Professional Documents
Culture Documents
Thesis
Thesis
Napoleao Nepomuceno
PHD THESIS
to obtain the title of
Defended by
Napoleão N EPOMUCENO
Acknowledgments
I would like to record my gratitude to my Ph.D advisors. I am indebted to them
more than they know.
It is a pleasure to convey my acknowledge to my coauthors, and I hope to keep
up our collaboration in the future.
I have also benefited by advice, guidance, and help from people from
SME 3Roam, and I am very grateful to them.
I am immensely thankful to my Ph.D. referees for their critical and constructive
comments on this thesis.
I gratefully acknowledge to my colleagues for providing a stimulating and fun
environment in which to work and learn.
I would like to thank my family for the support and love they provided me
throughout my entire life.
And, in particular, I wish to thank Tarciana for being there and for being so
nice. Obrigado, Pretinha.
Network optimization for wireless microwave backhaul
Abstract: Technological breakthroughs have transformed the telecommunications
industry aiming at providing capacity and efficiency to support the increasing de-
mand for wireless broadband services. With the advances in access technologies,
the capacity bottleneck of cellular networks is gradually moving from the radio
interface towards the backhaul – the portion of the network infrastructure that
provides interconnectivity between the access and core networks. The ability for
microwave to be rapidly and cost-effectively deployed is being a crucial point for
successfully tackling the backhaul bottleneck problem.
However, backhaul solutions available with this technology have received little
attention from the scientific community. Nevertheless, the growth of microwave
backhaul networks and their increasing complexity give rise to many interest-
ing optimization problems. In fact, unlike wired networks, the capacity of a
microwave radio link is prone to variations, either due to external factors (e.g.,
weather) or by the action of the network operator. This fundamental difference
raises a variety of new issues to be addressed appropriately. Therefore, more re-
fined approaches for dealing with network optimization in wireless microwave
backhaul need to be conceived.
In this thesis, we investigate network optimization problems related to the de-
sign and configuration of wireless microwave backhaul. We are concerned with
a general class of problems expressed in terms of minimum cost multicommodity
flows with discontinuous step increasing cost functions on the links of the net-
work. These problems are among the most important and challenging problems
in network optimization. Generally, they are computationally very difficult and,
in practice, can only be solved approximately. We introduce mathematical mod-
els for some of these problems and present solution approaches essentially based
on general mixed integer programming, chance-constrained programming, relax-
ation techniques, cutting plane methods, as well as hybrid metaheuristics.
This work was done in collaboration with the SME 3Roam, and partially
developed within the scope of the joint project RAISOM (Réseaux de collecte
IP sans fil optimisés), among INRIA Sophia Antipolis, SME 3Roam, and SME
Avisto. This thesis was developed under joint PhD thesis supervision between the
University of Nice-Sophia Antipolis and the Federal University of Ceará.
List of Figures xv
1 Introduction 1
1.1 Context and motivation . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Our contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Outline of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Microwave communications 9
2.1 Microwave radio system . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Capital and operational costs . . . . . . . . . . . . . . . . . . . 11
2.1.2 Radio spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.3 Performance analysis . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Link power budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Basic terminology . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.2 Power budget calculations . . . . . . . . . . . . . . . . . . . . . 19
2.2.3 Vigants–Barnett model . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.4 Crane rain fade model . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 3Link . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 Models, input and output . . . . . . . . . . . . . . . . . . . . . 22
2.3.2 About 3Link . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Bibliography 103
List of abbreviations
3G/4G Third- and Forth-Generation Cellular Networks
BER Bit Error Rate
BPSK Binary Phase Shift Keying
BSC Base Station Controller
BTS Base Transceiver Station
CAPEX Capital expenditure
FCC Federal Communications Commission
FDD Frequency-Division Duplexing
IDU Indoor Unit
ILP Integer Linear Programming
IP Internet Protocol
ITU International Telecommunications Union
LOS Line-of-Sight
LTE Long Term Evolution
MCF Multicommodity Flow
MCMCF Minimum Cost Multicommodity Flow
MPLS Multiprotocol Label Switching
MSC Mobile Switching Center
ODU Ourdoor Unit
OPEX Operating expenditure
PSTN Public Switched Telephone Network
QAM Quadrature Amplitude Modulation
QPSK Quadrature Phase Shift Keying
RF Radio Frequency
SNR Signal-to-Noise Radio
WDM Wavelength-Division Multiplexing
WiMAX Worldwide Interoperability for Microwave Access
List of Figures
Introduction
BSC/
Radio MSC INTERNET
Interface
Microwave
Fiber / Copper
BTS
Microwave Microwave
Fiber / Copper Fiber / Copper
Radio
Interface Microwave BSC/
Fiber / Copper MSC
BTS BSC/
MSC
PSTN
The backhaul basically comprises two functionalities within the cellular net-
work. The first, often referred to as last mile, is the connection between the base
transceiver stations (BTS) – which furnish radio coverage over a geographical area
to support communications with individual mobile handsets – and the base station
2 Chapter 1. Introduction
the demand for wireless broadband services. The ability for microwave backhaul
networks to be rapidly and cost-effectively deployed is being a crucial point for
successfully tackling the backhaul bottleneck problem, and it is not a coincidence
that over 50% of the world’s base transceiver stations are connected using point-
to-point microwave technologies [Lit09, San09].
Nevertheless, backhaul solutions available with this technology have received
little attention from the scientific community. Apart from original problems in
wireless communications (e.g., channel characterization, power control, and fre-
quency assignment), several interesting problems from network optimization (e.g.,
facility location, capacity planning, and traffic routing) arise in this developing
area when microwave backhaul networks become larger and more complex. Due
to fundamental differences between wireless networks and classical wired net-
works, a variety of questions in this area remain to be adequately addressed and,
doubtlessly, some additional considerations inherent to wireless communications
must be taken into account for better responding some of them.
First, the design of microwave links requires special engineering considera-
tions related to the wireless channel. In fact, while wired channels are stationary
and predictable – links are always characterized by a given capacity and other pa-
rameters –, wireless links are time-varying by nature – e.g., weather conditions can
introduce instantaneously variations into the communication channel – and have
a dynamic behavior – e.g., modulation and coding can be adjusted according to
weather conditions to keep some performance criteria.
In addition, the radio spectrum is a limited natural resource which requires
sensible allocation to many different applications and systems, and coordination
to promote its efficient use. Bandwidth assignment, which is closely related to
network capacity planning, is a highly specialized task requiring a tremendous
amount of engineering and normally subject to renewal upon payment of renewal
fees. Besides, frequency planning plays a critical role to avoid interference within
one’s own network and between other operators’ networks [AHK+ 07, HTRS10].
Furthermore, the design of microwave backhaul networks presents particular
constraints. It is essential for microwave links to have a clear line-of-sight (LOS)
– i.e., there is a direct path without any obstruction (such as buildings, trees, or
mountains) between the communication endpoints –, which strongly dictates the
network topology. In addition, microwave backhaul networks usually have a sig-
nificant dependency on the core network, which impacts the decision of where to
install network nodes and links in order to interconnect all the points of interest to
the core network [Dar07].
Therefore, more refined approaches for dealing with network optimization in
wireless microwave backhaul networks need to be conceived. Besides the techni-
cal task of microwave link engineering, we are especially concerned about network
optimization problems that consider the random variations of wireless channels
and the dynamic behavior of microwave links. This is the main issue that we ad-
dress in this thesis, which was partially developed within the scope of the joint
project RAISOM (Réseaux de collecte IP sans fil optimisés), among INRIA Sophia An-
4 Chapter 1. Introduction
tipolis, SME 3Roam, and SME Avisto, and financed by the SME 3Roam and the
Région PACA through a PhD grant. This thesis was developed under joint PhD
thesis supervision between the University of Nice-Sophia Antipolis and the Fed-
eral University of Ceará.
Chapter 3 covers the minimum cost bandwidth assignment problem under out-
age probability constraints. We introduce (joint) chance-constrained mathematical
programs to tackle this problem and derive their integer linear programming (ILP)
counterparts. To enhance the performance of ILP solvers, we propose cutset-based
valid inequalities. We present a comparative study on the performance of the dif-
ferent formulations and illustrate the price of reliability.
Appendix A focuses on a different research work with which we have been in-
volved during the PhD program. It covers the routing reconfiguration problem
through a game played on digraphs. In this game, we study the tradeoff between
two conflicting parameters: the total number of interruptions and the maximum
number of simultaneous interruptions. We show that minimizing one of these pa-
rameters while the other is constrained is NP-complete. We also prove that there
exist digraphs for which minimizing one of these parameters arbitrarily impairs
the quality of the solution for the other one. Conversely, we exhibit classes of di-
graphs for which good tradeoffs can be achieved. We finally detail the relationship
between this game and the routing reconfiguration problem.
C HAPTER 2
Microwave communications
phone networks, commercial and government mobile radio systems, and oneway television and
radio broadcast systems are all examples of point-to-multipoint microwave networks. Usually, it
consists of one or more broad-beam antennas that are designed to radiate toward multiple end-user
terminals. [And03, Leh10]
2 Wireless technology continues to expand the frequency range at which commercially viable com-
munication systems can be built and deployed. Millimeter-wave is a new generation of point-to-
point radio communication operating at very high frequencies, typically including 71–76 GHz, 81–
86 GHz, and 92–95 GHz. Frequencies up to 300 GHz are also the subject of wireless communications
research. [Man09, Leh10]
10 Chapter 2. Microwave communications
consists of three basic components: the indoor unit (IDU) which performs all dig-
ital processing operations, containing the baseband and digital modem circuitry
and, optionally, a network processing unit that provides advanced networking ca-
pabilities such as routing and load balancing; the outdoor unit (ODU) which houses
all the radio frequency (RF) modules for converting a carrier signal from the mo-
dem to a microwave signal; and the antenna used to transmit and receive the sig-
nal into/from free space, which is typically located at the top of a communication
tower, as illustrated in Figure 2.1. Antennas used in microwave links are highly
directional, which means they tightly focus the transmitted and received energy
mainly into/from one specific direction. To avoid waveguide losses, the antenna
is directly attached to the ODU which, in turn, is connected to the IDU by means
of a single coaxial cable. The distance between the indoor and outdoor equipment
can sometimes be up to 300 meters. Two microwave radios are required to estab-
lish a microwave link (usually operating in duplex mode3 ) between two locations
that can be several kilometers apart. It should be noted that a single IDU can sup-
port multiple ODUs in a same site and, thus, multiple microwave links between
different locations.
waves in the directions determined by the design and orientation of the antenna.
As a radio wave travels through the atmosphere, it experiences different propa-
gation phenomena – e.g., free-space loss, reflection, diffraction, and scattering –
which negatively impact the perceived energy at the receiving antenna. Besides
the transmitted signal, the electromagnetic fields from the interference and noise
sources are also converted to power at the RF receiver, likely leading to impre-
cise interpretation of the transmitted signal. Finally, the RF receiver processes this
power in an effort to recover exactly the source information that was originally
transmitted.
and economic use of frequency bands. In conjunction with ITU regulations, na-
tional legislation instruments establish the availability of frequency bands for spe-
cific applications and the procedures for issuing licenses, as well as the rights and
obligations resulting from using the spectrum.
It is important to distinguish two terms related to the radio spectrum utiliza-
tion: allocation and assignment. Allocation refers to the administration of the ra-
dio frequency spectrum performed by the ITU, which sets out its use by one or
more communication services. An allocation then is a distribution of frequencies
to radio services. Assignment is the authorization given by an administration for a
radio station to use a radio frequency under specified conditions. An assignment
then is a distribution of a frequency to a given radio station [ICT10].
There are a number of frequency bands that have been allocated throughout
the world for use by licensed fixed broadband services. Within the general ITU
band designations, individual countries may elect to implement polices, such as
those by the Federal Communications Commission (FCC) [FCC10] in the United
States, that allow those frequencies to be licensed and used within their country
boundaries. In a given band, there may be requirements for maximum radiated
power levels, particular efficient modulation types, and even standards for the
radiation patterns of directional antennas. These criteria are established to reduce
or minimize interference among systems that share the same spectrum, and to
ensure that the spectrum efficiency is sufficiently high to justify occupying the
spectrum [And03].
2.1. Microwave radio system 13
R = l · b f · lb · es · K (2.1)
where
For example, the annual renewal fee for a bandwidth of 7 MHz at the 8 GHz
frequency band is about €1,800, while for a bandwidth of 28 MHz at the same
frequency band is about €7,000.
first step is the link power budget (See Section 2.2). The result is an estimation
of the signal-to-noise ratio (SNR) value, from which we can obtain some impli-
cations in terms of channel capacity and bit error rate (BER). Given the assigned
channel bandwidth B and the signal-to-noise ratio value S/N, expressed as a lin-
ear power fraction, we can determine an upper bound for the channel capacity C,
in bits per second (bps), assuming that the BER approaches zero when the data
transmission rate is below the channel capacity, according to Shannon’s capacity
theorem [Sha48]:
S [W ]
C [bps] = B[ Hz] · log2 1 + (2.2)
N [W ]
The degree to which a microwave radio system can approximate this limit de-
pends on receiver noise and modulation techniques [ZP00]. The receiver noise is
generated by components used to implement the communication system, and it is
closely tied to the bandwidth. The receiver noise can be calculated by:
N = kTB (2.3)
where
Other sources of noise may arise externally to the system, such as electrical noise
from industry machinery and interference from other users of the radio channel4 .
With regard to the modulation technique, there are several features that influ-
ence the preference for some modulation scheme, such as bandwidth efficiency
and power efficiency. These requirements are conflicting, and existing modula-
tion schemes do not simultaneously perform both of them. Power efficiency de-
scribes the ability of the system to reliably send information at the lowest practical
power level. Bandwidth efficiency describes the ability of a modulation scheme to
accommodate data within a limited (as small as possible) bandwidth. The most
power-efficient modulation methods, like binary phase shift keying (BPSK) and
quadrature phase shift keying (QPSK), present a rather modest bandwidth effi-
ciency. Since spectrum is a very expensive and highly regulated commodity (as
previously discussed in Subsection 2.1.2), good bandwidth efficiency with low
BER is a priority for designers of microwave radio systems [And03, Man09, Leh10].
The most commonly used modulation method for microwave radio systems is
quadrature amplitude modulation (QAM), which employs a combination of am-
4 Noise is usually differentiated from interference in that it may not be identifiable to a given
source and does not carry any useful information. Interference generally refers to identifiable man-
made transmissions [And03].
2.1. Microwave radio system 15
plitude and phase techniques5 . Instead of transmitting one bit at a time, two or
more bits are transmitted simultaneously according to the QAM scheme in use.
An m-QAM scheme presents m different combinations of amplitude and phase,
each one representing a n-bit pattern called a symbol (with n = log2 m and inte-
ger). In practice, given the assigned channel bandwidth B and the m-QAM scheme
in use, we can determine the channel capacity C by:
Figure 2.2: The rectangular representation for QPSK and 16-QAM constellations.
5 A radio wave has three fundamental characteristics – frequency, amplitude, and phase – which
Table 2.3: Bandwidth efficiency and SNR requirement for different QAM schemes.
good practice to keep some power reserve available to flexibly respond to fade or
interference from other links.
Circulator loss. Normally, the output of a transmitter is combined with the re-
ceiver input when the transmitter and receiver share the same antenna. This du-
plexing device has some loss associated with it. This loss value in dB is obtained
from the equipment manufacturer.
Line loss. This is the loss in dB from the transmission line connecting the trans-
mitter with the antenna (or the antenna with the receiver). This loss depends on
the waveguide type and length.
Connector loss. This is the loss in dB introduced by the use of connectors be-
tween transceivers devices, transmission lines, and antennas. This loss depends
on the quantity and the quality of the connectors that are used.
Antenna gain. The antenna gain value, usually given in dBi (dB relative to an
isotropic radiator), is obtained from the antenna manufacturer. It depends on the
antenna type and link operating frequency. The antenna is one of the link system
elements that the design engineer can change to improve link performance.
Antenna radome loss. Special covers for antennas called radomes are available
to protect the horn feed and reduce the wind loading on the tower. The radome
cover on the antenna introduces some loss. This loss value in dB is obtained from
the antenna manufacturer.
Effective radiated power. This is the sum of the transmitter output power and
transmitter antenna gain minus the losses from the circulator, waveguide, and
radome. It is normally expressed in dBm.
Frequency. This is the operating frequency of the link, usually expressed in MHz.
Normally, the nominal center frequency of the channel is used to find the path loss.
Path length. This is the distance, usually expressed in km, from the transmitting
antenna to the receiving antenna.
Free space path loss. This is the loss in signal strength, normally expressed in
dB, that occurs when an electromagnetic wave travels over a line of sight path in
free space.
Atmospheric absorption loss. This is the total atmospheric absorption loss along
the link path. This loss, usually expressed in dB, depends on the path length and
link operating frequency.
Total path loss. This is the sum of the free space and other loss factors along the
path, expressed in dB.
Received signal level. This is the signal strength, usually measured in dBm, at
the input of the radio receiver.
Fade margin. This is the difference, measured in dB, between the received signal
level and the receiver sensitivity. This is an important result for the link availability
calculations and one of the ultimate results of the link budget calculation.
Signal-to-noise ratio. This is the ratio, usually measured in dB, between the re-
ceived signal level and the receiver noise.
Required SNR. This is the SNR value needed at the input of the radio receiver
to achieve a specific BER (normally 10−6 ). This value is found on the equipment
specifications for each modulation scheme.
Link availability. This represents the probability that the BER is at or below a
given quality threshold level (normally 10−6 ). Conversely, an outage is the time
when the link is not available. It can occur for a variety of reasons, including
multipath fades, rain fades, and equipment failures.
2.2. Link power budget 19
where
For a line-of-sight microwave radio system, the primary source of loss is the
decrease of the signal power due to uniform propagation, proportional to the in-
verse square of the distance. The free space path loss equation can be written as
follows:
where
The received signal level PRX can be used to assess other important parameters,
such as the signal-to-noise ratio and fade margin. As previously discussed, the
SNR value is a determinant of the channel capacity. The fade margin, in turn, is
used to calculate the link availability under fading conditions. Fading is a broad
term that is applied to a wide range of variations observed in the signal amplitude,
phase, and frequency characteristics [And03]. Fading phenomena are described
in statistical terms, and the probability of fades of a particular magnitude can be
evaluated through analytical techniques [Bar72, Vig75, Cra80, Cra96], from which
we can estimate the probability of outage and hence the link availability.
20 Chapter 2. Microwave communications
−A
PF = 6.0 × 10−10 C f d3 10 10 (2.7)
where
The Crane rain method [Cra80, Cra96] is used to predict the attenuation by rain on
terrestrial propagation path. The degree of attenuation due to rain is a function of
the rain intensity or rain rate. Rain rates are given in terms of the probability that
the average millimeters of rain that fall over an hour will be exceeded. Calculating
the outage is an iterative process that is based on the calculation of the attenuation.
The rain attenuation is calculated by the Crane method as follows:
where
b = 2.3R−
p
0.17
The rain rate values in mm/h are directly related to the geographical area
where the link will operate. There exist maps and tables that provide an indi-
cation of the appropriate rain rate values for the area where the link will be de-
ployed [And03]. To find the rain outage percentage with the Crane rain model, A R
is iteratively calculated with increasing rain rates for the appropriate zone until
the value of rain attenuation A R equals the fade margin. The percentage of time
corresponding to that rain rate is the rain outage percentage PR .
For path lengths greater than 22.5 km, the calculation is done for D = 22.5 km
and the final adjusted outage probability is calculated as:
D
PR = P22.5 (2.10)
22.5
where PR is the final rain outage percentage and P22.5 is the rain outage percentage
for a path length of 22.5 km. The annual outage probability may be multiplied by
the number of seconds in a year to yield the total number of seconds per year that
the link is unavailable because of rain outages.
2.3 3Link
The design of microwave radio systems commonly makes use of software tools
that implement appropriate and widely accepted propagation prediction models.
Most microwave design tools are developed by radio manufacturers and therefore
are typically biased toward the manufacturers’ own equipment. Many commercial
tools, such as Pathloss 5.0, are also available upon purchase. These tools basically
calculate the performance of a given configuration and, finally, determine whether
it meets the client’s objectives. If the modelling process shows that the system per-
formance is inadequate, then the configuration can be adjusted until the predicted
performance meets the service objectives (if possible), what can take several itera-
tions.
SME 3Roam decided to develop its own software tool, 3Link, that is provided
with its Wireless Ethernet Starter Kit. The main advantage of 3Link lies in the
fact that, besides implementing the classical link power budget to assess the mi-
crowave link performance and availability, it determines the different configura-
22 Chapter 2. Microwave communications
tions available for the microwave link to be deployed according to customers re-
quirements. Essentially, 3Link helps microwave link engineers to answer these
following questions:
1. Will the microwave link transmit data fast enough for the user?
Input. The basic input data are parameters related to the scenario and required
performance targets for the new microwave link to be deployed (See Figure 2.4).
Some of them are informed by the user, while others are retrieved from the 3Link
database. The parameters entered by the user are:
Then, 3Link can determine other parameters, such as the set of bandwidth
choices according to the frequency, the set of modulation schemes depending on
the frequency and bandwidth, and the applicable range of output power for each
frequency, bandwidth and modulation combination.
2.4. Conclusion 23
Output. The basic output computed by 3Link are all valid configurations for the
new microwave link (See Figure 2.5). Each configuration consists of the choice
of antennas, bandwidth, modulation scheme, and output power. In addition, for
each valid configuration, 3Link presents a performance summary that shows how
well the microwave link is predicted, as follows:
Models. The 3Link flowchart diagram is illustrated in Figure 2.6. For each config-
uration (antenna, bandwidth, output power, modulation, etc) 3Link first performs
the noise (Equation (2.3)), path loss (Equation (2.6)), and power budget (Equa-
tion (2.5)) calculations, from which we can estimate the received signal level and
receiver noise. The signal level is used to compute the SNR (which depends on re-
ceiver noise) and fade margin. According to the SNR requirements (Table 2.3) and
the estimated SNR value, we can determine which modulations are available and,
thus, the link’s capacity (Equation (2.4) multiplied by an encoding factor less than
1). The fade margin is used to compute the link’s availability (Vigants–Barnett and
Crane models). Finally, 3Link determines which configurations successfully meet
the performance requirements.
2.4 Conclusion
In this chapter, we presented the main considerations related to the technical task
of conceiving a microwave link. Along with the fundamentals of wireless commu-
nications and the basic blocks of microwave radio systems, we introduced 3Link,
an optimization tool that we have developed for the designing and planning of mi-
crowave links. This work, developed under SME 3Roam supervision, represents
Chapter 2. Microwave communications
24
Figure 2.4: 3Link input screen: scenario and performance parameters.
2.4. Conclusion
25
an important contribution of this thesis, and the effort devoted to it has provided
an indispensable technical background used throughout this thesis.
The methods and analyses presented in this chapter are also used for design-
ing microwave backhaul networks, and it would be very interesting to conceive
a complete network design tool driven by technical constraints – e.g., LOS con-
siderations, sites locations, frequency availability, etc – and performance criteria –
e.g., network CAPEX and OPEX costs, traffic requirements, network reliability, etc
– for determining network topology and configuration solutions. In general, the
optimization problems involved in this task are NP-hard [Dar07]. From a technical
viewpoint, these problems also present high complexity [HART10].
In the following chapters, we focus on two different optimization problems
related to the planning and configuration of microwave backhaul networks. We
consider the main parameters (e.g., channel bandwidth, modulation scheme, and
transmission power) that determine the capacity of a microwave link and impact
the network costs (e.g., renewal fees of licenses and power utilization), while re-
garding different scenarios of optimization.
C HAPTER 3
limited natural resource which has been regulated worldwide to promote its ef-
ficient use. Second, the radio channel is a random and difficult communication
medium. Actually, environment conditions can introduce instantaneously varia-
tions into the communication channel, likely leading to outage events. Therefore,
capacity planning for wireless microwave backhaul requires additional reliability
investigation.
In a single microwave link design, a basic calculation of the link’s performance
can be used to determine whether it meets the required service and reliability. Es-
sentially, the link availability under fading conditions can be determined from the
fade margin derived from the link budget. However, determining the reliability of
a network with several microwave links is a much more difficult task. In fact, the
overall system’s reliability is not just a function of the availability of single links,
but it also depends on the network capacity and data connections to be established.
The capacity of microwave links is basically determined by the channel band-
width and the modulation scheme used to transmit data. On the one hand, the as-
signed bandwidth for each link is a network engineer’s decision subject to obtain-
ing licenses upon payment of renewal fees whose values are usually in accordance
with the amount of spectrum (e.g., in MHz) with which a license is associated. The
amount of spectrum is generally available in modules, resulting in discontinuous
step increasing cost functions on the links.
On the other hand, in response to channel fluctuations, we assume that the
modulation scheme is a random factor. In fact, to overcome outage events, mod-
ern wireless communication systems employ adaptive modulation which has been
shown to considerably enhance radio link performance [GC97, GC98]. Adaptive
modulation refers to the automatic modulation (and other radio parameters) ad-
justment that a wireless system can make to prevent weather-related fading from
causing communication on the link to be disrupted.
Since communication signals are modulated, varying the modulation also
varies the amount of traffic that is transferred per signal. Therefore, in order to
keep the quality of the communication in terms of BER, this technique entails the
variability of the links’ capacity, as shown in Figure 3.1. For instance, 256-QAM
modulation can deliver approximately four times the throughput of QPSK.
Microwave backhaul solutions, although having limited bandwidth and suf-
fering channel impairments, must degrade gracefully as environment conditions
degrade. As a common practice, backhaul operators highly overprovision band-
width during capacity planning to avoid traffic bottlenecks under adverse scenar-
ios (when the performance of some links deteriorates). This approach, however,
incurs additional investments that do not result in resource- and cost-efficient net-
works, besides leading to the inefficient use of the radio spectrum.
Therefore, establishing better wireless backhaul solutions is not just a matter
of adding bandwidth, but it also entails a complex decision aiming at enhancing
network reliability to cope with channel fluctuations.
3.2. Chance-constrained programming 31
P (h j ( x, ζ ) ≥ 0) ≥ 1 − ε, j ∈ J (3.1)
P (h j ( x, ζ ) ≥ 0, j ∈ J ) ≥ 1 − ε (3.2)
P = probability measure
x = vector of decision variables
ζ = vector of random parameters
h( x, ζ ) ≥ 0 = finite system of inequalities
J = index set of the inequalities
ε = infeasibility tolerance ∈ [0, 1]
and u 6= v. Let δ+ (v) denote the set of outneighbors of v and δ− (v) the set
of inneighbors of v. Let Wuv be the number of bandwidth choices available for
arc uv ∈ E. Each bandwidth buvw , for w = 1, . . . , W , is associated with its cost cw
uv uv
w
and a random variable ηuv that represents the number of bits per symbol of the
current modulation scheme (remember that the modulation scheme varies in re-
sponse to channel fluctuations). Let ε uv > 0 be the infeasibility tolerance (typically
near zero) on link uv chosen by the network engineer. The traffic requirements
are defined by K oriented pairs of nodes (sk , tk ), with sk , tk ∈ V and sk 6= tk , and
expected demand dk of pair k = 1, . . . , K.
We want to determine the bandwidth assignment and the traffic flows that
minimize the total bandwidth cost. Let yw uv be the binary decision variable in-
w
dicating whether the bandwidth buv , w = 1, . . . , Wuv , is assigned or not for the
k denote the fraction of dk , k = 1, . . . , K, routed on
arc uv ∈ E. The flow variables f uv
the arc uv ∈ E. The optimization problem can be formulated as follows:
Wuv
min ∑ ∑ cwuv ywuv (3.3)
uv∈ E w=1
−1,
if v = sk ,
∀v ∈ V,
s.t. ∑− k
f uv − ∑+ k
f vu = 1, if v = tk ,
k = 1...K
(3.4)
u∈δ (v) u∈δ (v)
0, otherwise
!
K Wuv
P ∑ dk f uv
k
≤ ∑ ηuv
w w w
buv yuv ≥ 1 − ε uv ∀uv ∈ E (3.5)
k =1 w =1
Wuv
∑ ywuv = 1 ∀uv ∈ E (3.6)
w =1
k
f uv ∈ [0, 1], yw
uv ∈ {0, 1} (3.7)
In this formulation, the objective function (3.3) represents the total bandwidth
cost that is to minimize. The flow conservation property is expressed by (3.4).
It provides the routes for each demand pair, guaranteeing that the traffic require-
ments are entirely fulfilled. Constraints (3.5) ensure that the available capacity on
each link (taking into account the bandwidth choice and the random modulation)
supports all the traffic to be routed through it with (high) probability 1 − ε uv . Fi-
nally, the bandwidth assignment is determined by (3.6). For each link, it forces a
single selection among the possible bandwidths.
Since we have to deal with a finite number of scenarios, this probabilistic pro-
gram can be equivalently written as a standard integer linear program. However,
this equivalent model is highly intractable due to the very large number of scenar-
ios to be considered.
Let us examine different alternatives through an example. Consider a single
microwave link and a unique demand of 70 Mbps. Assume that we have two
bandwidth choices, 10 MHz and 20 MHz, and consider the same discrete proba-
34 Chapter 3. Bandwidth assignment for reliable backhaul
bility distribution function for both bandwidth choices, given in Figure 3.2. Let
ε = 0.05 be the infeasibility tolerance (which corresponds to a reliability level of
0.95).
We can easily calculate the expectation value of the random variable η, E(η ) =
0.80 × 8 + 0.10 × 7 + 0.02 × 6 + 0.04 × 5 + 0.03 × 4 + 0.01 × 2 = 7.56. Note that,
if we use the expectation value for estimating the link’s capacity, the bandwidth
choice of 10 MHz is sufficient to carry all the demand (since the estimated capacity
will be 7.56 × 10 = 75.6 Mbps). However, P (10 · η ≥ 70) = P (η ≥ 7) = 0.90 ≤
0.95. Thus, this solution does not respect the infeasibility tolerance ε = 0.05.
Now let us assume the worst case (lowest modulation scheme) for estimating
the link’s capacity. This way, a feasible solution for this case will be more expensive
but necessarily feasible for all other cases. However, assuming the QPSK modula-
tion scheme (with bandwidth efficiency of 2 bps/Hz), we do not dispose enough
capacity to satisfy the demand of 70 Mbps even if we consider the largest available
bandwidth of 20 MHz (the estimated capacity will be 2 × 20 = 40.0 Mbps).
Here, we use the idea of basic scenarios from Klopfenstein [Klo10] and take ad-
vantage of the problem structure to obtain, in a more efficient manner, the deter-
ministic counterparts of constraints (3.5):
K Wuv
∑ dk f uvk ≤ ∑ nwuv buv
w w
yuv ∀uv ∈ E (3.8)
k =1 w =1
where for each microwave link and bandwidth, the constant nw uv can be easily com-
puted from the cumulative probability distribution of the random variable ηuv w . It
represents the maximum number of bits we can assume taking into account the
infeasibility tolerance ε uv , i.e., the highest modulation scheme whose cumulative
probability value is greater than 1 − ε uv (see Figure 3.3 for an illustration). Note
3.3. Mathematical formulations 35
!
K Wuv
P ∑ dk f uv
k
≤ ∑ w w w
ηuv buv yuv ∀uv ∈ E ≥ 1−ε (3.9)
k =1 w =1
by the arc uv with respect to the bandwidth choice w. Let ρwm uv be the cumulative
w
probability value of the random variable ηuv regarding the modulation m. Now
wm represents the capacity on the arc uv for a given bandwidth choice w and a
buv
specific modulation m. In addition, the binary decision variables y obtain a new
index m that incorporates the assumption on the modulation scheme. The problem
can be rewritten as:
w
Wuv Muv
min ∑ ∑ ∑ cw wm
uv yuv (3.10)
uv∈ E w=1 m=1
−1,
if v = sk ,
∀v ∈ V,
s.t. ∑− k
f uv − ∑+ k
f vu = 1, if v = tk ,
k = 1..K
(3.11)
u∈δ (v) u∈δ (v)
0, otherwise
w
K Wuv Muv
∑ dk f uv
k
≤ ∑ ∑ wm wm
buv yuv ∀uv ∈ E (3.12)
k =1 w =1 m =1
w
Wuv Muv
∏(∑ ∑ ρwm wm
uv yuv ) ≥ 1 − ε (3.13)
uv∈ E w=1 m=1
w
Wuv Muv
∑ ∑ ywm
uv = 1 ∀uv ∈ E (3.14)
w =1 m =1
k
f uv ∈ [0, 1], ywm
uv ∈ {0, 1} (3.15)
Note that now, in the capacity constraints (3.12), we assume explicitly a hy-
pothesis on the modulation scheme. Obviously, for a given link and bandwidth,
the lower the modulation scheme is, the lower will be the capacity assumed to this
link and the higher will also be the probability that the effective capacity on this
link supports all the traffic to be routed through it. In other words, more conser-
vative hypotheses on the modulation scheme will lead to more reliable solutions.
Constraint (3.13) denotes formally this relation. According to the bandwidth
assignment and the hypotheses on the modulation scheme for each microwave
link, this constraint guarantees that the reliability of the solutions is greater than
1 − ε. Note that constraint (3.13) is not linear. Nevertheless, we can derive
an equivalent linear constraint exploiting the properties of logarithmic functions
as well as constraints (3.14) stating that exactly one bandwidth and modulation
scheme are chosen. First, by the monotonicity of any logarithmic function, (3.13)
is equivalent to
3.4. Valid inequalities 37
w !
Wuv Muv
log ∏ ∑ ∑ ρwm wm
uv yuv ≥ log(1 − ε) (3.16)
uv∈ E w=1 m=1
Next, we employ that the logarithm of a product is equal to the sum of the
logarithms:
w !
Wuv Muv
∑ log ∑ ∑ ρwm wm
uv yuv ≥ log(1 − ε) (3.17)
uv∈ E w =1 m =1
The resulting formulation still results in large scale integer linear programs,
which are very hard to solve. In what follows, we introduce some cutset-based
valid inequalities to improve the solving performance for this formulation.
This inequality denotes that there should be enough capacity on the arcs of
the cutset in order to satisfy the demands that must be routed through it. Cutset
inequalities are necessary for a capacity vector to be feasible, but it is well-known
that they are not sufficient in general [CCG09]. In the sequel, we present two types
of strong inequalities to our problem obtained from cutset inequalities by applying
Chvátal-Gomory cutting plane methods (cf. [Wol98]).
38 Chapter 3. Bandwidth assignment for reliable backhaul
Figure 3.4: A cutset in a network consists of the arcs between the shores.
wm , with
Type 1. Given a cutset (S, S), let auv be the smallest capacity coefficient buv
respect to the arc uv ∈ (S, S),
wm
auv = minw=1,...,Wuv minm=1,...,Muv
w b
uv .
where (
1 wm > a
if buv uv
1wm
uv = .
0 otherwise
In general the linear programming (LP) relaxation of (3.10)–(3.12), (3.14), (3.15),
(3.18) does not satisfy (3.21) although all integer solutions have to satisfy it. Hence,
the inequality is valid and can enhance the solving of the ILP. Under certain con-
ditions, (3.21) defines a facet of the convex hull of feasible solutions indicating the
importance of this inequality (cf. [RKOW10] for conditions for similar results).
Type 2. Instead of subtracting auv for all coefficients in (3.19), we can also apply
Chvátal-Gomory rounding directly to (3.19) for a certain value a0 . If we take
a0 = max auv
uv∈(S,S)
w & '
Wuv Muv wm
d(S,S)
buv
∑ ∑ ∑ a0
wm
yuv ≥
a0
(3.22)
uv∈(S,S) w=1 m=1
and has at least |(S, S)| coefficients equal to one. Moreover, the following equal-
ity can be easily obtained as a sum of constraints (3.14) associated with the cut-
set (S, S):
w
Wuv Muv
∑ ∑ ∑ ywm
uv = |( S, S )|, (3.23)
uv∈(S,S) w=1 m=1
which is weaker than (3.21) with a0 instead of a. The same procedure can be ap-
plied for other values a0 . Given a cutset (S, S), let a0uv be the second smallest ca-
wm , with respect to the arc uv ∈ ( S, S ), of the inequality (3.19)
pacity coefficient buv
associated with this cutset, and let a0 be the largest coefficient among all a0uv . The
result of Chvátal-Gomory rounding is again (3.24), but this time at least two co-
wm
l m
buv
efficients a0 − 1 for every arc uv ∈ (S, S) are equal to 0, and the cutset in-
equality is indeed different from (3.21). We refer to the inequalities derived this
way as Type 2. Again, it can be shown that (3.24) defines a facet of the convex hull
of feasible solutions under certain conditions (beyond the scope of our study).
In what follows, we focused on the integer linear program associated with the
joint chance-constrained formulation, since it is in general more appropriate to
cope with practical instances of this problem. All the computations were carried
out on a Linux machine with 3.20 GHz Intel Xeon W5580 CPU (8 Threads) and
64 GB RAM, using IBM ILOG CPLEX 12.1 [CPL10] as underlying solver.
Price of reliability. As we assume the same availability for microwave links us-
ing the highest-level modulations and under the hypothesis that the lowest-level
modulations are sufficiently robust to guarantee an availability of 100%, instead of
explicitly setting the infeasibility tolerance ε, we can specify the maximum num-
ber of links N that we assume using highest-level modulations. To prove that, let
us rewrite (3.18) considering these assumptions (ρw1 uv = 1, cumulative probability
w2
value for lowest-level modulation, and ρuv = ρ, cumulative probability value for
highest-level modulation):
Wuv
∑ ∑ (log(1)yw1 w2
uv + log( ρ ) yuv ) ≥ log(1 − ε ) (3.25)
uv∈ E w=1
Wuv
log(ρ) ∑ ∑ yw2
uv ≥ log(1 − ε ) (3.26)
uv∈ E w=1
Wuv
log(1 − ε)
∑ ∑ yw2
uv ≤
log(ρ)
=N (3.27)
uv∈ E w=1
3.5. Computational results 41
Note that, in this scenario, the parameter N directly implies the infeasibility
tolerance ε and, as the former increases the latter also increases, i.e., the reliability
of the solutions can become worst. Therefore, in order to observe the evolution
of the bandwidth cost as a function of the infeasibility tolerance, we ran tests for
N = 0, 10, . . . , 80.
The solutions were obtained by solving the (enhanced) formulation D (as fur-
ther described in Table 3.1 and below) to optimality, which took several hours
(even days) of computation for each instance. As illustrated in Figure 3.6, the total
bandwidth cost decreases as we admit larger values for the infeasibility tolerance
(N augments). For N = 0, when we have to assume the lowest-level modulations
for all radio links, the network cannot provide enough capacity to satisfy all the
traffic demands, hence this problem is infeasible. For N = 10 (ε = 0.01), the band-
width cost is 38.6% higher than the bandwidth cost for N = 80 (ε = 0.077) and
68.4% higher compared to the case where we do not use any optimization (i.e., we
assign 28 MHz for every microwave link). For N = 60, 70, 80, the probabilistic con-
straint does not really affect the cost of the solutions because the number of links
we need to consider using the highest-level modulations to satisfy all the traffic
requirements is smaller than 60 for these instances.
400
bandwidth cost ( x 1000 )
350
300
250
10 20 30 40 50 60 70 80
parameter N
{v1 , v2 , v3 , v4 , v5 , v6 , v7 , v8 , v9 , v10 },
{v4 , v5 , v9 , v10 , v14 , v15 , v19 , v20 , v24 , v25 },
{v1 , v2 , v3 , v6 , v7 , v11 },
{v1 , v2 , v6 , v7 , v11 , v12 },
{v18 , v19 , v20 , v21 , v22 , v23 , v24 , v25 }.
Due to computational limitations, first a limit of 30,000 nodes (LP relaxations)
of the branch-and-bound process was set for solving each instance. All other solver
settings were preserved at their defaults. In addition, as the size and the complex-
ity of the LP relaxations vary according to each formulation, instead of imposing
a limit on the number of nodes, we also performed tests where we set a time limit
of 1 hour of computation for each instance, preserving all other solver settings at
their defaults.
Figure 3.7 illustrates, for each instance, the optimality gap (given by the ILP
solver) achieved for the different formulations considering first a limit on the num-
ber of nodes (Figure 3.7(a)) and then on the execution time (Figure 3.7(b)), along
with the best feasible solution (Figure 3.7(c)) and lower bound (Figure 3.7(d)) val-
ues obtained for the first scenario (limit on the number of nodes).
With respect to the optimality gap, the adding of valid inequalities improves
the performance of the ILP solver in both scenarios. Formulation B, which consid-
ers only base cutset inequalities, performs significantly better than the basic prob-
lem formulation A. Formulation C, which introduces only the valid inequalities
Type 1 and Type 2, also improves the basic problem formulation A for most cases,
but it does not perform as well as formulation B. Actually, the valid inequalities
Type 1 and Type 2 are more useful in conjunction with base cutset inequalities. In
fact, formulation D presents the best results in terms of the achieved optimality
gap for most cases.
3.6. Conclusion 43
15% 15%
A B C D A B C D
10% 10%
optimality gap
optimality gap
5% 5%
0% 0%
10 20 30 40 50 60 70 80 10 20 30 40 50 60 70 80
parameter N parameter N
(a) Gap within nodes limit (b) Gap within time limit
400 400
A B C D A B C D
solution value ( x 1000 )
300 300
250 250
10 20 30 40 50 60 70 80 10 20 30 40 50 60 70 80
parameter N parameter N
Figure 3.7: Optimality gap, best solution and lower bound values achieved for the
different formulations.
Nevertheless, there are not significant differences among feasible solutions (see
Figure 3.7(c) – lower values mean better solutions) found by each formulation to
explain the better performance of formulations B and D. In Figure 3.8, we show
the performance of the different formulations according to the number of LP re-
laxations for these instances. Note that, in general, the optimality gap decreases
rather quickly until finding a barrier in a given level, when the lower bounds de-
fined by the LP relaxations improve rather slowly, suggesting that much compu-
tational effort is made to prove the optimality of the current feasible solutions.
Formulations B and D coped better to the task of finding tighter lower bounds (see
Figure 3.7(d) – higher values mean better lower bounds), and this explains in part
why these formulations provide lower optimality gaps.
3.6 Conclusion
In this chapter, we have presented a chance-constrained programming approach
to tackle the problem of assigning bandwidth for reliable microwave backhaul net-
works. We introduced different mathematical formulations and proposed cutset-
44 Chapter 3. Bandwidth assignment for reliable backhaul
30% 30%
A B C D A B C D
25% 25%
20% 20%
optimality gap
optimality gap
15% 15%
10% 10%
5% 5%
0% 0%
0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000 0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000
1 1 1 1 2 2 2 2 3 1 1 1 1 2 2 2 2 3
# nodes # nodes
(a) N = 10 (b) N = 20
30% 30%
A B C D A B C D
25% 25%
20% 20%
optimality gap
optimality gap
15% 15%
10% 10%
5% 5%
0% 0%
0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000 0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000
1 1 1 1 2 2 2 2 3 1 1 1 1 2 2 2 2 3
# nodes # nodes
(c) N = 30 (d) N = 40
30% 30%
A B C D A B C D
25% 25%
20% 20%
optimality gap
optimality gap
15% 15%
10% 10%
5% 5%
0% 0%
0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000 0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000
1 1 1 1 2 2 2 2 3 1 1 1 1 2 2 2 2 3
# nodes # nodes
(e) N = 50 (f) N = 60
30% 30%
A B C D
A B C D
25% 25%
20% 20%
optimality gap
optimality gap
15% 15%
10% 10%
5% 5%
0% 0%
0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000 0 2500 5000 7500 0000 2500 5000 7500 0000 2500 5000 7500 0000
1 1 1 1 2 2 2 2 3 1 1 1 1 2 2 2 2 3
# nodes # nodes
(g) N = 70 (h) N = 80
based valid inequalities for this problem. In our computational studies, we dis-
cussed the price of reliability and compared the performance of the different math-
ematical formulations. The results show that there is a tradeoff between band-
width assignment costs and network reliability. In addition, the cutset-based valid
inequalities improved the solving performance of the basic problem formulation.
Many challenges related to this problem remain unaddressed. For example, be-
sides the modulation scheme, we could consider other radio parameters (e.g., cod-
ing and transmission power) that vary in response to the conditions of the channel
and that have a great impact on the capacity of a microwave link. Obviously, more
realistic models while being more accurate also bring more difficulty on the prob-
lem solving.
In addition, it would be interesting to compare our solution to other ap-
proaches used to reduce the vulnerability of wireless networks (e.g., equipment
and frequency diversity). Actually, conceiving reliable solutions for microwave
backhaul networks involves many different design decisions. For example, in or-
der to reduce the impact of signal fluctuations, it could be more cost-efficient to
use an expensive antenna with very high gain than obtain licenses for using a
great amount of spectrum.
As a final remark we want to note that, in contrast to wired networks where
libraries of standardized benchmark instances are accessible [OPTW10], a data li-
brary with realistic data sets for microwave backhaul networks is not available so
far. A characterization of the topologies and traffic behavior of microwave back-
haul networks would be a great contribution for helping engineers who deal with
network optimization in this challenging area.
C HAPTER 4
ably set on the fly –, the traffic fluctuation over the time offers an opportunity to
power mitigation when microwave links are underused.
Even though terrestrial microwave links are not typically concerned with
power efficiency from an economical viewpoint, since they have plenty of power
available and the transmitter output power is very low – usually a few dozen milli-
watts of RF energy produced by the radio transmitter (which is equivalent to a few
hundred watts of RF energy effectively radiated from a highly directional antenna)
–, from a technical point of view it is a good practice to use low transmission power
levels. In fact, reducing transmission power increases the potential for frequency
reuse within a system and reduces possible interference to neighboring systems.
This is especially important in congested urban areas (containing many ran-
domly oriented microwave links), where operators opt for equipment collocation1 .
The collocation trend in the industry can actually create compliance challenges that
operators otherwise would not have encountered. The reason is they must sub-
mit compliance records for their own equipment and for the equipment owned
by collocation tenants at the site. In addition, there exists a public perception of
the safety of microwave radiation2 . In certain situations, and depending on the
site accessibility to the public, if emissions exceed the maximum allowed exposure
levels, any company that contributes to the RF emissions in that area is responsible
for mitigating the problem [Leh10].
Wireless networks have been intensively studied in recent years with a spe-
cific focus on capacity or other QoS parameters and installation costs [ZWZL05,
GPR08, MPR08, ORV10]. Many researches have especially focused on min-
imizing energy consumption in wireless networks, such as minimum energy
broadcasting, backbone construction or monitoring in sensor and ad-hoc net-
works [SRSW05, FKNP07]. Since a huge majority of these works consider energy-
constrained devices (sensors, ad-hoc nodes, etc), most existing solutions are per-
device power optimization. In our settings, in which a global optimum is sought
out, system-wide approaches might be more relevant.
In what follows, we introduce a mathematical programming approach to con-
ceive power-efficient microwave backhaul networks. Under this scenario, we can
define a power-efficient configuration as the minimum transmission output power
needed to achieve the SNR requirement associated with a given modulation. Every
microwave link holds a finite set of power-efficient configuration points, charac-
terized by a modulation scheme and a transmission power level (as illustrated in
Figure 4.1). Note that power-efficient configurations depend on several system’s
1 Collocation is a general concept that refers to multistation sites consisting of numerous transmit-
ters and receivers installed within a limited geographical area. The site often consists of a number
of antennas, typically belonging to different operators, that are all mounted on a common tower or
distributed among a small number of closely positioned towers [Leh10].
2 It should be noted that microwaves fall within the nonionized portion of the electromagnetic
spectrum, below the ionizing portion where X, gamma, and cosmic rays reside. Microwave energy
thus has insufficient energy to ionize atoms and so is unable to change the DNA composition of
human tissue or cause cancer [Man09].
4.2. Mathematical formulations 49
parameters (such as the antenna gain, line losses, and distance between sites) that
define the scenario of application.
90
power-efficient points 79.34
80
70
60
Power (mW)
50
42.81
40
30
18.47
20
11.10
10
4.20
0.88
0
0 28 56 84 112 140 168 196 224
Capacity (Mbps)
arc selection (represented by a continuous line in contrast with the dashed lines).
The basic idea is to set, for each link, the lowest configuration level that supports
all the traffic to be routed through it. The arc selection, conditioned by the traffic
volume on the link, explicitly determines its capacity and power utilization.
the arc uv under the configuration m with respect to the traffic requirement k. The
optimization problem can be formulated as follows:
Muv
min ∑ ∑ cm m
uv yuv (4.1)
uv∈ E m=1
k if v = sk ,
−d ,
Muv Mvu
∀v ∈ V,
s.t. ∑− ∑ mk
xuv − ∑+ ∑ mk
xvu = dk , if v = tk ,
k = 1...K
(4.2)
u ∈ δ ( v ) m =1 u ∈ δ ( v ) m =1
0, otherwise
K
∀uv ∈ E,
∑ xuv
mk m m
≤ buv yuv
m = 1 . . . Muv
(4.3)
k =1
Muv
∑ ym
uv ≤ 1 ∀uv ∈ E (4.4)
m =1
mk
xuv ∈ R+ , y m
uv ∈ {0, 1} (4.5)
In this formulation, the objective function (4.1) represents the total power uti-
lization that is to minimize. For each link, it counts the energy consumption due
to the radio operation at a given transmission power level, defined by its configu-
ration. The flow conservation property is expressed by (4.2). It provides the routes
for each demand pair, guaranteeing that the traffic requirements are entirely at-
tended. By (4.3), it is assured that, on each link, the available capacity according to
its configuration supports all the traffic to be routed through it. Finally, the link’s
configuration choice is determined by (4.4). For each microwave link, it forces a
single selection among the possible power-efficient configurations.
This formulation results in large scale integer linear programs, which are very
hard to solve in practical cases. In addition, solution methods for this problem
have received little attention in the literature. In [GM97], a relaxation that com-
bines both column and constraint generation is used to derive lower bounds to
this problem. In [AT02], a difference of convex function algorithm is applied to
provide feasible solutions. These studies consider general step increasing func-
tions, where convexification may derive poor approximations.
In our application, however, we perceive that the energy cost per bit rises as
we assume higher modulation levels. Taking advantage of this fact, we introduce a
convexification-based relaxation that takes advantage of the inherent convex shape
of the energy cost functions on the links to obtain lower bounds on the power
utilization and determine the network’s configuration.
52 Chapter 4. Power-efficient radio configuration
70
1.305
60
Power (mW)
50
40
0.869
30
20
0.263
0.247
10
0.016 0.059
0
0 28 56 84 112 140 168 196 224
Capacity (Mbps)
Figure 4.4: Piecewise linear convex energy cost function on the links.
due to the decimal format representation of the marginal energy costs). Con-
versely, if we want a capacity of 150 Mbps, the power utilization will be given by
(56 − 0) × 0.016 + (112 − 56) × 0.059 + (140 − 112) × 0.247 + (150 − 140) × 0.263 =
13.746 mW. Note that this later in reality does not represent a possible power-
efficient radio configuration.
The problem can be rewritten as a MCMCF with piecewise linear convex cost
functions. Consider the problem statement, as in the previous subsection, and the
following modifications: now buv m represents the maximum increment of capacity
on the arc uv when we move from the configuration m − 1 to the immediate higher
level m, and cmuv denotes the marginal energy cost into this configuration. As the
marginal cost for routing an amount of traffic over higher QAM schemes is always
increasing, the modulation and the transmission power for each radio link can be
implicitly determined by the variable x of highest configuration level and nonzero
value, i.e. by the flow at the highest QAM scheme. The problem can be then
formulated as follows:
Muv K
min ∑ ∑ ∑ cmuv xuv
mk
(4.6)
uv∈ E m=1 k =1
k if v = sk ,
−d ,
Muv Mvu
∀v ∈ V,
s.t. ∑− ∑ mk
xuv − ∑+ ∑ mk
xvu = dk , if v = tk ,
k = 1...K
(4.7)
u ∈ δ ( v ) m =1 u ∈ δ ( v ) m =1
0, otherwise
K
∀uv ∈ E,
∑ xuv
mk m
≤ buv
m = 1 . . . Muv
(4.8)
k =1
mk
xuv ∈ R+ (4.9)
The total energy cost is now given by a continuous linear function (4.6). The
flow conservation constraints (4.7) remain as in the previous model and implic-
itly provide, besides the routes for each demand pair, the network’s configuration.
54 Chapter 4. Power-efficient radio configuration
Finally, by (4.8), we guarantee that, through every link, the flow over each config-
uration level does not exceed its capacity.
This formulation gives rise to continuous linear programs and can be easily
solved even if we have to deal with very large problem instances. Despite the
fact that the resulting optimal solution of the associated linear program is not a
practical one, it yields lower bounds on the energy consumption. Furthermore,
quite satisfactory solutions can be obtained by means of simple heuristics based
on the fractional optimum.
Particularly, we consider a direct heuristic algorithm that contemplates the op-
timal solution of the relaxation and reassigns, for each microwave link, the lowest
configuration level that supports the network’s flows (obtained by the continu-
ous linear program) through it. In other words, this heuristic establishes feasible
solutions by rounding up the links’ capacity to meet a possible power-efficient con-
figuration. In some cases, it can lead to the underuse of many links that operate
at higher modulation levels to carry a small extra amount of traffic that could be
routed through other links. In what follows, we present an algorithm to overcome
this problem and improve the configuration solutions.
tions research techniques into more powerful algorithms (commonly referred to as hybrid metaheuris-
tics [Rai06]) have been proposed. These algorithms have been shown a very promising strategy, as
they typically represent complementary perspectives over the problem solving process as a whole.
Some of them mainly aim at providing optimal solutions in shorter time, while others primarily
focus on getting better heuristic solutions [PR05].
4 Metaheuristics (e.g., simulated annealing, genetic algorithms, and ant colony optimization) have
succeeded in providing a suitable balance between the efficiency and effectiveness criteria while
tackling many nontrivial optimization problems, even though they cannot furnish any sort of guar-
antee of finding optimal solutions. This success can be evidenced by the diversity of works about
this topic found in the literature (see [BR03, GK03] and references therein).
4.3. Hybrid algorithm 55
each, AB, AC, CD, DB, as illustrated in Figure 4.6. Let us also assume the power-
efficient configuration data for the link scenario used in our simulations (see Ta-
ble 4.1). The relaxation routes the demands on different links, for an overall cost
of 4 × 10 × 0.016 = 0.64 mW. Note that the relaxation computes the energy cost
assuming a capacity of 10 Mbps for each link. Actually, this solution does not rep-
resent a possible radio configuration. Thus, in order to obtain a feasible solution,
the heuristic rounds up the capacity of all links from 10 Mbps to 56 Mbps, for a
total energy cost of 4 × 0.88 = 3.52 mW. The exact model, however, avoids the use
of the link AB, routing the demand AB through the links AC, CD, DB, for an over-
all energy cost of 3 × 0.88 = 2.64 mW. Starting from the relaxation solution, the
hybrid algorithm can also find this optimal solution by rounding up the capacity
of the links AC, CD, DB, while pruning down the capacity of the link AB.
Initial
Solution
Final
Solution
consider the two power-efficient configurations that are derived from the round-
ing up and pruning down operations. Thus, the search space of the metaheuristic
becomes limited to a subset of possible network’s radio configurations. More pre-
cisely, the cardinality of the search space is given by 2n , where n is the number of
links whose flows do not meet the capacity of a power-efficient configuration.
At each iteration of the simulated annealing process, a new network config-
uration is generated from the current solution by switching the configuration of
a single link randomly chosen (among those that present two possible configu-
rations). Then, we apply the filter to check if this new configuration leads to a
feasible solution. In the affirmative case, we attribute the total power utilization
as energy score value of this solution. Otherwise, we generate another network
configuration. We then follow the original description of the simulated annealing
process [KGV83], where the algorithm replaces the current solution by the new so-
lution with a probability that depends on the difference between the correspond-
ing energy score values and a global parameter T (called the temperature).
4.4. Computational results 57
• Distance: 1,000 m.
We used IBM ILOG CPLEX 12.1 [CPL10] as ILP solver to execute both the ex-
act formulation and the model relaxation, as well as to implement the filter of the
hybrid algorithm. As we adopted a time limit of two hours of computation for the
exact formulation, the best feasible solutions found within this limit (from now on
referred to as integer solutions) are not supposed to be optimal. The same time
limit was assumed to the hybrid algorithm. On the other hand, few seconds of
58 Chapter 4. Power-efficient radio configuration
computation are required to solve the model relaxation and generate basic heuris-
tic solutions. The relaxation solutions yield lower bounds on the power utilization,
however they do not represent possible network radio configurations. Conversely
the integer, heuristic and hybrid solutions correspond to practical network radio
configurations.
In order to observe the evolution of the power utilization as a function of the
traffic amount, we have multiplied the demand matrix by a traffic volume factor λ,
initiated at 0.05 and increased by 0.05 until the network’s infrastructure does not
support all the traffic volume anymore. As illustrated in Figure 4.8, the total power
utilization evolves exponentially as the traffic volume increases.
3000
2000
1500
1000
500
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
1,25
1,30
(a) Grid 5 × 5
6000
4000
3000
2000
1000
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
(b) Grid 10 × 10
Figure 4.9 illustrates the gap of the hybrid solutions compared to both the re-
laxation and integer solutions. Remember that the relaxation solutions just give
lower bounds on the power utilization, and the integer ones are the best feasible
solutions found for the exact formulation within the time limit assumed.
350%
Relaxation Integer
300%
250%
200%
Solution gap
150%
100%
50%
0%
-50%
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
1,25
1,30
Traffic volume multiplier
(a) Grid 5 × 5
350%
Relaxation Integer
300%
250%
200%
Solution gap
150%
100%
50%
0%
-50%
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
(b) Grid 10 × 10
Note that the gap between the power utilization of the hybrid and relaxation
solutions abruptly decreases as the network becomes more charged. The hybrid
algorithm also induces a very satisfactory integrality gap (difference between the
power utilization of the hybrid and integer solutions in terms of percent of the
power utilization of the integer solution), particularly for instances on which the
amount of traffic is not small.
The exact approach performed well for all 5 × 5 grid instances, providing good
integer solutions within 2 hours of computation. However, for the 10 × 10 grid
instances, integer solutions were found just for the first six cases. Moreover, the
quality of these solutions was not satisfactory. These problem instances are associ-
ated with huge search spaces, and the ILP solver could not keep up with the task of
exploring well some promising regions within the time limit. None of the integer
solutions was proven to be optimal.
Heuristic feasible solutions based on the fractional optimum of the model re-
laxation were generated for all problem instances, and the execution time has
never exceeded a few seconds. As a drawback, instances on which the network’s
traffic was small have not presented good heuristic solutions. The hybrid algo-
rithm, in turn, clearly improved the heuristic solutions. In addition, for large
instances, the hybrid algorithm achieves solutions that are even better than the
integer ones.
These results are evidenced by the graphical representation of the solutions, as
illustrated in Figures 4.10– 4.13. The figures show the different solutions, side by
side, achieved by each method. In this representation, each arc is labelled by its
capacity. The dash-dotted arcs (in red) of the heuristic solutions represent the links
that have their capacity pruned down by the hybrid algorithm. On the other hand,
the dashed arcs (in blue) of the hybrid and integer solutions represent the links of
each of these solutions that operate at higher configuration levels with respect to
each other.
For a small traffic volume (Figure 4.10), the basic heuristic is inefficient mainly
because the relaxation blindly spreads the traffic among several radio links. There-
fore, by rounding up the links’ capacity, many links may use a higher modulation
level to carry a small extra amount of traffic that could be routed through other
links. The hybrid algorithm, in turn, has overcome this difficulty by aggregating
the traffic flow. Moreover, despite the fact that the hybrid and integer solutions
present very different link distributions, they are both basically composed of con-
nected oriented circuits.
As the traffic volume increases (Figures 4.11 and 4.12), the network configura-
tions become more and more similar. Nevertheless, the hybrid algorithm can still
fine-tune the heuristic solutions to achieve results as good as the integer method.
Finally, for large instances (Figure 4.13), the hybrid algorithm copes better with the
problem complexity than the integer approach.
In general, the computational results show that the hybrid method over-
achieves the heuristic algorithm (reducing the total power utilization of the ba-
sic heuristic solutions) and the exact approach (solving instances that are not
4.4. Computational results 61
reachable by executing the exact mathematical model and finding better solutions
within the same execution time for large instances).
56 56 56 56 56
56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56
56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56
000 000 000 000 000 000
56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56
56 112 112 56 56 112 112 112 56 56 112 56 56 56 56 112 112 56 112 112 56 56 56 56 56 112 56 112
112 140 112 112 112 112 112 112 112 112 112 112
56 56 112 56 56 56 112 56 56 56 112 56
112 56 140 112 112 112 112 140 112 112 112 56 140 112 112 112 112 140 112 112 140 56 140 112 112 112 56 112 112 112
112 112 112 112 56 112 112 112 112 112 112 56
56 56 56 56 56 56 56 56 56 56 56 56
112 56 112 112 112 112 112 140 112 140 112 56 112 112 112 112 112 140 112 112 140 56 112 56 112 112 112 112 56 140
112 56 112 112 56 56 112 56 56 56 112 112
112 112 112 112 112 112 112 112 112 112 112 112
112 112 56 112 112 112 112 112 112 56 112 112 56 112 112 56 112 112 112 56 112 56 56 112 112 56 112 112 112 56
56 56 112 112 56 112 112 56 56 112 112
112 112 112 112 112 112 112 112 112 112 112 56
112 224 224 196 112 224 224 168 112 224 224 168
112 112 112 112 112 112 112 112 112
112 168 112 168 112 140 112 168 56 168 112 168 112 140 112 140 112 168 56 168 112 168 112 140 112 140 112 168 56 168
168 224 224 196 168 224 224 196 168 224 224 196
56 112 168 112 56 112 168 112 56 112 168 112
224 112 224 196 224 224 196 224 196 196 224 112 224 196 196 224 196 224 196 196 224 112 224 196 224 224 168 224 196 196
168 196 196 196 168 196 196 196 168 196 196 196
168 112 140 112 168 112 140 112 168 112 112 112
224 56 224 168 224 224 196 224 112 224 224 56 224 168 224 224 168 224 112 224 224 56 224 168 224 224 168 224 112 224
140 56 196 112 112 56 196 112 112 56 196 112
196 196 168 168 168 196 140 168 168 196 140 168
196 168 140 196 196 140 168 168 196 112 196 168 140 196 196 140 168 168 196 56 196 168 140 196 196 140 168 168 196 56
112 196 168 112 196 168 112 196 168
224 168 196 168 224 168 196 168 224 168 196 168
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 112 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56 56 56 56 56 56 56 56
56 56 56 56 56 56 56 56 56
4.5 Conclusion
In this chapter, we have presented mathematical formulations for determining fea-
sible radio configurations in microwave backhaul networks, focusing on power
efficiency. We introduced a joint optimization of data routing and radio config-
uration that minimizes the total power utilization while handling all the traffic
requirements simultaneously. In particular, we proposed an approximation of the
power utilization on the links by a piecewise linear convex cost function. A heuris-
tic based on the fractional optimum was applied to generate feasible solutions. Fi-
nally, we introduced a hybrid algorithm to improve radio configuration solutions.
The basic idea behind our approach is to provide a power-aware solution that
takes advantage of the traffic fluctuation to keep power-efficient radio configura-
tions. Besides increasing the potential for frequency reuse and reducing possible
interference to neighboring systems, this approach can also improve the perfor-
mance of the operator’s own system since it favors the use of more robust mod-
ulation schemes whenever possible. In view of the scenarios to come, given the
expected high density of sites for 4G systems, this approach is of critical impor-
tance to ensure the proper functioning of the telecommunications infrastructure
necessary to support the growing demand for wireless broadband services.
Nevertheless, note that this method presents rather modest energy savings
compared to those achieved by switching off communication devices [CMN09,
IOR+ 10]. In fact, the operation of outdoor and indoor units represents the bulk
of the energy consumption of microwave radio systems. The impact of saving en-
ergy is significant, particularly in the developing world where energy is a scarce
resource. In this sense, we have some preliminary work, in collaboration with
David Coudert and Issam Tahiri from Mascotte team. Basically, we investigate on
4.5. Conclusion 63
saving energy by selectively turning off idle ODU and IDU communication de-
vices in low-demand scenarios.
Finally, given the computational complexity of the problem studied in this
chapter and knowing the necessity of obtaining efficient solutions in a more rea-
sonable time, we conducted a performance study to investigate alternative models
and solution methods to cope with the task of finding power-efficient configura-
tions in microwave backhaul networks. This study is presented in the following
chapter.
C HAPTER 5
Reoptimizing power-efficient
configurations
Muv
min ∑ ∑ cm m
uv yuv (5.1)
uv∈ E m=1
k k
−d , if v = s ,
Muv Mvu
∀v ∈ V,
s.t. ∑− ∑ mk
xuv − ∑+ ∑ mk
xvu = dk , if v = tk ,
k = 1...K
(5.2)
u ∈ δ ( v ) m =1 u ∈ δ ( v ) m =1
0, otherwise
K
∀uv ∈ E,
∑ xuv
mk m m
≤ buv yuv
m = 1 . . . Muv
(5.3)
k =1
Muv
∑ ym
uv ≤ 1 ∀uv ∈ E (5.4)
m =1
mk
xuv ∈ R+ , y m
uv ∈ {0, 1} (5.5)
In this formulation, the number of continuous and binary variables can be re-
spectively defined by | E| · M · K and | E| · M (assuming that Muv = M, ∀uv ∈ E).
The parameters | E| and K are closely related to the size of the network, while the
parameter M is defined for the number of modulation schemes available to the
microwave radio system (normally a small set). Therefore, for each microwave
link, the radio configuration is associated with a discrete space of power-efficient
solutions (as illustrated in Figure 5.1).
mk
xuv ∈ R+ , y m
uv ∈ [0, 1] (5.6)
Muv K
min ∑ ∑ ∑ cmuv xuv
mk
(5.7)
uv∈ E m=1 k =1
k if v = sk ,
−d ,
Muv Mvu
∀v ∈ V,
s.t. ∑− ∑ mk
xuv − ∑+ ∑ mk
xvu = dk , if v = tk ,
k = 1...K
(5.8)
u ∈ δ ( v ) m =1 u ∈ δ ( v ) m =1
0, otherwise
K
∀uv ∈ E,
∑ xuv
mk m
≤ buv
m = 1 . . . Muv
(5.9)
k =1
mk
xuv ∈ R+ (5.10)
68 Chapter 5. Reoptimizing power-efficient configurations
The main advantage of this relaxation consists in reducing the number of vari-
ables (| E| × M × K, instead of | E| · M · (K + 1) of the classical linear relaxation)
while keeping the problem structure. Note that now every feasible link configura-
tion for this relaxation – which again can or not correspond to a valid configuration
for the exact formulation and then a valid radio configuration in practice – can rep-
resent an optimal solution, each of them associated with a given level of capacity
on the link (as illustrated in Figure 5.3).
5
Linear Convexification-based
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
1,25
1,30
Traffic volume multiplier
(a) Grid 5 × 5
700
Linear Convexification-based
600
500
Execution time (s)
400
300
200
100
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
(b) Grid 10 × 10
lowing, we still use the fact that the marginal energy cost for higher configurations
is always increasing, which ensures an utilization of the arcs from the lowest con-
figuration level to the highest one (considering the progressive multi-arc represen-
tation of a microwave link). But now xuv k represents the flow on the arc uv with
Muv
min ∑ ∑ cm m
uv zuv (5.11)
uv∈ E m=1
−dk ,
if v = sk ,
∀v ∈ V,
s.t. ∑− k
xuv − ∑+ k
xvu = dk , if v = tk ,
k = 1...K
(5.12)
u∈δ (v) u∈δ (v)
0, otherwise
K Muv
∑ k
xuv = ∑ zm
uv ∀uv ∈ E (5.13)
k =1 m =1
∀uv ∈ E,
zm m
uv ≤ buv (5.14)
m = 1 . . . Muv
k
xuv ∈ R+ , z m
uv ∈ R
+
(5.15)
The total energy cost is given by a continuous linear function (5.11) of the vari-
ables z. The flow conservation constraints (5.12) provide the routes for each de-
mand pair. By (5.13), we say that the level of capacity on each microwave link
must be equal to the total flow on it. Finally, by (5.14), we guarantee that, through
every link, the flow on each configuration level does not exceed its capacity. The
number of continuous variables of this formulation can be defined by | E| · (K + M ).
We also applied a widely used approach related to the definition of commodi-
ties when they present the same routing cost (equal to zero in our problem). In this
case, instead of considering an individual commodity for every demand pair, de-
mands are assumed to be aggregated at their source nodes. This approach signif-
icantly reduces the problem size. Figure 5.5 shows a comparison of the execution
time for the original convexification-based relaxation (Relaxation 1), the reformu-
lation presented in this section without demand aggregation (Relaxation 2), and
the reformulation with demand aggregation (Relaxation 3). The refinements intro-
duced in the original convexification-based formulation produce considerable im-
provements on the problem solving, particularly in high-demand scenarios where
the execution time of the original relaxation tends to increase.
1,0
0,6
0,4
0,2
0,0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
1,25
1,30
Traffic volume multiplier
(a) Grid 5 × 5
300
200
150
100
50
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
(b) Grid 10 × 10
Muv K Muv
min ∑ ∑ ∑ (cmuv + λmuv )xuv
mk
− ∑ ∑ λm m
uv buv (5.16)
uv∈ E m=1 k =1 uv∈ E m=1
k if v = sk ,
−d ,
Muv Mvu
∀v ∈ V,
s.t. ∑− ∑ mk
xuv − ∑+ ∑ mk
xvu = dk , if v = tk ,
k = 1...K
(5.17)
u ∈ δ ( v ) m =1 u ∈ δ ( v ) m =1
0, otherwise
mk
xuv ∈ R+ (5.18)
72 Chapter 5. Reoptimizing power-efficient configurations
Figure 5.6 shows a comparison of the power utilization bound obtained by the
original convexification-based formulation (Relaxation) and the lagrangian relax-
ation with different parameters and stopping criteria (Lagrangian(α, β)). On the
one hand, the lagrangian relaxation with subgradient optimization method clearly
converges to the original relaxation solution. On the other hand, we did not obtain
time efficiency on the problem solving. As illustrated in Figure 5.7, the original
relaxation is solved much faster than the lagrangian relaxation. Actually, different
5.3. Lagrangian relaxation 73
reasons can explain this behavior. Firstly, we have used the highest network con-
figuration level to obtain the upper bound cost, and a too large value could make
the steps too long, and hence slowed down the convergence. Secondly, it is not
easy to find the best values for α and β. In general, as these values decrease, the
execution time also decreases. However, we have rapidly lost the quality of the
relaxation solution for too small values for α and β.
2500
Relaxation Lagrangian (5, 30) Lagrangian (5, 50)
Lagrangian (6, 30) Lagrangian (6, 50)
2000
Power utilization (mW)
1500
1000
500
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
1,25
1,30
Traffic volume multiplier
(a) Grid 5 × 5
5000
Relaxation Lagrangian (5, 30) Lagrangian (5, 50)
Lagrangian (6, 30) Lagrangian (6, 50)
4000
Power utilization (mW)
3000
2000
1000
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
(b) Grid 10 × 10
250
Relaxation Lagrangian (5, 30) Lagrangian (5, 50)
Lagrangian (6, 30) Lagrangian (6, 50)
200
Execution time (s)
150
100
50
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
1,25
1,30
Traffic volume multiplier
(a) Grid 5 × 5
5000
Relaxation Lagrangian (5, 30) Lagrangian (5, 50)
Lagrangian (6, 30) Lagrangian (6, 50)
4000
Execution time (s)
3000
2000
1000
0
0,05
0,10
0,15
0,20
0,25
0,30
0,35
0,40
0,45
0,50
0,55
0,60
0,65
0,70
0,75
0,80
0,85
0,90
0,95
1,00
1,05
1,10
1,15
1,20
(b) Grid 10 × 10
Note that, in the exact formulation (5.1)– (5.5), the variables x are easily han-
dled as they are continuous in the problem formulation. However, the variables
y are constrained to be binary, leading to some difficulty on the problem solving.
The basic idea is to partition the original problem into two subproblems (an inte-
ger linear program in y and a continuous linear program in x) and use an iterative
method to generate only a subset of the constraints for the first subproblem.
Suppose that the vector y is fixed to a given network radio configuration y.
In this case, the objective function reduces to a constant cy, and we arrive to a
simple feasibility program. In other words, the original problem can be rewritten
as follows:
( )
Muv
min
y∈{0,1}
∑ ∑ cm m
uv yuv + min+ {0x : (5.2) – (5.3)} : (5.4)
x ∈R
(5.20)
uv∈ E m=1
K Muv
max ∑∑ f (v, k )θvk − ∑ ∑ (buv
m m
yuv )µm
uv (5.21)
v ∈V k =1 uv∈ E m=1
∀uv ∈ E,
s.t. ∑ g(uv, i )θik − µm
uv ≤0 m = 1 . . . Muv , (5.22)
i ∈V k = 1...K
θik ∈ R, µm
uv ∈ R
+
(5.23)
where
k if v = sk
−d ,
f (v, k) = dk , if v = tk
0, otherwise
and
−1, if i = u
g(uv, i ) = 1, if i = v
0, otherwise
K Muv
∑∑ f (v, k )θvk − ∑ ∑ (buv
m m
yuv )µm
uv ≤ 0, ∀(θ, µ) ∈ PD (5.24)
v ∈V k =1 uv∈ E m=1
76 Chapter 5. Reoptimizing power-efficient configurations
We call these the Benders’ inequalities (also called a feasibility cut [CCG09]),
since they are used as cuts in the Benders’ decomposition algorithm. Note that PD
is a full-dimensional polyhedral cone. Any solution in PD is called a ray.
In the exact formulation (5.1)– (5.5), we can replace (5.2)– (5.3) by the condi-
tion (5.24) that states the feasibility of ( P). This leads to the pure integer Benders’
master problem ( M):
Muv
min ∑ ∑ cm m
uv yuv (5.25)
uv∈ E m=1
K Muv
s.t. ∑∑ f (v, k)θvk − ∑ ∑ (buv
m m
yuv )µm
uv ≤ 0, ∀(θ, µ) ∈ PD (5.26)
v ∈V k =1 uv∈ E m=1
Muv
∑ ym
uv ≤ 1 ∀uv ∈ E (5.27)
m =1
ym
uv ∈ {0, 1} (5.28)
5.5 Conclusion
In this chapter, we have revisited mathematical formulations for determining fea-
sible radio configurations in microwave backhaul networks. We introduced sev-
eral refinements to the previous formulations aiming at reducing the execution
time of the problem solving. Computational results corroborated the gain in time
5.5. Conclusion 77
robust and adaptive network solutions. All these challenges involve interdisci-
plinary expertise in communications, signal processing, and network design.
In particular, as future research, we are interested in responding some open
questions related to the problems discussed in Chapters 3 and 4. For example,
with respect to bandwidth assignment for reliable microwave backhaul networks,
we consider the random variations into the communications channel, but neglect
traffic uncertainty (assuming a static demand matrix). As a natural step, we intend
to apply robust optimization approaches to cope with this traffic uncertainty. In
addition, we envisage to study other technological solutions used to reduce the
vulnerability of wireless networks, such as equipment and frequency diversity.
Actually, we think that a global framework which considers the numerous (or at
least the main) decisions impacting the design of microwave backhaul networks
could be conceived to tackle instances of reasonable size.
With respect to the power-efficient radio configuration problem, many techni-
cal questions need to be answered. First, it should be noted that, in reality, adaptive
modulation techniques are based on real-time information of the wireless com-
munication channel. Moreover, decisions on the radio configuration are basically
taken at the network nodes. Therefore, our approach could be used as a centralized
background process (using global information of the network, especially current
traffic requirements) to determine the optimal modulation scheme for each mi-
crowave link. This network configuration target would be sent to the nodes of the
network, while keeping the decisions on the actual radio configuration (based on
both information of the communications channel and the configuration target) at
the network nodes.
In addition, in collaboration with David Coudert and Issam Tahiri from Mas-
cotte team, we are currently investigating the potential for energy savings in mi-
crowave backhaul networks. Since outdoor and indoor units represent the bulk of
the energy consumption of microwave radio systems, we aim at saving energy by
selectively turning off idle ODU and IDU communication devices in low-demand
scenarios. This problem basically relies on a fixed-charge capacitated network de-
sign, which is very hard to optimize.
Finally, we have a particular enthusiasm to study alternative models and so-
lution methods to cope better with these challenging optimization problems. The
work presented in Chapter 5 is a very first step in this direction. Besides, we intend
to carry out theoretical analyses in particular network topologies, such as star, tree,
and ring, to obtain optimal solutions and/or bounds to the problems studied here.
A PPENDIX A
Tradeoffs in routing
reconfiguration problems
We consider a variant of the graph searching games that models the routing recon-
figuration problem in WDM networks. In the digraph processing game, a team
of agents aims at processing, or clearing, the vertices of a digraph D. We are in-
terested in two different measures: 1) the total number of agents used, and 2) the
total number of vertices occupied by an agent during the processing of D. These
measures respectively correspond to the maximum number of simultaneous con-
nections interrupted and to the total number of interruptions during a routing
reconfiguration in a WDM network.
Previous works have studied the problem of independently minimizing each
of these parameters. In particular, the corresponding minimization problems are
APX-hard, and the first one is known not to be in APX. In this work, we give sev-
eral complexity results and study tradeoffs between these conflicting objectives.
In particular, we show that minimizing one of these parameters while the other
is constrained is NP-complete. Then, we prove that there exist some digraphs for
which minimizing one of these objectives arbitrarily impairs the quality of the so-
lution for the other one. We show that such bad tradeoffs may happen even for a
basic class of digraphs. On the other hand, we exhibit classes of graphs for which
good tradeoffs can be achieved. We finally detail the relationship between this
game and the routing reconfiguration problem. In particular, we prove that any
instance of the processing game, i.e. any digraph, corresponds to an instance of
the routing reconfiguration problem.
A.1 Introduction
In this work, we study the digraph processing game, analogous to graph searching
games [FT08]. This game aims at processing, or clearing, the vertices of a contami-
nated directed graph D. For this, we use a set of agents which are sequentially put
and removed from the vertices of D. We are interested in two different measures
and their tradeoffs: the minimum number of agents required to clear D and the
minimum number of vertices that must be covered by an agent. The digraph pro-
cessing game has been introduced in [CPPS05] for its relationship with the routing
reconfiguration problem in Wavelength Division Multiplexing (WDM) networks.
In this context, the goal is to reroute some connections that are established between
pairs of nodes in a communication network, which can lead to interruptions of
82 Appendix A. Tradeoffs in routing reconfiguration problems
R2 Remove an agent from a vertex v of D if all its outneighbors are either pro-
cessed or occupied by an agent, and process v;
A digraph whose vertices have all been processed is said processed. A sequence
of such operations resulting in processing all vertices of D is called a process strat-
egy. Note that, during a process strategy, an agent that has been removed from a
(processed) vertex can be reused. The number of agents used by a strategy on a
digraph D is the maximum number of agents present at the same time in D during
the process strategy. A vertex is covered during a strategy if it is occupied by an
agent at some step of the process strategy.
Figure A.1 illustrates two process strategies for a symmetric digraph D of 7
vertices. The strategy depicted in Figure A.1(a) first puts an agent at vertex x1
(rule R1 ), which let y1 (rule R3 ) be processed. A second agent is then put at r (rule
R1 ) allowing the vertex x1 to be processed, and the agent on it to be removed (rule
R2 ). The procedure goes on iteratively, until all the vertices are processed. The
depicted strategy uses 2 agents and covers 4 vertices. Another process strategy is
depicted in Figure A.1(b) that uses 3 agents and covers 3 vertices. Note that this
latter strategy consists in putting agents at the vertices of a feedback vertex set1 of
minimum size.
Clearly, to process a digraph D, it is sufficient to put an agent at every vertex of
a feedback vertex set F of D (rule R1 ), then the vertices of V ( D ) \ F can be sequen-
tially processed using rule R3 , and finally the vertices of F can be processed and
all agents can be removed (rule R2 ). In particular, a Directed Acyclic Graph (DAG)
can be processed using 0 agent and thus covering no vertices. Indeed, to process
a DAG, it is sufficient to process sequentially its vertices starting from the leaves
(rule R3 ). Note that any process strategy for a digraph D must cover all the vertices
1A set F of nodes of D is a feedback vertex set if the removal of all nodes in F makes D acyclic.
A.1. Introduction 83
1 1 1 2
x1 r x3
y1 x2 y3
y2
2 2 2 2
1 1
2 1 2 1 2
x1 r x3
2 2
y1 x2 y3
y2
Process Number The problem of finding the process number of a digraph D, was
introduced in [CPPS05] as a metric of the routing reconfiguration problem (see
Section A.5). Formally,
Definition 1 The process number of D, denoted by pn( D ), is the smallest p such that
there exists a ( p, ∞)-process strategy for D.
For instance, the digraph D of Figure A.1 satisfies pn( D ) = 2. Indeed, Fig-
ure A.1(a) describes a process strategy using 2 agents, and it is easy to check that
there is no process strategy using at most 1 agent. Digraphs whose process number
is equal to 0 or 1 can easily be identified, as they respectively correspond to acyclic
digraphs, and to graphs whose strongly connected components have a feedback
vertex set of size at most 1 (which can be checked in linear time [CS07]). In [CS07] is
also given an polynomial algorithm to recognize digraphs whose process number
is equal to 2. However the problem of computing the process number of general di-
graphs is NP-complete and not in APX (i.e., admitting no polynomial-time approx-
imation algorithm up to a constant factor, unless P = NP) [CPPS05]. A distributed
polynomial-time algorithm to compute the process number of trees (or forests)
with symmetric arcs has been proposed in [CHM08]. Furthermore, a general
heuristic to compute the process number of a digraph is described in [CHM+ 09].
In [Sol09], Solano conjectured that computing the process number of a digraph can
be solved, or approximated within a constant factor, in polynomial time if the set
of covered vertices is given as part of the input. We disprove this conjecture, show-
ing that computing the process number of a digraph remains not in APX (and so
is NP-complete) in this situation (see Theorem 1).
When considering symmetric digraphs, which can be thought of as a directed
version of an undirected graph, one notices that the process number is closely
related to two other graph invariants, the node search number and the pathwidth.
The node search number of a graph G, denoted by sn( G ), is the smallest p such
that rules R1 and R2 (R3 is omitted) are sufficient to process G using at most p
agents. See [Bre67, DPS02, FT08, KP86, Par78] for more details. The pathwidth
of a (undirected) graph G, denoted by pw( G ), was introduced by Robertson and
Seymour in [RS83]. It has been proved in [EST94] by Ellis et al. that the pathwidth
and the node search number are equivalent, that is for any graph G, pw( G ) =
sn( G ) − 1. The relationship between these parameters and the process number has
been described in [CPPS05]: pw( G ) ≤ pn( G ) ≤ pw( G ) + 1 (and so sn( G ) − 1 ≤
A.1. Introduction 85
pn( G ) ≤ sn( G )), where pn( G ) is the process number of the digraph built from G
by replacing each edge by two opposite arcs. Since computing the pathwidth of
a graph is NP-complete [MHG+ 88] and not in APX [DKL87], determining these
parameters is as hard.
Definition 2 Let m f vs( D ) denote the smallest q such that there exists a (∞, q)-process
strategy for D.
there exist some digraphs for which minimizing one of these objectives arbitrarily
impairs the quality of the solution for the other one. In the following, we define
formally the tradeoff metrics we will now study.
Tradeoff Metrics We introduce new tradeoff metrics in order to study the loss
one may expect on one parameter when adding a constraint on the other. In par-
ticular, what is the minimum number of vertices that must be covered by a process
strategy for D using pn( D ) agents ? Similarly, what is the minimum number of
agents that must be used to process D while covering m f vs( D ) vertices ?
mfvs pn(D)
q = mfvs p(D)
mfvs(D)
p
pn(D) p = pnq (D) pn mfvs(D)
starting from the optimal number of agents pn( D ), each extra agent added allows
to strictly decrease the number of covered vertices, until the optimal, m f vs( D ), is
reached. Let Hn be the symmetric directed star with n ≥ 3 branches of length 2
(for instance, H3 is the digraph of Figure A.1), and let Gk be the graph that con-
sists of the disjoint union of H3 , · · · , Hk , k ≥ 3. Then, for any 0 ≤ i ≤ k − 2,
( pn( Gk ) + i, m f vs( Gk ) + k − 2 − i ) = (2 + i, (k(k + 1)/2) − 5 + k − i ) are minimal
values (this can be easily proved using the easy results described in Section A.2.1).
Our results consist in an analysis of the behaviour of the two given tradeoff mea-
sures both in general digraphs and in symmetric digraphs. As mentioned above,
in general, no process strategy minimizes both the number of agents and the num-
ber of covered vertices (see example in Figure A.1). Hence, we are interested in the
loss on one measure when the other is constrained. In particular, we are interested
m f vs pn (D) m f vs ( D )
in the ratios pn (D)
and m f vspn( D) . This study involves various theorems on the
complexity of estimating this loss (Section A.2) and the existence of digraphs for
which it can be arbitrarily large (Section A.3). We also study in Section A.4 the case
of symmetric digraphs. Finally we describe in Section A.5 the relation between the
routing reconfiguration problem and the processing game.
More precisely, we begin by disproving a conjecture from Solano [Sol09] (The-
orem 1). Then, we prove that for all α ∈ [0, 1], the problems of determining the
parameters α.pnm f vs ( D ) + (1 − α) pn( D ) and α.m f vs pn ( D ) + (1 − α)m f vs( D ) are
NP-complete (Theorem 2). In particular, the problem of determining pnm f vs ( D ) is
not in APX and the problem of determining m f vs pn ( D ) is APX-hard (Theorem 2).
pnm f vs+q ( D )
Then, we prove that for any q ≥ 0 (resp. for any p ≥ 0), the ratio pn( D )
(resp.
m f vs pn+ p ( D )
m f vs( D )
) is not bounded even in the class of bounded process number digraphs
88 Appendix A. Tradeoffs in routing reconfiguration problems
m f vs ( D )
(Theorem 3 and Theorem 4). However we prove that m f vspn( D) ≤ pn( D ) for any
symmetric digraph D (Lemma 1).
In Section A.5, we detail the relationship between the processing game and
the reconfiguration routing problem. In this context, any instance of the routing
reconfiguration problem may be represented by a directed graph, called the de-
pendency digraph of this instance, such that the routing reconfiguration problem
is equivalent to the processing of this digraph. We prove the reverse, that is, any
digraph is the dependency digraph of an instance of the reconfiguration problem
(Theorem 5).
A.2.2 NP-completeness
Before proving that computing the tradeoff parameters introduced in Section A.1.1
are NP-complete, we disprove a conjecture of Solano about the complexity of com-
puting the process number of a digraph D.
Indeed a possible approach for computing the process number, proposed by
Solano in [Sol09], consists of the following two phases: 1) finding the subset of
vertices of the digraph at which an agent will be put, and 2) deciding the order
in which the agents will be put at these vertices. Solano conjectures that the com-
plexity of the process number problem resides in Phase 1 and that Phase 2 can be
solved, or approximated within a constant factor, in polynomial time [Sol09]. We
disprove this conjecture :
Theorem 1 Computing the process number of a digraph is not in APX (and thus NP-
complete), even when the subset of vertices of the digraph at which an agent will be put is
given.
Proof. Let D be any symmetric digraph. Let us consider the problem of computing
an optimal process strategy for D̂ when the set of vertices covered by agents is con-
strained to be V. By the remark in Section A.2.1, such an optimal strategy always
exists. It is easy to check that this problem is equivalent to the one of comput-
ing the node search number (and so the pathwidth) of the underlying undirected
graph of D which is NP-complete [MHG+ 88] and not in APX [DKL87].
Theorem 2 Let α ∈ [0, 1] be fixed. The problem that takes a digraph D as an input and
that aims at determining:
Proof. The two cases for α = 0 clearly holds from the literature.
Corollary 1 For an input digraph D and two integers p ≥ 0 and q ≥ 0, the problems of
determining:
Theorem 3 For any C > 0 and any integer q ≥ 0, there exists a digraph D such that
pnm f vs+q ( D )
pn( D )
> C.
A.3. Behaviour of ratios in general digraphs 91
x11 y1,1
1 y2,1
1
x21
x12 y1,1
2 y2,1
2
x22
x1n y1,1
n y2,1
n
x2n
1,1 2,1 3,1 k,1
x1n+1 ISn ISn ISn ISn x2n+1
P2n,k K1n+1 P1n,k K2n+1 P2n,k
(a) Dn,k of Theorem 4 and Corollary 3 (Case k odd). The red symbol
represents the inexistence of arcs between these subgraphs. The arcs from
V ( Dn,k ) \ V (Kn1 +1 ) to V (Kn1 +1 ) are not represented.
here go
(b) D2,5 in Cor. 3 where the arcs from all vertices to one triangle K31 have been omitted.
Proof. Consider the symmetric directed star Hn defined in Section A.2.1. Let now
D be the digraph consisting of q + 1 pairwise disjoint copies of Hn . So D has q + 1
strongly connected components. We get m f vs( D ) = (q + 1)n. By definition, any
( pnmv f s+q ( D ), m f vs( D ) + q)-process strategy for D covers at most q(n + 1) + n
nodes. Therefore, there exists at least one of the q + 1 strongly connected com-
ponents for which at most n nodes must be covered. Thus to process this com-
ponent, n agents are required. Indeed (n, n) is a minimal value of Hn , and by
definition we cannot decrease the first value without increasing the second one.
pnm f vs+q ( D )
Hence, pnm f vs+q ( D ) = n while pn( D ) = 2. Taking n > 2C, we get pn( D )
> C.
pnm f vs ( D )
Corollary 2 For any C > 0, there exists a digraph D such that pn( D )
> C.
Theorem 4 For any C > 0 and any integer p ≥ 0, there exists a digraph D such that
m f vs pn+ p ( D )
m f vs( D )
> C.
92 Appendix A. Tradeoffs in routing reconfiguration problems
Proof. Let n ≥ 2 and let k ≥ 1 be an odd integer. Let us consider the digraph
Dn,k built as follows. Let ISn1 , · · · , ISnk be k independent sets, each ISnt (1 ≤ t ≤
k) having n vertices: y1t , y2t , . . . , ytn . Let Pn,k be the digraph obtained from the k
independent sets ISnt (1 ≤ t ≤ k) by adding the arcs from yit to ytj +1 , for 1 ≤ j ≤
i ≤ n and t = 1, 3, . . . , k − 2, and from yit to ytj +1 , for 1 ≤ i ≤ j ≤ n and t =
2, 4, . . . , k − 1. Let Kn+1 be the symmetric clique with n + 1 nodes: x1 , x2 , . . . , xn+1 .
The digraph Dn,k is obtained from two copies Pn,k 1 , P2 of P
n,k n,k and two copies
Kn+1 , Kn+1 of Kn+1 , by adding the following arcs. In what follows, yt,a
1 2
j denotes the
jth vertex in the tth independent set of Pn,ka , where j ≤ n, t ∈ {1, k }, a ∈ {1, 2}, and
x aj denotes the jth vertex of Kna , where j ≤ n + 1, a ∈ {1, 2}. There are arcs from xia
k,a
to y1,a b
j , for 1 ≤ i ≤ j ≤ n and a = 1, 2, and from yi to x j , for 1 ≤ i ≤ j ≤ n, a = 1, 2
and b = 3 − a. Finally there is an arc from each node of V ( Dn,k ) \ V (Kn1 +1 ) to each
node of V (Kn1 +1 ). Note that these last arcs are not needed to obtain the results but
help make the proof less technical.
Figure A.3(a) shows the general shape of Dn,k , where the red symbol rep-
resents the inexistence of arcs between these subgraphs. D2,5 is depicted in Fig-
ure A.3(b). For not overloading the figures, the arcs from V ( Dn,k ) \ V (Kn1 +1 ) to
V (Kn1 +1 ) are not represented.
Clearly, m f vs( Dn,k ) = 2n, and any MFVS consists of { x11 , . . . , xn1 } plus n vertices
of Kn2 +1 .
First, note that to process one vertex of Kn1 +1 , there must be a step of any process
strategy for Dn,k where n agents are simultaneously occupying n nodes of Kn1 +1 .
Hence, pn( Dn,k ) ≥ n. Note that, similarly, any process strategy for Dn,k must
occupy n vertices of Kn2 +1 . Moreover, because of the arcs from V ( Dn,k ) \ V (Kn1 +1 )
to V (Kn1 +1 ), any agent that is placed at some vertex in V ( Dn,k ) \ V (Kn1 +1 ) can only
be removed when all vertices of Kn1 +1 are occupied or processed. Consider any
process strategy S for Dn,k (in particular, S uses at least n agents) and let s0 be the
first step of S that does not consist in placing an agent at some vertex of Kn1 +1 . By
above remark, after step s0 − 1 of S, n agents are occupying n vertices of V (Kn1 +1 ).
Up to reorder the first s0 − 1 steps of S, we obtain a process strategy for Dn,k that
starts by placing n agents at n vertices of V (Kn1 +1 ), without increasing the number
of agents used nor the number of vertices occupied by S. Moreover, if the vertex of
V (Kn1 +1 ) that is not occupied is xi1 with i < n + 1, it means that an agent is placed
at xn1 +1 during the first n steps of the strategy. Replacing this operation by the
placement of an agent at xi1 instead of xn1 +1 does not modify the remaining part of
the strategy (but the operation "remove the agent from xn1 +1 " which is replaced by
"remove the agent from xi1 ") since the vertex xn1 +1 can be processed immediately
when the n other vertices of Kn1 +1 are occupied. Hence, we may assume that S
starts by placing agents at { x11 , . . . , xn1 } and then processes xn1 +1 .
Second, any process strategy for any graph can easily be modified, without
increasing (possibly decreasing) the number of used agents nor the number of
occupied vertices, in such a way that the strategy processes all possible vertices
A.3. Behaviour of ratios in general digraphs 93
before placing or removing agents. In other words, the rule R3 can be first applied
without increasing the considered parameters. Therefore, any process strategy
S for Dn,k can be modified , without increasing the number of agents used nor
the number of vertices occupied by S, into a strategy that first places n agents at
{ x11 , . . . , xn1 }, then processes xn1 +1 and all vertices of Pn,k
2 , and finally that mimics S.
Since S is good, it first places n agents at { x11 , . . . , xn1 }, then processes xn1 +1 and
2 . We set x 1 = y0,1 , for all i ≤ n. Let S = S0 . Let 0 ≤ j < k and
all vertices of Pn,k i i
let S j be the strategy that mimics the j first phases of S∗ and then performs in the
same order those movements of S0 that concern the unprocessed vertices at this
step. We prove by induction on j < k that S j can be transformed into the good
process strategy S j+1 for Dn,k satisfying the desired properties without increasing
the number of touched independent sets. Clearly, S0 is a good process strategy for
Dn,k that satisfies these properties.
Assume that, for some 0 ≤ j < k − 1, S j is a good process strategy that satisfies
the desired properties. Then, S j starts by occupying the vertices of { x11 , . . . , xn1 },
processes xn1 +1 and the vertices of Pn,k 2 and then occupies and processes succes-
j,1
sively all vertices of ISnr,1 , r = 1 . . . j until all vertices of ISn are occupied. Let s j be
j+1,1
the step of S j when it occurs. We first prove that S j touches ISn . Indeed, if j is
j+1,1 j+1,1
even, there are n vertex-disjoint paths from yn (resp., from y1 if j is odd) to
2 2 j+1,1 j+1,1
x1 , . . . , xn . While yn (resp., from y1 if j is odd) is not processed, no agent in
94 Appendix A. Tradeoffs in routing reconfiguration problems
j,1
ISn can be removed, and thus only p + 1 ≤ n − 1 agents are available. Therefore,
j+1,1 j+1,1
the only way to process yn (resp., from y1 if j is odd) is to place an agent at
j 0
it. Hence, there is a step of S (hence, of S ) that consists of placing an agent at
j+1,1 j+1,1 j+1,1
yn (resp., y1 if j is odd). Hence, S0 touches ISn . To conclude, we modify
S j by adding after step s j the j + 1th phase of S∗ . That is, after step j, the strategy
j+1,1 j,1
successively occupies the vertices of ISn removing the agents at ISn until all
j+1,1 j,1
vertices of ISn are occupied and all vertices of ISn have been processed. Then,
the strategy mimics the remaining steps of S j . The strategy obtained in such a way
is clearly S j+1 that satisfies all desired properties. In particular, the obtained strat-
egy is a good process strategy for Dn,k that touches the same independent sets as
S0 .
Note that there exists a ( pn( D ) + p + 1, m f vs( D ))-process strategy for the di-
graph Dn,k described in the proof of Theorem 4 whereas the minimum q such that
a ( pn( D ) + p, q)-process strategy for Dn,k exists, is arbitrarily large.
m f vs pn ( D )
Corollary 3 For any C > 0, there exists a digraph D such that m f vs( D )
> C.
Proof. Without loss of generality, we prove the lemma for a connected digraph D.
Let S be a ( pn( D ), m f vs pn ( D ))-process strategy for D = (V, E). Let O ⊆ V be the
set of vertices occupied by an agent during the execution of S. Let F be a MFVS of
D. Let us partition V into (Y, X, W, Z ) = (O ∩ F, O \ F, F \ O, V \ (O ∪ F )). Since
D is symmetric, V \ F is an independent set because it is the complementary of
a MFVS. Since the vertices not occupied by S have all their neighbors occupied,
V \ O is an independent set. Given V 0 ⊆ V, N (V 0 ) denotes the set of neighbors of
the vertices in V 0 . The partition is illustrated in Figure A.4.
First, note that | N (W ) ∩ X | ≤ pn( D )|W |, because, for any vertex v ∈ W to be
processed, all its neighbors must be occupied by an agent. Thus, the maximum
degree of v is pn( D ).
A.4. Behaviour of ratios in symmetric digraphs 95
MFVS V \ MFVS
Y X
occupied
vertices
unoccupied
vertices Independent
W Z
Independent
x1 y1 z1
b x2 y2 z2 d
a f
c e
xn yn zn
K 1,2 IS 1n v IS 2n K 2,1
Figure A.5: Symmetric digraph SDn of Lemma 2 (Figure A.5(a)) and instance of
SDn when n = 5 (Figure A.5(b)). The red symbol represents the absence of arcs.
Lemma 2 For any given ε > 0, there exists a symmetric digraph D such that 3 − ε ≤
m f vs pn ( D )
m f vs( D )
< 3.
Proof. Let n ≥ 1. Let us consider the digraph SDn built as follows. Let ISn1 and
ISn2 be two independent sets of n nodes each: respectively x1 , . . . , xn and z1 , . . . , zn .
Let Kn+1 be a symmetric clique of n + 1 nodes y1 , . . . , yn , yn+1 = v. The digraph
SDn is built starting from the disjoint union of ISn1 , ISn2 , Kn+1 and 6 isolated vertices
{ a, b, c, d, e, f } by adding the following arcs. There are symmetric arcs between the
nodes xi and y j and the nodes zi and y j , for any 1 ≤ i ≤ j ≤ n. Furthermore, all
symmetric arcs of the complete bipartite graph with partitions {b, c} and ISn1 are
added. Similarly, all symmetric arcs of the complete bipartite graph with partitions
{d, e} and ISn2 are added. Finally, the symmetric arcs ( a, b), ( a, c), (d, f ), (e, f ) are
added. The general shape of SDn is depicted in Figure A.5(a). The digraph SD5 is
represented in Figure A.5(b).
Note that the set F = {y1 , . . . , yn , b, c, d, e} is a feedback vertex set of SDn , with
| F | = n + 4. Thus m f vs(SDn ) ≤ n + 4 (actually, one can easily check that F is a
minimum feedback vertex set of SDn ). Clearly, pn(SDn ) ≥ n. In what follows, we
prove that any strategy using n + 1 agents needs to cover at least 3n + 2 vertices,
and we present a (n + 1, 3n + 2)-process strategy for SDn . Since m f vsn+1 ( D ) ≤
m f vs pn ( D ) for any digraph D, the result follows.
First, we prove by contradiction that all process strategies for SDn using n + 1
agents must start by processing either the nodes b and c or the nodes d and e, and
so by placing the n + 1 agents either at vertices a and x1 , . . . , xn or at vertices f and
z1 , . . . , z n .
A.4. Behaviour of ratios in symmetric digraphs 97
Suppose that the first vertex to be processed is either a or belongs to ISn1 , and it
is processed at step s. Therefore, the vertices b and c must be occupied by agents at
this step (such that a can be processed thereafter). Without loss of generality, let us
assume that b is processed, say at step s0 , before c. Since at most n − 1 agents are
available while c and b are occupied, no vertex of the clique Kn+1 can be processed
before step s0 . On the other hand, at step s0 , all vertices of ISn1 are processed or
occupied by agents such that b can be processed. Let X be the subset of vertices
of ISn1 that are occupied at step s0 , and let Y = V ( ISn1 ) \ X. For any xi ∈ Y, yi
must be occupied at step s0 (since xi is processed and yi is not). Hence, at step s0 ,
at least 2 + | X | + |Y | = n + 2 agents are occupying some vertices, a contradiction.
By symmetry, f and any vertex of ISn2 cannot be the first vertex to be processed.
Now suppose that the first vertex to be processed is yi ∈ Kn+1 , i ≤ n + 1. Note
that all vertices of Kn+1 , but yn+1 = v, have at least n + 2 outneighbors. Therefore,
i = n + 1. When v is processed, the n vertices of Kn+1 \ {v} must be occupied,
leaving at most one free agent. But now, all vertices of Kn+1 but v have at least 2
unprocessed outneighbors. Whatever be the placement of the last agent, no other
vertex can be processed and no agents can be released. Hence, the strategy fails, a
contradiction.
Hence, any process strategy using n + 1 agents must start by processing b, c, d
or e. Without loss of generality, (by symmetry), let us assume that the first vertex
to be processed is b. Hence, the strategy must start by placing agents at any vertex
in { a} ∪ V ( ISn1 ). At this step, the strategy processes b and c without covering
them. Then a can be processed and the agent at it is released. At this step, no
other vertex can be processed. Moreover, the only move that can be done is to
place the free agent at yn . Indeed, any other move would let all agents blocked.
Then the free agent is placed at node yn and xn can be processed and the agent
occupying it can be released. Similarly, the strategy sequentially places an agent at
yn−i , processes xn−i and removes the corresponding agent, for 1 ≤ i ≤ n − 1. It is
easy to check that any variation of this would make the strategy immediately fail.
Once all vertices y1 , · · · , yn are occupied, then v can be processed without being
covered. Then, the strategy goes on being highly constrained: for 1 ≤ i ≤ n, the
free agent occupies zi , allowing to process yi and to free the agent occupying it.
Finally, when all vertices of ISn2 are occupied, the free agent must occupy f , and
all remaining vertices may be processed. Again, all these moves are forced for,
otherwise, the strategy would be blocked.
Such a strategy covers 3n + 2 nodes. Therefore, m f vs pn (SDn ) ≥ m f vsn+1 (SDn )
m f vs (SD )+2 m f vs (SD )
= 3n + 2. Hence, m f vspn(SDn n) ≥ 3n 10 pn n
n+4 . For n > ε − 4, we get m f vs(SDn ) ≥ 3 − ε.
Moreover, since SDn has 3n + 7 vertices, we get m f vs pn (SDn ) ≤ 3n + 6, and so
m f vs pn (SDn )
m f vs(SDn )
< 3.
m f vs pn ( D )
Conjecture 1 For any symmetric digraph D, m f vs( D )
≤ 3.
98 Appendix A. Tradeoffs in routing reconfiguration problems
a b a b
(h,c) (h,i) (e,j)
c d e f c d e f
(d,c) (d,b) (i,j)
(e,b)
g h i j g h i j
(a) Old set of routes S1 . (b) New set of routes S2 . (c) Dependency digraph D
from S1 to S2 .
connections, tear down their old routes, and finally, set up the new paths of connec-
tions (h, c), (d, b), (e, j). Such a strategy interrupts a total of 3 connections and these
ones are interrupted simultaneously. Another strategy may consist of interrupt-
ing the connection (h, i ), then sequentially: interrupt connection (h, c), reconfigure
(d, c) without interruption for it, set up the new route of (h, c), then reconfigure
in the same way first (d, b) and (e, b) without interruption for these two requests,
and then (e, j) and (i, j). Finally, set up the new route of (h, i ). The second strategy
implies the interruption of 4 connections, but at most 2 connections are interrupted
simultaneously.
Indeed, possible objectives are (1) to minimize the maximum number of con-
current interruptions [CHM+ 09, CPPS05, Sol09, SP09], and (2) to minimize the
total number of disrupted connections [JS03]. Following [CPPS05, JS03], these
two problems can be expressed through the theoretical game described in Sec-
tion A.1.1, on the dependency digraph [JS03]. Given the initial routing and the
new one, the dependency digraph contains one node per connection that must be
switched. There is an arc from node u to node v if the initial route of connection v
uses resources that are needed by the new route of connection u. Figure A.6 shows
an instance of the reconfiguration problem and its corresponding dependency di-
graph. In Figure A.6(c), there is an arc from vertex (d, c) to vertex (h, c), because
the new route used by connection (d, c) (Figure A.6(b)) uses resources seized by
connection (h, c) in the initial configuration (Figure A.6(a)). Other arcs are built in
the same way.
Given the dependency digraph D of an instance of the problem, a ( p, q)-process
strategy for D corresponds to a valid reconfiguration of the connections where p
is the maximum number of concurrent disruptions and q is the total number of
interruptions. Indeed the three rules can be viewed in terms of reconfiguration of
requests:
R2 Remove an agent from a vertex v of D if all its outneighbors are either pro-
cessed or occupied by an agent, and process v;
Route an interrupted connection when final resources are available;
The next theorem proves the equivalence between instances of the reconfigu-
ration problem and dependency digraphs.
Theorem 5 Any digraph D is the dependency digraph of an instance of the routing re-
configuration problem.
100 Appendix A. Tradeoffs in routing reconfiguration problems
C 2j C 2j
R0 R0
Rk Rk
vk,2j vk,2j+1 vk,2j vk,2j+1
Ri R i
Proof. Roughly, consider a grid network where each initial lightpath of any con-
nection is some row of the grid. If two connections i and k are linked by an arc (i, k )
in the dependency digraph, then we build the new lightpaths of both connections
as depicted in Figure A.7 which actually create the desired dependence. Note that
the lightpath of connection k is deported on an additional row, i.e., a row corre-
sponding to no connection. For each arc of the dependency digraph, we can use
different columns of the grid-network, in such a way that these transformations
may be done independently.
More formally, let D = (V, A) be a digraph with V = {c1 , · · · , cn } and A =
{ a1 , · · · , am }. Let us define the network G as a (n + 2) × (2m) grid such that each
edge of which has capacity one. Let Ri denotes the ith row of G (0 ≤ i ≤ n + 1) and
Ci its ith column (1 ≤ j ≤ 2m), and let vi,j ∈ V ( G ) be the vertex in Ri ∩ Cj . For any
i, 1 ≤ i ≤ n, connection i, corresponding to ci in D, occurs between vi,1 ∈ V ( G )
the leftmost vertex of Ri and vi,2m ∈ V ( G ) the rightmost vertex of Ri , and let the
initial lightpath of connection i follows Ri . Now, we present an iterative method
to build the new lightpath of each connection. Initially, for any i, 1 ≤ i ≤ n,
the new lightpath Pi0 of connection i equals the old lightpath Ri . Now, after the
j −1
( j − 1)th step (0 < j ≤ m) of the method, let Pi be the current value of the
new lightpath of connection i and assume that in the subgraph of G induced by
j −1
columns (C2j−1 , · · · , C2m ), Pi equals Ri . Consider a j = (ci , ck ) ∈ A and let us do
j j −1
the following transformation depicted in Figure A.7. For any ` ∈
/ {i, k }, P` = P` .
j j −1
Now, is defined by replacing the edge (vi,2j−1 , vi,2j ) in
Pi Pi
by the shortest path
from vi,2j−1 to vk,2j−1 (following C2j−1 ), the edge (vk,2j−1 , vk,2j ), and the shortest
j
path from vk,2j to vi,2j (following C2j ). Similarly, Pk is defined by replacing the edge
j −1
(vk,2j−1 , vk,2j ) in Pk by the shortest path from vk,2j−1 to vn+1,2j−1 if i < k (resp., to
v0,2j−1 if i > k), the edge (vn+1,2j−1 , vn+1,2j ) (resp., (v0,2j−1 , v0,2j )), and the shortest
path from vn+1,2j to vk,2j (resp., from v0,2j to vk,2j ). It is easy to check that the grid G,
the sets of initial lightpaths { R1 , · · · , Rn } and final lightpaths { P1m , · · · , Pnm } admit
D as dependency digraph.
A.6. Conclusion 101
Note that a digraph may be the dependency digraph of various instances of the
reconfiguration problem. Since any digraph may be the dependency digraph of a
realistic instance of the reconfiguration problem, Theorem 5 shows the relevance
of studying these problems through dependency digraph notion.
A feasible reconfiguration may be defined by a ( p, q)-process strategy for the
corresponding dependency digraph. Problem (1) is equivalent to minimize p (Sec-
tion A.1.1) and Problem (2) is similar to the one of minimizing q (Section A.1.1).
Consider the dependency digraph D of Figure A.6. From Section A.1.1, we can not
minimize both p and q, that is the number of simultaneous disrupted requests and
the total number of interrupted connections. Indeed there does not exist a (2, 3)-
process strategy while (2, 4) and (3, 3) exist (Figure A.1(a) and Figure A.1(b)).
It is now easy to make the relation between tradeoff metrics introduced in Sec-
tion A.1.1 and tradeoffs for the routing reconfiguration problem. For example,
pnm f vs introduced in Definition 3 represents the minimum number of requests
that have to be simultaneously interrupted during the reconfiguration when the
total number of interrupted connections is minimum. Also Section A.2 shows that
the problems of computing these new tradeoffs parameters for the routing recon-
figuration problem are NP-complete and not in APX. Finally Section A.3 proves
that the loss one can expect on one parameter when minimizing the other may be
arbitrarily large.
A.6 Conclusion
In this work, we address the routing reconfiguration problem through a game
played on digraphs. We introduce the notion of ( p, q)-process strategy and some
tradeoff metrics in order to minimize one metric under the constraint that the other
is fixed. We proved that the problems of computing these parameters are APX-
hard and some are not in APX. We also proved that there exist digraphs for which
minimizing one parameter may increase the other arbitrarily. For further research,
we plan to continue our study for symmetric digraphs in order to (dis)prove Con-
jecture 1. Moreover, it would be interesting to design exact algorithms and heuris-
tics to compute ( p, q)-process strategies.
Bibliography
[AHK+ 07] K. Aardal, S. V. Hoesel, A. M. C. A. Koster, C. Mannino, and A. Sas-
sano. Models and solution techniques for frequency assignment
problems. Annals of Operations Research, 153(1):79–129, 2007. 3
[And03] H. Anderson. Fixed broadband wireless system design. John Wiley &
Sons, 2003. 9, 12, 14, 15, 16, 19, 20, 21, 22
[CGB09] S. Chia, M. Gasparroni, and P. Brick. The next challenge for cellular
networks: backhaul. IEEE Microwave Magazine, 10(5):54–66, 2009. 2
[CHM+ 09] D. Coudert, F. Huc, D. Mazauric, N. Nisse, and J-S. Sereni. Routing re-
configuration/process number: Coping with two classes of services.
In 13th Conf. on Optical Network Design and Modeling (ONDM), 2009.
83, 84, 98, 99
[CPPS05] D. Coudert, S. Perennes, Q.-C. Pham, and J.-S. Sereni. Rerouting re-
quests in WDM networks. In 7ème Rencontres Francophones sur les As-
pects Algorithmiques des Télécommunications (AlgoTel’05), pages 17–20,
mai 2005. 81, 83, 84, 98, 99
[DXC+ 08] L. Dai, Y. Xue, B. Chang, Y. Cao, and Y. Cui. Optimal routing for wire-
less mesh networks with dynamic traffic demand. Mobile Networks
and Applications, 13(2):97–116, 2008. 47
106 Bibliography
[EST94] J.A. Ellis, I.H. Sudborough, and J.S. Turner. The vertex separation and
search number of a graph. Information and Computation, 113(1):50–79,
1994. 84
[GC98] A. Goldsmith and S.-G. Chua. Adaptive coded modulation for fading
channels. IEEE Transactions on Communications, 46(5):595–602, 1998.
30
[PM04] M. Pióro and D. Medhi. Routing, flow, and capacity design in communi-
cation and computer networks. Morgan Kaufmann, 2004. 4
[ZP00] J. Zyren and A. Petrick. Tutorial on basic link budget analysis. Tech-
nical Report AN-9804.1, Intersil, 2000. 14
[ZWZL05] J. Zhang, H. Wu, Q. Zhang, and B. Li. Joint routing and schedul-
ing in multi-radio multi-channel multi-hop wireless networks. In 2nd
IEEE International Conference on Broadband Networks (BROADNETS’05,
pages 678–687, 2005. 48