Ph.D.

Dissertation

Wireless Sensor Networks in the Future Internet of

Things: Density, Mobility, Heterogeneity and Integration

Marc Barceló Lladó

Thesis Advisors: José López Vicario and Antoni Morell Pérez

Department of Telecommunications and Systems Engineering,
Escola d’Enginyeria,
Universitat Autònoma de Barcelona (UAB).

September, 2015
 To my parents and my brother,

“They did not know it was impossible, so they did it.”

- Mark Twain

i
 Abstract

Wireless sensor networks (WSNs) are expected to revolutionize the way we live, work,
and interact with the physical environment. Although WSNs have been in the spotlight
of the research community for the last decade, their performance in practical implemen-
tations is still far behind the theoretical results. This is mainly due to the practical
issues that arise in real-life scenarios. As a result, WSNs are generally limited to simple
environmental sensing applications. The aim of this thesis is to reduce the gap between
the theoretical and real potential of WSNs, and therefore increase their integration in
society. In particular, this thesis focuses on the following four practical obstacles: high
node density, node mobility, traffic heterogeneity and integration into the future Inter-
net of Things (IoT). First, we deal with the interference problem in high density sensor
deployments. We address this issue proposing a pragmatic joint routing, transmission
power control and channel allocation approach, built upon the well-known RPL (Routing
Protocol for Low-Power and Lossy Networks). This reduces the average packet collisions
and the energy consumption of WSNs. Second, we address the low communication reli-
ability and robustness in WSNs with mobile nodes. In particular, we propose a solution
that combines RPL with a position-based routing approach based on Kalman filtering.
This provides the efficiency and reliability of RPL, and also includes mobility support for
non-static nodes. Third, we study the problem of QoS (Quality of Service) provisioning
in WSNs managing heterogeneous traffic. With this in mind, we propose a multi-tree
approach based on the construction of multiple RPL Instances. This constructs multiple
virtual topologies to address the particular requirements of each traffic flow individu-
ally. Finally, we focus on the efficient integration of wireless sensors into Cloud-based
IoT platforms. In particular, we propose a formulation to orchestrate the resource uti-
lization of the whole network, taking advantage of the recent advances in virtualization
and mobile cloud computing. This optimizes the overall consumption, considering the
capabilities and limitations of each node, while satisfying the service requirements and
the individual users’ demands.

iii
 Acronyms

API Application Programming Interface

AODV Ad Hoc On-demand Distance Vector
bps bits per second
BS Base Station
C-RPL Cooperative RPL
CSDP Cloud Service Distribution Problem
CSMA-CA Carrier Sense Multiple Access - Collision Avoidance
CTP Collection Tree Protocol
DAO Destination Advertisement Object
DIO DODAG Information Object
DIS DODAG Information Solicitation
DODAG Destination Oriented Acyclic Graph
DSR Dynamic Source Routing
EO End Office
ETX Expected Transmissions
GBR Gradient-Based Routing
GEAR Geographic and Energy Aware Routing
GPS Global Positioning System
HO Head Office
HR Hierarchical Routing
IEEE Institute of Electrical and Electronics Engineers

v
IETF Internet Engineering Task Force
ILP Integer Linear Programming
IO Intermediate Office
IoT Internet of Things
KP-RPL Kalman Positioning RPL
LEACH Low-Energy Adaptive Clustering Hierarchy
LLN Low-Power and Lossy Network
MAC Medium Access Control
MANET Mobile Ad-hoc Network
MaxPDR Maximum PDR
MinAP Minimum Aggregated Power
MIPS Million of Instructions Per Second
MP2P Multipoint-to-Point
OF Objective Function
pdf probability density function
P2P Point-to-Point
P2MP Point-to-Multipoint
PDR Packet Delivery Ratio
PRR Packet Reception Rate
PL Path Loss
QoS Quality of Service
RPL Routing Protocol for Low-Power and Lossy Networks
ROLL Routing Over Low-Power and Lossy networks
SDN Software-Defined Networking
SDP Service Distribution Problem
SHM Structure Health Monitoring
RSSI Receive Signal Strength Indicator
TSCH Time-Slotted Channel Hopping
VANET Vehicular Ad-hoc Network

vi
vFSN Virtual Fog Service Network
WSN Wireless Sensor Network
WSSN Wireless Sensor Surveillance Network

vii
 Contents

1. Introduction 1
1.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1. Node Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2. Node Mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.3. Traffic Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.4. Integration into the Internet of Things . . . . . . . . . . . . . . . . 6
1.2. Outline and Research Contributions . . . . . . . . . . . . . . . . . . . . . 6

2. Background 15
2.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2. Gradient-Based Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1. RPL: Routing Protocol for Low-Power and Lossy Networks . . . . 18
2.3. Realistic Multi-Channel and Transmission Power Control . . . . . . . . . 21
2.3.1. Multi-Channel Control . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.2. Transmission Power Control . . . . . . . . . . . . . . . . . . . . . . 22
2.4. Position-Based Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.5. Management of Heterogeneous
Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5.1. Data Traffic Classification . . . . . . . . . . . . . . . . . . . . . . . 30
2.6. From Cloud to Fog Computing . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6.1. Overview of Cloud Computing . . . . . . . . . . . . . . . . . . . . 32
2.6.2. Distributed Cloud Computing . . . . . . . . . . . . . . . . . . . . . 34
2.6.3. Fog Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3. Joint Routing, Channel Allocation and Power Control to Support

High-Density Deployments 37
3.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.1. Motivation and Previous Work . . . . . . . . . . . . . . . . . . . . 38
3.2.2. Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.3. Organization of the Chapter . . . . . . . . . . . . . . . . . . . . . . 41

ix
 3.3. Joint Routing, Channel Allocation and Power Control . . . . . . . . . . . 42
3.3.1. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3.2. MinAP: Minimum Aggregated Power . . . . . . . . . . . . . . . . . 45
3.3.3. MaxPDR: Maximum Packet Delivery Ratio . . . . . . . . . . . . . 48
3.3.4. Multi-Channel Control . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4. Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4.1. Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.2. Energy Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.4.3. Reliability vs. Consumption Tradeoff . . . . . . . . . . . . . . . . . 57
3.4.4. Collision Probability . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5. Experimental Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5.1. WSN Testbed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5.2. Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.6. Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4. Position Assisted Routing to Support Node Mobility 67

4.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2.1. Motivation and Previous Work . . . . . . . . . . . . . . . . . . . . 68
4.2.2. Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2.3. Organization of the Chapter . . . . . . . . . . . . . . . . . . . . . . 72
4.3. KP-RPL: Kalman Positioning - RPL . . . . . . . . . . . . . . . . . . . . . 72
4.4. Positioning Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4.1. Positioning based on Confidence Regions . . . . . . . . . . . . . . 74
4.4.2. Enhanced Positioning using Kalman Filtering . . . . . . . . . . . . 78
4.5. Blacklisting Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.6. Routing Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.6.1. Anchor to Anchor Routing . . . . . . . . . . . . . . . . . . . . . . 83
4.6.2. Mobile to Anchor Routing . . . . . . . . . . . . . . . . . . . . . . . 84
4.7. Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.7.1. Robustness against Channel Conditions . . . . . . . . . . . . . . . 88
4.7.2. Number of Concurrent Positioning Nodes . . . . . . . . . . . . . . 90
4.7.3. Density of Anchor Nodes . . . . . . . . . . . . . . . . . . . . . . . 93
4.7.4. Energy Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.7.5. ETX vs PDR Tradeoff . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.8. Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

5. Multi-Tree Routing to Manage Heterogeneous Traffic 101

5.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.2. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.2.1. Motivation and Previous Work . . . . . . . . . . . . . . . . . . . . 102
5.2.2. Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.2.3. Organization of the Chapter . . . . . . . . . . . . . . . . . . . . . . 106
5.3. Preliminary Results on Multi-Tree Routing . . . . . . . . . . . . . . . . . 107

x
 5.4. C-RPL: Cooperative - RPL . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.4.1. Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.4.2. Rank Computation and Parent Selection . . . . . . . . . . . . . . . 114
5.4.3. Coalition Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.4.4. Cooperation Game among RPL Instances . . . . . . . . . . . . . . 117
5.4.5. Example of C-RPL . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.5. Fairness in Networks with Multiple Instances . . . . . . . . . . . . . . . . 124
5.6. Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.6.1. Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . . 127
5.6.2. Cooperation Parameter . . . . . . . . . . . . . . . . . . . . . . . . 130
5.6.3. Impact of Node Density . . . . . . . . . . . . . . . . . . . . . . . . 133
5.6.4. Impact of Traffic Load . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.6.5. Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.7. Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of

Things 141
6.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.2. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.2.1. Motivation and Previous Work . . . . . . . . . . . . . . . . . . . . 142
6.2.2. Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.2.3. Organization of the Chapter . . . . . . . . . . . . . . . . . . . . . . 146
6.3. IoT-Cloud in Smart Cities . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.4. System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.4.1. Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.4.2. Service Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.4.3. Service Requirements . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.5. The IoT-Cloud Service Distribution Problem . . . . . . . . . . . . . . . . 153
6.6. Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.6.1. Smart City Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.6.2. Simulation Details . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
6.6.3. Augmented Reality Service . . . . . . . . . . . . . . . . . . . . . . 165
6.6.4. Autonomous Sensing and Actuation Service . . . . . . . . . . . . . 168
6.6.5. City Monitoring Service . . . . . . . . . . . . . . . . . . . . . . . . 170
6.7. Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

7. Conclusions and Future Work 177

7.1. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.2. Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

xi
 Chapter 1

Introduction

1.1. Motivation

Wireless sensor networks (WSNs) have attracted much attention of researchers during
the last decade. Despite their variety, all sensor networks have certain fundamental fea-
tures in common. These are composed of low cost devices that are networked via low
power wireless communications, to which traditional sensors (e.g., mechanical, thermal,
biological, chemical, optical, magnetic) may be attached to measure physical phenom-
ena. These sensor nodes, which may be deployed randomly all over the sensing area,
collaborate among themselves to reach a common goal. Some of the benefits of wireless
sensing versus wired approaches are: lower installation and maintenance costs, increased
flexibility, broader sets of applications, and larger deployments.

Since the very beginning, this technology promised to revolutionize the way we live, work,
and interact with the physical environment. In the present, wireless sensors are being de-
ployed on roads, in cities, forests, factories, houses, and many other scenarios. This large
scale sensing platform enables the user to have access to a wide range of applications,
particularly in the fields of positioning, distributed detection and monitoring. Some
examples are environmental monitoring [Rao15], industrial automation [Gun13], agri-
culture [San15], natural disaster prediction [Arj15], homeland security [Bar13b], smart

1
 1.1. Motivation

buildings [Sur15], healthcare [Zha14], transportation [Hu15] and structure health moni-
toring [Liu15].

Besides the advantages of wireless sensors, their limited communication, computing,

storage, and energy resources, has motivated the research community to consider WSNs
as a new paradigm in contrast with general data networks. In particular, the communi-
cation constraints (e.g., 250 kbps in IEEE 802.15.4) narrow the feasible solutions to those
that do not need to use the radio resources extensively. The computational constraints
of commercial wireless sensors (e.g., typical microcontroller speed around 8-48 MHz)
do not allow strategies demanding heavy computation. The memory limitation (e.g.,
typically around 128-512 KB) restrict the use of strategies that need significant storage
resources. Finally, their limited battery capacity (e.g., typical battery resources provide
around 27 kJ) limit their duty cycle. On the other hand, the well known wireless chan-
nel dynamics severely affect them, due to their low transmission power (e.g., around 3
dBm), the traffic congestion at their frequency bands, and their low-gain omnidirectional
antennas.

The particularities and constraints of WSNs have inspired a vast literature to exploit
this technology as efficiently as possible. Since wireless sensors may be deployed in in-
accessible locations in which it is not possible to replace or recharge their batteries for
practical reasons, the research community has mainly focused on reducing their energy
consumption [Pan14]. It is well known that the radio transceiver is the most energy de-
manding part of the wireless sensor [Wan06]. Therefore, energy efficient communications
are imperative to increase the network lifetime of WSNs. The most relevant mechanisms
to reduce their energy consumption are:

Sleep scheduling [Guo12]: The wireless sensor radios generally have four operation
modes: transmission, reception, idle and sleep. Idle, transmission and reception
modes consume a similar amount of energy. Therefore, sleep scheduling techniques
put nodes into sleep mode (i.e., radio off) whenever possible, since it consumes
much less energy.

2
1. Introduction

Energy efficient MAC (Medium Access Control) protocols [Hua12]: Their objective
is to avoid collisions so that two interfering nodes do not transmit at the same time,
and therefore avoid wasting energy resources with retransmissions.

Data-driven approaches [Jia15]: Their goal is to reduce the amount of bits trans-
mitted by reducing the data redundancy. These take advantage on the spatial and
temporal correlation of measurements. Some examples are data prediction, data
compression and data aggregation.

At the beginning, WSNs where planned to be dedicated systems, where highly specialized
protocols were used within the WSN, and vendor-specific gateways were used to provide
network connectivity with the IP world. Therefore, gateways and sensors often have to
be from the same vendor in order to be compatible [Ish13]. Recently, the RPL (Routing
Protocol for Low-power and lossy networks) standard has been proposed by the IETF
(Internet Engineering Task Force) ROLL (Routing Over Low power and Lossy networks)
working group to bring IPv6 to WSNs, so that sensor nodes can be natively addressed
and connected through the IP protocol with the rest of the Internet. As a result, a new
breed of WSN services and applications are expected in the near future, since they are
anticipated to be a key element of the future Internet of Things (IoT). In this thesis, we
analyze which are the key issues that need to be addressed to enable these services and
propose solutions to throw some light onto these problems. In particular, we visualize
future WSN deployments as highly dense, mobile, multi-purpose networks, perfectly
integrated into the rest of the IoT. Then, we focus on each of these features throughout
this thesis.

1.1.1. Node Density

The node density of wireless sensor deployments is expected to remarkably increase, par-
ticularly in urban scenarios. Note that larger WSNs, probably from different providers,
are expected to coexist in the same sensing areas [Xia13]. Therefore, it is crucial to
reduce the interferences among sensors, without decreasing the network throughput.

3
 1.1. Motivation

Although some commercial wireless sensors have multi-power and multi-channel capa-
bilities [Atm09], which certainly alleviate the problem of interferences among sensors,
sometimes these cannot be efficiently exploited due to practical issues, such as hardware
or software incompatibilities.

1.1.2. Node Mobility

Thanks to the reduction of size and weight of commercial wireless sensors, these will
be attached to mobile entities more easily. However, communication protocols generally
assume that WSNs are static [Pan14], and hence most of these strategies are not practical
in scenarios with mobile nodes. Note that in this case the network must dynamically
adjust itself to variations on the distances among nodes, sudden obstacles, disabled
nodes, and so on. This introduces additional challenges that need to be considered in
order to enable robust communications in practical scenarios [Pen10].

1.1.3. Traffic Heterogeneity

Although current wireless sensors are generally application-specific nodes developing

simple tasks, the increasing processing capabilities of commercial sensors will enable
more advanced services, which may combine multiple independent tasks. Moreover,
wireless sensors are envisioned as infrastructure resources, which means that they may
run multiple third-party applications on-demand [Mit12]. Note that each application
has its own particularities in terms of data generation (e.g., continuous, intermittent),
traffic pattern (e.g., intermittent, periodic, bursty) and QoS requirements (e.g., lifetime,
reliability, delay) [Wan10]. Therefore, the management of the traffic generated by each
application must be addressed individually.

4
1. Introduction

1.1.4. Integration into the Internet of Things

The advances in mobile computing and software defined networking, enable WSNs to be
connected with the Cloud and also be part of the Internet of Things (IoT). Then, the
extensive processing resources of the Cloud can be combined with the ubiquity of wire-
less sensors. Moreover, the data sensed by different WSNs may be shared on a bigger
scale, enabling a more efficient exploitation of the physical infrastructure [Mis14b]. This
will extend the capabilities of WSNs far beyond sensing and tracking applications, par-
ticularly in smart cities, grids and transportation networks. However, the heterogeneity
of devices and services complicates the efficient orchestration of these platforms.

In this thesis, we address the above mentioned challenges from a pragmatic perspec-
tive, taking into account not only the hardware limitations of wireless sensors, but also
the issues that may arise with their implementation in real-life WSNs. Moreover, the
compatibility of the proposed solutions with the existent networking standards is also
a major requirement. With this in mind, we design them to be compatible with the
well-known RPL protocol, which is an energy efficient protocol that has been already
standardized by the IETF (Internet Engineering Task Force).

1.2. Outline and Research Contributions

In this section, we outline the contents of the thesis and the major contributions. Chapter
1 discloses the main challenges that need to be addressed to enable the forthcoming WSN
applications and their integration in society (See Figure 1.1). Chapter 2 illustrates the
background for the following chapters. Chapter 3 introduces a pragmatic joint routing,
power control and channel allocation approach to enable higher density deployments.
Chapter 4 introduces a robust routing strategy for WSNs with both static and mobile
nodes to deal with the issues that arise in real-life WSNs with mobile nodes. Chapter
5 introduces a multi-tree routing approach to efficiently manage multiple traffic types

5
 1.2. Outline and Research Contributions

Integrated into the IoT

(Chapter 5)

Higher Traffic Heterogeneity

Integration of WSNs
(Chapter 4)

in Socitety
Higher Node Mobility
(Chapter 3)

Higher Node Density

(Chapter 2)
Current
Deployments

Practical Challenges

Figure 1.1. Growth of the integration of wireless sensor networks in society thanks to the development
of practical strategies.

in a single WSN. Chapter 6 introduces a minimum cost flow formulation for IoT-Cloud
platforms to address the integration of WSNs into the future IoT. Finally, Chapter 6
concludes the thesis and prospects our future work. In the following, we discuss the
details of the contributions.

Chapter 3 - Joint Routing, Channel Allocation and Power Control to Support

High Density Deployments

Nowadays, commercial wireless sensors have multi-power and multi-channel capabilities.

These effectively reduce the network consumption and avoid the high collision probability
that may arise in convergecast networks. However, practical issues arise in cross-layer
implementations. In this chapter, we enhance RPL to obtain a joint routing, trans-
mission power control and channel allocation solution for real-life WSNs. Two different
strategies (MinAP and MaxPDR) are designed and implemented in a WSN testbed with
commercial motes. We compare these strategies with the original RPL and also with
other standardized routing protocols using simulations. We also provide experimental
results to show that the network performance is improved both in terms of reliability

6
1. Introduction

and energy consumption.

The contributions of this chapter have been included in a journal article:

Barcelo, M.; Correa, A.; Vicario, J.L.; Morell, A., “Joint routing, channel alloca-
tion and power control for real-life wireless sensor networks”, in Transactions on
Emerging Telecommunication Technologies, January 2014.

This has been inspired by the previous conference publications:

Barcelo, M.; Correa, A.; Vicario, J.L.; Morell, A., “Joint routing and transmission
power control for Collection Tree Protocol in WSN”, IEEE International Sympo-
sium on Personal Indoor and Mobile Radio Communications (PIMRC), September
2013.

Barcelo, M.; Vicario, J.L.; Seco-Granados, G.; Puig, J.M.; Laborda, J.M., “Multi-
channel routing algorithm for cluster-tree wireless sensor networks in aerospace
applications”, IEEE Fly By Wireless (FBW), June 2011.

Chapter 4 - Position Assisted Routing to Support Node

Mobility

Mobility is still one of the greatest challenges in WSNs, since energy efficient routing
strategies are generally designed for static networks. In particular, RPL has been proven
to be very efficient in static WSNs. However, its slow response to topology changes
and its huge signalling cost to keep up-to-date routes in the presence of mobile nodes
makes it inefficient in mobile scenarios. In this chapter, we introduce KP-RPL (Kalman
Positioning - RPL), a novel routing strategy for WSNs with both static and mobile nodes.
The objective of KP-RPL is to provide robust routing, taking into account the issues
that arise in real-life WSNs with mobile nodes (e.g., sudden obstacles, interferences,
estimation errors). This extends RPL with a new metric for mobile nodes that combines
Kalman positioning and blacklisting. The simulation results show that the reliability
and the robustness of the network in bad channel conditions is enhanced, compared to

7
 1.2. Outline and Research Contributions

existent approaches. Moreover, thanks to the Kalman filter, the minimum rate that is
necessary for positioning is reduced and the network lifetime extended.

The contributions of this chapter have been submitted for journal publication in:

Barcelo, M.; Correa, A.; Vicario, J.L.; Morell, A.; Vilajosana, X.,“Position Assisted
RPL using Kalman Filtering in Wireless Sensor Networks with Mobile Nodes”,
(submitted to) IEEE Sensors Journal, 2015.

A preliminary position-aware routing protocol was included in the following conference

publication:

Barcelo, M.; Correa, A.; Vilajosana, X.; Vicario, J.L.; Morell, A., “Novel routing
approach for the TSCH mode of IEEE 802.15.4e in wireless sensor networks with
mobile nodes”, IEEE Vehicular Technology Conference (VTC), September 2014.

Chapter 5 - Multi-Tree Routing to Manage Heterogeneous

Traffic

Advanced WSN applications may need to develop multiple tasks that involve sensing,
processing and gathering data from different sensing units. This heterogeneous data may
have multiple and sometimes opposite sets of requirements. In these scenarios, differ-
ent objective functions must be combined, and therefore traditional single-tree routing
approaches are not efficient. On the contrary, RPL virtually splits the network into
multiple RPL Instances, which transport each kind of data according to its particular
requirements. However, this protocol does not define any mechanism to decide the nodes
that must belong to each instance, and this decision has a strong impact in the network
energy consumption and performance. With this in mind, in this chapter we introduce
C-RPL (Cooperative - RPL), which creates multiple instances following a cooperative
strategy among nodes with different sensing tasks. As a result, the energy consumption,
the complexity and the cost of the nodes is reduced compared to RPL, since they are
active less time, perform fewer tasks and are equipped with less sensing hardware. In

8
1. Introduction

this chapter, we also propose a novel fairness analysis for networks with multiple in-
stances, showing that C-RPL achieves a better tradeoff in terms of performance and
energy consumption than RPL.

The contributions of this chapter have been submitted for journal publication in:

Barcelo, M.; Correa, A.; Vicario, J.L; Morell, A., “Cooperative interaction among
multiple RPL instances in wireless sensor networks”, (submitted to) Computer
Communications, 2015.

Preliminary multi-tree and cooperative multi-tree routing approaches were introduced

in the following conference publications:

Barcelo, M.; Correa, A.; Vicario, J.L; Morell, A., “Cooperative multi-tree sleep
scheduling for surveillance in wireless sensor networks”, IEEE Military Communi-
cations Conference (MILCOM), November 2013.

Barcelo, M.; Correa, A.; Vicario, J.L; Morell, A., “Multi-tree routing for heteroge-
neous data traffic in wireless sensor networks”, IEEE International Conference on
Communications (ICC), June 2013.

Chapter 6 - IoT-Cloud Formulation to Integrate WSNs into the Future In-

ternet of Things

The confluence of the Cloud and the Internet of Things (IoT) enables a new breed of
applications that create real-time valuable information via the analysis of live data from
a wide range of sensing devices. Wireless sensors are planned to be a key element in
this new paradigm, and therefore their efficient integration is essential. In this chapter,
we address this as a minimum cost mixed-cast flow problem. This is mathematically
formulated using only linear constraints and solved via integer linear programming. We
refer to this formulation as the IoT-Cloud SDP (Service Distribution Problem), since it
finds the placement of virtual service functions over an IoT-Cloud platform. We focus on
energy consumption as the major driver of todays network and cloud operational costs.

9
 1.2. Outline and Research Contributions

The heterogeneous set of IoT-Cloud network components are characterized according to

their associated sensing, compute and/or transport capacity and energy efficiency. We
solve the IoT-Cloud SDP for an illustrative set of smart city services, in which users
provide and request information using their smart devices, showing that this is a flexible
solution to efficiently orchestrate IoT-Cloud networks.

The contributions of this chapter are included in a journal article that is being prepared
for publication:

Barcelo, M.; Correa, A.; LLorca, J.; Vicario, J.L; Morell, A.; Tulino, A., “IoT
Cloud Service Optimization in Smart Cities”. (Manuscript in preparation).

The SDP was first formulated for cloud networks in the following conference publica-
tion:

Barcelo, M.; Llorca, J.; Tulino, A.; Narayan, R., “The Cloud Service Distribution
Problem in Distributed Networks”, IEEE International Conference on Communi-
cations (ICC), June 2015.

Additional contributions

Besides the research contributions covered in this thesis, some other topics related to
WSNs have also been considered during this period.

The study of a PCA-based data aggregation technique has been submitted for publication
in the following journal article:

Morell, A.; Correa, A.; Barcelo, M.; Vicario, J.L., “Data Aggregation and Principal
Component Analysis in WSNs” (submitted to) IEEE Wireless Communications,
2015.

The contributions related with the design of positioning systems based on Kalman fil-
tering are included in the following journal and conference publications:

10
1. Introduction

Correa, A.; Barcelo, M.; Morell, A.; Vicario, J.L., “Enhanced inertial-aided indoor
tracking system for wireless sensor networks: A review,” IEEE Sensors Journal,
September 2014.

Correa, A.; Barcelo, M.; Morell, A.; Vicario, J.L., “Distance-based tuning of the
EKF for indoor positioning in WSNs,” European Signal Processing Conference
(EUSIPCO), September 2014.

Correa, A.; Morell, A.; Barcelo, M.; Vicario, J.L., “Navigation system for elderly
care applications based on wireless sensor networks,” European Signal Processing
Conference (EUSIPCO), August 2012.

The results regarding the exploitation of multiple receivers around the body for indoor
positioning systems have been published in the following conference publication:

Correa, A.; Barcelo, M.; Morell, A.; Vicario, J.L., “Indoor pedestrian tracking
system exploiting multiple receivers on the Body” International Conference on
Indoor Positioning and Indoor Navigation (IPIN), October 2014.

Finally, an extension of this article including additional details and experimental results
has been submitted for journal publication in:

Correa, A.; Barcelo, M.; Morell, A.; Vicario, J.L., “Indoor pedestrian tracking
with on-body multiple receivers” (submitted to) IEEE Sensors Journal, 2015.

11
 Chapter 2

Background

2.1. Summary

In this chapter, we provide a background overview of the most relevant topics in the
following chapters. First, we briefly introduce gradient-based routing with a focus on the
RPL protocol, since the routing strategies presented in Chapters 3-5 are built upon this
protocol. Second, we introduce the multi-channel and multi-power approaches, which
are effective solutions to deal with packet collisions in high density WSNs (Chapter 3).
Third, we present position-based routing as an efficient mechanism to deal with the
mobility of nodes (Chapter 4). Fourth, we discuss the practical challenges that appear
in WSNs managing heterogeneous traffic, and also a general classification of the most
relevant kinds of traffic in WSNs (Chapter 5). Finally, we introduce cloud computing and
motivate the transition to fog computing, which is a key concept to efficiently integrate
wireless sensors into the future IoT (Chapter 6).

2.2. Gradient-Based Routing

In this section, we motivate the use of gradient-based routing in general, and the RPL
protocol in particular, in front of other routing approaches.

13
 2.2. Gradient-Based Routing

WSNs are mainly used for monitoring purposes. Therefore, the communication in these
networks is frequently many-to-one (i.e., the base station collects data from many sensor
nodes). In this scenario, tree-based routing is extensively used because it can reduce the
amount of sending data by using data aggregation techniques. Many strategies can be
used to create a routing tree. In general, the most interesting approaches can be classified
in clustering protocols, position-based routing protocols and self-organizing coordinate
protocols [Wat11]. Clustering protocols, such as low-energy adaptive clustering hierarchy
(LEACH) [Hei00], impose a structure to balance the energy consumption of nodes. Each
node determines in each round whether it becomes a cluster head or a leaf node using a
stochastic algorithm. Then, leaf nodes transmit to cluster heads, which aggregate and
compress the data and forward it to the sink. These techniques assume that all nodes
can communicate with a cluster-head, but the accuracy of this assumption depends on
the transmission ranges of the wireless sensors and the particular scenario. As a result,
the performance of the network may be reduced due to the use of unreliable links. On
the other hand, position-based routing, such as geographic and energy aware routing
(GEAR) [Yu01], constructs the routes according to the position of nodes. This reduces
the complexity of the routing process, but the obtention of the geographic information
may not be possible. Moreover, the inevitable positioning errors reduce the network
performance [Pen11]. In contrast, self-organizing coordinate protocols, such as gradient-
based routing (GBR) [Man01], do not need to estimate the position of the nodes. This
avoids using GPS chips, indoor positioning systems, or programming the position of each
node manually. This last category, and more precisely the gradient-based approach, has
been found the most suitable mechanism for convergecast WSNs by the IETF, through
the ROLL working group [Wat11].

Gradient-based routing is based on the transmission of control messages (also referred to

as pilots or beacons) to estimate the link qualities. According to the link quality estima-
tor, the network can be optimized in terms of reliability, energy consumption, latency,
bandwidth, and so on. GBR is the classical gradient routing protocol, but there exist
more advanced gradient-based implementations, such as CTP (Collection Tree Proto-

14
2. Background

col) [Gna09], RPL (Routing Protocol for Low-Power and Lossy Networks) [Win12] and
the hierarchical routing defined in ZigBee [Cuo07]. CTP is a routing protocol specially
designed for relatively low traffic rates. It is widely used in practical WSN implemen-
tations. In fact, CTP has been included in TinyOS 2.x and also in OMNET++. For
this reason, it has been widely used in the recent years for both research [Vil10] and
commercial products. For instance, CitySee and GreenOrbs are using this protocol
for ecological surveillance purposes, and Powernet for energy consumption monitoring.
RPL is a more recent protocol that provides a similar performance in terms of relia-
bility, energy consumption and protocol overhead. In addition to the features of CTP,
this can successfully communicate with IP devices in the greater Internet using IPv6
addresses [Ko11]. Therefore, we have adopted this protocol in the design of our routing
algorithms.

2.2.1. RPL: Routing Protocol for Low-Power and Lossy Networks

RPL is an energy efficient and reliable protocol designed for (LLNs) Low-Power and
Lossy Networks, such as WSNs. This protocol constructs convergecast trees, referred
to as DODAGs (Destination Oriented Acyclic Graphs), with one or multiple roots that
act as sinks, according to a predefined objective function (OF). The OF defines how
RPL nodes select and optimize routes, and should be defined according to the network
requirements. This also defines the concept of RPL Instance that groups one or multiple
DODAGs with a common OF (See Figure 2.1). For the sake of simplicity, in this thesis
we assume that each RPL Instance includes only one DODAG, and each DODAG has a
single sink.

According to the particular OF defined for each DODAG, the nodes compute their
relative distance to the sink, referred to as Rank [Vil14]. This metric is also used to
avoid and detect loops in order to ensure that packets make forward progress within
the DODAG, since it must monotonically decrease towards the DODAG destination.
The typical OF in RPL is based on the minimum number of expected transmissions

15
 2.2. Gradient-Based Routing

Internet (IPv6)

DODAG Root 1 DODAG Root 2 DODAG Root 3

Rank = 0
1 7 12

Rank = 1
2
13 Rank = 2
8 9 Rank = 2
16
6 14
3
10 Rank = 2 Rank = 3
11 17
4 5 15

DODAG 1 DODAG 2 DODAG 3

Instance 1 Instance 2

Figure 2.1. A RPL network with three DODAGs in two instances.

(ETX). Basically, this metric indicates how many times a message must be transmitted
on average to reach its final destination. For instance, the i-th anchor node estimates its
end-to-end ETX, when forwarding to the j-th anchor node, as the ETX to reach the j-th
anchor node (ET Xi,j ), plus the end-to-end ETX of this node (ET Xj ), that is:

ET Xi = ET Xi,j + ET Xj , (2.1)

where the ETX between two nodes is computed as the reciprocal of their packet delivery
ratio (PDR). Then:
1
ET Xi = + ET Xj . (2.2)
P DRi,j

RPL defines three different control packets: DODAG Information Object (DIO), DODAG
Information Solicitation (DIS) and Destination Advertisement Object (DAO):

DIO packets are sent in broadcast and propagate Ranks in order to construct and
maintain the DODAG. A DIO packet includes information that allows a node to
discover a RPL Instance, to learn its configuration parameters, and also to select
a DODAG parent set. The most relevant fields in a DIO packet are the 128-bit

16
2. Background

DODAGID, which uniquely identifies a DODAG using an IPv6 address, the 8-

bit RPLInstanceID, which identifies the instance in the network, and the 16-bit
Rank. The rate of DIOs is adjusted dynamically by a trickle timer according to
the network needs [Lev11].

DIS packets are sent in unicast to solicit DIOs from a neighbour node, for instance
to discover nearby DODAGs. When a node receives a DIS packet from another
node, it must unicast a DIO packet to this node.

DAO packets are sent in unicast upwards along the DODAG to the selected parent
in storing mode, or to the DODAG root in non-storing mode [Gad12]. This is a
mechanism to collect information about the network topology.

Multiple kinds of traffic are supported in RPL: multipoint-to-point (MP2P), point-to-

multipoint (P2MP), and point-to-point (P2P). MP2P is the traffic pattern of applica-
tions collecting information from multiple sources at a single location. This is the most
common traffic in WSNs. P2MP traffic is generated in applications that disseminate in-
formation from a single node to multiple nodes. P2P traffic is also supported to enable
nodes to communicate among themselves.

In terms of security, RPL supports message confidentiality and integrity. In particular,

RPL has three basic security modes: unsecured, preinstalled and authenticated. In the
first mode, RPL control packets are sent without additional security mechanisms. In
the second mode, RPL Instances have preinstalled keys that are applied to generate
secured RPL messages. In the third mode, nodes may join a RPL Instance as a leaf,
but they need to obtain a preinstalled key from an authentication authority to act as a
router.

The interested reader is referred to [Win12] for further details on RPL and illustrative
examples.

17
 2.3. Realistic Multi-Channel and Transmission Power Control

2.3. Realistic Multi-Channel and Transmission Power Con-

trol

In this section, we introduce the multi-channel and multi-power strategies, which define
the frequency channel and the transmission power of wireless sensors, respectively. These
have been used in Chapter 3 to both reduce the energy consumption and alleviate the
packet collision problem in highly dense WSNs.

2.3.1. Multi-Channel Control

Channel allocation has been extensively studied in the literature to increase the band-
width of single channel MAC protocols by parallelizing transmissions among neighboring
nodes [Hua12]. In single channel WSNs under heavy load, many messages are lost due to
collisions among wireless sensors [Jov11]. However, using multiple channels, the poten-
tial interferences are avoided assigning different channels to nodes within a distance of
two-hops (See Figure 2.2). Moreover, network throughput can potentially be increased.
Some examples of multi-channel MAC protocols are MMSN [Zho06], TMCP [Wu08] and
MC-LMAC [Dur08].

In practical implementations, several issues arise with many multi-channel protocols

[Sai14]. Some of them involve heavy computation that cannot be implemented in com-
plexity constrained nodes, such as wireless sensors [Haj11]. In other cases, they require
additional hardware resources, such as multiple transceivers [Li11]. Moreover, it is also
important to take into account the additional overhead that these protocols introduce,
which may be very high. Due to these issues, most of multi-channel protocols cannot be
implemented in practice. It is also important to mention that node-independent channel
allocation may not be feasible in tree-based networks, since nodes would be required to
adapt their reception channel for each son node. Note that channel hops can be a major
source of transmission errors due to channel mismatches and time synchronization errors,

18
2. Background

Ch1

Ch3
Ch2

Ch6

Ch5
Ch4

Ch7

Cluster

Figure 2.2. Example of a channel assignment scheme in WSNs that allocates a single transmission and
reception channel to each node using clusters.

since both nodes should be set on the same channel at the same time [Hua12].

Summarizing, multi-channel protocols for WSNs must not require complex processing,
should not introduce significant overhead, and should not enforce particular hardware
requirements in order to make possible its implementation in commercial motes. More-
over, it is recommended to allocate the channels according to the network topology in
order to benefit from the node distribution and avoid frequent channel changes.

2.3.2. Transmission Power Control

The purpose of transmission power control techniques is mainly twofold: i) limiting mul-
tiuser interference to increase single-hop throughput, and ii) reducing power consumption
to prolong battery lifetime. On one hand, the interferences among nodes should be min-
imized to reduce collisions and increase the throughput. On the other hand, since the
radio transceiver is the main source of energy consumption in WSNs, the transmission
power should be defined according to the current needs. Some examples of power control
strategies proposed in the literature are presented in [Bra09], [Meg11], [Fu12].

19
 2.4. Position-Based Routing

The reliability of a link is closely related to the received signal strength [Ram06]. Accord-
ing to the channel conditions, the transmitter should increase or decrease the transmit
power, either to increase the signal strength at the receiver, or to reduce the energy
consumption and channel interferences with other nodes (See Figure 2.3). It is also
important to consider that additional routes can be discovered using a higher transmit
power, which could be more energy efficient than the routes already discovered. On the
other hand, a reliable link reduces the number of average retransmissions due to un-
received packets, decreasing the channel utilization of that link and therefore potential
collisions. Therefore, in practice the transmit power should not be adjusted without
considering routing information. Moreover, the complexity and the additional overhead
of the transmission power assignment should also be considered. Finally, the particular
range and granularity of power levels has also a relevant impact on the efficiency of these
strategies [Cot14], and this may vary in each commercial platform.

Therefore, a pragmatic transmission power control scheme for WSNs should take into
account the particular hardware characteristics. Moreover, it should be decentralized
and simple to reduce the communication overhead and spend the minimum computing
resources.

2.4. Position-Based Routing

In this section, we introduce position-based routing that is an efficient solution for mobile
WSNs, highlighting the impact of realistic conditions in its performance. In Chapter 4,
we have combined this technique with gradient-based routing to provide robust and
reliable routing in WSNs with mobile nodes.

Several WSN applications require the combination of static nodes, also referred to as
anchors, and nodes attached to mobile entities (e.g., people, equipment, goods). Some
examples are elderly health monitoring, animal tracking, search and rescue and vehic-
ular networks. However, providing robust and reliable routing in this kind of networks

20
2. Background

p3
p2
pN
p1

Figure 2.3. Transmission range of a wireless sensor using different transmit power levels (assuming
omnidirectional antennas).

is still one of the greatest challenges in WSNs [Cad13]. Note that additional issues
must be considered, since mobility may lead to frequent node disconnections, and a high
signalling cost is required to create and repair routes. Moreover, this information may
be frequently out of date at the moment of taking the routing decisions, leading to in-
efficient routing decisions [Gad14]. Routing strategies based on the position of nodes,
referred to as position-based routing or geographical routing, alleviate these problems by
eliminating the need for topology maintenance [Kar00]. Since routing decisions are not
based on routing tables, they do not need to exchange and maintain routing informa-
tion, reducing the memory requirements of nodes, and also the reaction time to sudden
topology changes [Sar14].

The main assumption that relies on this protocols is that sensors are aware of their
position, the position of their neighbors, and the position of the destination. The position
of anchor nodes can be assumed to be error free (i.e., sophisticated positioning systems
can be used, since their position has to be set only once). However, the position of mobile
nodes is unknown, and therefore it has to be frequently estimated. Since GPS (Global

21
 2.4. Position-Based Routing

Greedy area

Transmission range

Relay

Destination Source

Figure 2.4. Greedy packet forwarding.

Positioning System) positioning may not feasible in WSNs, due to the absence of a GPS
module or to GPS reception problems, indoor positioning systems are frequently used.
Then, sensors broadcast their position to the rest of nodes.

The classical geographic routing algorithm uses a greedy strategy that forwards packets
to the neighbour that is closer to the destination (See Figure 2.4). Then, in a multi-
hop scheme, the packets advance towards the destination in each hop. This avoids the
creation of loops by strictly forwarding packets to nodes located closer to the destination
than themselves. However, since forwarding decisions are locally optimal, these may not
lead to globally optimal paths. On the other hand, nodes may not find any candidate
neighbor, resulting in a dead-end node. Then, recovery routines must be applied to avoid
these situations. Face routing is an alternative to greedy routing in order to guarantee
delivery [Kim05]. This uses the concept of planarization, which constructs planar sub-
graphs that contain no intersecting edges. However, it has a low energy efficiency and
it is not efficient in networks that are not located in planar surfaces. This motivates
hybrid greedy-face routing [Kuh03], which generally use greedy forwarding and switch
to face routing when nodes cannot find any candidate neighbour (i.e., any node that is
closer to the sink than itself).

Most of position-based routing strategies, assume accurate location information, but

22
2. Background

positioning systems have a limited accuracy that depends on the technology and the
environment. On the other hand, the so-called unit disk graph model is generally used,
assuming that nodes may reach any node located at a distance equal or lower than its
transmission range. However, in real implementations the transmission ranges are not
uniform, the antennas are not omnidirectional, the links are not bidirectional, and the
reliabilities of links are not boolean. Both positioning [Pen11] and transmission model
[Zam08] inaccuracies should be considered by the routing algorithm, since these cannot
be completely avoided. Note that these have a strong impact in the network reliability,
particularly in the presence of mobile nodes.

2.5. Management of Heterogeneous

Traffic

In this section, we illustrate the problem of managing multiple kinds of traffic in a

WSN. Moreover, a general classification of the main kinds of traffic in WSN applications
is provided. We have addressed the problem of considering the QoS requirements of
heterogeneous traffic in Chapter 5.

With the increasing computational capacities of wireless sensors, a higher number of

WSN applications will combine multiple tasks, and different kinds of measurements,
providing more advanced and complex services. This generates multiple traffic flows
coexisting at the same network, each of them with its particular QoS requirements (See
Figure 2.5). Moreover, these requirements may even change over time in order to satisfy
the users’ demands. For instance, SHM (Structure Health Monitoring) [Har10] systems
may need to collect information coming from different sensing units, such as pressure,
vibration or temperature. Moreover, they also need to send alarm messages in case of
broken sections or systems failures. In addition, continuous messages are broadcast for
external monitoring and calibration. This heterogeneity should be taken into account
to meet the service requirements, as well as for the proper utilization of the limited

23
 2.5. Management of Heterogeneous
Traffic

Flow 1
Flow 2

Figure 2.5. Example of a network with two convergecast trees.

resources of wireless sensors.

Traditional routing approaches are generally focused on a particular QoS metric, such as
delay, reliability, bandwidth, security, energy consumption, and so on. In order to con-
sider multiple QoS metrics, multi-objective routing [Alw13] combines them in a single
metric. However, due to the contradictory relationship among the QoS requirements, it
is not always possible to find a solution that satisfies all the parameters [Mie03]. There-
fore, these strategies look for a balanced solution among the different objective functions.
However, this may not be feasible in applications imposing hard-QoS constraints, such
as a minimum delay or a maximum end-to-end reliability. In this case, data packets
must be categorized into different levels or classes and forwarded accordingly [Ari13]. In
convergecast networks, multi-tree routing schemes are an efficient mechanism to provide
this differentiation at the network layer [Bar13a]. Multi-tree routing algorithms may
provide congestion control, load balancing and QoS differentiation, creating an indepen-
dent tree for each traffic flow according to its particular requirements. Then, each node
has a different parent assigned for each tree in the network.

The RPL protocol considers the multi-tree solution to deal manage heterogeneous traffic.

24
2. Background

In particular, this protocol creates multiple logically independent RPL Instances that
may run concurrently. Each instance may be optimized with a particular objective
function, which should be correlated with the traffic requirements. The identifier of
the instance is defined in the RPLInstanceID field of the data and control packets.
Nodes may belong to multiple RPL Instances, and they could act as a router in some
instances, and as a leaf node in others, but they can only join to one DODAG per
instance. However, it is worth noting that RPL does not specify how to create and
operate with them, and this is set for future specifications [Win12].

2.5.1. Data Traffic Classification

Although a low energy consumption, a high reliability and a reduced latency is always
desired, each type of data has its particular requirements. Moreover, the transmission
pattern may also be different (e.g., bursty transmissions, periodic transmissions, or in
some cases without transmissions during days or months). Nevertheless, the main kinds
of data can be grouped in event detection, non-critical monitoring and critical monitoring
[Wan10].

2.5.1.1. Event Detection Traffic

In event detection tasks, it is crucial to receive the information as soon as possible in

order to turn on the actuators or send an alarm. Therefore, a reduced delay is the main
requirement of this kind of traffic. Since nodes do not need to send information if the
event has not been detected, this traffic is generally isolated. On the other hand, the
transmission consumption is not critical since these occur just occasionally. Furthermore,
note that the reliability is neither as important as delay since many packets can be sent
in the case of alarm. Then, it is enough to receive one of these packets, since all of them
contain the same information. Although the latency of the network may depend on many
factors, such as the sleeping strategy or the transmission scheduling, we assume that the

25
 2.5. Management of Heterogeneous
Traffic

latency in multi-hop networks is proportional to the number of hops, since the packet
must be decoded, processed and coded again in each hop. However, in a particular
implementation the delay metric should be adapted to consider the particularities of the
MAC layer.

2.5.1.2. Non-Critical Monitoring Traffic

Certain tasks, such as ambient conditions monitoring, require periodic transmissions

from every node to the sink. In this case, the delay and the reliability are not critical.
Instead, the energy consumption becomes the most critical parameter, since they require
periodic data transmissions. For instance, temperature, light or pressure measurements,
are highly correlated in time and space. The measurements lost during the transmission
can be easily estimated using the past measurements (i.e., time correlation) or even
with the measurements coming from other motes (i.e., space correlation). Therefore, the
routing strategy should prioritize the energy consumption in this case.

2.5.1.3. Critical Monitoring Traffic

Some applications require a high reliability. This is the case of applications managing
critical data, such as positioning systems or vital signs monitoring. In these applications,
it is essential to assure that as many packets as possible reach the sink. Moreover, their
traffic pattern strongly depends on the application requirements. For instance, in a
positioning application, the mobile node should continuously transmit information as
long as the mobile entity is moving. It is easy to notice that the performance of these
applications is strongly related to the network reliability, since each packet contains
relevant information. Therefore, the routing algorithm should provide the maximum
packet reliability.

26
2. Background

2.6. From Cloud to Fog Computing

In this section, we introduce cloud computing and also motivate the transition from
cloud computing to fog computing in order to move processing closer to the end-users.
In Chapter 6, we address the integration of WSNs into Cloud-based IoT platforms taking
into account the advantages of fog computing.

2.6.1. Overview of Cloud Computing

Cloud computing is a cutting edge technology that makes use of a shared pool of con-
figurable computing resources that can be rapidly provisioned and released to store and
process data in third-party servers. Then, personal devices become interfaces to power-
ful data centers. This model reduces the necessity of people and companies of buying
dedicated hardware using a pay-as-you-go model. The main advantages of cloud comput-
ing are infrastructure cost reduction, greater flexibility, elasticity and optimal resource
utilization. Then, major companies, such as Amazon, Google, IBM, Cisco and Alcatel-
Lucent have invested in cloud computing and offer cloud-based solutions for individuals
and businesses.

Cloud computing offers many different services, which form the cloud computing stack
(See Figure 2.6). They describe how cloud services are made available to clients. These
can be categorized in [Asl12]:

IaaS: Infrastructure as a Service. This is the lowest level of the stack. It provides
infrastructure components to clients, such as virtual machines, servers, networks
and firewalls. Amazon Web Services is a large IaaS provider.

PaaS: Platform as a Service. This layer provides application programming inter-

faces (APIs) to clients to interact with databases and web servers, which automat-
ically scale according to the users’ requirements. Google AppEngine is a popular
PaaS provider.

27
 2.6. From Cloud to Fog Computing

Cloud Clients

End Users

Level of Abstraction
SaaS

Level of Control
Application
Developers PaaS

Network
IaaS
Architects

Figure 2.6. Cloud computing stack.

SaaS: Software as a Service. This provides complete online software solutions

already developed, such as virtual desktops or social media platforms. SaaS is
currently a huge market and still has a strong growth potential. According to
Forrester’s annual industry outlook, global SaaS software revenues are forecasted
to reach $106 billions in 2016.

In practice, most of current cloud services are a combination of these services, and
therefore the line between them is becoming blurred.

2.6.2. Distributed Cloud Computing

Already, connected devices have reached 9 billion, and this number is expected to grow
more rapidly and reach 24 billion by 2020 [Gub13]. Internet traffic will soon be domi-
nated by the consumption of resource and interaction intensive cloud services and ap-
plications from resource-limited communication end points. Moreover, the number of
latency-sensitive applications, such as real-time services, is also growing. Both end user
experience and overall network efficiency will push for a fundamental transformation of
todays centralized network and cloud architectures towards a highly distributed con-

28
2. Background

verged cloud-network platform, composed of a large number of cloud nodes distributed

across an increasingly meshed converged metro-core [Bar15]. In this massively dis-
tributed cloud network, and with the introduction of network functions virtualization
(NFV) and software-defined networking (SDN), virtual cloud service functions can be
dynamically and elastically instantiated over commodity servers at multiple cloud loca-
tions close to the end users and interconnected via a programmable network fabric. This
way, a cloud network operator can host a variety of services and applications over a com-
mon distributed physical infrastructure, reducing both capital and operational expenses,
while delivering high quality of experience (QoE).

2.6.3. Fog Computing

With this in mind, fog computing, also referred to as edge computing, extends the cloud
computing paradigm to the edge of the network. In general, a fog network is a highly
virtualized platform, composed of a huge number of heterogeneous devices, that provides
computation, storage, and networking services between end devices and traditional cloud
computing data centers. Fog nodes may be resource-poor devices, such as set-top-boxes,
access points, routers, base stations, smart devices, or low complexity cloud nodes, such
as cloudlets [Sat09] or micro-clouds [Shi13], placed at the edge of the network. The
integration of end-devices into the Cloud, motivates the definition of additional service
models besides IaaS, PaaS and SaaS, such as SAaaS (Sensing and Actuation as a Service),
which includes sensors and actuators.

In addition to the benefits of cloud computing, fog computing provides support for
latency-sensitive applications, since it allows applications to run as close as possible to
the end-users, mobility support and location awareness [Sto14]. Moreover, this reduces
the utilization of the cloud resources when these are not strictly necessary, partially
storing and processing at edge devices. Some scenarios in which fog computing has a huge
potential are smart grids, smart buildings, smart cities, connected vehicles, augmented
reality and mobile big data analytics [Bon12], [Yi15].

29
 2.6. From Cloud to Fog Computing

Although fog computing has several advantages in front of cloud computing, the design
and management of fog networks is still a challenging task. In [Yi15], the following issues
are identified:

Network connectivity: Emerging techniques, such as SDN and NFV, are required
to provide flexible services that integrate heterogeneous devices, such as smart
devices and wireless sensors.

QoS requirements: The individual capacities of fog nodes and the particular service
requirements must be considered.

Computation offloading: Offloading decisions must react to the network changes

as quickly as possible.

Provisioning and resource management: Efficient resource management models are

required to avoid wasting storage, computing and energy resources.

Other issues, such as accounting, billing, interfacing model and security. These are
out of the scope of this thesis.

The identified issues mainly arise due to the heterogeneity of fog nodes and services, and
the dynamism of fog networks in terms of nodes, network access and resources.

30
 Chapter 3

Joint Routing, Channel Allocation and

Power Control to Support High-Density
Deployments

3.1. Summary

The multi-channel and multi-power capabilities of commercial sensors are planned to

be very important in future WSN implementations. This is due to the increasing sen-
sor density in some scenarios, such as smart cities. Note that the interferences among
sensors reduce the network performance, due to packet losses, and increase its energy
consumption, due to packet retransmissions. However, incompatibilities may arise when
multi-channel and multi-power strategies are designed regardless of the algorithms ap-
plied to the rest of layers. However, the complexity of cross-layer approaches may be
too high to be applied in WSNs.

In this chapter, we propose a pragmatic joint routing, transmission power control and
channel allocation scheme to address this problem. This is integrated in the well-known
RPL (Routing Protocol for Low-Power and Lossy Networks) to make it compatible with
this standardized current networking standard. We propose two different approaches:
MinAP (Minimum Aggregated Power) and MaxPDR (Maximum Packet Delivery Ra-
tio). The first approach is designed for applications that demand a very low energy

31
 3.2. Introduction

consumption, while the second approach is for applications with high reliability require-
ments.

3.2. Introduction

3.2.1. Motivation and Previous Work

WSNs are composed of a large number of distributed sensor nodes. These share the
same communication channel, and thus interferences and collisions are prone to happen.
Since the number of WSN applications is constantly growing, this problem is expected
to be very significant in the near future. On one hand, wireless sensors are generally
designed to be application specific (i.e., deployed to perform a particular task), and
therefore multiple WSNs may need to coexist to provide advanced services [Raw14]. On
the other hand, network operators provide their own sensing platform [Mis14b]. As a
result, an increasingly number of wireless sensors will coexist, particularly in densely
populated scenarios, such as urban areas.

Packet collisions reduce both the performance and the energy efficiency of the network.
In order to deal with this problem, many energy efficient communication strategies have
been proposed. These can be implemented in the physical layer (e.g., reduce the trans-
mission power), the medium access control (MAC) layer (e.g., reduce the collision prob-
ability) or the network layer (e.g., find energy efficient routes) [Sha13].

Concerning the network layer, tree-based routing is extensively used in convergecast

networks, (i.e., networks where all traffic is sent to a single node). One of its advantages
is the possibility of setting up the network in a completely distributed manner, reducing
the traffic overhead compared to centralized approaches. This also allows the reduction
of the amount of sending data by using data aggregation and compression techniques.
One of the most interesting solutions for the construction of this kind of networks is
the gradient-based routing approach [Wat11]. This technique uses control messages

32
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

to evaluate the quality of the wireless links. The link quality estimator metric may
focus on different criteria, such as reliability, number of hops, node battery level, energy
consumption, and so on.

On the other hand, concerning the physical and MAC layers, transmission power control
and channel allocation strategies can be very effective to increase the network lifetime
and reduce the collision probability of the network [Dur12]. Power control strategies
adjust the transmission power of each node at the minimum level that guarantees a
tolerable reliability at the receiver [Bra09], [Meg11], [Fu12], [Cor08]. On the other hand,
multi-channel approaches distribute the frequency resources available [Zha12], [Mor13],
[Bac10], [Hua12]. As a result, multiple nodes can transmit in parallel, increasing the
throughput, and alleviating collisions and interferences.

The combination of individual solutions at each layer may not be straightforward and
could lead to inconsistencies [Bar13c]. Therefore, some cross-layer strategies have been
proposed in the literature [ElB11], [Li11], [Luo10]. However, most of these solutions
are too complex to be applied in practical WSNs, unless drastic simplifications are
considered. Note that the wireless dynamics and the limited hardware capabilities of
wireless sensors must be considered in real-life implementations.

In this chapter, we propose a pragmatic cross-layer approach for real-life WSNs. The
objective is to design an implementable joint routing, transmission power control and
channel allocation solution. Instead of designing an unfeasible cross-layer approach, we
enhance the standardized routing protocol RPL (Routing Protocol for Low-Power and
Lossy Networks) [Win12]. This protocol has been selected because it considers many
important issues that appear in real-life WSNs and it can successfully communicate with
devices outside the WSN [Ko11]. More specifically, we propose two different approaches
to enhance RPL, MinAP (Minimum Aggregated Power) and MaxPDR (Maximum Prob-
ability Delivery Ratio). They modify the link estimator, the path selection and the
packet forwarder of RPL, and can be introduced in this protocol without any additional
signalling cost.

33
 3.2. Introduction

3.2.2. Contributions

The main contributions of this chapter are:

We propose a pragmatic cross-layer strategy combining routing, transmission power

control and channel allocation for real life WSNs. Although these problems are
extensively studied separately, from the best of our knowledge a joint solution for
real-life WSNs has not been proposed before.

We design a metric for applications that prioritize the energy consumption (Mi-
nAP), and another for applications with high reliability requirements (MaxPDR).
Moreover, the second approach adjusts the tradeoff between the network energy
consumption and reliability using a configurable parameter.

We implement MinAP and MaxPDR in a commercial WSN platform in order

to provide experimental results of their performance in a real environment, and
compare them with existent solutions.

3.2.3. Organization of the Chapter

The remaining of this chapter is organized as follows: Section 3.3, presents an overview of
the proposed joint routing, channel allocation and transmission power control solution.
This also introduces the MinAP and MaxPDR metrics and the multi-channel mechanism.
Section 3.4, evaluates the performance of the proposed solution using simulations. This
compares MinAP and MaxPDR with LEACH, GBR and the hierarchical routing of
ZigBee in terms of reliability, energy consumption and collision probability for different
scenarios. Section 3.5, provides experimental results of the proposed approach in a
commercial WSN platform composed of 13 IRIS motes. Finally, Section 3.6 summarizes
this chapter and presents the main conclusions.

34
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

3.3. Joint Routing, Channel Allocation and Power Con-

trol

The problems of routing, channel allocation and transmission power control are mostly
studied separately. This is because the optimal cross-layer solution becomes a complex
optimization problem [ElB11]. Since WSNs can only deal with low complexity tasks, a
different approach is necessary. We design a pragmatic cross-layer solution built upon
RPL. This avoids both the incompatibility issues that may arise with the combination of
several individual techniques, and the difficulties that may appear in the implementation
of unpractical cross-layer approaches in commercial WSNs.

3.3.1. Overview

The proposed RPL extension modifies the link quality estimation, the path selection
and the packet forwarder of this protocol. More precisely, the link estimator considers
a different metric. Instead of using an ETX-based metric, two alternative metrics are
defined, referred to as MinAP and MaxPDR. The first one focuses on energy consumption
and the second one on network reliability. The path selection has also been modified
according to the estimates of the two proposed metrics. Finally, the packet forwarder is
also modified since packets can be transmitted using multiple transmission power levels
and multiple frequency channels. Therefore, the energy consumption and the packet
collisions can be reduced. The overall differences between the original RPL and the
proposed RPL-based approach are illustrated in Figure 3.1.

Following the structure of RPL, the proposed approach uses DIO packets to compute
the Rank of each node. Each node composes a list with potential parents and their
corresponding Ranks. According to these Ranks, routes are formed by the most preferred
parent of each node. As in RPL, the rate of these packets is dynamically adjusted by
the Trickle algorithm [Lev11]. This algorithm reduces the DIOs rate if the topology is

35
 3.3. Joint Routing, Channel Allocation and Power Control

RPL Proposed Approach

Link Estimation Link Estimation

- ETX - MinAP / MaxPDR

Path Selection Path Selection

- Minimum Energy Consumption /
- Minimum ETX
Maximum Reliability

Packet Forwarding Packet Forwarding

- Fixed Power - Adaptable Transmission Power Control
- Single Channel - Multiple Channels

Figure 3.1. Comparison between the original RPL and the proposed approach.

stable, and increases this rate whenever a change in the topology needs to be propagated.
However, instead of broadcasting DIOs at a unique transmission power level, they are
transmitted at multiple power levels (LDIO ), being LDIO = 0 and LDIO =Lmax the
minimum and maximum available levels, respectively.

The tree construction process also establishes the transmission power and the trans-
mission channel of each node. These two parameters are used to send unicast packets,
such as data messages towards the sink, control packets going downward, ACKs (Ac-
knowledgements), and so on. However, DIOs are transmitted in broadcast at multiple
power levels, from 0 to Lmax . On the other hand, these packets are broadcast using a
previously defined common channel in order to assure that all nodes are able to listen
these packets. Therefore, nodes listen to this channel periodically in order to maintain
their Rank and avoid potential loops in the DODAG. In order to schedule the transmis-
sion and reception of data and control packets (DIOs and DISs), nodes divide the total
time in data and control periods. The time dedicated to each period must be defined
according to the particular network requirements (i.e., a longer data period increases
the network throughput, and a longer control period increases the adaptability of the
network to sudden changes due to interferences, obstacles, and so on). In particular, a
control period duration of 5% of the superframe suffices in our implementation, since
nodes are static.

36
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

New nodes can join the DODAG at any time. To do so, the new node broadcasts a DIS
packet using the common channel. Then, nodes receiving this packet will reset their
Trickle timer, hence start transmitting DIOs with the lowest interval. Finally, the new
node uses these DIO packets to compute the Rank associated with the rest of nodes
based on the Objective Function. Note that nodes are continuously broadcasting DIOs
at multiple transmission powers. Therefore, new nodes will periodically receive DIOs at
the highest power level from each node located at one hop distance.

Since nodes compose a list of potential parents, the process of restoring the tree con-
nectivity due to new obstacles, interferences, node failures and so on, is straightforward.
Once a node detects that it has lost the connectivity with its parent, because it does not
receive its ACKs, the list of potential parents is updated and the new parent is selected
accordingly. Finally, the transmission power level and the transmission channel are also
updated.

3.3.2. MinAP: Minimum Aggregated Power

This strategy is mainly focused on reducing the transmission power as much as possible
to minimize the energy consumption and reduce the interferences among nodes. Then,
DODAGs are constructed using the minimum end-to-end aggregated transmission power.
The block diagram of this strategy can be found in Figure 3.2.

In this strategy, each node selects its own transmission power, being B the transmitter
and A the receiver (LtxA,B ), according to:

LtxA,B = bLmax (1 − P DRA,B )c, (3.1)

where P DRA,B is the link PDR between nodes A and B, that is defined as:

RA,B
P DRA,B = , (3.2)
PB

37
3.3. Joint Routing, Channel Allocation and Power Control

DIO message
received at node
A from node B

PDRA,B = RA,B / P
LtxA,B = ⎣Lmax(1-PDRA,B)⎦
APA,B = W[Ltx] + AP*B

No
APA,B < AP*A

Yes

ParentA = B
AP*A = APA,B

END

Figure 3.2. Block diagram of MinAP.

38
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

where RA,B is the number of DIO messages correctly received at node A from node B,
and PB is the total number of DIO messages sent by B.

The aim of this power selection mechanism is to select the minimum transmission power
to ensure reliable reception of packets. According to (3.1), the transmission power of
node A to node B decreases with the number of DIOs correctly received from B. For
instance, if node A has received the 50% of DIOs sent by node B, it selects its transmission
power level as Lmax /2 (or the level below), assuming that it has received all DIOs sent
with a higher transmission power level. Note this assumes the high correlation among
the PDR of a link with the RSSI measured at the receiver shown in [Ram06]. This
particular mechanism has been proposed due to its simplicity and compatibility with
the RPL protocol.

According to the radio transceiver specifications, each transmission power level (LtxA,B )
has an associated value in Watts (WA,B ). Since these equivalences are known, the total
aggregated power (AP ) from node A to the sink, using node B as the first relay (APAB ),
can be computed as:

X
APAB = Wi , (3.3)
i∈pA,B
pA,B = {l1 , l2 , ..., lS } , (3.4)

where pA,B is the path from A to the sink, using B as the first relay, and l1 , l2 , ..., lS , are
the set of links included in this path, being l1 the link from A to B and lS the last one to
reach the sink. However, the knowledge of all this information in A would require a lot
of signalling. Alternatively, this is computed in a decentralized manner as follows:

APAB = Wl1 + APB∗ , (3.5)

where APB∗ is the aggregated power of the best path from B to the sink, and it is
broadcast by B in its control messages. Note that this is important to reduce the
communication overhead.

39
 3.3. Joint Routing, Channel Allocation and Power Control

DIO message
received at node
A from node B

PDRA,B[LDIO] = RA,B[LDIO] / PB[LDIO]

PDRAB[LDIO] = PDRA,B[LDIO]·PDR*B

No Yes
LDIO = Lmax

No No
ParentA = B PDRAB[LDIO] > PDR*A

Yes
Yes
No
PDRAB[LDIO] ≥ β·PDR*A
ParentA = n
Yes PDR*A = PDRAB[LDIO]
Ltx = LDIO
Ltx = LDIO

END

Figure 3.3. Block diagram of MaxPDR.

Finally, node A defines its parent as the node that provides the lowest aggregated power
value (APA∗ ).

3.3.3. MaxPDR: Maximum Packet Delivery Ratio

This strategy is mainly focused on maximizing the network reliability. In this case, nodes
select the path with the highest end-to-end PDR. Then, the transmission power of each
node is reduced as much as possible according to a configurable parameter β. The block
diagram of this strategy can be found in Figure 3.3.

40
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

In multi-hop networks, the path from each node to the sink must be considered in order
to estimate the reliability of the network. In this strategy, only the DIO messages sent
using LDIO = Lmax are considered to define the routes. This has been considered due to
the highest reliability of RSSI measurements at higher transmission powers. The rest of
DIOs are used to reduce the transmission power of wireless sensors as much as possible.
To do so, a different PDR value is calculated for each transmission power level.

B ) is:
The end-to-end PDR from A using B as the first relay (P DRA

Y
B
P DRA = P DRi , (3.6)
i∈pA,B
pA,B = {l1 , l2 , ..., lS } , (3.7)

For each transmission power level, this can also be calculated in a decentralized manner
as follows:
B ∗
P DRA [LDIO ] = P DRl1 [LDIO ] · P DRB , (3.8)

∗ is the PDR associated with the best path from B to the sink, and it is
where P DRB
∗ ) is set as
broadcast by B. Finally, the node with the highest end-to-end PDR (P DRA
the parent of A.

Finally, if node B has been selected by A as its parent, node A defines its transmission
power as the lowest one that fulfills the following condition:

B ∗
P DRA [LDIO ] ≥ β · P DRA , (3.9)

where β is a threshold parameter that can be adapted, from 0 to 1, according to the re-
liability requirements. This adjusts the willingness of nodes to reduce their transmission
power. Note that high β values are more demanding in terms of reliability, whereas low
values tend to relax it.

41
 3.3. Joint Routing, Channel Allocation and Power Control

3.3.4. Multi-Channel Control

The objective of multi-channel strategies is to increase the spectral efficiency of the

network in order to reduce the high collision probability that may arise in single-channel
tree networks.

In this section, we present a channel allocation process that uses the signalling of the rout-
ing process. Note that the scope of the chapter is not to define a new multi-channel MAC
protocol. Instead, the objective is to include a channel assignment mechanism in RPL
without increasing its signalling cost. Moreover, potential issues that may arise with the
combination of multiple algorithms in the MAC and network layers are avoided.

The proposed method distributes the frequency resources using the cluster structures
automatically created in the routing process, which are composed by a single parent
node and all its child nodes. In order to avoid time-frequency synchronization errors, a
single frequency channel is associated to each cluster (See Figure 2.2). In particular, the
transmission channel (TxCh) of each node is selected according to the reception channel
(RxCh) of its parent node:

T xCh = RxCh [P arent] . (3.10)

From the DIO packets sent by neighbour nodes, each node knows which reception and
transmission channels have been selected by those nodes. This is valuable information,
since these are the channels that are used by nodes within a distance of two hops, and
they must be avoided to reduce network collisions. With this information all the nodes
in the network define their own Channel Utilization vector(ChU t) as:

ChU t = {u1 , ..., uk , ..., uF }, (3.11)

where each value corresponds to the number of neighbour nodes that use each of the
F parallel channels. Using this vector, RxCh is selected as the channel with the lowest

42
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

number of nodes associated:

RxCh = arg min(ChU t). (3.12)

k

At the end of the training process, each node has two channels assigned (TxCh and
RxCh). In order to reduce time-frequency synchronization errors, non-leaf nodes remain
in RxCh unless they need to transmit their own information or to relay information
coming from other nodes.

Figure 3.4 summarizes the routing, transmission power control and channel allocation
processes.

3.4. Simulation Results

This section evaluates the performance of the proposed methods and compares them
with the original RPL [Win12] in terms of energy consumption, reliability and collision
probability. For the sake of comparison, we also include LEACH [Hei00], GBR [Man01]
and the hierarchical routing used in ZigBee [Cuo07] (for simplicity referred here to as
ZigBee) (See Section 2.2). Simulations are conducted in Matlab.

The scenario consists of N sensing nodes transmitting temperature measures through the
DODAG to a single sink node. They are randomly deployed on a LxL m2 surface, and the
sink is situated in the middle of the square. Data aggregation has not been considered,
hence each packet is transmitted independently at each hop. Nodes operate at the 16
different channels defined by the IEEE 802.15.4-2003 standard [IEE11] in the 2.4 GHz
frequency band. They use channel 11 to broadcast and receive DIOs and the rest of
channels (12 to 26) to send data and other unicast messages. A total of 200 realizations
are averaged. In each realization, nodes send their temperature measurements to the
sink through the network.

First, a collision-free multiple access scheme has been considered to evaluate the re-

43
 3.4. Simulation Results

ROUTING AND TRANSMISSION POWER CONTROL

Listen DIO packets

Store channels used by other nodes

MaxPDR MinAP

Link Quality Estimation Link Quality Estimation

Calculate PDR values Calculate AP values

Set Parent Node Set Parent Node

Highest PDR Lowest AP

Set Tx Power Set Tx Power

Reduce Tx Power Select Tx Power
according to β according to Parent

Channels
Parent used by
Node neighbour
nodes

CHANNEL ALLOCATION

Set TxCh
TxCh = RxCh of Parent Node

Create ChUt vector

ChUt = {u1, …, uk, …, uF}

Select RxCh
RxCh = argmink (ChUt)

Figure 3.4. Overview of the routing, transmission power control and channel allocation processes.

44
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

liability and the energy consumption of each strategy. After that, we have used a
CSMA-CA scheme to show the collision probability reduction due to the multi-channel
approach.

The path loses (P L) are modeled using the one slope log-distance path loss model [MS05],
which is commonly employed in current simulation frameworks (e.g., ns-2):

P L(dB) = P L0 (dB) − 10αlog(d) + γ, (3.13)

where P L0 is the path loss at the reference distance d0 (1 m), α is the path loss exponent,
d is the communication distance, and γ is zero mean Gaussian noise with variance σ 2
that models the attenuation caused by flat fading. In these simulations, the following
physical parameters are assumed, σ 2 =6 [Rap01], α=3 [Pu12] and P L0 =50 dB (empir-
ically measured using IRIS motes). Moreover, the sensitivity value (i.e., -91 dBm) and
the multiple transmission power levels of the RF230 transceiver [Atm09] (i.e., 16 levels
from -17.2 dBm to 3 dBm) are considered.

3.4.1. Reliability

Figure 3.5 compares the reliability in terms of end-to-end PDR, considering 30 nodes
and different area deployment dimensions. In these simulations, the threshold parameter
of MaxPDR (β) is set to 1. The impact of this parameter is evaluated in Section
3.4.3. The results show that MaxPDR, with and without transmission power control,
obtains the highest reliability, even higher than the original RPL and ZigBee. Note
that these construct the routes using end-to-end ETX estimates, while MaxPDR uses
PDR estimates. For instance, in a deployment area of 80x80 m2 , MaxPDR achieves
a PDR of 0.97 (0.98 without power control), while RPL and ZigBee obtain 0.85 and
0.92, respectively. On the other hand, both GBR and MinAP present a lower PDR in
the same situation (i.e., 0.61 and 0.68, respectively). This is because these construct
the routes based on the number of hops and the aggregated power, respectively. The

45
 3.4. Simulation Results

0.9

0.8

0.7
PDR

0.6

0.5

0.4 LEACH
GBR
ZigBee
RPL
0.3 MaxPDR (β = 1)
MaxPDR (no Power Control)
MinAP
0.2
10 20 30 40 50 60 70 80 90 100
L (m)

Figure 3.5. Reliability in terms of PDR for different deployment areas.

poor performance of LEACH in large deployment areas is because it considers that

all nodes can communicate with each other. Therefore, cluster-heads are not selected
according to the particular scenario. This is the reason why gradient routing approaches
generally have a higher reliability than clustering approaches, particularly in large-scale
deployment areas.

3.4.2. Energy Consumption

Figure 3.6 presents the average aggregated power of 30 nodes for different deployment
areas. From this figure, it can be observed that the gradient-based strategies adjust
their power consumption according to the scenario, but LEACH does not. In most
cases, MaxPDR with β=1 and MinAP present a lower energy consumption than GBR,
ZigBee and RPL thanks to their transmission power control mechanism. In fact, MinAP
can reduce the aggregated power of RPL in more than 2.5 mW. The transmission power
control is particularly useful in small deployment areas, since the average aggregated

46
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

8
LEACH
GBR
7 ZigBee
RPL
MaxPDR (β = 1)
MaxPDR (no Power Control)
6 MinAP
Aggregated Power (mW)

0
10 20 30 40 50 60 70 80 90 100
L (m)

Figure 3.6. Energy consumption in terms of aggregated power (mW) for different deployment areas.

power is 99% lower.

In Figure 3.7, the aggregated power is shown for different network densities. The sensing
area is 100x100 m2 . While GBR, ZigBee , RPL and MaxPDR with β=1 show a similar
consumption, it is interesting to observe that MinAP has a considerably lower slope
compared with the rest of strategies. Thus, its energy consumption is substantially
much lower, particularly in high density networks. For instance, MinAP reduces the
aggregated power of RPL more than 75% in the case of networks composed of 100 nodes.
Note that the aggregated power of MaxPDR without power control is very high. This
is because the PDR metric is very demanding in terms of energy consumption, since it
finds the most reliable routes without considering their energy consumption associated.
Hence, it is very important to reduce this consumption by adjusting the transmission
power of wireless sensors.

47
 3.4. Simulation Results

7
LEACH
GBR
ZigBee
6 RPL
MaxPDR (β = 1)
MaxPDR (no Power Control)
MinAP
5
Aggregated Power (mW)

0
10 20 30 40 50 60 70 80 90 100
N

Figure 3.7. Energy consumption in terms of aggregated power (mW) for different node densities.

3.4.3. Reliability vs. Consumption Tradeoff

In convergecast networks, the packets relayed toward the sink may need several transmis-
sions. This increases the energy consumption and reduces the reliability of the network.
On one hand, the energy consumption can be reduced by lowering the transmission power
of each hop. On the other hand, a high transmission power increases the reliability of
each link and also reduces the number of hops to reach the sink. Consequently, there is
a tradeoff between these two parameters, since both should be considered.

In Figure 3.8, the PDR vs. aggregated power of each strategy is represented in the
particular case of 30 nodes deployed in a 100x100 m2 area. The confidence intervals
(90%) of each value are included. Note that these intervals are large because nodes are
randomly deployed. In this figure, multiple β values (from 0.1 to 1) have been considered
for MaxPDR. Unfortunately, MaxPDR does not achieve a higher reliability than MinAP
using equal or less aggregated power. However, it is important to highlight that MaxPDR
increases the flexibility of the network, since this can be adjusted according to the specific

48
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

0.9

0.8

0.7

0.6
PDR

0.5

0.4

0.3
LEACH
0.2
GBR
ZigBee
RPL
0.1
MaxPDR (β = 1)
MaxPDR (no Power Control)
MinAP
0
1 2 3 4 5 6 7 8
Aggregated Power (mW)

Figure 3.8. Reliability vs. energy consumption tradeoff in terms of PDR vs. aggregated power (mW).

network requirements of each application using β. We can observe that for β values lower
than 1, MaxPDR obtains a lower PDR than RPL and ZigBee, but its aggregated power
is much lower. For instance, using β=0.9 the PDR of MaxPDR is 6% and 12.5% lower,
respectively, while the aggregated power is 39% lower.

3.4.4. Collision Probability

In Figure 3.9, the collision probability of MaxPDR and MinAP is calculated with and
without the proposed channel allocation mechanism. In addition, RPL, GBR and Zig-
Bee, and a multi-channel version of these, have also been included to illustrate the
problem of single-channel networks. Multiple transmission probabilities have been con-
sidered to represent different network traffic patterns, (i.e., a transmission probability of
1 means that nodes transmit continuously). A total of 30 nodes are randomly deployed
in a 100x100 m2 area.

The results show that the single-channel versions have a very high collision probability,

49
 3.5. Experimental Evaluation

0.9
MaxPDR (Single-Channel)
MinAP (Single-Channel)
0.8 RPL (Single-Channel)
GBR (Single-Channel)
ZigBee (Single-Channel)
0.7 MaxPDR (Multi-Channel)
MinAP (Multi-Channel)
RPL (Multi-Channel)
0.6 GBR (Multi-Channel)
ZigBee (Multi-Channel)
Collision Probability

0.5

0.4

0.3

0.2

0.1

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Tx Probability

Figure 3.9. Collision probability in MaxPDR for different transmission probabilities.

especially in scenarios with high throughput requirements. Note that using a multi-hop
approach, a single packet may require several transmissions to reach the sink. Conse-
quently, the collision probability is increased and the network throughput is reduced.
On the other hand, using a multi-channel approach the collision probability becomes
much lower. In fact, using MaxPDR or MinAP with the proposed channel allocation,
the collision probability is always below 12%. As a result, the collision probability can
be reduced up to 85%, compared with the original RPL, by adding this strategy. Note
that the remaining collisions mostly come from the multi-channel hidden terminal prob-
lem. Therefore, they could be avoided assuming an alternative MAC protocol instead
of CSMA-CA [Bac10].

3.5. Experimental Evaluation

In this section, MinAP and MaxPDR are implemented in commercial wireless sensors
and evaluated using a WSN testbed. The objective is to prove their implementability

50
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

and evaluate their performance in a real life application. In the version of MaxPDR
without power control, the minimum, medium and maximum transmission power levels
have been used.

3.5.1. WSN Testbed

The WSN testbed is composed of 13 IRIS motes from MEMSIC [MEM] (See Figure
3.10), which are compatible with the IEEE 802.15.4 standard, distributed as shown in
Figure 3.11. IRIS motes are equipped with the Atmel ATMega 1281 8-bit microcontroller
[Atm14], 8 KB of RAM and the RF230 transceiver [Atm09]. This operates at the 2.4
GHz band, it uses up to 16 channels, and its transmission power range is adjustable
from -17.2 dBm to 3 dBm and divided in 16 steps. This transceiver has two operating
modes: the basic mode, and the extended mode. In our implementation, the basic
mode has been chosen in order to better observe the impact of each strategy, since the
extended mode includes additional features that are out of the scope of this chapter (e.g.,
automatic address filtering, automatic acknowledgement). In particular, transmissions
are scheduled according to a simple time synchronized mechanism based on the ID of
each node. Therefore, the proposed approach is also compatible with more advanced
time synchronized mechanisms.

Each sensing mote transmits temperature and light measurements to the sink node
every second. Nodes have been deployed without any line of sight to favor multi-hop
transmissions. Moreover, the network has been located in a region with interferences
coming from other systems operating at the same frequency band, such as WiFi and
Bluetooth.

3.5.2. Experimental Results

The mean values of PDR and aggregated power obtained using each implementation
are shown in Table 3.1. These results indicate that MaxPDR provides the maximum

51
 3.5. Experimental Evaluation

Figure 3.10. IRIS mote [MEM].

7m

24 m

Figure 3.11. WSN testbed composed of 12 sensing motes (red spots) and a sink mote (grey square).

52
3. Joint Routing, Channel Allocation and Power Control to Support
High-Density Deployments

Table 3.1. Experimental Results. PDR and aggregated power (mW).

MaxPDR (no Power Control) MaxPDR

Pmax Pmed Pmin (β=1)
PDR 0.95 0.89 0.50 0.94
AP (mW) 4.489 1.256 0.049 0.805

MinAP RPL ZigBee

PDR 0.65 0.91 0.92
AP (mW) 0.077 2.539 2.826

reliability when the maximum transmission power is used (i.e., 0.95). However, this is
reduced as the transmission power is reduced, obtaining 0.89 and 0.5 using the medium
and minimum levels, respectively. On the other hand, applying the power control with
β=1, a high degree of reliability is also achieved (i.e., 0.94). In fact, this reliability is even
higher than the reliability obtained by RPL and ZigBee (i.e., 0.91 and 0.92, respectively),
confirming the simulation results presented in Section 3.4. Moreover, MaxPDR with β=1
also has a low aggregated power, reducing more than 3 times the aggregated power of
the original RPL, and more than 5.5 times compared with MaxPDR without power
control.

Obviously, the lowest aggregated power is found using MaxPDR with the minimum
transmission power. In contrast, it shows poor performance in terms of reliability. On
the other hand, MinAP has also a very low aggregated power (97% lower than the
original RPL), and it adjusts the transmission powers according to the scenario. As a
result, this strategy is the best option for applications with very low energy consumption
requirements, and no strict reliability constraints. For instance, this is the case of peri-
odic environment measuring applications, such as temperature or light monitoring. The
reason is that these kinds of measures are highly correlated in time and space. Thus, un-
received measures can be efficiently estimated using previous information or even using
measures received from other nodes.

53
 3.6. Summary and Conclusion

3.6. Summary and Conclusion

In this chapter, we address the node density problem in WSNs, enhancing the widely
implemented RPL routing protocol. We propose a pragmatic joint routing, channel allo-
cation and transmission power control solution for real life WSNs. Two different routing
metrics are proposed, referred to as MinAP and MaxPDR, which combine reliability and
energy efficiency criteria. Both include a power control mechanism to adapt the trans-
mission power according to the channel conditions and the QoS requirements. This uses
the DIO packets broadcast in RPL to not increase the overall signalling cost. Moreover,
a simple multi-channel mechanism is also included. This assigns a single pair of chan-
nels for transmission and reception to each node to avoid time-frequency synchronization
errors. MinAP and MaxPDR have been compared using simulation and experimental re-
sults with the original RPL, and also with LEACH, GBR and the gradient-based routing
protocol adopted in ZigBee.

The simulation results show that the proposed approach increases the reliability, and
reduces the aggregated power and the collision probability of RPL. More specifically,
MaxPDR can increase by more than 10% its PDR and MinAP can reduce its aggre-
gated power up to 99%. It is also worth noting that MaxPDR can be adapted using
the parameter β to deal with the reliability vs. energy consumption tradeoff. The ex-
perimental results show that MaxPDR achieves a PDR of 0.94, compared to 0.91 for
the original RPL, and it also reduces its aggregated power by 68%. On the other hand,
MinAP reduces its aggregated power by 97%, being a very interesting solution if the
energy consumption of the network is more important than its reliability. These results
have verified the implementability of the proposed RPL-based approach and confirmed
the promising results obtained through simulations.

54
 Chapter 4

Position Assisted Routing to Support

Node Mobility

4.1. Summary

In the near future, commercial wireless sensors will be easily attachable to mobile entities,
such as people, wild animals, vehicles, and so on. However, node mobility adds significant
challenges to the network routing due to the frequent topology changes. Therefore,
routing strategies designed for static WSNs are not efficient in these scenarios. On the
other hand, the routing protocols designed for mobile WSNs generally assume perfect
positioning information, but the positioning accuracy is limited, particularly in indoor
environments. Therefore, practical strategies are necessary for WSNs with mobile nodes
to further enhance their integration in society.

In Chapter 3, we introduced a RPL-based routing approach for energy efficient routing

in WSNs. The results have shown that this is a reliable and efficient routing strategy
for WSNs with static nodes. Unfortunately, the slow response to topology changes of
RPL and its high signalling cost to keep up-to-date routes in the presence of mobile
nodes makes it not appropriate for mobile scenarios. In this chapter, we further extend
RPL by introducing KP-RPL (Kalman Positioning - RPL), a novel routing strategy
for WSNs with both static and mobile nodes. The objective of KP-RPL is to provide

55
 4.2. Introduction

robust routing, taking into account the issues that arise in real-life WSNs with mobile
nodes. This combines RPL with Kalman positioning and blacklisting to develop a new
RPL metric. The simulation results show that the reliability and the robustness of the
network in harsh conditions is enhanced, compared to existent approaches. Moreover,
thanks to the Kalman filter, the rate of packets for positioning can be reduced to extend
the network lifetime.

4.2. Introduction

4.2.1. Motivation and Previous Work

Wireless sensor networks are planned to be dynamic network infrastructures, combin-

ing static and mobile nodes. However, the implementation of wireless sensor networks
in scenarios with mobile nodes is particularly challenging [Cad13]. This is because the
mobility of nodes deteriorates the network performance due to the continuous changes
in the network topology [Zam08]. Since energy efficient routing strategies are generally
designed for static WSNs [Pan14], they are not practical in mobile scenarios without sig-
nificant changes. Moreover, the additional issues that appear in real-life scenarios (e.g.,
obstacles, interferences, estimation errors), reduce the robustness of the communications
[Pen11].

The problem of mobility in wireless networks has been extensively studied in mobile ad-
hoc networks (MANETs). Some examples are the protocols ad hoc on-demand distance
vector (AODV) [Per99] and dynamic source routing (DSR) [Joh01]. However, routing
algorithms designed for MANETs are not recommended for WSNs due to their low energy
efficiency [Gad06]. One of the main reasons is that they consider many-to-many routes,
while communications in WSNs are generally many-to-one. The high energy efficiency of
clustering approaches (e.g., LEACH [Hei00]), has motivated researchers to adapt them
in order to support node mobility in WSNs (e.g., LEACH-M [Kim06], LEACH-ME

56
4. Position Assisted Routing to Support Node Mobility

[Kum08], CBR [Aww09]). Nevertheless, these suffer from the typical reliability problems
of clustering approaches (See Section 2.2). Another routing approach that supports
mobility is geographical routing [Kar00], which uses a greedy forwarding mechanism
whereby each node forwards a packet to the neighbor that is closest to the destination.
Since this method does not explicitly consider the link reliabilities, it is not recommended
for lossy networks. In [Zam08], the authors introduce the PRRxdistance metric, which
considers both the relative position and the Packet Reception Rate (PRR). This strategy
increases the network reliability compared to traditional geographical routing. However,
this strategy can still lead to the construction of unreliable paths, since it exclusively
uses local information in the forwarding decisions. Therefore, a different approach is
necessary to overcome the lack of reliability in this kind of scenarios.

Gradient-based protocols, such as RPL [Win12] or CTP (Collection Tree Protocol)

[Gna09], are reliable and efficient strategies for static WSNs. However, as shown in
[Gun13] and [Cob14], the mobility of nodes deteriorates their performance. Note that
since routes are defined in a proactive fashion, nodes need to keep up-to-date routing
tables. Therefore, the signalling required to maintain up-to-date routes is too costly in
terms of energy resources and overhead. Recently, some works tried to enhance RPL
in mobile environments. In [Lee12], RPL is implemented in Vehicle Ad-hoc Networks
(VANETs), and the impact of various RPL parameters is studied in these networks. In
particular, the authors modify the trickle algorithm and propose a new loop avoidance
mechanism. In [Saf12] and [Saa11], the mobility problem is addressed using mobile sinks.
They adapt the rate of control messages according to the speed of mobile nodes in order
to increase the energy efficiency of the network. In [Kor12], RPL is modified to differen-
tiate between mobile and static nodes, prioritizing routing through static nodes rather
than mobile nodes. In [Tia13], geographical information is included in RPL for WSNs
assisting VANETs. The authors propose a new routing metric to improve RPL in terms
of packet delivery ratio and overhead. In [Gad14], the authors propose Co-RPL, which
extends RPL with the Corona mechanism for the localization of mobile nodes in WSNs.
In Co-RPL, mobile nodes forward to the best neighbor within the same corona (i.e., cir-

57
 4.2. Introduction

cular area around the DODAG roots). In general, the previously mentioned approaches
enhance RPL in scenarios with mobile nodes, since they take into account important
issues that appear in these scenarios. However, they do not consider the inevitable posi-
tioning errors existent in real-life networks. Note that robustness is essential in WSNs,
since low power communication links are very sensitive to channel dynamics.

In this chapter, we introduce KP-RPL (Kalman Positioning - RPL) in order to enhance

RPL using positioning information from Kalman filtering. KP-RPL divides the network
routing into: i) routing among static nodes and ii) routing among mobile and static
nodes. On the one hand, we use RPL among static nodes, using ETX (Expected Trans-
missions) as the link quality metric. On the other hand, the routing among mobile and
static nodes follows a position-based approach. First of all, each mobile node generates
its own confidence region, which includes its most likely positions, taking into account
the channel conditions. Then, this position is refined using Kalman filtering. Previous
to the routing decision, mobile nodes generate a blacklist to discard nodes that may not
be reachable due to positioning errors. Finally, nodes use end-to-end ETX estimates to
select the best routing paths in each transmission time slot.

4.2.2. Contributions

The main contributions of this chapter are:

We propose a novel routing approach, referred to as KP-RPL, that combines RPL

with position-based routing. RPL provides reliable communications and position-
based routing provides mobility support. This is complemented with Kalman
filtering to reduce the impact of channel dynamics.

A blacklisting approach that improves the network reliability in the presence of

positioning errors. This is applied taking into account the accuracy of positioning
estimates.

58
4. Position Assisted Routing to Support Node Mobility

4.2.3. Organization of the Chapter

The remaining of this chapter is organized as follows: Section 4.3, presents an overview
of KP-RPL, differentiating the different: phases, positioning, blacklisting and routing.
Section 4.4, explains the positioning method of mobile node, which is based on confidence
regions and Kalman filtering. Section 4.5, motivates blacklisting and explains how this
is applied to reduce the impact of positioning errors. Section 4.7, evaluates KP-RPL
and compares it with other position-based routing approaches in terms of positioning
accuracy, PDR and ETX for different scenarios. Finally, Section 4.8 summarizes this
chapter and presents the main conclusions.

4.3. KP-RPL: Kalman Positioning - RPL

Currently, the RPL routing protocol does not incorporate efficient mechanisms for han-
dling mobile nodes, since it was designed to meet the QoS requirements of static low-
power and lossy networks (LLNs), such as wireless sensor networks. Its slow response to
topology changes results in the selection of suboptimal paths to the root and connectivity
losses, which may lead to severe degradation of the network performance.

In this chapter, we introduce KP-RPL in order to include mobility support in RPL

using position information. The block diagram of KP-RPL is shown in Figure 4.1. This
divides the routing problem into two different subproblems: anchor to anchor and mobile
to anchor routing. On one hand, the anchor to anchor links are constructed using the
traditional RPL algorithm [Win12]. On the other hand, the mobile to anchor links are
selected combining end-to-end ETX estimates, which are obtained using the positions
estimated by the Kalman filter, with blacklisting. Note that we avoid constructing links
among mobile nodes, since otherwise there would be two sources of positioning errors,
which may lead to frequent network disconnections.

First of all, each mobile node defines its own confidence region using RSSI (Receive

59
 4.3. KP-RPL: Kalman Positioning - RPL

Anchor to Anchor Mobile to Anchor

Create the confidence regions

POSITIONING

Estimate the position of Estimate the position of mobile

anchors nodes

Refine their position using

Kalman filtering

BLACLISTING
ROUTING

Create the routes using

Create the routes among
position-based ETX
anchors using RPL
estimations

Figure 4.1. Block diagram of KP-RPL.

60
4. Position Assisted Routing to Support Node Mobility

Gateway

Figure 4.2. WSN composed of static and mobile sensors, which are attached to people.

Signal Strength Indicator) measurements from anchor nodes. This region is the area
where this node is most likely to be located, taking into account the uncertainty of
RSSI measurements. Then, each mobile node defines its position as the location with
the highest probability within its own confidence region. This position is refined using
Kalman filtering, which combines it with velocity estimates. Note that the low memory
and computational requirements of Kalman filtering make it suitable for low power
wireless sensors [Rib10]. Before making the forwarding decision, each mobile node creates
a blacklist to discard potentially unreliable links, taking into account the variance of
the position estimated. Finally, the end-to-end ETX values of each possible route are
estimated, and the routing decisions are taken accordingly.

In the following sections, the positioning, blacklisting and routing phases are detailed.

4.4. Positioning Phase

Static nodes, also referred to as anchor nodes, are located in a well-known position so it
can be assumed that the information about their location do not contain errors. On the
other hand, the positions of mobile nodes have to be estimated frequently (See Figure
4.2). In this section, we introduce the positioning approach that is used by mobile nodes
prior to their routing decisions.

61
 4.4. Positioning Phase

4.4.1. Positioning based on Confidence Regions

In this section, we explain how mobile nodes estimate their position by overlapping the
confidence region of various anchor nodes, which are computed using RSSI measurements
as we will show in this section. This approach has been chosen instead of well-known
methods, such as least squares, because it defines the influence of each anchor node
individually. Then, the RSSI measurements from closer anchor nodes have a stronger
impact in positioning than those measured from distant anchors.

In order to create these regions, we assume that mobile nodes know the statistics of the
wireless channel that is time-varying and follows the one slope log-distance path loss
model, as assumed in Chapter 3. The value of an instantaneous RSSI measurement at
the k-th time period following this model is [Pat03]:

RSSI k = P1m − 10α log10 d − γ, (4.1)

where P1m is the received power at 1 meter from the transmitter, α is the path-loss
exponent, d is the transmission distance and γ ∼ N 0, σγ2 is zero mean Gaussian noise

that models the shadowing effects.

Following this model, the maximum likelihood estimator of the distance, between the
mobile node and an anchor node, using RSSI measurements is [Pat03]:

RSSI k −P1m
dˆkRSSI = 10 10α . (4.2)

Note that dˆkRSSI does not follow a Gaussian distribution due to the exponential relation-
ship between the distance and the RSSI value. In fact, the distance estimation follows
a log-normal distribution, that is:

 
ln dˆkRSSI ∼ N ln dk , σd2 , (4.3)

62
4. Position Assisted Routing to Support Node Mobility

where dk is the real distance to the anchor node and σd = (σγ ln 10) / (10α) is the
standard deviation of the distance estimation [Li07]. Then, the probability density
function (pdf ) of this estimation is:

2
ˆ

(ln d−ln d)
   √ 1 exp − 2 if dˆ > 0
fd dˆ = 2πσd dˆ 2σd
(4.4)
0 if dˆ ≤ 0,


where d is the real distance. From the pdf of the distance estimation, we create the
confidence region of the a-th anchor node, referred to as Ca , as the area where the
mobile node is located with confidence probability qLN . Then, assuming omnidirectional
antennas, Ca is a circle with radius r that fulfills Pr{dk ≤ r} = qLN , centered on the
a-th anchor node location, where qLN is defined as:
  2 
Z r
1 ln x − ln dˆkRSSI
qLN (r) ' √ exp −  dx. (4.5)
 
0+ 2πσd x 2σd2

Then, the m-th mobile node defines its confidence region by overlapping the confidence
regions of the L closest anchor nodes (Figure 4.3), that is:

Cm = C1 ∩ C2 ∩ ... ∩ CL . (4.6)

Within its confidence region, the mobile node estimates its position, referred to as p̂m ,
evaluating a set of discrete points uniformly distributed inside Cm , defined as Cm . As
far as the probability is almost constant in small areas, a density of 2 points/m2 suffices
in our particular implementation. Then, the mobile node estimates the probability to
be located in each point of Cm using (4.4). Assuming that the pdfs of different anchor
nodes are independent, the mobile node computes its position p̂m as:

Y
p̂m = max qLN (dc,a ) ∀c ∈ Cm , (4.7)
c
a∈Ap

63
 4.4. Positioning Phase

"m
^
#2RSSI
^1
# RSSI
1 ^
^
! m #3RSSI

Figure 4.3. Confidence region of a mobile node created overlapping the confidence regions of three
anchor nodes.

where Ap is the set of anchor nodes considered for positioning and dc,a is the euclidean
distance from the c-th point to the a-th anchor node (i.e., kc − pa k2 ).

4.4.2. Enhanced Positioning using Kalman Filtering

Once the position of the mobile node has been estimated from its confidence region, KP-
RPL combines it with velocity estimates using Kalman filtering [Kal60]. The Kalman
filter is a set of mathematical equations that provides an efficient computational solution
of the least-squares method [Wel95]. In the context of WSNs, this has been extensively
used in numerous applications, particularly for tracking and navigation [Cor14]. Its
low processing and memory requirements [Rib10] and the positioning accuracy obtained
with this filter motivate us to implement it in KP-RPL. In this section, we introduce the
basics of Kalman filtering, and the interested reader is referred to [May79] for a detailed
explanation.

A Kalman filter is articulated in two steps: i) the prediction step, and ii) the correction

64
4. Position Assisted Routing to Support Node Mobility

step. In the prediction step, the state vector xk−1 , which includes the information of the
mobile node at the (k-1)-th time instant, is updated using the transition matrix F. On
one hand, we model the state vector at the k-th time instant xk in a two dimensional
space by means of its position and velocity as:

h iT
xk = pkx pky vxk vyk , (4.8)

where pkx , pky and vxk , vyk represent the position and the velocity in Cartesian coordinates,
respectively. On the other hand, the transition matrix models the movement of the node,
which in our case is assumed to have constant velocity in x and y. Then, this matrix is
defined as:  
1 0 T 0
 
0 1 0 T
 
F=

,
 (4.9)
0 0 1 0
 
0 0 0 1

where T is the transmission period of the mobile node. Using this matrix, the state
vector xk is computed as:
xk = Fxk−1 + vk , (4.10)

where vk is a zero mean Gaussian noise with covariance matrix Q.

On the other hand, the state covariance matrix at the k-th time instant Pk , which
estimates the covariances of the positioning and velocity estimates in xk , is updated
as:
Pk = FPk−1 FT + VQVT , (4.11)

where V is the state noise covariance matrix, which models possible acceleration changes,

65
 4.4. Positioning Phase

that is defined as:  

0.5T 2 0
 
 0
 2
0.5T 

V=
 .
 (4.12)
 T 0 
 
0 T

Note that we have not considered the position and velocity measurements at the k-th
instant yet, since xk−1 is an input coming from the previous iteration. In the correction
step, we take into account the vector of measurements at the k-th instant zk that is
defined as:
h iT
k k k k k
z = p̃x p̃y ṽx ṽy , (4.13)

where p̃kx , p̃ky and ṽxk , ṽyk are the position and velocity measurements in Cartesian coor-
dinates, respectively. Note that the m-th node obtains p̃kx and p̃ky from p̂m , while ṽxk , ṽyk
are measured by its accelerometer.

Then, we correct the state vector and the state covariance matrix with zk as [Wel95]:

xk = xk + Kk (zk − Hxk ), (4.14)

and
Pk = (I − Kk H)Pk , (4.15)

where H is the measurement matrix, which maps the measurements into the state vector.
The Kalman gain K is computed as [Wel95]:

Kk = Pk HT (HPk HT + R)−1 , (4.16)

where R is the covariance matrix of measurements. Since the covariance of positioning

is difficult to estimate, this matrix has to be defined with a calibration process.

Finally, from the first two components of xk we obtain the new position estimation p̂∗m ,
and from Pk we obtain the estimated variance of this position σ̂k . In Section 4.5, we

66
4. Position Assisted Routing to Support Node Mobility

 k−! k−! k−! k−! T


x y x y

T
 pɶ kx pɶ ky 
Confidence
Regions KALMAN
FILTER p̂*m
T
vɶ kx vɶ ky 
σˆ k
Accelerometer

Figure 4.4. Inputs/outputs of the Kalman filter.

use the output of the Kalman filter to discard the nodes that may be unreachable from
the routing process. In Section 4.6, the quality estimation of links is computed with
p̂∗m .

Figure 4.4 shows the inputs and outputs of the Kalman filter.

4.5. Blacklisting Phase

Due to the limitations of positioning algorithms, it is impossible to avoid position inac-

curacies, which can cause a relevant degradation in the performance of position-based
routing algorithms [Pen11].

In order to diminish their impact in the reliability of the network, in KP-RPL mobile
nodes create a blacklist before making any routing decision. This list includes the nodes
that could be unreachable, taking into account the variance of position estimates. Basi-
cally, we force that the candidate anchor nodes of mobile nodes must be reachable from
any point inside their confidence region.

67
 4.5. Blacklisting Phase

After the Kalman filter, mobile nodes construct their confidence regions again, taking
into account the refined position and its variance. It is important to note that the
distances between this position and anchor nodes do not follow a log-normal distribution.
Assuming that the position error of p̂∗m is independent in x and y and Gaussian N 0, σ̂k2 ,

these distances follow a Rician distribution [Pap02]. Then, the confidence regions after
Kalman filtering, with confidence probability qR , are circles with radius r that fulfill
P r{dk ≤ r} = qR , centered on the anchor node location. The confidence probability qR
is defined as:
2 ! !
r
x2 + dˆK xdˆK
Z
x
qR (r) ' exp − I0 dx, (4.17)
0+ σ̂k2 2σ̂k2 σ̂k2

where dˆK is the euclidean distance from the position estimated by the Kalman filter to
the anchor node, and I0 is the modified Bessel function of the first kind with zero order.
∗ , from the intersection of the newly generated
Then, the m-th mobile node defines Cm
confidence regions, as in Section 4.4.1. Then, its set of candidate anchor nodes Km
includes the anchor nodes that fulfill the following condition:

∗
kc − pa k2 ≤ rtx , ∀c ∈ Cm , a ∈ A, (4.18)

where A is the set of anchor nodes in the network, and rtx is the transmission range of
mobile nodes.

Thanks to the newly generated confidence regions, only the anchors that can be reachable
by the mobile node (i.e., nodes in K), taking into account its position and velocity
estimates, are considered in the routing phase. Then, the probability of reaching the
anchor node selected is increased, since the impact of positioning errors is lower, and
therefore the robustness of the network is enhanced.

68
4. Position Assisted Routing to Support Node Mobility

4.6. Routing Phase

The routing problem in KP-RPL is decomposed into: i) anchor to anchor routing and
ii) mobile to anchor routing. This strategy exploits the simplicity and reliability of RPL
among static nodes, and provides mobility support to specific nodes using position-based
routing.

4.6.1. Anchor to Anchor Routing

Gradient-based routing is an energy efficient and reliable solution to construct converge-

cast graphs, as we have shown in Chapter 3. Following the same approach, we use the
RPL protocol for the routing among anchor nodes, since this is the most prominent
standard for LLNs. RPL constructs a convergecast directed acyclic graph rooted at the
sink node, referred to as DODAG (Destination Oriented Acyclic Graph), using control
messages referred as DIOs (DODAG Information Objects) (See Section 2.2.1).

The Rank of anchor nodes is the number of expected transmission (ETX). Basically, this
metric indicates how many times a message must be transmitted on average to reach
its destination. The i-th anchor node estimates its end-to-end ETX, when forwarding
to the j-th anchor node, as the ETX to reach the j-th anchor node (ET Xi,j ), plus the
end-to-end ETX of this node (ET Xj ). The ETX between two nodes is computed as the
reciprocal of their packet delivery ratio (PDR), then:

1
ET Xi = + ET Xj . (4.19)
P DRi,j

4.6.2. Mobile to Anchor Routing

Unfortunately, RPL is designed for static networks, and the mobility of nodes deterio-
rates its performance [Cob14]. In addition, the signalling required to maintain up-to-
date routes is too costly in terms of energy resources and overhead [Gun13]. In order to

69
 4.6. Routing Phase

overcome these issues, in KP-RPL mobile nodes make their routing decisions using posi-
tioning information, which is obtained from the Kalman filter. Moreover, mobile nodes
only consider their set of candidate anchor nodes (i.e., K) as possible parents.

Following the RPL approach, the mobile nodes select their parents according to the
ETX from themselves to the sink. Then, the end-to-end ETX of the m-th mobile node
(ET Xm ), when transmitting through the a-th anchor node, is estimated as the ETX
from itself to this anchor node (ET Xm,a ), plus the ETX from the anchor to the sink
(ET Xa ):
ET Xm = ET Xm,a + ET Xa . (4.20)

On one hand, ET Xa can be obtained from the Rank of the a-th anchor node, which is
computed using (4.19). Ranks are included in the Rank field of DIO packets, and the mo-
bile node may solicit this information using a DIS packet, as explained in [Win12].

On the other hand, the ET Xm,a is estimated as the reciprocal of P DRm,a . According
to [Ram06], this PDR can be estimated as the probability of receiving a packet at the
k-th time period with RSSI higher than the sensitivity of the transceiver (RSSIth ), that
is:
k
P DRm,a = P r{RSSIm,a ≥ RSSIth }. (4.21)

Assuming that RSSI measurements decay with distance following the log-distance path
loss model in (4.1), then:

P DRm,a = P r{γ ≤ P1m − 10α log10 dm,a − RSSIth }. (4.22)

Since γ is Gaussian N 0, σγ2 , we can estimate P DRm,a as:

X
γ2
Z  
1
P DRm,a = Pr{γ ≤ X} = √ exp − 2 dγ, (4.23)
−∞ 2πσγ 2σγ

where X has been used for the sake of clarity instead of P1m − 10α log10 dˆm,a − RSSIth .

70
4. Position Assisted Routing to Support Node Mobility

p̂ m p̂*m

RSSI Confidence Kalman Routing am

Regions Filter Decision
vm σˆ k

Blacklisting

Accelerometer

Figure 4.5. Overview of the communication among modules in KP-RPL (mobile nodes).

Note that dm,a has been substituted by dˆm,a , since we estimate the distances between
mobile and anchor nodes using the position estimated by the Kalman filter (p̂∗m ).

Finally, the m-th mobile node selects its parent am , from the set of candidate anchor
nodes Km , using (4.20) as:
 
1
am = arg min + ET Xk , ∀k ∈ Km . (4.24)
k P DRm,k

Note that (4.24) strongly depends on the accuracy of positioning, since P DRm,k is
estimated based on p̂∗m . Moreover, it also depends on the effectiveness of the blacklisting
strategy, since this decides the anchor nodes in Km .

In Figure 4.5, we summarize the communication among modules in KP-RPL, from the
creation of confidence regions to the final routing decision.

4.7. Simulation Results

In this section, the performance of KP-RPL, with and without blacklisting, is evaluated
using Matlab and compared with two existent geographical routing approaches. A ver-
sion of KP-RPL without Kalman filtering, referred to as P-RPL, is also included in the
simulations to observe the impact of Kalman filtering. In this case, the blacklisting and
routing decisions are taken using the position estimates obtained from the confidence
regions. We have considered the typical geographical routing forwarding criterion (i.e.,

71
 4.7. Simulation Results

nodes transmit to the node closest to the sink, as long as it is reachable) [Pen11], and
also the strategy proposed in [Zam08] (PRR x distance). In the latter case, each mobile
node multiplies the reliability of each link by the distance improvement towards the des-
tination, and then it selects the node that maximizes this value. We assume that both
geographical routing approaches use the well-known least squares algorithm to estimate
the position of mobile nodes. For the sake of a fair comparison, geographical routing
is only considered for the mobile to anchor communications, being anchor to anchor
communications always managed by the RPL protocol.

The received power at 1 m has been empirically measured using IRIS motes (i.e., P1m =
-47 dBm). A path loss exponent α=3 [Pu12] and a path loss variance σγ2 =6 [Rap01]
are assumed. Moreover, the sensitivity value and the maximum transmission power of
the IRIS transceiver (i.e., RF230 [Atm09]), have been considered (-91 dBm and 3 dBm,
respectively).

The scenario is composed of a mobile node, which moves across a 150x150 m2 sensing
area, and 100 anchor nodes uniformly distributed within this area in a square grid
formation. The mobile node moves with a speed of 2 m/s and transmits a packet per
second. In particular, we consider the trajectories of the mobile node in Figure 4.6. An
accelerometer provides speed estimates in x and y. The error of this measurements in
each direction follows a Gaussian distribution N 0, σs2 , with zero mean and σs = 0.5

m/s. In order to better observe the performance of each algorithm, we assume that the
MAC layer is able to avoid collisions among packets.

The mobile node considers its 4 closest anchor nodes for positioning, selected according
to its instantaneous RSSI measurements. Note that a higher number of anchor nodes
may reduce the positioning error, but it would also increase the computing complexity
of positioning. In both P-RPL and KP-RPL, the confidence regions of anchor nodes are
created using the 95% confidence intervals.

The most relevant parameters considered in these simulations are summarized in Table
4.1.

72
4. Position Assisted Routing to Support Node Mobility

150
Sink
Anchor Nodes
Trajectory

100
y (m)

50

0
0 50 100 150
x (m)

(a) Square trajectory.

150
Sink
Anchor Nodes
Trajectory

100
y (m)

50

0
0 50 100 150
x (m)

(b) Zig-zag trajectory.

Figure 4.6. Trajectories of the mobile node.

73
 4.7. Simulation Results

Table 4.1. Simulation Parameters

Parameter Value
Sensing area 150x150 m2
Anchor nodes 100
Speed of mobile nodes 2 m/s
Transmission period 1s
Transmit power 3 dBm
Sensitivity of the receiver -91 dBm
Power received at 1 m (P1m ) -47 dBm
Path loss exponent (α) 3
Path loss variance (σγ2 ) 6

4.7.1. Robustness against Channel Conditions

First of all, we evaluate the impact of the channel conditions in KP-RPL. In Figure 4.7,
we compare the positioning error of KP-RPL and P-RPL for different σγ . For the sake
of comparison, we also include the positioning error of the least squares algorithm. As
shown in this figure, we can obtain a better positioning accuracy using confidence regions
than using least squares. The reason is that confidence regions are created taking into
account the channel conditions, and the impact of RSSI measurements depends on the
distance from the mobile node to each anchor node. We can also observe that Kalman
filtering significantly enhances the positioning robustness thanks to the correction of
RSSI measurements using the velocity estimates from the accelerometer and the pre-
diction model. For instance, if σγ = 7 the positioning error in both square and zig-zag
trajectories is reduced more than 5 meters compared to least squares.

In Figure 4.8, we evaluate the reliability in terms of PDR for different channel conditions.
This shows that, even without blacklisting, both P-RPL and KP-RPL are more reliable
than geographical routing. Moreover, blacklisting increases their PDR up to 12%. This
figure also shows that in some cases P-RPL has a higher PDR than KP-RPL (See Figure
4.8b). It is important to remark that the routing metric of both protocols is ETX, and
although the PDR and ETX of a link are reciprocal values, in multi-hop networks they
are not equivalent. This is because ETX also depends on the number of hops (i.e.,

74
4. Position Assisted Routing to Support Node Mobility

8
Least Squares
P-RPL
7 KP-RPL

5
Position error (m)

0
1 2 3 4 5 6 7
N

(a) Square trajectory.

9
Least Squares
P-RPL
8 KP-RPL

6
Position error (m)

0
1 2 3 4 5 6 7
σγ

(b) Zig-zag trajectory.

Figure 4.7. Position error for different σγ .

75
 4.7. Simulation Results

ET Xa,b,c = ET Xa,b + ET Xb,c , P DRa,b,c = P DRa,b P DRb,c ). Therefore, in order to

increase the ETX of the network, KP-RPL reduces its PDR compared to P-RPL in
these cases. We evaluate this tradeoff in Section 4.7.5.

4.7.2. Number of Concurrent Positioning Nodes

In RPL, the nodes use the trickle algorithm [Lev11] to adapt the rate of DIOs according
to the current needs. With this algorithm, the energy consumption of these nodes
can be reduced when the network routes do not need to change frequently. However,
position-based routing needs frequent RSSI measurements, since the mobile node needs
to estimate its position before making any forwarding decision. As a result, the rate
of DIOs must be kept constant, leading to high battery consumption of anchor nodes.
Then, it is important to evaluate the impact of reducing the number of concurrent anchor
nodes that broadcast DIO packets to reduce the network consumption.

Assuming that anchor nodes do not coordinate their DIOs, in Figure 4.9 we evaluate the
average ETX for different activation probabilities (i.e., active anchor nodes broadcast
DIOs). In both trajectories, we observe a higher robustness of KP-RPL in front of low
activation probabilities than the rest of approaches. In fact, we can observe that KP-
RPL only needs an activation probability of 0.2 to achieve the same performance than
keeping all the anchor nodes always active. On the contrary, the rest of approaches need
an activation probability higher than 0.5 to achieve a similar performance.

4.7.3. Density of Anchor Nodes

A large density of anchor nodes enhances the positioning accuracy and the network
reliability, but it also increases the network cost. In order to observe the impact of the
density of anchor nodes, Figure 4.10 and Figure 4.11 show the positioning error and
the ETX for different densities, respectively. In this case, we have assumed a random
deployment of anchor nodes, since these cannot always be distributed in a square grid

76
4. Position Assisted Routing to Support Node Mobility

0.85
Geo Routing (d)
Geo Routing (PRR x d)
P-RPL (no Blacklisting)
0.8 P-RPL
KP-RPL (no Blacklisting)
KP-RPL
0.75

0.7
PDR

0.65

0.6

0.55

0.5
1 2 3 4 5 6 7
σγ

(a) Square trajectory.

0.9
Geo Routing (d)
Geo Routing (PRR x d)
P-RPL (no Blacklisting)
0.85 P-RPL
KP-RPL (no Blacklisting)
KP-RPL
0.8

0.75
PDR

0.7

0.65

0.6

0.55
1 2 3 4 5 6 7
σγ

(b) Zig-zag trajectory.

Figure 4.8. Packet delivery ratio for different σγ .

77
 4.7. Simulation Results

40
Geo Routing (d)
Geo Routing (PRR x d)
35 P-RPL (no Blacklisting)
P-RPL
KP-RPL (no Blacklisting)
KP-RPL
30

25
ETX

20

15

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Activation Probability (%)

(a) Square trajectory.

35
Geo Routing (d)
Geo Routing (PRR x d)
P-RPL (no Blacklisting)
30 P-RPL
KP-RPL (no Blacklisting)
KP-RPL
25

20
ETX

15

10

0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Activation Probability (%)

(b) Zig-zag trajectory.

Figure 4.9. Expected number of transmissions for different activation probabilities.

78
4. Position Assisted Routing to Support Node Mobility

3.5
Least Squares
P-RPL
KP-RPL

3
Position error (m)

2.5

1.5

1
100 120 140 160 180 200 220 240 260 280 300
N

Figure 4.10. Position error for different anchor node densities (Square Trajectory). The anchor nodes
are distributed randomly.

formation due to their number. We can observe that both in terms of position error
and ETX, the density of anchor nodes can be lower using KP-RPL than with the rest of
strategies for a given target performance. Note that KP-RPL obtains a lower position
error using 100 anchor nodes than least squares using 300 anchor nodes. This means
that Kalman filtering is particularly effective in low density deployments.

4.7.4. Energy Consumption

In Figure 4.12, we compare the average power consumption spent for transmitting data
packets from the mobile node to the sink. We assume the transmission consumption of
the RF230 transceiver (i.e., 16.5 mA), the typical size of IEEE 802.15.4 data packets (i.e.,
127 bytes), and a bitrate of 250 kbps. Besides the packet rate, this consumption mainly
depends on the average ETX of the network. In the simulation results, we can observe
that in most cases KP-RPL has the lowest energy consumption. This is particularly lower
in scenarios with bad channel conditions, thanks to the higher robustness of KP-RPL

79
 4.7. Simulation Results

5.8
Geo Routing (d)
Geo Routing (PRR x d)
P-RPL (no Blacklisting)
5.6 P-RPL
KP-RPL (no Blacklisting)
KP-RPL
5.4

5.2
ETX

4.8

4.6

4.4
100 120 140 160 180 200 220 240 260 280 300
N

Figure 4.11. ETX for different anchor node densities (Square Trajectory). The anchor nodes are
distributed randomly.

shown in Section 4.7.1. Therefore, KP-RPL reduces the overall energy consumption of
the network thanks to: i) the lower power that is necessary to transmit data packets, and
ii) the lower rate of positioning packets that is required to achieve accurate positioning,
as we have shown in Section 4.7.3.

4.7.5. ETX vs PDR Tradeoff

We have observed in Figure 4.8 that in some cases KP-RPL has a lower PDR than P-
RPL. This is because of the ETX vs. PDR tradeoff caused by blacklisting. On one hand,
the ETX of the network may be reduced if the confidence regions are too conservative
(i.e., mobile nodes may not be considering reliable paths). On the other hand, Figure
4.8 has shown that blacklisting increases the average PDR. Therefore, the size of the
confidence regions has a relevant impact on both ETX and PDR, and this impact may
be different in KP-RPL and P-RPL. In order to evaluate this tradeoff, Figure 4.13 shows
the values of ETX and PDR for different blacklisting criteria. We create the confidence

80
4. Position Assisted Routing to Support Node Mobility

0.07
Geo Routing (d)
Geo Routing (PRR x d)
P-RPL (no Blacklisting)
0.068 P-RPL
KP-RPL (no Blacklisting)
KP-RPL
Average Power Consumption (W)

0.066

0.064

0.062

0.06

0.058

0.056
1 2 3 4 5 6 7
σγ

(a) Square trajectory.

0.062
Geo Routing (d)
Geo Routing (PRR x d)
P-RPL (no Blacklisting)
0.06 P-RPL
KP-RPL (no Blacklisting)
KP-RPL
Average Power Consumption (W)

0.058

0.056

0.054

0.052

0.05

0.048
1 2 3 4 5 6 7
σγ

(b) Zig-zag trajectory.

Figure 4.12. Average power consumption (W) for different channel conditions.

81
 4.8. Summary and Conclusions

regions of anchor nodes using different confidence probabilities (q = qLN = qR ), where

low values create small blacklisting areas, while high values tend to create larger areas in
order to be more restrictive. We can see that the tradeoff can be managed by adjusting
the size of the confidence regions. Note that certain applications may require a minimum
PDR, and therefore q has to be increased, even if it also increases the average ETX.

4.8. Summary and Conclusions

The implementation of WSNs with mobile nodes in real-life scenarios is a challenging

task, particularly in environments with bad channel conditions, such as industrial sce-
narios. In this chapter, we address this issue with KP-RPL, which further enhances RPL
enabling it to cope with both static and mobile nodes. KP-RPL combines RPL with a
robust, yet simple positioning algorithm, based on the construction of confidence regions
and Kalman filtering. Moreover, a blacklisting strategy is applied to avoid considering
unreliable links. We have shown that Kalman filtering improves the routing decisions of
mobile nodes, since it enhances their position estimates. On the other hand, blacklisting
provides robustness in front of the inevitable positioning errors of mobile nodes, which
are due to the inaccuracies of RSSI measurements and velocity estimates.

In general, the simulation results show that KP-RPL is a reliable routing approach that
is also robust to channel conditions. We have evaluated the average ETX of KP-RPL
in different network conditions and mobile node trajectories, observing that it increases
the reliability of the network up to 25%, compared to geographical routing. The results
have shown that blacklisting enhances the network reliability, but it may also increase
the average ETX of the network. This generates a tradeoff that can be managed with
the size of the confidence regions, where the ETX is improved with smaller areas, and
larger areas increase the PDR. We have also evaluated KP-RPL for different anchor
node densities, since this has an impact on the positioning accuracy of the network. The
results show that the number of anchor nodes can be reduced using KP-RPL, compared

82
4. Position Assisted Routing to Support Node Mobility

3.36

P-RPL
KP-RPL
3.34

3.32

3.3
ETX

3.28

3.26

3.24

3.22
0.72 0.74 0.76 0.78 0.8 0.82 0.84 0.86
PDR

(a) Square trajectory.

3.05
P-RPL
KP-RPL

2.95
ETX

2.9

2.85

2.8

2.75
0.78 0.79 0.8 0.81 0.82 0.83 0.84 0.85 0.86 0.87
PDR

(b) Zig-zag trajectory.

Figure 4.13. Tradeoff between PDR and ETX using different confidence probabilities: q = 0.382 (0.5σ)
(bottom-left), q = 0.682 (1σ), q = 0.866 (1.5σ), q = 0.954 (2σ), q = 0.988 (2.5σ) and q = 0.998 (3σ)
(top-right).

83
 4.8. Summary and Conclusions

to geographical routing approaches. Thus, the infrastructure cost of the network is

lower. In terms of energy consumption, we have shown that KP-RPL enhances the overall
network consumption. On one hand, using KP-RPL the number of simultaneously active
anchor nodes can be reduced, since this needs less nodes for positioning. On the other
hand, the average energy consumption due to the transmission of data packets is also
lower, thanks to its higher reliability. As a result, the lifetime of both anchor and mobile
nodes is extended.

84
 Chapter 5

Multi-Tree Routing to Manage

Heterogeneous Traffic

5.1. Summary

The QoS requirements of the traffic generated by WSNs collecting a single phenomenon
can be easily managed by designing the routing protocol properly. However, the increas-
ing sensing and computational capabilities of wireless sensors enables the implementation
of more advanced services, which may combine data from different sensing units. More-
over, in the near future WSNs are planned to be shared infrastructures that may be
used by multiple on-demand services. This generates multiple traffic flows in the same
WSNs with their own QoS requirements.

In Chapters 3 and 4, homogeneous traffic is assumed. In order to manage heterogeneous

traffic, RPL may virtually split the network into multiple RPL Instances, which are
constructed according to a predefined objective function. However, this protocol does
not define any mechanism to decide the nodes that must belong to each instance, and this
decision has a strong impact in the network energy consumption and performance. With
this in mind, in this chapter we introduce C-RPL (Cooperative - RPL), which creates
multiple instances following a cooperative strategy among nodes with different sensing
tasks. As a result, the energy consumption, the complexity and the cost of the wireless

85
 5.2. Introduction

sensors is reduced compared to RPL, since they are active less time, perform fewer tasks
and are equipped with less sensing hardware. In this chapter we also propose a novel
fairness analysis for networks with multiple instances, showing that C-RPL achieves
a better tradeoff, in terms of performance and energy consumption, than RPL with
non-cooperative instances.

5.2. Introduction

5.2.1. Motivation and Previous Work

Many advanced WSN applications need to develop multiple tasks. For instance, SHM
(Structure Health Monitoring) [Har10] systems may need to collect information com-
ing from different sensing units, such as pressure, vibration or temperature. Moreover,
they also need to send alarm messages in case of broken sections or systems failures.
In addition, continuous messages are broadcast for external monitoring and calibration.
Wireless Sensor Surveillance Networks (WSSNs) may also need to combine multiple
traffics, such as event detection traffic, to detect intrusions or smoke, and positioning in-
formation of mobile entities [Bar13b]. All this heterogeneous traffic needs to be properly
managed by the network. For instance, latency is a critical requirement in event detec-
tion applications [Bag10]. On the contrary, critical monitoring tasks may admit a certain
delay in some cases, but they require a high reliability [Luo11]. On the other hand, am-
bient monitoring applications may not have strict delay or reliability constraints, but a
low energy consumption becomes crucial because these measurements are transmitted
periodically [Pan14].

Multi-objective routing approaches, such as [Mal06], [Alw13], consider multiple criteria

simultaneously to handle different QoS requirements. These strategies find a tradeoff
solution taking into account multiple objective functions at the same time. For instance,
in [Che10] the authors propose a solution to increase the lifetime and throughput of the

86
5. Multi-Tree Routing to Manage Heterogeneous Traffic

network, while reducing its latency. In [Yet12], the aggregate energy consumption and
delay are optimized using genetic algorithms. However, the requirements of different
kinds of traffic may not be satisfied with multi-objective routing, since they are not
addressed individually. Note that frequently, these requirements have contradictory
relationships among them [Mie03]. For instance, the lowest delay to the destination
is generally found using the minimum number of hops, while the solution with the
lowest energy consumption may require many shorter (i.e., low-power) transmissions
(See Section 2.5).

Generally, the individual QoS requirements of multiple traffic flows are addressed through
buffering or prioritized Medium Access Control (MAC) mechanisms [Ari13]. At the net-
work layer, it is also possible to provide QoS differentiation through multiple routing
trees. Although this mechanism has been mainly used in the literature for load balancing
[Chu11], or to increase the network robustness in front of faulty links [Wei07], it can also
be used to efficiently manage the QoS requirements of heterogeneous traffic [Lon13]. In
fact, the well-known RPL (Routing Protocol for Low-Power and Lossy Networks) pro-
tocol [Win12] has adopted this approach. In order to address multiple traffic flows, this
protocol virtually splits the network into multiple RPL Instances, where each instance
uniquely defines a set of one or more DODAGs (Destination Oriented Directed Acyclic
Graphs) with a common objective function. Then, each instance can be used to address
the particular requirements of a specific kind of traffic. In [Raj14], two RPL instances
are defined to individually manage latency constrained and high priority traffic in Smart
Grids. In particular, they construct each RPL Instance according to the minimum num-
ber of hops and the minimum expected number of transmissions (ETX), respectively. In
[Lon13], the authors differentiate between nodes for monitoring purposes, and nodes for
high priority traffic (e.g., alarms). Then, they create one RPL Instance that is only com-
posed of nodes from the first group, and also a second instance that groups nodes from
both groups. However, not any of these strategies define how to dynamically select the
nodes that belong to each instance, and therefore this decision must be taken in advance.
Note that the best solution depends on the objective function of each instance and the

87
 5.2. Introduction

particular network conditions. Therefore, RPL could either follow a low-consumption

strategy (i.e., each node belongs only to the instance associated to its tasks) or a high-
reliability strategy (i.e., any node may belong to any instance as long as it senses or
forwards the kind of traffic associated to that instance). On one hand, the first strategy
may reduce the network performance, due to the reduction of the node density of each
instance. On the other hand, the second strategy does not consider that nodes associ-
ated to different tasks may have different duty cycles. As a result, this solution may not
be energy efficient, since it does not prioritize the communication among nodes with the
same duty cycle. Then, many nodes have to extend their active time, thus increasing
their energy consumption.

In this chapter, we propose a cooperative RPL-based strategy (C-RPL) to manage this

tradeoff. This defines the nodes that belong to each instance, referred in C-RPL to
as C-RPL Instances, following a cooperative strategy. Taking into account the selfish
nature of nodes, coalitions are created according to a utility function that considers the
tradeoff between the performance and the energy consumption. From a game theoretical
perspective, the solution of this cooperation problem, such as the solution of the WSN
cooperation problem in [Md13], has been found to be equivalent to the well-known
prisoner’s dilemma game [Axe81]. Briefly, this is a two person zero game that describes
a situation where two players increase their utility if they both cooperate, but if a
player decides not to cooperate while the other cooperates, its utility gain is even higher
than cooperating. As a result, players will never cooperate (i.e., Nash equilibrium of
the prisoner’s dilemma game). This game is suitable for studying complex interactions
among players, such as the cooperation among RPL instances, since rational actions do
not cause the Pareto optimality.

On the other hand, since multiple performance criteria may be involved (e.g., latency,
bandwidth, energy consumption), it is important to distribute the network resources
in a “fair” manner (i.e., considering the requirements of each traffic in the network).
Although many definitions can be found in the literature to evaluate fairness [Shi14], such

88
5. Multi-Tree Routing to Manage Heterogeneous Traffic

as weighted fairness, max-min fairness or proportional fairness, to the best of the authors
knowledge, none of these definitions have been used before when different objective
functions are considered in the same network. In this chapter, we propose a metric to
evaluate the overall network fairness in networks with multiple instances, which may
have different objective functions.

5.2.2. Contributions

The main contributions of this chapter are:

We address the performance and energy efficiency issues that may appear in RPL in
the presence of heterogeneous traffic. Then, we propose a novel approach (C-RPL)
that coordinates the RPL Instances to form energy efficient coalitions according
to their individual objective functions and the network conditions.

We propose a mechanism to evaluate fairness in networks with multiple RPL In-

stances. This evaluates the distribution of the existent network resources to address
the different and sometimes contradictory objectives of each instance.

5.2.3. Organization of the Chapter

The remaining of this chapter is organized as follows: Section 5.3 provides some prelim-
inary results to validate multi-tree routing approaches in real-life scenarios. Section 5.4,
motivates and presents C-RPL, detailing the mechanism to create the C-RPL Instances.
This also provides an illustrative example of C-RPL applied to a WSN with three differ-
ent kinds of traffic. Section 5.5, proposes a metric to evaluate fairness in networks with
multiple instances. Section 5.6 compares the performance of C-RPL (with and without
the cooperation game) with RPL for different scenarios. Finally, Section 5.7 summarizes
this chapter and presents the main conclusions.

89
 5.3. Preliminary Results on Multi-Tree Routing

Table 5.1. Instances Coexisting in the Network

Traffic Objective Function

Positioning information Maximum PDR
Ambient conditions & Battery Minimum aggregated power
Fall detection Minimum number of hops

5.3. Preliminary Results on Multi-Tree Routing

First of all, we evaluate the performance of a simple multi-tree routing strategy in a real
scenario. The objective of this is section is to validate this approach in a commercial
WSN platform, since to the best of the authors knowledge, it has not been implemented
before.

We develop a habitat monitoring application that provides: i) the position of a mobile

entity, ii) habitat environment conditions, and iii) movement and falling detection of
the mobile entity, using an assembled accelerometer. In order to efficiently manage
each traffic flow, multiple trees are constructed according to their main requirements,
which are defined in Section 2.5 (See Table 5.1). The interested reader is referred to
[Cor12] in order to find detailed information about the positioning algorithm of the
mobile node.

The platform is composed of 14 MEMSIC IRIS sensors, which are equipped with the
RF230 transceiver chip [Atm09]. The mobile node is attached to the mobile entity. This
receives information from the rest of nodes, computes its own position, and transmits
it to the nearest node. The sink node, which is connected to the computer, gathers the
information coming from the network. The rest of nodes are distributed strategically to
cover all the points needed for positioning and ambient conditions monitoring. Once the
nodes have been deployed, a different tree is created for each traffic flow. Figure 5.1 is
a representation of the application, showing the position of the mobile node (red spot),
temperature information, battery status and alarms (movement and falling of the mobile
entity). The ambient conditions measurements are sent every 2 seconds, the positioning

90
5. Multi-Tree Routing to Manage Heterogeneous Traffic

Node 1 25.1 ºC

Node 2 25.3 ºC

1 2 3 4 5 Node 3 25.1 ºC

Node 4 25.5 ºC

Node 5 24.8 ºC

Node 6 25.3 ºC

SINK Node 7 25.2 ºC

6 7
Node 8 24.6 ºC

Node 9 25.0 ºC

Node 10 25.4 ºC

Node 11 25.2 ºC

Node 12 24.9 ºC
8 9 10 11 12
Movement: Yes
Falling: No

Figure 5.1. Example of a habitat monitoring application

Table 5.2. Experimental Results of a Multi-Tree Application

Alarms Temp. & Battery Positioning

Hops (#) 1,25 1,92 2,25
AP (mW) 2,49 0,077 4,489
PDR (%) 85,60 65,26 95,20

messages every 0.5 seconds, and the event detection messages are only sent when the
event occurs. The average results obtained through the experimental evaluation are
presented in Table 5.2.

It can be observed that it is possible to construct a tree optimized for each kind of traffic.
Using the proposed solution, the lowest delay and a reliability of 85.6% is guaranteed
for the alarm messages. On the other hand, the ambient conditions monitoring has a
very low consumption per packet associated. This is very important because of their
periodicity. Therefore, the reduction of the aggregated power associated to these packets
significantly extends the network lifetime. Finally, the positioning measures are received
with a PDR higher than 95%, which is considered to fulfill the high reliability demands
of positioning applications.

Although this simple multi-tree approach has been shown to be efficient managing multi-
ple QoS requirements, we have assumed that all nodes participate in all trees. Note that

91
 5.3. Preliminary Results on Multi-Tree Routing

in a real application, the tasks of each node should be defined according to the service
needs, to avoid wasting energy resources. This enables the network to put nodes into
sleep mode according to sleep scheduling strategies. In general, the radio transceivers
of wireless sensors have four states: send, receive, idle and sleep. In the receive and
send states nodes are sending or receiving information, respectively. In the idle state,
nodes monitor the wireless channel for incoming packets, and in the sleep state nodes
are inactive. Although, the transmit state consumes the highest amount of energy (e.g.,
the RF230 transceiver consumes 16.5 mA in the transmit state, and 15.5 mA in the
receive and idle states [Atm09]), the idle consumption is the most relevant in WSNs,
since wireless sensors generally spend much more time waiting for packets than trans-
mitting or receiving them. For instance, a node that periodically transmits a packet per
second, with a typical packet duration of 5 ms, spends around 99.5% of its time in the
idle state. On the contrary, the sleep state consumes much less energy (e.g., the RF230
transceiver consumes 20 nA in this state). Since the idle consumption wastes important
energy resources, we should reduce the time that wireless sensors remain unnecessarily
active in order to increase the network lifetime.

In order to illustrate this issue, in Figure 5.2 we show the consumption spent in idle and
send states of a WSNs (assuming the network parameters in [Bar13b]), using Matlab.
In particular, we compare the consumption of a single-tree and a multi-tree strategy,
with and without cooperation among trees. We can observe that the consumption for
transmitting may be an order of magnitude lower than the idle consumption in a single-
tree strategy. However, if we create multiple trees and allow them to cooperate, the
idle consumption can be reduced without affecting the network performance, since the
average active time of sensors is reduced.

In addition, we also need to consider the compatibility of the multi-tree strategy with
the current networking standards. With this in mind, in Section 5.4 we develop C-RPL
(Cooperative-RPL), which is a multi-tree routing approach build upon RPL.

92
5. Multi-Tree Routing to Manage Heterogeneous Traffic

5
Transmit
4.5 Idle (Single Tree)
Idle (Multi-Tree)
Idle (Collaborative Multi-Tree)
4

3.5
Power Consumption (Watts)

2.5

1.5

0.5

0
0 20 40 60 80 100 120 140 160 180 200
N

Figure 5.2. Comparison of the power consumption in transmit and idle states in a multi-tree network.
The network parameters in [Bar13b] are assumed.

5.4. C-RPL: Cooperative - RPL

In order to efficiently reduce the active time of sensors and address the multiple QoS
requirements of WSNs with heterogeneous traffic, we divide the network in multiple trees,
which group nodes according to their tasks in the network, and make them collaborate
only if it is necessary (See Figure 5.3). In particular, C-RPL uses the capability of RPL
to construct multiple RPL Instances, and then creates coalitions among them according
to their needs. This assumes selfish instances, which aim to maximize their performance
with the minimum energy consumption.

We assume that nodes support slotted listening duty cycling [Tse02], which is a simple
mechanism to reduce the idle energy consumption of the network. Basically, nodes peri-
odically wake-up and remain fully operative for a specific time period. The percentage of
time that a node is active defines its duty cycle, which must be set according to the QoS
requirements, where high duty cycles are required for urgent data, while low duty cycles

93
 5.4. C-RPL: Cooperative - RPL

Sink

Data Flow 1
Data Flow 2

Figure 5.3. Example of a wireless sensor network divided in two different trees, which cooperate among
them (red arrows).

may be used for non-critical data. Accordingly, we assume that each kind of traffic has
its own duty cycle associated, and they are not overlapped in time.

In this section, we first discuss the issues of RPL with heterogeneous traffic that motivate
C-RPL. Then, we explain how instances evaluate their utility related to each possible
coalition. Finally, we explain the cooperation game among RPL Instances that is used
to define the C-RPL Instances.

5.4.1. Problem Statement

RPL Instances operate independently from each other, and therefore their construction
is also independent. The nodes that may belong to each instance need to be defined in
advance, reducing the flexibility of the routing process. Then, a particular RPL Instance
may be either composed of: i) the nodes with tasks associated with this instance, or ii)

94
5. Multi-Tree Routing to Manage Heterogeneous Traffic

all the nodes in the network. In the first case, nodes only need to be active during the
active time assigned to their respective instance or instances. In the second case, any
node may belong to any instance as long as it senses or forwards traffic associated to
that instance. Therefore, in this case nodes need to be active during the time assigned
to its traffic, and also during the active time assigned to the rest of instances in which
they participate. Clearly, the first strategy may reduce the network performance due
to the reduction of the node density of each instance. On the other hand, the second
strategy may waste energy resources because of including nodes with different duty cycle
in the same DODAG. As a result, many nodes have to extend their active time, and
thus increasing the average energy consumption of the network.

In C-RPL, the nodes that belong to each instance do not need to be predefined, and
therefore C-RPL Instances can be constructed dynamically according to the objective
function of each instance, the location of nodes and the particular network conditions.
In general, C-RPL defines the nodes of each C-RPL Instance following a cooperative
approach among the nodes that develop the same task in the network. The objective of
each instance is to maximize its utility, taking into account the possible coalitions that
can be formed with the rest of instances in the network. A cooperation strategy among
instances is adopted, rather than a cooperation strategy among nodes, to prioritize the
group interests in front of individual interests. Note that local decisions among nodes
may also have an impact in the rest of the network. For example, if a node increases
its transmission rate due to a collaboration with another node, it is also increasing the
transmission rate of the nodes that forward its packets to the sink through the DODAG.
This increases their energy consumption and it may even cause congestion problems in
these nodes.

5.4.2. Rank Computation and Parent Selection

Following the RPL approach, each node selects its parent node using the concept of
Rank, which is defined according to the objective function. In this sense, the main

95
 5.4. C-RPL: Cooperative - RPL

difference between both approaches is that in C-RPL nodes compute additional ranks
to evaluate the possible coalitions that may be created in the network. These coalitions
may be composed of one or multiple instances.

In general, the n-th node defines its Rank and parent related to each coalition and
objective function, as:
∗,C k,C
rn,o = min rn,o ∀C ∈ P, o ∈ O, (5.1)
k∈C

pC k,C
n,o = arg min rn,o ∀C ∈ P, o ∈ O, (5.2)
k∈C

where C is the coalition, C is the set of nodes that belong to this coalition, o is the
objective function, P is the set of possible coalitions in the network, and O is the set of
objective functions.

In networks with more than three instances, we limit the cooperation game to the coali-
tions between two instances, and the coalition between all the instances (referred to as
the grand coalition). This is to prevent the number of possible coalitions to grow ex-
ponentially with the number of instances. Note that when the node density is low, the
instances will tend to create the grand coalition in order to improve their performance.
On the other hand, when the node density is high enough, the instances will tend to
create small coalitions in order to reduce their energy consumption. This is shown in
Section 5.6.

In C-RPL, all nodes need to broadcast the information associated with each possible
coalition. Currently, in RPL each DIO (DODAG Information Object) packet has a
unique RPLInstanceID and Rank fields. This may constrain the construction of mul-
tiple RPL Instances since the signalling overhead has to be increased. In fact, the
management of multiple instances has been identified as one of the open issues in RPL
that needs to be addressed in the following versions [Win12]. Although this is out of the
scope of this thesis, we propose a more flexible size of DIO packets, allowing to include
multiple RPLInstanceID and Rank fields. Then, although the size of each DIO is slightly
increased, the nodes do not need to broadcast a different DIO for each possible C-RPL

96
5. Multi-Tree Routing to Manage Heterogeneous Traffic

Instance.

5.4.3. Coalition Utilities

The creation of larger C-RPL instances may increase the performance of the network.
However, since coalitions are energy consuming, a utility function is used to decide which
coalitions, if any, are created.

When two instances agree to collaborate, they both share their nodes. Since they have
different duty cycles, the nodes from different instances must coordinate their active
times in order to communicate. In a slotted listening duty cycle scheme, this means
that either the transmitter or the receiver must remain active during the active time
assigned to the other instance, increasing its duty cycle. We define that the node that
adjusts its duty cycle in C-RPL is always the transmitter. On the other hand, the
nodes forwarding packets from other instances increase their transmission probability,
since they have to transmit additional packets. In general, both instances increase their
energy consumption by forming a coalition due to the additional duty cycle of their
nodes. Then, there exists a tradeoff between performance and energy consumption that
all instances need to evaluate before forming a coalition. The performance of each node
is characterized by its Rank, and its energy consumption is estimated taking into account
its duty cycle and average transmission rate in that coalition. In particular, we consider
the utility of each node in the coalition C, for a given objective function o, using linear
aggregation as:
∗,C ∗,C
uC


n,o = − ρr̄n,o + (1 − ρ)ēn,o , (5.3)

∗,C
where r̄n,o and ē∗,C ∗,C ∗,C
n,o are the mean-variance normalization of rn,o and en,o , respectively.

This is applied to homogenize the Rank and energy consumption values. Note that
the utility is inversely proportional to both Rank and energy consumption. A weighted
linear utility has been chosen, since it allows the network designer to easily adjust the
willingness of nodes to increase their Rank at the expenses of increasing their energy

97
 5.4. C-RPL: Cooperative - RPL

consumption using the cooperation parameter ρ. This must be configured according to

the application requirements.

Then, the instance I computes the utility of the coalition C as the minimum utility,
when its associated objective function oI is considered, among the nodes in I (i.e., set
of nodes that belong to this instance):

C
UI,o I
= min uC
n,oI . (5.4)
n∈I

Note that the minimum utility is considered to avoid unfair solutions among nodes, such
as instances that overload a particular node.

5.4.4. Cooperation Game among RPL Instances

Once the utilities of each possible coalition have been computed, the instances must
decide which coalitions are created in order to form the C-RPL Instances. We model
instances as selfish entities that aim to maximize their performance with the minimum
energy consumption. In this section, we show that without any coordination, the in-
stances would never agree to cooperate among them.

Following a game theoretical approach, we can model this game similarly to the prisoner’s
dilemma game [Axe81] (Table 5.3). This game defines R (reward) as the utility when
both players cooperate, S (sucker) as the utility when a player cooperates but the other
player does not, T (temptation) stands for the utility when a player does not cooperate
but the other player does. Finally, if neither cooperate, they both obtain a reward of P
(punishment). The relation among them is T > R > P > S.

Table 5.3. Prisoner’s Dilemma Game

Player 2
Cooperate Not Cooperate
Cooperate R1 ,R2 S1 ,T2
Player 1
Not Cooperate T1 ,S2 P1 ,P2

98
5. Multi-Tree Routing to Manage Heterogeneous Traffic

Comparing the prisoner’s dilemma game with the cooperation game among two RPL
Instances (Table 5.4), we can observe that they are very similar. In this game, each
instance is a player, where I1 is the instance that requests the collaboration and I2
is the instance that decides whether to collaborate or not. U c is the utility when an
instance cooperates (i.e., shares its nodes), and U n when it does not. Note that the
Not Cooperate - Cooperate case is not considered here because this would mean that I2
accepts a collaboration from I1 that has not been requested.

Table 5.4. Cooperation Game among RPL Instances

I2
Cooperate Not Cooperate
Cooperate U1c ,U2c U1c ,U2n
I1
Not Cooperate - U1n ,U2n

Since the utility function considers both the performance and the energy consumption,
U c can be either higher or lower than U n . However, analyzing this game, we can find
that the Nash equilibrium is the Not Cooperate - Not Cooperate solution. This is because
either if I2 is willing to cooperate or not, I1 always increases its utility by defecting to
cooperate, standing that I2 cooperates (i.e., the Rank remains equal and the consumption
is lower). As a result, even if I2 accepts to cooperate, I1 will also defect to cooperate
because in this new situation its utility is higher if it does not share its nodes. Since this
may not be the solution that maximizes their utilities, it is necessary to mediate in the
construction of coalitions.

In C-RPL, the cooperation game is managed by the instances, but supervised by the
sink. The closest node of each instance to the sink (referred to as instance head),
computes the utility of its respective instance for each possible coalition. Then, each
instance head finds the best coalition C ∗ and defines the C-RPL Instance, broadcasting
this decision to the rest of nodes, which set their parent nodes accordingly. In case
that the rest of instances in C ∗ do not join the coalition (i.e., this is not their best
coalition), the instance head updates C ∗ and broadcast it again to the rest of nodes,
creating a different C-RPL Instance. Instance heads periodically evaluate if their C ∗

99
 5.4. C-RPL: Cooperative - RPL

Table 5.5. Instances Coexisting in the Network

Instance Traffic Objective Function

IH Event detection H: Min. Hops
IE Non-critical monitoring E: Min. ETX
IP Critical monitoring P: Max. PDR

is still the best possible coalition. On the other hand, the coalitions are supervised by
the sink, which has knowledge about the performance that each instance should obtain.
Using these values as a reference, the sink can detect if any instance in a coalition does
not collaborate. In order to punish this instance, the sink should not acknowledge its
packets. In this case, that instance would have to collaborate again in order to continue
receiving ACKs. This is summarized in Figure 5.4.

5.4.5. Example of C-RPL

Finally, we present an illustrative example of C-RPL in a network with three instances,

which manage three different kinds of traffic. In order to cover the main kinds of traffic
that are present in WSNs (see Section 2.5), we consider event detection, non-critical
monitoring and critical monitoring traffic. A different objective function is defined for
each traffic according to their main requirements (i.e., latency, consumption and relia-
bility). In particular, in this example we consider the instances in Table 5.5.

The computation of Ranks is related to their objective function. Then, each node com-
putes its own Ranks as follows:

 
∗,C k,C ∗,C
rn,H = min rn,H = min Hopskn + rk,H ∀C ∈ P, (5.5)
k∈C k∈C

 
∗,C k,C ∗,C
rn,E = min rn,E = min ET Xnk + rk,E ∀C ∈ P, (5.6)
k∈C k∈C
 
∗,C 1
rn,P = k,C
min rn,P = min r∗,C ∀C ∈ P. (5.7)
k∈C k∈C P DRnk k,P

The value of Hopskn , ET Xnk or P DRnk is estimated by the n-th node using the DIOs

100
5. Multi-Tree Routing to Manage Heterogeneous Traffic

Compute the Utility

Associated to each
Possible Coalition

Find C* from all

Possible Coalitions

Join C*

Have All
Groups in C*
Yes
Joined
No
Is C* the Best
Coalition No

Remove C* from
Yes
Possible Coalitions

Is the Sink No
Responding
ACKs

Yes

Figure 5.4. Block diagram of the construction of C-RPL Instances

101
 5.4. C-RPL: Cooperative - RPL

∗,C ∗,C ∗,C

coming from the k-th node [Win12]. On the other hand, the values of rk,H , rk,E or rk,P
are broadcast by the k-th node within its DIO packets. Note that Ranks increase in the
direction of the leaf nodes and decrease in the direction of the sink node.

Figure 5.5 illustrates the possible coalitions evaluated by IE . This instance evaluates
the C-RPL Instance constructed using only its own nodes (Figure 5.5a), the C-RPL
Instances constructed in coalition with each other group: {IE , IP } and {IE , IH } (Figure
5.5b), and also the C-RPL Instance that groups the whole network: {IE , IP , IH } (Figure
5.5c). For each of these possible coalitions, the n-th node in IE computes a different
Rank:

∗,1 k,1
rn,E = min rn,E ∀k ∈ IE , (5.8)
k
∗,2 k,2
rn,E = min rn,E ∀k ∈ {IE , IP } , (5.9)
k
∗,3 k,3
rn,E = min rn,E ∀k ∈ {IE , IH } , (5.10)
k
∗,4 k,4
rn,E = min rn,E ∀k ∈ {IE , IP , IH } . (5.11)
k

On the other hand, the rest of the instances also evaluate all their possible coalitions.
Then, the nodes in IE compute some additional Ranks, in this case using the metrics of
the other instances. In particular, the Rank using the nodes in IE and the nodes of each
of the other instances: {IE , IP } and {IE , IP } (Figure 5.6.a), and also the Rank using the
whole network considering each respective metric {IE , IP , IH } (Figure 5.6b and Figure
5.6c). Then, the n-th node of IE computes the following additional Ranks:

∗,5 k,5
rn,P = min rn,P ∀k ∈ {IE , IP } , (5.12)
k
∗,6 k,6
rn,H = min rn,H ∀k ∈ {IE , IH } , (5.13)
k
∗,7 k,7
rn,P = min rn,P ∀k ∈ {IE , IP , IH } , (5.14)
k
∗,8 k,8
rn,H = min rn,H ∀k ∈ {IE , IP , IH } . (5.15)
k

∗,1 ∗,4
Summarizing, Ranks rn,E to rn,E are used by IE to evaluate the possible coalitions that

102
5. Multi-Tree Routing to Manage Heterogeneous Traffic

SINK

ETX

IP IE IH
∗,1
(a) rn,E
SINK

ETX ETX

IP IE IH
∗,2 ∗,3
(b) rn,E ,rn,E
SINK

ETX

IP IE IH
∗,4
(c) rn,E

Figure 5.5. Possible coalitions evaluated by IE . Circles indicate the nodes available to construct the
C-RPL Instance. The color of the circle indicates the Rank metric used to evaluate the coalition

103
 5.5. Fairness in Networks with Multiple Instances

can be formed with IP and IH , and are computed using the metric ETX. On the other
∗,5 ∗,8
hand, Ranks rn,P to rn,H are used to inform the rest of instances about the performance
that they would obtain by creating a coalition with IE . Note that these are computed
using the objective functions of the other instances, which are the minimum number of
hops and the maximum end-to-end PDR, in this case.

5.5. Fairness in Networks with Multiple Instances

The collaboration among instances can improve their performance, but it also increases
their energy consumption. Without control, these collaborations may overload a partic-
ular instance, depleting the batteries of its nodes too fast. This situation may cause the
loss of specific sensing capabilities. On the other hand, it is also important that these
collaborations try to satisfy all their different requirements. Therefore, it is important
that the available resources are shared in a “fair” manner among instances. Neverthe-
less, it is important to notice that the objective here is not to distribute the resources
uniformly, since instances perform different tasks with different energy consumption de-
mands.

In order to evaluate fairness, many definitions can be found in the literature (e.g.,
weighted fairness, max-min fairness, proportional fairness) [Shi14]. However, to the
best of the authors knowledge, not any of these definitions has been used before to eval-
uate fairness in networks with multiple instances. In this chapter, we propose to use the
following metric: !
1 X
F = min (xn − x∗n ) ∀I ∈ K, (5.16)
I |I|
n∈I

where K is the set of instances in the network, xn is the utility achieved by the n-th
node, and x∗n is the optimum utility of this node. Since we evaluate fairness among
instances, the average deviation of the nodes in each instance is computed. Note that
the maximum value of fairness is found when xn = x∗n ∀n (i.e., xn is always equal or
lower than x∗n ). Then, F is always zero or negative.

104
5. Multi-Tree Routing to Manage Heterogeneous Traffic

SINK

PDR hops

IP IE IH
∗,5
(a) rn,P
SINK

PDR

IP IE IH
∗,6 ∗,7
(b) rn,H ,rn,P
SINK

hops

IP IE IH
∗,8
(c) rn,H

Figure 5.6. Possible coalitions that involve IE evaluated by other instances. Circles indicate the nodes
available to construct the C-RPL Instance. The color of the circle indicates the Rank metric used to
evaluate the coalition.

105
 5.6. Simulation Results

In order to consider the tradeoff between performance and energy consumption, we

evaluate fairness in both senses. In particular, the fairness in terms of Rank (Fr ) and
energy consumption (Fe ) are defined as follows:
!
1 X  ∗,∗ ∗,C ∗

Fr = min r̄n,oI − r̄n,o ∀I ∈ K, (5.17)
I |I| I
n∈I

!
1 X  ∗,∗ ∗

Fe = min ēn,oI − ē∗,C
n,oI ∀I ∈ K, (5.18)
I |I|
n∈I

∗,∗
where oI is the objective function of I, C ∗ is the coalition selected by I, and r̄n,o I and

ē∗,∗
n,oI are the optimum mean-variance normalized Rank and energy consumption of the

n-th node, respectively. These are the best values among all the possible coalitions in
terms of Rank and energy consumption.

This metric evaluates how the network resources are balanced among the instances in the
network, taking into account the deviation of each node from its best possible solution.
In particular, a very low value in Fr means that there is an instance that is not receiving
enough resources, while a low value in Fe means that the network is overloading a
particular instance in terms of energy consumption. Since we want to avoid both extreme
situations, it is important to consider the tradeoff between both metrics.

5.6. Simulation Results

In this section, we compare the performance of C-RPL and RPL in a WSN with hetero-
geneous traffic using Matlab. In particular, we evaluate their performance for different ρ
values (i.e., cooperation parameter), node densities and traffic loads. Finally, we compare
each strategy in terms of fairness.

106
5. Multi-Tree Routing to Manage Heterogeneous Traffic

5.6.1. Simulation Environment

In RPL, the nodes that may belong to each instance must be defined in advance. In this
context, we consider two different versions of RPL, which construct their RPL Instances
using different criteria. In the first version, simply referred to as RPL, RPL Instances
may be composed of any node in the network, regardless of its specific tasks. In the
second version, referred to as RPL II (RPL with Independent Instances), nodes belong
only to the instance related to its tasks, and therefore to a unique RPL Instance. For
the sake of comparison, we also include a second version of C-RPL, referred to as C-
RPL NCG (C-RPL with No Cooperation Game), in which each instance constructs the
C-RPL Instance that maximizes its utility independently. Note that this strategy is not
equivalent to RPL, since this considers not only the Rank, but also the energy consump-
tion of each possible C-RPL Instance using (5.4). This has been included to show the
lack of fairness that may arise without the cooperation game among instances.

We assume a slotted listening duty cycling mechanism that adjusts the duty cycle of
each node according to its tasks in the network. Therefore, nodes remain active only
if they are sensing, sending or expecting to receive data. Moreover, in order to better
observe the impact of each routing strategy in the network performance, we assume that
there are no collisions among packets.

The simulation scenario considers a total of N sensing nodes, randomly deployed in a

150x150 m2 area. The information collected by them is gathered at the sink node, which
is located in the middle of the sensing area.

The path losses (P L) are modeled following the log-distance path loss model:

P L(dB) = P L0 (dB) − 10αlog(d) + γ, (5.19)

where P L0 is the path loss at the reference distance d0 (1 m), α is the path loss exponent,
d is the communication distance, and γ is the attenuation caused by flat fading, which

107
 5.6. Simulation Results

has zero mean and variance σ 2 . The value of P L0 has been empirically measured using
IRIS motes (i.e., P L0 = 50 dB). A path loss exponent α=3 [Pu12] and a path loss variance
σ 2 =6 [Rap01] are assumed. Moreover, we assume the sensitivity value (-91 dBm), the
maximum transmission power (3 dBm) and the current consumptions in sleep, idle and
transmit states (20 nA, 15.5. mA and 16.5 mA, respectively) of the IRIS transceiver
(i.e., RF230 [Atm09]). Table 5.6 summarizes the main simulation parameters.

Table 5.6. Simulation Parameters

Parameter Value
General
Sensing area (L x L) 150 x 150 m2
Path loss at 1 m (P L0 ) 50 dB
Path loss exponent (γ) 3
Flat fading variance (σ 2 ) 6
Packet size 127 bytes
Radio Transceiver
Maximum data rate 250 kbps
Current consumption:
Transmit state (Itx ) 16.5 mA
Idle state (Iidle ) 15.5 mA
Sleep state (Isleep ) 20 nA
Transmission power 3 dBm
Sensitivity -91 dBm
Supply voltage (VDD ) 3V

In particular, the sink collects event detection, critical monitoring and non-critical mon-
itoring measurements, coming from the wireless sensors, which generate up to 250 kbps.
The size of packets is assumed to be 127 bytes (i.e., standard packet size in IEEE
802.15.4). Each kind of data is sensed by a different instance, which are composed of
the same number of nodes. Since instances may sense different magnitudes, the traffic
originated at different instances cannot be generally aggregated. On the other hand,
the traffic from the same instance can be compressed using data aggregation techniques,
since they may be spatially and temporally correlated. In particular, we assume a com-
pression rate of 80% [Cap12] (i.e., nodes can combine up to 5 packets into a single
packet on average). The duty cycle, the average packet transmission rate and the objec-

108
5. Multi-Tree Routing to Manage Heterogeneous Traffic

tive function of each instance are defined according to the particularities of each kind of
data (See Table 5.7).

Table 5.7. Traffics Managed by the Network

Traffic Duty Cycle Packets/s Objective

Event detection 20 % 0.01 Min. Hops
Non-critical monitoring 5% 0.1 Min. ETX
Critical monitoring 10 % 0.2 Max. PDR

5.6.2. Cooperation Parameter

In Figures 5.7 to 5.9, we show the impact of the cooperation parameter ρ in the network
performance, in terms of the particular objective function of each instance. Although
individual ρ values could have been considered for each node or instance, we assume a
uniform ρ, for the sake of simplicity. Each instance is composed of 40 nodes. In general,
we can observe that instances do not cooperate until a minimum ρ is considered. This is
because the network prioritizes the energy consumption in front of Rank (i.e., Hops, ETX
and PDR) when using low ρ values. On the other hand, when the cooperation parameter
is increased, C-RPL and C-RPL NCG tend to the RPL solution, since instances are more
prone to collaborate with other instances, as long as they increase their performance. We
can also observe that without the cooperation game, the instances start creating larger
C-RPL Instances using lower ρ values. This is because these are decided individually
by each instance. In this case, the best performance is always found using RPL, but in
situations with congestion problems this may not be always the case, as we discuss in
Section 5.6.4.

In Figure 5.10, the average energy consumption is shown for different ρ values. As we
have observed before, the network prioritizes the energy consumption if low ρ values are
considered. The average power consumption in C-RPL increases with ρ, from 5.5 mW to
8.5 mW, due to the additional communication among instances. This causes that more
nodes increase their duty cycles, and therefore their energy consumption also increases.

109
 5.6. Simulation Results

5.5
RPL
RPL II
C-RPL
5 C-RPL NCG

4.5
Hops

3.5

2.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ρ

Figure 5.7. Average number of hops of IH .

5
RPL
RPL II
4.8 C-RPL
C-RPL NCG

4.6

4.4

4.2
ETX

3.8

3.6

3.4

3.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ρ

Figure 5.8. Average ETX of IE .

110
5. Multi-Tree Routing to Manage Heterogeneous Traffic

0.95

0.9

0.85
PDR

0.8

0.75

0.7

RPL
0.65 RPL II
C-RPL
C-RPL NCG
0.6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ρ

Figure 5.9. Average PDR of IP .

For the same reason, C-RPL NCG, has a higher consumption than C-RPL. In fact, this
may be even higher than in RPL.

5.6.3. Impact of Node Density

In Figures 5.11 to 5.13, we evaluate the performance of RPL and C-RPL for different
network densities, considering ρ=0.9 and the same number of nodes for each instance
(N/3). In networks with a low density of nodes, the best performance is achieved by
RPL. On the contrary, when instances are independent from each other (i.e., RPL II), the
number of possible routes is reduced, and therefore their performance is lower. When
the density of the RPL Instances increases, RPL II improves its performance, since
the communication distances are shorter. In general, the performance of C-RPL is an
intermediate value between both solutions. Note that the performance of each instance
could be individually adjusted with individual ρ parameters.

As we can observe in Figure 5.13, a node density above 130 nodes generates network

111
 5.6. Simulation Results

8.5

Average Power Consumption (mW) 8

7.5

6.5

RPL
5.5 RPL II
C-RPL
C-RPL NCG
5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ρ

Figure 5.10. Average power consumption (mW) of the network for different ρ values.

congestion in RPL, reducing the network performance. Note that IP is the instance
that generates the highest amount of traffic, and therefore it is congested at lower node
densities than the rest of instances. However, the impact of node density is different
in C-RPL, since C-RPL avoids creating C-RPL Instances with overloaded nodes. We
discuss the network congestion problem in the next section.

In general, C-RPL reduces the average number of coalitions with the node density, since
instances do not need to create large C-RPL Instances to have a good performance. As
a result, the average energy consumption of C-RPL can be always lower than RPL. This
can be observed in Figure 5.14, where C-RPL reduces the average power consumption
of RPL around 17%.

5.6.4. Impact of Traffic Load

The packet rates in the previous sections were not high enough to cause congestion
problems with 120 nodes. However, bottlenecks may frequently appear in convergecast

112
5. Multi-Tree Routing to Manage Heterogeneous Traffic

6.5
RPL
RPL II
6 C-RPL
C-RPL NCG

5.5

5
Hops

4.5

3.5

2.5
90 100 110 120 130 140 150 160 170 180
N

Figure 5.11. Average number of hops of IH .

6
RPL
RPL II
C-RPL
5.5
C-RPL NCG

5
ETX

4.5

3.5

3
90 100 110 120 130 140 150 160 170 180
N

Figure 5.12. Average ETX of IE .

113
 5.6. Simulation Results

0.9

0.8
PDR

0.7

0.6

0.5 RPL
RPL II
C-RPL
C-RPL NCG
0.4
90 100 110 120 130 140 150 160 170 180
N

Figure 5.13. Average PDR of IP .

8.5

8
Average Power Consumption (mW)

7.5

6.5

RPL
5.5 RPL II
C-RPL
C-RPL NCG
5
90 100 110 120 130 140 150 160 170 180
N

Figure 5.14. Average power consumption (mW) for different node densities.

114
5. Multi-Tree Routing to Manage Heterogeneous Traffic

1
RPL
RPL II
0.9 C-RPL
C-RPL NCG

0.8

0.7
PDR

0.6

0.5

0.4

0.3

0.2
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Packets/s

Figure 5.15. Average PDR of IP for different packet rates.

networks when the traffic is moderate, since the traffic is many-to-one. In this case, some
nodes may not be able to forward all packets during their corresponding duty cycle. In
Figure 5.15, we show the impact of traffic congestion using different packet rates in
IP , while the packet rates of the rest of instances are not modified. We can observe
that RPL is the most affected strategy due to the higher number of communications
among nodes from different instances. As a result, a large amount of traffic cannot be
aggregated, causing congestion problems in the network. For example, when each node
in IP generates 4 packets/s, the PDR is around 0.3. On the other hand, when instances
are completely independent, the traffic can always be aggregated and therefore its impact
is much lower (above 0.6 in the same case).

C-RPL manages this issue adjusting the cooperations according to the traffic conditions.
When the traffic is low, the instances tend to form larger C-RPL Instances to achieve
a better performance by collaborating among them. However, when the traffic is high
enough to cause congestion problems, the C-RPL Instances tend to be smaller in order
to reduce the amount of traffic that cannot be aggregated.

115
 5.7. Summary and Conclusions

5.6.5. Fairness

In applications with multiple sets of requirements, the network must not be focused in
one specific task. Instead, it must distribute the available resources in a fair way to
satisfy the requirements that each traffic demands. In Figures 5.7 to 5.14 we have shown
that C-RPL and C-RPL NCG are intermediate solutions among RPL and RPL-II in
terms of performance and energy consumption, which can be adapted using ρ. In this
section, we show that C-RPL distributes the network resources more fairly than RPL.
Figure 5.16 shows the fairness of each strategy in terms of energy consumption and Rank
for different values of ρ. As we can observe, using RPL II the consumption fairness is
close to zero, but the worst Rank fairness is obtained (around -1.17). On the other
hand, using RPL we can obtain a better Rank fairness, but this is much worse in terms
of consumption fairness (around -1.45). We can also observe here that using C-RPL
we can obtain intermediate solutions according to the ρ parameter. For instance, using
ρ = 0.8 C-RPL obtains a fairness around -0.75, both in terms of Rank and consumption.
Note that intermediate solutions mean that the resources are fairly shared, and not any
instance is overloaded.

In this figure, we can also compare the fairness degree of C-RPL when the cooperation
game is applied or not. In general, we can observe that C-RPL avoids solutions with a low
consumption fairness thanks to the cooperation game. On the other hand, C-RPL NCG
tends to provide a better Rank fairness, regardless of its consumption fairness.

5.7. Summary and Conclusions

In this chapter, we have presented C-RPL, a multi-tree routing approach to deal with
the traffic heterogeneity in WSNs. The objective is to adapt the active time of wireless
sensors according to the network needs, and also satisfy the particular QoS requirements
of each traffic flow. As we did in the previous chapters, we base our strategy on the

116
5. Multi-Tree Routing to Manage Heterogeneous Traffic

-1.8
RPL
RPL II
-1.6 C-RPL
C-RPL NCG
-1.4

-1.2
Consumption Fairness

-1

-0.8

-0.6

-0.4

-0.2

0
0 -0.2 -0.4 -0.6 -0.8 -1 -1.2
Rank Fairness

Figure 5.16. Rank and consumption fairness for ρ = 0, 0.5, 0.6, 0.7, 0.8, 0.9 and 1, where the smallest
ρ is located on the right bottom side.

RPL protocol for the sake of compatibility with the current networking standards. In
particular, we use the capability of RPL to create virtual subnetworks, referred to as
RPL Instances, which can be constructed with different objective functions. A multi-
tree strategy has been adopted due to its simplicity and compatibility with the RPL
protocol. This is particularly interesting in WSNs managing traffics with opposite sets
of QoS requirements, and therefore they need to be addressed individually. Note that
otherwise, multi-objective routing could be applied.

C-RPL creates energy efficient instances, referred to as C-RPL Instances, following a

cooperative strategy among nodes with different sensing tasks. Since nodes would never
agree to cooperate, C-RPL coordinates them to create coalitions that improve the trade-
off between their own performance and energy consumption using the cooperation param-
eter ρ. This can be individually defined for each DODAG according to the application
requirements.

The simulation results show that C-RPL adjusts the C-RPL Instances according to the

117
 5.7. Summary and Conclusions

individual objective function of each traffic flow and the current network conditions.
In particular, we have observed that C-RPL tends to create large instances when the
node density is low. Otherwise, it constructs smaller instances to reduce the energy
consumption of the network and avoid congestion problems. Besides, we have proposed
a fairness analysis for networks with multiple instances. This measures the distribution
of the network resources when multiple objective functions coexist in the network. The
results show that C-RPL obtains a more balanced fairness than RPL-based approaches
with non-cooperative instances.

118
 Chapter 6

IoT-Cloud Formulation to Integrate

WSNs into the Future Internet of Things

6.1. Summary

The confluence of distributed cloud networking and the IoT will enable a new range of
services in “smart” environments (e.g., smart cities, smart grids, smart transportation
systems). Wireless sensors are planned to be a fundamental part in this new paradigm
and therefore their efficient integration is critical. On one hand, the IoT-Cloud augments
the computing resources of sensors and extends their battery life. On the other hand, the
data gathered by different sensing platforms can be shared on a bigger scale, enabling a
more efficient exploitation of the physical infrastructure.

In Chapters 3, 4 and 5, WSNs were assumed to be isolated networks that merely collect
specific data to be gathered at a central server. However, thanks to the recent advances
in network programability and virtualization, wireless sensors can be integrated into
IoT-Cloud networks as distributed sensing and computing resources. In this chapter, we
formulate this problem as a minimum cost flow problem using only linear constraints.
The objective is to the find the optimal placement of virtual functions over the IoT-
Cloud that meets user requests, satisfies network resource capacities, and minimizes
overall network cost. We solve this problem for an illustrative set of smart city services,

119
 6.2. Introduction

where users interact with the city using their smart devices.

6.2. Introduction

6.2.1. Motivation and Previous Work

The Internet of Things (IoT) connects many heterogeneous devices, such as wireless
sensors, RFID devices, smartphones, wearables and connected vehicles, which produce
and/or consume information in real-time. When connecting the IoT to the Cloud, in-
formation collected from multiple locations can be processed and analyzed to produce
meaningful information, which can be accessed remotely [Ala13]. At the same time, the
intrinsic limitations of lightweight mobile devices (e.g., battery life, processing power,
storage capacity) can be alleviated by taking advantage of the extensive resources in the
Cloud using offloading techniques [Kum10]. The resulting IoT-Cloud paradigm enables
the network designers to implement a new breed of services and applications (e.g., health
system monitoring, traffic control, energy management, vehicular networking), which are
expected to define the essence of next generation smart environments (e.g., smart cities,
smart homes, smart grids [Bot14]).

With the increasing number of services and connected devices, IoT traffic is expected to
grow dramatically in the coming years [Wen14]. In order to satisfy the QoS requirements
of end users and increase the network efficiency, cloud networks are becoming increas-
ingly distributed (i.e., composed of a large number of dispersed cloud nodes). This has
motivated the placement of low complexity cloud nodes in close proximity to the device
layer, such as cloudlets [Sat09] or micro-clouds [Shi13]. In this context, and with the
particularities of the IoT in mind, in [Bon12] the authors propose to expand the Cloud
paradigm up to the device layer, creating the so-called Fog, making the analogy with a
cloud close to the ground. In fog networks, the end-layer can handle part of the compu-
tational tasks [Zhu15]. As a result, the latency and resource utilization of the network

120
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

resources can be improved compared to traditional cloud computing [Aaz14].

Wireless sensors are expected to assume a crucial role in IoT-Cloud networks thanks to
their ubiquity and low cost. Therefore, it is very important to efficiently integrate them
into the Fog. This means that they can be orchestrated together with the rest of cloud
nodes and IoT devices. Note that although wireless sensors have been conceived to be
application-specific, these can become cloud ready infrastructures using software-defined
networking (SDN) [Nas14]. The platforms integrating WSNs into the Cloud are referred
in the literature to as Sensor-Cloud architectures [Ala13]. These combine the extensive
processing resources of cloud servers with the ubiquity of wireless sensors. Moreover,
the data gathered by different WSNs, which generally belong to different operators,
may be shared on a bigger scale [Xia13], [Mis14b]. In [Luo12], the authors propose
Sensor-OpenFlow using the separation of the control and data planes applied in Open-
Flow [McK08]. This makes the underlying network (i.e., data plane) programmable by
manipulating a user-customizable flow table for each sensor. In [Mit12], the authors
present a Sensor-Cloud architecture that makes use of the Contiki operating system
[Dun04]. This aims to provide a platform that can obtain any type of data from dif-
ferent heterogeneous sensing infrastructures. In [Asl12], the authors introduce a service
oriented architecture, referred to as WSN-SOrA, which is designed to orchestrate the
service provisioning in large scale WSNs. In [Mis14a], the importance of the gateway
in Sensor-Cloud architectures is highlighted. Then, an optimal gateway selection ap-
proach is proposed to minimize the network delay. In [Mis14b], the authors study the
performance enhancements that can be obtained using Sensor-Cloud platforms over tra-
ditional WSN architectures. The main security issues that may appear in these networks
are addressed in [Ahm14].

The previously cited works focus on the architectural and implementation challenges
associated to Sensor-Cloud platforms. However, it is still necessary to mathematically
formulate the optimization of services in IoT-Cloud networks, taking into account the
heterogeneity of sensing, transmission, and computing resources across the physical in-

121
 6.2. Introduction

frastructure, as well as the tight IoT QoS requirements. Then, in this chapter we for-
mulate the problem of finding the optimal resource allocation in IoT-Cloud networks
that meets user requests, satisfies network resource capacities and minimizes overall net-
work consumption. We refer to this problem as the service distribution problem (SDP).
In the context of cloud networks, we introduced the cloud service distribution problem
(CSDP) in [Bar15]. The CSDP is formulated as a minimum cost mixed-cast flow prob-
lem [Ahu93], in which cloud services are represented by a service graph that encodes the
relationship between input and output information flows. This solution jointly optimizes
the use of compute, storage and transport resources in arbitrary cloud network topolo-
gies. Moreover, this is able to capture flexible service chaining, resource consolidation
savings, unicast and multicast delivery, and latency constraints. However, the generality
of the CSDP, and in particular of its cost function, while valuable in its intention to
capture a wide range of business models, leads to a mixed integer linear program of
exponential complexity.

In this chapter, we focus on the optimization of IoT services in IoT-Cloud platforms. We

adopt a linear pay-as-you-go cost model, in which the cost of a service is proportional to
its use of the physical infrastructure. In addition, we include specific models that char-
acterize the particularities of IoT-Cloud networks (wireless access, wireline cloud metro)
and end devices (e.g., battery life, processing power, sensing capability, radio technology)
of essential importance in IoT environments, and not yet characterized in the existing
literature. We provide an efficient linear programming formulation to the IoT-Cloud
SDP that can be solved in polynomial time, enabling the dynamic and elastic config-
uration of IoT-Cloud resources for real-time IoT services based on short-term demand
estimations. Then, we solve the IoT-Cloud SDP for an illustrative set of smart city ser-
vices. In particular, we consider three illustrative services that require the orchestration
of the sensing cloud and device layers.

122
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

6.2.2. Contributions

The main contributions of this chapter are the following:

We mathematically formulate the service distribution problem in IoT-Cloud net-

works as a minimum cost mixed-cast flow problem. This can be efficiently solved
via linear programming.

We solve the IoT-Cloud SDP for an illustrative set of smart city services that
allow users the consumption of augmented information resulting from the real-
time processing of live smart city streams.

6.2.3. Organization of the Chapter

The remaining of this chapter is organized as follows: Section 6.3, motivates IoT-Cloud
platforms in smart cities and its advantages in relation to centralized cloud strategies.
Section 6.4, introduces the system model and presents the concept of service graph.
Section 6.5, formulates the service distribution problem in IoT-Cloud platforms. Sec-
tion 6.6, shows the energy consumption of illustrative smart city services using the
IoT-Cloud SDP. Finally, Section 6.7 summarizes this chapter and presents the main
conclusions.

6.3. IoT-Cloud in Smart Cities

Smart cities are a clear example of the huge potential of the IoT. These connect many
wired and wireless devices to provide services that enhance the wellbeing of people (Fig-
ure 6.1). Smart cities aim to improve many city services, such as the system health mon-
itoring, traffic mobility, waste management, energy management and healthcare.

The general structure of smart cities includes a device layer, a transmission layer and an
application layer (Figure 6.2). The device layer provides raw data from sensing devices

123
 6.3. IoT-Cloud in Smart Cities

Cloud
Network
Health-Care,
Critical Infrastructures
Monitoring
Video Streams

Smart Devices

Public Transport,
Pollution, Traffic
Monitoring

Figure 6.1. Smart city architecture.

124
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

Application Layer:
Analyzes and processes input data
to generate useful information.
Elements: servers.

useful
raw data
information

Transport Layer:
Transports data and information
to / from end devices.
Elements: base stations, routers.

useful
raw data
information

Device Layer:
Generates raw data and requests
useful information.
Elements: RFID, sensors, smart
devices, vehicles.

Figure 6.2. Smart cities layers

and also requests processed information. In general, there are nodes that only provide
data (e.g., distributed sensors, cameras, RFID, GPS), nodes that only request infor-
mation (e.g., actuators, public screens), and nodes that may both provide and request
information (e.g., smart devices). The network layer transports both the raw data to
the cloud servers, and the processed data to the end devices. This layer combines wired
and wireless links, with different technologies. Finally, the application layer processes
and analyzes raw data to generate the information requested by the device layer.

Traditionally, a centralized cloud scheme in smart cities was considered to be the most
efficient solution, since the consolidation of processing resources in a single location re-
duces the operational cost of the network. However, recent works indicate that the

125
 6.4. System Model

network performance can be improved by doing more processing close to the end users
[Sat15], [Ha13]. In this context, [Sat09] and [Shi13] are initiatives to place distributed
near-user cloud nodes. Thanks to the recent advances on mobile computing [Lei13] and
virtualization [Nas14], the device layer can also be integrated into the cloud network.
Then, smart cities can be modeled as converged IoT-Cloud networks in order to effi-
ciently manage their resources, enhance their QoS, and reduce their energy consumption
[Bon12].

In this scenario, sensors, smartphones, connected vehicles, and the rest of mobile devices,
are not simple endpoints, but storing, sensing and computing resources. With this
level of abstraction, they can be orchestrated together, and also interact with the cloud
resources. Some of the advantages of the IoT-Cloud paradigm in smart cities are:

Low latency: the processing can be placed at the edge of the network in order to
support latency sensitive applications.

Geographical distribution: It allows widely distributed cloud systems, since the

device layer is an active part of the Cloud.

Large scale sensor networks: The interoperability of WSNs from different providers
creates larger virtualized sensing platforms.

Location aware and mobility support: The current location of users can be used
to provide service mobility.

In the following section, we introduce the system model considered in this chapter,
which captures the high heterogeneity of IoT-Cloud platforms in terms of services and
devices.

6.4. System Model

In this chapter, we represent an IoT-Cloud network as a graph. Note that we do not

model the network as we did in the previous chapters, since we must consider different

126
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

types of nodes (e.g., wireless sensors, smart devices, cloud servers) and links (e.g., ZigBee,
Wifi, 4G, optical). Moreover, we also need to capture the different classes of services in
“smart” environments, which may collect and/or distribute data. Then, using graph-
based representations the network can be mathematically formulated in a more formal
way, and therefore its optimization is simplified.

6.4.1. Network Model

We consider a network modeled as a directed graph G = (V, E) with V vertices and E

edges representing the set of network nodes and links, respectively. Each node is charac-
terized by its particular energy resources (e.g., power grid, battery), processing resources
(e.g., processor, microprocessor) and data acquisition resources (e.g., camera, sensors,
I/O interfaces). These are connected by wired and wireless links, which are character-
ized by their particular transport capacity. In particular, we denote with cpr sn
u and cu

the data processing and acquisition capacities (in bps) at node u ∈ V. Analogously, we
denote with epr sn
u and eu the data processing and acquisition efficiencies (in Watts per

bit) at node u ∈ V. Finally, ctr tr

u and evu denote the capacity (in bps), and the efficiency

(in Watts per bit) of the link (v, u) ∈ E. Besides the particularities of the physical links,
these also consider the transmitter and receiver efficiencies and capacities.

6.4.2. Service Model

We denote with O the set of information objects or flows that can be processed, captured,
or transported over the network, and with P the set of virtual functions that can be
implemented. An information object o is, in general, the output of a function po ∈ P
that requires the set of objects Z(o) ⊂ O as input. For example, in a personalized
video streaming application, the processed video stream o results from the combination
of multiple video streams (i.e., Z(o)) via the video processing function po .

We represent a cloud service φ by a rooted tree Tφ = (Aφ , Oφ ), referred to as the service

127
 6.4. System Model

o3 o5
p3
o1 o2
p1
p2

o4
Figure 6.3. An example of a service graph, Tφ = (Aφ , Oφ ), with |Oφ | = 5 objects, |Aφ | = 4 edges, and
|P| = 3 virtual functions.

tree. For any node o ∈ Oφ , there is a directed edge (z, o) ∈ Aφ for all z ∈ Z(o), as
shown in Figure 6.3. Hence, the set of objects Z(o) ⊂ O required to generate object o
via function po are represented as the children of o in the service tree Tφ . In particular,
the root of the service tree rφ ∈ O represents the final information object that needs
to be delivered to the end user(s). Note that rφ has no outgoing links in Tφ , as it is
not the input to any other object for cloud service φ, and the leaves of the rooted tree
have no incoming links, as they represent source information that cannot be created by
a cloud service function. The service tree hence encodes the relationship between all the
objects necessary to deliver the final product to the end user(s) via the required cloud
functions. It is important to note that each object o is uniquely characterized by the
pair (po , Z(o)). This allows different objects to share the same input, but be created via
different functions, or to share the same function, but have different inputs.

6.4.3. Service Requirements

When a user requests a cloud service φ, the user is, in essence, requesting the final
information object or flow represented by the root of the service graph, rφ ∈ R. We
denote with Bo the bitrate (in bps) associated with content object o ∈ O.

In terms of QoS, we denote with hvu and rvu the transport delay (in seconds) and relia-
bility (in terms of packet delivery ratio) associated with the link (v, u) ∈ E, respectively,

128
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

and with hu the processing delay at cloud node u ∈ V. We use Hd,o to denote the
maximum delay allowed for the delivery of content object o ∈ O at destination d ∈ V.
Moreover, Eu (in Watts) stands for the maximum power consumption of node u ∈ V,
taking into account its expected lifetime. Note that Eu may be ∞ for non battery
powered nodes.

Once introduced the system model, in the following section we formulate the IoT-Cloud
SDP.

6.5. The IoT-Cloud Service Distribution Problem

The IoT-Cloud SDP is formulated as a minimum cost mixed-cast flow problem. This is
characterized by the following flow variables:

User-object flows: are characterized by a triplet (d, o, z), which indicates that the
given flow is carrying information of object z ∈ O, used to deliver final product
tr,d,o,z sn,d,o,z
o ∈ O at destination d ∈ V. In particular, fvu , fvu , fupri ,d,o,z and fupro ,d,o,z
are binary variables that indicate whether object z is carried (i.e., tr), captured
(i.e., sn) or processed (i.e., pr) by link (v, u) ∈ E or node u ∈ V for final product
o ∈ O at destination d ∈ V. Note that we differentiate between fupri ,d,o,z and
f pro ,d,o,z , which denote the input and output flows of the processing unit at node
u ∈ V associated with triplet (d, o, z). Figure 6.4 illustrates the network flows
associated with a given triplet (d, o, z) at node u ∈ V, where pr represents the
processing unit that hosts virtual functions and qud,o,z is a binary demand parameter
that indicates if node u ∈ V requests object z ∈ O. Note that qud,o,z = 0 if u 6= d or
z 6= o, since users only request final information objects for themselves.

tr , f sn and f pr , determine the total amount of information flow

Global flows: fvu u u

carried, captured or processed at a given physical node/link, respectively.

129
 6.5. The IoT-Cloud Service Distribution Problem

fusn, d, o, z qud, o, z

fwr, d, , z u fuvtr, d, o, z

fupro, d, o, z fupri , d, o, z

pr

Figure 6.4. Generalized flow conservation at node u ∈ V, where pr represents the processing unit that
hosts virtual functions at node u.

Objective Function

We define a generic cost associated with the capturing (kusn ), processing (kupr ), and trans-
tr ) of traffic over the physical resources of the network. For instance, these may
port (kvu
be defined in terms of capital cost ($/bit), energy consumption (W/bit), environmental
impact (CO2 /bit), or a combination of multiple costs. Then, the objective function is
characterized by the global flows and determines the total cost of the network:

X
tr tr


minimize kvu fvu +
(v,u)∈E
X
(kusn fusn + kupr fupr ) . (6.1)
u∈V

Next, we include the problem constraints, which assure the conservation of the total flow,
capture the capabilities and capacities of nodes/links, model the unicast and multicast
traffic, and define the QoS requirements.

130
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

Generalized Flow Conservation Constraints

User-object flows must satisfy demand and flow conservation. By modeling the demand
qud,o,z as part of the outgoing flow of node u, we can use the following generalized flow
conservation constraints:

X
qud,o,z + tr,d,o,z
fuw + fupri ,d,o,z =
w∈N + (u)
X
tr,d,o,z
fvu + fusn,d,o,z + fupro ,d,o,z ∀u, d, o, z. (6.2)
v∈N − (u)

Flow conservation constraints state that the outgoing flow associated with a given triplet
(d, o, z) must be equal to the incoming flow for that same triplet, for any node u. As
illustrated in Figure 6.4, the outgoing flow is composed of the outgoing transport flows,
the processing flow leaving node u towards the processing unit, and the demand flow;
while the incoming flow is composed of the incoming transport flows, the capturing flow,
and the processing flow going out of the processing unit.

In addition, each processing unit must satisfy the following flow conservation con-
straint:

fupro ,d,o,z ≤ fupri ,d,o,y ∀u, d, o, z, y ∈ Z(z). (6.3)

This constraint makes sure that in order to have a processed flow z for demand (d, o),
a flow associated with each of the input objects required to generate z, y ∈ Z(z), for
demand (d, o), must be present at the input of the processing unit.

131
 6.5. The IoT-Cloud Service Distribution Problem

Function Availability Constraints

These constraints allow restricting the set of virtual functions that can be implemented
at a given node:

fupro ,d,o,z = 0 ∀u, d, o, z, pz ∈

/ Pu , (6.4)

where Pu ⊂ P is the set of virtual functions available at node u.

Source Constraints

We denote the set of objects that are available in the network as source information and
hence input for cloud services as S ⊂ O. We define the set Ou ⊂ S as the objects that
are captured by node u:

fusn,d,o,z = 0 ∀u, d, o, z ∈
/ Ou . (6.5)

(6.6)

In addition, we need to make sure that source objects S ⊂ O are not created in the
network:

fupro ,d,o,z = 0 ∀u, d, o, z ∈ S. (6.7)

Mixed-cast Constraints

Mixed-cast constraints allow modeling the unicast or multicast nature of flows via the
corresponding relationship between user-object and global flows. Since a single captured
object can be used to satisfy multiple demands, capturing flows are said to be multicast.
Hence, user-object capturing flows for the same object, but for different destinations,

132
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

are allowed to overlap:

fusn,d,o,z ≤ fusn,z ∀u, d, o, z, (6.8)

X
fusn,z Bz = fusn ∀u. (6.9)
z∈O

where the user-object flows are weighted by the size of the object Bz .

On the other hand, the transport and processing flows may be unicast or multicast. If
these are multicast:

tr,d,o,z tr,z
fvu ≤ fvu ∀u, d, o, z, (6.10)
tr0
X
tr,z
fvu Bz = fvu ∀u, (6.11)
z∈O

fupro ,d,o,z ≤ fupr,z ∀u, d, o, z, (6.12)

X
fupr,z Bz γz = fupr ∀u. (6.13)
z∈O

where output processing flows are scaled by a factor γz that captures the changes (in
bps) associated with the generation of object z from its input Z(z) via function pz .

If transport and processing are unicast, user-object flows cannot overlap and must be
added across both objects and destinations:

0
XXX
tr,d,o,z tr
fvu Bz = fvu ∀(v, u), (6.14)
d∈V o∈O z∈O
XXX
fupro ,d,o,z Bz γz = fupr ∀u. (6.15)
d∈V o∈O z∈O

0
tr instead of f tr to denote that the global transport flow in (6.11)
Note that we use fvu vu

and (6.14) does not include the possible packet retransmissions. These are considered in
(6.20).

133
 6.5. The IoT-Cloud Service Distribution Problem

QoS Constraints

In this section, we formulate the constraints that define the QoS requirements in terms
of latency, reliability and energy consumption.

Latency

We use δd,o,z to denote the (local) delay associated with a particular user-object flow
(d, o, z). This is computed as the weighted sum of the delay associated with the transport
and processing of (d, o, z) flows:

X X
tr,d,o,z
δd,o,z = fvu hvu + f pro ,d,o,z hu ∀d, o, z. (6.16)
(v,u)∈E u∈V

In (6.16), without loss of generality, we assume that input processing flows have zero
delay and capture all processing delay with the output processing flows. We then use
ag
δd,o,z to denote the aggregate delay associated with user-object flow (d, o, z), computed
as the sum of the local delay, δd,o,z , plus the maximum across the aggregate delays of all
ag
of its input flows δd,o,y , ∀y ∈ Z(z), as:

ag
δd,o,z = δd,o,z ∀d, o, z ∈ S, (6.17)
ag ag
δd,o,z + δd,o,y ≤ δd,o,z ∀d, o, z, y ∈ Z(z). (6.18)

Finally, the aggregate delay associated with the delivery of final object o at destination
d is constrained according to the service requirements:

ag
δd,o,o ≤ Hd,o ∀d, o. (6.19)

134
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

Reliability

The reliability of links has a relevant impact on the traffic flow due to the retransmissions
caused by packet losses, particularly in low power wireless links. The average number
of retransmissions required in link (v, u) ∈ E is modeled as the reciprocal of its packet
delivery ratio:

tr tr 0
fvu = fvu /rvu ∀(v, u). (6.20)

The value of rvu is between zero (i.e., disconnected nodes) and 1 (i.e., ideal link), and
therefore it increases the actual transport flow.

Energy Consumption

The total energy consumption due to transport, processing and acquisition at node u
must be lower than its maximum energy consumption:

X X
tr tx tr rx
fuw eu + fvu eu +
(u,w)∈E (v,u)∈E

fusn esn pr pr
u + fu eu ≤ Eu ∀u. (6.21)

tr with etx and erx , which

Note that we separate the impact of the transport flow fvu v u

denote the transmission and reception efficiency (in Watts per bit) of nodes u and v,
respectively.

135
 6.6. Simulation Results

Capacity Constraints

Global flows must satisfy capacity constraints, as:

tr
fvu ≤ ctr
vu ∀(v, u), (6.22)

fusn ≤ csn
u ∀u, (6.23)

fupr ≤ cpr
u ∀u. (6.24)

Integer/Fractional Flow Constraints

User-object flows can either be binary or fractional. Accordingly, these can be defined
as:

tr,d,o,z
fvu , fupri ,d,o,z , fupro ,d,o,z , fusn,d,o,z ∈ {0, 1} ∀(v, u), u, d, o, z, (6.25)

(6.26)

or

tr,d,o,z
fvu , fupri ,d,o,z , fupro ,d,o,z , fusn,d,o,z ∈ [0, 1] ∀(v, u), u, d, o, z. (6.27)

As shown in [Llo13], binary flow variables in mixed-cast flow problems can be relaxed
if: a) all flows are unicast, b) the network topology is a tree, or c) the network can
perform intra-session network coding (e.g., random linear coding). Then, the polynomial
complexity of the IoT-Cloud SDP may be reduced in these practical cases.

6.6. Simulation Results

In this section, we solve the IoT-Cloud SDP to minimize the overall network consumption
of three illustrative smart city services via the linear programming solver Xpress-MP. In

136
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

HO

IO IO

EO EO EO

Smart Devices

WSN1 WSN2 WSN3

Virtualized WSN

Figure 6.5. Structure of the smart city.

particular, we analyze: i) the virtual fog service network solution (vFSN), which allows
tasks to be processed anywhere in the network, ii) a distributed cloud approach, which
places tasks in any cloud node, and iii) a fully distributed approach, which allocates all
the processing at the devices level.

6.6.1. Smart City Scenario

We consider the smart city network in Figure 6.5 that includes wireless sensors, smart
devices (e.g., smartphones, tablets, smart glasses), base stations and cloud nodes. We
assume a hierarchical cloud infrastructure that is composed of end offices (EOs), inter-
mediate offices (IOs) and a Head Office (HO) [Bar15]. Base stations are equipped with
a network gateway and also with a low complexity cloud instance, such as a cloudlet
[Sat09]. Three different WSNs collect environmental information. Each WSN is com-
posed of 25 sensors, uniformly distributed over a sensing area of 250x250 m2 .

Precisely modeling the processing capacity (in MIPS) and efficiency (in MIPS/W) of

137
 6.6. Simulation Results

the smart city nodes is difficult, due to the limited information disclosed by network
operators. Since this is out of the scope of this chapter, we compute approximate values
using information extracted from [Vis15] (i.e., cloud equipment), [Alt15] (i.e., smart
devices), and [Xia10] (i.e., wireless sensors), which are presented in Tables 6.1 and 6.2.
Note that we assume that the efficiency of cloud nodes increases at higher layers, thanks
to the consolidation of processing tasks. Since the capacity and efficiency of smart
devices vary considerably depending on the kind of device, we consider different realistic
values in the simulations.

6.6.2. Simulation Details

The processing complexity of tasks (in instructions per bit), depends on the particular
application. In this section, we assume that simple tasks, such as data aggregation, have
a complexity of 100 instructions per bit. On the other hand, the tasks that require a
more complex data analysis require a higher number of instructions per bit (i.e., 500-5000
in our simulations).

We assume that the statistics of the wireless channel are known. This is time-varying
and follows the two-slope log-distance path loss model described in [Kyö07] for urban
micro-cell scenarios. Then, the path losses (P L) are modeled as follows:

P L(dB) = P L0 (dB) + 20 − 12.5α + (6.28)

10α log(d) + 3 log(fc /5) + γ,

where P L0 is the path loss at the reference distance d0 (1 m), α is the path loss exponent,
d is the communication distance, fc is the system frequency (in GHz) and γ ∼ N 0, σγ2

is zero mean Gaussian noise that models the shadowing effects. This model assumes
α=2.8 and σ=4 dB. Moreover, a collision-free MAC layer is assumed to obtain results
independent from the specific MAC mechanism. Please note that in this chapter we
have considered a slightly different path loss model than in previous chapters to better

138
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

model smart city scenarios.

Taking into account the general characteristics of ZigBee transceivers, we assume fc =2.4
GHz, P L0 (dB)=35 dB [MS05], a transmit power of 3 dBm, a sensitivity value of -91
dBm [Atm09] and omnidirectional antennas. We also consider that the packets gener-
ated at wireless sensors have a length of 127 bytes (i.e., standard packet size in IEEE
802.15.4).

The maximum energy consumption of wireless sensors is limited by their expected bat-
tery duration. Note that the replacement of their batteries is a costly operation and
therefore, nodes should not deplete their batteries before the scheduled maintenance.
With this in mind, in the simulations we assume a minimum lifetime of each sensor of 1
year. This is introduced in (6.21) of the IoT-Cloud SDP formulation, taking into account
that they are equipped with a battery that provides 27 kJ [Siv10].

The maximum latency may be as low as 10 ms in latency sensitive applications, such as

cognitive assistant services [Sat15]. On one hand, strict delay requirements pushes the
processing closer to the end user. On the other hand, using the most powerful nodes and
the fastest links the computation and transmission times can be reduced, and these are
generally closer to the network core. In these simulations, we narrow the set of feasible
solutions of the IoT-Cloud SDP using (6.19) taking into account a maximum end-to-end
delay of 10 ms.

The range of applications that smart cities can provide is very wide. In this section, we
focus on the following examples: augmented reality, autonomous sensing and actuation,
and city monitoring.

6.6.3. Augmented Reality Service

The real world and a computer generated view are combined in augmented reality ser-
vices to create the perception of a new reality. Recent products providing augmented
reality are Google Glass, Sony SmartEyeglass and Microsoft HoloLens. In the con-

139
 6.6. Simulation Results

Table 6.1. Capacities and efficiencies of fog nodes

Capacity Efficiency
Cloud Node (HO) 53.5 Million MIPS 500 MIPS/W
Cloud Node (IO) 26 Million MIPS 200 MIPS/W
Cloud Node (EO) 13 Million MIPS 133 MIPS/W
Cloud Node (BS) 6.5 Million MIPS 100 MIPS/W
Wireless Sensor 1 MIPS 4167 MIPS/W

Table 6.2. Capacities and efficiencies of links

Capacity Efficiency
Optical 4480 Gbps 12.6 nJ/bit
4G (Down/Up) 72/12 Mbps 76.2/19 µJ/bit
WiFi 150 Mbps 300 nJ/bit
ZigBee 250 kbps 100 nJ/bit

text of smart cities, the real-time video stream captured by a smart device may be
combined with an informative view of the city. In general, smartphones contain the
necessary elements to implement augmented reality services (i.e., internet connexion,
sensors, camera). However, the high computational requirements of these applications
and the limited battery resources of smartphones motivates the network operators to
consider offloading approaches [Alt15]. In this context, fog computing can extend the
capabilities of end devices, without introducing the high latency of centralized cloud
approaches.

In this section, we consider a scenario in which end users request augmented video
streams using their smart devices. This is the combination of a real-time video stream,
captured with the camera of the smart device, personalized information, downloaded
from the Internet, and data from the city gathered by the wireless sensors. We assume
that 2000 instructions per bit are required to generate the augmented video. Half of
users access the network using 4G and the other half use a WiFi connexion. Note that
the video captured and the augmented video must be send in unicast, but the data from
wireless sensors can be sent in multicast to avoid duplicate packets.

In Figure 6.6, we compare the average power consumption for different processing effi-

140
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

18
Smart device
Cloud
16 vFSN
Average Power Consumption per User (W)

14

12

10

0
500 300 200 100 50
Processing Efficiency (MIPS/W)

Figure 6.6. Average power consumption (W) of an augmented reality application for different smart
device efficiencies (MIPS/W).

ciencies. These are considered to capture the multiple kinds of end devices sharing the
smart city infrastructure. On average, end users request an augmented reality video that
results from the combination of 200 kbps of video, 50 kbps of personalized information
and 1 packet/s from each wireless sensor. The simulation results show that vFSN ad-
justs the offloading decisions according to the processing efficiency of end devices. The
processing tends to be allocated at the devices layer if their efficiency is high enough,
while low efficient devices offload their processing to the cloud servers. In general, users
accessing the network using WiFi, tend to use the cloud resources frequently. However,
the access cost of 4G networks, push the processing to the end devices, as observed also
in [Vis15].

In Figure 6.7, the overall power consumption of the network for multiple augmented
video bitrates is shown. These represent different video qualities defined according to
the users’ preferences. We assume that the end devices are smartphones with a processing
efficiency of 1000 MIPS/W and 4.5 Wh. We constrained the solutions of the IoT-SDP

141
 6.6. Simulation Results

35
Smart device
Cloud
vFSN
30

Average Power Consumption per User (W)

25

20

15

10

0
50 150 300 500
Video Bitrate (kbps)

Figure 6.7. Average power consumption (W) of an augmented reality application for different video
qualities (kbps). Smartphones must provide at least 5 hours of service.

to support at least 5 hours of augmented video service before depleting the smartphone
batteries. In this figure, we observe that the video is processed by the smartphones,
even with low video bitrates. This is due to the higher energy efficiency of smartphones,
compared to the efficiency of cloud servers. The difference in terms of consumption
between local processing and cloud processing increases with the bitrate. Besides the
additional processing cost, note that to process the video at the Cloud, this must be
transported from the smartphone to a cloud node and then transported back to the
smartphone, once the additional information has been included. However, due to the
limited energy resources of smartphones, users demanding an augmented video with
high quality (e.g., 500 kbps) must offload their processing to the Cloud. Otherwise,
these cannot provide the minimum hours of service.

142
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

6.6.4. Autonomous Sensing and Actuation Service

Wireless sensors and actuators have been traditionally application specific. However,
considering them as fog nodes, their tasks in the network can be adapted to the service
requirements, increasing or reducing their activity dynamically.

In this section, we consider a WSN that monitors a particular phenomenon in the city
and activates the actuators accordingly. There are 5 of these sensors that are also actu-
ators. Nearby wireless sensors may combine their measurements using data aggregation
techniques. This reduces traffic by a compression rate that depends on the spatial and
temporal correlation of data. A realistic compression rate of 50% [Cap12] is assumed in
this simulations. Based on these measurements, a simple decision of turning on/off each
actuator is taken.

In Figure 6.8, we compare the power consumption for different packet generation rates
at wireless sensors. Note that a high rate increases the accuracy of the service, but also
increases the energy consumption of wireless sensors. Then, in a real application this
should be defined according to the service requirements. As we can observe, the sensing
platform can manage the actuators without the assistance of the cloud infrastructure as
long as the packet rate is low (i.e., lower than 2 packets per second). However, at higher
rates this processing has to be offloaded to the Cloud in order to keep the nodes alive
for at least 1 year without the replacement of their batteries.

In Figure 6.9, we consider different processing complexities, representing different kinds

of data aggregation. Similarly to what has been observed in Figure 6.8, for low aggre-
gation complexities the wireless sensors process everything themselves to avoid using
the cloud resources. However, this is not possible if the processing cost is higher than
a certain value (i.e., 70 instructions per bit). Then, this has to be placed at the cloud
layer.

In both cases, the vFSN solution places the processing at the sensing layer as long as it
is possible to take advantage of its higher efficiency. In case that the packet rate or the

143
 6.6. Simulation Results

0.4
Wireless Sensor
Cloud
0.35 vFSN

0.3
Average Power Consumption (W)

0.25

0.2

0.15

0.1

0.05

0
0.5 1 1.5 2 5
Packets/s

Figure 6.8. Average power consumption (W) of an autonomous sensing actuation platform for different
packet generation rates (packets/s).

processing complexity increase above a certain threshold, this is automatically placed at

the cloud layer.

6.6.5. City Monitoring Service

Thanks to the deployment of large scale wireless sensor networks, it is possible to pro-
vide real-time information about the city. In this example, users request personalized
information that is generated from the analysis of the data gathered by three different
WSNs. For instance, in a mobility assistant service, these WSNs could provide the speed
of vehicles, pollution levels and congestion at the city access, respectively. Then, differ-
ent mobility options could be obtained from this data (e.g., the fastest route, the most
eco-friendly route or the best public transport alternative). Note that the measurements
from wireless sensors can be sent in multicast, since they can be used to generate mul-
tiple information objects. However, the personalized information can only be sent in
multicast for these users requesting the same information at the same time. Otherwise,

144
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

0.09
Wireless Sensor
Cloud
0.08 vFSN

0.07
Average Power Consumption (W)

0.06

0.05

0.04

0.03

0.02

0.01

0
10 50 70 80 100
Data Aggregation Complexity (Instructions/bit)

Figure 6.9. Average power consumption (W) of an autonomous sensing actuation platform for different
data aggregation complexities (instructions/bit).

this must be sent in unicast.

We assume smart devices with a processing efficiency of 1000 MIPS/J. The processing
complexity of analyzing the data from each WSN is 500 instructions per bit, while the
complexity of generating the personalized information is 5000 instructions per bit. An
average rate of 1 packet per second is generated by each wireless sensor.

In Figure 6.10 and Figure 6.11, we compare the power consumption per user for different
average number of users requesting simultaneously the same information using WiFi and
4G, respectively. The bitrate of this information is assumed to be 200 kbps. Note that if
this is processed and transported in multicast through the cloud layer, the processing of
this information can be consolidated in a specific cloud location. Nevertheless, we assume
that users receive the information in unicast from the base stations, since 4G multicast
is not available commercially yet. In Figure 6.10, we observe that if there are more than
2 simultaneous WiFi users, it is more efficient to consolidate the processing at a cloud
location than processing it locally, even if the transport cost of the requested information

145
 6.6. Simulation Results

1
Smart device
0.9 Cloud
vFSN

0.8
Average Power Consumption per User (W)
0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4
Simultaneous Wifi Users

Figure 6.10. Average power consumption (W) of a city monitoring application for different WiFi users
requesting the same information at the same time.

is higher. However, for users accessing the network using 4G (Figure 6.11), the additional
transport cost, which is much higher in 4G networks, cannot be compensated with
a relatively low number of users. Note that this cost could be reduced using energy
efficient 4G multicast streaming schemes [Yoo14].

In Figure 6.12, we compare the average power consumption for different bitrates of
personalized information. A high bitrate represents applications providing detailed in-
formation, while low bitrates indicate that the application is only providing the most
relevant information. We consider the same number of WiFi and 4G users. The same
information is simultaneously requested by 4 users on average. As we can observe, if
users request a low bitrate, the processing tends to be placed at the cloud layer to reduce
the processing cost by consolidating it in a single location. However, the cloud approach
becomes less efficient at high bitrates due to the transport cost from the Cloud to the
smart devices, particularly to 4G devices. Then, vFSN moves the processing closer
to the users in order to reduce this cost, at the expenses of reducing the processing

146
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

18
Smart device
Cloud
16 vFSN
Average Power Consumption per User (W)

14

12

10

0
1 2 4 8
Simultaneous 4G Users

Figure 6.11. Average power consumption (W) of a city monitoring application for different 4G users
requesting the same information at the same time.

consolidation.

6.7. Summary and Conclusions

The future internet will combine the Internet of Things (IoT) and the Cloud. In this
new paradigm, wireless sensors become cloud ready infrastructures and therefore, they
can be orchestrated together with the rest of the IoT.

In this chapter, we formulate the service distribution problem (SDP) in IoT-Cloud net-
works, referred here to as the IoT-Cloud SDP. This finds the optimal placement of
virtualized functions over the network, taking into account the heterogeneous capaci-
ties and limitations of end devices. We have implemented the IoT-Cloud SDP in three
illustrative smart city services: augmented reality, autonomous sensing and actuation,
and city monitoring. In the first example, the simulation results show that vFSN ad-
justs the offloading decisions according to the processing efficiencies of end devices. In

147
 6.7. Summary and Conclusions

12
Smart device
Cloud
vFSN
10

Average Power Consumption per User (W)

0
10 20 50 100 300
Personalized Information Rate (kbps)

Figure 6.12. Average power consumption (W) of a city monitoring application for different personalized
information rates (kbps).

particular, the augmented reality video of highly efficient devices is processed locally,
to reduce latency and traffic going through the network, while the processing of low
efficient devices tend to be offloaded to the Cloud. In the second example, we have
observed that vFSN moves the processing of the information from wireless sensors with
a high data generation rate to the Cloud in order to extend their lifetime. On the other
hand, the processing is placed at the sensing platform when this rate and the processing
complexity are low. In the third example, we have shown that vFSN takes the offload-
ing decisions of a city monitoring application according to the personalized information
bitrate and the number of simultaneous users requesting the same information. While a
high bitrate moves the processing closer to the end user in order to reduce its transport
cost, a high popularity pushes the processing back to the Cloud in order to consolidate
the processing tasks.

In general, we have shown that the IoT-Cloud SDP captures the tradeoffs that appear
in IoT-Cloud platforms due to the heterogeneity of services, network technologies and

148
6. IoT-Cloud Formulation to Integrate WSNs into the Future Internet of
Things

devices. Then, using a fully virtualized approach, individual placement decisions can be
taken in order to minimize the energy consumption of these networks, and hence reduce
their cost and environmental impact.

149
 Chapter 7

Conclusions and Future Work

In this thesis, we have addressed several challenges that are expected to be critical in
the integration of WSNs into the future Internet of Things. In Chapter 3, we have
enhanced RPL with a pragmatic channel allocation and transmission power control.
Both techniques are expected to be crucial in future WSNs, due to the increasing density
of wireless sensors in many scenarios, such as smart cities. In Chapter 4, we have designed
a robust routing approach that deals with the challenges originated by the mobility of
nodes. Note that with the reduction of the size of wireless sensors, they are going to be
easily attached to mobile entities for monitoring, localization or tracking applications. In
Chapter 5, we have addressed the traffic heterogeneity problem with a cooperative multi-
tree routing strategy. The increasing complexity of applications and the collaboration
among different WSNs originates independent traffic flows with individual requirements.
Finally, in Chapter 6 we have formulated the integration of wireless sensors into Cloud-
based IoT platforms. As a result, wireless sensors can extend their hardware limitations
and share their measurements in a bigger scale.

7.1. Conclusions

First of all, in Chapter 1 we have discussed the main motivations and presented the
outline of this thesis. We have also introduced the main research contributions, which

151
 7.1. Conclusions

are contained in Chapters 3-6.

In Chapter 2, we have presented the background for the following chapters. We have
first introduced gradient-based routing, with emphasis on the standardized RPL proto-
col, since this is the core of the proposed solutions in Chapters 3-5. Second, we have
introduced the multi-channel and multi-power mechanisms for WSNs from a pragmatic
perspective. Third, we have explained position-based routing and the impact of realistic
hardware and channel conditions on its reliability. Fourth, we have presented the prob-
lem of managing heterogeneous traffic in a single WSN, and a general classification of the
most important traffics in WSNs. Finally, we have motivated the transition from cloud
computing to fog computing, its most relevant advantages, and also some challenges that
need to be addressed to enable this paradigm shift.

In Chapter 3, we have proposed a pragmatic joint routing, channel allocation and trans-
mission power control solution to enable high density wireless sensor deployments. This
provides multi-channel and multi-power capabilities to RPL that can be adjusted accord-
ing to the channel conditions and the QoS requirements. We have shown that it efficiently
reduces the energy consumption and the packet collisions of the network, without adding
significant complexity to RPL. Two different routing metrics are proposed, referred to as
MinAP and MaxPDR, which combine reliability and energy efficiency criteria. MinAP is
more suitable for applications with tight energy consumption constraints, and MaxPDR
is recommended for applications that demand a high network reliability. These have
been compared through simulations with RPL and also with LEACH, GBR and the
gradient-based routing protocol adopted in ZigBee. These simulations have shown that
the proposed extension increases the reliability, and reduces the aggregated power and
the collision probability of RPL. It is also worth noting that MaxPDR can be adapted
using a configurable parameter β to deal with the reliability vs. energy consumption
tradeoff. Moreover, the proposed strategies have also been validated in a WSN testbed
with commercial motes. The experimental results have confirmed their implementability
and also the promising results obtained through simulations.

152
7. Conclusions and Future Work

In Chapter 4, we have addressed node mobility in WSNs with KP-RPL, which further
extends RPL enabling it to cope with mobile nodes. This combines RPL with a robust,
yet simple positioning algorithm, based on the construction of confidence regions and
Kalman filtering. Moreover, blacklisting is applied to avoid disconnections due to posi-
tioning errors. Thanks to these additional algorithms, KP-RPL reduces the impact of
the positioning inaccuracies that are introduced by the RSSI measurements and velocity
estimates. Thus, mobile nodes can make better routing decisions. We have evaluated
KP-RPL in different network conditions and trajectories of the mobile nodes, observ-
ing that it increases the reliability of the network, compared to geographical routing.
Moreover, the impact of bad channel conditions in KP-RPL is significantly lower. We
have observed that blacklisting increases the network reliability, but it also increases
the average ETX of the network, since this is a conservative strategy. This generates a
tradeoff that can be managed with the size of the confidence regions, where the ETX is
improved using smaller areas, and larger areas increase the PDR. We have also evaluated
KP-RPL in scenarios where anchor nodes are not continuously broadcasting positioning
packets. In this situation, KP-RPL is particularly efficient because it needs a lower num-
ber of positioning nodes than geographical routing thanks to the Kalman filer. Then,
the activity of anchor nodes can be reduced in order to increase their lifetime. Moreover,
the lower ETX achieved by KP-RPL reduces the energy consumption of mobile nodes
for transmitting data packets. Therefore, KP-RPL extends the overall lifetime of the
network.

In Chapter 5, we have extended RPL with C-RPL to deal with the QoS requirements of
multiple traffic flows. This creates energy efficient instances, referred to as C-RPL In-
stances, following a cooperative strategy among nodes with different sensing tasks. Since
nodes would never agree to cooperate, C-RPL coordinates them to create coalitions that
improve the tradeoff between their own performance metric and energy consumption,
using a cooperation parameter ρ. The simulation results show that C-RPL efficiently
creates C-RPL Instances according to their individual objective functions and the par-
ticular network conditions. In particular, we have observed that C-RPL tends to create

153
 7.1. Conclusions

large instances when the node density is low. Otherwise, it constructs smaller instances
to reduce the energy consumption of the network and also to avoid congestion problems.
Besides, we have proposed a fairness analysis for networks with multiple instances that
measures the distribution of the network resources when multiple objective functions
coexist in the network. The results show that C-RPL obtains a higher fairness degree
than RPL approaches with non-cooperative instances.

Finally, in Chapter 6 we have formulated the IoT-Cloud SDP to efficiently solve the
service distribution problem in Cloud-based IoT platforms. This considers the novel
paradigm of fog computing, which extends the Cloud up to the device layer. The IoT-
Cloud SDP orchestrates wireless sensors together with the rest of devices in the IoT to
find the placement of virtual functions across the network that minimizes the overall
network cost, meeting users’ requests and satisfying the network resource capabilities.
We have solved the IoT-Cloud SDP for three illustrative smart city services: augmented
reality, autonomous sensing and actuation, and city monitoring. The results show that
the task offloading is adjusted according to the energy efficiency of end devices. Highly
efficient devices can reduce the communication costs processing tasks locally, while low
efficient devices tend to offload their tasks to not waste battery resources. The generation
rate of wireless sensors also has an impact on the solution, since it introduces traffic in
the network that needs to be processed somewhere. This is processed by the wireless
sensors as long as their lifetime constraints are satisfied, but it has to be processed at the
Cloud otherwise. On the other hand, a high bitrate of personalized information moves
the processing closer to the end user in order to reduce its transport cost. However,
highly popular information tends to be processed at the Cloud in order to consolidate
the processing tasks at a single location.

154
7. Conclusions and Future Work

7.2. Future Work

In general, the contributions presented in this thesis can be further developed as fol-
lows:

Considering state-of-the-art low-power MAC protocols in our simulations.

Validating the simulation results obtained using Matlab with network simulators,
such as ns-3.

Extending the experimental results with additional scenarios and larger deploy-
ments.

Combining the RPL-based techniques proposed in Chapters 3 to 5.

Moreover, there are some topics derived from the research presented in each chapter that
remain as future work.

In Chapter 3, it would be interesting to explore the combination of other multi-channel

and multi-power mechanisms with a low signalling cost and RPL. Moreover, the col-
lision problem among sensors could be experimentally studied using a larger WSN
testbed.

In Chapter 4, KP-RPL has been studied for mobile nodes following a predefined tra-
jectory, and it could be interesting to observe its performance in scenarios with un-
predictable moving patterns. In addition, an experimental evaluation of KP-RPL in a
commercial WSN testbed has not been provided yet.

In Chapter 5, it would be interesting to study a more distributed C-RPL that applies the
cooperation game among nodes instead of groups of nodes. This would reduce its sig-
nalling cost and also increase the granularity of its solutions. Moreover, an experimental
comparison among RPL and C-RPL would also be very significant.

In Chapter 6, a more precise energy efficiency model of cloud servers, smart devices and
wireless sensors would improve the accuracy of the simulation results. In addition, the

155
 7.2. Future Work

IoT-Cloud SDP could also be applied to other smart environments, such as smart homes,
smart grids and smart transportation systems.

156
 Bibliography

[Aaz14] Aazam, M. and Huh, E.N. “Fog computing and smart gateway based com-
munication for cloud of things”. In “Future Internet of Things and Cloud
(FiCloud), 2014 International Conference on”, pages 464–470. Aug 2014. doi:
10.1109/FiCloud.2014.83.
[Ahm14] Ahmed, M.; Cui, H.; and Huang, X. “Smart integration of cloud computing
and mcmc based secured wsn to monitor environment”. In “Wireless Commu-
nications, Vehicular Technology, Information Theory and Aerospace Electronic
Systems (VITAE), 2014 4th International Conference on”, pages 1–5. May 2014.
doi:10.1109/VITAE.2014.6934449.
[Ahu93] Ahuja, R.K.; Magnanti, T.L.; and Orlin, J.B. Network Flows: Theory, Algo-
rithms, and Applications. Prentice-Hall, Inc., Upper Saddle River, NJ, USA,
1993. ISBN 0-13-617549-X.
[Ala13] Alamri, A.; Ansari, W.S.; Hassan, M.M.; Hossain, M.S.; Alelaiwi, A.; and
Hossain, M.A. “A survey on sensor-cloud: Architecture, applications, and ap-
proaches”. International Journal of Distributed Sensor Networks, 2013(917923),
2013.
[Alt15] Altamimi, M.; Abdrabou, A.; Naik, S.; and Nayak, A. “Energy cost mod-
els of smartphones for task offloading to the cloud”. Emerging Topics in
Computing, IEEE Transactions on, PP(99):1–1, 2015. ISSN 2168-6750. doi:
10.1109/TETC.2014.2387752.
[Alw13] Alwan, H. and Agarwal, A. “MQoSR: A multiobjective qos routing protocol
for wireless sensor networks”. ISRN Sensor Networks, 2013:12, 2013.
[Ari13] Arifuzzaman, M.; Matsumoto, M.; and Sato, T. “An intelligent hybrid MAC
with traffic-differentiation-based QoS for wireless sensor networksqos for wireless
sensor networks”. Sensors Journal, IEEE, 13(6):2391–2399, June 2013. ISSN
1530-437X. doi:10.1109/JSEN.2013.2252163.
[Arj15] Arjun, D.; Bala, A.; Dwarakanath, V.; Sampada, K.; Prahlada Rao, B.; and
Pasupuleti, H. “Integrating cloud-wsn to analyze weather data and notify

157
 BIBLIOGRAPHY

saas user alerts during weather disasters”. In “Advance Computing Con-

ference (IACC), 2015 IEEE International”, pages 899–904. June 2015. doi:
10.1109/IADCC.2015.7154835.
[Asl12] Aslam, M.; Rea, S.; and Pesch, D. “Service provisioning for the wsn cloud”.
In “Cloud Computing (CLOUD), 2012 IEEE 5th International Conference on”,
pages 962–969. June 2012. ISSN 2159-6182. doi:10.1109/CLOUD.2012.132.
[Atm09] Atmel. “AT86RF230: Low power 2.4 ghz transceiver for zigbee, IEEE 802.15.4,
6LoWPAN, RF4CE and ISM applications”, 2009.
[Atm14] Atmel. “Atmega 1281datasheet”, February 2014.
[Aww09] Awwad, S.; Ng, C.; Noordin, N.; and Rasid, M. “Cluster based routing protocol
for mobile nodes in wireless sensor network”. In “Collaborative Technologies
and Systems, 2009. CTS ’09. International Symposium on”, pages 233–241. May
2009. doi:10.1109/CTS.2009.5067486.
[Axe81] Axelrod, R. and Hamilton, W.D. “The evolution of cooperation”. Science,
211(4489):1390–1396, 1981.
[Bac10] Bachir, A.; Dohler, M.; Watteyne, T.; and Leung, K. “MAC essentials for
wireless sensor networks”. Communications Surveys Tutorials, IEEE, 12(2):222
–248, quarter 2010. ISSN 1553-877X. doi:10.1109/SURV.2010.020510.00058.
[Bag10] Baghyalakshmi, D.; Ebenezer, J.; and Satyamurty, S. “Low latency and energy
efficient routing protocols for wireless sensor networks”. In “Wireless Commu-
nication and Sensor Computing, 2010. ICWCSC 2010. International Conference
on”, pages 1 –6. jan. 2010. doi:10.1109/ICWCSC.2010.5415892.
[Bar13a] Barcelo, M.; Correa, A.; Lopez Vicario, J.; and Morell, A. “Multi-tree routing
for heterogeneous data traffic in wireless sensor networks”. In “Communications
(ICC), 2013 IEEE International Conference on”, pages 1899–1903. June 2013.
ISSN 1550-3607. doi:10.1109/ICC.2013.6654799.
[Bar13b] Barcelo, M.; Correa, A.; Lopez Vicario, J., J.; and Morell, A. “Cooperative
multi-tree sleep scheduling for surveillance in wireless sensor networks”. In
“Military Communications Conference, MILCOM 2013 - 2013 IEEE”, pages
200–205. Nov 2013. doi:10.1109/MILCOM.2013.43.
[Bar13c] Barcelo, M.; Correa, A.; Vicario, J.L.; and Morell, A. “Joint routing and
transmission power control for collection tree protocol in WSN”. In “Personal
Indoor and Mobile Radio Communications (PIMRC), 2013 IEEE 24th Inter-
national Symposium on”, pages 1989–1993. Sept 2013. ISSN 2166-9570. doi:
10.1109/PIMRC.2013.6666470.
[Bar15] Barcelo, M.; Llorca, J.; Tulino, A.M.; and Raman, N. “The cloud service
distribution problem in distributed cloud networks”. In “Communications

158
BIBLIOGRAPHY

(ICC), 2015 IEEE International Conference on”, pages 344–350. June 2015.
doi:10.1109/ICC.2015.7248345.
[Bon12] Bonomi, F.; Milito, R.; Zhu, J.; and Addepalli, S. “Fog computing and its role
in the internet of things”. In “Proceedings of the First Edition of the MCC
Workshop on Mobile Cloud Computing”, MCC ’12, pages 13–16. ACM, New
York, NY, USA, 2012. ISBN 978-1-4503-1519-7. doi:10.1145/2342509.2342513.
[Bot14] Botta, A.; de Donato, W.; Persico, V.; and Pescape, A. “On the integration
of cloud computing and internet of things”. In “Future Internet of Things and
Cloud (FiCloud), 2014 International Conference on”, pages 23–30. Aug 2014.
doi:10.1109/FiCloud.2014.14.
[Bra09] Bravos, G. and Kanatas, A.G. “Integrating power control with routing to
satisfy energy and delay constraints in sensor networks”. European Trans-
actions on Telecommunications, 20(2):233–245, 2009. ISSN 1541-8251. doi:
10.1002/ett.1248.
[Cad13] Cadger, F.; Curran, K.; Santos, J.; and Moffett, S. “A survey of ge-
ographical routing in wireless ad-hoc networks”. Communications Sur-
veys Tutorials, IEEE, 15(2):621–653, Second 2013. ISSN 1553-877X. doi:
10.1109/SURV.2012.062612.00109.
[Cap12] Cappiello, C. and Schreiber, F.A. “Experiments and Analysis of Quality- and
Energy-Aware Data Aggregation Approaches in WSNs”. In “QDB 2012, 10th
International Workshop on Quality in Databases”, 2012.
[Che10] Chen, Y.; Nasser, N.; Salti, T.E.; and Zhang, H. “A multipath QoS routing
protocol in wireless sensor networks”. Int. J. Sen. Netw., 7(4):207–216, 2010.
ISSN 1748-1279. doi:10.1504/IJSNET.2010.033204.
[Chu11] Chung, T.P.; Lin, T.S.; Zheng, X.Y.; Yen, P.L.; and Jiang, J.A. “A load
balancing algorithm based on probabilistic multi-tree for wireless sensor net-
works”. In “Sensing Technology (ICST), 2011 Fifth International Confer-
ence on”, pages 527 –532. 28 2011-dec. 1 2011. ISSN 2156-8065. doi:
10.1109/ICSensT.2011.6137035.
[Cob14] Cobarzan, C.; Montavont, J.; and Noel, T. “Analysis and performance evalua-
tion of rpl under mobility”. In “Computers and Communication (ISCC), 2014
IEEE Symposium on”, pages 1–6. June 2014. doi:10.1109/ISCC.2014.6912471.
[Cor08] Correia, L. and Nogueira, J. “Transmission power control techniques for mac
protocols in wireless sensor networks”. In “Network Operations and Manage-
ment Symposium, 2008. NOMS 2008. IEEE”, pages 1049 –1054. april 2008.
ISSN 1542-1201. doi:10.1109/NOMS.2008.4575277.
[Cor12] Correa, A.; Morell, A.; Barcelo, M.; and Vicario, J. “Navigation system for
elderly care applications based on wireless sensor networks”. In “Signal Pro-

159
 BIBLIOGRAPHY

cessing Conference (EUSIPCO), 2012 Proceedings of the 20th European”, pages

210–214. Aug 2012. ISSN 2219-5491.

[Cor14] Correa, A.; Barcelo, M.; Morell, A.; and Lopez Vicario, J. “Enhanced inertial-
aided indoor tracking system for wireless sensor networks: A review”. Sen-
sors Journal, IEEE, 14(9):2921–2929, Sept 2014. ISSN 1530-437X. doi:
10.1109/JSEN.2014.2325775.

[Cot14] Cotuk, H.; Bicakci, K.; Tavli, B.; and Uzun, E. “The impact of transmission
power control strategies on lifetime of wireless sensor networks”. Computers,
IEEE Transactions on, 63(11):2866–2879, Nov 2014. ISSN 0018-9340. doi:
10.1109/TC.2013.151.

[Cuo07] Cuomo, F.; Della Luna, S.; Monaco, U.; and Melodia, F. “Routing in zigbee:
Benefits from exploiting the ieee 802.15.4 association tree”. In “Communica-
tions, 2007. ICC ’07. IEEE International Conference on”, pages 3271–3276. June
2007. doi:10.1109/ICC.2007.542.

[Dun04] Dunkels, A.; Gronvall, B.; and Voigt, T. “Contiki - a lightweight and flexible
operating system for tiny networked sensors”. In “Proceedings of the 29th
Annual IEEE International Conference on Local Computer Networks”, LCN ’04,
pages 455–462. IEEE Computer Society, Washington, DC, USA, 2004. ISBN
0-7695-2260-2. doi:10.1109/LCN.2004.38.

[Dur08] Durmaz Incel, O.; Jansen, P.; and Mullender, S. “MC-LMAC: A multi-channel
mac protocol for wireless sensor networks”, 2008.

[Dur12] Durmaz Incel, O.; Ghosh, A.; Krishnamachari, B.; and Chintalapudi, K.
“Fast data collection in tree-based wireless sensor networks”. Mobile Com-
puting, IEEE Transactions on, 11(1):86 –99, jan. 2012. ISSN 1536-1233. doi:
10.1109/TMC.2011.22.

[ElB11] ElBatt, T. and Andersen, T. “A cross-layer framework for multiple access

and routing design in wireless multi-hop networks”. Wireless Communica-
tions and Mobile Computing, 11(8):1155–1167, 2011. ISSN 1530-8677. doi:
10.1002/wcm.902.

[Fu12] Fu, Y.; Sha, M.; Hackmann, G.; and Lu, C. “Practical control of transmission
power for wireless sensor networks”. In “Network Protocols (ICNP), 2012 20th
IEEE International Conference on”, pages 1 –10. 30 2012-nov. 2 2012. doi:
10.1109/ICNP.2012.6459981.

[Gad06] Gadallah, Y. “A comparative study of routing strategies for wireless sensor net-
works: Are manet protocols good fit?” In “Proceedings of the 5th International
Conference on Ad-Hoc, Mobile, and Wireless Networks”, ADHOC-NOW’06,
pages 5–18. Springer-Verlag, Berlin, Heidelberg, 2006. ISBN 3-540-37246-6,
978-3-540-37246-2. doi:10.1007/118147643.

160
BIBLIOGRAPHY

[Gad12] Gaddour, O. and Koubı́A, A. “Survey rpl in a nutshell: A survey”.

Comput. Netw., 56(14):3163–3178, September 2012. ISSN 1389-1286. doi:
10.1016/j.comnet.2012.06.016.

[Gad14] Gaddour, O.; Koubaa, A.; Rangarajan, R.; Cheikhrouhou, O.; Tovar, E.; and
ABID, M. “Co-rpl: Rpl routing for mobile low power wireless sensor net-
works using corona mechanism”. In “Industrial Embedded Systems (SIES),
2014 9th IEEE International Symposium on”, pages 200–209. June 2014. doi:
10.1109/SIES.2014.6871205.

[Gna09] Gnawali, O.; Fonseca, R.; Jamieson, K.; Moss, D.; and Levis, P. “CTP: Robust
and efficient collection through control and data plane integration”. Tech. rep.
sing-09-01, Stanford University, April 2009.

[Gub13] Gubbi, J.; Buyya, R.; Marusic, S.; and Palaniswami, M. “Internet of things
(iot): A vision, architectural elements, and future directions”. Future Gener.
Comput. Syst., 29(7):1645–1660, September 2013. ISSN 0167-739X. doi:
10.1016/j.future.2013.01.010.

[Gun13] Gungor, V.C. and Hancke, G.P. Industrial Wireless Sensor Networks: Appli-
cations, Protocols, and Standards. CRC Press, Inc., Boca Raton, FL, USA, 1st
edition, 2013. ISBN 1466500514, 9781466500518.

[Guo12] Guo, P.; Jiang, T.; Zhang, Q.; and Zhang, K. “Sleep scheduling for crit-
ical event monitoring in wireless sensor networks”. Parallel and Distributed
Systems, IEEE Transactions on, 23(2):345–352, 2012. ISSN 1045-9219. doi:
10.1109/TPDS.2011.165.

[Ha13] Ha, K.; Pillai, P.; Lewis, G.; Simanta, S.; Clinch, S.; Davies, N.; and Satya-
narayanan, M. “The impact of mobile multimedia applications on data center
consolidation”. In “Cloud Engineering (IC2E), 2013 IEEE International Con-
ference on”, pages 166–176. March 2013. doi:10.1109/IC2E.2013.17.

[Haj11] Hajiaghayi, M.; Dong, M.; and Liang, B. “Optimal channel assignment and
power allocation for dual-hop multi-channel multi-user relaying”. In “INFO-
COM, 2011 Proceedings IEEE”, pages 76–80. April 2011. ISSN 0743-166X.
doi:10.1109/INFCOM.2011.5935298.

[Har10] Harms, T.; Sedigh, S.; and Bastianini, F. “Structural health monitor-
ing of bridges using wireless sensor networks”. Instrumentation Measure-
ment Magazine, IEEE, 13(6):14 –18, december 2010. ISSN 1094-6969. doi:
10.1109/MIM.2010.5669608.

[Hei00] Heinzelman, W.; Chandrakasan, A.; and Balakrishnan, H. “Energy-efficient

communication protocol for wireless microsensor networks”. In “System Sci-
ences, 2000. Proceedings of the 33rd Annual Hawaii International Conference
on”, page 10 pp. vol.2. jan. 2000. doi:10.1109/HICSS.2000.926982.

161
 BIBLIOGRAPHY

[Hu15] Hu, X.; Yang, L.; and Xiong, W. “A novel wireless sensor network frame for
urban transportation”. Internet of Things Journal, IEEE, PP(99):1–1, 2015.
ISSN 2327-4662. doi:10.1109/JIOT.2015.2475639.
[Hua12] Huang, P.; Xiao, L.; Soltani, S.; Mutka, M.; and Xi, N. “The evolution
of MAC protocols in wireless sensor networks: A survey”. Communica-
tions Surveys Tutorials, IEEE, PP(99):1 –20, 2012. ISSN 1553-877X. doi:
10.1109/SURV.2012.040412.00105.
[IEE11] IEEE. IEEE Standard 802.15.4, ”Part 15.4: Wireless Medium Access Control
(MAC) and Physical Layer (PHY) Specifications for Low-Rate Wireless Per-
sonal Area Networks (WPANs), June 2011.
[Ish13] Ishaq, I.; Carels, D.; Teklemariam, G.K.; Hoebeke, J.; Abeele, F.V.d.; Poorter,
E.D.; Moerman, I.; and Demeester, P. “Ietf standardization in the field of the
internet of things (iot): A survey”. Journal of Sensor and Actuator Networks,
2(2):235, 2013. ISSN 2224-2708. doi:10.3390/jsan2020235.
[Jia15] Jian, X.; Geng, Y.; Zhengyu, C.; and Qianqian, W. “A survey on the
privacy-preserving data aggregation in wireless sensor networks”. Com-
munications, China, 12(5):162–180, May 2015. ISSN 1673-5447. doi:
10.1109/CC.2015.7112038.
[Joh01] Johnson, D.B.; Maltz, D.A.; and Broch, J. “Ad hoc networking”. chapter DSR:
The Dynamic Source Routing Protocol for Multihop Wireless Ad Hoc Networks,
pages 139–172. Addison-Wesley Longman Publishing Co., Inc., Boston, MA,
USA, 2001. ISBN 0-201-30976-9.
[Jov11] Jovanovic, M.; Djordjevic, G.; Nikolic, G.; and Petrovic, B. “Multi-channel
media access control for wireless sensor networks: A survey”. In “Telecom-
munication in Modern Satellite Cable and Broadcasting Services (TELSIKS),
2011 10th International Conference on”, volume 2, pages 741–744. Oct 2011.
doi:10.1109/TELSKS.2011.6143217.
[Kal60] Kalman, R.E. “A new approach to linear filtering and prediction problems”.
Transactions of the ASME–Journal of Basic Engineering, 82(Series D):35–45,
1960.
[Kar00] Karp, B.N. “Geographic routing for wireless networks”. Technical report, Har-
vard University, 2000.
[Kim05] Kim, Y.; Govindan, R.; Karp, B.; and Shenker, S. “On the pitfalls of geographic
face routing”. In “in Proc. of DIALM-POMC”, pages 34–43. ACM Press, 2005.
[Kim06] Kim, D.S. and Chung, Y.J. “Self-organization routing protocol supporting mo-
bile nodes for wireless sensor network”. In “Computer and Computational Sci-
ences, 2006. IMSCCS ’06. First International Multi-Symposiums on”, volume 2,
pages 622–626. June 2006. doi:10.1109/IMSCCS.2006.265.

162
BIBLIOGRAPHY

[Ko11] Ko, J.; Dawson-Haggerty, S.; Gnawali, O.; Culler, D.; and Terzis, A. “Evalu-
ating the Performance of RPL and 6LoWPAN in TinyOS”. In “Proceedings of
the Workshop on Extending the Internet to Low power and Lossy Networks”,
IP+SN ’11. April 2011.
[Kor12] Korbi, I.; Ben Brahim, M.; Adjih, C.; and Saidane, L. “Mobility en-
hanced rpl for wireless sensor networks”. In “Network of the Future (NOF),
2012 Third International Conference on the”, pages 1–8. Nov 2012. doi:
10.1109/NOF.2012.6463993.
[Kuh03] Kuhn, F.; Wattenhofer, R.; and Zollinger, A. “Worst-case optimal and average-
case efficient geometric ad-hoc routing”, 2003.
[Kum08] Kumar, G.; Vinu Paul, M.; and Jacob, K. “Mobility metric based leach-
mobile protocol”. In “Advanced Computing and Communications, 2008. AD-
COM 2008. 16th International Conference on”, pages 248–253. Dec 2008. doi:
10.1109/ADCOM.2008.4760456.
[Kum10] Kumar, K. and Lu, Y.H. “Cloud computing for mobile users: Can offloading
computation save energy?” Computer, 43(4):51–56, April 2010. ISSN 0018-
9162. doi:10.1109/MC.2010.98.
[Kyö07] Kyösti, P.; Meinilä, J.; Hentilä, L.; Zhao, X.; and Jämsä, T. “Winner ii channel
models”. Technical report, EBITG, TUI, UOULU, CU/CRC, NOKIA, 2007.
[Lee12] Lee, K.; Sudhaakar, R.; and Ning, J. “A comprehensive evaluation of RPL
under mobility”. International Journal of Vehicular Technology, 2012:10, 2012.
[Lei13] Lei, L.; Zhong, Z.; Zheng, K.; Chen, J.; and Meng, H. “Challenges
on wireless heterogeneous networks for mobile cloud computing”. Wire-
less Communications, IEEE, 20(3):34–44, June 2013. ISSN 1536-1284. doi:
10.1109/MWC.2013.6549281.
[Lev11] Levis, P.; Clausen, T.; Hui, J.; Gnawaliand, O.; and Ko, J. “Rfc 6206: The
trickle algorithm”. In “Internet Engineering Task Force”, March 2011.
[Li07] Li, X. “Collaborative localization with received-signal strength in wireless sen-
sor networks”. Vehicular Technology, IEEE Transactions on, 56(6):3807–3817,
2007. ISSN 0018-9545. doi:10.1109/TVT.2007.904535.
[Li11] Li, J.; Guo, X.; and Guo, L. “Joint routing, scheduling and channel assignment
in multi-power multi-radio wireless sensor networks”. In “Performance Comput-
ing and Communications Conference (IPCCC), 2011 IEEE 30th International”,
pages 1 –8. nov. 2011. ISSN 1097-2641. doi:10.1109/PCCC.2011.6108087.
[Liu15] Liu, X.; Cao, J.; Song, W.Z.; Guo, P.; and He, Z. “Distributed sensing for
high-quality structural health monitoring using wsns”. Parallel and Distributed
Systems, IEEE Transactions on, 26(3):738–747, March 2015. ISSN 1045-9219.
doi:10.1109/TPDS.2014.2312911.

163
 BIBLIOGRAPHY

[Llo13] Llorca, J. and Tulino, A. “The content distribution problem and its complexity
classification”. Technical report, Alcatel-Lucent, 2013.

[Lon13] Long, N.T.; Uwase, M.P.; Tiberghien, J.; and Steenhaut, K. “QoS-aware cross-
layer mechanism for multiple instances RPL”. In “Advanced Technologies for
Communications (ATC), 2013 International Conference on”, pages 44–49. Oct
2013. ISSN 2162-1020. doi:10.1109/ATC.2013.6698074.

[Luo10] Luo, J.; Rosenberg, C.; and Girard, A. “Engineering wireless mesh networks:
Joint scheduling, routing, power control, and rate adaptation”. Networking,
IEEE/ACM Transactions on, 18(5):1387 –1400, oct. 2010. ISSN 1063-6692.
doi:10.1109/TNET.2010.2041788.

[Luo11] Luo, H.; Tao, H.; Ma, H.; and Das, S. “Data fusion with desired reliability in
wireless sensor networks”. Parallel and Distributed Systems, IEEE Transactions
on, 22(3):501 –513, march 2011. ISSN 1045-9219. doi:10.1109/TPDS.2010.93.

[Luo12] Luo, T.; Tan, H.P.; and Quek, T. “Sensor openflow: Enabling software-defined
wireless sensor networks”. Communications Letters, IEEE, 16(11):1896–1899,
November 2012. ISSN 1089-7798. doi:10.1109/LCOMM.2012.092812.121712.

[Mal06] Malakooti, B. and Thomas, I. “A distributed composite multiple criteria routing

using distance vector”. In “Networking, Sensing and Control, 2006. ICNSC ’06.
Proceedings of the 2006 IEEE International Conference on”, pages 42–47. 2006.
doi:10.1109/ICNSC.2006.1673115.

[Man01] Mani, C.S. and Srivastava, M.B. “Energy efficient routing in wireless sensor
networks”. In U. McLean, editor, “Military Communications Conference Pro-
ceedings on Communications for Network-Centric Operations: Creating the In-
formation Force”, volume 1, pages 357–361. October 2001.

[May79] Maybeck, P.S. Stochastic Models: Estimation and Control: v. 1: Estimation

and Control:, volume 1 of Mathematics in Science and Engineering. Elsevier
Science, 1979. ISBN 9780080956503.

[McK08] McKeown, N.; Anderson, T.; Balakrishnan, H.; Parulkar, G.; Peterson, L.;
Rexford, J.; Shenker, S.; and Turner, J. “Openflow: Enabling innovation in
campus networks”. SIGCOMM Comput. Commun. Rev., 38(2):69–74, March
2008. ISSN 0146-4833. doi:10.1145/1355734.1355746.

[Md13] Melo de, P.; Cunha, F.; and Loureiro, A. “A distributed protocol for cooperation
among different wireless sensor networks”. In “Communications (ICC), 2013
IEEE International Conference on”, pages 6035–6039. June 2013. ISSN 1550-
3607. doi:10.1109/ICC.2013.6655566.

[Meg11] Meghji, M. and Habibi, D. “Transmission power control in multihop

wireless sensor networks”. In “Ubiquitous and Future Networks (ICUFN),

164
BIBLIOGRAPHY

2011 Third International Conference on”, pages 25 –30. june 2011. doi:
10.1109/ICUFN.2011.5949130.
[MEM] MEMSIC. “IRIS datasheet”.
[Mie03] Mieghem, P.V. and Kuipers, F.A. “On the complexity of QoS routing”. Com-
puter Communications, 26:376–387, 2003.
[Mis14a] Misra, S.; Bera, S.; Mondal, A.; Tirkey, R.; Chao, H.C.; and Chattopadhyay, S.
“Optimal gateway selection in sensor-cloud framework for health monitoring”.
Wireless Sensor Systems, IET, 4(2):61–68, June 2014. ISSN 2043-6386. doi:
10.1049/iet-wss.2013.0073.
[Mis14b] Misra, S.; Chatterjee, S.; and Obaidat, M. “On theoretical modeling of sensor
cloud: A paradigm shift from wireless sensor network”. Systems Journal, IEEE,
PP(99):1–10, 2014. ISSN 1932-8184. doi:10.1109/JSYST.2014.2362617.
[Mit12] Mitton, N.; Papavassiliou, S.; Puliafito, A.; and Trivedi, K. “Combining cloud
and sensors in a smart city environment”. EURASIP Journal on Wireless Com-
munications and Networking, 2012(1):247, 2012. doi:10.1186/1687-1499-2012-
247.
[Mor13] Morell, A.; Vilajosana, X.; Vicario, J.L.; and Watteyne, T. “Label switching
over ieee802.15.4e networks”. Transactions on Emerging Telecommunications
Technologies, pages n/a–n/a, 2013. ISSN 2161-3915. doi:10.1002/ett.2650.
[MS05] Martinez-Sala, A.; Molina-Garcia-Pardo, J.M.; Egea-Ldpez, E.; Vales-Alonso,
J.; Juan-Llacer, L.; and Garcia-Haro, J. “An accurate radio channel model for
wireless sensor networks simulation”. Communications and Networks, Journal
of, 7(4):401–407, Dec 2005. ISSN 1229-2370. doi:10.1109/JCN.2005.6387982.
[Nas14] Nastic, S.; Sehic, S.; Le, D.H.; Truong, H.L.; and Dustdar, S. “Provisioning
software-defined iot cloud systems”. In “Future Internet of Things and Cloud
(FiCloud), 2014 International Conference on”, pages 288–295. Aug 2014. doi:
10.1109/FiCloud.2014.52.
[Pan14] Pantazis, N.; Nikolidakis, S.; and Vergados, D. “Energy-efficient rout-
ing protocols in wireless sensor networks: A survey”. Communications
Surveys Tutorials, IEEE, PP(99):1–41, 2014. ISSN 1553-877X. doi:
10.1109/SURV.2012.062612.00084.
[Pap02] Papoulis, A. and Pillai, S. Probability, Random Variables, and Stochastic Pro-
cesses. McGraw-Hill series in electrical engineering: Communications and signal
processing. Tata McGraw-Hill, 2002. ISBN 9780070486584.
[Pat03] Patwari, N.; Hero, A.; Perkins, M.; Correal, N.; and O’Dea, R. “Rel-
ative location estimation in wireless sensor networks”. Signal Processing,
IEEE Transactions on, 51(8):2137–2148, 2003. ISSN 1053-587X. doi:
10.1109/TSP.2003.814469.

165
 BIBLIOGRAPHY

[Pen10] Peng, B. and Kemp, A. “Impact of location errors on geographic routing in re-
alistic wsns”. In “Indoor Positioning and Indoor Navigation (IPIN), 2010 Inter-
national Conference on”, pages 1–7. Sept 2010. doi:10.1109/IPIN.2010.5647292.

[Pen11] Peng, B. and Kemp, A. “Energy-efficient geographic routing in the presence

of localization errors”. Computer Networks, 55(3):856 – 872, 2011. ISSN 1389-
1286. doi:http://dx.doi.org/10.1016/j.comnet.2010.10.020.

[Per99] Perkins, C. and Royer, E. “Ad-hoc on-demand distance vector routing”.

In “Mobile Computing Systems and Applications, 1999. Proceedings. WM-
CSA ’99. Second IEEE Workshop on”, pages 90–100. Feb 1999. doi:
10.1109/MCSA.1999.749281.

[Pu12] Pu, C.C.; Lim, S.Y.; and Ooi, P.C. “Measurement arrangement for the es-
timation of path loss exponent in wireless sensor network”. In “Computing
and Convergence Technology (ICCCT), 2012 7th International Conference on”,
pages 807–812. 2012.

[Raj14] Rajalingham, G.; Gao, Y.; Ho, Q.D.; and Le-Ngoc, T. “Quality of service differ-
entiation for smart grid neighbor area networks through multiple RPL instances-
rpl instances”. In “Proceedings of the 10th ACM Symposium on QoS and Secu-
rity for Wireless and Mobile Networks”, Q2SWinet ’14, pages 17–24. ACM, New
York, NY, USA, 2014. ISBN 978-1-4503-3027-5. doi:10.1145/2642687.2642695.

[Ram06] Raman, B.; Chebrolu, K.; Madabhushi, N.; Gokhale, D.Y.; Valiveti, P.K.;
and Jain, D. “Implications of link range and (in)stability on sensor network
architecture”. WiNTECH ’06, pages 65–72. ACM, New York, NY, USA, 2006.
ISBN 1-59593-539-8. doi:10.1145/1160987.1161000.

[Rao15] Rao, S. and Mehta, N. “Energy harvesting wsns for accurately estimating the
maximum sensor reading: Trade-offs and optimal design”. Wireless Communi-
cations, IEEE Transactions on, 14(8):4562–4573, Aug 2015. ISSN 1536-1276.
doi:10.1109/TWC.2015.2422811.

[Rap01] Rappaport, T. Wireless Communications: Principles and Practice. Prentice

Hall PTR, Upper Saddle River, NJ, USA, 2nd edition, 2001. ISBN 0130422320.

[Raw14] Rawat, P.; Singh, K.D.; Chaouchi, H.; and Bonnin, J.M. “Wireless sensor net-
works: A survey on recent developments and potential synergies”. J. Supercom-
put., 68(1):1–48, April 2014. ISSN 0920-8542. doi:10.1007/s11227-013-1021-9.

[Rib10] Ribeiro, A.; Schizas, I.; Roumeliotis, S.; and Giannakis, G. “Kalman filtering
in wireless sensor networks”. Control Systems, IEEE, 30(2):66–86, April 2010.
ISSN 1066-033X. doi:10.1109/MCS.2009.935569.

[Saa11] Saad, L. and Tourancheau, B. “Sinks mobility strategy in IPv6-based WSNs for
network lifetime improvement”. In “New Technologies, Mobility and Security

166
BIBLIOGRAPHY

(NTMS), 2011 4th IFIP International Conference on”, pages 1–5. Feb 2011.
ISSN 2157-4952. doi:10.1109/NTMS.2011.5720597.

[Saf12] Safdar, V.; Bashir, F.; Hamid, Z.; Afzal, H.; and Pyun, J.Y. “A hybrid routing
protocol for wireless sensor networks with mobile sinks”. In “Wireless and
Pervasive Computing (ISWPC), 2012 7th International Symposium on”, pages
1–5. July 2012. doi:10.1109/ISWPC.2012.6263665.

[Sai14] Saifullah, A.; Xu, Y.; Lu, C.; and Chen, Y. “Distributed channel alloca-
tion protocols for wireless sensor networks”. Parallel and Distributed Sys-
tems, IEEE Transactions on, 25(9):2264–2274, Sept 2014. ISSN 1045-9219.
doi:10.1109/TPDS.2013.185.

[San15] Santoshkumar and Udaykumar, R. “Development of wsn system for precision

agriculture”. In “Innovations in Information, Embedded and Communication
Systems (ICIIECS), 2015 International Conference on”, pages 1–5. March 2015.
doi:10.1109/ICIIECS.2015.7192904.

[Sar14] Sara, G. and Sridharan, D. “Routing in mobile wireless sensor network: a

survey”. Telecommunication Systems, 57(1):51–79, 2014. ISSN 1018-4864. doi:
10.1007/s11235-013-9766-2.

[Sat09] Satyanarayanan, M.; Bahl, P.; Caceres, R.; and Davies, N. “The case for vm-
based cloudlets in mobile computing”. Pervasive Computing, IEEE, 8(4):14–23,
Oct 2009. ISSN 1536-1268. doi:10.1109/MPRV.2009.82.

[Sat15] Satyanarayanan, M.; Schuster, R.; Ebling, M.; Fettweis, G.; Flinck, H.; Joshi,
K.; and Sabnani, K. “An open ecosystem for mobile-cloud convergence”. Com-
munications Magazine, IEEE, 53(3):63–70, March 2015. ISSN 0163-6804. doi:
10.1109/MCOM.2015.7060484.

[Sha13] Shamsan Saleh, A.M.; Ali, B.M.; Rasid, M.F.A.; and Ismail, A. “A survey on
energy awareness mechanisms in routing protocols for wireless sensor networks
using optimization methods”. Transactions on Emerging Telecommunications
Technologies, pages n/a–n/a, 2013. ISSN 2161-3915. doi:10.1002/ett.2679.

[Shi13] Shiqiang, et al, W. “Mobile micro-cloud: Application classification, mapping,

and deployment”. In “Annual Fall Meeting of ITA (AMITA)”, 2013.

[Shi14] Shi, H.; Prasad, R.; Onur, E.; and Niemegeers, I. “Fairness in wire-
less networks:issues, measures and challenges”. Communications Sur-
veys Tutorials, IEEE, 16(1):5–24, First 2014. ISSN 1553-877X. doi:
10.1109/SURV.2013.050113.00015.

[Siv10] Sivagami, A.; Pavai, K.; Sridharan, D.; and SatyaMurty, S. “Estimating the
energy consumption of wireless sensor node: Iris”. International Journal on
Recent Trends in Engineering and Technology, 3(4):141–143, May 2010.

167
 BIBLIOGRAPHY

[Sto14] Stojmenovic, I. and Wen, S. “The fog computing paradigm: Scenarios and
security issues”. In “Computer Science and Information Systems (FedCSIS),
2014 Federated Conference on”, pages 1–8. Sept 2014. doi:10.15439/2014F503.

[Sur15] Suryadevara, N.; Mukhopadhyay, S.; Kelly, S.; and Gill, S. “Wsn-based smart
sensors and actuator for power management in intelligent buildings”. Mecha-
tronics, IEEE/ASME Transactions on, 20(2):564–571, April 2015. ISSN 1083-
4435. doi:10.1109/TMECH.2014.2301716.

[Tia13] Tian, B.; Hou, K.M.; Shi, H.; Liu, X.; Diao, X.; Li, J.; Chen, Y.; and Chanet,
J.P. “Application of modified rpl under vanet-wsn communication architecture”.
In “Computational and Information Sciences (ICCIS), 2013 Fifth International
Conference on”, pages 1467–1470. June 2013. doi:10.1109/ICCIS.2013.387.

[Tse02] Tseng, Y.C.; Hsu, C.S.; and Hsieh, T.Y. “Power-saving protocols for IEEE
802.11-based multi-hop ad hoc networks”. In “INFOCOM 2002. Twenty-First
Annual Joint Conference of the IEEE Computer and Communications Societies.
Proceedings. IEEE”, volume 1, pages 200–209 vol.1. 2002. ISSN 0743-166X. doi:
10.1109/INFCOM.2002.1019261.

[Vil10] Vilajosana, X.; Llosa, J.; Pacho, J.C.; Vilajosana, I.; Juan, A.A.; Vicario, J.L.;
and Morell, A. “Zero: Probabilistic routing for deploy and forget wireless
sensor networks”. Sensors, 10(10):8920–8937, 2010. ISSN 1424-8220. doi:
10.3390/s101008920.

[Vil14] Vilajosana, X. and Pister, K. “Minimal 6TiSCH configuration”. Internet-draft,

6TiSCH, April 2014.

[Vis15] Vishwanath, A.; Jalali, F.; Hinton, K.; Alpcan, T.; Ayre, R.; and Tucker, R.
“Energy consumption comparison of interactive cloud-based and local applica-
tions”. Selected Areas in Communications, IEEE Journal on, 33(4):616–626,
April 2015. ISSN 0733-8716. doi:10.1109/JSAC.2015.2393431.

[Wan06] Wang, A. and Sodini, C. “On the energy efficiency of wireless transceivers”. In
“Communications, 2006. ICC ’06. IEEE International Conference on”, volume 8,
pages 3783–3788. June 2006. ISSN 8164-9547. doi:10.1109/ICC.2006.255661.

[Wan10] Wang, Q. “Traffic analysis and modeling in wireless sensor networks and their
applications on network optimization and anomaly detection”. Network Proto-
cols and Algorithms, Macrothink Institute, 2(1), 2010.

[Wat11] Watteyne, T.; Molinaro, A.; Richichi, M.; and Dohler, M. “From manet to
ietf roll standardization: A paradigm shift in wsn routing protocols”. Com-
munications Surveys Tutorials, IEEE, 13(4):688–707, 2011. ISSN 1553-877X.
doi:10.1109/SURV.2011.082710.00092.

168
BIBLIOGRAPHY

[Wei07] Wei, W. and Zakhor, A. “Multiple tree video multicast over wireless ad hoc
networks”. Circuits and Systems for Video Technology, IEEE Transactions on,
17(1):2–15, Jan 2007. ISSN 1051-8215. doi:10.1109/TCSVT.2006.885719.
[Wel95] Welch, G. and Bishop, G. “An introduction to the kalman filter”. Technical
report, Chapel Hill, NC, USA, 1995.
[Wen14] Wenge, R.; Zhang, X.; Dave, C.; Chao, L.; and Hao, S. “Smart city ar-
chitecture: A technology guide for implementation and design challenges”.
Communications, China, 11(3):56–69, March 2014. ISSN 1673-5447. doi:
10.1109/CC.2014.6825259.
[Win12] Winter, T.; Thubert, P.; Brandt, A.; Clausen, T.; Hui, J.; Kelsey, R.; Levis,
P.; Pister, K.; Struick, R.; and Vasseur, J. “Rpl: Ipv6 routing protocol for
low-power and lossy networks”. In “IETF RFC 6550”, March 2012.
[Wu08] Wu, Y.; Stankovic, J.; He, T.; and Lin, S. “Realistic and efficient multi-
channel communications in wireless sensor networks”. In “INFOCOM 2008.
The 27th Conference on Computer Communications. IEEE”, pages 1193 –1201.
april 2008. ISSN 0743-166X. doi:10.1109/INFOCOM.2008.175.
[Xia10] Xiao, Y.; Chen, H.; and Haizhon., F. Handbook on Sensor Networks. World
Scientific Publishing Co., 2010. ISBN 978-981-283-730-1.
[Xia13] Xiao, Y.; Simoens, P.; Pillai, P.; Ha, K.; and Satyanarayanan, M. “Lowering
the barriers to large-scale mobile crowdsensing”. In “Proceedings of the 14th
Workshop on Mobile Computing Systems and Applications”, HotMobile ’13,
pages 9:1–9:6. ACM, New York, NY, USA, 2013. ISBN 978-1-4503-1421-3. doi:
10.1145/2444776.2444789.
[Yet12] Yetgin, H.; Cheung, K.; and Hanzo, L. “Multi-objective routing optimization
using evolutionary algorithms”. In “Wireless Communications and Networking
Conference (WCNC), 2012 IEEE”, pages 3030–3034. April 2012. ISSN 1525-
3511. doi:10.1109/WCNC.2012.6214324.
[Yi15] Yi, S.; Li, C.; and Li, Q. “A survey of fog computing: Concepts, applications and
issues”. In “Proceedings of the 2015 Workshop on Mobile Big Data”, Mobidata
’15, pages 37–42. ACM, New York, NY, USA, 2015. ISBN 978-1-4503-3524-9.
doi:10.1145/2757384.2757397.
[Yoo14] Yoon, J.; Zhang, H.; Banerjee, S.; and Rangarajan, S. “Video multicast with
joint resource allocation and adaptive modulation and coding in 4g networks”.
Networking, IEEE/ACM Transactions on, 22(5):1531–1544, Oct 2014. ISSN
1063-6692. doi:10.1109/TNET.2013.2279887.
[Yu01] Yu, Y.; Govindan, R.; and Estrin, D. “Geographical and energy aware rout-
ing: A recursive data dissemination protocol for wireless sensor networks”. In
“UCLA Computer Science Department Technical Report”, pages 1–11. 2001.

169
 BIBLIOGRAPHY

[Zam08] Zamalloa, M.Z.n.; Seada, K.; Krishnamachari, B.; and Helmy, A. “Effi-
cient geographic routing over lossy links in wireless sensor networks”. ACM
Trans. Sen. Netw., 4(3):12:1–12:33, June 2008. ISSN 1550-4859. doi:
10.1145/1362542.1362543.
[Zha12] Zhang, G.a.; Gu, J.y.; Bao, Z.h.; Xu, C.; and Zhang, S.b. “Joint routing and
channel assignment algorithms in cognitive wireless mesh networks”. Trans-
actions on Emerging Telecommunications Technologies, pages n/a–n/a, 2012.
ISSN 2161-3915. doi:10.1002/ett.2560.
[Zha14] Zhang, Y.; Sun, L.; Song, H.; and Cao, X. “Ubiquitous wsn for healthcare:
Recent advances and future prospects”. Internet of Things Journal, IEEE,
1(4):311–318, Aug 2014. ISSN 2327-4662. doi:10.1109/JIOT.2014.2329462.
[Zho06] Zhou, G.; Huang, C.; Yan, T.; He, T.; Stankovic, J.; and Abdelzaher, T.
“Mmsn: Multi-frequency media access control for wireless sensor networks”.
In “INFOCOM 2006. 25th IEEE International Conference on Computer Com-
munications. Proceedings”, pages 1–13. April 2006. ISSN 0743-166X. doi:
10.1109/INFOCOM.2006.250.
[Zhu15] Zhu, C.; Sheng, Z.; Leung, V.; Shu, L.; and Yang, L. “Toward offering more
useful data reliably to mobile cloud from wireless sensor network”. Emerging
Topics in Computing, IEEE Transactions on, 3(1):84–94, March 2015. ISSN
2168-6750. doi:10.1109/TETC.2014.2364921.

170