Energy Consumption of Internet of Things Applications and Services Chrispin Gray ORCID iD: 0000-0002-5223-4314 Submitted in partial fulfilment of the requirements of the degree of Doctor of Philosophy Department of Electrical and Electronic Engineering THE UNIVERSITY OF MELBOURNE August 2018 Produced on archival quality paper.
253
Embed
Energy Consumption of Internet of Things ... - Minerva Access
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Energy Consumption of Internet ofThings Applications and Services
Chrispin GrayORCID iD: 0000-0002-5223-4314
Submitted in partial fulfilment of the requirements of the degree of
Doctor of Philosophy
Department of Electrical and Electronic EngineeringTHE UNIVERSITY OF MELBOURNE
All rights reserved. No part of the publication may be reproduced in any form by print,photoprint, microfilm or any other means without written permission from the author.
Abstract
THE Internet of Things (IoT) is a new paradigm of interconnectivity that has recently
garnered attention in the field of ICT, with an estimated proliferation of 50-200
billion connected devices (i.e. IoT/smart devices) by the end of the decade. This expo-
nential device growth raises concerns as it elicits potential risks including an increase in
global energy consumption arising from the deployment of such numbers of devices, the
additional network energy cost for handling potential IP traffic increment and the poten-
tial impact on the global energy consumption and carbon footprint of the ICT industry.
However, due to the development/deployment of many IoT services being in their em-
bryonic stage, there is little research on the characterisation of energy consumption of
these services in the literature.
In this thesis, we aim to investigate and gain a better understanding of the energy
consumption of IoT network applications and services. We do so by developing energy
consumption models and in turn, energy-efficient network architectures for the delivery
of IoT services. To achieve this goal, we employ and model a few case studies including
two of the most well-known and widely deployed IoT services, home automation and
security (HAS) and video surveillance services.
For the assessment of energy consumption of an IoT service, we obtained a range of
IoT products including a consumer-off-the-shelf (COTS) HAS system, as a representative
example. We analyse and model (through direct measurements) the energy consumption
of each component and the complete system including an IoT attributable share of the
home gateway energy consumption. Our results reveal that the energy consumption of
a simple COTS IoT service is non-trivial (more than one-third) when compared to the
annual energy usage of a mid-size suburban home. HAS energy consumption globally
iii
becomes substantial, in comparison with the ICT industry’s energy consumption projec-
tions, when IoT service numbers are scaled using published deployment estimates.
The IoT leverages a number of existing and emerging technologies to provide a com-
plete end-to-end service, one of which is short-range wireless network protocols. We
obtained, measured and analysed the energy-efficiency of five of the most popular COTS
wireless protocol modules, Bluetooth Classic & Low Energy, ZigBee, Wi-Fi and RF 433 MHz.
We compare these technologies through their application in a simple domestic stock-
control IoT service with three communication paradigm options. The results demonstrate
that careful consideration should be given to the choice of a communication mode and
wireless interface in IoT application development. Such decision should be driven by the
volume of traffic exchange and frequency of transmission of the application/service.
The emergence of edge/fog computing as an alternative to cloud computing promises
to tackle some critical pitfalls of cloud including energy consumption. To investigate
the energy efficiency of IoT network architectures, the data-intensive video surveillance
IoT service is employed as a case study. Using the end-to-end energy models devel-
oped, we investigate four (Local, Edge and Cloud) dissimilar network architectures for
the delivery of IoT services. We show that it is more energy-wise to adopt an edge-
based architecture for on-demand streaming applications but both live streaming and
computationally-intensive applications are more energy-efficient when designed with a
local access architecture.
We further study a number of access network technologies for the IoT. They include
very-high bit rate digital subscriber line (VDSL2), passive optical network (PON), point-
to-point optical network (PtP), fourth generation long term evolution (4G LTE), low power
wide area networks (LPWA) and Wi-Fi access (Shared and Unshared). We show that for
low data access rates, LPWA is more energy-efficient while a Shared Wi-Fi access with
PON backhaul is most energy-efficient for medium to higher data access rates.
The findings in this thesis reinforce the need for careful design consideration when
developing future IoT solutions.
iv
Declaration
This is to certify that
1. the thesis comprises only my original work towards the PhD,
2. due acknowledgement has been made in the text to all other material used,
3. the thesis is less than 100,000 words in length, exclusive of tables, maps, bibliogra-
phies and appendices.
Chrispin Gray, August 2018
v
Preface
This thesis is submitted for the degree of Doctor of Philosophy at the University of Mel-
bourne. The content of this thesis is completely original except where acknowledgement
and references are made to previous work. The research work described herein was
conducted by the student under the supervision of Laureate Professor Emeritus Rodney
Tucker, Dr Kerry Hinton, Mr Robert Ayre and Dr Leith Campbell. The student benefited
from technical comments and guidance from his supervisors through one-on-one interac-
tions and group meetings alike. All of the work towards this thesis was carried out after
enrolment into the degree and the thesis has not been submitted for any other degree or
qualification at any other university. In the preparation of this thesis, no third-party ed-
itorial assistance was solicited. Financial support for this undertaking was provided by
the Australian Government and University of Melbourne through the International Post-
graduate Research Scholarship (IPRS) and the Australian Postgraduate Award (APA).
The student was also a recipient of a top-up scholarship from the Centre for Energy-
Efficient Telecommunication (CEET) in collaboration with Alcatel-Lucent and Bell-labs.
Part of this work has been presented for publication. The following lists the publica-
tions and contributions of its authors:
• F. Jalali, S. Khodadustan, C. Gray, K. Hinton and F. Suits, ”Greening IoT with Fog:
A Survey,” IEEE International Conference on Edge Computing (EDGE), Honolulu, HI,
2017, pp. 25-31.
Contributions: The first three authors provided most of the content and co-wrote
the manuscript. The fourth and fifth authors provided technical comments, advice
and proofreading of the manuscript up to its publication.
vii
• C. Gray and L. Campbell, ”Should my toaster be polled? Towards an energy-
efficient Internet of Things,” 26th International Telecommunication Networks and Appli-
cations Conference (ITNAC), Dunedin, 2016, pp. 26-31. (Best Student Paper Award)
Contributions: The first author developed the main content of the work including
experiments, modelling and analysis and manuscript writing. The second author
provided the introduction, conclusion, proofreading and technical comments.
• C. Gray, R. Ayre, K. Hinton and R. S. Tucker, ”Power consumption of IoT access
network technologies,” IEEE International Conference on Communication Workshop
(ICCW), London, 2015, pp. 2818-2823.
Contributions: The first author carried out the modelling and analysis work and
developed the manuscript. The second and third authors mainly supervised the
development of this work and together with the fourth author, provided technical
comments and suggestions, including proofreading the manuscript prior to publi-
cation.
viii
Acknowledgements
Completing a PhD is a fulfilling and rewarding achievement that is impossible without
the help and support of a few amazing people. These people were instrumental in nudg-
ing me ever so slightly to the right path and supportive through challenging times.
Foremost, I would like to thank my supervisors, Laureate Professor Rodney Tucker,
Dr Kerry Hinton, Mr Robert Ayre and Dr Leith Campbell for their inspiration, guidance
and support throughout my PhD candidature. I deeply appreciate Kerry for shaping
the framework of this thesis and particularly grateful to Robert for his practical exper-
tise, mentorship and careful guidance especially during the experimental phase of this
study. Special thanks to Leith for his invaluable contribution towards this thesis. Thanks
to my advisory committee chair Professor Christopher Leckie for his time, feedback and
words of advice. I am thankful for the financial assistance I received from the Australian
Government, the University of Melbourne and the Center for Energy-Efficient Telecom-
munications (CEET).
It is a pleasure to also thank my friends and colleagues from CEET including Dinuka
Kudavithana, Fatemeh Jalali, Sascha Sussspeck, Hamid Khodakarami, Ashrar Matin and
Olivia Zhu. The friendship we share is one I would forever cherish.
I am indebted to my late dad Crispin Gray and my mum Paulina Gray for their un-
flinching love and support, without which none of this would be possible. Special thanks
and appreciation is also extended to my sister Crispina Lot-Thomas.
Most importantly, I would like to express my deepest gratitude to my loving wife
Mina for her love, patience and kind encouragement throughout this journey and for
affording me all the help I could ever ask for. Finally, thanks to my son Leon for putting
a smile on my face every time I walk through the door.
5.7 Descriptions of parameters defined in the energy models. . . . . . . . . . . 128
5.8 Parameters of the Raspberry Pi 3 Model B with attached Pi Camera. . . . . 141
6.1 Power consumption and data rate of unshared network elements . . . . . . 164
6.2 Power Consumption and data rates of shared network elements . . . . . . 164
6.3 Energy per bit values of shared network elements . . . . . . . . . . . . . . 165
6.4 Power consumption values for 4G LTE base station components from [23] 169
6.5 Traffic level distribution of a dense urban area over a 24 hour period . . . 170
xxvi
List of Acronyms
ADC Analogue to Digital ConversionADSL Asynchronous Digital Subscriber LineBNG Border Network GatewayBS Base StationCSMA/CA Carrier Sense Multiple Access with Collision AvoidanceDAC Digital to Analogue ConversionDC Data CentreDSLAM Digital Subscriber Line Access MultiplexerFTTN Fibre to the NodeFTTP Fibre to the PremisesHTTP Hyper-Text Transfer ProtocolICT Information and Communication TechnologyIoT Internet of ThingsIP Internet ProtocolISM Industrial Scientific and MedicalISP Internet Service ProviderLAN Local Area NetworkLPWA Low Power Wide Area NetworkLTE Long Term EvolutionM2M Machine-to-MachineMCU Microcontroller UnitPC Personal ComputerPON Passive Optical NetworkPtP Point-to-Point Optical NetworkPUE Power Usage EffectivenessQoS Quality of ServiceVDSL Very-high-bit-rate Digital Subscriber LineWSN Wireless Sensor Networks
xxvii
Chapter 1
Introduction
THE Internet of Things (IoT) is a novel paradigm of interconnectivity that has re-
cently garnered momentum in the field of modern telecommunications. The IoT
ecosystem is expected to usher in an era of ubiquitous presence of uniquely identifiable
physical objects or ”things” (referred to as IoT devices or smart devices), connected via
the Internet to measure, report and, in some cases, perform actions autonomously [24,25].
There are tens of IoT use-case applications and services across many industries in-
cluding but not limited to E-Health, Smart Building, Smart Manufacturing, Smart Agri-
culture, Connected Vehicle, Environmental Monitoring and Home Automation & Secu-
rity (HAS)/Smart Home [6,24,26]. A number of these IoT services are becoming increas-
ingly popular with users, partly aided by products and services from leading technology
pioneers like Apple, Google and Amazon. They include HAS (e.g. Apple HomePod,
Google Home, Amazon Echo), video surveillance (e.g. Google Nest Cam, Netgear Arlo)
and E-Health or Wearable applications (e.g. FitBit) to name a few. Many of these services
are still in the embryonic state of deployment and are yet to be fully characterised in the
literature.
It is estimated that the IoT will host between 50-200 billion connected devices by 2020
[2, 4–6]. These devices support a variety of IoT services. While there are many bene-
fits and opportunities arising from the adoption of these services [24, 25], those benefits
could come with potential risks, one of which is the additional electrical energy required
for powering these devices and their collective standby energy consumption to maintain
network connectivity and/or their ”smartness” [22, 27]. As the number of devices grows
exponentially, so too will the number of IoT services they support, the volume of traffic
1
2 Introduction
generated by said devices, the different types of traffic flows and the anticipated aggre-
gate network traffic across the Internet [28]. This growth further raises concerns for the
potential increase in energy consumption of existing and future networks to support IoT
service traffic volumes.
Growth in energy consumption of the information and communication technology
(ICT) industry is also of increasing concern. Some estimates suggest that the ICT industry
is accountable for 2-4% of global carbon emissions [29, 30], a value that could increase
with the deployment of billions of IoT devices. An International Energy Agency (IEA)
report [31] in 2013 forecast global energy consumption of network-enabled devices to
grow from 615 TWh in 2013 to 1,140 TWh in 2025, if no improvements are made to the
ICT industry’s energy usage. A similar study by Andrae et al. [32] in 2015 shows the
electricity usage trend of consumer devices is estimated to reach about 1,330 TWh by 2022
if only 1% annual improvement in energy efficiency is effected from 2010. A critical point
to mention is that the above estimates are based on network-enabled devices including
PCs, Smart TVs, mobile phones and home entertainment systems. They do not include
IoT devices such as IoT gateways, sensors, actuators, IP cameras and smart consumer
appliances (e.g. connected refrigerator). Thus, while there have been many studies of
trends in global energy consumption, and more specifically in ICT energy consumption,
the future impact of billions of IoT devices on ICT energy consumption has not been fully
explored.
The basic IoT service generally includes hardware IoT devices (i.e. sensors, actuators,
cameras, etc.) at the network edge, the IoT gateway with ”typical” gateway functions
(e.g. network access and traffic control, aggregation, etc.) and a cloud service for data
processing, storage, management and control. Communication between IoT devices and
their designated IoT gateway would likely be wireless [24, 33], provided by one of many
enabling wireless network protocols including Bluetooth Classic (BT), Bluetooth Low En-
ergy (BLE), Wi-Fi, ZigBee and Radio Frequency (RF) 433 MHz. The energy consumption
of two or more of these protocols has been studied and compared in the literature [34–39].
For example, the authors in [34] showed that Wi-Fi consumes 5-times more power than
BT when considering similar example chipsets with the same data rate for a single con-
1.1 Energy Consumption of IoT Services 3
nection, while the study in [35] revealed energy per unit data transferred for BLE to be
2.5-times better than ZigBee.
Between the IoT gateway and its parent cloud service, there are a number of network
segments including the access, metro/edge and core networks [7]. Research on the en-
ergy consumption of these network segments revealed the access network as the least
energy-efficient [7, 40, 41]. For IoT services, there exist a variety of wireline or wireless
network access technology choices ranging from xDSL (digital subscriber line) and opti-
cal network technologies, 4G/5G mobile technologies, to low-power wide area networks
(LPWA). Studies on the energy efficiency of access technologies have mostly considered
higher data access rates (> 1 Mb/s) and do not include more recent access network tech-
nologies like LPWA [42, 43]. IoT service access rates tend to be lower than other user
services [26, 33].
In this thesis, we aim to investigate and develop energy consumption models and
energy-efficient network architectures for the delivery of IoT services. To achieve this
goal, we employ a few case studies including two of the most well-known and widely
deployed IoT services, HAS and video surveillance systems.
1.1 Energy Consumption of IoT Services
Connected devices in HAS worldwide is expected to reach about half (50%) of all con-
nected devices by 2020 [28]. Hence the popularity of HAS in particular has attracted
significant research attention on mitigating potential energy consumption increase and
designing energy-efficient systems and protocols. Most studies on energy efficiency of
HAS or other IoT applications tend to take a more device-centric approach [22,27,44–50];
some, a more network-centric focus [51,52]; while others focus on processing and storage
centres [53]. These studies offer a variety of techniques for improving energy efficiency,
including duty-cycle optimisation, sleep-modes, efficient node placement, routing mech-
anisms, lower information transfer rate, virtualization and a move towards distributed
computing. A common theme among these studies is a focus on the energy efficiency of
the end-devices, their local networks and gateways/aggregators, with little attention to
4 Introduction
the energy consumed by other segments of the service including:
(i) (a share of) the energy consumed by the home gateway modem - which facilitates
Internet connectivity - attributable to IoT data traffic.
(ii) the energy consumed by the transport networks (including access, metro/edge and
core networks) for transmission of IoT data, events messages or commands and the
data centres which processes and stores the said data.
Another IoT service that is increasingly being embraced by users is the video surveil-
lance service, propelled by the ever increasing need for both private and public safety
and security. The modern IoT video surveillance service - unlike its legacy counter-
parts [54] - benefits immensely from an integrated (i.e. video camera, video encoder and
server/storage in one unit) and less complex system design (e.g. new generation net-
work IP camera [55]), and the availability of cheap, reliable and secure cloud computing
services [56, 57]. Some estimates suggest that video surveillance services would likely be
the most data-intensive IoT application [28, 58]. Given such estimates, with video trans-
mission being the most energy-demanding part of the system [59], there is an incentive
to minimise the number and size of video data frames transmitted using compression
algorithms [60]) and to develop more energy-efficient network architectures. Hence, the
research work in this area includes the design of energy-efficient video sensor devices
[12, 61], efficient multi-camera coverage overlap [62], local video processing as opposed
to cloud [63, 64] and balancing energy consumption with key parameters such as band-
width, delay and frame rate [65, 66]. From the survey of literature, there is a general lack
of extensive research on an end-to-end characterisation of IoT video surveillance services.
A related study is the 2016 IEA report [22], which estimated the standby energy con-
sumption of mains-powered IoT devices and IoT gateways for a number of IoT use-case
Many sensor devices operate at very low duty-cycle (e.g. battery-powered temperature
sensor [116]) and are only active for a short burst of time, often with the majority of
their operational lifetime spent in the standby or sleep state. Most sensor devices can
be classified into one of two main types based on their differences in application and
communication methods. These two types are referred to as: (i) time-based sensor device
and (ii) event-driven sensor device in this model.
Time-Based Sensor Device
Time-based sensor devices perform periodic operations (e.g. once a minute) with fixed
intervals over time (e.g. temperature, pressure sensor devices). Figure 3.3 depicts the
PDEV
tCOM
Time (t)tMCU
1 Cycle (TC)
tSEN
0
PSEN
PMCU
PCOM
PSBY
Communications (COM)
Processing (MCU)
Sensing (SEN)
Standby (SBY)
NR
Figure 3.3: Example power consumption time evolution of a time-based sensor deviceoperating periodically with fixed intervals between transmissions.
3.3 Home Automation and Security System Energy Model 45
power consumption time evolution of a time-based sensor device. This figure emulates the
measured behaviour of example sensor devices (see Appendix A) on which the model is
developed. Each operational cycle consists of the three main tasks:
(i) Collect each data sample within a timeframe tSEN. A total of NSEN samples are
collected in one operational cycle. Depending on the application and the particular
device, only one, or many sample(s) may be collected per cycle.
(ii) Process or analyse the data samples collected using instructions/code in memory
within time tMCU. The data is also prepared for onward transmission via a commu-
nications module.
(iii) Transmit collected data to a parent device in time tCOM with NR number of retrans-
missions per message (e.g. NR = 3 as shown in Figure 3.3). A parent device can be
an IoT gateway or intermediate routing devices in a mesh network.
Figure 3.3 also includes a standby phase between the three main tasks when the system
is inactive. The figure shows PSBY continues through the SEN, MCU and COM periods
because the functions that make up this power consumption (e.g. basic clock, memory
keep-alive) continue concurrently with the main tasks. For a given sensor device, it is
assumed that the energy consumed for a particular task x (i.e. Ex = 〈Px〉× tx) is constant.
An expression of the energy consumption of a time-based sensor device (ts) for 1 complete
operational cycle (E(ts)DEVC) with a duration TC is therefore given by:
E(ts)DEVC= NSENESEN +NMCUEMCU +NCOM
(NRECOM
)+ ESBYC (3.5)
where NR > 1 and ESBYC = PSBYTC. The total energy consumption of a time-based sensor
device with NC cycles of operation over a long operating time T is then given as:
E(ts)DEV(T ) = NC(T )× E(ts)DEVC(3.6)
46 Energy Consumption of Home Automation and Security Systems
Event-Driven Sensor Device
An event-driven sensor device performs tasks at irregular intervals in response to an ex-
ternally observed phenomenon (e.g. motion/door sensor devices). This is known as an
event. A power consumption time evolution of an event-driven sensor device is shown
on in Figure 3.4. Similar to its time-based counterpart, an event-driven sensor device per-
PDEV
Time (t)tMCU
1 Event
tSEN
0
Trigger
PMCU
PSEN
PSBY
Communications (COM)
Processing (MCU)
Sensing (SEN)
Standby (SBY)
tCOM
PCOM
NR
Figure 3.4: Example power consumption time evolution of an event-driven sensor deviceoperating on randomly triggered events.
forms three main tasks. In most cases, the sensing task (with time tSEN) is continuous,
so that PSEN and PSBY are indistinguishable. The diagram and the model to follow allow
for this case, and for any in which the sensing period is interrupted, for example, so as
to prevent repetitive triggers arising from the same event, as is the case with some PIR
sensor devices (see Appendix A). Alternatively, for some other event-driven sensor de-
vices, the sensing process may be a continuous low-level communication process (i.e. a
radio receiver in listening mode) that receives a trigger or poll (an event) from its parent
gateway. The sensing process can be active with relatively low power draw, or passive
with negligible power draw (e.g. PIR). When an event is detected, the MCU is activated
3.3 Home Automation and Security System Energy Model 47
to process or analyse the data received from the sensor unit in time tMCU and determines
whether or not the data should be sent off via the communication module in time tCOM.
To model an event-driven sensor, consider a device having NE random number of events
detected in a total operating time T . Let us assume the sensor unit is always in the active
state (i.e. tSEN = T and PSEN = constant), and that each triggered event is reported to
the gateway device (i.e. NMCU = NCOM = NE). Based on measurement of some event-
driven sensor devices (see Appendix A), in most cases, PSEN = PSBY or alternatively,
PSEN PSBY such that PSEN + PSBY ≈ PSEN. Assuming that the events are repetitive
with a similar amount of data being processed and transmitted, applying (3.4), the total
energy consumption of an event-driven sensor device (es), over a long operating time T is
expressed as:
E(es)DEV(T ) ≈ ESEN +NE(T )(〈EMCU〉+NR〈ECOM〉
)(3.7)
for NE > 0 and NR > 1, where ESEN = PSENT .
3.3.3 Modelling an Actuator Device
In general, the operation of an actuator contains a period of communication, a period
of processing followed by the excitement of the actuating unit. More advanced devices
may have status reporting, ACK and ”keep alive” functions. The RF 433 MHz-based ac
actuator devices (see Appendix A) measured in this work did not show discernible power
level changes for the different tasks as seen with sensor devices, because the changes are
too small compared to the measuring resolution of the ac power meter (1 sec). However,
measurement of similar RF 433 MHz transmitter and receiver modules (in Appendix A)
used by these devices show quite distinct changes for the listening, transmitting and
receiving tasks. In addition, the power consumption of the actuator device itself is also
much greater than what was measured for the RF 433 MHz receiver module.
Figure 3.5 shows a conceptual model for the power consumption time evolution of
an actuator device receiving a random command which triggers an actuating event. The
actuator device can be regarded as the dual of an event-driven sensor, with the actua-
48 Energy Consumption of Home Automation and Security Systems
PDEV
Event
tACT
tMCU
tCOM
Time (t)
Command
ListeningListening
PCOM
PMCU PACT
PSBY
Communications (COM)
Processing (MCU)
Actuate (ACT)
Standby (SBY)
0
Figure 3.5: Conceptual model of the power consumption time evolution of an actuatordevice. The vertical rise of the rectangles demonstrates instantaneous power consump-tion and the horizontal, elapsed time.
tor’s COM and SBY functions being an equivalent to the SEN and SBY functions of the
event-driven sensor, and the ACT function being equivalent to the COM function of the
event-driven sensor. An event begins with the receipt of a command signal via the com-
munication module in ”listening mode” (i.e. radio receiver is ”always-on”) during time
tCOM, with a power draw PCOM. An interpretation or processing of each command by
the MCU follows during time tMCU (and power draw PMCU), and the event ends when
the MCU triggers the actuator unit for a period tACT, drawing power PACT.
Let us consider an actuator device having NE events triggered during an operating
time T . Assume that a part of the communication module (i.e. receiver in listening mode)
is always in an active state (i.e. tCOM = T andPCOM = constant) and that every event results
in an activation of the actuating unit (i.e. NMCU =NACT =NE), whereNACT in the number
of times the actuator unit is energised. It is also assumed that, in most cases,PCOM = PSBY
or alternatively, PCOM PSBY such that PCOM + PSBY ≈ PCOM. Thus, adapting equation
(3.4), the total energy consumption of an actuator device for a long operational period T
3.3 Home Automation and Security System Energy Model 49
is therefore given as:
E(a)DEV(T ) ≈ ECOM +NE(T )(〈EMCU〉+ 〈EACT〉
)(3.8)
where NE > 0 and ECOM = PCOMT .
3.3.4 Modelling a Smart Appliance
Smart appliances have separate and different primary functions (e.g. Toaster, Oven, LED
bulbs etc.) but are fitted with network-enabled communication modules (mostly wire-
less) to control, actuate or manage their basic internal functions, for example by using a
smartphone or an IoT web service via the Internet. Smart appliances are tethered to the
main electric grid, consuming a constant standby/idle power. The energy consumed to
maintain smart appliances in the ”listening” or ”standby” state can be significant [22].
In modelling the energy consumption of smart appliances, only the additional energy
required to maintain the network-enabled communication module is included here. The
energy consumed by the appliance in performing its core functions is not considered. We
used manufacturers’ datasheet values for communication modules and reported mea-
surements in the literature.
3.3.5 Modelling Energy Consumption of IoT Gateway
This section presents an energy model for a representative IoT gateway (IGW) device
shown in Figure 3.2. A common practice is the use of popular off-the-shelf development
platforms like BeagleBone Black (BB) [15] and Raspberry Pi (RPi) [117] as an IoT gateway.
The authors in [117] conducted power measurements of a Raspberry Pi Model B device
under network load of 0 - 90 Mb/s. Their results show a small but near linear network
load dependence on power consumption. The power consumption variation was only
approximately 70 mW corresponding to a less than 5% increment when compared to
the RPi idle power of ≈ 2 W. In the project work behind this thesis, the Ninja Block
system is employed as a representative HAS [15], although this product was discontinued
during the course of this project. Our measurement of the Ninja Block IoT gateway device
50 Energy Consumption of Home Automation and Security Systems
(which used a BeagleBone Black as the gateway processor and is described in section
3.5.5) revealed an insignificant variation in the gateway power consumption under light
network load (i.e. 0.1 - 2 Mb/s).
From the general expression of the power consumption of a network element ex-
pressed as equation (2.1) in Chapter 2, the power consumption (PIGW) of a representative
IoT gateway device is:
PIGW(t) = Pidle + ERIGW(t) (3.9)
Since the IGW does not generate a significant amount of data traffic, and its idle power
(Pidle) can be 95% [117] or more of its maximum power Pmax, i.e. Pidle ≥ 0.95Pmax, then
we make the approximation E = (Pmax−Pidle)/Cmax ≈ 0. Under this approximation, the
energy consumption of an IoT gateway (EIGW) for a period of time T is given as:
EIGW =
∫ T
0PIGW(t)dt ≈ Pidle × T (3.10)
3.3.6 Modelling the Energy Consumption of a Home Gateway Modem
Generally, a network element can be classified either as a lightly shared network ele-
ment (e.g. HGW or Optical Network Unit) or heavily shared network equipment (e.g.
Core Router) as described in [70]. A lightly shared network element deals with traffic
dedicated to a single-user or single-household, like a HGW for an xDSL service. To define
a single-household in this scenario, consider an IoT service installation in a home setting
(e.g. the HAS use-case) where the service shares a HGW with a number of traffic gen-
erating services (e.g. Facebook, YouTube, etc.). In modelling the energy consumption of
the HGW, careful consideration should be given to all applications and services gaining
network access, either intermittently or continuously, over a period of time.
It is assumed that the IoT devices report throughout the day resulting in a continuous
data flow from the IGW to the HGW. This continuous data flow could hinder poten-
tial energy saving from the implementation of some sleep-mode techniques [118], as the
HGW deals with the traffic flow with little or no idle periods. This consequential po-
tential energy penalty can therefore be accounted for by allocating it to the IoT service.
3.3 Home Automation and Security System Energy Model 51
Hence, in modelling the HGW, a shared power model [70] is adopted, which allocates to
each service part of the no-load/idle power as a proportion of the total traffic, in addition
to the incremental power consumption for each service as a function of its respective data
rate R, given as:
P (R) =
(Pidle
ρRmax+Pmax − Pidle
Rmax
)R (3.11)
where Pidle and Pmax are the no-load and maximum load power consumption of the
shared network element, Rmax its maximum data rate and ρ is the percentage utilisation
of the network element.
Let us consider a HAS in a home setting, having an IGW with a data rate RIoT such that
RIoT =∑
iRdevi, for i = 1, 2, ..., N , whereRdevi
represents the data rate of an IoT device i,
and N , the total number of IoT devices connected to the IGW. Let the background traffic
generated by all other services within the home be denoted by Rbgd, given that Rbgd =∑sRsvcs , for s = 1, 2, ..., Ns, where Rsvcs represents the data rate of a unique service s
and Ns being the total number of services accessing the HGW. Hence, the aggregated
total traffic RHGW [i.e. ρRmax in (3.11)], is given as:
RHGW(t) = Rbgd(t) +RIoT(t) (3.12)
Next, we estimate the background traffic (Rbgd) of a representative single-household. To
Figure 3.6: Traffic profile of a tier-2 ISP for a week [8].
achieve this, the diurnal traffic data of a tier-2 ISP with a predominately retail customer
52 Energy Consumption of Home Automation and Security Systems
base was obtained for 1 week in May 2017 [8], as shown in Figure 3.6.
The data in Figure 3.6 is for the ensemble of customers of that ISP, and we use this
relative traffic over time as a proxy for the traffic of an ”average” customer, recognising
that each individual customer’s traffic will differ. Furthermore, this study aims to later
estimate the total IoT energy consumption over many households (e.g. USA, Europe),
hence the rationale for the use of a profile that is the sum of many households.
The data is already quantised into 0.5 hour blocks. For ease of tabulation, we further
quantise the data into 2-hour blocks, and represent each block by one of six relative traffic
levels (i.e. steps). There is nothing particularly significant in the choice of six levels; other
researchers have used similar techniques [13] with different numbers of levels.
Figure 3.7: Average diurnal traffic profile of a tier-2 ISP network with step load levelswhich are percentages of the average load over a day [8].
Using this week-long data, the average hourly traffic for one diurnal cycle is repre-
sented by the traffic profile given by the blue trace of Figure 3.7, while the 2-hour average
traffic load levels (six steps) are represented by a fitted step profile in the same figure (red
trace). The six average traffic load levels µl, are expressed as a percentage of the average
load 〈Rave〉 (i.e. averaged over a day) where l = 1, 2, ..., 6. The step profile shows a swing
from 40% of average load at the early hours of a day, to 160% of average load during the
3.3 Home Automation and Security System Energy Model 53
Table 3.1: Diurnal traffic load levels for a tier-2 ISP. The hourly average data rates for theHGW are calculated using its estimated average daily traffic and load levels (µl).
busy-hours as given in Table 3.1. These relative load levels are subsequently used as an
approximation of a single-household hourly data rate for a 24 hour period relative to the
household’s average daily data rate 〈Rave〉 in the model. Tl in Table 3.1 is the duration in
hours of traffic load level µl over 1 diurnal cycle T , where T =∑6
l=1 Tl. An expression for
the household’s HGW background traffic (Rbgdl) in interval of traffic level l is therefore
given as:
Rbgdl= µl〈Rave〉 (3.13)
The average data volume usage of all Australian fixed-line households was 168 GB
per month (assuming 10% upload volume) [119] for the last quarter of 2017. Since the
household data volume usage is increasing at an annual rate of more than 50% [119]
(doubling every 2 years), we assume a household monthly data usage of 400 GB. For
this usage volume, the calculated daily average data rate 〈Rave〉 = 1.22 Mb/s (for 12.9
GB daily usage and 31 days per calendar month). Based on this daily usage, hourly aver-
age data rate or background traffic (Rbgdl) is calculated using (3.13) and given in Table 3.1.
With the estimated HGW background traffic, using equations (3.11), (3.12) and (3.13),
the average power consumption of the HGW attributable to the HAS with average data
rate RIoT, at a specific time of day with background traffic, Rbgdl, is given by:
PHGWl(RIoT) =
Pidle ×RIoT
µl〈Rave〉+RIoT+
(Pmax − Pidle
Rmax
)RIoT (3.14)
The total energy share, EHGW, consumed by the HGW over 24 hours, that is attributable
54 Energy Consumption of Home Automation and Security Systems
to the HAS is therefore given by:
EHGW =
6∑l=1
PHGWl(RIoT)× Tl (3.15)
3.4 Experiment Methodology and Measurement Setup
A detailed explanation of the experiment methodology, the experimental set-up, mea-
surement tools and applications are given in this section.
3.4.1 Methodology
Before the commencement of any experiment or measurement, the characteristics and
functionalities of the each test device was thoroughly studied. For IoT devices, the char-
acteristics considered here include the device’s bit-rate, transmit sequence, protocol stack,
energy consumption states (i.e. active, idle, standby, etc.) and trigger mechanisms. De-
pending on the operating mode of a device (i.e. always-on or triggered), the measure-
ments were conducted after a few minutes of operation. More than 10 iterations were
taken for a complete duty cycle of each device considered. The average power consump-
tion is calculated from at least 10 iterations. The power consumption results of each IoT
device measured is given in section 3.5.1.
To measure the power consumption of the IoT devices and gateways, 2 sets of power
meters were used:
i Powermate PM10AHDS ac Power Meter
ii Custom-built USB 3V, 5V and 9V dc Power Meters
A general block diagram of the measurement setup used is shown in Figure 3.8. The test
device is powered via the power meter which measures the power draw of the device by
sensing voltage drop across a low-value resistor that the meter inserts into the line. Data
traffic measurement is achieved by connecting a PC running the packet sniffing software,
3.4 Experiment Methodology and Measurement Setup 55
”Wireshark”, to the local area network. A detailed description of the ac and dc power
meters employed in this work is given below.
Figure 3.8: Schematic diagram of the measurement setup for recording both power con-sumption and data throughput of a test device.
3.4.2 AC Power Meter
The Powermate PM10AHDS ac power meter was utilised in the measurement of the NB
gateway, the HGW modem, the IP camera and ac power actuators which operate from
the mains power. The PM10AHDS has a 0.5% accuracy error for voltage ratings of 170 -
270 V RMS and a current rating up to 10 A RMS. It can accurately measure power as low
as 10 mW with a time granularity of 1 sec. Data logging of power values is achieved via
an RS-232 port. An RS-232 to USB adapter was used for connection to the data logging
PC. Table 3.2 lists the relevant parameters of the PM10AHDS and dc power meters, which
are discussed in the next section.
56 Energy Consumption of Home Automation and Security Systems
Table 3.2: Parameters of the ac and dc power meters used in the experiments
Parameter ac Power Meter dc Power MetersSupply Voltage (V) 230 V 3 V 5 V 9 VSampling Rate 1 sec 5 ms 5 ms 5 msAccuracy 10 mW 10 µW 100 µW 10 µW
3.4.3 DC Power Meter
Four (4) dc power meters, each with an in-built power supply, were used in this work: a
3V, 9V and two 5V (low current and high current) power meters. The power supply pro-
vides a controlled voltage to the IoT devices via an internal power supply circuitry and
simultaneously measures the device current draw using a range-specific current sensing
resistor. Measurements were conducted with a 10 µW and 100 µW accuracy and with 5
Figure 3.9: dc power supply and power meter block diagram. A detailed diagram isgiven in the appendix section.
ms granularity. Time granularity was commonly 5 ms, but could be re-programmed to
a granularity of below 1 ms where that granularity was important. Figure 3.9 shows a
simple illustration of the custom built power supply and power meter setup and Table
3.2 lists the important parameters of the dc power meters. The broken lines signify the
section of the setup that applies specifically to variants of the meters that were used with
3.5 Traffic and Energy Measurements of a Ninja Block HAS - A Case Study 57
USB devices. A more detailed diagram and description of the dc power meters can be
found in the Appendix B.
3.5 Traffic and Energy Measurements of a Ninja Block HAS - ACase Study
The Ninja Block HAS (NB) [15] was selected as a representative off-the-shelf example of
typical HAS, and used as a test-bed in this work. The NB system consists of wireless
sensor and actuator devices (i.e. IoT devices), a NB gateway unit (i.e. IoT gateway) and a
cloud service - hosted by Amazon EC2 servers (data centre) - through which control and
management of all connected IoT devices is facilitated. It uses the ISM band 433 MHz
RF protocol for wireless communication with end-devices. The network architecture of
the NB is similar to that shown in Figure 3.2. A detailed functional description of the NB
system is given in Appendix A.
3.5.1 Sensor Energy Consumption Measurements
This section presents the energy consumption measurements of the different IoT devices
considered as part of the NB system tested in this work. Detailed functional description
of these sensor devices is given in Appendix A.
Temperature and Humidity Sensor
The Clas Ohlson WT450H temperature and humidity sensor (T&H) [116] is a time-based
sensor device considered here. The T&H has a transmit-only 433 MHz communications
module which sends sensor readings to the NB gateway once every minute. This device
is normally powered by a 3V battery, but for these measurements, power was supplied
by the power meter. Power consumption measurements were conducted using the 3V dc
power meter. Figure 3.10 is a power plot of several operational cycles of the T&H. The
plot shows the T&H taking 8 temperature and humidity measurement samples (includ-
58 Energy Consumption of Home Automation and Security Systems
PSEN
tSEN
PMCU
PTx
tTx
tMCU
Figure 3.10: Power plot of a Clas Ohlson temperature and humidity sensor transmittingonce per minute. The small spikes occurring every 7-8 seconds are the sensing processesfor collecting the temperature and humidity readings.
Table 3.3: Measurement values for a WT450H T&H sensor for 1 cycle (1 min)
The energy consumption of the D&W sensor device is therefore calculated using equation
(3.7) as ≈ 63 mJ per event, ≈ 7 J for 10 events per day and 12 J for 100 events per day.
Figure 3.13 gives the energy consumption of D&W sensor device for a range of events
per day.
Effect of the Number of Events/Cycles of the Sensor Devices
Figure 3.13 shows a plot of energy consumption as a function of the number of cycles or
events per day for the PIR, D&W and T&H sensor devices. The figure shows minimal
100 101 102 103 104
Number of cycles/events per day
10-2
10-1
100
101
102
103
104
Energy Consumption (Joules)
PIR Sensor
D&W Sensor
T&H Sensor
Figure 3.13: Energy consumption of sensor devices for a 24 hours period.
62 Energy Consumption of Home Automation and Security Systems
change in energy consumption for less than 10 events per day for the event-driven D&W
sensor device (6.3 - 7 J), with the sensing phases being the dominant contributor to the
total energy. Between 10 and 100 events per day, a slight increase in energy consumption
can be seen as the contribution for message transmission increases. Above 100 events
per day, the plot shows sharp increase in energy consumption of the D&W sensor device,
dominated by the message transmission energy, due to a higher number of events. The
plot of the PIR sensor device (blue curve) shows a linear increment with increase in the
number of events. This is due to the same amount of energy consumed for each event,
since there is negligible power draw in its sensing phase. For the time-based T&H sensor
(red curve), a fairly linear increment is also seen because the same amount of energy is
consumed every cycle. Therefore, for time-based sensor devices and some types of event-
driven sensor devices, a lower energy consumption can be achieved by minimising the
standby power.
100 101 102 103 104
Number of events per day
0
20
40
60
80
100
Percentage of energy (%)
D&W Sensor (sensing)
D&W Sensor (event)
Figure 3.14: Percentage of sensor device energy consumption by phase per day.
To further understand the relationship between the energy consumption of the two
major energy consuming phases of D&W sensor device, a plot of percentage energy con-
3.5 Traffic and Energy Measurements of a Ninja Block HAS - A Case Study 63
sumption as a function of number of events per day is given in Figure 3.14. From the
figure we can see that the communication (transmit phase) becomes the major contribu-
tor beyond about 100 events per day for the D&W sensor device. The reason for this is
that the daily energy consumption of the sensor is fixed while that for message transmis-
sion increases with the number of events per day. This can assist in determining sensor
battery usage prediction based on an approximate number of events over a period of
time.
3.5.3 Actuator Energy Consumption Measurements
An example actuator device that operates with the NB system is the Watts Clever Remote
Control Socket. The controlled socket is a wireless 230V power socket that can be used
Figure 3.15: Power plot of a Watts Clever controlled socket in ON/OFF no-load states.
to remotely turn on or off household appliances via the Internet. It includes a 433 MHz
wireless receiver which facilitates the actuation of a relay, opening or closing the power
circuit. Using the PowerMate ac power meter, its no-load average power consumption
was recorded as 702.4 mW in its ”Socket OFF” state and 670.1 mW in the ”Socket ON”
64 Energy Consumption of Home Automation and Security Systems
state. A power trace of the measurement given in Figure 3.15 shows the socket consuming
slightly more energy when switched ”OFF” than when switched ”ON”.
3.5.4 IP Camera Energy Measurements
The FOSCAM FI9831W IP camera (IPcam) [55] is used as a video surveillance IoT de-
vice. The IPcam is an integrated wireless video surveillance solution with an in-built
web-server that supports High Definition (HD) quality video (H.264 compression) and
images up to a maximum of 1280 x 960 resolution. The IPcam connects to the NB gateway
through Ethernet or Wi-Fi access via the HGW.
For this measurement, the IPcam is configured for a frame size of 1280 x 720 pixels at 25
frames per second. Using the ac power meter and Wireshark (see Fig 3.8), the IPcam was
tested when accessing the network via Ethernet through a Netgear DS108 dual speed
hub, and Wi-Fi through a Linksys BEFW11S4 2.4 GHz Wireless Broadband router, with
the measurement results given in Table 3.6. The 30 kb/s difference in data rate could be
Table 3.6: Data rate and power consumption measurements of the FOSCAM FI9831W IPCamera.
Network Access Data Rate (Mb/s) Power (W)Ethernet 1.90 3.25Wireless (Wi-Fi) 1.93 3.75
attributed to data re-transmissions due to a lower reliability in a wireless link, in con-
trast with a wired Ethernet connection. The notable difference in power is due to the
extra energy cost for running the wireless interface card. A more comprehensive suite of
measurements was made, and the details are included in Chapter 5.
3.5.5 IoT Gateway Traffic and Energy Consumption Measurement
The NB gateway consists of a BeagleBone Black (BB) main processor and an Arduino
microcomputer daughter-board or cape as one unit (detailed description in Appendix
A). To determine its energy profile, the gateway was set up with 3 T&H sensors, 2 PIR
sensors, a wireless electronic button (like a door bell), and 3 ac controlled socket actuators
3.5 Traffic and Energy Measurements of a Ninja Block HAS - A Case Study 65
(cameras were added later). Traffic generated by the NB gateway was observed and
recorded using the packet sniffer Wireshark, running on a test PC (see Fig 3.8) while its
power consumption was recorded on a data logging laptop.
Idle-mode Measurements
Figure 3.16 shows a component-level power consumption of the NB gateway unit as
listed in Table 3.7. Although the idle-mode power consumption of the complete NB gate-
way unit is around 1.2 W when powered via a 5V USB 2.0 port on the dc power meter,
a significant difference (2 W) was observed when the unit is mains-powered (measured
with the ac power meter) showing a powerpack inefficiency of about 40%.
Figure 3.16: Power consumption of the components of the NB gateway unit.
Table 3.7: Power consumption values of the main components of the NB gateway unit.
Ninja Block Gateway Components Power (mW)BeagleBone Black (BB) 585BB + Arduino Cape 745BB + Arduino Cape + Wi-Fi 1135BB + Arduino Cape + Ethernet 1181BB + Arduino Cape + Ethernet + ac Power Pack 2005
66 Energy Consumption of Home Automation and Security Systems
Active-mode Measurements
A description of active-mode in the context of this study refers to a successful communi-
cation link between the NB gateway, its IoT devices and the NB cloud server application,
with the IoT devices actively transmitting/receiving data. The observed traffic and corre-
sponding power consumption are listed in Table 3.8. An average power consumption of
2 W (same as idle-mode) was measured during the tests. Hence the energy consumption
of the IGW (EIGW) is calculated using equation (3.10). The observed traffic plots of the
0 20 40 60 80 100 120Time (seconds)
0
50
100
150
200
250
300
350
Data
Rate
(kb
/s)
(a)
Uplink (IPcam)
Downlink (IPcam)
Uplink (Webcam)
Downlink (Webcam)
0 20 40 60 80 100 120Time (seconds)
0
500
1000
1500
2000
2500
3000
3500
Data
(kb
yte
s)
(b)
Upload (IPcam)
Download (IPcam)
Upload (Webcam)
Download (Webcam)
0 20 40 60 80 100 120Time (seconds)
0
1
2
3
4
5
6
7
Data
Rate
(kb
/s)
(c)
Uplink (sensors)
Downlink (sensors)
0 20 40 60 80 100 120Time (seconds)
0
10
20
30
40
50
60
70
Data
(kb
yte
s)
(d)
Upload (sensors)
Download (sensors)
Figure 3.17: Observed traffic from a Ninja Block linking sensors and cameras to the cloud.Plot (a) & (b) shows the data rate (kb/s) and data volume (kb) of the cameras; Plot (c) &(d) shows the data rate (kb/s) and data volume (kb) of the sensors.
NB gateway are given in Figure 3.17 and the results recorded in Table 3.8. An average
3.5 Traffic and Energy Measurements of a Ninja Block HAS - A Case Study 67
Table 3.8: Measured average data rate and power consumption of NB gateway.
Background Traffic Level40% of daily average traffic level50% of daily average traffic level90% of daily average traffic level110% of daily average traffic level140% of daily average traffic level160% of daily average traffic level
Figure 3.19: ADSL2+ HGW power consumption attributable to the NB system traffic as afunction of its data access rate for 6 different background traffic level profiles.
3.6 Estimating the Energy Consumption of Home Automationand Security Systems
This section describes the methods applied in estimating the energy consumption of an
installed HAS for a mid-size family home, followed by a global estimate of the energy
consumption of HAS devices using market-based forecast of future numbers of installed
IoT devices. The section concludes with an estimation of the annual energy consumption
of HAS in the top two smart homes markets (i.e. North America and Europe).
3.6.1 Estimating the Energy Impact on a Mid-Size Home
As an illustrative example of the additional energy cost of a HAS and its potential impact
on global energy consumption, a simple home installation scenario is considered. An
average mid-size home is envisaged with a multi-function or whole-home installation
(i.e. sensors, actuators and IoT gateway) as an example HAS.
For this estimate, consider a 3 bedroom 2-storey house fitted with a HAS. The system
70 Energy Consumption of Home Automation and Security Systems
consists of door sensors (front, back and garage doors) having an average of 10 events per
day, door lock actuators (front and back doors), window sensors having 2 events per day,
Coffee Maker, etc.), security cameras (e.g. FOSCAM in section 3.5.4), an IoT Gateway
(Ninja Block [15]) and an ADSL2+ home gateway modem [9]. Table 3.10 lists the number
Table 3.10: IoT devices of a multi-function HAS and their annual energy consumption.The table contains a mixture of battery and mains-powered IoT devices.
IoT DeviceNo. of Events Power Energy AnnualUnits per day (W) /day (kJ) Energy (kWh)
* This value represents a share of the daily energy consumption of the HGW allocated to the HAS as afraction of its average data rate, and is calculated using equation (3.15).
of each type of IoT device, their daily and annual energy consumption (assuming 365
days per calendar year) calculated using equation (3.16).
Annual Energy Consumption (kWh) =365 Days× Energy/Day (kJ)
3600(3.16)
Energy consumption values for the 433 MHz sensors and actuators described in the pre-
ceding sections are used in this analysis. Only the standby/idle power consumption
values are considered for the mains-powered devices. For smart light bulbs, the average
3.6 Estimating the Energy Consumption of Home Automation and Security Systems 71
number of lamps per household is 24 in EU [121], 36.6 in Australia [122] and 67.4 in the
US [123]. We assume 40 smart light bulbs for a mid-size home which is about average
of the 3 regions. Since no measurements were made for the smart light bulbs and smart
appliances, manufacturer datasheets and reported measurements in the literature were
applied. The idle power consumption of a LIFX LED smart light bulb is reported as 2 W
[120]. For the smart appliances, it is assumed that all devices are fitted with a Wi-Fi mod-
ule. The idle power consumption of the 802.11n Wi-Fi module is 1 W [124]. This takes
account of the inefficiencies of many power supply units which can be more than 30%
(see Figure 3.16).
(a) (b)
Figure 3.20: Percentage share of annual energy consumption of a HAS when the installedIoT devices are: (a) Mains and battery powered, and (b) Mains powered only.
Two scenarios were considered: (a) Mains & battery powered devices and (b) Mains pow-
ered devices only.
For scenario (a), all sensors are battery powered while all other devices are mains
powered. Using the data from Table 3.10 and applying equation (3.1), the calculated an-
nual additional energy consumption of this mid-size home installed with an off-the-shelf
HAS is ≈ 1,200 kWh. This corresponds to about 30% of the annual energy consumption
of a ”typical” mid-size (3-bedroom) suburban home in Victoria, Australia, estimated as
4,067 kWh [125]. The results are given in Table 3.11.
For scenario (b), all devices are mains powered. The average power consumption
of sensors and actuators were obtained from the recent IEA report [22] (sensor = 0.6 W;
72 Energy Consumption of Home Automation and Security Systems
(a)
0
100
200
300
400
500
600
700
800
Ener
gy C
on
sum
pti
on
(kW
h)
0
100
200
300
400
500
600
700
800
Ener
gy C
on
sum
pti
on
(kW
h)
(b)
Figure 3.21: Annual energy consumption by IoT devices type for a: (a) Mains and batterypowered HAS, and (b) Mains powered HAS.
Table 3.11: Annual energy consumption of a multi-function HAS installed in a mid-sizehousehold.
HAS System Annual Energy % of Ave. Mid-SizeConsumption (kWh) Household Energy Usage
Mains and Battery Powered 1,200 30%Mains Powered 1,434 35%
actuator = 1 W). The calculated annual energy consumption is slightly higher at 1,434
kWh, corresponding to just over one-third (≈ 35%) the annual energy consumption of
the same home. Pie charts breaking down the percentage share of each IoT device type
in both scenarios are given in Figure 3.20(a) and 3.20(b). The main difference between
the two charts is due to a lower share of energy consumed by the sensors when they
are battery powered as opposed to mains powered, as can be seen from Figure 3.21. This
difference may be relatively trivial for a single-household but is significant on a global scale
as demonstrated in section 3.6.2. However, for a ”fair” comparison, the battery life and
whole-of-life energy cost of the batteries should be included, but this information is not
readily available.
Potential Energy Savings
Smart light bulbs (LED) are clearly the least energy-efficient from the charts above and
will potentially reduce the energy savings achieved from replacing incandescent light
bulbs with their LED equivalent. To put this in perspective, a 60 W incandescent light
3.6 Estimating the Energy Consumption of Home Automation and Security Systems 73
bulb switched on for 4 hours/day consumes 87.6 kWh/year but a 10 W non-smart LED
equivalent consumes 14.5 kWh/year with ≈83% energy savings. However, a 10 W smart
LED equivalent having 2 W standby power (e.g. LIFX) consumes 32.1 kWh over the same
usage time, resulting in ≈63% energy savings only, the difference of which is the energy
cost of its ’smartness’. Beyond the ability to automate or remotely power on/off smart
light bulbs, which would depend on daily usage behaviours, it is unclear how much
energy savings there may be when compared with non-smart LED bulbs.
Other smart devices (e.g. Smart Washing Machine) may also provide discernible en-
ergy savings with the ability, for example, to reschedule their operation at an opportune
time, in coordination with a smart grid. Where such applications can be identified, the
absolute energy savings due to its IoT smart technology would need to be studied sepa-
rately and assessed to establish the overall energy benefit of the smartness.
3.6.2 Estimating the Global Energy Impact of HAS System Devices
Global forecasts of IoT device deployments tend to be coarse with very wide variations
between the views of different consultants [2,4,126,127]. To estimate the energy impact of
HAS deployment, market forecasts from different sources were used. ABI Research fore-
casts - used in the 2016 IEA Report on energy-efficiency of the IoT [22] - numbers of HAS
devices to grow from 700 million in 2015 to 1.9 billion by 2020 with a Compound Annual
Growth Rate (CAGR) of 22%. Cisco’s 2016 Zettabyte Era report forecasts a growth in HAS
devices from 2.4 billion in 2015 to 5.7 billion in 2020, a CAGR of 19.5% [28]. While the
more conservative ABI Research forecast forms the basis of this work, the Cisco forecast
is also applied to show a potential worst-case scenario or upper bound of recent forecasts.
To estimate the global energy impact over a longer time span, the ABI Research fore-
cast was extrapolated to 2025, giving a 10-year estimate of 5.1 billion devices, assuming
the same CAGR. For a fully-equipped HAS as given in Table 3.10, the distribution of
installed devices by device type (e.g. sensor, IP camera) is shown in Figure 3.22(a). Us-
ing the energy consumption values in Table 3.10, assuming mains powered sensors (unit
74 Energy Consumption of Home Automation and Security Systems
!"#$%&
%
(a)
!"#$%&
%
(b)
Figure 3.22: (a) Global forecast of installed HAS devices (2015 - 2025); (b) Global energyconsumption estimate of HAS from 2015 - 2025.
standby power = 0.6 W [22]), the global energy consumption for HAS is estimated to rise
from 7.8 TWh in 2015 to 57 TWh in 2025 as depicted in Figure 3.22(b).
Applying the same principle to the Cisco projection, the data was extrapolated to
2025 using the same CAGR giving a device number increase to 14 billion by 2025 (see
Figure 3.23(a)). The global energy consumption is estimated to increase from 26 TWh
in 2015 to 156 TWh in 2025 as shown in Figure 3.23(b). The disparity is considerable as
the Cisco projection is almost 3-fold higher than that of ABI Research, which shows the
coarse nature of today’s estimates.
To further demonstrate the significant energy impact of installing mains powered sen-
sors over battery-powered sensors as mentioned in section 3.6.1, consider the above ex-
(a) (b)
Figure 3.23: (a) Cisco’s global forecast of installed HAS devices; (b) Global energy con-sumption estimate of HAS.
3.6 Estimating the Energy Consumption of Home Automation and Security Systems 75
trapolated ABI Research projections. The energy consumption for mains and battery
powered devices as compared to mains powered is depicted in Figure 3.24. The plot
shows a potential energy savings of 11.7 TWh for battery over mains powered sensors.
This analysis however does not take into account the energy to manufacture and safely
dispose of the batteries. Neither does this illustrative example include potential reduc-
tions from the development of more energy efficient IoT devices in the future. From
Figures 3.20 and 3.21, the stand-out example of where more efficient devices would im-
prove the efficiency of such an extensive IoT implementation would be the development
of more efficient ”smart light bulbs.”
3.6.3 Estimating the Energy Impact of Smart Homes
A recent market-based report on smart homes by Berg Insight focuses on smart homes
or HAS that have smartphone apps or web portal as a user interface, thus excluding
legacy systems. Focusing on two major smart home markets (i.e. North America (NA)
and Europe), the report forecasts installed smart homes in North America to grow from
16.9 million in 2015 to 46.2 million in 2020, a CAGR of 30%. Similarly, a 54% CAGR in
Figure 3.24: Annual energy consumption estimate for mains-powered HAS devices vs.mains and battery powered HAS devices.
76 Energy Consumption of Home Automation and Security Systems
Europe (EU28+2) increasing from 6.6 million smart homes in 2015 to 44.9 million smart
homes in 2020. Since these projections end at 2020, an extrapolation of the data was done
for 2021 to 2025 but with half the CAGR - using the same CAGR would exceed the total
number of households. Hence, a 10-year forecast with an effective CAGR of ≈ 22% and
≈ 39% shows the number of smart homes in NA and EU28+2 to reach 92 million and 146
million respectively. To estimate the smart homes’ energy impact, an assumption of the
Table 3.12: Deployment scenarios of an IoT HAS.
HAS System DeploymentDevices Large-Scale Medium-Scale Small-ScaleSensors 44 25 10Actuators 20 10 4Smart LED Lamps 40 20 5Smart Appliances 15 6 3IP Camera 4 2 1IoT Gateway 1 1 1
number of devices per household is required. From the Berg Insight report, it was shown
that between 12 - 17% of the 2015 installed smart home systems were multi-function or
whole-home systems while the remainder were point solutions (i.e. function specific). To
cater for this skewed adoption pattern, 3 deployment scenarios (i.e. large, medium and
small) were proposed as given in Table 3.12. The large-scale deployment is effectively a
multi-function, whole-home HAS similar to that in Table 3.10, assuming mains-powered
sensors (average power = 0.6 W [22]).
The annual energy consumption of a large, medium and small-scale smart home sys-
tem, calculated using equation (3.1) and energy consumption values from Table 3.10, is
given as 1434 kWh, 730 kWh and 263 kWh respectively. Assuming the 3 deployment
scenarios are distributed such that 20% of smart homes are installed with large scale sys-
tems, 40% with medium-scale and 40% with small scale, the total energy consumption
estimate for smart home systems in NA and EU28+2 is given in Figure 3.25. Figure 3.25
shows that the additional energy consumption estimate for potential smart homes instal-
lations in NA and EU28+2 is scheduled to increase from 8.7 TWh and 3.6 TWh in 2015
to 63 TWh and 100 TWh in a decade, if installation rates keep up at the projected pace.
To put this in perspective, it will require about nine (9) large power plants [128] (2 GW
3.7 Discussion and Conclusions 77
Figure 3.25: Annual energy consumption estimate for smart home systems in NorthAmerica and Europe.
output power) to support smart home systems only in the two regions considered. Also,
the data suggest EU28+2 is on track to become the most users of smart home systems
worldwide.
3.7 Discussion and Conclusions
In this chapter, a measurement-based first-order estimate of the potential energy impact
of home automation and security systems has been presented. An example HAS was
modelled for a ”typical” mid-size household. A step-by-step approach of energy mod-
elling of the different components of the HAS is presented, starting with the sensors,
actuators, IP camera, down to the IoT gateway and home gateway/modem. A traffic-
based proportion of the HGW energy consumption is allotted to the HAS because it is
shared by other user applications and services in the home.
Measurement results for sensors indicate two main types: ones with autonomous
and regular reporting and others triggered by events. This difference is critical in un-
derstanding the energy usage patterns of IoT devices, especially the battery powered
devices. The results indicate HAS exhibit low data throughput (< 10 kb/s) when sen-
78 Energy Consumption of Home Automation and Security Systems
sors, actuators and other smart devices are connected but moderate data throughput for
an image streaming service if more than one camera is connected. Energy consumption
estimates of a complete HAS with mains and battery powered IoT devices could be as
high as 30% more than the average annual energy consumption of a typical mid-size
suburban household. This number could increase to about 35% (≈ 1,434 kWh) if the
entire HAS is mains-powered. Smart light bulbs stood out as the least energy efficient
among the IoT devices considered. Table 3.10 and Figure 3.20 showed that the ”smart-
ness” of smart LED bulbs and smart power sockets were the largest contributors to this
energy consumption growth. Smarter design of these products should be the first focus
for product development, and the balance between the potential energy saving through
”smartness” as against the energy cost of their ”smartness” considered carefully in their
deployment.
The potential global energy impact of HAS device deployment is non-trivial. From a
recent market-based device shipment forecast, an energy consumption between 57 TWh
and 156 TWh is possible by 2025. This could require up to 9 large power plants [128]
(2 GW power output) to support. However, the results indicate an opportunity for poten-
tial savings of about 12 TWh if battery powered sensors are chosen over mains powered
sensors. This must be explored further as the analyses provided here does not consider
embedded energy.
While this work presents a first-order estimate of the potential energy impact of one
IoT application use-case, it gives an indication that there is a price to pay if the market
projections of IoT devices by 2025 [2–4, 126] are to come to fruition. One caveat is the
expectation that more efficient energy usage behaviours could be achieved by the use of
the data and control capabilities enabled by the IoT. These potential energy savings have
not been considered here.
Chapter 4
Energy Consumption of IoT WirelessNetwork Protocols
4.1 Introduction
AS discussed in Chapters 2 and 3, the Internet of Things (IoT) [24] is not just about
new sensors and actuators but will include many consumer devices/physical ob-
jects that are not currently networked and do not draw electricity when they are not in
use. Bringing these devices into the IoT eco-system will require an addition of communi-
cations interfaces which will mostly be wireless. A naive approach to adding communica-
tions interfaces (e.g. a Wi-Fi module on each device) could lead to substantial additional
energy consumption without corresponding benefit in functionality, if an inappropriate
design option is chosen. While this effect may be marginal for a single device, energy
efficiency in communications design will be important for individual households and
in aggregate may have considerable effect, as discussed in Chapter 3 (for example, see
chapter 14 of [129]).
A variety of wireless communications interfaces have been considered as enablers of
the IoT with a wide range of characteristics (e.g. bit rate, topology and energy consump-
tion) as discussed in Chapter 2. In this chapter, we compare the energy consumption char-
acteristics of five common consumer off-the-shelf (COTS) wireless interfaces while con-
sidering three different communication paradigms (event-driven, broadcast and polling).
The five wireless interfaces considered here represent today’s dominant COTS equipment
79
80 Energy Consumption of IoT Wireless Network Protocols
for IoT application. They include Bluetooth Classic (BT), Bluetooth Low Energy (BLE),
ZigBee, Wi-Fi and 433 MHz module (referred to as RF433 in this study). Power consump-
tion measurement of the COTS interface options was conducted for their various states
of operation, and utilised to determine their energy use in different scenarios.
To examine the options for energy-efficient communication using the COTS devices,
we employ a simple stock control application involving a domestic toaster communi-
cating with a gateway hosting an inventory system. This application is rich enough
to encompass the full range of design options while being simple enough to enable a
straightforward analysis of energy consumption. The application choice is in-line with
developing interest in ’smart kitchens’ [130] and automated stock control for online re-
ordering of basic grocery items [131, 132] as part of assisted living. Also, a toaster is now
considered as the first ‘thing’ of the Internet-of-Things [133].
While the example given here may be somewhat contrived, it illustrates the point that
careful design for energy usage and efficiency in the IoT will be important. The type and
volume of data, the frequency of data transmission, and the type of communications in-
terface should all be considered. The chapter is organised as follows: Section 4.2 gives a
brief description of the five wireless network communication protocols employed in this
study. Section 4.3 describes the measurement setup and presents the RF modules used.
The power consumption measurements are given in Section 4.4 followed by a descrip-
tion of the domestic stock control IoT application in Section 4.5. The energy consumption
model is given in Section 4.6 and a comparison of energy consumption discussed in Sec-
tion 4.7. Section 4.8 discusses energy-efficient design based on the results and Section 4.9
concludes the chapter.
4.2 Wireless Network Communication Protocols
In Chapter 2, a number of wireless network communication protocols that can be consid-
ered a part of the enabling technologies of the IoT are presented. In this section, a brief
description of the five commonly employed wireless network communication protocols
for IoT applications is given. They include Bluetooth Classic, Bluetooth Low Energy,
4.2 Wireless Network Communication Protocols 81
Wi-Fi, ZigBee and Radio Frequency 433 MHz (RF433).
4.2.1 Bluetooth Classic
Bluetooth technology, based on the IEEE 802.15.1 standard, is designed to support short-
range, ad-hoc connectivity amongst devices. Traditional or Bluetooth Classic (BT) oper-
ates in the license-free 2.4 GHz Industrial Scientific and Medical (ISM) band using adap-
tive Frequency Hopping Spread Spectrum (FHSS) channel access (1600 hops/sec), with
79 defined channels of 1 MHz channel width, 32 of which are used for device discovery
[34]. BT devices can operate either as a master or slave with one master interconnecting
up to seven active slave devices - hence a star topology. BT is capable of data rates up
to 2 Mb/s. However, BT has some major disadvantages which limits widespread im-
plementation in IoT applications. These disadvantages include a relatively high power
consumption, large data packets with huge overhead, complex protocol stack with large
memory demand, long connection time and limited range (∼ 10 m) [134]. Additionally,
while BT has some low power modes (e.g. sniff mode), it lacks an effective sleep-mode
regime which is critical for minimising energy consumption for battery-operated IoT de-
vices.
4.2.2 Bluetooth Low Energy
Bluetooth Low Energy (BLE) is defined in the Bluetooth Specification 4.0 [135] as the low
energy version of BT, designed for IoT-like applications - low cost and complexity, low
duty-cycle and infrequent transmission. Like BT, BLE (or Bluetooth Smart) is based on
the 2.4 GHz ISM band and uses FHSS but with 40 channels (including 3 advertisement
channels for device discovery) of 2 MHz channel width, and a data rate of 1 Mb/s. BLE
can either operate as a central (master) or peripheral (slave) device, with a slave able to
connect to only one master device at a time (i.e. star topology). An important feature
of BLE is its sleep-mode capability which, in combination with a low duty-cycle appli-
cation, significantly reduces the device energy consumption. A slave may broadcast a
limited amount of data in an advertisement packet, or a communications link between
82 Energy Consumption of IoT Wireless Network Protocols
a master device and a slave device may be established for dedicated or higher volume
traffic applications. While BLE has a longer range (∼ 30 - 50 m) than BT, its deployment
is limited to IoT applications with short-range requirements.
4.2.3 Wi-Fi
Based on the IEEE 802.11 standard for wireless local area network (WLAN), Wi-Fi op-
erates on multiple frequency bands (e.g. 2.4 GHz, 5 GHz, 60 GHz) and is a widely
used short-range wireless protocol. The Wi-Fi protocol has undergone several revisions
(802.11a/b/g/n/ac) with its latest version (802.11ac) capable of achieving broadband
speeds in excess of 500 Mb/s [136]. Wi-Fi supports two operating modes in its topology
formation: Peer-to-Peer (Ad-hoc) network topology and Star (Infrastructure) network
topology. While Wi-Fi was originally designed for high-speed, short-range (up to 100 m)
communication, its ubiquity in homes and places of work has resulted in an increasing
use in a number of IoT applications (e.g. Smart Light Bulbs or Smart Appliances).
4.2.4 ZigBee
Based on the IEEE 802.15.4 standard, which defines the PHY and MAC layers, ZigBee has
an established set of specifications for low power, low data-rate and short-range appli-
cations. The ZigBee Alliance defines the Network and Application Support layers [18].
ZigBee operates on the license-free 868 MHz, 915 MHz and 2.4 GHz bands with data
rates of 20 kb/s, 40 kb/s and 250 kb/s, respectively. In addition to its star and cluster-
tree networking capabilities, the ZigBee protocol can be configured for mesh networking
which offers long reach by means of data relay. ZigBee defines three types of devices:
(a) ZigBee Coordinator - acts as a gateway/hub and is the most resourceful, (b) ZigBee
Router (ZR) - acts as intermediate router in addition to application functions, (c) ZigBee
End-Device (ZED) - performs application functions and is the least resourceful device.
ZC and ZR are classified as fully functional devices while ZED is a reduced functional
device. The ZigBee protocol was designed for low complexity and low energy usage with
a sleep-mode capability.
4.3 Measurement Setup 83
4.2.5 RF433
Based on the ISM 433 MHz band, RF433 can be used for wireless data transmission, in-
cluding in wireless sensors and home automation systems (see Chapter 3). Its frequency
range is 433.05 - 434.79 MHz. RF433 is not a standardised protocol like Bluetooth, Zig-
Bee and Wi-Fi but is a widely employed, inexpensive wireless communication option for
many IoT devices and applications. Based on this backdrop and the general belief that
RF433 will be a common hardware selection in future IoT applications, it is included in
this study.
4.3 Measurement Setup
In this study, we measure the power consumption (energy consumption) characteristics
of each of the above communication technologies, in order to be able to compare their
performance in typical applications. A description of the experiment and measurement
setup is given in this section. It is assumed that each COTS wireless interface commu-
nicates with an always-on gateway device (i.e. master, ZC, etc.) that is common to all
scenarios.
4.3.1 Description of Measurement Setup
The components used in the measurements include two COTS short-range RF modules
for each communication protocol, an Arduino Duemilanove (Due) board [137] equipped
with Atmel’s Atmega 328P-PU microcontroler (MCU), a dc power supply and power
meter unit (described in detail in Chapter 3), a test computer (PC) and smartphone.
A block diagram of the experiment setup is shown in Figure 4.1. Earlier RF mod-
ules might include separated transceiver circuitry, antenna, MCU and serial interface but
modern RF modules today are mostly designed as a System-on-Chip (SoC) with inte-
grated transceiver module and MCU. Hence, the RF modules in this study are measured
as one SoC unit, given that, in practice, they may be deployed as such.
The power meter records power consumption, current draw (mA) and voltage (V)
84 Energy Consumption of IoT Wireless Network Protocols
RF Module
Arduino DuePC/
Smartphone
Mains (240V)
Power Meter
Slave/End-device Master/Coordinator
2.4 GHz / 433 MHz
RF ModulePower Supply
Measurement Unit
Data Logging PC
(DUT)
(Data Sink)(Data Source)
Figure 4.1: Block diagram of experiment setup
of the device under test (DUT)/end-device, and the measurements are logged on a PC
for post-processing. Measurements are recorded at an average of 1 ms intervals with an
accuracy of 10 µA. With the exception of the RF433 modules, which are either transmit-
only or receive-only, each RF module has 4 leads connected: +ve power input (VCC),
Ground (GND), transmit (Tx) and receive (Rx). Figure 4.2 displays an image showing
Arduino Duemilanove Board
XBee ZigBee RF Module
Power Supply & Power Meter
USB Cable to PC(Arduino Power)
GND
Tx
Rx
GND
VCC
Figure 4.2: Image of measurement setup for a XBee ZigBee module
4.3 Measurement Setup 85
this setup with an XBee ZigBee RF module as an example. The same principle applies
to DUT modules with USB connection. For accurate measurements of the DUT power
consumption, the VCC and GND leads are directly connected to the dc power supply via
a power meter, while the Tx and Rx leads are either connected to the transmit and receive
rail (terminals 1 & 2) on the Arduino board or a USB port. A steady voltage of 5V (or 3.3V
where applicable) is supplied to the modules during the tests.
The master/coordinator (gateway device) modules are powered from the power rails
on an Arduino board, the USB interface of a PC or the smartphone battery. Packets re-
ceived from the DUT are recorded with timestamps. Measurements are collected over a
5 minutes duration. In one case a digital oscilloscope (e.g. Digilent Analog Discovery
USB Oscilloscope) in a setup with a 10 Ω current metering shunt resistor was used for the
measurement.
In the experiment, the DUT - positioned about 1 meter away from the master/ coor-
dinator - is programmed to send the pre-set 10-byte (10B) data payload every 5 seconds.
The master/coordinator receives and displays the transmitted data on a serial monitor
using a terminal emulator such as Tera Term or Bluetooth Smart Data app. The Arduino
Due generates the 10B test data and is independently powered (not measured).
4.3.2 RF Module Power Measurement Setup
Table 4.1 gives details of the RF modules used in the experiments. For BT measurements,
a JY-MCU HC-06 module, compliant with Bluetooth v2.1 standard was used. The serial
interface of the BT DUT (slave) was configured at 9600 Baud. Bluetooth v2.0 standard
lacks the very low power sleep-mode; therefore, it is assumed that the module, when not
operated continuously, is turned off after every Tx and Rx operation of the DUT.
For the measurement of BLE, the Redbear BLE Nano v2 [138] was used. The BLE
Nano is a breakout board that simplifies prototyping or small-scale production of new
IoT products, based on the Nordic nRF52832. The latter is a SoC including the RF module
and an ARM Cortex-M4F SoC at its core. The BLE Nano comes with Bluetooth v4.2
protocol stack (Bluetooth 5 ready) and can be configured as a central or peripheral device.
86 Energy Consumption of IoT Wireless Network Protocols
Table 4.1: RF modules used in experiment.
Standard Device RF Module RF Module Firmware OperatingName Manufacturer Number Version Voltage (V)
Bluetooth JY-MCU Linvor HC-06 1.8 5Bluetooth LE BLE Nano v2 Nordic nRF52832 - 3.3
Table 4.3: Duration of operation in different phases.
RF ModulePhase Duration (ms)Wake-up Tx Rx
Bluetooth 1836 1 1Bluetooth LE 4 1 1
ZigBee 5 2.5 3.5Wi-Fi 3165 1 1RF433 1 126 406
88 Energy Consumption of IoT Wireless Network Protocols
4.4.1 BT Measurements
A Bluetooth link controller has a number of connection states including standby, inquiry
(scan/advertise), paging, connected/active, sniff, park and hold states [16]. Figure 4.3(a)
shows the power consumption of the BT module in a non-connected standby state de-
picting the advertisement and scanning phases. The device sends advertisements on
Advertisements Scanning
(a)
Standby Pairing Connected (Active) Sniff State (Low Power)
(b)
Figure 4.3: Power trace of BT module (a) in standby state showing the scanning and ad-vertising phases and (b) showing the state transitions from standby to pairing, connectedand sniff states.
different channels, each round lasting about 5 ms. During advertisement, the BT module
draws an average of 79 mW, consuming about 0.4 mJ each round. During scanning with
nearly 100% duty cycle (duration≈ 320 ms), the module draws about 208 mW consuming
67 mJ, which is over 50 times more than its advertisement phase. Such a huge disparity
in energy usage explains the reasoning for slave/end-devices - which are often energy
constrained - being limited to sending advertisements while master devices (often with
more resources, e.g. PC/ Smartphone) engage in scanning during the discovery process.
Figure 4.3(b) gives a much wider view of the major stages a BT device undergoes.
The figure shows the standby state, when scanning and advertisements are performed,
the pairing stage, when paging and synchronisation process occurs, the connected stage,
which includes data Tx, Rx and ACK, and lastly the low-power sniff state (i.e. idle),
when the module is less-active but listens for transmissions at set intervals (125 ms in
this case), depicted by the spikes on the right-hand-side of the figure. The average power
4.4 Power Consumption Measurement 89
consumption in the low-power sniff mode is about 16 mW. For this power draw, an Alka-
line AA battery of 1800 mAh capacity [140] can power this module for about 23 days only
in its low-power state. BT communication transmits complete data frames per timeslot
(1 timeslot = duration in 1 frequency hop) in 0.625 ms, with an ACK packet received in
a subsequent timeslot. Since the power meter can only resolve variations down to 1 ms,
the minimum phase duration is assumed as such.
4.4.2 BLE Measurement
The measurement was conducted for the BLE non-connected and connected/active states.
Figure 4.4(a) shows the power consumption trace of a BLE peripheral device advertising
at 200 ms intervals and in connected mode with connection events depicted on the right-
hand-side. Each advertisement payload was 30B long and was transmitted to three ad-
AdvertisementsConnection Events
(a)
Scanning
(b)
Post-processing
Figure 4.4: Power consumption trace of BLE module configured as (a) peripheral (slave)device showing advertisements and connection events (b) central (master) device in scan-ning mode.
vertisement channels. For one advertisement round, the peripheral consumed about 0.08
mJ, lasting about 2 ms (0.625 ms per timeslot). If an example BLE peripheral module only
broadcasts its advertisement packet once every second, a 1800 mAh AA battery [140] can
power the module for over 2040 days.
Figure 4.4(b) shows the power consumption trace of a non-connected BLE central
device scanning three advertisement channels with a 20% duty cycle (i.e. scan window
90 Energy Consumption of IoT Wireless Network Protocols
= 200 ms, scan interval = 1s). The steps in Figure 4.4(b) can be attributed to its post-
processing activities drawing ≈ 23 mW. For scanning the central consumes 14.4 mJ per
cycle. The energy consumption of a BLE module in its non-connected states ultimately
depends on the choice of a number of parameters (e.g. connection interval) based on an
application usage scenario. The measured power draw and duration of BLE connected
phases are given in Tables 4.2 and 4.3.
4.4.3 ZigBee Measurement
With ZigBee, the measurement was carried out on a configured XBee ZigBee end-device
as the DUT, which connects to an XBee ZigBee coordinator (i.e. gateway). The DUT was
configured for cyclic sleep-mode with reverse polling. In reverse polling, the end-device
(in sleep-mode) wakes up at regular intervals (default = 100 ms but can be modified in
firmware) and polls its coordinator to request any data sent to its address while sleeping.
The end-device sends a poll once after its wake-up sequence and again before going to
sleep. In sleep-mode, the XBee draws a maximum of 50 µA at 3.3V [139]. Figure 4.5 shows
CSMA/CA
(RX Mode)Rx
(ACK)
Tx
Wake-up
Sequence
Sleep
Rx to Tx & Tx
to Rx switch
(100 µs)MCU Shutdown
Sequence
Time (ms)
Cu
rren
t (m
A)
Shutdown pollData
Sleep
6 12 18 24 30 36 42 480-6 54
10
20
30
40
50
60
70
0
-10
-20
-30
Figure 4.5: Current draw of the XBee ZigBee end-device module during a connectionevent.
a detailed plot of a ZigBee end-device module current draw for one connection event.
4.4 Power Consumption Measurement 91
The plot shows the wake-up sequence which includes MCU wake-up, pre-processing
and synchronisation. The initial poll is shown in the next phase, starting with channel
access process, which is the contention based CSMA/CA as defined by the 802.15.4 MAC
specification. The duration of the CSMA/CA process may vary depending on channel
availability. The power consumption and duration values are reported in Tables 4.2 and
4.3 for an operating voltage of 3.3V. A similar example ZigBee module transmitting a
message once every minute, and is powered by a 1800 mAh AA battery would operate
for about 953 days.
4.4.4 Wi-Fi Measurement
In measuring the Wi-Fi module, an infrastructure operating mode was used. The Wi-Fi
module and the AP were configured for the IEEE 802.11g standard, with a maximum
theoretical data rate of 54 Mb/s (highest of all the RF modules considered). At start-up
the Wi-Fi module scans the channels for periodic beacons from the AP before connection
process is initiated. The measured duration for establishing a connection was about 3165
ms. The Wi-Fi module power draw in different states is reported in Table 4.2 and their
duration in Table 4.3. As with BT, the phase durations are assumed to be the limit of the
power meter (1 ms). We note that the idle power consumption of Wi-Fi (61 mW) makes it
unreasonable to use batteries. The same 1800 mAh AA battery described in the previous
section will operate this module for nearly 6 days only when in the idle state.
4.4.5 RF433 Measurement
The RF433 transmitter and receiver modules were measured separately. Figure 4.6(a)
shows the power trace of the 433 MHz Tx module sending 10B of data and Figure 4.6(b)
shows the Rx module receiving the same amount of data at 5 sec intervals.
The transmitter module consumes a small amount of power (≈ 0.03 mW) when idle
and about 45 mW on average when transmitting. This is depicted by the spikes in Fig-
ure 4.6(a), indicative of its on-off keying modulation. Tables 4.2 and 4.3 list the power
draw and duration values of the RF433 modules. Because the Tx and Rx modules lack an
92 Energy Consumption of IoT Wireless Network Protocols
(a)
TX
(b)
RX Listening RX RXListening
Data (RX)
Post-processing
Figure 4.6: Power trace of an RF433 (a) transmitter module sending a 10B data (b) receivermodule in receiving the same amount of data.
integrated MCU, there is no substantial wake-up time for the transmitter and very little
for the receiver (about 3 ms). The receiver module however maintains a steady but rela-
tively higher power consumption level (≈ 29 mW) when idle (i.e. listening mode) as can
be seen from Figure 4.6(b). The RF433 module has the longest over-the-air transmission
time due to its low data transfer rate (9.6 kb/s max) which is several orders of magnitude
less than that of the other RF modules. Furthermore, while the Tx and Rx duration are
similar for the same data transfer, the receiver remained at a higher power level for an
additional duration of about 280 ms for post-processing or data verification. A combined
transmitter and receiver pair power by a 1800 mAh AA battery will operate for nearly 13
days.
4.5 Domestic Stock Control IoT Application - A Case Study
4.5.1 Application Architecture
The basic application architecture of the IoT application is considered to include a large
number of end devices (sensors and domestic appliances), communicating via a short-
range wireless medium to a gateway device (eg. a ’home gateway’) and being controlled
by that gateway. In some cases, the home gateway will communicate through an external
network with cloud-based applications. While the popular wireless interface options
4.5 Domestic Stock Control IoT Application - A Case Study 93
discussed in the preceding sections have propagation or performance characteristics that
may be important in particular applications, the focus of this case study is on the energy
efficiency of each option in the context of a domestic operational paradigm.
Let us consider a simple domestic stock control application, envisaging a toaster re-
porting to the home gateway the number of bread slices it has toasted up to a particular
time. Coupled with reporting of the bread quantity when it first enters the household
(and other events), this can enable automated stock control and replenishment of bread
supply by a commercial supplier.
To support this application, a wireless communications interface must be added to
the toaster assuming the home gateway is already equipped with suitable interfaces. The
toaster interface will consume power in addition to the normal operation of the toaster
and will, in some circumstances, add a few percentage points to the power consumption
of the toaster. For this application, it is assumed that the gateway is within the ’smart
kitchen’ so that the communication range is a few metres at most. The toaster need only
report the number of slices of toast in a given period, so the payload is only a few bytes.
All the wireless options in the experiments are capable of providing the necessary range
and throughput. The energy consumption of the application over a period of one week
is considered.
4.5.2 Communication Paradigms
There are three basic communications paradigms which have different energy implica-
tions. These are:
(i) Broadcast
(ii) Polling
(iii) Event-driven
Broadcast Mode
In Broadcast mode, the toaster reports its status (number of bread slices toasted) at reg-
ular intervals (e.g. every minute). The communications interface of the toaster will be
94 Energy Consumption of IoT Wireless Network Protocols
active periodically, perhaps frequently. Because the application will require reliable com-
munications between the toaster and the gateway, the communications interface will re-
main active after wake-up from its sleep/off state until an acknowledgement has been
successfully received (if possible). It can then return to an inactive ’sleep’ state until the
next communications cycle. In Broadcast mode, the natural cycle adopted for this appli-
cation is hourly: that is, the stock control algorithm determines when more bread will be
required and schedules a delivery at some future hour.
Polling Mode
In Polling mode, the gateway asks the toaster at particular times to report its status and
the toaster does so. The communications interface should be in a listening state at all
times and should become fully active when a polling request is received. When respond-
ing to a polling request, the communications interface is fully active until an acknowl-
edgement has been successfully received. It can then return to a listening state. The
Polling mode is assumed to be governed by a standard stock-control application in which
the economic order quantity (EoQ) is proportional to the square root of the demand, i.e.
EoQ = k√D, where D is the demand and k is a constant (see, for example, [141]). Figure
4.7 shows a plot of polling instances against demand. The red curve is a scaled square
root of demand and the blue curve, the average number of polling instances. Given the
assumption of EoQ, the next polling event is set at one-half of the estimated time to ex-
haustion of bread supply, with the restriction that polling should be no more frequent
than once per hour and no less frequent than once per day. This is a minimal polling
design: a ’simpler’ design (e.g. polling once per hour) could generate many more polling
events. In this design, the toaster is polled only once per day for low numbers of toasting
events. The number of polling events increases only as the square root of demand and,
with a suitable value of k, can be kept below 25 polling events per week even for large
numbers of toasting events.
4.6 Energy Consumption Model 95
Square root of demand (scaled)
Average number of polling instances
Figure 4.7: Plot of polling instances against demand.
Event-driven Mode
In Event-driven mode, the toaster reports only when a toasting operation begins. The
toaster interface becomes fully active when new bread is introduced to the toaster. It
remains active until an acknowledgement has been successfully received (or communi-
cation is deemed successful) before returning to inactive state. The number of events is
usage driven.
Which of these communications paradigms uses least energy for a specific wireless
interface depends both on the relative energy usage of the interface in its various states
and on the frequency of use, that is, on the traffic at the toaster. No one communications
paradigm is optimal over the full traffic range.
4.6 Energy Consumption Model
The main operational modes/states of the wireless interfaces (RF modules) for which
power consumption measurement was conducted include:
• Inactive or sleep mode - in cases without explicit sleep mode, this is when the interface
96 Energy Consumption of IoT Wireless Network Protocols
is turned off.
• Wake-up mode - Transitions between sleep and active and from active to inactive.
• Listening/Idle mode - when the interface can detect and receive a message but may
not be able to transmit.
• Transmit mode - when the interface is transmitting data.
• Receive mode - when the interface is receiving data.
The total energy consumption Etotal of a short-range RF module is a summation of the
ith wake-up (Ewui), transmit (Etxi) and receive (Erxi) energies, in addition to the energy
consumed in the sleep and idle states, Es and Eid, expressed as:
Etotal = Es + Eid +N∑i=1
(Ewui + Etxi + Erxi
)(4.1)
Etotal = PsDs + PidDid +
N∑i=1
(PwuiDwui + PtxiDtxi + PrxiDrxi
)(4.2)
where Ps, Pid, Pwu, Ptx and Prx are the power consumption of the RF module in the
sleep, idle, wake-up, Tx and Rx modes and Ds, Did, Dwu, Dtx and Drx their respective
durations. N is the number of communications: i.e. number of events, broadcasts or
polling requests.
4.6.1 Energy Measurement
The power draw and duration measurements given in Tables 4.2 and 4.3 are applied
to the model. The calculated energy is the product of the power draw and duration.
With polling communication (home gateway to end-device), the RF module remains in an
idle/listening state in order to detect incoming packets. Therefore, for Wi-Fi, BT (in sniff
mode), BLE and RF433, there is no sleep and wake-up power consumption component in
equation (4.2). In a weekly cycle, the time taken in the idle mode is the total weekly period
(604,800 seconds) less the time for all polling request transmissions. ZigBee, unlike the
other three interfaces, can go into sleep-mode during this phase, and reverse polls (end-
device to home gateway) its coordinator/master at regular intervals.
4.6 Energy Consumption Model 97
When designed for broadcast-only communication, a 10B payload is sent by the end-
device (toaster) once every hour. Consequently, there will be 168 transmissions (N = 168)
per week irrespective of the frequency of toast slices made. In an event-driven communi-
cation, however, the energy consumed by the modules scales linearly with the number N
of toasts made. The energy consumption of Wi-Fi and BT modules can be greatly influ-
enced by the time required (see Table 4.3) to power on and re-establish connection with
the master/coordinator. For the BT module, an average of 640 ms was measured for a
complete inquiry scan and advertisement (stages of BT protocol), assuming the interface
is completely turned off at the end of each transmission. It takes an average of 1836 ms
for the BT wake-up phase (including turn on, inquiry scan, paging and synchronization)
during which the HC-06 consumes an average of 96 mW. The nano Wi-Fi dongle takes a
little longer (3165 ms) in the wake-up phase but with lower average power consumption
(66 mW).
4.6.2 Processing and Storage Energy
A simple IoT hardware application like the toaster example requires some processing
and storage of data bits. Although the main focus of this work is on the measurement
and assessment of the communication energy consumption, an IoT toaster needs some
memory to store toast counts, which can be incremented step-wise until its reset point. It
is therefore necessary to consider its potential processing and storage energy usage.
For the same MCU and clock crystals on a circuit board, the processing energy usage
may be small but similar for the three communication methods. With regard to stor-
age, while an event-driven method may not necessarily require storage, the broadcast
and polling methods do. Recent RF modules are designed as a SoC, with a few kilo-
bytes/megabytes of static random access memory (SRAM) or NAND flash memory. For
example, the toaster could be designed with a 32-bit ARM Cortex M4F MCU (as in [142]),
which has 64 KB SRAM and 512 KB flash memory. The M4F consumes 1.2 µA (2V op-
erating voltage) in a low-power state with full SRAM retention. Hence, a toaster may
consume about 1.5 J per week to retain the toast count data of less than 10B, which is far
smaller than the capacity of the SRAM. Utilizing flash memory, on the other hand, does
98 Energy Consumption of IoT Wireless Network Protocols
not require power to retain data (non-volatile). A study in [143] shows that a NAND
flash memory consumes 4.72 µJ and 38.04 µJ to read and write data, respectively, to a 2
KB memory page within a 128 KB memory block. For 250 read/write operations, a flash
memory may consume as much as 10.7 mJ of energy.
The choice of memory is dependent on the type of application, the frequency of data
requests and amount to data per operational duty cycle. We can, however, conclude that
the energy use for running the processor and storing or retaining the data bits is small
and can be ignored if flash memory is utilized for this application.
4.7 Comparison of Energy Consumption
The plots in Figure 4.8 show the weekly energy usage comparison of BT, BLE, ZigBee,
Wi-Fi and RF433 for the three communication paradigms considered in this study. For
event-driven and broadcast modes, it is assumed that the interface is turned off between
communications events. In all modes, the Wi-Fi module is the most energy hungry, with
BLE being the most energy-efficient for all but one (i.e. Polling). ZigBee is more energy-
efficient than BLE for polling due to its reverse polling capability. This allows the ZigBee
module to sleep for longer periods while BLE periodically listens to the central (home
gateway) for polling requests. This result demonstrates an incentive for the communica-
tion to be driven by the end device. Whilst it is energy-wise (visibly from Figure 4.8 (a))
to choose event-driven mode for lower usage rates, the broadcast mode is more energy-
efficient beyond 168 servings per week. The energy usage of polling mode, however, is
several orders of magnitude higher than both broadcast and event-driven modes, since
the RF modules must stay in idle mode to receive incoming polling requests from the
home gateway. Note that the variation of the polling energy numbers is masked on the
plot due to the use of a logarithmic scale.
As an example, if toasting 20 slices of bread in a week, the BLE interface, configured in
polling, event-driven or broadcast mode, will consume 36.9 kJ, 4.2 J or 35.3 J, respectively.
Hence polling mode uses nearly 9000-times more energy (about 40000-times more for
BLE). To put this in perspective, an 800 W toaster will consume 72 kJ of energy in one
4.8 Energy-Efficient Design 99
BT
BLE
ZigBee
Event-drivenWi-Fi
RF433
(a)
BT
BLE
ZigBee
Broadcast Wi-Fi
RF433
(b)
BTBLE
ZigBee
Polling
Wi-Fi RF433
(c)
Figure 4.8: Energy consumption per week for BT, BLE, ZigBee, Wi-Fi and RF433 usingcommunications paradigms (a) Event-driven, (b) Broadcast and (c) Polling.
toasting operation (1.5 minute average toasting time). That is, adding an always-on Wi-
Fi interface to a toaster in this instance adds about 2.6% (1 slice per toasting event) or
5.1% (2 slices per toasting event) to the power consumption.
4.8 Energy-Efficient Design
This section draws some conclusions about energy-efficient design using the application
architecture from section 4.5 and power measurements from section 4.4. In particular, we
show that the most energy-efficient communications paradigm for each wireless interface
100 Energy Consumption of IoT Wireless Network Protocols
depends on the number of events reported (or amount of traffic) over the interface.
4.8.1 Interfaces Always On
The comparisons in Section 4.7 indicated that polling is not an energy-efficient option.
This assumed, however, that the interfaces would be turned off between events in event-
driven and broadcast modes. A naive design would just add a standard wireless interface
to an end-device and assume that the interface is always on. Figure 4.9 shows the total
energy per week used by an RF433 interface with the three communications paradigms
if the RF module stays on the entire time. In this case (and in general, if the interfaces are
Broadcast
Polling
Event-driven
RF433 (Always On)
Figure 4.9: Energy consumption per week of an RF433 interface (Always On).
always on), event-driven communication is preferred in the case where the number of
servings is very small. Once the number of toasting events exceeds a threshold (about 7
for this example), polling mode is more energy efficient. Broadcast mode will eventually
become the most energy efficient mode when the number of servings is very large (not
shown in the figure). Although this is a somewhat contrived example, it shows that the
energy-efficient design of the local wireless communications for the Internet of Things
can depend both on the choice of wireless interface and the amount of traffic (or applica-
tion) to be carried.
4.9 Conclusions 101
4.8.2 Interfaces Powered Down When Not In Use
For most traffic levels, sleep-mode is important for energy-efficient communications. For
very high traffic levels, always-on polling will eventually be preferable to event-driven
communication, but in this case broadcast will be the most energy efficient paradigm.
Figure 4.10 shows the total weekly energy used by a BT interface when toggling on/off
Polling
Event-driven
Broadcast
BT (Toggled On/Off)
Figure 4.10: Energy consumption per week of the BT interface (toggled on/off).
in the broadcast and event-driven modes. (Note the logarithmic scale in order to accom-
modate all three communication modes.) In this case, the event-driven mode is preferred
until the amount of traffic exceeds the constant broadcast frequency, after which broad-
cast mode is preferred.
4.9 Conclusions
In this chapter we have measured the energy use of five popular wireless interfaces when
they are in their inactive, listening and transmitting states. Using these measured values,
we have calculated the total energy usage for three different communications paradigms
– broadcast, polling and event-driven. In each case, it has been shown that the most
energy-efficient communications paradigm depends both on the relative energy usage by
102 Energy Consumption of IoT Wireless Network Protocols
the interface in its various modes and on the frequency and volume of traffic transmitted
over the interface.
These results suggest that for an IoT application developer, careful consideration of
the applications to be run over the wireless interfaces will be required to determine an
energy-efficient design. In many cases, the level of traffic will be uncertain (or will change
over time). The results suggest that applications should be designed to adapt to traffic
levels by selecting a different communications paradigm when it becomes more energy
efficient to do so.
Note also that a least-capital-cost solution may not be the most energy efficient (and
hence the least cost over the lifetime of the interface). Bluetooth and Wi-Fi interfaces,
for example, may be lowest cost and most readily available because of their volume pro-
duction but may consume more energy than alternatives. Their many communications
features may also be of no benefit for the specific application for which the communi-
cations interface has been added. A design process that takes into account total cost of
ownership (including both initial and operating costs) will be preferable.
Chapter 5
Energy-Efficient Architecture for IoTVideo Surveillance Services
5.1 Introduction
V IDEO surveillance systems and services are increasingly becoming a significant
application of the Internet-of-Things (IoT). IoT video surveillance application in-
volves generating streams from one or more video sources (cameras) at an end-user’s
premises, some level of signal compression and in some cases aggregation at those premises,
transmission through the public network to another user premises (live streaming), or
to a storage facility (e.g. data centre) from which the signal can be delivered to other
premises on demand.
Many emerging IoT services like video surveillance are still in the embryonic state of
deployment and yet to be fully characterised in the literature. Hence, there is little study
and analysis of their energy consumption and implications for the networks on which
the service is built. There is therefore a need for an end-to-end analysis of the application
data flow, taking into account the energy associated with every network segment. Fur-
thermore, a good understanding of step-wise energy consumption can be important in
determining an energy-efficient network architecture for the service.
Cloud computing has been pivotal to the growth of IoT services including video
surveillance applications. However, cloud computing may not always be an optimal so-
lution for processing and distribution of video data due to many quality-of-service (QoS)
103
104 Energy-Efficient Architecture for IoT Video Surveillance Services
and other concerns including latency, bandwidth utilisation, mobility support, location
awareness and energy consumption [56, 57]. Fog computing, an emerging alternative to
cloud, promises to address these concerns by bringing cloud intelligence and resources
much closer to the user with the implementation of distributed computing [67, 68]. It
has been shown that Fog computing could aid in the reduction of energy consumption,
both in the network architecture and cloud [63, 69]. The study in [63] shows that the
power consumption of an industrial IoT video streaming service could be improved by
more than 50% compared with cloud computing and storage alternatives. This could
be achieved through local processing in a local (fog) server as oppose to cloud, a result
achieved by applying machine learning techniques to analyse and prioritise the video
data.
In this chapter, an investigation of the energy trade-off between the network architec-
tures of IoT video surveillance systems is carried out. The study considers use cases of
live video streaming, on-demand video streaming from a fog [i.e. local (direct) or edge]
or cloud data storage, and processing of video data (using a face recognition applica-
tion). As with the rest of the work described in this thesis, the focus is on a domestic
(or small-business) IoT application context. Further, this chapter deals specifically with
streaming, storage and processing of an end-user or customer’s video data, and not the
larger scale streaming of video traffic from service providers such as ISPs or companies
such as Netflix.
A network IP camera (IPcam) based video surveillance service is employed as a use-
case model for a generalised IoT video surveillance service. Such a service might, as an
example, be part of a security service, or part of an in-home aged care monitor service,
etc. The contributions in this chapter are as follows:
(i) A detailed power consumption modelling of a network IP camera taking into ac-
count the effect of video frame rate, pixel rate and video bit rate on its power usage.
(ii) An end-to-end equipment-level energy consumption model for network architec-
tures of IoT video surveillance services and their video streaming and storage en-
ergy consumption implications.
(iii) An assessment of the energy implication of video processing when done locally as
5.2 Motivation for study 105
opposed to processing at an edge data centre or a more remote cloud data centre,
using as an example a face-recognition application.
Our empirical analysis, modelling and results offer insight into the energy efficiency
of video surveillance applications and services in an IoT context, and could help appli-
cation developers and architects make informed decisions in optimising and deploying
more energy efficient services.
The chapter is organised as follows. Section 5.2 gives the motivation for this study
and 5.3 introduces an IoT video surveillance system. Section 5.4 details experiments and
measurements conducted for this work and a model for power consumption of a net-
work IPcam is detailed in section 5.4.3. Section 5.5 gives a general network architecture
for IoT video streaming and the network element energy models are given in 5.6. Sections
5.7 and 5.8 discuss modelling for live streaming and on-demand streaming architectures
respectively and section 5.9 evaluates a face recognition application as an example of
surveillance-oriented video data processing. A conclusion and insights into the implica-
tions of the results is given in section 5.10.
5.2 Motivation for study
The principal component of a video surveillance application is the streaming of data-
intensive Standard Definition (SD), High Definition (HD) or Ultra High Definition (UHD)/
”4K” video across the public network infrastructure to Data Centres (DC), irrespective of
the geographical location of the service-specific DC. Today, a number of service providers
offer and actively promote 24/7 HD (1080p) or UHD (2160p) live video streaming ser-
vices with pay-per-use cloud storage (for on-demand access) for up to 30 days (e.g. Nest
Cam1, Amazon Cloud Cam2) or more. Furthermore, it is also actively promoted that
users can share - directly or via embedded web-links on a website - access to live video
streams or stored video files through the IPcam’s management application. Some of these
features could potentially have significant network energy implications for the service as
1https://nest.com/cameras2https://www.amazon.com
106 Energy-Efficient Architecture for IoT Video Surveillance Services
a whole. Investigating these possible energy consumption implications on the network
infrastructure (e.g. due to traffic increase) will form the basis of this work. To put the
ensuing traffic volumes of these services into perspective, let us consider a back-of-the-
envelope calculation for a representative example video surveillance system.
A 2 Megapixel HD IPcam operating at 30 frames per second (f/s) requires a bit rate
of 4-6 Mb/s for an acceptable video quality [11, 144]. An IPcam continuously streaming
at this rate will initiate and upload a data traffic volume of ≈ 1.6 TB/month across the
Internet to a DC. Such a traffic volume is about 10-12 times the monthly average Aus-
tralian household data consumption [119, 145] and exceeds many ISP broadband data
plans today. A full UHD video stream requires a bit rate 2-3 times more than a HD and
8-9 times more than a SD video stream [28]. Assuming UHD streaming, it will take just
under 15 million annually subscribed camera units (noting the projected shipment of 98
million camera units in 2017 [146]) to exceed the global IP traffic (≈ 1.06 Zettabytes) in
2016 [28]. Clearly, backhaul of UHD IoT video streams may come at a significant net-
work infrastructure energy cost and could degrade network QoS if this usage scenario
is adopted. A good understanding of the related energy consumption is needed, and
possible mitigation strategies are yet to be fully studied.
5.3 IoT video surveillance system
An IoT video surveillance system can be described as an all-in-one commercial off-the-
shelf IoT solution for personal or commercial video surveillance (e.g. Nest Cam1, Ama-
zon Cloud Cam2, Foscam3, Netgear Arlo4). A typical video surveillance system com-
prises of one or more network IPcam(s) and a bundled cloud service, with web-based
front-end applications through which live video streams from the IPcam, or on-demand
historical video clips can be accessed [10, 11, 144]. In contrast to older video surveil-
lance technologies, network IPcams are cost-effective and simpler to set up. They benefit
from the use of the standard public IP-based network infrastructure and relatively less
costly DC storage and processing services, enabling secured world-wide access to any
110 Energy-Efficient Architecture for IoT Video Surveillance Services
rate of 30 f/s and a maximum configurable bit rate of 4 Mb/s. A frame in this context
refers to one image within a video stream. The IPcam’s video encoder runs a standard
H.264 video compression algorithm [60], which restricts all of the experiment measure-
ments to H.264 encoding. A 16 GB class 10 SD card was used as the local data storage
device for some of the experiments.
5.4.1 Measurement Setup
The measurement setup is as shown in Figure 5.3. For convenience, the IPcam is posi-
tioned facing a target scene with little motion - tests with constant motion did not show
a difference in power. The IPcam power consumption during each experiment was mea-
sured and recorded through a USB-connected data logging laptop (see Figure 5.3), using
a digital AC power meter, Power-Mate PM10AHDS [148]. The meter has an accuracy of
± 0.01 W and granularity of 1 second. Data traffic generated by the video stream was
captured using a packet sniffer application, ”Wireshark”, on the client PC, as shown in
Figure 5.3.
Power Meter
Gateway
Wireless Router
Power Data Logging Laptop
Ethernet/Wi-Fi
USBMains(240 V)
Web Browser & Packet Sniffer (Wireshark)
IP Camera(Web-server & Local Storage)
Pow
er
Cloud
Figure 5.3: Measurement setup comprising IP camera, power meter and data loggingPCs.
5.4 Experimental Parameters and Measurements 111
To determine the power consumption profile of the IPcam, a number of experiments
were conducted with varying load - load in this context referring to the video frame size,
frame rate and video bit rate. These settings were made through the IPcam’s web-based
video management application (version 2.11.1.120). Configuration settings are then writ-
ten to the IPcam’s video encoder and a pre-test wait time of 2 minutes was allowed prior
to commencing each test. The test-series configurations are given in Table 5.2. The first
Table 5.2: Experimental configuration of the IP Camera.
Frame Frame Size Number Aspect Bit RateDescriptor (pixels) of Pixels Ratio Limit (b/s)
180p 320 x 180 57600 16:9 200k240p 320 x 240 76800 4:3 256k360p 640 x 360 230400 16:9 512k480p 640 x 480 307200 4:3 1M720p 1280 x 720 921600 16:9 2M960p 1280 x 960 1228800 4:3 4M
column in Table 5.2 denotes a convenient shorthand frame descriptor employed in this
study to refer to the different frame sizes. A test-series may include the live stream-
ing frame size (i.e. 320 x 180 to 1280 x 960 pixels), the designated frame rate (i.e. 5 -
30 f/s in steps of 5 f/s), the maximum allowable bit rate (200 kb/s to 4 Mb/s) and the
H.264 key frame interval (i.e. time between 2 full frames). Applying industry best practice
[10, 11, 144], the key frame interval was set at 2 seconds for all test-series (e.g. for a 30 f/s
configuration, one full frame (i-frame) is produced every 60 frames while delta/p-frames
are produced between full frames. See [60] for more details.) The minimum configurable
key frame interval is 10 frames - hence the minimum frame rate measured is 5 f/s.
For each test-series, an elapsed video streaming time of 5 minutes was employed and
the reported power values are an average of 300 power readings (1 per second). Each test
was conducted using both Ethernet and Wi-Fi network access. Tests were also conducted
when recording and storing video files locally on the SD card and remotely to a DC. For
streaming and/or storage via a DC, power consumption models for cloud services were
developed using the IPcam’s parent cloud service (i.e. Foscam) and a third-party cloud
112 Energy-Efficient Architecture for IoT Video Surveillance Services
service (i.e. camCloud5) as examples. Both services use Amazon EC2 data centre servers
based in North America.
5.4.2 Power Consumption Measurements
The results of the power consumption measurements are presented in this section. Two
key parameters were varied in the tests for a set of frame sizes: (i) video frame rate and
(ii) video bit rate.
Power Consumption vs Frame Rate
Power consumption measurements were recorded while varying the IPcam frame rate
from 5 f/s to 30 f/s in steps of 5 f/s, for a maximum bit rate setting and key frame inter-
val (2 sec), when streaming video with frame size 180p, 240p, 360p, 480p, 720p, 960p
pixels as given in Table 5.2. Figure 5.4 (a) & (b) show results of power consumption
measurements as a function of video frame rate for Ethernet and Wi-Fi network access,
respectively. Each curve corresponding to a designated frame size (in pixels) consists
3.0
3.2
3.4
3.6
3.8
4.0
0 5 10 15 20 25 30 35
Pow
er (
W)
Frame rate (f/s)
240p
720p480p
960p
Ethernet
180p 360p
Linear ModelMeasurement
No-load / baseline power
(a)
3.0
3.2
3.4
3.6
3.8
4.0
0 5 10 15 20 25 30 35
Pow
er (
W)
Frame rate (f/s)
180p
720p
360p
960p
No-load / baseline power
Wi-Fi
240p
480p
Linear ModelMeasurement
(b)
Figure 5.4: Power consumption as a function of frame rate of a network IP camera with(a) Ethernet access and (b) Wi-Fi access.
of a baseline component (y-intercept), and an additional component that is linearly load-5https://www.camcloud.com
5.4 Experimental Parameters and Measurements 113
dependent, a behaviour analogous to that of some network devices and modern personal
computers, and corroborated by a number of studies [95,96]. The slope of the lines in Fig-
ure 5.4 increases with an increase in frame size due to the higher energy cost of encoding
and transmitting larger frames.
Table 5.3: Line of best-fit equations for power consumption as a function of frame rate ofthe IPcam, where x = frame rate.
Frame Number Power (Watts)Size of pixels Ethernet Wi-Fi180p 57600 3.07 + 0.5× 10−3x 3.51 + 0.9× 10−3x
240p 76800 3.07 + 1.2× 10−3x 3.52 + 1.2× 10−3x
360p 230400 3.08 + 1.6× 10−3x 3.52 + 1.4× 10−3x
480p 307200 3.08 + 2.3× 10−3x 3.53 + 2.2× 10−3x
720p 921600 3.06 + 4.5× 10−3x 3.51 + 4.3× 10−3x
960p 1228800 3.05 + 6.6× 10−3x 3.53 + 6.1× 10−3x
Using the linear least squares method, the lines of best-fit for each curve were plotted.
Thus for a given frame size, at x f/s, the IPcam power consumption is given as: bf +mfx,
where bf (in watts) denotes the y-intercept which corresponds to the no-load/baseline
power, mf is the slope/load-dependent power component representing the energy con-
sumption per frame and expressed in joules per frame (J/f) - the subscript f refers to
frame rate. Equations of the lines of best-fit for each curve are given in Table 5.3.
From Figure 5.4(a) [Ethernet] and Table 5.3, the baseline power consumption, bf ≈
3.07± 0.01 W represents more than 90% of the peak power draw at the maximum frame
rate. Table 5.3 lists the baseline power bf, and the incremental energy per frame mf as
the coefficient of frame rate x. Standard Deviation (SD) and Standard Error (SE) of bf are
≈ 0.01 and 0.005 respectively. Similarly for Wi-Fi network access, [see Figure 5.4(b)], the
baseline power draw, denoted by bf ≈ 3.52 ± 0.01 W [SD = 0.01, SE = 0.004], shows an
additional 0.45 W compared with that of the baseline with Ethernet. This difference is
attributable to the Wi-Fi network card. Note that the Ethernet port was in the on-state
during Wi-Fi experiments. The incremental energy per frame for the Wi-Fi connected
IPcam, mf is also given in Table 5.3.
114 Energy-Efficient Architecture for IoT Video Surveillance Services
From measurement, under low light conditions, the IPcam consumes an additional
baseline power of ≈ 2 W to drive its infrared camera functionality.
Estimating the Energy per pixel
The trends in the results above with increasing pixel counts imply that energy per pixel
is a unifying concept for the varying frame rates. To determine the per-pixel energy
consumption - expressed in joules per pixel (J/px), the energy per frame results presented
in Table 5.3 were utilised. Fig 5.5(a) and (b) show plots of the incremental energy per
frame (mf and mf) as a function of frame sizes (in megapixels). Using the linear line of
0
2
4
6
8
10
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Millions
Incr
emen
tal E
ner
gy p
er F
ram
e (m
J)
Frame Size (Mega pixels)
Ethernet
(a)
Linear ModelMeasurement
Slope = Energy per pixel
0
2
4
6
8
10
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Millions
Incr
emen
tal E
ner
gy p
er F
ram
e(m
J)
Frame Size (Mega pixels)
Wi-Fi
(b)
Linear ModelMeasurement
Slope = Energy per pixel
Figure 5.5: Incremental energy per frame plotted as a function of frame size (in megapixels) for a network IP camera with (a) Ethernet access and (b) Wi-Fi access.
best-fit equation, the incremental energy per framemf = bpx +mpxy and mf = bpx +mpxy,
where y represents the pixel count in the frame. The vertical intercepts (bpx and bpx) are
at 567 nJ with Ethernet and 693 nJ with Wi-Fi access. Similarly, the slopes/incremental
energy consumption per pixel (mpx and mpx) are 4.7 and 4.2 nJ/px with Ethernet and Wi-
Fi access respectively. This difference in energy per pixel might more likely be ascribed
to measurement uncertainty, since the measuring device had an accuracy of±10 mW (i.e.
10 mJ for 1 sec granularity), or to the Ethernet and Wi-Fi connectivity within the Foscam
being implemented with completely different driver chipsets.
5.4 Experimental Parameters and Measurements 115
Power Consumption vs Bit Rate
The IPcam configuration allows setting of its maximum video bit rate independently of
the frame size and frame rate; this is achieved by adapting its video encoding parameters.
Investigating the IPcam power consumption as a function of the video bit rate required
keeping the frame rate and number of key frames constant while varying the bit rate for
each configurable frame size. Only the top four bit rate steps were considered (i.e. 0.5,
1, 2, and 4 Mb/s) as there was no discernible difference in the consumption for lower bit
rate steps (i.e. 100 to 256 kb/s). A frame rate of 25 f/s and a key frame interval of 2 seconds
were used.
3.0
3.2
3.4
3.6
3.8
4.0
0 1 2 3 4 5
Pow
er (
W)
Bit rate (Mb/s)
240p720p
480p
960p
Ethernet
180p 360p
(a)
Linear Model
Measurement
3.0
3.2
3.4
3.6
3.8
4.0
0 1 2 3 4 5
Pow
er (
W)
Bit rate (Mb/s)
240p720p
480p
960p
Wi-Fi
180p360p
(b)
Linear Model
Measurement
Figure 5.6: Power consumption as a function of bit rate of a network IP camera with (a)Ethernet access and (b) Wi-Fi access.
The plot of power consumption as a function of bit rate is given in Figure 5.6(a) and
(b) for Ethernet and Wi-Fi access respectively. These figures demonstrate the net result of
employing an increasing coding effort with consequential energy cost to achieve lower
bit rate and thus achieving reduced energy cost for transmission. A similar linear charac-
teristic was observed with slight slopes for higher frame sizes, while the slopes of lower
frame sizes were substantially flat. This difference can be ascribed to the Foscam al-
ways employing a minimal level of encoding. For example, from the figure, a 240p video
stream at 25 f/s and 2 sec key frame interval requires no more than 1.2 Mb/s.
116 Energy-Efficient Architecture for IoT Video Surveillance Services
The equations of linear best-fit for each curve in Figure 5.6 are given in Table 5.4, using
the format br +mrz for a bit rate of z b/s, where br and mr are the baseline and incremen-
tal energy per bit respectively. Whereas the tests of power consumption against frame
rate, which showed a common baseline value among the frame sizes (Figure 5.4), these
results show a distinct difference in baseline consumption for different frame sizes. This
difference is mainly due to the additional power required for encoding at higher frame
sizes. The slopes for each of the traces, representing the incremental energy consumption
per bit, are also relatively consistent.
Table 5.4: Line of best-fit equations for power consumption as a function of bit rate of theIPcam, where z is the bit rate.
To model the power consumption of an IPcam, refer to Figure 5.1 which shows the cam-
era consisting of 3 main subsystems: the CSU, the IPU and the web server. The IPcam’s
power consumption, P , can be expressed as a sum of the power draw of these subsys-
tems, given as:
P = Pcsu + Pipu + Psrv (5.1)
where Pcsu, Pipu and Psrv represent the power consumed by the camera sensor unit, the
image processing unit and the web server respectively.
Studies have shown that the CSU, IPU (including the video encoder) and web server
exhibit power consumption to pixel rate, frame rate or bit rate relationships that can
be represented by a load-independent component and a load-proportional component
5.4 Experimental Parameters and Measurements 117
[149–151]. This behaviour is consistent with the measurements given in section 5.4.2.
Given that there are no ways to separate the power of Pcsu, Pipu and Psrv from external
measurements, we can therefore bring together their load-independent components and
their load-proportional components and rewrite (5.1) for a given time t as:
P (t) = Pidle +(EpQ
)f + EbR(t) (5.2)
where Pidle is the sum of the components of the camera power consumption that are
independent of the video parameters pixel rate, frame rate, and output bit rate, Ep is the
energy per pixel processed and Eb, the energy per bit, while Q and R are the number
of pixels per frame and video bit rate respectively. This equation is consistent with the
result presented in Tables 5.3 and 5.4 for an IPcam streaming video at f frames per second
with a bit rate R.
5.4.4 Traffic Measurements (Abit)
To measure the traffic volume,Abit (in bits), between the IPcam’s web-server and a client’s
web-browser, the packet sniffing tool (Wireshark in ”promiscuous mode”) was used to
record end-to-end traffic flows. Using the frame size : bit rate configuration combination
given in Table 5.2 at 25 f/s and a 2 sec key frame interval, live video streams (including a
128 kb/s audio stream) of a target scene were initiated and the resultant data traffic flow
recorded for 1 minute.
In a second experiment the camera was equipped with a 16 GB SD card storage and
the live stream recorded on the card for 1 minute, creating an Audio Video Interleave
(AVI) video file. Each session was repeated 5 times and the mean traffic volume and file
sizes calculated.
Figure 5.7(a) shows the measured traffic volume plotted against video file sizes for
a 1 minute video stream. The video file sizes range from 3 MB to 28 MB while the cor-
responding bit stream traffic volume was ≈ 1 MB higher on average, attributed to ACK
packets, framing (e.g. IP and Ethernet header), retransmission, and network/link man-
agement messages. Figure 5.7(b) and (c) show the measured traffic volume and bit rates
118 Energy-Efficient Architecture for IoT Video Surveillance Services
0
5
10
15
20
25
30
0 10 20 30
Mea
sure
d T
raff
ic V
olu
me
(MB
)
Video File Size (MB)
Upstream Traffic
Downstream Traffic
Total Traffic
(a)
0
5
10
15
20
25
30
180p 240p 480p 720p 960pM
easu
red
Tra
ffic
Vo
lum
e (M
B)
Frame Size
Upstream Traffic
Downstream Traffic
Total Traffic
(b)
0
1
2
3
4
5
180p 240p 480p 720p 960p
Bit
Rat
e (M
b/s
)
Frame Size
Upstream Bit Rate
Downstream Bit Rate
TotalBit Rate
(c)
0
5
10
15
20
25
30
35
180p 240p 480p 720p 960p
Nu
mb
er o
f D
ata
Pack
ets
('0
00
)
Frame Size
Upstream Packets
Downstream Packets
Total Packets
(d)
Figure 5.7: Measured traffic volume generated by the IPcam during live streaming exper-iment with frame rate of 25 f/s and key frame interval of 2 sec, plotted against (a) video filesizes and (b) video frame sizes. Bottom left and right: (c) Bit rate (Mb/s) and (d) Numberof IP packets, plotted against video frame sizes.
for the studied frame sizes (180p to 960p), while (d) depicts the packet distribution for the
traffic flow. On close examination of the Wireshark logs, the upstream traffic consisted
of mostly 1514 bytes TCP packets shortly followed by 54 bytes ACK downstream pack-
ets. The upstream/downstream packet ratio approached parity for lower frame sizes but
rapidly increases with increase in frame size (e.g. > 2 for 960p).
5.5 IoT Video Surveillance Network Architecture 119
5.5 IoT Video Surveillance Network Architecture
An example equipment-level network architecture for IoT video surveillance services is
as depicted by the upper half of Figure 5.8. Typical network architecture consists of 4
main segments: the access network, the metro/edge network, the core network and a
data centre for storage and processing [41, 94]. The IoT Device Network (IDN) - shown
Access Network
OLTEthernetSwitch
BNG
BNG
Edge Router
Edge Router
Core Router
Core Router
Core Router
Edge Router
Metro & Edge Network Core Network Cloud Data Centre
Server
Core Router
Edge Data Centre
Server
Storage
IDNs
Storage
EthernetSwitch
EthernetSwitch
ONU
ONU
ONU
HGW
HGW
HGW
HGW
IoT Device Network (IDN)
Figure 5.8: An example network architecture (with PON) of an IoT video surveillanceservice spanning from the IP camera to data storage centres.
in the breakout diagram below the main figure - may contain many IoT devices but the
focus here is the IPcam, which includes an IoT device and IoT gateway as one unit (see
Figure 5.1). For the access network, a Fibre-to-the-Premises (FTTP) architecture using
Passive Optical Network (PON) technology, which has been shown in [42] to be the most
energy efficient access network technology, is adopted. Hence, network access to the IP-
cam is facilitated via a HGW (Ethernet or Wi-Fi) and an Optical Network Unit (ONU),
both usually located within the customer’s premises, which ends at the broken line. With
PON access network, a single fibre from the Optical Line Terminal (OLT) feeds a num-
ber of subscriber (32/64 are common) ONUs via an optical splitter (i.e. the black dot)
[89]. The metro/edge network includes an Ethernet aggregation switch which combines
traffic from many subscribers while the Border Network Gateway (BNG) acts as the gate
keeper, providing authentication and authorisation services. Edge routers within the
metro/edge network segment represent the gateway to the global Internet. The core net-
120 Energy-Efficient Architecture for IoT Video Surveillance Services
work consists of several large routers, generally located in major population centres, plus
optical transmission systems linking those centres. The Cloud Data Center (cDC) is usu-
ally in a centralised location and consists of storage, servers, aggregation switches and
edge routers, the latter linking the cDC to the core network.
Data storage and processing facilities to the left of cDC in Figure 5.8 are generally
considered a part of fog computing [68]. The lower half of Figure 5.8 represents an Edge
Data Center (eDC) akin to Content Delivery Network (CDN) infrastructure [68,152]. The
eDC is modelled using equipment employed by CDNs. Local storage is represented by
the SD card attached to the IPcam.
5.6 Network Energy Consumption Modelling
A description of the modelling approach employed in this study and subsequent energy
consumption models applied in characterising network equipment depicted in Figure 5.8
is detailed in this section. Specifications for representative commercial network equip-
ment gathered from manufacturer datasheets and other resources in the public domain
(e.g. published papers) are used in the models.
5.6.1 Modelling Approach
In modelling the energy consumption of a network segment and its elements, a common
”bottom-up” approach is to characterise energy consumption of each network element
based on its number of users (i.e. a ”per user” model) [40–42]. Another useful approach
allocates energy consumption as a function of application traffic flows (i.e. a ”capacity-
based” model) through the network element [69, 70]. The latter approach is adopted in
this study for two main reasons:
(i) Video surveillance applications are data intensive, generating many traffic flows
and potentially accounting for a notable proportion of global Internet traffic [28, 58,
153].
(ii) It allows for the apportionment of energy consumption of all network elements in
5.6 Network Energy Consumption Modelling 121
the network (regardless of size, capacity or energy consumption), across each of the
services that use the network element.
5.6.2 Energy Model
For live video streaming, the IPcam must be ”always on”. Since live streaming is es-
sentially the primary function of the IPcam, its full power draw is accounted for in the
energy model. Considering an IPcam streaming from time t1 to t2. Given its power con-
sumption P (t), as expressed in equation (5.2), the energy consumption of the IPcam Ecam
can be expressed as:
Ecam =
∫ t2
t1
P (t)dt (5.3)
Network elements such as routers and switches shown in Figure 5.8 exhibit a linear
power-to-load relationship as discussed in Chapter 2 and corroborated by many studies
[95, 96, 151]. This relationship is expressed as P (R) = Pidle + EincR, where Pidle (which
can be > 90% of max power [95, 96]) is the no-load power component, and EincR, the
incremental power component for a bit rate R. Einc is the incremental energy per bit
given as: (Pmax−Pidle)/Rmax, where Pmax and Rmax are the maximum power consump-
tion and bit rate respectively. In addition to the traffic dependent power consumption
EincR, the ”capacity-based” modelling approach [70] apportions the device idle power
consumption across its traffic streams, using the expression Pidle/URmax, where U = av-
erage utilisation. Therefore, the average energy per bit, λ, of a network element is given
as:
λ =PidleURmax
+ Einc (5.4)
Many services concurrently access network elements at any given time. Given a ser-
vice that generates, transmits and/or receives a total traffic volume of Abit (in bits) across
the network, the additional energy consumption of a network element n, attributable to
such service (En) is given as:
En =
(PidleURmax
+ Einc
)Abit = λAbit (5.5)
122 Energy-Efficient Architecture for IoT Video Surveillance Services
Storage device energy consumption can be modelled as a function of the volume of
data stored over a period of time. For large storage facilities like DCs, data is often stored
on shared spinning disks that are accessible to many users/services through server re-
quests. Similar to some network elements, the power consumption of storage devices
contains an idle component (due to spinning disks) and a dynamic/incremental compo-
nent due to data input/output (IO) operations. In modelling a shared storage device, a
power per bit approach, which allocates the idle power consumption across the total data
volume stored, in addition to the incremental power for data IO, is adopted.
Consider a storage device with an idle power consumption Pidle and a maximum
power consumption and storage capacity Pmax and Bmax (in bits) respectively. Assum-
ing an average utilisation Us, the energy consumption (Es) attributable to a service with
stored video file(s) of size Abit, for a duration ts (in sec) can be expressed as:
Es =
(Pidle × tsUsBmax
+ Einc,s
)Abit (5.6)
where Einc,s is the incremental energy per bit for data IO operations.
For ts ≫ UsBmaxEinc,s/Pidle, then Pidlets/UsBmax + Einc,s ≈ Pidlets/UsBmax. Therefore
(5.6) can be written as:
Es ≈ µtsAbit (5.7)
where µ represents the average power per bit of the storage device, given by µ = Pidle/UsBmax.
5.6.3 Network Equipment Modelling
A selection of the representative network equipment applied in characterising the differ-
ent network architectures, modelling techniques and assumptions is presented below.
Customer Premises Equipment
The customer premises equipment (CPE) refers to energy consuming network devices
within a subscriber’s premises, that are a part of the network architecture being mod-
5.6 Network Energy Consumption Modelling 123
elled. They include the IPcam, home gateways and the ONU.
A) IP Camera: Applying equation (5.2) for a configuration of 720p (1280×720) pixels at
25 f/s (2 sec key frame interval) and 2 Mb/s video bit rate, the IPcam power draw P is
calculated as 3.18 W and 3.62 W for Ethernet and Wi-Fi network access. For a one minute
live video streaming, the energy consumption contribution of the IPcam is calculated
using (5.3) as 191 J and 217 J respectively. Wi-Fi network access is used in the calculations
below.
Table 5.5: Energy per bit of network equipment in the transport network.
Network ElementMax Capacity Idle Max Energy
Model [DL/UL]* Power Power per bit(Gb/s) (Watts) (Watts) (nJ/bit)
Bringing the content source closer to the user edge (i.e. fog computing) involves pro-
visioning what are described as ”fog nodes” or eDC, which can be edge routers (with
attached storage), servers, modems or gateway devices [67, 68]. An eDC can be readily
co-located within an ISP’s facility, adopting a similar architecture employed by CDNs at
the edge network. In this model, it is assumed that the eDC is populated with servers
and network equipment akin to that of a well-known CDN provider, Akamai [160]. Us-
126 Energy-Efficient Architecture for IoT Video Surveillance Services
ing Akamai’s quarterly reported carbon intensity for bandwidth, the study in [155] cal-
culated its energy per bit for data traffic as 0.89 nJ/bit. This value is assumed to include
a factor for redundancy.
For the eDC, a PUE of 1.8 is assumed; this value is closer to the average PUE of typical
facilities today [156,161]. The difference in PUE values between cDC and eDC is owing to
notion that newer cDC facilities are much more likely to be energy efficient (lower PUE)
than a central office or ISP facilities for example.
In modelling edge storage for the eDC, similar equipment used for cloud storage is
employed but of a smaller scale. Given the equipment specifications in Table 5.6, with
an assumed utilisation of 50%, the power per bit stored for an edge storage device is
calculated as 1.04 pW/bit.
5.7 Modelling IoT Live Video Streaming Architectures
With live video streaming, real-time video frames emanating from the IPcam are trans-
mitted directly to an end-user’s device without being stored. The energy consumed by
this service is therefore dependent on the IPcam, transport network (access, edge, core
networks), DCs (for some applications) and the end-user device on which video frames
are rendered. The last is not considered in this analysis as the focus is on the service
network architecture. Three models of IoT live video streaming can be considered; (i)
Direct Live Streaming (also known as HTTP/RTSP Streaming) (ii) Cloud Streaming and
(iii) Edge Streaming. The last two may interchangeably utilise HTTP or FTP protocols.
5.7.1 Direct Live Streaming
Direct live streaming is intuitively the default option when considering real-time video
streaming. Two possible user types are modelled; (i) Local User and (ii) Remote User.
(i) Local User (LU): This refers to a user within the same LAN accessing live streams from
the IPcam via the HGW. An example of such usage is the baby monitor application. Over
5.7 Modelling IoT Live Video Streaming Architectures 127
a time interval t the total energy consumed by a live (l) streaming service when accessed
by a local user (E(l)LU) is given as:
E(l)LU = Ecam +NuλhgwAbit (5.8)
whereNu is the number of concurrent users accessing live video streams from the IPcam,
λhgw is the energy per bit of the HGW,Abit is the total bits per exchange (upload) over the
interval, andEcam is the total energy consumed by the IPcam over the interval, calculated
using (5.2) and (5.3).
(ii) Remote User (RU): This refers to a user accessing direct live video streaming service
from the same or a different Autonomous System (AS), outside the IPcam’s LAN. The
latter is modelled here noting that only a small energy contribution from the core network
is missing from cases where the RU is close to the source (e.g. same ISP).
Since the data source is at the user premises, the video stream flows through the
access and edge networks twice. Two RU types can be defined; one in which the IPcam
and RU are within the same geographical region (e.g. same city/country) referred to as
RU-local, and the other in which they are in different geographical regions, RU-global. For
RU-local, the video data transits through its parent ISP’s access and edge networks, and
at least one core router (i.e. Hcn = 1), then through the edge and access networks again
to the destination user device. For RU-global, the traffic traverses multiple core network
elements in addition to the edge and access networks. The total energy consumed by a
live streaming service when accessed by a remote user (E(l)RU) can be expressed as:
E(l)RU = Ecam + 2
[λacc + ησn
(λes + λbng +Henλer +
Hcnλcr
2
)]NuAbit
= Ecam +(2λTPe +Hcnησnλcr
)NuAbit (5.9)
where λacc, λes, λbng, λer, λcr represent the energy per bit of the access network, Ether-
net switch, BNG, edge router and core router respectively; the factor of 2 in the second
term represents the number of times video data traverses the access and edge network
128 Energy-Efficient Architecture for IoT Video Surveillance Services
segments; η is a factor for redundancy, assumed to be 2, and σn, a factor for cooling and
power conversion overhead for the transport network equipment - akin to PUE of a DC
- assumed to be 1.8 [161]. Hen and Hcn are the number of hops/routers traversed in the
edge and core networks. Hcn = 1 for RU-local andHcn = 12 (from traceroute - discussed in
section 5.7.2) for RU-global. λTPe is the combined energy per bit of the transport network,
up to the edge, given by:
λTPe = λacc + ησn(λes + λbng +Henλer
)(5.10)
For the access network, λacc = λhgw + λonu + σnλolt, where λhgw, λonu, λolt are the energy
per bit of the HGW, ONU and OLT respectively. The HGW and ONU are self-cooled (no
cooling/overhead factor). All defined parameters in this section are listed in Table 5.7.
Table 5.7: Descriptions of parameters defined in the energy models.
Parameter DescriptionEcam Energy consumption of the IPcam streaming for duration t in secondsλhgw Energy per bit of the home gateway deviceλonu Energy per bit of the access network optical network unitλolt Energy per bit of the access network optical network terminalλes Energy per bit of the metro-grade Ethernet switchλbng Energy per bit of the border network gatewayλer Energy per bit of the edge routerλcr Energy per bit of the core routerλTPe Combined energy per bit of the transport network up to the edgeλTPc Combined energy per bit of the transport network up to the coreλeDC Total energy per bit of a edge data centreλcDC Total energy per bit of an cloud data centreNu Number of users streaming live video from the IPcamHen Number of hops/routers traversed in the edge networkHcn Number of hops/routers traversed in the core networkη factor for network redundancy or protectionσn Overhead factor for cooling and power conversion of network equipmentσe PUE of an edge data centreσc PUE of a cloud data centreAbit Total exchanged bits between the IPcam and a user device
5.7.2 Cloud Live Streaming
For cloud live streaming, the live video data is transmitted to the centralised cDC before
being routed to the user device, often with a some delay [68]. This model is conventional
5.7 Modelling IoT Live Video Streaming Architectures 129
for many IoT video streaming services today predicated on safe, worldwide accessible
storage of 24/7 video footage, captured by an IPcam. The energy consumed by a cloud
live streaming service is dependent on the energy consumed by the cDC server in addi-
tion to that of the transport network and the IPcam. Video data traverses the end-to-end
transport network twice (factor of 2). Therefore, the total energy consumption of a live
streaming service that is routed via a cDC (E(l)cloud) is given as:
E(l)cloud = Ecam + 2
[λacc + ησn
(λes + λbng +Henλer +Hcnλcr
)+σcλcDC
2
]NuAbit
= Ecam +(2λTPc + σcλcDC
)NuAbit (5.11)
where λcDC represent the energy per bit of the cDC server and its attached network ar-
chitecture, σc is the PUE of a cDC and λTPc , the combined energy per bit of the transport
network up to the core, given by:
λTPc = λacc + ησn(λes + λbng +Henλer +Hcnλcr
)(5.12)
Measurements using Foscam’s cloud services showed that its live streaming feature em-
ploys a direct HTTP streaming architecture with a delay 6 1 sec. With a third-party cloud
service (camCloud), however, the measurements revealed that video data is routed via a
cDC using FTP, with delay ≈ 8 - 13 sec. Both services use Amazon’s EC2 servers from a
similar DC location in the United States. To determine the path of video streams to the
cDC, the popular utility traceroute was employed from an end-user device to the desig-
nated servers. Both cloud services had similar edge and core network hop counts (i.e.
Hen ≈ 1, Hcn ≈ 12) but total hops between 23 and 28 - the difference attributed to ad-
ditional hops within Amazon’s AS. These values for number of hops (Hen and Hcn) are
used throughout this work.
130 Energy-Efficient Architecture for IoT Video Surveillance Services
5.7.3 Edge Live Streaming
With edge live streaming, video data is assumed to be streamed to an eDC closer to the
user before being re-routed to end-user device. Similar to direct and cloud streaming,
the live video stream passes through the access and edge network segments twice (factor
of 2). It is further assumed that most end-users access video streams within the same
geographical region [162], avoiding the need for routing across the core network. The
total energy consumed, E(l)edge, by service for live streaming via the eDC is expressed as:
E(l)edge = Ecam +
(2λTPe + σeλeDC
)NuAbit (5.13)
where λeDC is the energy per bit of the eDC, σe is the PUE of the eDC and λTPe , the
combined energy per bit of the transport network for an edge DC architecture.
Video File Size (MB)0 5 10 15 20 25 30
Ene
rgy
Con
sum
ptio
n (J
oule
s)
0
500
1000
1500
2000
2500
Direct (LU)
Direct (RU-local)
Edge
Direct (RU-global)
Cloud (Low)
Cloud (High)
Figure 5.9: Energy consumption of network architectures for a live video streaming ser-vice.
5.7 Modelling IoT Live Video Streaming Architectures 131
5.7.4 Evaluation
For this analysis, consider a single user scenario (Nu = 1) and an IPcam configured with
different frame sizes (180p to 960p), live streaming at 25 f/s (2 sec key frame interval) for
1 minute (t = 60 sec). The corresponding video traffic volume (Abit) is shown in Fig-
ure 5.7(a). The energy per bit and power per bit values for the network equipment are
taken from Tables 5.5 and 5.6 respectively, and the number of hops in the edge and core
networks areHen = 1 andHcn = 12.
Figure 5.9 shows a plot of the energy consumption of direct (LU and RU), edge and
cloud (lower and upper bound) live streaming architectures, as a function of video traffic
volume (0 - 30 MB). The values plotted include energy consumed by the IPcam, the trans-
port network and in some cases the DC servers and their attached network infrastructure
as described by equations (5.8) to (5.13). The plots scale linearly with higher video traffic
volumes.
From Figure 5.9, it is clear that live video streaming via cloud servers consumes the
greatest amount of energy, while the absence of the transport network energy in direct
streaming for a local user i.e. Direct (LU) ensures the least energy usage. An edge-
based architecture is more energy efficient than a cloud-based architecture but less energy
efficient than direct access. Evidently, the energy cost of routing live streams via a DC is
the underlying differentiator, given that there is no significant difference in the energy
cost for data transport among the three. At Abit = 0 (i.e. y-axis), the energy consump-
tion is only attributed to the IPcam. It is very likely that most live streaming requests
will come from outside the home LAN. Hence the figure indicates that it is more energy
efficient to access direct live streaming from the IPcam, avoiding additional energy cost
of DC servers. Since storage of video data at a DC is optional for video surveillance
services, adopting an architecture that always routes live video streams through a DC is
energy costly and inefficient when storage of the video footage is not required.
A breakdown of the energy consumption by the individual segments composing the
various network architectures is given in Figure 5.10.
132 Energy-Efficient Architecture for IoT Video Surveillance Services
Video File Sizes (MB)0 5 10 15 20 25 30
Ene
rgy
Con
sum
ptio
n (J
oule
s)
0
500
1000
1500
2000
2500
IPcamAccess Network
Edge & Core Network
eDC
cDC (Low)
cDC (High)
Figure 5.10: Energy contributions of the different network segments in a live streamingnetwork architecture.
5.8 Modelling IoT On-Demand Video Streaming Architectures
With IoT on-demand video streaming, the real-time video streams are recorded to storage
devices as video files, from where they can be accessed or downloaded on-demand at
any time. Therefore, IoT on-demand video streaming models include storage energy in
addition to the IPcam, transport and DC energy consumption. Three sources of video file
storage (i.e. SD card, edge and cloud storage) associated with the three on-demand video
streaming architectures: (direct, edge and cloud streaming) are analysed.
5.8.1 Energy Consumption of IPcam for On-Demand Streaming
In section 5.7, for a live streaming architecture, the energy consumption of the IPcam
(with no SD card) was calculated using its full power (baseline and incremental) con-
sumption. For on-demand streaming, however, the optional SD card is required. Fur-
thermore, on-demand streaming functionality operates independently of the IPcam’s pri-
5.8 Modelling IoT On-Demand Video Streaming Architectures 133
mary live streaming function. Hence, in this model, only the additional power consump-
tion when the SD card is installed (baseline and incremental) is considered in calculating
the IPcam’s energy consumption.
3.40
3.50
3.60
3.70
3.80
0 20 40 60 80
Pow
er C
onsu
mpti
on (
Wat
ts)
Time (sec)
Baseline power
Start
(t1)
Wi-Fi
End
(t2)
Baseline + Encoder (load) + SD Card
Baseline + Encoder (load)
PSD
δPcam
Figure 5.11: Power consumption of IPcam with attached SD card when streaming a 1minute video file on-demand.
To determine the IPcam incremental energy consumption for on-demand streaming,
the measurement setup described in section 5.4.1 is adopted. Figure 5.11 shows an exam-
ple of the measured power consumption of the Wi-Fi-connected IPcam when streaming
a 16.1 MB recorded one minute video file (1280 x 720 pixel, 25 f/s, 2 sec key frame interval)
from the SD card to a PC web browser. Figure 5.11 is a plot of power consumption as
a function of time (blue curve), and displays the average baseline power with light en-
coder load (lowest pixel rate format), the baseline plus high encoder (load) power (high
pixel rate format) with and without the SD card attached. The difference between the
latter two values (PSD ≈ 70 mW as given in Table 5.6) gives the baseline power of the SD
card which is attributed to the energy for storage. The encoder (load) is an indication of
the additional power consumed when the IPcam is configured for higher pixel rates (e.g.
134 Energy-Efficient Architecture for IoT Video Surveillance Services
720p from Table 5.2 in this case). The blue curve in Figure 5.11 shows the start (t1) and
end (t2) points of the streaming process.
The incremental energy consumption of an IPcam for the on-demand streaming pro-
cess, Ecam(inc), contains two components: one for storage and the other for streaming,
expressed as:
Ecam(inc) = ESD +NdEcam(str) (5.14)
where ESD is the storage energy given by PSDts, with ts being the length of time (in sec)
the video file is stored; Nd is the number of times the video is downloaded; Ecam(str) is
the energy consumed for streaming the video file, which is the integral (from t1 to t2) of
the difference between the blue curve and the line representing the baseline power for the
IPCam with encoder (load) and SD card (δPcam(t)) that is immediately below the curve.
For the 1 minute video file considered above, Ecam(str) is calculated as ≈ 3 joules.
5.8.2 Direct On-Demand Streaming
With direct on-demand video streaming, users can access video files that were previ-
ously recorded and stored on the SD Card attached to the IPcam. Since the source of
video streams is the same as with live streaming, a similar architecture applies. The total
energy consumption for a LU and RU (E(od)LU and E
(od)RU ) accessing direct on-demand (od)
streaming for Nd times within a time period ts is expressed as:
E(od)LU = ESD +Nd
(Ecam(str) + λhgwAbit
)(5.15)
E(od)RU = ESD +Nd
[Ecam(str) +
(2λTPe +Hcnησnλcr
)Abit
](5.16)
whereHcn = 1 for RU-local andHcn = 12 for RU-global.
5.8 Modelling IoT On-Demand Video Streaming Architectures 135
5.8.3 Edge & Cloud On-Demand Streaming
As with the previous on-demand architectures, streaming via edge and cloud services
incurs additional energy for DC storage of video files. The total energy consumption of
an on-demand video streaming service that is accessed from an eDC and cDC, E(od)edge and
E(od)cloud are given as:
E(od)edge =
[(Nd + 1
)(λTPe + σeλeDC
)+NR
(ησeµets
)]Abit (5.17)
E(od)cloud =
[(Nd + 1
)(λTPc + σcλcDC
)+NR
(ησcµcts
)]Abit (5.18)
where NR represents the number of edge or cloud DCs with stored replicas of the video
file of size Abit; µe and µc are the power per bit for edge and cloud storage devices. The
factor (Nd + 1) indicates that the video file must first be uploaded to a DC before it can
be streamed Nd times. λTPc and λTPc represents the per-bit transport energy to reach an
edge and cloud DC respectively, defined by equation (5.10) and (5.12).
5.8.4 Evaluation
To evaluate and compare the energy efficiency of IoT on-demand video streaming archi-
tectures described and modelled in the preceding section, let us consider a 1280 x 720
pixel, 25 f/s, 16.1 MB video file (Abit = 128.8 Mbits) that is 1 minute in length. Energy-
per-bit values for the transport network and DCs, and power-per-bit values for storage
devices are sourced from Tables 5.5 and 5.6 respectively. The PUE of the cDC and eDC
(σc and σe) are taken as 1.2 and 1.8. It is assumed the video file was stored for 1 minute
(ts = 60 sec) for a short-term storage scenario and located at a single DC facility (NR = 1).
Figure 5.12 shows plots of energy consumption of direct (LU and RU), edge and cloud
(low and high) network architectures for IoT on-demand video streaming services, as a
function of video file sizes (in MB). As with the live streaming case, both high and low
136 Energy-Efficient Architecture for IoT Video Surveillance Services
Video File Size (MB)0 5 10 15 20 25 30
Ene
rgy
Con
sum
ptio
n (J
oule
s)
0
500
1000
1500
2000
2500
3000
Direct(RU-local)Direct
(RU-global)
Cloud (Low)
Cloud (High)
Edge
Direct (LU)
Figure 5.12: Energy consumption of network architectures for IoT on-demand videostreaming services.
estimates for a cloud architecture are far less energy efficient than edge and direct on-
demand streaming mainly due to the energy cost of cDC and cloud storage in addition
to the transport energy cost. An edge architecture is a more energy efficient option than
a cloud architecture but less efficient than direct streaming. Figure 5.12 also shows no
significant difference between a RU streaming stored video files from the IPcam (i.e. lo-
cal fog) within the same geographical location (RU-local) and that from a distant user
(RU-global). Overall, direct streaming from the IPcam is the most energy efficient archi-
tecture for on-demand streaming provided the gateway can handle the load. An edge
architecture could be considered the most energy optimal if data protection, redundancy
and reliability are paramount to the service.
Number of Downloads
To further investigate the influence of storage energy cost on the energy efficiency of a
given network architecture, let us consider a 10 minute long 1280 x 720 pixel, 25 f/s video
5.8 Modelling IoT On-Demand Video Streaming Architectures 137
file of size 161 MB. Consider that the file is uploaded and stored in the DCs for 1 calendar
month (30 days x 86400 sec/day), hence ts = 2.592× 106 sec.
Number of downloads per month10-1 100 101 102 103
Ene
rgy
Con
sum
ptio
n (J
oule
s)
104
105
106
107
Direct (LU)
Direct (RU-global)
Cloud (Low)
Cloud (High)
Edge
Direct(RU-local)
Figure 5.13: Energy consumption as a function of number of downloads per month foran on-demand video streaming service.
Figure 5.13 shows the energy consumption of network architectures as a function of
number of downloads per month for an IoT on-demand video surveillance service. The
plot shows that for few download per month, total energy consumption of cloud and
edge-based architectures is largely dominated by the storage energy, but significantly
increases with an increase in the number of downloads per month, mainly due to the
energy cost of data transport and the DC servers. The cost of local storage (SD card) is
about 10 times that of edge and cloud, which is shown to be a major factor for direct
on-demand streaming being a less energy efficient option in this case. While there is
no clear difference between direct (RU-local) and direct (RU-global), they can be seen
as being more energy efficient than cloud, only for higher numbers of downloads (>
35) per month; this arises as the higher local storage cost becomes offset by the higher
transport costs involved in the transport network to a distant DC. A total number of
138 Energy-Efficient Architecture for IoT Video Surveillance Services
downloads of 1000 per month (limit of Figure 5.13) is an average of 1.4 per hour which
is within the performance capacity of the IPcam or a similar IoT gateway device (e.g.
Raspberry Pi). From Figure 5.14, an edge architecture can be clearly seen as the most
energy efficient for less than about 300 downloads per month, beyond which there is
hardly a clear distinction between an edge architecture and that of a direct streaming
architecture.
Energy per Download
A useful energy efficiency metric used to assess network architectures of a service is its
energy per download. Using the same 10 minute long video file as in the previous sec-
tion, the per-download energy can be determined for each of the network architectures
considered. For a direct (LU and RU), edge and cloud on-demand streaming service, the
per-download (pd) energy consumption E(pd)LU , E(pd)
RU , E(pd)edge and E(pd)
cloud, is given by:
E(pd)LU =
(ESD
Nd+ Ecam(str)
)+ λhgwAbit (5.19)
E(pd)RU =
(ESD
Nd+ Ecam(str)
)+(2λTPe +Hcnησnλcr
)Abit (5.20)
E(pd)edge =
[λTPe + σe
(λeDC +
NR(ηµets
)Nd
)]Abit (5.21)
E(pd)cloud =
[λTPc + σc
(λcDC +
NR(ηµcts
)Nd
)]Abit (5.22)
where NR = 1 and Nd > 1. All other parameters are defined in Table 5.7.
Figure 5.14 shows plots of energy per download as a function of number of down-
loads per month for an IoT on-demand video surveillance service. The figure depicts
plots for direct, edge and cloud-based network architectures for a range of downloads
per month. Figure 5.14 shows that for almost the entire range of number of downloads
5.8 Modelling IoT On-Demand Video Streaming Architectures 139
Number of downloads per month10-1 100 101 102 103
Ene
rgy
per
dow
nloa
d (J
oule
s)
103
104
105
106
Direct(RU-local)
Edge
Cloud (Low)
Direct (LU)
Direct(RU-global)
Cloud (High)
Figure 5.14: Energy per download as a function of number of downloads for on-demandstreaming of a 10 minute video file from a storage device.
per month, the energy consumption of the service is minimised when an edge-based
network architecture is adopted. As seen in the previous analysis, there is no distinct
difference when stored files are either accessed by remote users within the same city
(RU-local) or half way around the world (RU-global). This is attributed to the highly-
shared and highly efficient network elements in the core network. The absence of the
transport energy consumption for direct (LU) architecture makes it the most efficient but
it is limited to access within the IPcam’s LAN.
Effect of Video File Replication in Data Centres
To further understand the effect of replicating files in a number of DCs, a study of the per-
download energy consumption for a range of DC replicas is conducted. Using equations
(5.20) to (5.22) for a 10 minute (161 MB) video file that is replicated NR times, where NR
= 2, 20, 200 and 2000. While it is unlikely to have 200 or 2000 surveillance video files, the
140 Energy-Efficient Architecture for IoT Video Surveillance Services
Number of downloads per month10-1 100 101 102 103 104 105
Ene
rgy
per
dow
nloa
d (J
oule
s)
103
104
105
106
Edge(N
R= 2)
Cloud(N
R= 2000)
Edge(N
R= 2000)
Cloud(N
R= 200)
Edge(N
R= 200)
Cloud(N
R= 20)
Edge(N
R= 20)
Cloud(N
R= 2)
Direct(RU-global)
Figure 5.15: Energy per download comparison for replicating a 10 minute video file in 2,20, 200 and 2000 cloud and edge data centres.
model is indicative of other applications (e.g. VoD) with such file volume.
The plots in Figure 5.15 show energy per download as a function of number of down-
loads for edge and cloud-based network architectures with NR replicas of the video file.
The figure also shows a plot for direct (RU-global) on-demand streaming. Generally, the
plots show that an increase in the number of times a video file is downloaded comes with
an increase in the transport energy consumption. The results indicate that for a video file
that is accessed a few times a month (< 5), it is more energy efficient to store and serve
such a file from a cDC (4 µJ/bit) while maintaining replicas in two DCs. For low to
medium number of accesses (< 300 download per month), the total energy consumption
of the service is minimised when served from an eDC with two DC replicas. However,
if the number of downloads is high (exceeding about 300 downloads per month), the
results suggest that the total energy consumption of the service can be minimised with
a direct (RU-global) access architecture. Overall the energy consumption of the service
is minimised when video files are stored in a lower number of DCs much closer to the
5.9 Application with Video Processing Load 141
users, either in a distributed edge network architecture for fewer accesses per month or a
gateway device located at the customer’s premises for much higher accesses per month.
5.9 Application with Video Processing Load
In the previous sections, an analysis of the energy implications of a video surveillance
system is presented through the lens of energy consumed for a video transmission load
and a video storage load. In this section, an investigation of an energy efficient architec-
ture including a video processing load is presented. A comparison of the energy required
for local processing as opposed to cloud and edge processing is given, using a face detec-
tion and recognition application. An example use-case scenario is a facial identification
system utilised for gaining access to a home or office space or a video surveillance cam-
era capable of identifying subjects. In the place of the IPcam, a Raspberry Pi Model 3
(RPi) [163] with its attached pi camera was used (i.e. IoT gateway and IoT device respec-
tively). Specification of the RPi is given in Table 5.8. The face detection and recognition
Table 5.8: Parameters of the Raspberry Pi 3 Model B with attached Pi Camera.
Parameter ValueCPU Quad Core 1.2 GHzRAM 1 GBStorage (Patriot SDHC Card) 16 GBPi Camera 5 MP, 1080p @ 30 f/sNetwork Access Ethernet/Wi-FiOS Version Linux 8 (jessie)
application was programmed and ”trained” to detect and recognise facial features (using
the Eigenfaces algorithm) of 5 unique persons. The features were extracted from 12 im-
ages for each person. Hence, a face database of 60 images with a file size of 13 MB was
deployed on the RPi.
142 Energy-Efficient Architecture for IoT Video Surveillance Services
5.9.1 Energy Consumption of Raspberry Pi
To determine the energy consumed by the RPi when executing a face recognition process,
the measurement set up described in section 5.4.1 with a PowerMate power meter was
used. Figure 5.16 shows an example of a plot of power consumption as a function of time
0
1
2
3
4
0 5 10 15 20 25 30
Pow
er C
onsu
mpti
on (
Wat
ts)
Time (sec)
Start
(t1)
End
(t2)
Idle Power
Raspberry Pi
Figure 5.16: Power consumption of a Raspberry Pi 3 as a function of time when detectingand recognising a single face from a video frame. Time elapsed between time t1 and t2 is≈ 4.5 seconds.
for the RPi while detecting and recognising a known face (1 out of the 5 persons) from a
video frame.
The sequence of events for the face recognition process is as follows:
(i) with a prepared measurement set up, the pi camera was positioned facing the indi-
vidual to be identified.
(ii) the face recognition application was initiated at time t1 as shown in Figure 5.16; the
pi camera goes through a warm-up process (for 1 sec) prior to commencing video
capture; a video frame was grabbed and analysed for a positive match using the
face database.
(iii) an output of the recognition process was given at time t2 after which the RPi returns
5.9 Application with Video Processing Load 143
to its idle mode.
In calculating the incremental energy consumption of the RPi for a video processing
application, equation (5.3) is employed for a duration t1 to t2. Although the recognition
process at its core needs no storage capacity, the face database which holds the images
and features of known individuals requires some storage. Hence, the model should ac-
count for storage energy.
Unlike the IPcam, the RPi’s SD card is not dedicated to data storage as it also hosts
the Linux operating system upon which the RPi is based. It is impractical to separate
the energy consumption attributed to storage as was the case in the model for the IP-
cam. Hence, the energy consumption for face database storage on the RPi (using the
same SD card) is assumed to be similar to that of the IPcam as outlined in section 5.8.1.
The incremental energy consumption of a RPi (Epi(inc)) in conducting Np number of face
recognition processing operations is given as:
Epi(inc) = ESD +NpEprc (5.23)
where Eprc is the incremental energy for video processing, obtained from Figure 5.16 by
taking the integral (from t1 to t2) of the difference between the curve and the broken line
representing the RPi idle power.
5.9.2 Evaluation
Detailed information on the energy consumption for specific applications such as video
processing in a DC are not readily available in the literature. Furthermore, modelling
such processes for a DC is non-trivial. Therefore, in evaluating an energy efficient net-
work architecture for a video processing application, a different approach is needed.
It can be considered more energy efficient to run a video processing application ser-
vice locally (e.g. on the RPi), rather than an eDC or cDC if the following inequalities are
satisfied:
(ESD +NpEprc
)<
[Np(λTPe + σeλeDC
)+ σeηµets
]Abit (5.24)
144 Energy-Efficient Architecture for IoT Video Surveillance Services
(ESD +NpEprc
)<
[Np(λTPc + σcλcDC
)+ σcηµcts
]Abit (5.25)
where Abit is the video data volume (in bits) and Np, the number of face recognition
processing operations. All other parameters are defined in Table 5.7.
From (5.24) and (5.25), the left-hand-side of the inequalities represents the energy
allocated to storage of a local face database on the SD card (ESD) and the incremental
energy for local video processing. Without a means of separating the storage energy
allocated to a relatively small face database (13 MB) from that of the RPi Linux OS (> 1.5
GB), this estimate can be considered an upper limit. From the right-hand-side (RHS) of
(5.24) and (5.25), the terms in round brackets include both the transport energy incurred
in sending the video frames to a DC, and the energy consumed in the DC in handling the
video frames. The RHS also include the storage energy for a face database in the DC. It is
assumed that the DC face database is 100 MB in size.
Number of face recognition operations per day100 101 102 103 104
Ene
rgy
Con
sum
ptio
n (J
oule
s)
102
103
104
105
106
107
Local (RPi)
Edge
Cloud (Low)
Cloud (High)
Figure 5.17: Energy consumption as a function of number of face recognition operationsprocessed per day.
5.10 Conclusions 145
To run a face recognition application at a DC, video data frames must first be trans-
mitted to the DC. Let us assume the minimum video data required is one-second long.
Given that the pi camera outputs video of 5 megapixel frames at 30 f/s, the traffic volume
for a one-second long video stream with H.264 compression is ≈12 MB [164] (Abit = 96
Mbits). Figure 5.17 shows plots of energy consumption per day as a function of number
of face recognition operations per day for a local, edge and cloud-based service.
The results show that for local video processing, the energy cost of storage dominates
the total energy consumption of the service. The transport energy cost to transfer each
operation’s video data to a DC for processing in addition to the storage energy for the
DC face database is lower for an eDC than that of a cDC. Modelling with a much larger
DC face database (up to 1 TB) yielded no significant difference in the results. The results
further indicate that, for a low number of recognition operations per day (< 25), it is more
energy efficient to transport and process video data at an edge facility. However, the cost
of transport to a DC facility outweighs the cost of local processing for medium to higher
number of operations per day. Therefore, for a home setting (low usage scenario), it might
be deemed more energy efficient to process video data in an eDC with the assurance of
scale for more processor-intensive workloads. However, it is almost always more energy
efficient to process video data locally if the gateway device is capable of handling the
workload. If a face is not recognised from the local database (e.g. an infrequent visitor),
it might need to be referred to the eDC or cDC anyway.
5.10 Conclusions
In this chapter, an assessment has been made of the energy consumption of local (fog),
edge (fog) and cloud-based network architectures for IoT video surveillance services,
when considering video streaming, storage and processing energy implications. Using
energy consumption models for network equipment, experiments and measurements
where applicable, representative examples of network elements and relevant use-case
scenarios (live streaming, on-demand streaming and video processing), a first order esti-
mate of the energy consumption of the 4 main network architectures was presented. The
146 Energy-Efficient Architecture for IoT Video Surveillance Services
estimates were used as an indication of the energy efficiency of these network architec-
tures for IoT services.
A model for the power consumption of a network IP camera (IPcam), which takes into
account video parameters such as frame rate, video bit rate and pixel rate was given. The
results show that the power consumption of a network IPcam is linearly proportional to
its video pixel rate, but with an idle power that is greater than 90% of the total power.
The model can be used for similar new generation network IPcams.
To assess, evaluate and compare the energy consumption of IoT live streaming, on-
demand streaming and video processing, energy models for their respective network
architectures were developed. The results indicate that for accessing live video from an
IoT video surveillance system, it is more energy efficient to use direct (local) access from
the IPcam. Live streaming via an eDC or cDC is less efficient with no substantive benefit.
When considering on-demand streaming, however, the energy cost of local storage in
the IPcam can be 10 times higher than in a DC. However, storing video data closer to
the user (eDC) does not only reduce round trip latency, it is also shown to be a more
energy efficient option for data intensive applications like video surveillance services
and processor-intensive applications like face detection and recognition on video.
Deployment of the services for Internet of Things is ramping up with cloud comput-
ing being the de-facto option for many. The results shown here gives a clear indication
that fog computing may not only improve network QoS metrics for data intensive IoT
applications, but they also show that it can potentially save energy. In addition, the use
of fog computing may stave off core network and cDC capacity increments in the future.
Chapter 6
Power Consumption of IoT AccessNetworks
6.1 Introduction
IN previous chapters, it has been established that the IoT involves the use of many
sensing devices, actuators or ”things” connected over the Internet to measure, report,
and in some cases perform actions autonomously. Many of the ”devices” - referred to as
IoT devices in this thesis - within the IoT eco-system may be sensors, for example report-
ing periodically on their environment, sending small data packets at regular intervals,
with low-level data streams. However, some of the IoT devices may be video cameras,
sending a higher volume of data continuously, for which the traffic requirements may be
substantial.
Access networks provide the initial points of connection between the IoT network
and the Internet or the Cloud, enabling end-to-end bi-directional connectivity for the IoT
ecosystem. Some studies on energy consumption of the Internet (including the access,
edge and core network segment) have indicated that the access network is the least en-
ergy efficient segment of the Internet [40, 41, 43]. However, these studies tend to focus
on higher data access rates (above 1 Mb/s). Many IoT services or use-cases (e.g Smart
Farming or Environmental Monitoring) may involve low traffic volume per connected
device, with far greater numbers of connected devices [2, 24], each with low data access
rate demands, often outside the range considered by previous studies. A good under-
147
148 Power Consumption of IoT Access Networks
standing of the energy consumption of access network technologies for low bandwidth
requirements (as in IoT use-cases) is therefore important in assisting the ICT industry to
reduce its carbon footprint [165, 166] in future network deployments.
In modelling large networks (e.g. the Internet), it is impractical to perform direct
power measurements. A more practical approach is a combination of modelling and
measurement (where applicable) to determine the power consumption of large networks
[40, 70]. It is important to note that different access technologies provide different per-
user bandwidth levels, and have dissimilar power usage. In this chapter, a number of ac-
cess network technologies and architectures that are appropriate to IoT applications are
considered, then the more power-efficient access technologies for the IoT over a range
of traffic levels are identified. Using a range of data sources (equipment manufactur-
ers’ datasheets, some measurements and previous literature), an evaluation of current
wireline and wireless network architectures is presented. The power consumption and
energy-efficiency of the various technologies are compared for low bit rates (sub-1 Mb/s).
In the model, IoT device data traffic and their required protocol and signalling overheads
- which can be substantial for small-packet data flows - are treated as a whole. The chap-
ter concludes with suggestions on energy-efficient access network choices for future IoT
installations.
6.2 IoT Access Network Architecture
This section describes a range of access network architecture options enabling IoT con-
nectivity. The architectures considered here are shown in Figure 6.1. Access networks
provide connection from users and end-devices at the network edge, through to the In-
ternet core. Today, there are a number of access network technologies with different
transmission media (i.e. optical, copper, free-space wireless), traffic capacity and levels
of aggregation. The access network could include one or more wireline and wireless net-
work access technologies. It is assumed that the access network carrying aggregated IoT
traffic would broadly be the same or similar to today’s access networks, but handling dif-
ferent traffic loads. Hence the energy consumption of different types of access network
6.2 IoT Access Network Architecture 149
IoT Gateway
Wi-Fi
Internet Edge & Core
Networks
Passive Splitter
LPWA Base Station
LTE Base Station
DSLAM / MSAN
CPE Edge Node (EN)Remote Node (RN)
ONU
OMC
Modem
Modem
OLT
Ethernet Switch
IoT Gateway
IoT Gateway
IoT Gateway
IoT Gateway(IGW)
BLEANT+
ZigBee
IoT Device Network (IDN)
IDN
IDN
IDN
IDN
BLEANT+
ZigBee
BLEANT+
ZigBee
BLEANT+
ZigBee
IDN
PON
PtP
VDSL2
LTE
LPWA
LoRaWAN
NB-IoT
sigfox Wireless
Copper
Fibre
Figure 6.1: High-level diagram of the IoT access network architecture. There are fourmain nodes (left to right): the IoT gateway, the customer premises equipment (CPE) mo-dem, the remote node (RN) and the edge node (EN) located at the central office (CO) orlocal exchange. The nodes are vertically aligned while the technologies are horizontallyaligned.
technologies - both wireline and wireless - when handling traffic with IoT-like statistics
is considered.
In Wireline Access Networks, a physical connection joins the IoT gateway through
to the central office. Wireline access technologies benefit from a controlled medium with
high capacity, transfer speeds, low latency and security. For wireline access, considering
Passive Optical Network (PON), Point-to-Point Optical Network (PtP) and Very-high-bit-
rate Digital Subscriber Line (VDSL2), using a fibre to the node (FTTN) backhaul network.
With Wireless Access Networks, an IoT gateway connection to the Internet core is
established via a (at least one) wireless link. As seen in Figure 6.1, a wireless link can
be used as transport medium either between an IoT gateway and modem, or more com-
150 Power Consumption of IoT Access Networks
monly between the modem and base station (typical of mobile cellular network access),
transmitting traffic through to the central office. Wireless links are established using ra-
dio waves in free space with licensed or unlicensed frequency spectrum. Hence, they are
subject to the vagaries of an uncontrolled medium, including propagation losses, interfer-
ence, signal power attenuation, spectral inefficiencies and low capacity, and offer a range
of user access bit rates. Wireless technologies mostly differ in coverage range, frequency,
transmission power, multiplexing scheme and modulation. We consider recent wire-
less access technologies including Long Term Evolution (LTE), Low-Power Wide-Area
(LPWA) network and a shared (i.e. with hundreds of users) and unshared (i.e. single-user)
Wi-Fi network.
A detailed description of the afore-mentioned technologies are discussed in the sub-
sections below.
6.2.1 Passive Optical Network (PON)
PON is a widely used optical access network technology, owing to its low-cost, low main-
tenance sharing architecture [89]. With PON (see Figure 6.1), a cluster of customer sites
share a connection (single fibre line) to a network Optical Line Terminal (OLT) at the par-
ent exchange or CO via a passive splitter [87,89] (see Figure 6.1). Hence there is no power
requirement at the remote node. Common split ratios are 32-way (up to 20 km) or 64-way
(up to 10 km) [87, 89], but the selection of ratio depends on the planned traffic levels and
customer site density. An Optical Network Unit (ONU) at each customer site commu-
nicates with the OLT using Time Division Multiplexing (TDM). Gigabit Passive Optical
Network (GPON) is commonly deployed and can provide up to 2.4 Gb/s downstream
(DS) and 1.2 Gb/s upstream (US) bit rates between the ONU and OLT [89].
6.2.2 Point-To-Point Optical Network (PtP)
An optical Point-to-Point link provides a dedicated fibre connection from the customer
site to the CO [87]. As shown in Figure 6.1, an optical media converter is utilised as a
CPE modem connecting a single fibre line to an Ethernet switch, usually located at the
6.2 IoT Access Network Architecture 151
CO. This one-to-one relationship between customer site and fiber terminals (e.g. Ethernet
ports) at the CO makes a PtP first installation more costly and requires more powering
and maintenance [87]. PtP architecture also does not have a remote node, hence no power
is required. Such systems are more commonly used where customer traffic demands are
high (e.g. an office complex, hospital, or school).
6.2.3 VDSL2 using FTTN
VDSL2 (very-high-bit-rate digital subscriber line) - the latest of the xDSL generation -
takes advantage of existing copper pairs pre-installed for legacy telephony systems. In
DSL technology, fixed-line telephone service utilise lower frequency bands - to carry a 4
kHz voice signal - while higher frequency bands (up to 30 MHz) are allocated to broad-
band services [89,167]. Voice and data channels are separated at the customer site using a
low-pass filter. With VDSL2, customer sites connect via existing copper lines to a nearby
concentrating network element, often described as a DSLAM (Digital Subscriber Line Ac-
cess Multiplexer) or MSAN (Multi Service Access Node) as shown in Figure 6.1. In either
case, the node connects via optical fibre or FTTN network to its parent exchange. Shorter
copper loops as a result of a FTTN backhaul architecture provide increased bandwidth
capacity to the customer site. Using VDSL2, access rates of up to 100 Mb/s (symmetric)
are achievable [167].
6.2.4 Long Term Evolution (LTE) Wireless
Long-range wireless access is provided through the 4G LTE cellular network. LTE - a
recent generation (4G) of 3GPP’s cellular network technologies - is based on orthogonal
frequency division multiplexing (OFDM) in the downlink and single-carrier frequency-
division multiplexing in the uplink [90, 91]. In the model, LTE Rel-8 standard is used,
which offers higher peak data rates due to a large system bandwidth of up to 20 MHz
and higher order spacial multiplexing techniques [91]. In 4G LTE, the customer site mo-
dem connects to a local cellular base station/eNode-B, which in turn connects back to a
switch at the CO, typically via fibre (see Figure 6.1). The traffic capacity of 4G LTE de-
152 Power Consumption of IoT Access Networks
pends on the transmission path in terms of distance, topography, and interference from
other signals, and is generally lower than the theoretical figures quoted. Under ideal con-
ditions, downlink and uplink data transmission rates of up to 73.4 Mb/s and 20.6 Mb/s
respectively, are achievable for a 2x2 MIMO, 10 MHz bandwidth configuration [168].
6.2.5 Low-Power Wide Area (LPWA) Network
Low-Power Wide Area network is an emerging radio network technology specifically de-
signed for M2M/IoT communications with low power consumption and wide area cov-
erage being its most desired features. Operating in the unlicensed sub-1 GHz frequency
bands (i.e. 433 MHz, 868-915 MHz), LPWA caters for applications and services of partic-
ularly low data-rate (10 - 100 b/s typical; 100 kb/s max) demands and small daily traffic
volumes [92,93]. This limits the application of LPWA technology to a subset of M2M ser-
vices with infrequent small data transmissions, like smart meters, parking management
and asset/vehicle tracking. In a LPWA network, as depicted in Figure 6.1, a wireless
modem or module connected to or embedded in customer site end-devices (e.g. sensors)
links to the LPWA base station - known as a gateway/concentrator (referred to as a gate-
way from this point forward) - via the air interface. Traffic from hundreds/thousands
of these end-devices is backhauled from the gateway to an operator’s back-end network
and application servers (typically cloud applications) via Ethernet, 3G/4G/5G wireless
or optical IP links similar to those commissioned for LTE base stations. Depending on
the type of application, an LPWA end-device module may be deployed in remote loca-
tions without mains power and can be battery-powered with many years of continuous
operation [93]. Today, the LPWA deployment landscape is dominated by proprietary
technologies (e.g. sigfox, Narrowband IoT, Long Range (LoRa)) that are mostly incom-
patible with each other [169, 170].
LPWA is still a developing technology with presence in major markets around the
world [127,169]. Hence equipment information, traffic statistics and manufacturer datasheets
are not readily available in the literature.
6.3 Power Consumption Model 153
6.2.6 Wi-Fi Network
Wi-Fi network is a short-range wireless technology suited to cable-free connecting of user
devices to the customer site CPE modem. Wi-Fi is useful in providing connectivity be-
tween the IoT gateway and the network-facing modem, and in some instances between
the sensors or “things” and their gateway. Our model of an IoT access network - given
in Figure 6.1 - shows the IoT gateway connecting to the customer site CPE modem via
twisted-pair copper (i.e. Ethernet) for all wireline access technologies and via a USB mo-
dem dongle for 4G LTE Wireless. For LPWA network, a local IoT gateway is not required.
Although it is possible to utilise a Wi-Fi connection as a substitute for Ethernet to
connect the IoT gateway to the modem, the choice here is to model Wi-Fi connectivity
only with a passive optical network backhaul in the comparison. Wi-Fi operates in the
2.4 and 5 GHz ISM bands and under ideal conditions, can provide data rates in excess of
300 Mb/s [20].
6.3 Power Consumption Model
In this section, a description of a generic energy model for the IoT access network, dis-
cussed in the preceding section is presented. As seen from Figure 6.1, the IoT access net-
work includes four main nodes: the IoT gateway which aggregates traffic from the IDN,
the CPE modem which is technology-specific and located at the customer site, the remote
node (RN) which is usually at the street cabinet nearest to the customer site or base sta-
tion within few kilometres, and the exchange node (EN) located at the local exchange.
The power consumption per IoT gateway (Pi) for the complete IoT access network in
Figure 6.1 can be expressed as:
Pi = PIoT + PCPE +XRNPRN
NRN+XEN
PEN
NEN(6.1)
where PIoT, PCPE, PRN and PEN are the power consumed by the IoT gateway, the CPE mo-
dem, the remote node and the exchange node, while NRN and NEN represent the number
of users/ports of the remote and exchange nodes respectively. XRN and XEN are multi-
154 Power Consumption of IoT Access Networks
plier factors (where XRN, XEN > 1) used to account for the additional power consumed
for environmental control (i.e. cooling) of the facilities within which network nodes are
located plus overheads (e.g. power systems losses). These factors are akin to the power
usage effectiveness (PUE) ascribed to data centres [171]. The value of 1.5 is used as an
average overhead factor as in [42,70,171]. In the above model (6.1), the IoT gateway, CPE
modem and some remote nodes are naturally cooled.
The estimates presented in this chapter are developed using power consumption val-
ues for commercially available typical equipment, manufacturer data-sheets and some
experimental measurements (where applicable), to provide a first-order energy model.
The model in (6.1) thus provides an estimate of the power consumption of an IoT access
network considering IoT-like traffic statistics. Based on an in-house experimental mea-
surement of traffic levels of an example IoT service (i.e. a home automation system [15]),
a data access rate between 1 kb/s and 1 Mb/s is applied. Different access technologies
provide different data rates, which may also be dependent on the location of a customer
site with respect to its parent remote node. In order to achieve a fair comparison in the
analysis, practical and achievable data rates for each technology that are based on field
reports and data-sheets are used.
In the following section, a discussion of some power consumption modelling ap-
proaches used to represent the nodes of the IoT access network technologies described in
section 6.2 is given.
6.4 Network Element Modelling
This section describes the modelling of each network element in the IoT access network.
To aptly represent the different usage characteristics of networking equipment, they are
classified into three main types:
1. Unshared Network Elements
2. Fixed-user Shared Network Elements
3. Multi-user Shared Network Elements
6.4 Network Element Modelling 155
The unshared network element type refers to a ”single-user” equipment (e.g. modem),
hence its power consumption is allocated to one user in the model. Some shared network
elements may serve a fixed number of users (e.g. DSLAM) while others may service an
indeterminate number of users and services (e.g. Ethernet Switch). We therefore model
shared network elements as either ”fixed-user” equipment in which its power consump-
tion is allotted to a finite number of users or as ”multi-user” equipment in which power
consumption is allotted based on the access bit rate traffic generated by the users and
services traversing the network equipment. These energy models are discussed below.
6.4.1 Unshared Network Elements
An unshared network element refers to a network equipment that deals with traffic ded-
icated to a ”single-user” (e.g. modem for an xDSL service, an ONU). To define a ”single-
user”, consider two sets of use-cases; In the first set, an IoT service installation is in a
home or small business setting (e.g. HAS) where it shares a CPE modem with many
other traffic generating services (e.g. Facebook, Youtube, Outlook), often with higher
traffic flows than the IDN. However, in the second set (e.g. Environmental Monitoring,
Smart Agriculture), the IoT service (i.e. sensors and IoT gateway) installation may be
located in settings where the CPE modem is dedicated to carrying traffic flows generated
by the IoT service exclusively. The latter is considered here. Hence, for the modelling of
an unshared network element, a ”single-user” is defined as a single IoT service (i.e. one
IDN) comprising of many traffic generating IoT devices with an IoT gateway, communi-
cating via one CPE modem to the cloud. An unshared network element is thus referred
to as a single-user network element.
Consider a single IoT service with many IoT devices within its IDN. The ith IoT device
generates a steady stream of bits with rate Rdev,i(t) (bit/sec), where i = 1, 2, ..., n for n
number of IoT devices within the IDN. The sum of all data streams from IoT devices
within the IDN, RIoT(t) (bit/sec), can be expressed as:
RIoT(t) =
n∑i=1
Rdev,i(t) (6.2)
156 Power Consumption of IoT Access Networks
From the model of a generic network equipment in Chapter 2, we established that the
power consumption of a single-user network device can be represented by equation (2.1).
CPE modems tend to be ”always on” with continuous power consumed irrespective of
the traffic they handle. From an in-house measurement of an ADSL2+ modem [9], simi-
lar measurements in the literature [96] and the European Commission’s code of conduct
on energy consumption of broadband equipment [124] (referred to as the EUCoC), it is
known that the idle (no-load) power of a CPE modem (Pidle) can be a significant pro-
portion (90% or more) of its maximum power (Pmax); hence Pidle = βPmax (for β > 0.9)
where β is the fraction of maximum power consumed by the network equipment in its
idle state. For the equipment maximum capacity Rmax, the incremental energy portion
of equation (2.1) given by ERmax = (1 − β)Pmax << Pidle. The power consumption of a
single-user network element dealing with traffic generated by a single IoT service RIoT(t)
is given as:
Psu(t) = Pidle + ERIoT(t) ≈ Pidle (6.3)
Given that the power consumed by the network element Psu(t) varies only slightly with
load, the constant power consumption model given in (6.3) is adopted; hence power
consumption values from manufacturer data-sheets are used as representative of typical
single-user network elements in the modelling.
6.4.2 Fixed-user Shared Network Elements
A network element at the remote node or exchange node can be designed and dimen-
sioned to carry data traffic from a fixed number of users. For example, a multi-slot chassis
MSAN [172] contains management and switching cards plus a number of line cards fea-
turing several ports per line card. The number of ports provisioned on the downstream
(user) facing line cards often matches the number of users while the upstream ports back-
hauls user traffic to the Internet core. A schematic diagram of the main components of a
fixed-user shared network element is shown in Figure 6.2.
Consider a fixed-user shared network element (e.g. an MSAN) with a set number
6.4 Network Element Modelling 157
Switchfabric
Upstream port
Backplane
Management
Downstream (user) port
Line card
Line card
Line card
Line card
^PP
1
2
N
Figure 6.2: Schematic diagram of the internal structure of an MSAN
of user ports N (assuming 1 port per user modem) as shown in Figure 6.2. The power
consumption of the equipment can be divided into two main portions: 1) the dedicated
portion P (on the left hand-side) representing the power consumed by the line cards and
their downstream (user) ports and 2) the shared portion P (on the right hand-side) rep-
resenting the power consumed by the shared switch fabric, management and upstream
ports.
Consider a steady stream of traffic passing through one of the identical line card ports
j given as Rport,j , where j = 1, 2, ....,m, with m being the number of ports per line card
for L line cards. Assuming all network management traffic associated with each port is
an inclusive fraction of Rport,j , and for a fully-equipped network element with all ports
provisioned (i.e. number of ports = number of users), the total traffic through a fixed-
user network element Rtot can be expressed as: Rtot = (L×∑m
j=1Rportj). Equation (2.1),
expressed as a function of bit rate R (bit/sec) is then applied. The power consumption
of a fixed-user shared network element Pfu(Rtot) as a function of the total traffic (Rtot) is
calculated as the sum of the dedicated portion (P ) and shared portion (P ) (as indicated
158 Power Consumption of IoT Access Networks
in Figure 6.2) and expressed as:
Pfu(Rtot) = LmPidle + Lm∑j=1
EjRport,j + Pidle + ELm∑j=1
Rport,j
= LmPidle + Pidle + LmRport(E + E) (6.4)
where Pidle is the no-load power for each port per line card and Pidle, the no-load power
of the shared components (i.e. switch fabric, etc.), while E and E are the incremental
energy per bit per port and that of the shared components respectively.
Power consumption measurements of a similar network element [173] show that its
idle power can be about 92% or more of its full-load power, with the slope denoted as E
being very small. Thus even for a maximum bit rate of each port per line card Rport(max),
the expression in (6.4) may be simplified to:
Pfu(Rtot) ≈ NPidle + Pidle (6.5)
given that LmPidle + Pidle >> LmRport(max)(E + E) with the total number of ports N =
Lm. Under the same conditions Rtot = N〈Rport〉, where 〈Rport〉 (bit/sec) represents the
average traffic per port. Using equation (6.5), the expression for power consumption per
port (user) for a fixed-user shared network element Pfu(Rport) can be written as:
Pfu(Rport) ≈Pmax
N(6.6)
where Pmax ≈ (NPidle + Pidle) which is consistent with power consumption measure-
ments of a similar equipment described in [173]. Since a single IoT service uses one
modem which connects to one port, Rport = RIoT; hence Pfu(RIoT) ≈ Pmax/N .
6.4.3 Multi-user Shared Network Elements
A multi-user network element refers to an item of equipment dealing with traffic from
many (hundreds/tens of thousands) users, services and application data flows at any
given time. To determine the power consumption of a network user or service accessing
6.4 Network Element Modelling 159
a multi-user network element, a portion of the equipment power is allocated to that user
or service. To achieve this goal, the number of users or services traversing the network
element at any given time is an essential quantity. But determining a finite number of
users or services simultaneously accessing such an equipment may not be practical. One
method of modelling what can be described as a ”highly shared” network element is the
”traffic-based” approach described in [70].
Let us consider this approach, which allocates a portion of the total element power
consumption (idle and incremental power) P (Rtot) to a service with data traffic rate Rsrv,
as a fraction of the total element data rate Rtot. The power consumption attributable to
this service at time t is given as:
Psrv(t) =Pidle
Rtot(t)Rsrv(t) + ERsrv(t) (6.7)
whereE is the incremental energy per bit given byE = (Pmax−Pidle)/Rmax, Pmax and Pidle
are the network element maximum and idle power consumption, andRmax, its maximum
data rate in bits/sec.
We can calculate the energy consumption of the service by taking the integral of Psrv(t)
over a required period of time T (e.g. a diurnal cycle), given as:
Qsrv(T ) = Pidle
∫ T
0
Rsrv(t)
Rtot(t)dt+ E
∫ T
0Rsrv(t)dt
= Pidle
∫ T
0
Rsrv(t)
Rtot(t)dt+ EBsrv(T ) (6.8)
where Bsrv(T ) is the total data volume of the service in bits over a time duration T . We
can then obtain the average power consumption of the service 〈Psrv(T )〉, given by:
〈Psrv(T )〉 =Qsrv(T )
T= Pidle
(1
T
∫ T
0
Rsrv(t)
Rtot(t)dt
)+ E
(Bsrv(T )
T
)
= Pidle
⟨Rsrv
Rtot
⟩+ E〈Rsrv〉 (6.9)
given that the mean value 〈X〉 = 1/T∫ T0 X(t)dt.
160 Power Consumption of IoT Access Networks
In (6.9), the fractional data traffic rate of the service passing through the network element,
relative to the total data traffic is given by Rsrv(t)/Rtot(t), and its mean value 〈Rsrv/Rtot〉
over time T . This mean value can range from 0 to 1 as expressed by the inequality:
0 6
⟨Rsrv
Rtot
⟩6 1 (6.10)
Since the value of Rsrv(t)/Rtot(t) at a given time is highly dependent of the type of ser-
vice and its data flow rate, let us consider three example service scenarios to examine the
effect of the mean value 〈Rsrv/Rtot〉 on the attributable energy consumption allocation to
the service.
Scenario 1: For a given service with a data rate Rsrv(t) ≈ Rconst (e.g. video surveillance
service), where Rconst is a constant data rate, then we get:
⟨Rsrv
Rtot
⟩≈⟨
1
Rtot
⟩Rconst (6.11)
Then from (6.9),
〈Psrv〉 ≈(Pidle
⟨1
Rtot
⟩+ E
)Rconst (6.12)
Scenario 2: For a given service with a data rate that is approximately proportional to the
total data rate of the network element (a service with data rate similar to a typical diurnal
cycle), then Rsrv(t) ≈ kRtot(t) with 0 6 k 6 1, hence,
⟨Rsrv
Rtot
⟩≈ k (6.13)
Then from (6.9),
〈Psrv〉 ≈ Pidlek + E〈Rsrv〉 (6.14)
Scenario 3: For a given service with a data rate that is negligible in comparison with the
total data rate of the network element, such that Rsrv(t)/Rtot(t) 1, then,
⟨Rsrv
Rtot
⟩≈ 0 (6.15)
6.4 Network Element Modelling 161
Then from (6.9),
〈Psrv〉 ≈ E〈Rsrv〉 (6.16)
On the flip side, the service may have its maximum data traffic rate at a point when the
aggregate diurnal data traffic rate is at its minimum (i.e. Rsrv(t)/Rtot(t) ≈ 1), and vice
versa during the peak traffic period/busy hour, which will then give Rsrv(t)/Rtot(t) ≈ 0,
for which (6.16) applies. For the former instance, (6.9) can be written as:
〈Psrv〉 ≈ Pidle + E〈Rsrv〉 (6.17)
Both data and management traffic flowing through the network element - denoted as
background traffic (Rbgd) - can be expressed as the sum of individual service data flows
Rsrv,l at a given time t given by: Rbgd(t) =∑n
l=1Rsrv,l(t), where l = 1, 2, ...., n, for n
number of services. The traffic generated by an IoT service, RIoT(t) is regarded as an
Pidle
Pmax
Rmax
Pow
er
Con
sum
pti
on
(W
att
s)
Load (bit/sec)
Rbgd+RIoT
PIoT
Rbgd
Figure 6.3: Power consumption profile of a multi-user shared network element
addition to the background traffic Rbgd(t) as depicted in Figure 6.3; hence the aggregate
data traffic rate Rtot(t) passing through the network element at time t can be expressed
as:
Rtot(t) = Rbgd(t) +RIoT(t) (6.18)
162 Power Consumption of IoT Access Networks
Due to the burstiness and unpredictability of network traffic, operators manage the equip-
ment operating load utilization U and set by policy a maximum load utilization Umax,
such that 0 < U < Umax, in-order to mitigate possible congestion and minimise packet
loss. The total traffic Rtot(t) is a fraction of the maximum traffic level (Rmax) the network
element can support, such that (6.18) can be written as:
typical of a corporate or industrial setting with many users and services. The multi-user
shared network element model, given in (6.20) is employed for the shared Wi-Fi type and
the unshared network element model, given in (6.3) for the single-user type.
For the unshared Wi-Fi network model, an Ad-net GPON HGU ONT with built-in
Wi-Fi support [175] is used as an unshared Wi-Fi CPE modem with a reported power
consumption of 6 W. This Ad-net ONT represents a single IoT service in the model. PON
access is used as a backhaul network.
For a shared Wi-Fi network model, the Cisco ME 4624 Gateway [182] with built-in Wi-
Fi support, is employed as an Access Point (AP). The ME 4624 AP consumes a maximum
of 15 W. For shared Wi-Fi access modelling, equation (6.21) is applied to determine the
“per-bit” energy consumption of the AP, whilst considering a maximum bit rate of 150
Mb/s and an average utilization 〈U〉 of 20% (i.e. 〈Rbgd〉 ≈ 30 Mb/s). The energy-per-bit
of the shared Wi-Fi AP for 〈U〉 = 0.2 is thus calculated as 460 nJ/bit. Similar to the unshared
network, PON access is used for traffic backhaul to the network core.
Table 6.3 shows calculated energy-per-bit values for the AP and a mid-range Ethernet
Switch. The modelling of the Ethernet switch is discussed in section 6.6.5 below. The
6.6 Estimating the Power Consumption of Network Elements 165
values in table 6.3 are subject to traffic usage assumptions and are dependent on the
average utilization of the network element.
Table 6.3: Energy per bit values of shared network elements
Network Element Energy per bit (nJ/bit)Ethernet Switch [Mid-Range] 37Shared Wi-Fi Access Point 460
6.6.3 Passive Optical Network Access (PON)
In modelling a passive optical network, a GPON installed network with a 32-way split
ratio is considered. At the remote node a single fibre is passively split to serve up to 32
ONUs (28 ONUs in practice). Hence no power is required for the remote node (PRN = 0).
A Tellabs 1134 [154] optical line terminal is used as the exchange node for PON (see
Figure 6.1). An OLT can serve a determinate number of ONUs which depends on its
configuration; hence the fixed-user power model is used for the optical line terminal.
The 1134 OLT has 16 GPON downlink ports and twelve 1 GbE uplink interfaces. At
full configuration, with full port occupancy, the 1134 can service up to 512 customer sites
(i.e. 16 GPON ports x 32) with a 32-way split ratio per fibre line. The 1134 has a maximum
power consumption of 480 W [154]. Using the fixed-user shared network element model
in (6.6), the power per IoT service for the OLT is calculated as 0.9 W. The ONU used in
the model is the Zhone ZNID-2301 [177] with a power consumption of 3-5 W.
6.6.4 Point-to-Point Optical Network Access (PtP)
Point-to-Point optical access offers the highest bit rates to the subscriber by virtue of its
dedicated fibre links between the subscribers’ CPE modem and an exchange node (see
Figure 6.1). There is no remote node in a PtP installation, therefore no attributable power.
To model a PtP optical network, the low-range Cisco Catalyst 3750-24FS Ethernet
switch [183] is considered as the local exchange node. The 3750 switch provides 24 100FX
Ethernet ports and 2 SFP Gigabit Ethernet uplink ports, with a switching capacity of
166 Power Consumption of IoT Access Networks
32 Gb/s. The switching power consumption measured at 5% and 100% throughput is
reported as 54.2 W and 55.2 W [183] respectively and each port consumes 1.5 W [180].
Total idle and maximum power of the switch is 93.2 W and 94.2 W respectively, giving
a 1% load proportionality. PtP architecture ascribes 1 port per customer site (i.e. CPE
modem), hence the fixed-user shared network element model given in (6.6) is employed
for the low capacity Ethernet switch. The calculated power per IoT service is 3.9 W. A
CTC Union FRM-220 [178] optical media converter which consumes 4 W is selected as
the CPE modem for PtP architecture.
6.6.5 VDSL2 Access Network
A VDSL2 installation takes the form of a FTTN/FTTB network, terminating at the DSLAM
or MSAN, and a dedicated copper line from the DSLAM to the customer site. This model
is based on a fully-equipped ZyXel IES-5106 MSAN [172] with all ports occupied at the
remote node. At full configuration, the IES-5106 holds five VDSL2 line cards and a man-
agement switch card. Each VDSL2 line card supports 24 customer sites and consumes
about 66 W. Full-load power consumption of the IES-5106 MSAN is 391 W. The CPE mo-
dem for VDSL2 architecture is an Eltek V7601-A1 [174] which consumes 7 W.
For a model of an MSAN, a reported power consumption measurement of a similar
equipment showing its idle power being approximately 92% of its peak power is con-
sidered as a benchmark [173]. Furthermore, the number of port connections that can be
provisioned at the MSAN is fixed. With respect to the IES-5106, only 61 W of its maxi-
mum power 391 W can be shared amongst users on a traffic usage basis. Because most
of the IES-5106’s power is consumed in the idle state (> 84%), which is consistent with a
measured benchmark equipment, the slope E from Figure 2.5 for the MSAN is substan-
tially flat. Hence the power per IoT service for a VDSL2 remote node is calculated using
a fixed-user shared network element model given in (6.6), thus calculated as 3.3 W.
The exchange node is the mid-range Cisco ME 3800X-24FS Carrier Ethernet switch
providing 24 Gigabit Ethernet SFP ports with two 10 Gigabit SFP+ ports for traffic back-
haul, and a full-duplex switching capacity of 44 Gb/s [181]. The 3800X aggregates traffic
from multiple MSANs and base stations with an indeterminate number of users or ser-
6.6 Estimating the Power Consumption of Network Elements 167
vices [181]; hence a multi-user shared network element model is employed. In the model
an average utilization of 20% is assumed (i.e. 〈U〉 = 0.2), resulting in an average back-
ground traffic 〈Rbgd〉 = 4.8 Gb/s. Since the background traffic is far greater than that
from the IoT service (1 kb/s 6 RIoT 6 1 Mb/s), equation (6.21) is applied to calculate
the energy-per-bit of the Ethernet switch, which is calculated as 37 nJ/bit as given in
Table 6.3.
6.6.6 4G LTE Network Access
A simple 4G LTE access network involves a 4G LTE wireless modem at the customer site
connecting via a wireless link to the closest base station (i.e. the remote node), which in
turn connects to an Ethernet switch located at the local exchange (see Figure 6.1). Using
the 2012 state-of-the-art reference base station put forth by the Earth Project [23], an LTE
Rel-8 Macrocell BS, with 10 MHz bandwidth frequency division duplexing (FDD), is con-
sidered as the 4G LTE remote node. The Rel-8 BS has a 2x2 MIMO configuration with 2
transceivers per sector and a theoretical peak throughput of 73.4 Mb/s and 20.6 Mb/s in
the downlink and uplink respectively, achievable under ideal conditions [168]. A single
4G LTE BS can maintain hundreds or thousands of active users or gateway devices at any
given time and is therefore considered as multi-user shared network element.
A ZTE MF823 4G LTE wireless modem is utilized as the CPE modem. The plot in
Figure 6.4 shows the measured power consumption of the 4G Dongle having three dis-
tinct power levels: standby, idle and active, the latter accounting for power consumed
when transmitting and receiving data. The three levels are consistent with the EUCoC
[?], which sets out basic principles for component manufacturers, service providers and
network operators for improving energy efficiency of broadband equipment. In the ac-
tive state, the 4G dongle consumes an average of 1.4 W.
As shown in Figure 6.1, the BS connects to the core of the network via an Ethernet
switch at the local exchange. The Cisco 3800X is used to represent the Ethernet switch in
the model of a 4G LTE access network. Power consumption contribution of the exchange
node for a 4G LTE network access is hence calculated on a traffic volume basis using the
168 Power Consumption of IoT Access Networks
Figure 6.4: Power consumption measurement of a ZTE MF823 4G Wireless Dongle
energy-per-bit calculation in (6.21).
For the BS, it is assumed that the power consumption grows proportionally with the
number of transceivers. The total power consumption of a BS [23] with N transceivers
can be expressed as:
PBS = N × PPA + PRF + PBB
(1− σDC)(1− σMS)(1− σcool)(6.22)
where PPA, PRF and PBB account for the power draw by the power amplifier (PA), the
transceiver module and baseband unit (BBU), σDC, σMS, σcool account for the loss factors
due to the BS DC-DC power supply, main supply and cooling overheads (see table 6.4)
respectively. The radio frequency (RF) and signal processing components of a BS include
a number of functions common to both transmitter and receiver subsystems, for which
the power consumption is spilt equally, except where the data for each of the subsystems
is available separately. The RF power amplifier (PA) in the BS however is used only in
the transmitter, and its consumption is fully assigned to the downlink energy usage. The
transmitter PA power consumption is treated as load-proportional in accordance with
[23], whilst the consumption of the remaining components is considered to be indepen-
dent of load.
6.6 Estimating the Power Consumption of Network Elements 169
Table 6.4: Power consumption values for 4G LTE base station components from [23]
Base Station Component Transceiver (TRX)*RF Chain (TX / RX) 5.7 / 5.1 W
Baseband Unit 14.8 WPA Power (Idle / Max) 38.8 / 102.6 W
Main Supply Loss (σMS) 7%
Cooling Loss (σcool) 9%
DC-DC Loss (σDC) 6%
* A 2x2 MIMO base station described in [23] contains 2transceivers per sector, making a total of 6 transceiversfor a 3 sector base station. The above values apply to eachtransceiver.
Consider a single sector of the 2012 reference BS running in a 2x2 MIMO mode with
two transceivers per sector (N = 2). Using (6.22), the power consumption of the transmit
section (downlink) is calculated as 130 W and 291 W in the idle and full-load states,
respectively. For the receive section (uplink) of the BS there is no PA (hence PPA = 0).
The power consumed by the receiver of the BS is thus considerably lower than that of the
BS transmitter, hence the incentive for modelling the transmitter and receiver separately.
The power consumption of the receiver is calculate as 31 W.
Considering a one sector base station dealing with trafficRBS(t) at time t. Since the BS
is classified as a multi-user shared network element, equation (6.18) is applied to obtain
an expression for the aggregate traffic load RBS(t) handled by the BS sector as:
RBS(t) = Rbgd,BS(t) +RIoT(t) (6.23)
where Rbgd,BS(t) represent the background traffic of the BS sector and RIoT(t), the traffic
generated by an IoT service.
Wireless links exhibit uncontrollable channel characteristics with attributes such as
fading, interference and propagation losses, unlike a much more controlled medium in
the case of wireline. For 4G LTE, the average user access rate is dependent on the these
attributes, the user’s location with respect to the base station, the number of user requests
at that instant and available resource blocks to serve those users.
To estimate the BS sector’s uplink and downlink background traffic, consider a useful
170 Power Consumption of IoT Access Networks
approximation to a mobile network diurnal traffic profile given in [13], which is based
on data obtained from a telecoms operator. The BS hourly average traffic load levels, as a
percentage of the daily average for a dense urban area is listed in Table 6.5. A profile trace
of the BS hourly average traffic load, over a 24 hour period is shown in Figure 6.5. The
figure shows that the BS hourly traffic load utilization ρ(t) varies from 20% of the daily
average load Rave,BS during the early hours of a day to 140% of the daily average load
during busy-hours. Hence the BS sector background trafficRbgd,BS(t) = ρ(t)Rave,BS, given
that Rave,BS = 〈U〉Rmax,BS from equation (6.19), where 〈U〉 is the traffic load utilization
over the diurnal cycle and Rmax,BS the maximum throughput of the one sector BS.
Table 6.5: Traffic level distribution of a dense urban area over a 24 hour period
Traffic ProfileLevel ρ (%) Duration (hrs)
20 240 4100 4120 8140 6
Time of Day (hours)0 2 4 6 8 10 12 14 16 18 20 22 24
Rel
ativ
e Lo
ad L
evel
(%
)
0
20
40
60
80
100
120
140
160
Figure 6.5: Daily BS traffic load profile of a dense urban area. 100 % corresponds to thehourly traffic averaged at intervals over the 24 hour day [13].
Unlike other network elements listed in Table 6.2, the base station consumes unequal
6.6 Estimating the Power Consumption of Network Elements 171
power in each direction. A study in [184] reports that 87% of the BS power is consumed
in the downlink, with only 13% consumed in the uplink. These percentages are similar
to the BS (considered here) power consumption values as given in table 6.4 (i.e. ≈ 90%
downlink and 10% uplink). The substantial difference is due to the absence of the base
station PA in the uplink RF chain. For this reason, the incremental energy portion in
the uplink power model given below is disregarded. Using equation (6.20) the power
consumption for a one sector 4G LTE BS, PBSsect(RBS) can be written as:
PBSsect(RBS) =
(Pidle
RBS+Pmax − Pidle
Rmax,BS
)RBS (6.24)
where RBS is the aggregate average BS sector traffic load derived from (6.23), Pmax and
Pidle are the maximum and idle BS sector power and Rmax,BS the maximum BS sector
throughput.
Taking the average power for a 5:1 uplink to downlink traffic ratio (see Figure 3.17),
the expression for the 4G LTE BS sector uplink power PBSsect(RIoT,ul) attributable to an IoT
service can be expressed as:
PBSsect(RIoT,ul) =
(Pidle,ul
ρRave,ul +RIoT,ul
)RIoT,ul (6.25)
where RIoT,ul represents the uplink traffic generated by the IoT service, Rave,ul represents
the aggregate average uplink throughput of the BS and Pidle,ul represents the BS idle
power consumption of the uplink chain. As described above, there is no PA in the uplink
RF chain, hence the incremental energy portion of equation (6.24) is neglected.
Similarly, the expression for 4G LTE BS sector downlink power consumptionPBSsect(RIoT,dl),
attributable to an IoT service can be expressed as:
PBSsect(RIoT,dl) =
(Pidle,dl
ρRave,dl +RIoT,dl+Pmax,dl − Pidle,dl
Rmax,dl
)RIoT,dl (6.26)
where RIoT,dl represents the downlink traffic generated by the IoT service, Rave,dl repre-
sents the aggregate average downlink throughput of the BS, Rmax,dl represent the maxi-
mum downlink throughput of the BS and Pidle,dl and Pmax,dl represents the BS idle and
172 Power Consumption of IoT Access Networks
100 200 300 400 500 600 700 800 900 1000
IoT Data Access Rate (kb/s)
0
5
10
15
20
25
IoT
-rela
ted L
TE
BS
Secto
r P
ow
er
(W)
Background Traffic Level
20% of average background traffic level
40% of average background traffic level
100% of average background traffic level
120% of average background traffic level
140% of average background traffic level
Figure 6.6: 4G LTE BS sector power attributable to IoT traffic as a function of IoT dataaccess rate for 5 different background traffic level profiles.
maximum power consumption of its downlink chain. The aggregate average downlink
and uplink throughput of a 10 MHz 2x2 4G LTE BS sector is reported as 12 Mb/s and 6
Mb/s respectively, given typical deployment scenarios [168].
With the hourly average traffic utilization values ρ shown in Figure 6.5, the BS sector
aggregate average throughput Rave [168] and the expressions for power given in (6.25)
and (6.26), the BS sector power attributable to an IoT service accessing a BS remote node,
for throughputs between 1 kb/s and 1 Mb/s is calculated as:
Figure 6.6 shows the BS sector power attributable to IoT traffic for the five different traffic
utilization profiles of a typical weekday.
6.6 Estimating the Power Consumption of Network Elements 173
6.6.7 LPWA Access Network
The most recent of all the architectures discussed so far is a LPWA network architecture.
In a typical LPWA network, low-powered, low data-rate IoT devices transmit and/or re-
ceive small data packets to or from an LPWA gateway, using purpose-built physical layer
(PHY) modulation schemes via a wireless link. As described in section 6.2.5, a number
of LPWA network technology options are available today (e.g. LoRa) with similar net-
work architectures. In this section, the LoRa architecture is described and modelled as a
representative example of an LPWA network.
In LoRa technology [185], the physical layer (PHY) modulation technique is based
on chirp spread spectrum (CSS) with integrated forward error correction (FEC), while
the MAC layer is defined by the LoRaWAN protocol that is standardised by the LoRa Al-
liance [186]. While the PHY permits different protocol architectures (i.e. Mesh, Star, etc..),
the MAC layer is designed for a long-range star topology network [185]. The gateway, as
shown in Figure 6.7, internally consists of an RF front-end module and a ”host” micro-
controller unit (MCU) (similar to Raspberry Pi, Beagle Bone) or PC. The role of a MCU
in the gateway is to prepare data bits - wrapped as IP packets - for transport through the
Internet to the IoT services back-end servers. The MCU connects to an Ethernet switch
or a 3G/4G/5G network (see Figure 6.7) en-route to the Internet core. LoRaWAN defines
10 channels: 8 multi data-rate LoRa channels with 125 kHz demodulation bandwidth
and a range between 250 b/s to 5.5 kb/s, one high capacity channel (11 kb/s) for inter-
gateway backhaul purposes and one FSK channel with data rate up to 50 kb/s [185].
LoRa technology transceivers support 6 spreading factors (SF7 to SF12) corresponding to
different bandwidth options (SF7 being the best and SF12 the worse). The choice of SF is
dependent on the channel signal-to-noise ratio, the link budget and range [186].
In modelling an LPWA network, the IMST iC880A module [14] is adopted as the RF
front-end of a LPWA gateway and a Raspberry Pi [163] as the MCU. The iC880A is based
on Semtech’s SX1301 radio (LoRa technology [179]) which operates on the 868 MHz ISM
frequency band with 10 programmable channel paths. The SX1301 has a line-of-sight
range of up to 15 km but only a few kilometres in an urban environment. Wireless trans-
174 Power Consumption of IoT Access Networks
1
iC880A(Radio Module)
Host(MCU)
Higher data rate
Higher range
234
Base Station (Gateway)
IoT DevicesEthernet
3G/4G/5G
LoRaWANBack-end Servers
Access, Metro & Core Networks
Internet
Figure 6.7: Schematic diagram of a LPWA network structure [14]
mission is achieved via shared medium with uncontrollable characteristics including in-
terference, propagation loss and fading, which influence the link spectral efficiency and
bandwidth, resulting in a range of possible transmission data rates. The iC880A therefore
employs ”Dynamic Data Rate Adaptation” to achieve a channel data throughput range
from 0.3 kb/s, for IoT devices in ”difficult” locations, and up to 6 kb/s (for a good chan-
nel) is achievable by IoT devices closer to the gateway [14]. We assume all 8 channels
are programmed as LoRa channels, assuming SF7 for good channels with the maximum
per-channel bit rate of 6 kb/s. The combined maximum data-rate of the gateway for 8
channel paths simultaneously occupied by nodes close to the gateway is 48 kb/s, pro-
vided that propagation conditions are favourable.
Let’s consider a single LPWA modem embedded with a sensor or actuator (i.e. IoT de-
vice) that communicates with one LPWA gateway as shown in Figure 6.8. The IoT device
’a’ transmits with a data-rate ra,j , where device ’a’ belongs to an IoT service j (e.g. an
agricultural installation with sensors measuring soil moisture). IoT service j can have
many IoT devices within its IDN, such that a = 1, 2, ..., A. Similarly, the LPWA gate-
way can communicate with many IDNs, with each IDN belonging to an IoT service j,
for j = 1, 2, ..., J . The LPWA gateway contains a radio module with a total of K chan-
nels (where K = 8 for the iC880A radio module). Each channel k (where k = 1, 2, ...,K)
communicates with a single IoT device ’a’ from an IoT service j at any given time.
The aggregate data-rate of the IoT service j is then given asRLPWA,j(t) =∑Aj
a=1 ra,j(t).
6.6 Estimating the Power Consumption of Network Elements 175
Assuming the IoT devices transmits at the same bit rate, then the effective data-rate of the
service RLPWA,j = Aj × ra,j .
Raspberry Pi
Mo
d/D
emo
d(R
adio
)
SPI
SPI
Pac
ket
Han
dle
r
1
2
A1
1
2
A2
1
2
AJ
K
1
2
iC880A Radio Module Host (MCU)
r1,1
r2,1
rA,J
Channels
LPWA Gateway
IoT Service 1
IoT Service 2
IoT Service J
LoRa Wireless Link
ToLoRaWAN
Servers
Figure 6.8: Schematic diagram of the internal structure of an LPWA Gateway
If a number of IoT services (where 1 IoT service≡many sensors/actuators≡ 1 owner
/user) with data-rates RLPWA,j connects through the same LPWA gateway, the effective
data-rate of the gateway is given by:
Reff (t) =J∑j=1
K∑k=1
RLPWA,j,k(t) (6.28)
The effective gateway utilization U is then given as:
U =Reff (t)
Rmax=
∑Jj=1
∑Kk=1RLPWA,j,k(t)
Rmax(6.29)
where Rmax is the maximum bit rate of the gateway.
Since LPWA gateways provide wide coverage (up to 15 km for LoRa) and could reach
tens of thousands of IoT devices, which in turn may be part of hundreds of unique IDNs
owned by many users, the gateway is deemed as a ”highly shared” network element.
From Figure 6.7, it is shown that the gateway includes a Raspberry Pi which, from mea-
surements in Chapter 5, exhibits slight linear load-dependence similar to network ele-
ments described in [96]. Furthermore, current consumption characteristics of the IMST
176 Power Consumption of IoT Access Networks
iC880A module (see Figure 6.9) seem to scale linearly with number of channels, provided
the channels are fully utilised. Hence, a multi-user power model is applied for the gate-
way as given in (6.20).
Number of Paths / Channels0 2 4 6 8 10 12
Cur
rent
(m
A)
0
50
100
150
200
250
300
350
400
450
500
linear current (typical)
X: 0Y: 171.3
Figure 6.9: Current consumption characteristics of the iC880A RF module with a supplyvoltage of 5V from datasheet in the blue curve and the dash lines are an extrapolation ofthe datasheets values.
In the modelling of an LPWA access network (see Fig 6.1), the IoT device’s LoRa
transmitter modem, which can be embedded within the IoT device circuitry, is not con-
sidered. The power consumption of an LPWA gateway PL(RLPWA,j) for an IoT service j
with data-rate RLPWA,j is then given as:
PL(RLPWA,j) =
(Pidle∑J
j=1
∑Kk=1RLPWA,j,k
+
(Pmax − Pidle
Rmax
))RLPWA,j (6.30)
where Pidle and Pmax represent the idle and maximum power consumption of the gate-
way.
Current consumption characteristics of the iC880A module [14] is shown in Figure 6.9.
The supply voltage for ”typical” operation is 5 V. Based on Figure 6.9, the idle power
(Pidle) and maximum power (Pmax) consumption of the RF module is calculated as 0.9 W
6.7 Results and Discussion 177
and 2.2 W respectively. From measurements, the Raspberry Pi consumes about 2.2 W.
6.7 Results and Discussion
In this chapter, a first-order estimate of the power consumption of different access tech-
nologies, using information from network element datasheets, and from measurements
where applicable, is constructed. The results presented here are indicative of the current
state-of-the-art but may slightly vary depending on the choice of representative equip-
ment and future equipment energy efficiency improvements.
The power consumption of each network element involved in the transport of IoT
traffic is summed in accordance with (6.1). The graph in Figure 6.10 depicts the power
consumption of all access network technology options considered here for IoT gateway
data access rates between 1 kb/s and 1 Mb/s. In all but one of the architectures (i.e.
LPWA), an IDN is assumed to have an IoT gateway which connects to the Internet core
through either a PON, PtP optical, VDSL2, 4G LTE wireless or Wi-Fi network access.
For LPWA, however, the each sensor of the IDN connects directly to the LPWA gate-
way without an IoT gateway. The results indicate that the power consumption of the
fixed access network technologies is largely dominated by the power consumption of
their respective CPE modems. PON access, compared to VDSL2 and PtP, is relatively
more power efficient amongst the wireline technologies, by virtue of its passive remote
node and effective port sharing regime. Given that the IoT bandwidth requirement is
several orders of magnitude lower as compared to the bandwidth capacity of PtP access,
its power consumption variation is negligible. VDSL2, which uses fibre backhaul from
the remote node, appears to be the least power efficient for sub-1Mb/s data access rates.
This is due to the presence of an active power consuming network element at the remote
node added to the power of a VDSL2 modem.
In the case of 4G LTE wireless technology, the model reflects a range of background
traffic levels as shown in Figure 6.6. Due to high volume of traffic during busy-hours
between 4pm and 10pm (140% daily of average load), the power share attributed to 4G
LTE access is low, but increases greatly after midnight between 2am and 4am (20% of
178 Power Consumption of IoT Access Networks
IoT Data Access Rate (kb/s)100 101 102 103
Pow
er C
onsu
mpt
ion
(Wat
ts)
0
5
10
15
20
25
30
35
40
VDSL2
Unshared Wi-FiPtP
Shared Wi-Fi
LTE (100% of average)
LTE (40% of average)
LTE (20% of average)
PON
LPWA
Figure 6.10: Power consumption per IoT gateway for different access network technolo-gies. For LTE, power consumption varies according to the share of BS power consump-tion attributable to the background traffic.
daily average load) as the volume of background traffic drops. For IoT services using
4G LTE access, the power consumption apportioned to IoT traffic for the low, medium
and high background traffic volume is depicted in Figure 6.10. Power consumption of
4G LTE below 10 kb/s is relatively independent of traffic level due to the dominance of
the customer modem and gateway power consumption but increases beyond 10 kb/s.
The relative difference in power consumption between the low background traffic (20%
of daily average load) scenario and that of the high background traffic (100% of daily
average load) can be explained by the cost of sharing the idle power of a macro cell
base station between a small number of users/gateways. In the future, as the move to
5G networks with small cells and Narrow-Band IoT provisioning becomes a reality, a
situation like this will not apply due to low-powered small cells and improved resource
allocation for small packets as an example.
At higher IoT traffic levels and lower 4G LTE utilisation, IoT service via VDSL2 and
6.7 Results and Discussion 179
4G LTE consume similar amounts of power and there is no distinct power advantage
of one over the other. In contrast, at very low data rates, 4G LTE has a distinct power
advantage over any of the wireline technologies.
The difference between unshared Wi-Fi deployment with a single user and shared Wi-
Fi with many users is substantial. Although a shared Wi-Fi deployment may operate at
higher total traffic rates and consume more power, the sharing strategy ensures a very
small per-gateway or per-user power footprint. For higher IoT traffic levels the curve for
shared Wi-Fi in Figure 6.10 shows a slight increment in power as the per-bit power share
increases.
For low IoT traffic levels, LPWA appears to be the most power efficient of the wireless
and wireline access network technologies combined. This arises because it is optimised
for very low traffic levels, i.e. low data rate and infrequent transmissions. LPWA has a
distinct power advantage over shared Wi-Fi but for low data access rates (≤ 10 kb/s).
IoT Data Access Rate (kb/s)100 101 102 103
Ene
rgy
Effi
cien
cy in
ene
rgy-
per-
bit (
J/bi
t)
10-6
10-5
10-4
10-3
10-2
10-1
Unshared Wi-FiShared Wi-FiPONPtPVDSL2LTE (20% of average)LTE (40% of average)LTE (100% of average)LPWA
LTE (100% of average)
Shared Wi-Fi
LTE (40% of average)
LTE (20% of average)LPWA
VDSL2PtP
PON
Unshared Wi-Fi
Figure 6.11: Energy efficiency of different IoT access network technologies
180 Power Consumption of IoT Access Networks
The graph in Figure 6.11 shows the energy-per-bit of the different access technologies
considering sub-1 Mb/s IoT data access rates. It further reinforces the results presented in
Figure 6.10. From Figure 6.11, the energy-per-bit decreases sharply for all access technolo-
gies with increase in data access rate. More generally, it is seen that 4G LTE can provide
an energy efficient access technology for IoT services under conditions of low IoT data-
rate and a well utilised base station. Under these circumstances, 4G LTE will likely be
more energy efficient than most of the wireline access technologies. In particular, VDSL2,
which uses a powered remote node, is very likely to be the least energy efficient access
technology for IoT. LPWA is specifically designed for low power and is the most energy
efficient but only applicable for low data access rates. Shared Wi-Fi has the advantage
of being both shared and using a low power wireless technology for limited coverage,
which makes it the most energy efficient access technology for a wide range of IoT access
rates considered here.
6.8 Potential Power Savings with Sleep-mode
In Figure 6.10, the power consumption of the wireline access network technologies is
seen to be dominated (more than 64% in VDSL2 and PON) by the idle power of the
customer premises equipment (i.e. modem, ONU and IoT gateway). This presents an
opportunity for substantial power savings if automated sleep-mode techniques [187] are
applied. Although many IoT services may require round the clock network access, unlike
some user applications and services which may require access only during certain hours
of the day when users are active (e.g. Email, Facebook, etc...), potential energy savings
could be made by implementing micro sleep modes (i.e. Dozing, Fast Sleep, Deep Sleep),
Wake-on-LAN and other advanced techniques [118].
In calculating the power consumption of an LTE access network, the ’active’ power
state for the 4G LTE modem and the ONU is used in order to present upper-bound power
values as some IoT services may require continuous network access. As can be seen from
Figure 6.4, the 4G modem has three power states: active, idle and standby. The standby
state consumes about 4-5 fold less energy than the active state, essentially demonstrating
6.9 Conclusion 181
a sleep-mode implementation. The observed difference in power seems congruent with
the EUCoC. Hence, if an IoT service with a 10% uplink duty cycle between bursts with
lower bit rates is considered, energy savings of more than 75% could be realised for the
4G LTE modem and more than 35% for the ONU. The LTE power level curves in Figure
6.10 will also change if macro base station hibernation becomes a reality.
6.9 Conclusion
In this chapter, the development of a power consumption model that provides a first-
order estimate of the power consumption per IoT gateway for sub-1 Mb/s data access
rates is given. The results indicate that the power per IoT gateway for wireline access
technologies is dominated by their CPE modems for the access rates under the consider-
ation for IoT services. LTE models might change when small cells are implemented with
the 5G revolution. VDSL2 is seen as the least energy efficient for most of the assessed data
throughput range, while 4G LTE is seen as least energy efficient above several hundred
kilobits/second, depending on the traffic volume and time of day. Among the wireline
access technologies, PON is relatively energy efficient for the full range of IoT access rate
while LPWA, among the wireless access technologies, is more energy efficient but appro-
priate only for a smaller range of access rates. Shared Wi-Fi access with PON backhaul is
overall most energy efficient access technology for medium to higher data access rates,
provided that the shared Wi-Fi also carries a modest level of background traffic. For lower
access rates, there is no clear power advantage of 4G LTE over shared Wi-Fi where both
are available, but a marked difference for higher access rates as the share of total traffic
allocated to the IoT becomes significant in comparison with the cell background traffic.
The dominance of power consumption by the modems provides an opportunity for
significant power savings with the implementation of sleep-mode regimes and duty cycle
optimisation. Whilst the inefficiencies of some network architectures relative to others is
shown here, the choice of an IoT access technology will ultimately depend on the type of
application, the rate of data generation and the cost of deployment. Wireless access has
a significant advantage of being ubiquitous today and easy to augment once deployed;
182 Power Consumption of IoT Access Networks
hence LPWA may be a good choice for IoT applications with low data throughput de-
mands. However, shared Wi-Fi could be a better choice if an IoT application throughput
demand is high.
Chapter 7
Conclusions and Future ResearchDirections
THIS chapter summarizes the key contributions and findings of this thesis on the
energy consumption of Internet of Things (IoT) networks and services. The chapter
further discusses potential future research directions of this work.
7.1 Conclusions and Discussion
The IoT is a new paradigm of interconnectivity that extends the traditional Internet to
non-network-enabled objects or things and to networks (e.g. SCADA) that were not pre-
viously part of the Internet, with a view to create more accessible end-to-end services.
The IoT leverages a number of established and emerging technologies (i.e. WSN, Cloud
and Fog Computing) to deliver more accessible end-to-end services. While there could
be many benefits (e.g. home security, assisted living, etc.) derived from an IoT ecosys-
tem with billions of devices, such deployment elicits potential risks. These include an
increase in device standby energy consumption to maintain connectivity or perceived
”smartness”, additional network energy cost for handling possible IP traffic increment
due to high-traffic IoT applications (e.g. video surveillance), and the potential impact on
the global energy consumption of the ICT industry as a whole.
In this thesis, the energy consumption of IoT applications, networks and services was
investigated, with primary focus on a few well-known IoT use-cases, home automation
and security systems (HAS) and video surveillance services. Measurement of power con-
sumption and data traffic constituted the basis of our modelling approach; hence a num-
183
184 Conclusions and Future Research Directions
ber of IoT devices and gateway devices were measured. Such measurements of power
consumption, or similar data from other sources where necessary, were used to estimate
the overall energy consumption over time of the IoT devices and services described. The
key contributions of this thesis are as follows:
• New measurement-based energy models were introduced for IoT devices includ-
ing sensors, actuators and IoT gateways, and a shared energy model for the home
gateway device was developed (Chapter 3).
• Using developed energy models, the total energy consumption of HAS was inves-
tigated and estimates of its energy impact on the global ICT industry presented
(Chapter 3).
• New energy consumption measurements of the most common wireless network
communication protocols for IoT devices was presented. Using these measure-
ments in conjunction with an example IoT application, the energy implications of a
number of communication paradigms for IoT devices was analysed (Chapter 4).
• A measurement-based power consumption model for new generation web-based
network IP cameras was introduced (Chapter 5).
• The energy consumption of Local, Edge and Cloud-based network architectures for
IoT video surveillance services was investigated and evaluated for different video
streaming application models (Chapter 5).
• The power consumption of several IoT access network technologies was investi-
gated, modelled and evaluated for IoT-like (sub-1 Mb/s) data access rates (Chap-
ter 6).
These contributions can be broadly categorised into three areas: (i) ”Energy Consump-
tion of IoT Applications and Services”, (ii) ”Energy-Efficiency of IoT Network Architec-
tures” and (iii) ”Energy-Efficiency of IoT Access Network Technologies”. These areas are
discussed in the sections below.
7.1 Conclusions and Discussion 185
7.1.1 Energy Consumption of IoT Applications and Services
In Chapter 3, the total energy consumption of HAS or ”smart homes”, which includes
IoT devices, IoT gateways and smart appliances, was investigated. In order to conduct
this study, a representative example HAS (i.e. Ninja Block) was considered for a ”typical”
mid-size home. We carried out power consumption and traffic measurements of a variety
of devices including the following:
Temperature & Humidity Sensor (×2)(i) Passive Infrared Sensor(ii)
Door Sensor(iii) Window Sensor(iv)
Power plug Actuator(v) Logitech USB Webcam (×2)(vi)
Network IP camera(vii) Ninja Block IoT gateway(viii)
Raspberry Pi (IoT gateway)(ix) BeagleBone Black (IoT gateway)(x)