
)
Tr 1
i i log |KQi | Ci log |i | ,
KQi 0, for i = 1, 2, . . . , L,
(
)
Tr Vk VkH Pk , for k = 1, 2, . . . , G,
The above problem is convex over Vk and KQi , and can be

solved efficiently with polynomial complexity. We summarize
the proposed WMMSE-SCA algorithm as below:
Algorithm 1 WMMSE-SCA Algorithm
1:
2:
3:
4:
5:
6:
(
)
Initialize KQi = I, and Vk such that Tr Vk VkH = Pk .
G
2
i k=1 Hik Vk VkH HH
.
ik + i I + KQi )
(
1
H H
Uk
Hk Vk .
j=k Hj Vj Vj Hj + Kz + KQ
H
Wk I + Uk Hk Vk .
Fix i , Uk , and Wk , solve the convex optimization
problem (15). Set (Vk , KQi ) to be its optimal solution.
Repeat Steps 25, until convergence.
The WMMSE-SCA algorithm yields a nondecreasing sequence of objective values for problem (6). So the algorithm is
guaranteed to converge. Moreover, it converges to a stationary
point of the optimization problem. The convergence result is
stated in Theorem 2. Due to space limit, we omit the proof.
(0)
(0)
Theorem 2: From any initial point (Vk , KQi ), the limit
point (Vk , KQi ) generated by the WMMSE-SCA algorithm

is a stationary point of the weighted sum-rate maximization
problem (6).
B. Low-Complexity Beamforming and Quantization Noise Design Under High SQNR
Although locally optimal transmit beamformers and quantization noise covariance matrices can be found using the
WMMSE-SCA algorithm for any fixed user schedule, user priority, and channel condition, the implementation of WMMSESCA in practice can still be computationally intensive, especially when the channels are fast varying or when the scheduled
users in the time-frequency slots change frequently. In this
section, we aim at deriving near optimal transmit beamformers
and quantization noise covariance matrices in the high SQNR
regime. The main result of this section is that a simple separate
design which involves transmit beamformers matched to the
strongest channel singular vectors at the user side and scalar
quantizers with uniform quantization noise levels across the
antennas at each BS is approximately optimal for maximizing
the sum rate if an appropriate set of users are scheduled.
This leads to an efficient way for transmit beamforming and
backhaul compression design in the practical C-RAN systems.
The proposed approximate approach is derived by assuming

that SIC is implemented at the central receiver. Let H denote
the channel matrix between the users and the BSs and Gi
denote the channel matrix between the users and the ith BS
(which is assumed to be full rank). Denote Kxj = Vj VjH and
KX = diag(Kxj ). The sum-rate maximization problem can be
formulated as follows,

HKX HH + Kz + KQ
max
log
(16)
Kxj ,KQi
|Kz + KQ |

Gi KX GH + 2 I + KQi
i
i
s.t.
log
Ci
|KQi |
KQi 0, for i = 1, 2, . . . , L,
(
)
Tr Kxj Pj , for j = 1, 2, . . . , G,
where Kz = diag(i2 I), and KQ = diag(KQi ).
First, we provide a justification that the optimal quantization
noise levels should be set as uniform across the antennas
at each BS under high SQNR. Towards this end, we derive
the Karush-Kuhn-Tucker (KKT) condition for the optimization
problem (16) under the high SQNR assumption. To obtain the
KKT condition, form the Lagrangian

L(Kxj , KQi , i , j ) = log HKX HH + Kz + KQ
log |Kz + KQ |
+
i=1

2

i log Gi KX GH
i + i I + KQi
i=1
G
i log |KQi |
j Tr (Kxk )
(17)
j=1
where i is the Lagrangian dual variable associated with the

ith backhaul constraint, and j is Lagrangian multiplier for the
jth transmit power constraint.
Setting L/KQi to zero, we obtain the following optimality condition
(
)1 T ( 2
)1
Fi HKX HH + Kz + KQ
Fi i I + KQi
)1
(
1
2
+ i KQ
= 0 (18)
i Gi KX GH
i + i I + KQi
i
where Fi = [0, . . . , 0, IN , 0, . . . , 0] with only the ith N N
block being nonzero. It is easy to verify that the above
optimality condition can only be satisfied if 0 i < 1.
Furthermore, if the overall system is to operate at reasonably high spectral efficiency, where the received signal-tonoise ratios (SNRs) are high and the backhaul capacities
are large, we must have HKX HH + Kz + KQ Kz +
2
KQ and Gi KX GH
i +( i I + KQi KQi). Under this
1
high SQNR condition, HKX HH + Kz + KQ
0 and
(
)
1
2
Gi KX GH
+
I
+
K
0.
Then
the
optimality
conQi
i
(i 2
)1
1
dition becomes i I + KQi
i KQi , i.e.,
KQi
i
2 I
1 i i
(19)
where i [0, 1) is chosen to satisfy the backhaul capacity

constraints. The above result implies that uniform quantization
noise levels across the antennas at each BS are optimal at
1475
1.0
high SQNR, (although the quantization noise level may differ

from BS to BS depending on the background noise levels and
the backhaul constraints.) Note that this line of reasoning is
very similar to the corresponding condition for the SISO case
derived in [7].
Next, we consider the optimal transmit beamforming and
power allocation under high SQNR. Intuitively speaking, for
maximizing the sum rate each user should align its signaling direction with the strongest eigen-mode of the effective
channel and allocate power along this direction in a waterfilling fashion. Towards this end, we need to whiten the
combined quantization and background noise and interference,
then diagonalize the resulting channel to find its eigenmodes,
and iteratively perform the water-filling process among the
users [13]. As we see from (19), at high SQNR, the optimal
quantization noise covariance matrices are diagonal with entries
proportional to the background noise levels. Further, if we
choose d = min{M, N L/G}, i.e., if we let the total number
of user data streams be equal to the number of degrees of
freedom in the system, then multiuser interference would be
reasonably contained. In this case it is reasonable for each
user to transmit with maximum power and allocate the transmit
power uniformly among the transmit beamformers.
Based on the above intuition on beamforming design, we
propose a low-complexity beamformer design for C-RAN to
simply choose the transmit beamformers for each user to match
to its transmit channel. Specifically, for user k, first perform
singular value decomposition on Hk as Hk = k k H
k . Then
choose the transmit beamformer
Pk
Vk =
k (1 : d)
(20)
d
where k (1 : d) is a matrix formed by d columns of k
which corresponding to the d largest singular values of Hk .
This channel-matching beamformer together with per-antenna
scalar quantizer provide us a low-complexity approximation
scheme, which separately designs the transmit beamforming
and the backhaul compression. Simulation results in the later
section show that with SIC such a separate design performs
very close to the optimized joint design in the SQNR regime
of practical interest.
Define

G

2
+
I
+
I
k=1 Hik Vk VkH HH
i
i
ik
.
(21)
C(i ) = log
|i I|
We summarize the approximation scheme as below
Algorithm 2 Approximate Scheme
1: Perform
SVD on Hk , i.e., Hk = k k H
k . Set Vk =
Pk
d k (1
2:
3:
: d) for k = 1, . . . , G.
Use bisection in [min , max ] to solve for i in C(i ) =
Ci , for i = 1, . . . , L.
Set KQi = i I for i = 1, . . . , L.
Cumulative Distribution Function (CDF)
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Baseline: No central processor

Approx. scheme W/O SIC, C=120 Mbps
WMMSE-SCA W/O SIC, C=120 Mbps
Approx. scheme W/O SIC, C=320 Mbps
WMMSE-SCA W/O SIC, C=320 Mbps
0.2
0.1
0.0
0
Uplink User Rates (Mbps)
Fig. 2. Cumulative distribution of user rates without SIC

64
62
Sum Rate per Macro-Cell (Mbps)
60
58
56
54
52
50

Approx. scheme W/O SIC
WMMSE-SCA scheme W/O SIC
48
46
44
42
120
140
160
180
200
220
240
260
280
300
320
Average Backhaul Rate per Macro-Sector (Mbps)
Fig. 3. Per-cell sum rate vs. average per-sector backhaul capacity of the
proposed schemes without SIC.
IV. S IMULATION R ESULTS

In this section, the performances of the proposed WMMSESCA and approximation schemes are evaluated in a 19-cell 3sector/cell wireless network setup with central 7 cells (i.e., 21
sectors) forming a cooperating cluster. The users are randomly
located and associated with the strongest BS. Round-robin user
scheduling is used on a per-sector basis. Standard cellular
network parameters are used in simulation: the noise power
spectral density is 162 dBm/Hz; flat fading channel vectors
are chosen according to a distance-dependent path-loss model
128.1+37.6 log10 (d), where d is the distance in km, with 8 dB
lognormal shadowing and a Rayleigh component. The distance
between neighboring BSs is 500 m. An antenna gain of 14 dBi
is assumed. The maximum transmit power constraint for each
user is 23 dBm, and channel bandwidth is 10 MHz. Each BS
is equipped with N = 2 antennas; each user is equipped with
M = 2 antennas; G = N L users are scheduled simultaneously
in each time slot, where each user sends one data stream (i.e.,
d = 1) to the CP.
Fig. 2 and Fig. 3 compare the performance of the baseline
1476
the gap between the two proposed scheme vanishes. This

demonstrates the approximate optimality for separate design of
transmit beamforming and backhaul compression in the high
SQNR regime.
1.0
Cumulative Distribution Function (CDF)
0.9
0.8
0.7
0.6
V. C ONCLUSION
0.5
This paper studies transmit beamforming and backhaul compression design for an uplink MIMO C-RAN system. We propose a novel WMMSE-SCA algorithm to efficiently optimize
the transmit beamformer and quantization noise covariance
matrix for maximizing the weighted sum rate. Further, it is
shown that with SIC at the receiver, channel-matching beamforming with uniform power allocation at the users together
with scalar quantization with uniform quantization noise levels
across the antennas at each BS is near optimal for maximizing
the sum rate when SQNR is high. This leads to a lowcomplexity algorithm, with separate transmit beamforming and
backhaul compression design. Simulation results show that
with the optimized transmit beamformer and quantization noise
covariance matrix, the C-RAN architecture can significantly
improve the performance of wireless cellular networks, and
majority of the gain can be achieved by the proposed lowcomplexity scheme.
0.4
0.3

Approx. scheme, SIC, C=120 Mbps
WMMSE-SCA, SIC, C=120 Mbps
Approx. scheme, SIC, C=320 Mbps
WMMSE-SCA, SIC, C=320 Mbps
0.2
0.1
0.0
0
Uplink User Rates (Mbps)
Fig. 4. Cumulative distribution of user rates with SIC.

80
76
Sum Rate per Macro-Cell (Mbps)
72
68
64
60
56
R EFERENCES
Approx. scheme W/ SIC
WMMSE-SCA scheme W/ SIC
52
48
44
120
140
160
180
200
220
240
260
280
300
320
Average Backhaul Rate per Macro-Sector (Mbps)
Fig. 5. Per-cell sum rate vs. average per-cell backhaul capacity of the proposed
scheme with SIC.
system with the proposed beamforming and backhaul compression schemes implemented without SIC at the receiver. It
is shown that both of the WMMSE-SCA and approximation
schemes significantly outperform the baseline scheme without
multicell processing. The figure also shows that WMMSE-SCA
scheme always outperforms the approximation scheme when
SIC is not implemented. As the backhaul capacity increases, the
performance gap between these two schemes becomes smaller.
However, as shown in Fig. 3 such a gap does not vanish even
in the high SQNR regime, which demonstrates that the joint
design is strictly better than the separate design when the CP
does not implement SIC.
Fig. 4 and Fig. 5 show the performance of the proposed
schemes with SIC implemented at the receiver. As compared
to the case without SIC, the proposed WMMSE-SCA and
approximation schemes can achieve more gain on the overall
performance with SIC applied. The performance improvement
is more significant for the users with low rate (e.g. 10percentile users), as shown in 4. The figures also show that the
approximation scheme performs very close to the WMMSESCA scheme, especially when backhaul capacity is large. As
shown in Fig. 5, when the backhaul capacity goes to infinity,
[1] D. Gesbert, S. Hanly, H. Huang, S. Shamai, O. Simeone, and W. Yu,

Multi-cell MIMO cooperative networks: A new look at interference,
IEEE J. Sel. Areas Commun., vol. 28, no. 9, pp. 13801408, Dec. 2010.
[2] A. Sanderovich, S. Shamai, Y. Steinberg, and G. Kramer, Communication via decentralized processing, IEEE Trans. Inf. Theory, vol. 54,
no. 7, pp. 30083023, Jul. 2008.
[3] A. Sanderovich, O. Somekh, H. V. Poor, and S. Shamai, Uplink macro
diversity of limited backhaul cellular network, IEEE Trans. Inf. Theory,
vol. 55, no. 8, pp. 34573478, Aug. 2009.
[4] L. Zhou and W. Yu, Uplink multicell processing with limited backhaul
via per-base-station successive interference cancellation, IEEE J. Sel.
Areas Commun., vol. 31, no. 10, pp. 19811993, Oct. 2013.
[5] A. del Coso and S. Simoens, Distributed compression for MIMO
coordinated networks with a backhaul constraint, IEEE Trans. Wireless
Commun., vol. 8, no. 9, pp. 46984709, Sep. 2009.
[6] S.-H. Park, O. Simeone, O. Sahin, and S. Shamai, Robust and efficient
distributed compression for cloud radio access networks, IEEE Trans.
Veh. Commun., vol. 62, no. 2, pp. 692703, Feb. 2013.
[7] Y. Zhou and W. Yu, Optimized backhaul compression for uplink cloud
radio access network, IEEE J. Sel. Areas Commun., to appear 2014.
[8] , Approximate bounds for limited backhaul uplink multicell processing with single-user compression, in Proc. IEEE CWIT, Jun. 2013,
pp. 113116.
[9] S. S. Christensen, R. Agarwal, E. Carvalho, and J. M. Cioffi, Weighted
sum-rate maximization using weighted MMSE for MIMO-BC beamforming design, IEEE Trans. Wireless Commun., vol. 7, no. 12, pp. 4792
4799, Jul. 2008.
[10] Q. Shi, M. Razaviyayn, Z.-Q. Luo, and C. He, An iteratively weighted
MMSE approach to distributed sum-utility maximization for a MIMO
interfering broadcast channel, IEEE Trans. Signal Process., vol. 59,
no. 9, pp. 43314340, Sep. 2011.
[11] C. T. K. Ng and H. Huang, Linear precoding in cooperative MIMO
cellular networks with limited coordination clusters, IEEE J. Sel. Areas
Commun., vol. 28, no. 9, pp. 14461454, Sep. 2010.
[12] S.-H. Park, O. Simeone, O. Sahin, and S. Shamai, Joint precoding and
multivariate backhaul compression for the downlink of cloud radio access
networks, IEEE Trans. Signal Process., vol. 61, no. 22, pp. 56465658,
Nov. 2013.
[13] W. Yu, W. Rhee, S. Boyd, and J. M. Cioffi, Iterative water-filling
for Gaussian vector multiple-access channels, IEEE Trans. Inf. Theory,
vol. 50, no. 1, pp. 145152, 2004.
1477
Distributed Mobile Cloud Computing:

Joint Optimization of Radio and
Computational Resources
S. Sardellitti1 , S. Barbarossa1, G. Scutari2
1
Dept. of Information Engineering, Electronics and Telecommunications, Sapienza University of Rome, Rome, Italy
2
Dept. of Electrical Engineering, State University of New York at Buffalo, Buffalo, USA
AbstractWe consider a scenario composed by multiple mobile

users asking for computation offloading of their applications to
a set of cloud servers. A set of radio access points, small cell
base stations possibly coexisting with macro base stations, are
available to provide radio proximity access to fixed computational
resources. Our objective is to find the optimal assignment of
each mobile user to a cloud server through the most convenient
base station and, contextually, the optimal MIMO precoding
matrices and computational rates (virtual machines) to each
user, under latency constraints dictated by the users Quality
of Experience (QoE). The radio resources assigned to users
belonging to the same cell are orthogonal to each other, whereas
users of different cells might interfere against each other. The
latency constraint imposes a strict relationship between the time
spent for transferring the program execution from the mobile
device to the fixed server (and viceversa) and the time needed to
execute the computation. To properly exploit this relationship,
we formulate the computation offloading problem as a joint
optimization of the radio and computational resources, with
the objective of minimizing the overall energy consumption, at
the mobile terminal side, while meeting the latency constraints.
The resulting optimization problem is nonconvex in both the
objective function and in the constraints. Nevertheless, by hinging
on successive convex approximation techniques, we propose an
iterative algorithm able to converge to a local optimal solution
of the original nonconvex problem.
Index TermsComputation offloading, cloud computing, distributed resource allocation, successive convex approximation.
I. I NTRODUCTION
The main motivation underlying this work comes from two
big challenges of current development of wireless networks:
the scarcity of energy resources at the mobile terminals
necessary to run increasingly sophisticated applications under
latency constraints dictated by the users Qualty of Experience
(QoE) and the need of optimizing allocation of communication/computation resources in a dense network deployment
scenario. The way to overcome the limitations of cellphones
batteries is to offload computations from the mobile devices
to a cloud service provider. Computation offloading is a key
strategy to improve the computational capabilities of mobile
handsets or to reduce energy consumption at the mobile
terminal. A strong impulse to computation offloading has come
The work of Barbarossa and Sardellitti was funded by the European
Commission, within the 7th Framework Programme, Project ICT-TROPIC,
under grant no. 318784. The work of Scutari was supported by the USA NSF
under Grants CMS 1218717 and Career Award no. 1254739.
978-1-4799-7470-2/14/$31.00 2014 IEEE
through cloud computing. In particular, one of the key features

of cloud computing is virtualization, which makes possible to
run multiple operating systems and multiple applications over
the same machine (or set of machines), while guaranteeing
isolation and protection of the programs and their data. Several
offloading methods have been proposed in the computer science literature, mainly focusing on computational aspects (e.g.,
program/code partitioning, static versus dynamic offloading,
virtualization, etc.); examples are MAUI [1], and ThinkAir
[2]. The major limitations of current mobile cloud computing
[3] strategies over cellular networks are the mobile energy
consumption associated with the radio access as well as the
latency experienced in reaching the cloud server through a
wide area network (WAN). In general, in fact, accessing the
cloud through a WAN could make impossible, sometimes, to
meet the latency constraints dictated by the users QoE. At the
same time, the energy spent by the mobile terminal to access
the network through a macro base station might jeopardize
all the benefits offered by computation offloading. This is
particularly true for mobile users located at the boundaries of
macrocells. Clearly, providing a proximity radio access would
be beneficial, at least in terms of energy spent for accessing
the network. A possible way to alleviate this issue is to bring
the radio access closer to mobile users (MU) through the socalled cloudlets [4], where the radio access points are typically
Wi-Fi routers. However, WiFi does not guarantee any quality
of service and it implies the use of a double technology
in a cellular network (WiFi and LTE), which is somehow
undesired. Alternatively, in the European project TROPIC [5],
which constitutes the framework of this work, it has been
proposed to empower the small cell base stations, operating
with LTE technology, with additional cloud capabilities. This
idea takes advantage of a potential massive deployment of
small cell base stations, as a way to increase the spectral
efficiency of wireless networks through an intensive spatial
reuse of radio resources. Endowing small cell base stations
with cloud functionalities, albeit limited, gives rise to a truly
pervasive communication/computation system, able to improve
the capabilities of mobile devices enormously. Clearly, the
need of achieving an economy of scale in the deployment of
these computationally empowered base stations implies that
these base stations will have limited computational capabilities, well below typical those of cloud servers. However, we
1484
Fig. 1. Distributed cloud scenario.
need to take into account that the number of users served by a

small cell base station is also very limited because of the small
coverage radio. At the same time, properly interconnecting
these base stations gives rise to a federation of small clouds
that may possess, as a whole, a non negligible computational
capability.
The new network architecture is depicted in Fig. 1, reporting
a set of mobile users accessing cloud services through a
hierarchical system, comprising the big data center and a set of
distributed small cell cloud-enhanced e-Node B (SCceNB), or
a federation of SCceNBs, properly managed through what we
call a femto-cloud. This scenario represents a fully distributed
cloud computing system, where radio and computing resources
are orchestrated jointly in order to meet the users requirements
and take advantage of available resources and balance wireless
traffic and computational load across the system. In such a new
hierarchical architecture, the applications may run either on the
mobile handset, in a nearby SCceNB or federated SCceNB,
or in the distant cloud provider, depending on which of these
alternatives is more convenient for the mobile user. One of the
key features of the proposed scheme is scalability. The work in
this paper is formulated within the current European Project
TROPIC. However, while most of the activities in TROPIC
are concerned with a single cloud server, the emphasis in this
paper is the study of a multiuser system where multiple cloud
machines concur to serve a multitude of mobile users.
Clearly, the effectiveness of the offloading strategy depends
on both radio access and computational aspects; this calls for
a joint optimization of the radio and computational resources.
In [6], the authors proposed a method to jointly optimize
the transmit power, the number of bits per symbol, and the
CPU cycles assigned to each application, in order to minimize
the power consumption at the mobile side, under an average
latency constraint, in case of a single cell, interference-free,
SISO small-cell network, assuming a single cloud server. In
[7] it has been proposed a more general framework to find the
optimal radio/computational resources in a MIMO multi-cell
network wherein MUs of different cells may interfere against

each other and are served by a single cloud. A distributed
implementation across the cells of the resources allocation
algorithm proposed in [7] has been developed in [8]. In this
paper, we consider a much more general context, composed
of multiple cells and multiple cloud providers concurring
to serve multiple users. Each link is a MIMO channel. We
formulate the offloading problem as a joint optimization of
the mobile radio resourcesthe transmit covariance matrices
of the MUsand the computational resourcesthe CPU cycles/second assigned by the cloud to each MU. Furthermore,
we look for the best association of each mobile user to a
base station and to a cloud. The objective is to minimize
the overall energy consumption at the mobile sides, under
power budget and latency constraints (the latter constraint
is what couples the computation and communication optimization variables). As a by-product of this optimization, we
derive a novel handover strategy, where the association of a
mobile user to a base station comes to depend not only on
the wireless traffic, but also on the computational load of
the cloud servers closer (in terms of smaller latencies) to the
base stations involved. Differently from [6], the optimization
problem formulated in this work is nonconvex in both the
objective function and constraints. This creates of course more
difficulties with respect to [6]. Nevertheless, building on recent
Successive Convex Approximation (SCA) techniques [9][11]
and exploiting the structure of the problem, we propose an
iterative algorithm with provable convergence to local optimal
solutions of the nonconvex problem. The algorithm is also
suitable for a distributed implementation across the cells,
with limited signaling between the cells and the cloud server.
Numerical results show that the algorithm converges in a few
iteration to solutions having performance very close to the
solution achievable solving the combinatorial problem.
II. O FFLOADING
OVER FEMTOCLOUD NETWORKS
Let us consider a network composed of a set of Nc clouds,

Nb small cell access points (SCeNBs in LTE terminology)
and K mobile user equipments (MUEs). Each access point
n is connected to the cloud provider m through a backhaul
link assumed to be a high speed connection (e.g., fiber optics)
with latency TBnm . The k-th MUE and SCeNB n are equipped
with nTk transmit and nRn receive antennas, respectively. We
denote by I , {k : k = 1, . . . , K} the set of all the users.
In this scenario, each MUE k is willing to run an application
under a maximum latency Lk , dictated by the application.
From the offloading point of view, the application is characterized by a few features: i) the number bk of bits to be
transmitted from the mobile device to the remote server to
transfer the program execution; ii) the number of CPU cycles
k necessary to run the application; and iii) the number bok
of output bits (the result of the computation), to be sent
back to the mobile user. We assume that the program code is
available at the server, or otherwise that it can be downloaded
by the server through a high speed wired link. The mobile
user decides whether to offload computations or not depending
1485
on which strategy entails a lower energy consumption for

the mobile device. In case of offloading, the overall latency
experienced by the k-th MUE for accessing the network
through the base station n and being served by the cloud m
is given by
rx
knm = tkn + exe
mk + knm + TBnm
(1)
where: tkn is the time necessary to send the program state and
input (encoded with bk bits) necessary to transfer the program
execution from the k-th MUE to the n-th base station;
exe
mk = k /fmk
(2)
is the time for the server to execute k CPU cycles, having

denoted by fmk the number of CPU cycles/second of the
virtual machine running the application of user k over the
m-th cloud; rx
knm is the time necessary for the server to send
the result (encoded in bok bits) back to the k-th MUE. Finally,
the term TBnm is the delay over the backhaul used to transfer
the program state from the n-th base station to the m-th cloud.
Denoting by Qk the covariance matrix of the data transmitted
by user k and by Q , (Qk )kI the set of all covariance
matrices, the maximum achievable rate of MUE k accessing
the network through the n-th SCeNB is

1
rkn (Q) = log2 det I + HH
Qk Hkn
(3)
kn Rkn (Qk )
where Hkn is the channel matrix between the kth user and
the nth base station; the covariance matrix
X
Rkn (Qk ) = N0 I +
Hln Ql HH
(4)
ln
lI,l6=k
includes both noise and multi-user interference; Qk ,

(Ql )l6=k denotes the tuple of covariance matrices of all users
interfering with MUE k (i.e., all covariance matrices except
Qk ). Given each rkn (Q) the time tkn necessary for user k
to transmit the bk input bits of duration Tb to the n-th SCeNB
is
ck
(5)
tkn (Q) =
rkn (Q)
where ck = bk Tb . Our goal is to find the optimal strategy
to assign each user to a base station and to a cloud, in
order to minimize the overall energy consumption, under
latency constraints. The degrees of freedom are radio and
computational parameters, namely: i) the precoding matrix
Qk CnTk nTk for each user; ii) the number fmk of
CPU cycles/second for running the application of user k over
the m-th cloud; and iii) the rule for assigning each user
to a base station and then to a cloud. The assignment is
performed by selecting the binary values aknm {0, 1} for
k = 1, . . . , K, n = 1, . . . , Nb , m = 1, . . . , Nc , where the
subscripts k, n, and m denote, respectively, user, base station,
and cloud indexes. In principle, a user could be served by
multiple base stations, as in cooperative communications, and
by multiple clouds. However, this scenario would make the
overall computation management much more complicated. For
the sake of simplicity, we assume that, at each time frame, each
user is served by a single base station and a single cloud. As

a consequence, for each k, aknm = 1 if user k accesses the
network through the n-th BS and it is served by the m-th
cloud; all other values akn m , with n 6= n and m 6= m , are
set to zero.
Our objective is the minimization of a weighted
sum of the energies
spent by each mobile terminal:
PK
E(Q, a) ,
Ek (Q, ak ), with Ek (Q, ak ) =

k
k=1
Nb X
Nc
X
tr(Qk )
aknm tkn (Q), where a , (ak )kI , ak =
n=1 m=1
(aknm )n,m . The coefficients k are positive parameters that

could be varied dynamically to enforce some sort of fairness
among the users. From the expression of the overall latency
knm in (1), we may notice the interplay between radio
access and computational aspects. This motivates our joint
optimization of the radio resources, the transmit covariance
matrices Q , (Qk )kI of the MUEs and the computational
rates f , (fmk )m,k . The covariance matrices (or, equivalently
the precoding matrices) are constrained to respect a power
budget limit

Qk , Qk CnTk nTk : Qk 0, tr (Qk ) Pk ,
where Pk is the average transmit
Q power of user k. We will
denote by Q the joint set Q , kI Qk .
We are now able to formulate our optimization problem as
follows
min
Q,f ,a
s.t.
E(Q, a) ,
K
X
k Ek (Q, ak )
(P)
k=1
i) gknm (Q, fmk , aknm ) Lk , k, n, m

ii) tr(Qk ) Pk , Qk 0, k I
Nb
K X
X
iii) hm (f , a) ,
aknm fmk Fm , m, f 0
k=1 n=1
iv)
Nb X
Nc
X
aknm = 1, aknm {0, 1},
k, n, m
n=1 m=1
(6)
where
we
have
introduced
the
function

wk
ck
+ TBnm .
+
gknm (Q, fmk , aknm ) , aknm
rkn (Q) fmk
The constraints listed above have the following meaning:
i) the overall latency for each user k must be lower than the
maximum tolerable value Lk ;
ii) the total power spent by each user must be lower than its
total power budget Pk ;
iii) the sum of the computational rates fmk assigned by each
server cannot exceed the server computational capability Fm ;
iv) each mobile user should be served by one base stationcloud pair. Then to drive the solution towards the situation
where each user is served by a single base station and a
Nb X
Nc
X
single cloud, we enforce the constraint
aknm = 1,
n=1 m=1
for each k. For simplicity we have incorporated the term

rx
knm in the latency limit Lk .
Unfortunately, problem P is NP-hard in general, even if one
1486
focuses only on the optimization of the covariance matrices Q

More formally, for each user k we introduce the approxik (Zk ; Z ) defined as:
and CPU cycles f . Things are even more complicated because mation function E
of the integer nature of the coefficients aknm . Motivated by
Nb X
Nc
X
tr(Qk )aknm
k (Zk ; Z ) , ck
the difficulties of handling an NP-hard problem, possibly E
rkn (Qk , Qk )
in a scenario with a large number of users, we propose a
n=1 m=1

suboptimal optimization strategy, based on a specifically
aknm tr(Qk )
aknm tr(Qk )
+
+
tailored successive convex approximation, able to achieve
rkn (Qk , Qk ) rkn (Qk , Qk )
X
high-quality solutions for problem P. Let us start by using

Qk Ej (Z ), Qk Qk
+
the popular relaxation of the integer constraints, enabling the
jI,j6=k
binary variables aknm to belong to the following convex set
N
N
c
+k kQk Qk k2 +f k fk fk k2+a kak ak k2
b X
X
A = {(ak )kI : aknm [0, 1],
aknm = 1, k, n, m}.
(8)
n=1 m=1
Then, we can write the relaxed version of P as

min
E(Q, a)
s.t.
gknm (Q, fmk , aknm ) Lk ,
Q,f ,a
hm (f , a) Fm , m, f 0,
k, n, m
(7)
Q Q, a A.
We denote by X the feasible set of the optimization problem
in (7). Even by relaxing the binary variables a, problem in
(7) is still non-convex, since the objective function and the
constraints gknm and hm are not convex. In what follows,
we exploit the structure of (7) and building on some recent
advances on Successive Convex Approximation (SCA) techniques, as proposed in [10], [11], we devise an efficient iterative approximation algorithm converging to a local optimal
solution of (7).
III. A LGORITHMIC DESIGN
To solve the non-convex problem (7) efficiently, we adopt an
SCA-based algorithm where the original problem is replaced
by a sequence of strongly convex problems. To do this, we need
to find, at each iteration, a suitable convex approximation of
the nonconvex objective function E(Q, a) and the constraints
gknm (Q, fmk , aknm ), which are preliminarily discussed next.
where hA, Bi , tr(AH B) and Qk Ej (Z) denotes the conjugate gradient of Ej (Z) with respect to Qk evaluated at Z .
Note that in (8) we also added quadratic regularization terms,
making Ek (Zk ; Z ) uniformly strongly convex with respect
to Zk .
Based on each Ek (Zk ; Z ), given Z 0, we can now define
the candidate sum-energy approximation E(Z;

Z ) as
E(Z;
Z ) ,
K
X
k Ek (Zk ; Z ).
Note that this function enjoys some desirable properties such

as i) it has the same first-order behavior of the original
nonconvex function E(Q, a); ii) it is separable in the users
variables Zk and iii) it is uniformly strongly convex on X
(see [11] for further details).
B. Inner convexification of the constraints
Let us introduce in this section an inner convex approximation of the constraints gknm (Q, fmk , aknm ) around Q X ,
denoted by gknm (Q, fmk , aknm ; Qk ). To do so, let us exploit
first the concave-convex structure of the rate functions rkn (Q)
[cf. (3)]:
+
rkn (Q) = rkn
(Q) + rkn
(Qk ),
(10)
with
+
rkn
(Q) , log2 det Rkn (Qk ) + Hkn Qk HH
kn
(Qk ) , log2 det (Rkn (Qk ))
rkn
A. Approximant of E(Q, a)
Let Z , (Q, f , a), Z , (Q , f , a ) and Zk ,
c
(Qk , fk , ak ) with fk , (fmk )N
m=1 . Following [10], [11]
the main idea is to approximate around the current (feasible) iterate Z X the original nonconvex nonseparable
objective function E(Q, a) with a strongly convex func
tion, say E(Z;
Z ), that is separable in the MUEs variables and has the same first order behaviour of E(Q, a)
at Z . To motivate our choice observe preliminary that i)
for
P any
P given Qk = tQk each term Ek (Qk , ak ; Qk ) =
a
tr(Q
)
(Q
;
Q
)
is
biconvex
in
Q
k
k
k and
kn
k
n
m knm
convex
in
a
given
Q
;
and
ii)
the
other
terms
of
the
k
P
sum j6=k Ej (Qk ; aj , Qk ) are not convex in Qk . Thanks
to this structure a convex approximation of E(Q, a) can
be obtained for each MUE k by convexifying the term
E
by linearizing the non-convex part
Pk (Qk , ak ; Qk) and
,
Q
).
E
(Q
;
a
k
j
k
j6=k j
(9)
k=1
(11)
+
and Rkn (Qk ) defined in (4). Note that
Q rkn () and rkn () are
concave on Q and convex on Qk , m6=k Qm , respectively.
Using (10), and observing that at any feasible Q, f , a, it must
be rkn (Q) > 0 and fmk > 0 for all k, n and m, the constraints
gknm (Q, fmk , aknm ) in (7) can be rewritten as
+
gknm (Q, fmk , aknm ) = rkn
(Q) rkn
(Qk )
+ qknm (fmk , aknm )
(12)
where we introduce the convex function qknm (fmk , aknm ),

ck aknm fmk
. By retaining the convex
fmk (Lk TBnm aknm ) k aknm
part in (12) and linearizing the concave term rkn
(Qk ) we
obtain the inner convex approximation:
1487
+
gknm (Q, fmk , aknm ; Qk ) , rkn
(Q) rkn
(Qk )
E
X D
Qj rkn
(Qk ), Qj Qj + qknm (fmk , aknm ),
jI,j6=k
1
with Qj rkn
(Qk ) = HH
Hjn . Finally the
jn Rkn (Qk )
non-convex constraint hm (f , a) may be replaced by the following convex approximation
m (f , a; Z ) ,
h
Nb
K X
X
(aknm fmk + aknm fmk

).
(13)
k=1 n=1
C. Inner SCA algorithm

We can now introduce the proposed convex approximation
of the nonconvex problem (7) which consists in replacing the
nonconvex objective function E(Q, a) and the non-convex
constraints gknm , hm with their convex approximations. More
formally, given the feasible point Z X , we have
) , argmin E(Q,
f , a; Z )
Z(Z
Q,f ,a
s.t.
gknm (Q, fmk , aknm ; Qk ) 0, k, n, m

m (f , a; Z ) Fm , m
h
fmk 0, k, m
Q k Qk ,
k I,
aA
(P )
) , (Q(Z
), f (Z ), a
(Z )) the
where we denoted by Z(Z
unique solution of the strongly convex optimization problem.
The proposed solution method consists in solving iteratively
problem P , starting from a feasible Z0 . The formal description of the algorithm is given in Algorithm 1 below. Note that
in Step 2 of the algorithm we allow a memory in the update of
the iterate Z , (Q , f , a ). A practical termination criterion
in Step 1 is to stop the iterates when |E(Z+1 ) E(Z )| ,
where > 0 is the prescribed accuracy. The convergence proof
of Algorithm 1 is omitted because of the space limitation; it
consists in showing that all the conditions in [10, Th.1] are
satisfied.
Algorithm 1 : Inner SCA Algorithm for P
Data: Z0 , (Q0 , f 0 , a0 ) X ; { } (0, 1]; k , f > 0;
a > 0. Set = 0.
(S.1): If Z satisfies a suitable termination criterion, STOP
) , (Q(Z
), f (Z ), a
(Z )) [cf. P ];
(S.2): Compute Z(Z

) Z ;
(S.3): Set Z+1 = Z + Z(Z
(S.4): + 1 and go to (S.1).
IV. N UMERICAL
RESULTS
In this section we present some numerical results to assess

the performance of the proposed offloading strategy. Our
experiments are run under the following setup. We consider
a MIMO network with a number of base stations equal to
the number of clouds, i.e. Nb = Nc = 4. All transceivers
are equipped with nT = nR = 2 antennas. The other system
parameters are set as follows: Fm = 2 107 , Pk = 103 and
N0 = 1. In Fig. 2 we show an example of final base stations
assignment resulting as a solution of our algorithm, for K = 6
and Nb = Nc = 4. Note that the final base station selection
tends to assign each user to its nearest base station, while
distributing the computation among the clouds. Indeed, even if
we relaxed the binary values aknm to be real values belonging

to the interval [0, 1], in our numerical results we found that the
indexes assigning a user to a base station tend to be very close
to either 0 or 1, whereas the indexes assigning the user to a
cloud tend to assume intermediate values. This means that the
relaxed algorithm tends to a solution where each user accesses
the network through a single base station, but it is served by,
possibly, multiple clouds. This is indeed a quite interesting
result.
As mentioned in the introduction, our optimization strategy
provides also, as a by-product, a mechanism to handle handover, in case of low user mobility. As an example, with
reference to Fig. 2, we considered the case in which no users
are moving, except user 1, who is moving from nearby the
third base station towards the first base station. In Fig. 3, we
have plotted the energy consumption E1 of user 1, while it
moves, averaged over independent channel realizations, versus
the distance d1 between mobile user 1 and base station 3. In
the same figure, we report the result of the optimal method
solving a mixed integer programming problem, considering
all possible combinations of the integer values aknm , and
the results of our suboptimal approach, based on successive
convex approximations. Different curves refer to different total
numbers K of mobile users in the system. The two left and
right end-points refer to the situation where the user is close
to base station 1 or 3. We may observe that, as expected, the
energy consumption increases as the user moves away from
each base station. But the most striking result, shown in Fig. 3
is that the suboptimal approach provides results very close to
the benchmark algorithm, which has exponential complexity.
Finally, the slight increase of energy consumption with K is
the result of an increased interference level resulting from
an increasing number of users. The slight increase testifies
that the joint optimization is taking care of the interference
level properly. Clearly, the association of a mobile user to
a cloud depends on the radio load of each base station as
well as on the computational load of each cloud. Selecting
clouds non co-located with a base station may be a viable
solution if the backhaul link between a base station and the
cloud is sufficiently fast. To test this situation, we considered
again the scenario of Fig. 3, where user 1 is moving and the
backhaul link between base stations 1 and 3 is either congested
or relatively free. In Fig. 4 we have plotted the average global
energy consumption obtained by solving problem (7) in the
two extreme cases where the backhaul link among the two
clouds is congested or not. In the second case, the latency
along the backhaul is sufficiently low to essentially enable each
mobile to offload computation to any cloud in the network.
From Fig. 4, we may observe how the backhaul availability
brings, as expected, a network energy saving, as the energy
consumption increases with the number K of active users.
Finally, in Fig. 5 we report the average energy consumption
in the presence of K = 4 mobile users and a varying number
of base stations Nb . We compare again, the proposed suboptimal method with the optimal benchmark. We can see that, as
the number of base stations increase (denser deployment), the
1488
10
cloud2
MUE3
MUE4
BH12
BS2
cloud1
BH23
BH42
BH13
MUE2
MUE1
BS1
BH14
cloud4
cloud3
BH43
MUE5
10
K=4 congested backhaul

K=4 available backhaul
BS3
BS4
MUE6
10
0.5
1.5
2.5
d1
Fig. 2. Example of user-base stations assignment.
Fig. 4. Global energy consumption vs. d1 .
10
K =4
0.1
SCA algorithm
exhaustive search
0.09
0
10
0.08
E1
0.07
E1
0.06
1
K=4 SCA algorithm

K=4 exhaustive search
K=6 SCA algorithm
K=8 SCA algorithm
10
0.05
0.04
0.03
10
0.5
1.5
2.5
0.02
d1
0.01
2
Nb
Fig. 3. User energy consumption vs. d1 .
Fig. 5. Energy consumption of user 1 vs. the number of base stations.
increased number of degrees of freedom yields a considerable

energy saving. Furthermore, also in this case we can observe
how the proposed algorithm provides results very close to the
optimal ones.
V. C ONCLUSION
In conclusion, the joint optimization of base station/cloud
assignment, MIMO precoding and computational rates seems
to provide a powerful mechanism to handle computation
offloading. The proposed approach gives also a strategy to
handle hand-over, in case of low mobility, taking into account
not only the wireless traffic in each base station, but also the
computational load and the state of the backhaul connecting
different clouds.
R EFERENCES
[1] E. Cuervo, A. Balasubramanian, D. Cho, A. Wolman, S. Saroiu, R.
Chandra, and P. Bahl, MAUI: making smartphones last longer with code
offload, Proc. of the ACM Int. Conf. on Mobile Systems, Applications,
and Services, pp. 49-62, San Francisco, CA, USA, 15-18 June 2010.
[2] S. Kosta, A. Aucinas, P. Hui, R. Mortier, X. Zhang, ThinkAir: Dynamic
resource allocation and parallel execution in the cloud for mobile code
offloading, Proc. of INFOCOM 2012, pp. 945953, March 2012.
[3] N. Fernando, S. W. Loke, W. Rahayu, Mobile cloud computing: A
survey, Future Generation Computer Systems, vol. 29, pp. 84106, Jan.
2013.
[4] M. Satyanarayanan, P. Bahl, R. Caceres, N. Davies, The case for VMbased cloudlets in mobile computing, IEEE Pervasive Computing, pp.
1423, Oct.-Dec. 2009.
[5] FP7 European Project. (2012). Distributed computing, storage and radio
resource allocation over cooperative femtocells (TROPIC). [Online].
Available: http://www.ict-tropic.eu.
[6] S. Barbarossa, S. Sardellitti and P. Di Lorenzo, Joint allocation of
computation and communication resources in multiuser mobile cloud
computing, Proc. of IEEE 2013 Int. Workshop on Signal Process.
Advances in Wireless Commun. (SPAWC 13), Darmstadt, Germany, 16
19 June 2013.
[7] S. Sardellitti, G. Scutari, and S. Barbarossa,Joint optimization of radio
and computational resources for multicell MIMO mobile cloud computing, Proc. of IEEE 2014 Int. Workshop on Signal Process. Advances in
Wireless Commun. (SPAWC 14), Toronto, Canada, 2225 June 2014.
[8] S. Sardellitti, G. Scutari, and S. Barbarossa,Distributed joint optimization
of radio and computational resources for mobile cloud computing, Proc.
of IEEE Int. Conf. on Cloud Networking (CloudNet), Luxembourg, 810
October 2014.
[9] G. Scutari, F. Facchinei, P. Song, D.P. Palomar, and J.-S. Pang, Decomposition by partial linearization: Parallel optimization of multi-agent
systems, IEEE Trans. on Signal Process., vol. 63, no. 3, pp. 641656,
Feb. 2014.
[10] G. Scutari, F. Facchinei, L. Lampariello, and P. Song, Parallel and
distributed methods for nonconvex optimization, Proc. of IEEE Int. Conf.
on Acoustics, Speech, and Signal Process. (ICASSP 14), Florence, Italy,
49 May 2014,.
[11] G. Scutari, F. Facchinei, L. Lampariello, and P. Song, Parallel and
distributed methods for nonconvex optimization-Part I&II: Theory &
Applications, IEEE Trans. on Signal Process., (under review) 2014.
1489
An Improved Decoder for Cloud-Based Mobile

Networks under Imperfect Fronthaul
Jens Bartelt and Gerhard Fettweis
Vodafone Chair Mobile Communications Systems
Dresden University of Technology, Germany
Email: {jens.bartelt, gerhard.fettweis}@tu-dresden.de
AbstractA mobile network employing a so-called cloudbased architecture requires an additional fronthaul network to
forward digitized samples. With increasingly dense deployments
of small cells, the availability of high quality ber fronthaul
cannot be expected at all sites. Longer fronthaul distances or
the employment of wireless technologies renders fronthaul links
imperfect in the sense that they introduce bit-errors into the
digital samples. In this work, we build upon so-called hardware
noise resilient decoders to investigate decoders for imperfect
fronthaul that can achieve very good performance even under
high bit-error rates on the fronthaul.
Index TermsDecoding, soft information, cloud RAN, fronthaul, CPRI
Radio Access Link

Core
(erroneous) Fronthaul Link

Backhaul Link
User Equipment
Remote Radio Head
Baseband Unit (Cloud)
Core Network (EPC)
Fig. 1.
links.
Cloud-based mobile network architecture with imperfect fronthaul
I. I NTRODUCTION AND M OTIVATION

To satisfy the ever increasing demand in mobile trafc and
lower the cost for mobile networks, so-called cloud based
networks have recently received a lot of attention in industry as
well as academia [1], [2]. Such a network is depicted in Fig. 1.
The key idea of the cloud-networks is to ofoad some of the
processing that in a conventional network would be handled by
base stations to a centralized entity, referred to as the cloud.
The ofoaded processing can cover the whole protocol stack
and range from self-organizing network functionalities, radio
resource management and scheduling functionalities, down to
physical layer processing such as detection and decoding. In
the most extreme case, the base station is reduced to a socalled remote radio head (RRH) only comprising of antennas,
down-/upconverters, analog transceivers, and analog-to-digital
converters. The complete baseband processing is then handled
in the cloud, implementing a complete baseband unit (BBU).
The BBU and the RRH exchange baseband receive or transmit
signals in the form of digital In-phase/Quadrature-phase (I/Q)
samples via a fronthaul (FH) channel.
The promised advantages of such a cloud-architecture are
manifold and include cost savings through simplied base
stations and the use of off-the-shelf hardware in the cloud, joint
processing of multiple base stations through centralization,
and a higher utilization of the deployed hardware through
processor load balancing between BBUs.
These advantages come at the price of an additional transmission via the fronthaul channel. Recently, standardization
efforts have been made to standardize the transmission of
the I/Q samples, most prominently in the Common Public
Radio Interface (CPRI) consortium [3]. Most commonly, ber
links are used for fronthauling, which is then also referred
978-1-4799-7470-2/14/$31.00 2014 IEEE
to as digital Radio-over-Fiber. However, to satisfy the ver

increasing trafc demand, denser and denser deployments of
small cells have to be considered. These would be mounted
in locations like lamp posts and billboards where ber might
be not always be available. Therefore, high-capacity wireless
millimeter wave (mmWave) links have been proposed as an
alternative [4] and are also already available for commercial
use.
The CPRI standard requires the used FH technology to meet
very tight constraints, most notably a very low latency of 5 s
(without propagation) and bit-error rate (BER) of less then
1012 . Due to the strictness of these requirements, the FH
can almost be considered to be perfect, a term also used
in the 3GPP requirements for small cell backhaul [5]. While
these requirements are feasible for short range ber links, they
becomes more challenging at higher distances and for wireless
links. Especially outdoor mmWave links have to face varying
channel conditions due to rain and suffer from atmospheric
attenuation [6]. The only options to achieve the required BER
are to either transmit with sufciently high power to ensure a
high SNR, which reduces the energy efciency of the system,
or to ensure reliability by using an additional FH channel
code or automatic repeat request schemes, which introduce
additional latency.
To overcome these problems and to enable reliable and
highly energy efcient networks, we discuss in this work a
simple scheme that can improve the reliability of uplink communication in a cloud-based mobile network with imperfect
FH links, meaning that they have a BER much higher than
1012 . For this, we build on so-called error resilient decoders,
introduced in [7], [8]. We will recapture this previous work in
1478
Sec. II. as well as introduce our own system model. Sec. III.
describes the theoretic background and the improved decoder,
for which numerical results are given and discussed in Sec.
IV. The paper closes with our conclusions in Sec. V.
II. P REVIOUS W ORK AND S YSTEM M ODEL
In todays mobile networks, reliable communication is
ensured by modern forward error correcting (FEC) channel
codes such as turbo codes, which are, e.g., specied for the
3GPP LTE standard [9]. These codes can be decoded with
high performance thanks to the use of soft-information in
algorithms like the soft-Viterbi [10] or BCJR [11] algorithm.
The soft-information, often given in the form of log-likelihood
ratios (LLRs), is computed during demodulation/detection and
requires knowledge of the transmissions underlying random
process, which is usually assumed to be that of an additive
white Gaussian noise (AWGN) channel. This assumption
usually holds with sufcient precision, since other channel
effects of mobile channels such as multipath propagation and
Doppler effect are compensated by an equalizer.
In [12], the authors investigated how the distribution of
Gaussian received samples changes, if these sample are stored
in digital form in a faulty buffer memory. The background to
their work is that, due to aggressive voltage scaling in CMOS
technology, the samples will be incorrectly read from the
memory, resulting in uniformly distributed bit-errors. These bit
errors change the distribution of the samples in a deterministic
way. Building on this, the authors of [7], [8] developed error
resilient decoders that rely on a soft-information calculation
that no longer assumes a pure Gaussian, but the modied distribution. In this work, we now adapt their idea for application
in the uplink of cloud-based mobile networks.
RAN
UE
RRH

FH
BBU
Fig. 2.
Model for the uplink of a cloud-based mobile network.
Fig. 2 shows a discrete time baseband model for such a

network with one user equipment (UE), one RRH and one
BBU. The UE wants to send digital data in the form of a
bit-vector u = (u1 , , uk , , uK ) to the BBU. This data
is encoded in the encoder C1 with, e.g., a turbo code and
results in the bit-vector (v1 , , vl , , vL ). The resulting
code rate is K
L . This bit vector is modulated using the modulator M1 to one or multiple symbols x. These symbols are
transmitted over the radio access networks (RAN) channel,
which we model as an AWGN channel with noise variance
2 which results in an SNR of var(x)/ 2 . Note that, as stated
before, we assume other effects of a mobile network channel

to be perfectly compensated during equalization. The RRH
receives the noisy symbols x , which are quantized (analogto-digital converted) to an quantization index w that can be
represented in binary form as a vector of B quantization bits
w = (w1 , , wb , , wB ). The indices w are chosen from
an codebook C = {c1 , , cq , , c2B } of nite size 2B
according to some decision thresholds. Note that the quantization adds quantization noise and transforms the continuous
amplitude x into a discrete one. Numerous works deal with
the modeling of quantization errors and how to optimally
design a quantizer, e.g., the famous work of Lloyd [13].
However, our work does not deal with this effect, so for the
rest of the paper we assume that the bit-resolution B is chosen
high enough that the quantization noise can be neglected, and
the distribution of w can be assumed to be approximately
continuous.
The quantization bits of w are now sent over a digital
fronthaul channel. Depending on the technology used, this
might include an additional encoding, scrambling, modulation
or other transformation as well as the inverse operations at
the receiver. However, independent of the technology used
and the processing performed, this fronthaul transmission
will result in bit errors. We assume that these bit errors
can be modeled as being uniformly distributed on all bits,
which is a reasonable assumption because digital transmissions
usually employ scramblers that transform block errors to more
widely spread singular errors. The only inuence that the
technology and processing has is that it determines the bit error
probability, e.g., an additional channel code will decrease the
bit error probability. In conclusion, we model the fronthaul
channel as a bit symmetric channel (BSC) with a bit error
probability of Pe . This BSC is the equivalent of the erroneous
buffer memory in [8].
The bit-vector w is now received as an erroneous bit-vector
w at the baseband unit. There it is dequantized (digital-to
analog converted) by selecting a value from a dequantizer
codebook according to w as x = Q1 (w ). The dequantizer
codebook is known by the baseband unit and its design is
also part of designing an optimal quantizer, which, as pointed
out earlier, is not the focus of this paper. Note also, that the
reconstructed value would ideally be the received symbol x
but that it is distorted by quantization noise, making a perfect
dequantization impossible. Also, the wrong dequantization
value might be chosen because w is corrupted by bit errors.
The reconstructed received symbol x is now demodulated
by M 1 , meaning that the LLRs of its bits are calculated.
These LLRs v can then be fed into a soft-decision decoder
to determine the received user bits u . The core of developing
an improved decoder lies now within the calculation of the
LLRs. This is described in the next section.
III. I MPROVED D ECODING
In the following, we use the example of binary phase shift
keying (BPSK) modulation to illustrate the calculation of soft
1479
information, meaning that the bits 0 and 1 are mapped to

the amplitudes -1 and +1, respectively.
The received signal x after the AWGN channel follows the
conditional probability distribution
exp
(x 1)2
,
2 2
(1)
2 2
which is depicted as the green curve in Fig. 3. The corresponding LLRs are calculated as

f (x |x = 1)

.
(2)
L(x ) = ln
f (x |x = +1)
However, we want to determine the soft information not of x
but of x , which cannot be assumed to be Gaussian distributed
due to the effect of the BSC. To model its effect, we assume
in the following the use of a uniform linear mid-rise quantizer
with a step-width of x. The bit vector w is corrupted by
between 0 and B bit ips on the BSC, which means that after
dequantization the amplitude will be shifted depending on the
number and position of bit ips. For illustration, we assume
that the bit vector w received a single bit error on the b-th bit.
From here on we also skip the condition for x in our notation,
as the effect is similar for both values of x. The resulting
distribution can be calculated as [12]

f (x + 2b ), m even
1
(3)
fb (x ) =
f (x 2b ), m odd
with
m=
and
w
2Bb1
P (k) = Pek (1 Pe )(Bk)
(6)
so that we can write the new distribution that takes all possible
bit errors of one quantized received symbol into account as
f (x ) =
k=0
P (k)f (x ).
(7)
1.2
1
(5)
The parameter is the offset caused by the bit ip, e.g., a

ip in the sign bit causes a larger offset than a ip on a lower
value bit. The parameter m decides whether the shift is in the
upwards (ip 0 1) or downwards (ip 1 0) direction.
This is illustrated in Fig. 3 for a single bit ip on the fourth
bit (indicated in red) of a total of eight quantization bits. The
x-axis shows only the four most signicant bits. Note, how
for example the peak of the green curve, being represented by
the four bits 0101, is shifted to 0100.
The calculation for a larger number of bit ips is more
complicated as it depends on the exact combination of bit
ips. We refer the interested reader to the original works in
[12] for a more complete description. Here, we only want to
further note that each number of exactly k bit ips leads to a
distribution we write as f k (x ). The probability that exactly
k bit ips occur is given as
B

1.4
(4)
= 2B1 x.
These LLRs will then be fed into the decoder C 1 . In [8], it

was also noted that in the log-domain several simplications
can be made to lower the computational complexity of calculating the LLRs, which we will not reproduce here but which
are nonetheless applicable in our case as well.
PDF(x)
f (x |x = 1) =
Note that this also includes the possibility of having no bit

errors at all (k = 0). The authors of [8] noted, that for
Pe 1, the probabilities of multiple bit errors P (k) is small
compared to the probability of having no or a single bit error,
so that the distribution can be approximated by taking only a
single bit error per quantized received symbol into account.
As they showed, this has almost no inuence to the decoder
performance an thus, we follow this approach in our work
as well and use the approximate distribution from here on.
The resulting conditional distributions are shown in Fig. 4
as compared to the original one for Pe = 101 . As can be
easily seen, the distribution is quite different from the original
one, which means that a LLR calculation based on a purely
Gaussian distribution is widely inaccurate. We now improve
the decoding performance by calculating the LLRs according
to the new distribution as

f (x |x = 1)
.
(8)
L(x ) = ln
f (x |x = 1)
0.8
0.6
0.4
0.2
0
0001 0010 0011 0100 0101 0110 0111 1000 1001 1010
x
Fig. 3. Pure Gaussian distribution (green) as compared to the distribution if

the fourth bit has been ipped (blue)
IV. N UMERICAL R ESULTS

To show the performance of the improved decoder we
implemented the system model in Fig. 2 in Matlab using
QPSK modulation, the turbo code specied for the long term
evolution (LTE) standard in [9] with a rate matching module
to enable different code rates and a BSC with variable Pe .
The block length of the code was 256 and we simulated
the transmission of approximately K = 107 bits. We used
a uniform linear mid-rise quantizer with B = 8 bit resolution
and a step width of x = 10
28 .
A. Performance Improvements
The performance in terms of end-to-end BER (from u to
u ) is plotted in Fig. 5 for a varying SNR of the radio access
1480
1.4
f(x')
f(x'')
1.2
PDF(x)
0.8
0.6
0.4
0.2
0
-5
-4
-3
-2
-1
0
x
Fig. 4.
Pure Gaussian distribution of a BPSK signal (green) and after
forwarding the signal over a BSC with Pe = 101
channel and varying Pe s. The dashed lines show the improved

decoder as compared to the conventional decoder assuming
Gaussian distribution as solid lines for a code rate of 0.5. Note
that the black curve is the performance for Pe = 0, which
serves as baseline as it the same performance we would get
without an additional transmission or with a perfect fronthaul.
The error oor for the solid blue and red lines is a result of
the fact that the end-to-end transmission is limited by the BSC
and therefore cannot achieve a better performance for a higher
radio access SNR. Similar error oors can be expected from
the other curves, they are not visible here due to a limited
number of bits transmitted in the simulations.
0
10
-1
Pe=10 , conv.
-1
Pe=10 , adapted
-1
10
-2
Pe=5*10 , conv.l
-2
Pe=5*10 , adapted
BER
-2
10
-2
Pe=10 , conv.
-2
Pe=10 , adapted
-3
10
-3
Pe=10 , conv.
-3
Pe=10 , adapted
-4
Pe=0, conv.
10
4
6
SNR in dB
10
Pe=0, adapted
varying channel conditions in outdoor mmWave links. Thus,

it has to be measured or estimated in regular time intervals.
There are three principal possibilities for this:
Reference data
By sending out known data, the bit error rate of the
FH, which is identical to the bit error probability, can
be directly measured at the receiver. However, this
introduces overhead and thus, reduces the effective
throughput.
Reference symbols
Digital links usually require known reference symbols for, e.g., synchronization or channel quality
measurements for link adaptation. From the SNR
of a link the Pe can be indirectly derived by using
theoretic results for known modulation schemes.
Decode check
Depending on the code used, the number of bit
errors can be directly determined from the syndrome
calculation, or can be derived from an additional
CRC (circular redundancy check) code.
Any of this schemes is prone to a certain error so the calculated
distribution will never be exact. Therefore, we investigated
how such an error in Pe measurement would inuence the
performance of the improved decoder. For this, we simulated
the transmission with a certain Pe , but used a different
estimated Pe,est in the decoder. The results are displayed in
Fig. 6 for Pe = 5 102 and Pe,est ranging from 0.999 to
106 . To further illustrate the performance loss, we plotted
the difference in SNR required to achieve a BER of 104
over the normalized estimated error probability Pe,est /Pe in
Fig. 7.
We would like to point out that in the extreme cases of
Pe,est = {0, 1}, the SNR loss would be innite as in these
cases error oors appear above 104 which are not depicted
in the gures. As can be seen from the two gures, the
performance loss is very small when the estimation is within
one order of magnitude of the true Pe and even for four orders
of magnitude the loss is only about one dB. We conclude from
this result that the Pe does not need to be measured with great
precision, the knowledge of the order of magnitude should
sufce for good results.
C. Implications for Cloud-based Networks
Fig. 5. End-to-end bit error rate under different SNRs of the radio access
channel and different bit error probabilities of the fronthaul channel for a
conventional decoder assuming a Gaussian distribution (solid lines) and the
adapted decoder (dashed lines)
B. Performance under Estimated Bit-Error Probability

The improved decoder requires additional information compared to a conventional decoder. First, it requires knowledge
of the quantizer codebook C, however, this is also required
by a conventional system and usually does not change during
operation so it poses no new challenge to acquire this information. The second additional parameter, however, is the bit
error rate Pe of the FH channel, which can change due to, e.g.,
Our results qualitatively match those of [8], conrming

their ndings. However, we would like to point out that the
gains are quite signicant and of great importance for cloudbased networks. As can be seen, the curve for Pe = 103 is
almost identical to the baseline and even for Pe = 102 the
performance loss is less then 0.5 dB. This means that a Pe of
1012 , as required by the CPRI standard is in this case much
to strict and can be relaxed to about 103 .
Conventionally, low bit error probabilities can be achieved
by one of four methods: a more robust modulation, additional coding on the fronthaul, more transmit energy or lower
transmit distance. A more robust modulation reduces the
throughput, which is not desirable due to the high capacity
1481
10
-1
P e,est=P e=5 10
-2
BER
10
10
10
-3
P e,est=10
-2
P e,est=10
-3
P e,est=10
-4
P e,est=10
-5
P e,est=10
-6
BERs might still required if equalization or synchronization

is performed in the baseband unit (as actually intended in the
CPRI standard) or for, e.g., control signaling. Still, the given
results are quite remarkable and should be taken into account
when designing new fronthaul standards employing different
functional splits compared to CPRI as have been proposen in,
e.g., [16].
-2
V. C ONCLUSION
P e,est=0.999
-4
3.5
4.5
SNR in dB
5.5
Fig. 6. End-to-end bit error rate under different SNRs of the radio access
channel for the adapted decoder and Pe = 5 102 when Pe is worngly
estimated
1.4
1.2
SNR loss in dB
1
0.8
0.6
0.4
An imperfect fronthaul link introduces bit errors on digital

samples which modies the probability distribution of received
symbols in a deterministic way. By utilizing this modied
distribution for soft information calculation, a decoder can
be constructed that drastically improves the performance even
for low quality fronthaul link. In this work, we investigated
such a decoder by adapting so-called hardware noise resilient
decoders for the use in cloud-based mobile networks. We
showed the feasibility of using such a decoder in cloud-based
networks as well as its performance. We discussed means
of estimating the bit error probability on the fronthaul and
showed that even with a bad estimate great performance gains
can be achieved.
ACKNOWLEDGMENT
0.2
0 -5
10
10
-4
10
-3
10
-2
10
-1
10
10
Fig. 7. The loss in SNR to achieve an end-to-end BER of 104 when Pe

is wrongly estimated as Pe,est
demands of I/Q- sample forwarding. Additional coding also

reduces the throughput but more importantly increases the
latency of the overall transmission, which is also not desirably
due to the tight latency constraints mentioned earlier. Lower
transmit distance and higher transmit energy can be traded of
for one another, as the bit error probability depends on the
SNR at the receiver. It can be shown that the BER for a 4QAM transmission can be approximately calculated as [14]

SN R
1
Pe erfc
(9)
2
2
To achieve the Pe of 1012 required by CPRI, an SNR of
about 16.5 dB is needed as compared to only 7.3 dB for a Pe
of 103 , which would have the same performance according
to our simulations. Thus, by using a lower transmit power and
accepting a higher Pe , 9.2 dB or 88 % of transmit power could
be saved. Conversely, using the same transmit energy and
assuming a ber attenuation of 0.2 db/km [15], the range could
be extended by 46 km. Using wireless FH and assuming that
directive mmWave links are subject to free-space path loss, the
transmission distance could be increased by a factor of 2.9.
All these options would result in tremendous cost savings for
network operators.
Of course, the provided results apply only if the data transmitted over the fronthaul is pre-equalized user data. Stricter
The research leading to these results has received funding

from the European Union Seventh Framework Programme
(FP7/2007-2013) under grant agreement no. 317941. The
authors would like to acknowledge the contributions of their
colleagues in iJOIN, although the views expressed are those
of the authors and do not necessarily represent the project.
The authors would also like to thank Matthias Pingel for
his help in composing the presented work.
R EFERENCES
[1] H. Guan, T. Kolding, and P. Merz, Discovery of Cloud-RAN, in CloudRAN Workshop, Apr. 2010.
[2] P. Rost, C. Bernardos, A. D. Domenico, M. D. Girolamo, M. Lalam,
A. Maeder, D. Sabella, and D. Wubben, Cloud technologies for exible
5G radio access networks, IEEE Communications Magazine, vol. 52,
no. 5, May 2014.
[3] CPRI, Common Public Radio Interface (CPRI)Common Public Radio
Interface (CPRI); Interface Specication (V6.0), Tech. Rep., Aug. 2013,
online: http://www.cpri.info/.
[4] J. Bartelt and G. Fettweis, Radio-Over-Radio: I/Q-Stream Backhauling for Cloud-Based Networks via Millimeter Wave Links, in IEEE
Globecom Workshop (IWCPM 2013), Atlanta, USA, Dec. 2013.
[5] 3GPP, Scenarios and requirements for small cell enhancements for EUTRA and E-UTRAN, Tech. Rep. TR.36.932, Mar. 2013.
[6] FCC, Millimeter wave propagation: Spectrum management implications, Ofce of Eng. and Tech. Bulletin, no. 70, 1997.
[7] A. Hussien, M. Khairy, A. Khajeh, K. Amiri, A. Eltawil, and F. Kurdahi,
A combined channel and hardware noise resilient viterbi decoder, in
Signals, Systems and Computers (ASILOMAR), 2010 Conference Record
of the Forty Fourth Asilomar Conference on, Nov 2010, pp. 395399.
[8] A. Hussien, M. Khairy, A. Khajeh, A. Eltawil, and F. Kurdahi, A
class of low power error compensation iterative decoders, in Global
Telecommunications Conference (GLOBECOM 2011), 2011 IEEE, Dec
2011, pp. 16.
[9] 3GPP TS 36.212: LTE; Evolved Universal Terrestrial Radio Access (EUTRA); Multiplexing and channel coding , 3GPP Std., Apr. 2013.
[10] J. Hagenauer and P. Hoeher, A viterbi algorithm with soft-decision
outputs and its applications, in Global Telecommunications Conference
and Exhibition Communications Technology for the 1990s and Beyond
(GLOBECOM), 1989. IEEE, Nov 1989, pp. 16801686 vol.3.
1482
[11] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, Optimal decoding of linear

codes for minimizing symbol error rate (corresp.), Information Theory,
IEEE Transactions on, vol. 20, no. 2, pp. 284287, 1974.
[12] A. Khajeh, K. Amiri, M. Khairy, A. Eltawil, and F. Kurdahi, A
unied hardware and channel noise model for communication systems,
in Global Telecommunications Conference (GLOBECOM 2010), 2010
IEEE, Dec 2010, pp. 15.
[13] S. Lloyd, Least squares quantization in PCM, Information Theory,
IEEE Transactions on, vol. 28, no. 2, pp. 129137, 1982.
[14] A. Goldsmith, Wireless communications. Cambridge university press,
2005.
[15] W. Kiess, C. Choi, A. Khan, and K. Kozu, Cloud base stations and
xed mobile convergence for the next mobile network, in Globecom
Workshops (GC Wkshps), 2012 IEEE, Dec 2012, pp. 238243.
[16] U. Dotsch, M. Doll, H. P. Mayer, F. Schaich, J. Segel, and P. Sehier,
Quantitative Analysis of Split Base Station Processing and Determination of Advantageous Architectures for LTE, Bell Labs Technical
Journal, vol. 18, no. 1, pp. 105128, May 2013.
1483
Dependable Multi-Tenant Infrastructures Supporting

Cloud and Mobile Cloud Services
Markos. P. Anastasopoulos, Anna Tzanakaki and Dimitra Simeonidou
High Performance Networks Groups, University of Bristol, UK
{m.anastasopoulos, anna.tzanakaki, dimitra.simeonidou}@bristol.ac.uk
AbstractThis paper focuses on the design of dependable
Mobile Optical Virtual Networks (MOVNOs) to enable multitenancy over converged wireless, optical and computational
infrastructures supporting cloud and mobile cloud services.
Dependability is defined as the MOVNOs ability to offer
resilience in case of network or computing resource failures and
security guarantees in terms of isolation between MOVNOs that
share the common physical infrastructure. The VI planning
optimization objective is to minimize the overall energy
consumption satisfying at the same time specific end-to-end delay
constraints. To address these issues, together with the uncertainty
introduced by the mobility of the end devices we proposed a
MOVNO planning scheme based on non-linear programming.
Our modeling results illustrate interesting trade-offs between the
overall energy consumption of the MOVNOs, the mobility of the
end-devices, as well as the dependability impact on the energy
consumption and the utilization of the infrastructure resources.
Keywordsmobile optical virtual networks; mobile cloud; cloud
computing; energy efficiency
I.
INTRODUCTION
It is predicted that the number of end devices connected to

the Internet coupled with the rapid rise of Big Data and Cloudbased services, will result in increasingly vast volumes of data
(expected to exceed 1 zettabyte per year by 2016) and will
require to be transported to remote locations across an
interconnecting network infrastructure [1]. In addition, mobile
internet users are expected to experience a dramatic growth,
introducing a huge increase in mobile data, a big part of which
will come from Cloud Computing applications. To address these
emerging requirements the concept of Cloud and Mobile Cloud
Computing, where computing power and data storage are
moving away from the end user to remote computing resources,
has been introduced. Cloud computing facilitates customers to
access computing resources, on an on-demand basis. This
introduces new business models and facilitates new
opportunities for a variety of business sectors. At the same time,
it increases sustainability and efficiency in resource utilization
reducing capital and operational expenditures.
In view of this, the increasingly deployed cloud and mobile
cloud computing services are emerging as one of the fastest
growing business opportunities for Internet service providers
and telecom operators. To enable this emerging business
opportunity, there is a clear need for suitable network solutions
to interconnect distributed data centers (DCs) with fixed and
mobile end users. To address this need we have proposed a
converged network integrating optical metro and wireless access
978-1-4799-7470-2/14/$31.00 2014 IEEE
technologies with DCs in a common infrastructure providing

ubiquitous access to fixed and mobile end users [2].
In this heterogeneous infrastructure the optical metro
network solution supports frame based sub-wavelength
switching granularity and plays a key role, as it offers a very high
capacity, energy efficient, carrier grade transport with increased
dynamicity and flexibility. The Long Term Evolution (LTE)
wireless access technology adopted supports the requirement for
increased capacity in the access domain, improved Quality of
Service (QoS) capabilities and end users mobility.
Although optical networks offer almost unlimited capacity,
there are mostly available in urban environments and are limited
by long deployment times. On the other hand, spectrum efficient
wireless network technologies such as LTE can be effectively
used to provide a variety of services to a large pool of users.
However, the vast majority of wireless systems are currently
optimized for traditional voice and short message services
(SMS) since they offer higher profit margins to the Service
Providers (SPs) compared to more on-demand internet-enabled
content distribution (e.g., over-the-top) services [3]. To address
this challenge, novel technology solutions that allow SPs to
reduce their costs and assist them to achieve improved
utilization, multi-tenant efficiency, and economies of scale need
to developed.
Towards this direction, infrastructure virtualization ([4][12]), that allows physical resources to be shared and accessed
remotely on-demand, has been recently proposed as a key
enabling technology to overcome the strict requirements for
resource efficiency in cloud and mobile cloud environments.
Given the heterogeneous nature of the infrastructures suitable to
support efficiently cloud and mobile cloud services described
above, it is clear that cross-technology and cross-domain
virtualization is required. In this context, the introduction of
mobile optical virtual networks (MOVNOs) facilitates sharing
of heterogeneous physical resources among multiple tenants,
introducing new business models that suit well the nature and
characteristics of the Future Internet and enable new exploitation
opportunities for the underlying physical resources. However, to
ensure optimal performance and resource efficiency, MOVNOs
need to be appropriately designed and operated to address the
very dynamic and unpredictable traffic profiles and service
characteristics they are expected to support. Underestimating the
required computing and network resources may lead in inability
to satisfy end-users requirements, while overestimation may
1490
SOVNO/VDC #2
MOVNO#1
MOVNO#1
Second. Path
Primary Path
Metro
Wireless Access
Metro Optical
Metro
Wireless Access
Data Centers
Metro Optical
Data Centers
a)
b)
Fig. 1 MOVNO Planning with a) Resilience considerations: b) Security Considerations for the case where no common links are allowed between any two
MOVNOs
x VI security is addressed by ensuring maximum isolation

between the formed VIs. This is enforced by constraining the
designed VIs to have a limited number of common links that
they share between them. This approach follows the generic
security concept proposed by [18], [19]. In cases were
maximum security is required no common shared links
between VIs are allowed.
lead in resource over-provisioning and hence increased

operational and capital expenditures. In this context, optimal
MOVNO planning with respect to specific objectives and
challenges (e.g., seamless connectivity, latency, mobility,
dependability and elasticity) plays a key role.
In response to this, we propose the use of Non-Linear
Programming (NLP) to plan dependable MOVNOs, supporting
mobile cloud services, over heterogeneous wireless optical and
DC infrastructures. The objective of our optimization is to
minimize energy consumption, with the aim to address
sustainability issues and reduce the MOVNOs operational costs.
The proposed approach takes into account both mobility and
dependability issues extending our previous work on energy
efficient VI planning [13].
The rest of the paper is organized as follows. Section II
provides an overview of the network scenario under study and a
description of the relevant state-of-the-art. In Section III the
proposed virtual infrastructure (VI) planning scheme based on
NLP and the parameters assumed is presented. Section IV
examines the impact of the volume and type of traffic demands
as well as other parameters (e.g. mobility, end-to-end delay
guarantees etc) on the physical resources requirements.
Comparisons with conventional schemes are also presented.
Finally, Section V summarizes the paper conclusions.
II.
NETWORK SCENARIO AND RELATED WORK
The physical infrastructure considered in this paper is

illustrated in Fig. 1 and comprises an optical metro network
solution referred to as Times Shared Optical Network (TSON)
[14] deploying sub-wavelength switching granularity, LTE
wireless access technologies and DCs. Our study focuses on
optimal planning of dependable MOVNOs in terms of both
topology and resources over the physical infrastructure. To
identify the least energy consuming dependable MOVNOs, the
detailed power consumption models for the optical metro, the
wireless access networks [15] and IT resources are considered.
Mapping the virtual to physical resources and defining the VI
energy consumption parameters is also part of the VI planning
process. In our study dependability refers to infrastructure
resilience and security:
x resilience is addressed through protection for network and
computing resources [12], [13], [16], [17].
At the same time, in order to satisfy the low-latency

requirements of several content-rich mobile cloud computing
services such as high definition video streaming, online gaming
and real time language translation [20], emphasis during the
planning process should be given on defining acceptable degrees
of latency introduced by the various technology domains to
ensure guaranteed QoS levels.
So far, the concept of identifying the optimal virtual
resources has been extensively reported in the literature. A
detailed survey on the subject is reported in [17]. The role of
optical network virtualization in support of cloud services is
analyzed in [6]. In [4], a model describing the concept of
planning virtual infrastructures over of a converged network and
IT infrastructure is presented and an energy-aware VI planning
problem is formulated and compared to the a closest-IT
approach, providing savings of the order of 30% for single VI
establishment. In [5] the authors propose multiple VI energyaware planning, assuming global knowledge of the VI requests
and provide the optimal solution for the establishment of all VIs
concurrently, whereas in [11] a physical layer impairment-aware
virtual optical network (VON) composition algorithm for singleline-rate and mixed-line-rate network scenarios is proposed.
Other research efforts have focused on the problems of virtual
network mapping with resilience considerations (see [12]-[16]),
and on the optimal allocation of VONs over an all-optical
network substrate [21]. Another interesting work deals with the
problem of isolation and resource efficiency of VONs [22]. On
the topic of wireless network virtualization some preliminary
studies are provided in [23]-[24].
III.
MODEL DESCRIPTION
A. Problem Description
As already discussed, this paper focuses on planning of
MOVNOs over an integrated converged infrastructure
comprising a cellular LTE system for the wireless access part
and an optical metro network supporting frame based sub-
1491
wavelength switching granularity that interconnects the endusers with the computing resources. The objective of the
proposed planning problem is to identify the virtual
infrastructure topologies and resources required by the various
fixed and mobile tenants to facilitate them to provision services
and execute their applications over the common physical
infrastructure.
The VI planning model addresses also the requirements for
dependability of the planned VIs, offering at the same time
seamless connectivity to the mobile users. In this paper
dependability is supported through the adoption of resilience and
security mechanisms. In the VI planning phase, these
mechanisms are considered it terms of the suitable VI topology
selection and the additional resource requirements they
introduce. More specifically resilience in case of single link
failures is supported through the allocation of additional
resources required to facilitate 1:1 protection for all virtual
resources per MOVNO (Fig.1a). On the other hand, security is
supported according to [18]-[19], by keeping security sensitive
VIs disjoint to the rest of the VIs. By disjoint VIs we mean that
security sensitive VIs are allowed to have no or a limited number
of common physical links that they share with other VIs
(Fig.2a). The number of acceptable common links can be
defined depending on the security level that needs to be
supported. The rationale behind this policy is that, at any time
instance VIs will be realized through full or partially disjoint
physical layer paths, thus, minimizing possible effects from
other malicious VIs sharing the same physical substrate.
Furthermore, for each demand of , its network

demand volume is supported by means of a number of paths
assigned to the corresponding MOVNO paths. Taking into
account the demands in the network, let be the
candidate path list in the virtual layer for the paths required to
support demand of , and the corresponding
number of paths. The following demand constraints should be
satisfied for each VI:

Summing up the paths through each virtual link of the

, the necessary virtual capacity required for link
is determined through the following inequality

where is a binary variable taking value equal to 1 if link

of the optical network belongs to path realizing demand ,
0 otherwise.
Based on the same rationale, the capacity of each link in
is allocated by identifying the required paths in the PI. The
resulting PI paths determine the load of each link
of the PI, and hence its capacity at time . To mathematically
formulate the demand constraints for link , a similar approach
to that presented in (2) is adopted. Finally, during the VI to PI
mapping process the PI capacity constraint should be satisfied.
The inputs of the proposed VI planning algorithm include the

parameters specifying the details of the underlying physical
resources e.g., the network graph, the available bandwidth per
link, the energy consumption of the network, etc. In addition, the
traffic requests per VI are used as an input.
The planned MOVNOs are obtained by minimizing the

expected cost:
B. Basic Mathematical Formulation

The problem is formulated using a network composed of one
layer that contains the mobile virtual optical network requests
and one resource layer that contains the physical infrastructure
(PI). Each is modeled as an undirected graph
where , are used to denote the set of
nodes and virtual links, respectively, and is used to describe
the set of demands. These demands need to be served by an IT
server . The PI is modeled as a weighted undirected
graph where is the set of PI nodes and
the set of PI links. The objective of the planning process is to
map each to a subset of so that all MOVNO requests can
be satisfied. A similar multi-layer network optimization
approach can be found in [25].
subject to
x scheduling constraints described through (see eqs.1, 2)
The identification of the suitable IT resources is part of the

optimization output and the demand destinations for each
will be identified through the optimization performed by the
proposed model. To formulate this requirement the binary
variable is introduced to indicate whether demand of
is assigned to IT server or not. This variable takes
value equal to 1 if demand that is serviced by the the
is processed on server . For simplicity it is also assumed that
each demand can be assigned to only one server.

x capacity constraints in the VI (see eq.3)

x VI to PI mapping constraints
x capacity constraints in the PI
where is the cost for operating capacity of the PI link
a time and the cost for operating capacity of the virtual
link at time of .
C. Dependabiltiy and Mobility Considerations
1) Planning with Resilience Considerations: In this
subsection, a modeling approach using NLP suitable for the
optimal planning of resilient MOVNOs is proposed and
implemented, based on the analysis presented in [13]. As
described in Subsection III.A the resilience scheme considered
is that of 1:1 protection for network resources. Specifically, in
case of a single link failure, demands are routed to their
destination through a secondary (protection) path.
Similarly to the basic model, the MOVNO planning problem
with resilience considerations is formulated using a network that
is composed of one resource layer that contains the physical
1492
Wireless Access
Wireless Backhaul
Metro Optical
Data Centers
9000
8000
Additional Resources (%)
Total Power Consumption (Watt)
10000
7000
6000
5000
4000
DC
Metro Optical
Wireless Backhaul
Wireless Access
100
80
60
40
20
0
wo
without with without with without with

13770 Mbps
7650Mbps
10710 Mbps
Total Load (MCC/Cloud traffic =0.02)
with
wo
with
1
Call to mobility factor
wo
with
a)
b)
Fig. 2 a) Impact of traffic load on power consumption with and without network protection, b) Impact of mobility factor on the requested resources
infrastructure and will produce as an output the virtual

infrastructure layer. Again, in the PI a set of randomly selected
Source-Destination pairs is considered, with each source
generating demands that need to be routed at the destination
through a set of virtual links. In the proposed planning
algorithm, a possible failure of a link is treated by forwarding
demands to their destination via alternative paths. Therefore, in
order to protect the planned network from a possible failure of a
physical layer link , a link re-establishment mechanism is
introduced that routes demands to their destination via
alternative paths. A graphical representation of the protection
mechanism adopted in this work is provided in Fig.1a. It is
observed that MOVNO#1 is realized through a primary physical
layer path. However, in case of a PI link or DC failure, demands
will be routed through a secondary path to their destination.
Now let be the protection capacity that should be reserved
in the PI link at time . In order to identify the optimal
MOVNOs with resilience considerations, the objective function
given in (4) should be minimized, where is given by:

2) Planning with Security Considerations: VI security is

addressed by ensuring maximum isolation between the formed
VIs. As described in section II.A this is enforced by
constraining the designed VIs to have a limited number of
common links that they share between them. This approach
follows the generic security concept proposed by [18]. In cases
were maximum security is required no common shared links
between MOVNOs are allowed. A typical planning example for
the case where no common links are allowed between any two
MOVNOs is illustrated in Fig.2b. It is observed that both
MOVNOs are realized through different physical layer
resources leading in general to longer average path lengths and
hence increased network resource requirements.
3) Planning with Mobility Considerations: A major issue to
be taken into account is the distinction between traffic that
arises from fixed and mobile devices. The accurate estimation

of resources that should be reserved in the VI to ensure seamless
end-to-end service provisioning is based on the mobility model
that is adopted [26], the size of the LTE cells as well as the
traffic model used [2]. In the ideal case, a seamless handoff for
a mobile device can be 100% guaranteed only if the equivalent
amount of resources is reserved at all its neighbouring cells.
However, a more efficient approach would be to relate the
reserved resources in the neighbouring LTE cells with the
handoff probability, [26].
4) Planning with End-to-End Delay Considerations:
The previous constraints ensure that there are sufficient
network and processing capacities for all MOVNOs to support
the requested services. Apart from network bandwidth
requirements, end-to-end delay guarantees should be also
provided. Given that in highly loaded networks queuing delay
is the dominant part of the end-to-end delay, the MOVNO
planning scheme should be also capable of making queuing and
scheduling decisions across all the technology domains
involved. A typical queuing decision example includes the
reservation of a specific portion of the receivers/transmitters
queues by a VI either at a TSON node or at a wireless base
station (Fig. 1), with the objective to maintain the end-to-end
delay below a predefined threshold [28]. To achieve this, a
closed form approximation for the end-to-end delay for the
services that are provided by each can be extracted by
applying the Network of Queues theory [27].
IV. NUMERICAL RESULTS
The performance of the proposed VI planning scheme across
the multiple domains involved is studied based on the
infrastructure illustrated in Fig.1. For the PI, a macro-cellular
network with regular hexagonal cell layout has been considered
similar to that presented in [15]. The inter-site distance has been
1493
Additional network resources (%)
Resource Efficiency
1
0.9
0.8
0.7
0.6
6
0.04
0
0.06
0.3
0.2
0.1
0
2
1
Common links
2
1
Common links
Wireless Access
Wireless Backhaul
Metro Optical
0
W/O protection
0.02
2
Call to mobility factor
0.4
WIth protection
0.5
Fig. 4 Additional resources required (over the dynamic scheme with full
information) for different availability requirements
Network to compute ratio
Fig. 3 Impact of mobility and network to compute ratio to the maximum

number of tasks/s per mobile device that can be offloaded to the cloud.
set to 500m to capture to scenario of a dense urban network

deployment. Furthermore, a 22 MIMO transmission has been
considered, while the users are uniformly distributed over the
serviced area. Each site can process up to 115 Mbps and its
power consumption ranges from 885 to 1087W, under idle and
full load, respectively [27]. The wireless backhaul links are
treated as a collection of wireless micro wave links of 100 Mbps
capacity. For the computing resources, two Sun Oracle
Database Machine Basic Systems [28] have been considered
where each server can process up to 36Gbps of compressed flash
data. The physical TSON topology assumed is illustrated in Fig.
1, where the dimensions of the optical rings are below 5 km and
the supported data rate is 8.68Gbps with switching granularity
100Mbps. The power consumption of the TSON equipment is
measured to be 50W for the EDFAs and 100mW for the PLZT
chip.
1) Impact of traffic load on the average power consumption
with and without resilience
Fig. 2 a) illustrates the total power consumption of the converged
infrastructure (wireless access, wireless backhaul, optical
network and IT resources) with and without resilient
considerations. It is observed that the protection mechanism
introduces increase on the power consumption of the order of
20%. It is observed that for both schemes the average power
consumption increases almost linearly with the number of
demands. However, the relative benefit of the un-protected
design decreases slightly with the number of demands, when
approaching full system load. It can also be seen that the wireless
access technology is responsible for 50% of the overall power
consumption, while the optical network consumes less than 10%
of the total energy.
To further investigate how the service characteristics affect
the optimum planned MOVNOs, the impact of end user mobility
on the total power consumption (operational cost) of the VIs is
studied. To facilitate this, the call-to-mobility factor is defined
as the fraction of the service holding time over the cell residence
time. It is assumed that the cell residence time is exponentially
distributed with parameter and the service holding time is
Erlang distributed with parameters Fig 2 b) shows that
when mobility is higher (lower service-to-mobility factor)

additional resources are required to support the VI in the
wireless access domain. It is interesting to observe that this
additional resource requirement also propagates in the optical
metro network and the IT domain. The additional resource
requirements, across the various infrastructure domains,
effectively ensure availability of resources in all domains
involved (wireless access and backhauling, optical metro
network and DCs) to support the requested services and enable
seamless and transparent end-to-end connectivity. Therefore, to
achieve seamless end-to-end connectivity between end users and
computing resources additional resources should be reserved in
the MOVNOS. For fast moving mobile terminals and services
with high duration (low service-to-mobility factor) more than
90% additional optical network resources should be reserved by
the VI to ensure seamless service provisioning across the various
technology segments. For high mobility, the cases where
services are realized through different TSON gateways are
increased and, therefore, the resources in the optical domain
should be increased.
2) Impact of network to-compute ratio and mobility on the
infrastrucutre resource efficiency
Fig.3 illustrates the impact of communication cost and mobility
on the resource efficiency of the infrastructure. To derive this,
the network-to-compute resource ratio is introduced to capture
the communication cost associated with various types of
services. More specifically, it is used to capture the relationship
between computational and network bandwidth requirements
corresponding to specific cloud and mobile cloud service types.
In Fig. 3 it is observed that resource efficiency (defined as the
ratio of the mobile traffic load that can be supported over the
maximum mobile traffic load) increases when the network to
compute ratio decreases i.e. when the service demands impose
higher network resource requirements. Regarding end user
mobility as shown in Fig 3, when mobility is high (lower callto-mobility factor), additional resources are required to support
the VI in the wireless access domain. This additional resource
requirement also propagates in the optical metro network and the
DC domains in order to ensure availability of resources in all the
domains involved.
1494
Finally Fig. 4 shows that the higher the degree of disjointness

in the MOVNOs i.e. the tighter the security requirements, the
higher the network resource requirements are, while protection
introduces even higher resource requirements.
[7]
[8]
[9]
V. CONCLUSIONS
The increasingly deployed cloud and mobile cloud
computing services are emerging as one of the fastest growing
business opportunities. To enable this emerging business
opportunity, there is a clear need for suitable network solutions
to interconnect distributed DCs with fixed and mobile end users.
Our proposal involves a converged network integrating optical
metro and wireless access technologies with DCs in a common
infrastructure providing ubiquitous access to fixed and mobile
end users. To effectively and efficiently utilize the physical
infrastructure the model of multi-tenancy is proposed and is
facilitated through the adoption of infrastructure virtualization
and the creation of dependable Mobile Optical Virtual Networks
(MOVNOs). Dependability is defined as the MOVNOs ability
to offer resilience in case of network or computing resource
failures and security guarantees in terms of isolation between
MOVNOs that share the common physical infrastructure.
MOVNOs are planned with the objective to minimize the overall
energy consumption satisfying at the same time specific end-toend delay constraints. To address these issues, together with the
uncertainty introduced by the mobility of the end devices a
MOVNO planning scheme based on NLP has been developed.
Our modeling results illustrate trade-offs between the overall
energy consumption of the MOVNOs, the mobility of the enddevices, as well as the dependability impact on the energy
consumption and the utilization of the infrastructure resources.
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
ACKNOWLEDGMENT
This work was supported by the CONTENT (FP7-ICT318514) project funded by the EC through the 7th ICT FP and
the EPSRC grant EP/L020009/1: Towards Ultimate
Convergence of All Networks (TOUCAN) .
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
Cisco, The Zettabyte era, white paper, May 2012

A. Tzanakaki et.al., Virtualization of heterogeneous wireless-optical
network and IT infrastructures in support of cloud and mobile cloud
services, IEEE Communications Magazine. vol.51, no.8, pp.155-161,
Aug. 2013
H. Agusleo, N. Arora, The Road to Cloud Nine: How Service Providers
Can Monetize Consumer Mobile Cloud, CISCO, Feb. 2013.
A. Tzanakaki et.al.,Energy Efficiency in integrated IT and optical
network infrastructures: The GEYSERS approach, in Proc. of IEEE
INFOCOM WKSHPS, pp.343,348, 2011
A.Tzanakaki, M. Anastasopoulos, K. Georgakilas, D. Simeonidou,
Energy Aware Planning of Multiple VIs over Converged Optical
Network and IT Physical Resources, in Proc. ECOC 2011
M. Jinno, Y. Tsukishima, "Virtualized optical network (VON) for agile
cloud computing environment," in Proc. of OFC 2009.
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
1495
R. Nejabati, E. Escalona, P. Shuping, D. Simeonidou, Optical network

virtualization, in Proc. of ONDM, 2011
A. Wang, M. Iyer, R. Dutta, G. N. Rouskas, I. Baldine, "Network
Virtualization: Technologies, Perspectives, and Frontiers." IEEE/OSA J.
of Lightwave Technology, Dec. 2012
A. Tzanakaki, M. Anastasopoulos, K. Georgakilas, "Dynamic Adaptive
Virtual Optical Networks, in Proc. of OFC/NFOEC 2013 (invited).
N. Chowdhury, R. Boutaba, Network virtualization: state of the art and
research challenges, IEEE Commun. Mag., 47, no.7, pp.20,26, 2009
S. Peng, R. Nejabati, D. Simeonidou, Impairment-Aware Optical
Network Virtualization in Single-Line-Rate and Mixed-Line-Rate WDM
Networks, IEEE/OSA JOCN, Vol. 5, No.3, 2013.
M. F. Habib et.al.,Fault-tolerant virtual network mapping to provide
content connectivity in optical networks, in Proc. of OFC 2013.
A. Tzanakaki et.al., "Energy efficiency considerations in integrated IT and
optical network resilient infrastructures", in proc. of ICTON, 2011
G.S. Zervas et.al., Time Shared Optical Network (TSON): a novel metro
architecture for flexible multi-granular services, Opt. Express 19, B509B514, 2011
G. Auer, V. Giannini, "Cellular Energy Efficiency Evaluation
Framework", in Proc. Vehicular Tech. Conf., Hungary, May. 2011
A. Basta, B. Barla, M. Hoffmann, G. Carle, D. Schupke, Failure
Coverage in Optimal Virtual Networks, in Proc. of OFC 2013
D. Schupke, Multi-Layer and Multi-Domain Resilience in Optical
Networks, Proceedings of the IEEE, 100(5), May 2012.
N. A. Skorin-Kapov, J. B. Chen, L. b. Wosinska, A new approach to
optical networks security: Attack-aware routing and wavelength
assignment, IEEE/ACM Trans. Netw., Vol.18, No.3, pp.750-760, 2010.
M. Anastasopoulos, A. Tzanakaki, and D. Simeonidou, "Stochastic
Planning of Dependable Virtual Infrastructures Over Optical Datacenter
Networks," J. Opt. Commun. Netw. 5, 968-979, 2013.
H. Dinh et.al., A survey of mobile cloud computing: architecture,
applications, and approaches, Wirel. Commun. Mob. Comput. Oct. 2011
A. Pags, J. A. Garca-Espn, J. Perell, J. Ferrer Riera, S. Spadaro, S.
Figuerola, Optimal allocation of virtual optical networks for the future
Internet, in Proc. ONDM 2012
M. De Leenheer, J. Buysse, C. Develder, B. Mukherjee, Isolation and
resource efficiency of virtual optical networks, in Proc. of ICNC 2012
K. Katsalis et. Al., Content Project: Considerations Towards a CloudBased Internetworking Paradigm, in "proc. of IEEE SDN4FNS 2013
Keun-mo Park, Chong-kwon Kim, A framework for virtual network
embedding in wireless networks. In Proc. of CFI 2009
E. Kubilinskas, P. Nilsson, M. Pioro, Design Models for Robust MultiLayer Next Generation Internet Core Networks Carrying Elastic Traffic,
in proc. of DRCN 2003, 61-68 (2003).
Y. Fang, I. Chlamtac, Analytical Generalized Results for Handoff
Probability in Wireless Networks, IEEE Transactions on
Communications, Vol.50, No.3, pp. 369-399, Mar. 2002
G. Pujolle, W.Ai, A Solution for Multiserver and Multiclass Open
Queueing Networks, INFOR, Vol.24(3), .221-230, 1986
A. Tzanakaki et. al., A Converged Network Architecture for Energy
Efficient Mobile Cloud Computing'', \emph{in proc. of ONDM 2014
Oracle Data Sheet, Sun Oracle DataBaseMachine, Online:
http://www.oracle.com/us/products/database/database-machine069034.html
Meeting fronthaul challenges of future mobile

network deployments - the HARP approach
Lars Dittmann, Henrik L. Christiansen
Aleksandra Checko
DTU Fotonik, Network technologies and Service platforms

Technical University of Denmark - DTU
DK-2800 Lyngby, Denmark
ladit@fotonik.dtu.dk; hlch@fotonik.dtu.dk
MTI Radiocomp
DK-3400 Hillerd, Denmark
aleksandra.checko@mtigroup.com
Abstract In future mobile networks aggregation at

different levels is necessary but at the same time imposes
challenges that mandate looking into new architectures. This
paper presents the design consideration approach for a C-RAN
based mobile aggregation network used in the EU HARP project.
With this architecture fronthaul aggregation is performed which
might be an option for future generation of mobile networks.
Keywords
engineering.
C-RAN,
mobile,
aggregation,
SDN,
traffic
I. INTRODUCTION
High utilisation network
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Late aggregation
aggregat ion
aggregat ion
Low utilisation network
Low utilisation network
High utilisation network
Fig. 1. The balance between early vs. late aggregation from an optimisation
perspective (1 stage vs. 2 stage)
In multiservice networks the aggregation process is

commonly made in two steps: first step is service independent
as for single service networks while the second step is service
dependent. This concept ensures that as much traffic as
possible is aggregated as early as possible, but has the
drawback of preventing a quality based traffic engineering.
This concept for multiservice aggregation is as an example
used in 3G mobile network (fig. 2) in which the NodeB and the
Radio Network Controller (RNC) are aggregating traffic in a
service independent way prior to the mobile core network
where a service separation takes place.
II. EVOLUTION OF MOBILE NETWORKS

Even though the mobile network is usually split into a
(radio)access part and a core part it is from a generic
networking perspective all part of an access network with the
This work was partially sponsored by the 7th Framework programme for
research of the European Harp project, under the grant number HARP318489.
978-1-4799-7470-2/14/$31.00 2014 IEEE
Early aggregation
aggregat ion
The remainder of this paper is organized as follows. In

Section II we describe how aggregation in mobile network was
changing with the introduction of new generations of mobile
networks. In Section III we introduce a concept of a simple
fronthaul network. In Section IV we elaborate on EU HARP
project approach with aggregation in the fronthaul. In
Section V we outline future research directions towards
redefining aggregation in fronthaul network. Section VI
concludes the paper.
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
Client
aggregat ion
For the next generation mobile network called 5G as a

working title we might still see the need for a new radio
interface, but the majority of changes are probably to be seen in
fixed part of the network. Even though the bandwidth
efficiency can be improved in the existing frequency range for
cellular system and higher frequency band can be made useful
for small cell deployment, it cannot be avoided that the general
density of cells has to go significantly up in order to meet the
demand and expectation for unlimited capacity with
extremely low latency.
In a single service access network the aggregation part can

usually be made very simple and the number of stages will
often be given by the traffic characteristics of an area (fig. 1.).
The balance will typically be between aggregating as much
traffic as possible - to obtain the largest statistical gain - and
aggregating as early as possible in order to have the highest
level of utilization of as much of the network as possible. The
balancing is very much traffic profile dependent and some
flexibility in the aggregation networks is commonly an
advantage unless traffic can be predicted very accurately.
aggregat ion
While previous generations of mobile networks primarily

have been defined by the radio-interface, Long Term evolution
(LTE) and LTE-Advanced (LTE-A) have changed the picture
and to a larger extend included the Radio Access Network
(RAN) and core into the optimization process. This new focus
is related both to cost and performance (especially latency and
lower layer loss management) and even introduced the
fronthaul domain as a new area of research.
objective to connect user (and/or devices) with a global core

infrastructure in a reliable and cost efficient way.
1496
an overlay with small cells with coverage from a few meters to

a few hundred meters.
RNC
RNC
RNC
RNC
A density of typically ten or more eNodeBs per square

kilometer is both a challenge from a cost perspective and
regarding control and management. Furthermore, interference
is growing with increased cell density imposing a limit on
density. At one point
Packet
mobile
core
NodeB
RNC
NodeB
RNC
RNC
RNC
RNC
But the use of smaller cells also have a significant impact

on the traffic engineering perspective as the load of the smaller
cells varies much more than in larger cells as the number of
users within a small cell probably will be very small (and is
often zero).
circuit
mobile
core
NodeB
Traffic level aggregation
Service level aggregation
Fig. 2. Two level aggregation as used in 3G mobile networks.
With the evolution towards an all-IP based mobile network

it was made possible to simplify and collapse the mobile core
into a single packet based infrastructure, which in combination
with 4G/LTE also enabled a simplification of the RAN by
integrating RNC functions into the NodeB the eNodeB (fig.
3).
To cope with these challenges it has been proposed to split

up the eNodeB not back to a RNC and NodeB but into a
remote radio head (RRH) and a BaseBand Unit (BBU). The
initial gain in this architecture (e.g., as defined by the OBSAI
or CPRI industry standards[1]) is the ability to place the more
bulky and power consuming BBU in some distance from the
antenna system (fig.4.). This will make it possible to place
antenna systems at locations that otherwise would be
impossible or very expensive. It also lowers cost of network
operation as RRHs do not need air cooling that would
otherwise consume a lot of energy.
The physical separation of the eNodeB function is
commonly called the fronthaul network is initially made as a
fiber based link system.
The introduction of a fronthaul link enables a faster (and
cheaper) deployment of many more cells, but it does not solve
the issue of traffic aggregation and statistical resource sharing
close to the terminal/end-user.
eNodeB
eNodeB
eNodeB
Mobile core
RRH
eNodeB
Front-haul link
eNodeB
RRH
RRH
Traffic level aggregation
Service level aggregation

RRH
Fig. 3. Single level - single stage aggregation as used in LTE mobile

networks. The core is now fully packet based.
Front-haul link
Front-haul link
RRH
RRH
III. INTRODUCING THE FRONTHAUL NETWORK

One of the key drivers in the evolution from 3G mobile
networks towards 4G/LTE was improved network performance
in the RAN in order for Broadband mobile networks to match
the performance (especially with respect to latency) of wired
Broadband networks.
While latency issues was handled at protocol level the
continuous demand for more capacity is handled by combining
more frequency spectrum, better bandwidth utilization and
(especially) by introducing small cells and thereby a
significantly larger number of cells/eNodeBs.
A typical architecture is to maintain geographical coverage
using macro cells and expand the provided capacity by making
RRH
BBUs
Front-haul link
Traffic level
aggregation
Service level
aggregation
Fig. 4. Simple fronthaul link system with separation of RRH and BBU.
As the mobile network protocols have not been designed

for the use of a fronthaul link system, the maximum spacing
(due to latency in relation to protocol response time, e.g., the
HARQ retransmission) is limited to approximately 20-40 km [
1]. This enables to place the space/power consuming BBUs at
a common and possibly cheaper location than the location of
the RRH given by the cell location but also introduces a need
for a new physical layer link system with a bitrate capacity that
is typically significantly higher than the service level capacity
1497
(typically line rates in the Gbps range is required for a LTE cell
with 2x2 MIMO) [5].
Legend
RF
Logical
connection
RF
The use of a centralized location of several BBUs enables

the use of statistical sharing of the BBUs exploiting that not
all RRH will be active at the same time (fig.5.).
RF
RRH Aggregation
network
Access
network
BBU
Base
Base
Base
Base
pool
band
band
band
band
S1
R
F
X2
RF
RF
RRH Aggregation
network
RF
R
F
Fronthaul
BBU
Base
Base
Base
Base
pool
band
band
band
band
Aggregation
network
EPC
MME
S1
S5
SGW
Application server
Internet
S11
PGW
Application server
Backhaul
RRH
Front-haul link
Fig. 7. The advanced HARP architectural concept with agregation in the

fronthaul.
RRH
RRH
In the advanced architecture with a fronthaul aggregation

network rather than just fronthaul links, it is possible to
aggregate traffic from multiple RRH to the BBU pool and
potentially enabling a lower number of BBU than RRH. In this
way a RRH can be deactivated in case there is no traffic
demand in a certain cell.
Front-haul link
Front-haul link
RRH
RRH
RRH
Front-haul switch
with BBU pool
RRH
Front-haul link
Traffic level
aggregation
An RRH aggregation network, based on e.g., Ethernet,

additionally enables intelligent traffic aggregation. The
achievable multiplexing gain in the BBU pool is very much
dependent on the cells traffic profiles [4]. Therefore a flexible
mapping can optimize the use of resources in the BBU pool.
Service level
aggregation
Fig. 5. Advanced fronthaul network with statistical sharing of BBUs.
IV. THE HARP APPROACH TO FRONTHAUL NETWORK

AND C-RAN
The HARP project is on one hand researching and prototyping
RRH systems with advanced directional antenna systems and
on the other hand providing a concept and experimental
platform for highly flexible control and management based on
the idea of Cloud RAN (C-RAN) [3] and Software Defined
Networks (SDN) (fig.6. and fig. 7).
When opening up for a dynamical configuration of the RAN

based on traffic patterns a significant more flexible and
intelligent control platform is required. As in many other
network domain the HARP project is analyzing the use of
Network Function Virtualization (NFV) and SDN as a solution
for network control of front-haul and back-haul networks
(fig.8.).
As a centralized approach to network control, SDN has the
possibility of optimizing the performance not only within a
single BBU pool, but also across BBU pool and
incorporating offload technologies such as WIFI.
Traffic engineering is at the moment not a part of the SDN
framework several research project are currently promoting the
idea toward several SDOs.
Fig. 6. The basic HARP architectural concept based on classic C-RAN

approach with dedicated fiber lines[2]. The RAN is furthermore extended by
using special antennas proding massive MIMO and/or beamforming
capabilities.
Fig. 8. A generic approac towards a traffic based SDN controller supoorting
service and QoS differentiation.[5]. .
1498
V. FROM POINT-TO-POINT TO MULTIPOINT-TO-MULTIPOINT.

FUTURE CHALLENGES IN AGGREGATION NETWORK
CPRI standard was introduced together with deployments
of 3G/4G networks. It is optimized for a point-to-point
connections, which are a typical deployment scenario. The data
is usually sent over short distances, like basement (where
baseband units are placed) to the rooftop (where RRHs are
placed). Synchronization is carried from BBU to RRH along
with data utilizing synchronous nature of CPRI protocol, which
is the most popular on the discussed interface (fig. 9).
Functionality of RRH is limited to power amplification and
frequency filtering and signal processing to support them. This
simplicity is an advantage of a RRH, but creates a huge data
rates for the link between BBU and RRH - e.g. 2.5 Gbps are
needed to support 20 MHz, 2x2 MIMO, a typical configuration
of LTE cell. Such high data rates are not optimal solution for
C-RAN aggregation network.
Load
balancing
switches
RRHs
Virtual
BBU Pools
Fig. 10. Load balancing in C-RAN
GPS, 1588, ect
GPS?, 1588?,
new solution?
CPRI
BBU
RRH
Variable delay
BBU Pool++
Fig. 9. In point to point connection synchronization between RRH and BBU

is transmitted together with data via synchronous CPRi interface .
For C-RAN connection between RRHs and BBU pools will

be multipoint to multipoint to optimize transport network and
fully utilize load balancing between different BBU pools (fig
10). It will span over larger distances to fully leverage benefits
of multiplexing gain. That requires availability of a network
with huge capacity. As it is often not the case, functional split
between equipment at the cell site and central office needs to
be redesigned to reduce the data, as compression might not be
sufficient. Fully standardized solution supporting packet based
networks is desired for that purpose. That poses new challenges
on fronthaul aggregation network for C-RAN. Synchronization
is especially demanding, as variable network delay makes it
difficult to recover frequency and phase with variable network
delays (fig. 11). New study item has been recently adopted by
3GPP to initiate standardization work addressing those issues.
It is in line with the scope of HARP, where we investigate
solutions for Ethernet-based fronthaul (fig. 12).
Variable delay
BBU Pool++
RRH++
Fig. 11. In multipoint-to-multipoint fronthaul synchronization between RRH

and BBU will be affected by variable transmission delay in packet-based
switches. Functional split between RRH and BBu will most likely be changed,
hence RRH++ and BBU++.
Aggregation network
- fronthaul
Remote
Radio
Heads
CPRI over
Eth
Aggregation
network backhaul
BBU
Pool
CPRI2Eth
Gateways
CPRI /
OBSAI
CPRI2Eth
Gateway
Fig. 12. Packet-based fronthaul investingated in HARP.
1499
VI. CONCLUSION
By introducing C-RAN (in addition to directional oriented
antenna system) the EU project HARP demonstrates a cost
efficient and scalable solution for a front-haul network
supporting current and future mobile networks.
Even though the EU project HARP mainly is addressing
currently defined mobile networks standards, the C-RAN based
front-haul network might be even more relevant for the next
generation 5G network that currently is being research as a
concept to support an approach that can meet the demand in
terms of capacity, coverage, mobility, service support etc.
With the C-RAN approach in mind for 5G mobile network

the use of very simple RRH and an optical feeder network will
be very attractive both from CAPEX perspective and an
optimization platform for network control and management
including self-configuration of the provided flexibility of the
RRH.
In the future, functional split between RRH and BBU will
most likely be changed, redefining aggregation network.
Synchronization is one of the main challenges in multipoint-tomultipoint packet based fronthaul.
Even though there are many open issues regarding the 5G

networks there are some consensus that the 5G network will
have a very high density of access-points and that the current
frequency spectrum are insufficient to med the capacity
demand (even with a high reused enabled by the access point
density). A support for millimeter wavelength communication
in small cells is one of many ways to provide more capacity,
but the use of these much higher frequencies contains also new
challenges compared to the more traditional frequencies used
in mobile networks.
One of these major challenges is how to build a feeder
network (mobile front- and backhaul) that can support such
architecture.
1500
REFERENCES
[1]
[2]
[3]
[4]
[5]
www.cpri.info
A. Checko, L. Dittmann, L. Scolari, H. Christiansen, HARP - High
capacity network Architecture with Remote radio heads & Parasitic
antenna arrays, in proc. of EuCNC 2013, Lisbon, Portugal.
A. Checko, H.L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M.S.
Berger, L. Dittmann, Cloud RAN for Mobile Networks - a
Technology Overview accepted for publcation in IEEE
Communications Surveys and Tutorials
A. Checko, H. Holm, and H. Christiansen, Optimizing small cell
deployment by the use of C-RANs, in European Wireless 2014
(EW2014).
L. Dittmann, presentation ITU JCA-SDN meeting 11/7/14, Geneva,
Swiss.
Software-defined wired-wireless access network

convergence: the SODALES approach
Jordi Ferrer Riera, Carles Bock, Eduard Escalona
Michael C. Parker, Stuart Walker, Terry Quinlan
Fundaci i2CAT
Barcelona, Spain
{jordi.ferrer,carlos.bock,eduard.escalona}@i2cat.net
University of Essex
Colchester, United Kingdom
{mcpark,stuwal,quinlan}@essex.ac.uk
Volker Jungnickel, Kai Habel
Victor Marques
Fraunhofer Heinrich Hertz Institute

Berlin, Germany
{volker.jungnickel,kai.habel}@hhi.fraunhofer.de
Portugal Telecom Innovaao

Aveiro, Portugal
victor-m-marques@ptinovaao.pt
David Levi
Ethernity Networks
Israel
vidi.levi@ethernitynet.com
AbstractIt is envisaged that end-user access bandwidth
requirements will notably increase in the coming years; at the
same time, the number of connected mobile devices will also
exponentially grow as the fully-digital connected homes (Internet
of Things) becomes an everyday reality. This article presents an
active remote node (ARN) at an intermediate location between
the central office and end-user premises, as a flexible and futureproofed infrastructure topology approach for solving the
associated bandwidth and wired-wireless convergence issues. The
ARN represents the key architectural design innovation of the
SODALES (SOftware-Defined Access using Low-Energy
Subsystems) network. The rise of mobile communications, and
the trend for seamless convergence between fixed and wireless
networks is tending to make the purely passive approaches (e.g.
as exemplified by passive optical networking (PON) access
architectures) too restrictive, considering the modularity and
flexibility offered by an active remote node. We present a
performance analysis of the ARN node, to support the costeffectiveness of the proposed SODALES solution and to
demonstrate the potential benefits in terms network
performance, operational efficiency, and flexible functionality.
Looking forward, future ARN capabilities can also be expected to
include hierarchical caching, customer premises equipment
(CPE) virtualization, and nearer to the end-user location of
software-defined platforms supporting ubiquitous cloud services.
Keywordsaccess network convergence, active remote node,
performance analysis
I.
INTRODUCTION AND RELATED WORK
End-user access bandwidths requirements are constantly

increasing. It is expected that access bandwidths of 1 Gb/s will
be commonplace in the next two or three years, while 10 Gb/s
will be exploited by domestic applications within the next ten
978-1-4799-7470-2/14/$31.00 2014 IEEE
years [1]. With the rise of new content-rich services (e.g. HD

on-demand video, new television formats such as Super HD 4k
or Ultra HD 8k) it is expected that bandwidth requirements will
continue to steadily increase [2]. The expected service
bandwidth increase, jointly with the appearance in the arena of
business and mobile backhaul applications, is likely to create a
bottleneck in todays gigabit-class passive optical network
(PON) deployments [3]. Such PON deployments between the
central office (CO) and end-user have been the favored
solution up to the present. In addition to the traditional services
delivered both to wired and wireless end-user devices, the
authors in [3] also mention another set of trends driving up
bandwidth demand provision, such as: (i) the increasing
number of different devices connected to the Internet
(considering a fully-digital home, and the Internet of Things);
(ii) the accessibility and availability of a huge range of Cloud
services, for both business and residential users; (iii) the rise of
online content distribution; and (iv) the coexistence of
residential and business applications on a common access
platform, which may drive the requirement for bandwidthsymmetric deployments. There is thus a techno-economic
challenge to be addressed, in order to achieve such highbandwidth access segments.
A great deal of industry discussion has been focused on
what the technology beyond legacy GPON and XG-PON
systems [4,5,6] should be, considering the service bandwidth
requirements. Several options under consideration are 10G
EPON [7] or NG-PON [8], including both NG-PON1 and NGPON2 phases. However, it is becoming obvious that what
network operators require, in order to serve the new emerging
demands, is a more scalable, flexible, and future-proof
technology solution. From one perspective, the NG-PON2
system is targeting the fulfillment of this demand trend, as well
1501
as offering a cost-efficient PON upgrade path and a solution

that fully exploits both the low transmission loss and the
capacity of optical fibers. However, the critical technical and
functional requirement for wired-wireless convergence in
access networking is now becoming of essential importance.
Network convergence offers the prospect of optimizing the
total cost of ownership for the different network operators [3].
This reduction may be achieved by means of eliminating the
distinct network technology solutions in the access and
aggregation domains. Thus, fixed- and mobile-backhauling
must be considered in such future deployments. Converged
wired-wireless environments represent a key-enabler scenario
for the upcoming operators, since mobile subscribers are
expected to end up representing the major part of their future
business plans (considering the rise of connected mobile
devices). In fact, concepts for fifth generation (5G) mobile
networks are nowadays under intense discussion by both
industry and academia, such that 5G is now one of the key
topics within the research community. It is currently assumed
that by around the year 2020, a new generation of mobile
networks will have been deployed and be fully operative.
Recent research has identified major challenges, such as the
aforementioned massive growth in the number of connected
devices, and a wider and more diverse set of requirements,
starting from low data rates for control messages with a
moderate latency, to multi-gigabits per second for low-latency
interactive multimedia [9]. Besides these new radio concepts
envisaged as the part of the major challenges associated in 5G
research, it is clear that the high-capacity, flexibility, and
programmability requirements of 5G will have significant
impact on the network infrastructure underpinning the radio
part. It is commonly assumed that both the mobile and the
fixed access networks will therefore need to be optimized
jointly to meet all these high demands [10]. Therefore, full
technical and functional convergence of fixed optical with
wireless communications is one of the key requirements to be
addressed for future-proofed access networks.
It is from this perspective that this article presents the
SOftware-Defined Access using Low-Energy Subsystems
(SODALES) approach for a future-proofed, converged access
network architecture. The approach is based on the
introduction of an Active Remote Node (ARN), taking
advantage of the powering requirements of the existing base
stations (BSs) for radio access within such wired-wireless
converged infrastructures. The introduction of the ARN
represents the key structural innovation of the SODALES
network, with the focus on wired and wireless communications
convergence. As part of creating a new access architecture
paradigm is the additional realization that the presence of such
active nodes between end-users and the central office (CO) is
also likely to become an ever-more essential feature of nextgeneration access topologies, as intelligence and advanced
network functionalities are devolved ever closer to the endusers. In the SODALES network only very basic functionalities
such as statistical multiplexing at the Ethernet switches, basic
bandwidth service differentiation between residential and
business users, and a common backhaul infrastructure for a
RBS is considered. However, looking to the future, we can
expect ever greater network functions virtualization (NFV) to
take place at the ARN, with local caching as part of a

hierarchical content delivery network (CDN), more
sophisticated service differentiation (QoS/E) for a more
segmented end-user base, and additional software-defined
operation of the network to maximize exploitation (e.g. for
energy efficiency and operational expenditure (OpEx)
purposes) of available network resources. Thus we see the
SODALES ARN as the first tentative example of intelligent
functionality located at an active, street level, converged
network node, positioned intermediately between (fixed and
mobile) end-users and the central office (CO).
The rest of the paper is structured as follows: Section 2
contains a more detailed description of the SODALES
architecture, focusing on the ARN, and providing some
examples on the flexibility of such a functional node. Section 2
also contains technical descriptions of the different advantages
of introducing an active node in comparison with full-passive
approaches. Section 3 presents the performance analysis of the
ARN, via network simulations of the ARN, considering its
basic configuration. Finally, section 4 concludes and presents
some future work directions with respect to the ARN.
II. SODALES ARCHITECTURE: THE ACTIVE REMOTE NODE
From a physical architecture point of view, the presence of
an active remote node (ARN) [11,12] at an intermediate
location between the central office (CO) and end-users is the
key structural innovation of the SODALES network. The ARN
is relatively close (up to 300 meters) to end-users. Locating an
ARN intermediate to CO and end-users goes against the (up-tonow) conventional PON philosophy of removing all active
(electronic) equipment from the street level, so that there is a
purely passive optical link between CO and end-user.
Perceived benefits of the PON approach have been reduced
energy consumption, and reduced outside maintenance costs.
However, the rise of mobile communications, and the trend for
seamless convergence between fixed and wireless networks is
tending to make the purely PON approach too restrictive.
Indeed, an ultra-broadband 5G wireless network, e.g. evolving
from long-term evolution (LTE) technology, requires eversmaller cell sizes (from micro to pico and finally to femto cell
sizes), with base stations (BSs) and remote antenna units
(RAUs) in relatively near vicinities to the end-users. Such BSs
and RAUs need to be powered, e.g. either by a suitable line
from the national electricity grid, or preferably, via local
renewable (e.g. solar, wind) sources.
There have been previous examples of the use of active
intermediate nodes, e.g. as seen in the Ethernet active star and
Home Run architectures associated with an active optical
network (AON). However, these examples have been designed
with very limited functionality in a purely fixed nextgeneration optical access network. Likewise, the radio access
node (RAN) associated with mobile networks also captures
some of the functionalities incorporated within the SODALES
ARN, but only with mobile access networking in mind.
Convergence of fixed optical with wireless (mobile)
communications is a key functionality being investigated
within the SODALES project. As part of creating a new access
architecture paradigm is the additional realization that the
1502
presence of active nodes between end-users and the central

office is also likely to become an ever-more essential feature of
next-generation access topologies. Of course, this goes against
the PON viewpoint of only passive infrastructure between enduser and CO; but since mobile communications require
powered base stations at the intermediate street level, the logic
of the need for active remote nodes is becoming inescapable.
Indeed, by dimensioning the ARN appropriately, power/energy
saving techniques can also be employed to minimize the
carbon footprint. For example, limited statistical multiplexing
can be employed so as to exploit the available network
resources as optimally as possible.
The inevitable presence of such powered, active sites, and
the desire for seamless convergence between fixed and mobile
communications, allows us to take advantage of these active
nodes to create the SODALES architecture with its ARN close
to end-users. Such an innovative architectural approach has
additional critical advantages. In particular, these advantages
relate to the new opportunities for open access capabilities at
the L1 layer through to the L3 layer. This provides tremendous
flexibility to network providers (NPs), passive infrastructure
providers (PIPs), and service providers (SPs) to have their
business cases open to competitors, as increasingly required by
European competition legislation. On the one hand, for
incumbent operators this may present a short-term threat;
however, longer term, such flexibility in open access provision
is expected to also stimulate new SP and NP business models,
with greater efficiencies, greater consumer choice, and greater
technical innovation, such that overall, the digital citizens of
the future will gain greatly in the associated flexibility, scope
of technical advances, and in their individual ability to fully
exploit the opportunities anticipated to emerge from the esociety envisaged for tomorrow.
The existence of the ARN also provides a technical
advantage with respect to the Quality of Service (QoS) in
mobile communications. In particular, by extending network
intelligence closer to the end-user (i.e. out of the CO and into
the ARN) the relevant bandwidth allocations, and medium
access control (MAC) protocols relating to wireless access can
be handled much closer to the end-user, so reducing inherent
latencies caused by distance to the CO (particular where node
consolidation occurs with potentially >100km long-reach
PONs) as well as swifter contention resolution at the ARN.
This allows for a faster, more immediate, and higher quality
ultra-broadband access experience for the end-user. This is also
achieved by exploiting the relevant access network resources
between CO-ARN, as well as ARN-end-user, in a more
efficient manner. On the one hand, the ARN enables greater
energy efficiencies (as measured in joules per bit, J/b) to be
achieved; but also allows the inherent capacity of the access
network to be exploited in a more efficient fashion; i.e. from a
capital expenditure (CAPEX) point of view, less overprovisioning is required, so leading to a leaner, more efficient
deployment of network (and financial) resource. From these
perspectives, the SODALES architecture offers a technically
exciting and innovative perspective for future converged nextgeneration access networking.
Fig. 1. Basic SODALES architecture, serving 96 homes each with 1Gb/s,

and 3 SMEs with 10 Gb/s each, and one Remote Base Station (RBS) with 10
Gb/s bandwidth pipes.
Our basic SODALES architecture for an appropriately

designed ARN architecture is shown in Fig. 1. In this case, the
basic architecture has been designed to offer 1-Gb/s
statistically-multiplexed bandwidth to 96 residential homes, in
addition to providing dedicated 10-Gb/s bandwidth pipes to
each of three SME companies, and finally an additional 10Gb/s bandwidth pipe to a cellular remote base station (RBS)
site, e.g. for 3G, LTE wireless connectivity etc. In this case, we
have designed the ARN to consist of a 120G=340G Ethernet
switch chassis, where the first 40G rack has 4 output ports of
10G capacity each, connected respectively to the RBS and 3
SMEs. In this case, there is no contention or statistical
multiplexing. The other two 40G racks each have 48 output
ports, each offering a maximum 1-Gb/s bandwidth pipe. This
allows the 96 residential homes to be offered a final-drop
symmetrical upstream and downstream bandwidth of 1 Gb/s,
with a statistical multiplexing (over-subscription) ratio of only
1.2 (i.e. 20% over-subscription). The statistical multiplexing
ratio at the upstream port of the chassis is much higher at 6
(i.e. 500%) and represents a significantly useful degree of
optimal exploitation of network resource. As can be seen from
Fig. 1 a total of 20 Gb/s (equal to 2 @10 Gb/s/) light
paths are used between the ARN and the CO. These can be
considered as the allocated wavelengths in a XG-WDM-PON
backhaul system, such that rather than terminating the ONT at
the customer premises as would be conventionally expected,
the ONT is actually located at the ARN, and terminates with
the 120G Ethernet switch. Apart from this, the XG-PON
functions are conventional, with an intermediate arrayedwaveguide grating (AWG) wavelength-routing light paths to
other ARNs, or even allowing dedicated wavelengths to be
directly routed to a 10G ONT located at a RBS or SME
business premises. The XG-PON technology, which is fully
passive and uncontested between CO and ONT, is therefore
fully taken advantage of; but with the ONT located at the ARN,
this enables the flexible final-drop technologies: either wired
(fibre-optical, FTTH) or wireless (radio/mm-wave/optical) to
the residential home, and likewise convergence with mobile
communications. We note that at this stage, the functionality of
the ARN is kept to a minimum. Handover functions (e.g.
vertical, between macro/femto cells, for example; or horizontal,
between femto/femto cells etc.) are not performed at the ARN;
neither is a cognitive nor any co-operative function performed
locally here. As a means to be fully backwardly compatible
1503
with legacy mobile technology, and maintaining a simple (cost

inexpensive) ARN solution, any such network intelligence
functionality is assumed to be performed at the CO. Looking to
the future, once the ARN technological concept becomes more
familiar and common-place, we would expect that further
advanced and intelligent photonic network functionalities, e.g.
energy-efficient caching [13], and/or local routing of local
traffic will also be migrated closer to the end-user at the
actively-powered ARN.
The basic SODALES ARN architecture as indicated in Fig.
1 above has been configured for maximum flexibility with
respect to service provision to residential and light industry (i.e.
SME) end-users, and also to provide front/back-haul bandwidth
to a RBS. However, it could be the case that the ARN is
located in a purely residential area, without any SME or
industry requirements, e.g. in particular, this could be the case
for a block of flats, or other high-density residential location. In
which case, the ARN data bandwidth capacity totaling 30 Gb/s
that was previously allocated towards SME use can now be reallocated to residential users. In this case, a total of 168
residential homes can now each be served with symmetric 1Gb/s data bandwidth, and a RBS with 10 Gb/s.
use in Smart Cities, and machine-to-machine communications

required for the Internet of Things.
III. ARN SIMULATIONS
Network simulations have been carried out in order to
demonstrate the superior performance of the network when
introducing the ARN. The SODALES project is at present
prototyping the ARN to validate its performance in a real
environment and prior to this, simulations have confirmed its
performance. Figure 3 depicts the ARN basic data plane
utilized. The internal non-blocking architecture of the ARN
offers also complete resilience. With this configuration, the
limiting factor of the device will be the input / output ports.
Fig. 3. ARN Basic data plane
Fig. 2. Alternative SODALES ARN architecture, serving 168 homes each

with 1Gb/s, and one Remote Base Station with 10Gb/s bandwidth pipes
The number of additional residential homes served is a

function of the available port-counts of the respective Ethernet
switches. In particular, for the 40G switches, there is a
maximum 48 output ports available; whereas for the 10G
switches this reduces to 24 output ports. In which case, the
three 10G Ethernet switches no longer dedicated to SME
customers can now be reallocated to a total of 324=72
residential homes, so that the total number of homes served by
the ARN comes to 168 domestic units. The presence of a RBS
is still highly desirable, e.g. located on the roof of the block of
flats etc., so as to provide converged and seamless mobile
broadband access. Hence the RBS functionality is maintained
in this alternative ARN architecture as indicated.
Finally, we note that the presence of an ARN intermediate
the CO and end-users can provide important infrastructure
capability for network functions virtualization (NFV), e.g. for
customer equipment, as well as for the ARN equipment
resource optimization. In addition, the ARN also enables a
software-defined virtual platform [15], e.g. to support
ubiquitous ultra-high-speed wireless/wired access, for
providing connectivity and Cloud Services not just to endcustomers, but also to sensors and other actuators for potential
The basic configuration of the ARN offers 20-Gb/s uplink

capacity, by means of 210Gigabit Ethernet ports in a 1+1
redundant configuration. The residential customers are
connected by means of 96Gigabit Ethernet ports while
410 GE ports allow the connection of SMEs and RBS. Thus,
the potential customer capacity is up to 136 Gb/s. This means
that the oversubscription between the uplink and the customer
ports is lower than 1:7. Simulated performance of the
SODALES ARN was modelled by means of discrete-analysis
tools. Real traffic patterns were used as the starting point for
the network load, such that traffic profiles of existing FTTH
networks were used as a baseline for the simulations. These
consisted of triple play services combining voice (VoIP), video
(IPTV) and data. Figure 4 contains the basic simulation model
utilized for the ARN, considering the aforementioned data
plane.
ARN
Fig. 4. ARN simulation model
1504
Data services ranged from 100-Mb/s to 1-Gb/s symmetric.

Simulations were performed for different traffic loads, until the
performance of both systems saturated.
25"
Traffic out SODALES

Traffic out GPON
20"
15"
10"
5"
Tr
a
c"i
n
"( G
bp
s) "
1,
00
"
2,
00
"
3,
00
"
4,
00
"
5,
00
"
6,
00
"
7,
00
"
8,
00
"
9,
00
10 "
,0
0"
11
,0
0
12 "
,0
0
13 "
,0
0
14 "
,0
0
15 "
,0
0
16 "
,0
0
17 "
,0
0
18 "
,0
0
19 "
,0
0
20 "
,0
0
21 "
,0
0
22 "
,0
0
23 "
,0
0
24 "
,0
0
25 "
,0
0
26 "
,0
0
27 "
,0
0
28 "
,0
0
29 "
,0
0"
0"
Fig. 7. Network simulations and comparison of a GPON network and the

SODALES architecture (GPON limited to 2.5 Gb/s downstream capacity and
SODALES to the 210GbE aggregation interfaces)
Fig. 5. Network traffic of a real FTTH network offering 100 Mb/s to 500
subscribers. Note that downstream traffic tends to be limited by the server
side, while upstream reaches maximum customer capacity.
End-users were differentiated depending on their traffic

profile, between residential subscribers, SMEs and traffic
coming from the RBS. Each of the traffic patterns had different
loads depending on the time of the day, and were also taken
from real data. GPON systems provide a data stream of 2.5
Gb/s in the downstream direction, and a 1.25-Gb/s upstream
channel. These are shared typically between 64 users. The
theoretical capacity of the SODALES architecture is 8 times
more than GPON, and unsurprisingly, the total throughput of
both systems was limited by the uplink interface in the
SODALES architecture and the downstream / upstream
maximum data rate in GPON. GPON systems offer optimized
bandwidth allocation protocols that allow between 96% and
98% of the available channel capacity.
Fig. 8. Traffic load of a real FTTH deployment used for the simulations
In any case, the interesting results come from using real

traffic patterns within both architectures, and comparing their
performance. The fact of offering higher data rate input / output
ports allows the system to provide much higher scalability.
Statistical multiplexing factors increase with network capacity,
and therefore, the higher the data rate to the end-user is, the
higher the achieved multiplexing factor. When data rates
exceed 100 Mb/s, the channel utilization of typical end-users
dramatically decreases, as the devices (laptops, smartphones, or
even tablets) connected to the network do not generate enough
traffic to saturate the channel. Also, video streams offer
burstier traffic shapes, which have a better behaviour in a
multiplexing environment.
As more capacity is available, transmission times are
reduced, and this also reduces the time the channel is busy.
Lower times of service improve network latency and system
capacity. In GPON systems, we have simulated end-user
services up to 300 Mb/s with no loss of quality of experience
(QoE), while for the SODALES network a solid 1-Gb/s service
for residential and 10GbE for corporate users can be offered
without having the perception of having a shared media with
99.9% probability. The network uplink barely limits the
performance. Figure 7 contains the comparison between GPON
and SODALES, while Figure 8 depicts the traffic load used of
a real FTTH that has been used for the simulations.
Fig. 6. Traffic baseline for the different users
This envisages two interesting conclusions: the first one is

that GPON can not offer Gigabit data rates using 1:64 splitting
factors (to offer these data rates, the splitting factor should be
1:16 or below); and the second is that the SODALES network
is not limited by the uplink, and therefore more GbE ports can
be added to the ARN or some of them can be upgraded to offer
more 10GbE services. The point is therefore reached where the
1505
available network capacity becomes a commodity and the

channel remains inactive most of the time. This starts to
happen with data rates higher than 300 Mb/s. The fact that
currently there is no application that uses the available capacity
of a GbE access connection could imply that at present, the
killer app of Gigabit Networks is the speed itself and the
instantaneous response of the network (reduced latency).
IV. CONCLUSION AND FUTURE WORK
We have presented the SODALES approach for a
converged access network. The approach is based on the
introduction of an Active Remote Node (ARN), taking
advantage of the power required to feed the base station for
radio access within wired-wireless converged infrastructures.
The rise of mobile communications, and the trend for seamless
convergence between fixed and wireless networks is tending to
make the purely PON approach too restrictive. Indeed, an ultrabroadband 5G wireless network, e.g. evolving from long-term
evolution (LTE) technology, requires ever-smaller cell sizes
(from micro- to pico- and finally to femto-cell sizes), with base
stations (BSs) and remote antenna units (RAUs) in relatively
near vicinities to the end-users. Such BSs and RAUs need to be
powered, e.g. either by a suitable line from the national
electricity grid, or preferably, via local renewable (e.g. solar,
wind) sources. Thus, taking benefit of the power brought to the
BSs and RAUs, we introduce the ARN with further flexibility
and capabilities in order to address the emerging requirements.
We have presented numerous technical advantages of
introducing such an active node in section 2.
We have presented preliminary ARN performance
simulations in order to demonstrate how the ARN provides
higher network performance as compared with current
deployed PON solutions (i.e. GPON). The techno-economics
analysis of installing and deploying the active node in terms of
CapEx is out of scope of this article, although some work
demonstrating a lower cost per Mb/s was presented in [14].
The results presented show how the SODALES network,
including the ARN, is not limited by the uplink, and therefore
more GbE ports can be added to the ARN or some of them can
be upgraded to offer more 10GbE services and therefore
increase the end-user bandwidth or backhauling capacity,
depending on the configuration selected.
Considering the obtained results and the trends towards
ultra-broadband 5G networks to be deployed in 2020, the door
to future research work on the active remote node remains
clearly open. In this sense, there are two main directions on the
future work. On the one hand, there is the issue of ARN
interconnection (for resilience, and more localized traffic
management), which has not yet been rigorously considered.
Thus, one approach may be to consider the use of OFDM-PON
to interconnect the different remote nodes in a metro-access
convergent scenario. On the other hand, and again taking into
consideration the virtue of being powered, the ARN may
exploit the substantial operational and economic advantages in
aggregating traffic at intermediate points with standard

Ethernet switching technology, together with offering Cloud
Services, Content Caching within content delivery networks,
and virtual CPE (V-CPE) features in order to reduce traffic
aggregation at the central office.
ACKNOWLEDGMENT
The research leading to these results and the work
presented here has been partially funded by the European
Commission through the Seventh Framework Programme FP7ICT-2011-8 under the project Software-Defined Access using
Low-Energy Sub-Systems (SODALES), with grant agreement
n 318600.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
1506
L.G. Kazovsky, W.-T. Shaw, D. Gutirrez, N. Cheng, S.-W. Wong,

Next-Generation Optical Access Networks, J. Lightw. Technol., vol.
25, no. 11, Nov. 2007
Cisco Visual Networking Index: Hyperconnectivity and the
Approaching Zettabyte Era, Cisco Systems, San Jose, CA, 2010
Chanclou, P.; Cui, A; Geilhardt, F.; Nakamura, H.; Nesset, D., "Network
operator requirements for the next generation of optical access
networks," Network, IEEE , vol.26, no.2, pp.8,14, March-April 2012
doi: 10.1109/MNET.2012.6172269
ITU-T G.984 Series of Recs., Gigabit-Capable Passive Optical
Networks (G-PON).
ITU-T G.987 Series of Recs., 10-Gigabit-Capable Passive Optical
Networks (XG-PON).
Kani, J.-i.; Bourgart, F.; Cui, A; Rafel, A; Campbell, M.; Davey, R.;
Rodrigues, S., "Next-generation PON-part I: Technology roadmap and
general requirements," Communications Magazine, IEEE , vol.47, no.11,
pp.43,49, November 2009 doi: 10.1109/MCOM.2009.5307465
IEEE 802.3.av Available online at http://www.ieee802.org/3/av/
ITU-T G.989.1 40-Gigabit-capable passive optical networks (NGPON2): General Requirements
Osseiran, A., et al: Scenarios for the 5G Mobile and Wireless
Communications: the vision of the METIS project. IEEE
Communications Magazine, Vol. 52, no. 5, May 2014.
Jungnickel, V., et al: Software-defined open Network Architecture for
Front- and Backhaul in 5G mobile networks. 16th International
Conference on Transparent Optical Networks (ICTON 14), Graz,
Austria, July 2014.
C.Bock, S. Figuerola, M.C. Parker, T. Quinlan, S.D. Walker,
Convergent radio and fibre access architectures using low-energy
systems, ICTON12, Warwick, UK, June 2012
M.C. Parker, R. Martin, K. Guild, S.D. Walker, Hierarchical wireless
and optical access networking: Convergence and energy efficiency, 1st
Workshop on Green Optical Communications, ICTON11, Stockholm,
Sweden, June 2011
Hazem Gomaa, Geoffrey G. Messier, Robert Davies and Carey
Williamson, "Peer-Assisted Caching for Scalable Media Streaming in
Wireless Backhaul Networks," IEEE Global Telecommunications
Conference (GLOBECOM 2010), pp.1-5, Dec. 2010
Ferrer Riera, J., Bock, C., Mendes, T., Parker, M., Jungnickel, V., et al:
Techno Economics and Cost Analysis of Convergent Access
Networks. European Conference on Networks and Communications
(EuCNC 14), June 2014, Bologne, Italy
S. Figuerola, C. Bock, J. Ferrer Riera, E. Escalona, and J. Garcia-Espin,
"SODALES: An Integrated Wired-Wireless Open Access Architecture,"
in Asia Communications and Photonics Conference 2013, OSA
Technical Digest (online) (Optical Society of America, 2013), paper
AW3I.2.
Analytical Performance Model for Poisson Wireless

Networks with Pathloss and Shadowing Propagation
Jean-Marc Kelif1 , Stephane Senecal2 , Marceau Coupechoux3 ,
AbstractThe SINR (signal to interference plus noise ratio)

is a key factor for wireless networks analysis. Indeed, the SINR
distribution allows the derivation of performance and quality of
service (QoS) evaluation. Moreover, it also enables the analysis
of radio resources allocation and scheduling policies, since they
depend on the SINR reached by a UE (User Equipment).
Therefore, it is particularly interesting to develop an analytical
method which allows to evaluate the SINR, in a simple and quick
way, for a realistic environment. Considering a stochastic Poisson
network model, we establish the CDF (cumulative distributed
function) of the SINR. We show that the shadowing can be
neglected, in many cases, as long as mobiles are connected to
their best serving base station (BS), i.e. the BS which offers them
the most powerful useful signal. As a consequence, the analysis
of performance and quality of service, directly derived from the
CDF of SINR, can be established by using a propagation model
which takes into account only the pathloss. Moreover, we establish
that the Fluid network model we have proposed can be used to
analyze stochastic Poisson distributed network. Therefore, the
analysis of stochastic Poisson network can be done in an easy
and quick way, by using the analytical expression of the SINR
established thanks to the Fluid network model.
I. I NTRODUCTION
Mobile services demand is becoming more and more important. Therefore, the estimation of performance and quality
of service (QoS) has to be more and more precise in the
way to answer with accuracy to the demand. Performance
and quality of service evaluations of wireless networks can be
analyzed by using simulations or analytical models. Several
QoS parameters (like throughput, outage probability) can be
derived from the SINR distribution [1] [2] [3] [4].
Moreover the knowledge of the signal to interference plus
noise ratio (SINR) reached by a UE (User Equipment) allows to better perform the radio resource allocation and the
scheduling policy. Analytical models thus try to derive simple
SINR formula in order to quickly evaluate the performance of
a cellular network. Therefore, their analysis need tractable and
accurate models of networks. Two factors play an important
role in the evaluation of the SINR : the localization of BS
and the propagation phenomena. Among the network model
usually considered, the Hexagonal one is the most used.
However this model is based on a regular deployment of BS on
1 Jean-Marc
Kelif is with Orange Labs, France

Email: jeanmarc.kelif@orange.com
2 Stephane Senecal is with Orange Labs, France
Email: stephane.senecal@orange.com
3 Marceau Coupechoux is with Telecom ParisTech, France
Email: marceau.coupechoux@telecom-paristech.fr
4 Constant Bridon is with ENS Cachan and Orange Labs, France
Email: constant.bridon@ens-cachan.fr
978-1-4799-7470-2/14/$31.00 2014 IEEE
Constant Bridon4
an area. The Fluid model, which is another model of network,

considers the interfering base stations as a continuum of
infinitesimal interferers distributed in space. The main interest
of this model consists in its tractability, in the possibility
to establish closed form formula of the SINR, whatever the
location of a UE, and to establish the SINR distribution [5]
[6], too.
In this paper we consider a Poisson model network: the base
stations are randomly distributed according to a spatial Poisson
process [7] [8]. This model allows to take into account more
realistic environments than a hexagonal model of network. Indeed, the distances between base stations are not constant. The
propagation is generally modeled by a term which depends on
the distance from the transmitter, the pathloss. Since in a real
network, the power received at any point of the system also
depends on the local area, another term which characterizes
local area impact can be added. The last term, the shadowing,
is generally expressed as a lognormal distributed function [9].
Papers generally analyze the shadowing impact by considering
users connected to their nearest BS [9] [10]. Recent papers
[11] [12] analyze the impact of shadowing by considering
users connected to the BS which offers the highest useful
signal.
Our contribution: This paper focuses on the impact of the
pathloss and the shadowing on the CDF of the SINR, considering that users are connected to their best serving station : UEs
are connected to the BS which offers the highest useful signal.
We establish that in a Poisson network the CDF of the SINR
calculated by considering a propagation model which takes
into account the pathloss and the shadowing is very close to
the CDF obtained without considering the shadowing (only the
pathloss is considered). This result is very interesting. Indeed,
it allows to use the simple expression of the SINR given by
the Fluid network model. Therefore, the determination of the
quality of service and performance of a wireless network can
be done quickly and in a simple way.
The organization of the paper is as follows. In Section II,
we present the system model. We recall, in Section III, that
performance and QoS of Poisson model can be established
by using a Fluid model when the shadowing is not taken into
account. Section IV expresses the best server SINR. In Section
V, the impact of shadowing is analyzed, considering UEs
connected to the BS which offers the highest useful signal.
In Section VI, some precisions are given about the impact of
the analysis we developed on the study of realistic wireless
systems. A conclusion is given in Section VII.
1507
II. S YSTEM M ODEL

Let us consider a wireless network. We focus on the downlink transmission part. Our aim is to evaluate the performance
and the quality of service of a single user. We consider an
access technology in which the radio resources of a base
station (BS) are divided in a number of parallel, orthogonal,
non-interfering channels (subcarriers), i.e. OFDMA. Therefore, only inter-cell interference is considered, no intra-cell
interference.
A. Network Topology : Poisson Network
In a real network, the inter site distance is variable. The
Poisson model network, characterized by the density of BS,
allows to take it into account by considering a Poisson
distribution of base stations in a given area (Fig. 1). In this
configuration, the cells of the network form a Vorono diagram.
Therefore, it becomes necessary to analyze a wide zone, with
a great number of base stations, to determine the statistical
characteristics of the network in terms of performance and
quality of service.
Du = log2 (1 + u ).
(2)
Let us notice that there are alternative approaches like using

a modified upper bounded Shannon formula or throughputSINR tables coming from physical layer simulations. Moreover, expressions (1) and (2), calculated at any location of
the network, allow an evaluation of the CDF (Cumulative
Distribution Function) of the SINR (or the throughput). The
SINR CDF also provides the outage probability, i.e. the
probability that a user cannot be accepted in the network
since he cannot have a sufficient throughput. It is therefore
important to develop a method which allows to determine these
characteristics with a high accuracy, for a user at any distance
r from his serving BS.
Moreover, since the throughput allows to know the quality
of service that can be offered to a user, these methods make it
possible to determine this characteristic with a high accuracy
in a simple way. In particular, the minimum throughput,
obtained at cell edge, can be derived. By doing an integration
all over the cell range, the average throughput of the cell can be
calculated, too. Dynamical analysis also need the knowledge
of the SINR [13] as input.
III. R EMINDER ABOUT THE F LUID N ETWORK
A. Fluid Network
The fluid network model consists in replacing a given fixed
finite number of transmitters by an equivalent continuous
density of transmitters [14] [15]. Given an inter site distance
2Rc , interferers are characterized by a density BS of BS
starting at a distance of 2Rc from a BS (covering a zone of
radius Rc ), as illustrated on Fig. 2 (Rnw is the size of the
network). The interest of this model is to establish a simple
analytical expression of the SINR.
Fig. 1.
Poisson Network
Continuum
of base stations
Rnw
B. SINR of a user
We consider a single frequency network composed of N
base stations, transmitting at power P on each subcarrier. We
define gi (u) the path gain between BS i and user u on a
given subcarrier. The SINR u of user u served by BS i on
the considered subcarrier is given by:
u = P
P gi (u)
.
P gj (u) + Nth
Rc
2Rc
(1)
j6=i
with Nth the thermal noise on a subcarrier.
Fig. 2.
Fluid model: Network and cell of interest
C. Performance and quality of service

The knowledge of the SINR allows to calculate the throughput that may be reached by a user. Indeed, considering any subcarrier as an AWGN (Additive White Gaussian Noise) channel,
the SINR received by a mobile enables the determination of
the spectral efficiency Du (in bits/s/Hz) by using the Shannon
formula:
B. Calculation of the SINR

We consider a path gain gj (u) = Krj (u), where K is a
constant, rj (u) is the distance between user u and BS j and
the path loss exponent. Let us consider that a BS (eNode-B)
transmits at power P on each subcarrier. Denoting r = ri , we
can express (1) as (dropping u):
1508
r
= P ,
rj
A. Impact of the pathloss on the SINR

(3)
j6=i
We consider a urban environment, where the thermal noise

can be neglected.
By considering expression (3), we can see that the calculation of the analytical expression of the SINR of a user depends
on its location, on the location of its serving BS, and on the
location of each interfering BS.
Considering the fluid model, the SINR only depends on the
distance r of the user to its serving base station [16]:
(r) =
2
r
2BS (2Rc r)2
(4)
If BS is known, this analytical model allows to significantly simplify the calculation of the SINR: the only required
variable is the distance of the mobile to its serving BS. This
model has been proven to be reliable and close to the reality
for homogeneous hexagonal networks [14] [15], as well as for
heterogeneous networks [6].
The cell edge throughput can particularly be calculated
by setting r = Rc in (2) and (4). Therefore, the minimum
performance and quality of service offered to UE is evaluated
in a simple way. Moreover, a simple integration over the cell
range allows to calculate the average throughput of the cell.
These results on the fluid model are valid for a constant
inter site distance 2Rc . In the following section, we remind
that we established [17] a correspondence between a stochastic
Poisson network and the Fluid network model.
C. Comparison of the SINR CDF for the Poisson and Fluid
models
Since the CDF of the SINR characterizes the performance
and the quality of service of wireless systems, we established
a modified expression of the SINR given by the fluid model,
which allows to calculate the CDF of SINR for the Poisson
network reminded hereafter [17]:
ied
SIN RFmodif
= SIN RF luid (a + b)
luid
By considering expression (3), we can see that the calculation of the SINR of a user depends on its location. We notice
that in this case, users are connected to their best serving
station which is also their nearest BS.
B. Pathloss and Shadowing impact on the SINR
1) Propagation: Considering the power Pj transmitted by
the BS j, the power pj,u received by a mobile u can be written:
Yj,u ,
pj,u = Pj Krj,u
(7)
j,u
where Yj,u = 10 10 represents the shadowing effect. The

term Yj,u is a lognormal random variable characterizing the
random variations of the received power around a mean value.
j,u is a Normal distributed random variable (RV), with zero
mean and standard deviation, , comprised between 0 and
, where K is a constant, represents

10 dB. The term Pj Krj,u
the mean value of the received power at distance rj,u from the
transmitter (BSj ). The probability density function (PDF) of
this slowly varying received power is given by

2
ln(s) am
1
exp
(8)
pY (s) =
as
2a
where a = ln1010 , m = a1 ln(KPj rj,u

) is the (logarithmic)
received mean power expressed in decibels (dB), which is
related to the path loss and is the (logarithmic) standard
deviation of the mean received signal due to the shadowing.
2) SINR of a User connected to its nearest BS: Considering
the useful power P0 transmitted by its nearest base station
BS0 , the useful power p0,u received by a mobile u connected
to BS0 can be written:
p0,u = P0 Kr0,u
Y0,u .
(9)
For the sake of simplicity, we now drop index u and set r0,u =
r. The interferences received by u coming from all the other
base stations of the network are expressed by:
(5)
pext =
N
X
Pj Krj Yj .
(10)
j=1
which yields
ied
CDFP oisson CDFFmodif
luid
(6)
where a = 3 and b = -6, for a wide range of values of

(ranging from 2.2 to 4.2, usual range for the path-loss exponent
is comprised between 2.8 and 3.6)
We observed that for = 2.8, 3, 3.6, 3.8, the CDF of the
SINR established by the modified Fluid model and by the
Poisson model are very close: the differences between them
are less than 0.4 dB [17].
IV. SINR CALCULATION WITH PATHLOSS AND
S HADOWING
We consider N interfering base station (BS), a mobile u and
its nearest BS0 .
The SINR at user u is given by:

= PN
j=1
P0 Kr Y0
Pj Krj Yj + Nth
(11)
3) SINR of a User connected to its best serving BS: In this

case, the user is connected to the BS which offers the highest
useful signal. Therefore, the SINR at user u is given by:
maxN
j=0 (Pj Krj Yj )
= PN
.
j6=j Pj Krj Yj + Nth
(12)
where j is the BS which offers the highest useful signal to

the user u. We notice that in this case, due to shadowing, users
are connected to a BS which is not necessary the nearest one.
1509
V. S HADOWING I MPACT ON THE SINR

In this section, we analyze the shadowing impact on the
CDF on the SINR in a system modeled by a Poisson network.
A. Base stations distribution
In order to place the base stations, we set the expected
half inter site distance as Rc . This hypothesis fixes the
density of base stations BS . The values of BS and of the
studied surface SA give the Poissonian characteristic of the
network: the number of BS is drawn according to a Poisson
distribution of parameter BS SA . The surface is chosen to
obtain in average 50 stations in the area. It allows to have a
significant number of cells, representative of a realistic zone
covered by BS, and a significant number of interfering BS for
the computation of the SINR. Those BS are then placed in
the network, with no pairwise constraint (Fig. 1): distances
between neighboring base stations may be very low.
between the CDFs with shadowing and without shadowing is

less than 0.6 dB. Let us notice that for = 2.6, and for an
outage of 5%, the difference , between the curve without
shadowing and the curve with = 8 dB, may reach 1.3 dB.
Therefore, the impact of the shadowing is low in terms
of reachable throughput, coverage (probability of outage) and
QoS.
B. Poisson network SINR computation

The users are uniformly distributed on the whole area SA .
Then, the SINR of a UE is computed from its definition: the
best serving BS gives the power of the received signal, and all
the other stations generate interference. Several Monte Carlo
simulations are run. At each run, the number and locations
of the BS change, whereas the set of studied points (UE) is
fixed. As a result, for the set of studied points, we obtain
the corresponding SINR with different configurations of BS.
Therefore, it becomes easy to compute the CDF of the SINR
received by UE in this zone. Considering a toroidal shape of
the network allows to consider it as virtually infinite with no
edge effect for the computation of the SINR.
Fig. 3. CDF of the SINR with = 2.6 (left) and = 3 (right), for a Poisson
model network and standard deviations of the shadowing values comprised
between 0 dB (no shadowing) (curve in the left of each figure) and 8 dB
(curve in the right of each figure). For readability, only the curves drawn with
= 0 and 8 dB are indicated in the figure. The curves between these two
curves represent the cases = 3 dB, .6 dB. They are in the line thickness
Remark
We consider a model proposed in [18]. The Rayleigh fading,
not considered in this model, may be indistinguishable from
shadowing if the fading is sufficiently slow [18], as for
example if the mobile travels through a region of deep fades
at a very slow speed.
C. Cumulated Distributed Function of the SINR with Shadowing
We consider a Poisson model network. And we calculate
the CDF of the SINR, in the zone covered by this network,
by taking into account the shadowing. The UE are connected
to their best serving station, i.e. the BS which offers the best
signal. The SINR is calculated by using (12).
The figures 3 and 4 show the CDF of the SINR for different
values of the pathloss parameter and for standard deviation
of values 0 dB (no shadowing), 3 dB, 6 dB, and 8 dB. It
can be observed that for values of 3.5 and 4 (Fig. 4), the
curves between 0 and 8 dB are indistinguishable. Only for
= 2.6 (Fig. 3) it is hardly possible to distinguish between the
different curves. These curves show that for typical values of
the pathloss parameter , comprised between 2.6 and 4, and
values of the standard deviation of the shadowing ( 6 dB)
the impact of the shadowing is negligible : the difference
Fig. 4. CDF of the SINR with = 3.5 (left) and = 4 (right), for a Poisson
model network and standard deviations of the shadowing values comprised
between 0 dB (no shadowing) (curve in the left of each figure) and 8 dB
(curve in the right of each figure). For readability, only the values of 0 and 8
dB are indicated in the figure. The curves between these two curves represent
the cases 3 dB, 6 dB. They are in the. line thickness
D. Explanation of these results

In the aim to understand the phenomena which induce
such curves, we analyzed the impact of the shadowing on
the best useful received signal received by users, and on the
interferences (Fig. 5).
These curves show that the CDF of the useful signal
received by users is better with shadowing than without.
1510
taking into account the shadowing. Therefore it becomes possible to analyze realistic wireless networks without considering
the shadowing, in the standard range of values of propagation
pathloss parameter , and standard deviation of shadowing
. As a consequence the CDF of the SINR, established by
Monte Carlo simulation for the Poisson model network is very
close to the one calculated with the analytical expression of
the SINR given by the Fluid model network when a linear
function of the propagation parameter is applied. Therefore,
the analysis of performance, outage probability, throughput,
the radio resource allocation, the scheduling policy can be
done in an easy and quick way, by using the expressions
established with the Fluid Model network.
Fig. 5. CDF of the useful signal and the interferences, for = 2.6 (left)
and = 4 (right), without shadowing ( = 0 dB, blue continuous curves) and
with shadowing ( = 6 dB, red dotted
. .curves)
It is due to the fact that if a user receives a low signal

from its nearest BS, it has a non negligible probability to
receive a better signal from another BS due to shadowing, and
therefore to be connected to that BS than to be connected to
its nearest BS (case without shadowing). It can be observed
that the interferences increase, too. And the increase of the
interferences is at the same order as the increase of the useful
signal. Therefore the two increases compensate each other.
VI. I MPACT ON W IRELESS S YSTEM A NALYSIS
In the previous section, we established the CDF of the SINR
in a system modeled by a Poisson network, and by taking into
account the propagation in terms of pathloss and shadowing.
We showed that the shadowing impact is very low for typical
values of its standard deviation. Indeed, we showed that it
is equivalent, in terms of performance and QoS, to consider
the shadowing and a best serving BS policy, as to consider
a nearest serving BS policy without taking into account the
shadowing. Therefore, the analysis of performances, coverage/capacity and QoS of wireless networks can be done by
taking into account the propagation in a very simple way: by
only considering the pathloss.
Moreover, in Monte Carlo simulators, the implementation
of the shadowing is particularly greedy in simulation time. For
this reason also, it is interesting not to take it into account.
Therefore a simple analytical wireless model which does not
take into account the shadowing, is sufficient for most of the
analysis. The CDF of the SINR can be calculated in a easy
and quick way, by using the expressions (4), (5) and (6) given
by the Fluid network model.
VII. C ONCLUSION
In this paper, we analyzed the joint impact of the pathloss
and the shadowing in Poisson wireless networks by considering a best serving policy for users. We established that results
are very close, in terms of performance and QoS, to the ones
established by considering a nearest serving policy, without
ACKNOWLEDGMENT
This work was supported by the Seventh Framework Program for Research of the European Commission under grant
number HARP-318489.
R EFERENCES
[1] X. Lagrange, Principes et e volutions de lUMTS, Hermes, 2005.
[2] K. S. Gilhousen, I. M. Jacobs, R. Padovani, A. J. Viterbi, L. A. Weaver,
and C. E. Wheatley, On the Capacity of Cellular CDMA System, IEEE
Trans. on Vehicular Technology, Vol. 40, No. 2, May 1991.
[3] S. E. Elayoubi and T. Chahed, Admission Control in the Downlink of
WCDMA/UMTS, Lect. notes comput. sci., Springer, 2005.
[4] A. J. Viterbi, A. M. Viterbi, and E. Zehavi, Other-Cell Interference in
Cellular Power-Controlled CDMA, IEEE Trans. on Communications,
Vol. 42, No. 2/3/4, Freb/Mar/Apr. 1994.
[5] J-M. Kelif, M. Coupechoux, On the Dimensioning of Cellular OFDMA
Networks, Physical Communication Journal, Ref : PHYCOM118, doi :
10.1016/j.phycom.2011.09.008, online October 2011,
[6] J-M. Kelif, W. Diego and S. Senecal, Impact of Transmitting Power
on Femto Cells Performance and Coverage in Heterogeneous Wireless
Networks, Proc. of WCNC, 2012.
[7] F. Baccelli, B. Blaszczyszyn, and F. Tournois Downlink admission/congestion control and maximal load in CDMA networks, in Proc.
of IEEE INFOCOM03, 2003.
[8] A. Agarwal and A. K. Jagannatham, Distributed Estimation in Homogenous Poisson Wireless Sensor Networks, IEEE Wireless Communications Letters, Vol. 3, Feb. 2014
[9] J-M. Kelif, M. Coupechoux and P. Godlewski, Effect of Shadowing on
Outage Probability in Fluid Cellular Radio Networks, WiOpt 2008
[10] J-M. Kelif and M. Coupechoux, Impact of Topology and Shadowing on
the Outage Probability of Wireless Networks, ICC 2009
[11] M. Minelli, M. Coupechoux, J.-M. Kelif, M. Ma, SIR Estimation in
Hexagonal Cellular Networks with Best Server Policy, Wireless Personal
Communications Journal,2012
[12] B. Blaszczyszyn and M. Karray, Wireless Cellular Networks with
Shadowing, NetGCooP, Paris, 12-14 October 2011
[13] L. Rong, S. E. Elayoubi, and O. Ben Haddada. Performance Evaluation
of Cellular Networks Offering TV Services. IEEE Trans. on Vehicular
Technology, 60(2):644655, Feb. 2011.
[14] J-M. Kelif and E. Altman, Downlink Fluid Model of CDMA Networks,
Proc. of IEEE VTC Spring, May 2005.
[15] J-M. Kelif, M. Coupechoux, P. Godlewski, A Fluid Model for Performance Analysis in Cellular Networks, EURASIP Journal on Wireless Communications and Networking, Vol. 2010, Article ID 435189,
doi:10.1155/2010/435189
[16] M. Maqbool, P. Godlewski, M. Coupechoux and J-M. Kelif, Analytical
Performance Evaluation of Various Frequency Reuse and Scheduling
Schemes in Cellular OFDMA Networks, Performance Evaluation Journal, 2009.
[17] J-M. Kelif, S. Senecal, C. Bridon, M. Coupechoux, A Fluid Approach
for Poisson Wireless Networks, arXiv:1401.6336, 2014
[18] A. J. Viterbi, CDMA - Principles of Spread Spectrum Communication,
Addison-Wesley, 1995.
1511

Harp d2 1 1 Proceedings of First Special Session

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Harp d2 1 1 Proceedings of First Special Session

Uploaded by

Copyright:

Available Formats

HARP deliverable: D2.1.

1 Proceedings of first workshop/ special session at an international conference

D2.1.1 Proceedings of first workshop/ special session at international conference

D2.1.1 Proceedinngs of the first workshop/special session at

Introduction and summary ............................................................................................................... 3

General Information about the Workshop ....................................................................................... 4

Workshop Structure and Contributors ............................................................................................. 4

Overview of the Technical Contributions ......................................................................................... 5

Introduction and summary

General Information about the Workshop

Title: International Workshop on Wireless Optical Network Convergence in Support of Cloud

Workshop Structure and Contributors

Overview of the Technical Contributions

Workshop Promotion Activities

This appendix contains the following documents:

Load Modulated MIMO

Ralf Mller (FAU & NTNU)

Load Modulated MIMO

Massive MIMO philosophy:

Small Cell philosophy:

Massive MIMO vs. Small Cells

Ralf Mller (FAU & NTNU)

This work was supported in part by the EU-FP7 project

Constantinos Papadias, Vlasis Barousis

joint work with

Friedrich-Alexander Universitt Erlangen-Nrnberg (FAU)

Load Modulated MIMO:

Load Modulated MIMO

Ralf Mller (FAU & NTNU)

Load Modulated MIMO

Path loss overcome by array gain.

Larger distance to the users.

Massive MIMO philosophy:

Remote radio heads.

Small Cell philosophy:

Massive MIMO vs. Small Cells

Ralf Mller (FAU & NTNU)

Massive MIMO philosophy:

More base station sites required.

Small Cell philosophy:

Massive MIMO vs. Small Cells

Massive use of bandwidth

Load Modulated MIMO

Massive use of bandwidth

Massive use of antenna elements

Ralf Mller (FAU & NTNU)

Load Modulated MIMO

Both systems can operate in arbitrarily strong noise and interference.

Massive MIMO mimics the idea of spread spectrum.

Ralf Mller (FAU & NTNU)

Massive MIMO mimics the idea of spread spectrum.

Load Modulated MIMO

Ralf Mller (FAU & NTNU)

Load Modulated MIMO

For cell densification transmit arrays should be small.

Remote Radio Heads (RHHs)

Ralf Mller (FAU & NTNU)

For cell densification transmit arrays should be small.

Remote Radio Heads (RHHs)

Can massive arrays

Load Modulated MIMO

Ralf Mller (FAU & NTNU)

Load Modulated MIMO