Received January 24, 2017, accepted March 1, 2017, date of publication April 7, 2017, date of current version June

7, 2017.
Digital Object Identifier 10.1109/ACCESS.2017.2688279

Comparison of the Device Lifetime in Wireless

Networks for the Internet of Things
ÉLODIE MORIN1,2,3 , MICKAEL MAMAN2 , ROBERTO GUIZZETTI1 ,
AND ANDRZEJ DUDA3 , (Member, IEEE)
1 STMicroelectronics, 38140 Crolles, France
2 CEA-LETI, MINATEC, 38000 Grenoble, France
3 CNRS Grenoble Informatics Laboratory, Grenoble Institute of Technology, Grenoble Alps University, 38000 Grenoble, France
Corresponding author: Élodie Morin (morine@imag.fr)
This work was supported in part by the French Ministry of Research projects DataTweet under Contract ANR-13-INFR-0008-01 and in
part by the PERSYVAL-Laboratory under Contract ANR-11-LABX-0025-01.

ABSTRACT This paper presents a comparison of the expected lifetime for Internet of Things (IoT) devices
operating in several wireless networks: the IEEE 802.15.4/e, Bluetooth low energy (BLE), the IEEE 802.11
power saving mode, the IEEE 802.11ah, and in new emerging long-range technologies, such as LoRa and
SIGFOX. To compare all technologies on an equal basis, we have developed an analyzer that computes the
energy consumption for a given protocol based on the power required in a given state (Sleep, Idle, Tx, and Rx)
and the duration of each state. We consider the case of an energy constrained node that uploads data to a sink,
analyzing the physical (PHY) layer under medium access control (MAC) constraints, and assuming IPv6
traffic whenever possible. This paper considers the energy spent in retransmissions due to corrupted frames
and collisions as well as the impact of imperfect clocks. The comparison shows that the BLE offers the best
lifetime for all traffic intensities in its capacity range. LoRa achieves long lifetimes behind 802.15.4 and BLE
for ultra low traffic intensity; SIGFOX only matches LoRa for very small data sizes. Moreover, considering
the energy consumption due to retransmissions of lost data packets only decreases the lifetimes without
changing their relative ranking. We believe that these comparisons will give all users of IoT technologies
indications about the technology that best fits their needs from the energy consumption point of view. Our
analyzer will also help IoT network designers to select the right MAC parameters to optimize the energy
consumption for a given application.

INDEX TERMS Internet of Things (IoT), wireless sensor networks, 6LoWPAN, 802.15.4e, TSCH,
802.11ah, Bluetooth low energy, LoRa, SIGFOX, energy consumption model, clock drift.

I. INTRODUCTION energy consumption. However, the work by Tozlu [2] showed

The Internet of Things (IoT) aims at connecting small con- that some 802.11 devices can also benefit from good energy
strained devices to the Internet via the IP protocol, which efficiency in 802.11b Power Saving Mode (PSM) [3] and
enables new communicating applications. Ultra low energy even outperform 802.15.4. Bluetooth Low Energy (BLE) [4]
consumption is critical for IoT devices since they mostly has also added another competitor to the low-power stan-
operate on batteries and they need to reach lifetimes of dards, and besides the short-range solutions, emerging long-
the order of several years without battery replacement. Low range technologies such as 802.11ah [5], LoRa [6], and
power operation becomes even more challenging for nodes SIGFOX [7] also aim at high energy efficiency.
that harvest energy from the environment, for instance using In this paper, we present a comparison of the expected life-
solar panels. They can only consume a small amount of time for IoT devices operating in the wireless networks listed
energy over short time intervals that needs to be compensated in Table 1: beacon-enabled IEEE 802.15.4, the Time Slotted
by energy intake. An estimation of the energy consumption Channel Hopping (TSCH) variant of IEEE 802.15.4e [8],
and the device lifetime is thus crucial for choosing the most Bluetooth Low Energy (BLE), IEEE 802.11 PSM, and future
appropriate technology and finding the optimal settings of IEEE 802.11ah. We also consider the emerging long-range
configuration parameters. technologies (LoRa and SIGFOX) to see how they perform
Until recently, ultra low-power devices mainly relied on the compared to other protocols. We assume IPv6/6LoWPAN
802.15.4 standard [1] that was specifically designed for low protocols whenever possible (see Fig. 1) and carefully

2169-3536
2017 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 5, 2017 Personal use is also permitted, but republication/redistribution requires IEEE permission. 7097
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

TABLE 1. Overview of the analyzed technologies. duty-cycled IoT devices in a realistic way. Our goal was to
extend the recent realistic energy consumption models to all
considered technologies and make them available through an
open source tool allowing anybody to play with the network
parameters and quickly estimate lifetimes.
Finally, we want to evaluate the impact of pushing the IP
protocols to constrained IoT devices by taking into account
not only the hardware performance and duty-cycled MAC
operation, but also the protocol overhead (headers and frag-
mentation), along with the cost of keeping an active connec-
tion and time synchronization between two nodes. In this way,
we can evaluate the impact of all layers on energy consump-
tion and performance. The overall goal of this paper is to help
all users of IoT technologies to choose the technology that
best fits their needs from the perspective of energy efficiency.

B. ASSUMPTIONS, RESEARCH METHOD,

AND CONTRIBUTION
We consider the case of an energy constrained node (battery-
operated or energy harvesting) that generates and uploads
data to a sink (uplink traffic). We adopt the view of an IoT
FIGURE 1. Protocol stack of the considered technologies.
device that connects to the network and wants to stay alive
for the longest time. We limit the scope of the paper to a
single link case, because it is the current most common con-
analyze the energy consumption patterns for duty-cycled figuration setup, a kind of a baseline for comparisons (note
MAC layers to derive the lifetimes. that many papers limit the scope to a single link case, e.g. the
analytical energy consumption model of B-MAC proposed
A. MOTIVATION AND RESEARCH GOALS by Polastre et al. [24]). Moreover, the topology of all the
The motivation for this paper comes from the observation that considered networks except 802.15.4 is a star one in which
many different wireless technologies can be used for IoT and a node connects to or associates with a main-powered node
there is no exhaustive comparison in the literature of their to obtain Internet connectivity. In the future work, we plan
energy consumption that would help to make right decision to extend the analysis to multi-hop 802.15.4 and upcoming
about deployment of a given network. Although some char- mesh BLE networks.
acteristics of the considered technologies, such as the range Our research method consists of analyzing energy
or the used bandwidth, are not directly comparable, there is consumption of each technology by considering a real-
a need for a thorough comparison of the energy efficiency istic energy model in which the main energy con-
to identify their key factors and limitations. Many overviews sumers (radio transceiver, microcontroller) alternate between
of different wireless networks exist [9]–[13], but to the best active and inactive states (Tx, Rx, Idle, Sleep) with
of our knowledge, no previous work compares a large set of known power consumption patterns and timing con-
wireless networks for IoT from the energy efficiency point straints defined by the MAC layers. We derive the model
of view while setting them on an equal basis. Some papers from the previous proposals for 802.15.4/e [16]–[18],
just study two technologies [14], [15], but a comparison with extend them to other MAC technologies, and use it to analyze
other possible solutions is still missing. Such a comparison the technologies in function of different parameters.
with all technologies laid on the same basis can bring to For the presented analysis, we assume the application
light the respective advantages and drawbacks to drive the traffic workload composed of variable UDP (User Datagram
future developments for improving performance and energy Protocol) packet sizes over IPv6/6LoWPAN protocols when-
consumption. ever possible (see Fig. 1) or over a given MAC layer oth-
Another motivation of our work is to extend the energy erwise. The analysis takes into account the energy spent
consumption models that reflect the operation of duty-cycled in data frame retransmissions due to corrupted frames and
MAC layers, which is often neglected in energy consumption collisions. We consider the energy consumption for different
studies. Before recent papers that set the principles of the hardware platform parameters (i.e. energy consumption per
energy model we use [16]–[18], energy consumption models state) to find the parameters resulting in the minimal energy
only took into account the following main aspects: transmis- consumption. We then adopt the minimal energy hardware
sion power, distance between two nodes, packet size, and platforms to compute the lifetime of the considered technolo-
path loss [19]–[23]. This approach only modeled the behavior gies for application traffic with varying intensity, Packet Error
of the physical layer and it did not reflect the operation of Rate (PER), and clock precision. The premise is that even if

7098 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

such an idealized device does not yet exist, it represents the The four-state energy consumption model is not suffi-
best-in-class energy performance enabling the comparison ciently precise for some technologies, so we have added a
with other technologies. 6-state model for 802.15.4e TSCH in Section V to represent
The main new scientific contributions of the paper with the behavior of this technology better. Since some hardware
respect to the state of the art are the following: platforms do not operate according to the 6-state model, we
• we extend a realistic energy model that takes into cannot use it for all technologies. Nevertheless, we include
account the behavior of the PHY and MAC layers to all the results of both models for TSCH in Section V. More
considered technologies, research is further needed to take advantage of switching the
• we develop an open-source lifetime analyzer, a tool for microcontroller off during radio activity without spending to
estimating the lifetime of a given network based on the much energy in the wake-up phase.
main hardware and MAC parameters, The hardware model assumes instantaneous transitions
• we use the analyzer to study the impact of higher layer between states, but in reality, they take some time, which
protocols (fragmentation, protocol overhead), time syn- results in additional energy consumption. Finally, we use
chronization, and packet losses on the lifetime (main a linear battery current draw model, but the real battery
conclusions are summarized below), discharge process is much more complex. Nevertheless, the
• we reveal several new results: i) the power consumed linear model is not that far from complex recursive models
in the Sleep state becomes a determining factor of the (less than 10%) [25] and we compensate this difference by
device lifetime for low traffic intensities, ii) 802.15.4 taking into account parameterable current leakage.
consumes less energy than 802.11 PSM for low intensity
traffic, which restrains the findings by Tozlu [2], and D. STRENGTHS OF THE STUDY
iii) asynchronous technologies such as LoRa and SIG- The analyzer presented in this paper gives us a means for
FOX achieve long lifetimes when we assume realistic quick evaluation of the energy consumption in the main
clocks with drift. IoT wireless networks. It realistically models the behavior
• we discuss the suitability of the technologies for repre- of principal protocol layers and takes into account several
sentative applications corresponding to chosen use cases factors important for evaluating energy consumption: proto-
and study the feasibility of energy harvesting solutions. col overhead including fragmentation, power drained in the
Section VI provides the main conclusions of the sleep state, packet losses, time synchronization, and varying
comparison. application traffic.
Our objective of developing an analytical tool obviously
C. LIMITS OF THE STUDY leads to some simplifications, however, the comparisons
The comparison considers networks with different ranges with other studies show sufficient accuracy of its results.
varying from several meters (BLE) to several kilometers Considering all energy consumption factors would require a
(LoRa, SIGFOX). Moreover, they offer different transmis- full-fledged simulator.
sion robustness depending on the Modulation and Coding
Scheme (MCS) inherent to each technology and chosen trans- E. ORGANIZATION OF THE PAPER
mission power. We are aware that energy consumption is To provide an in-depth understanding of energy consumption
only one aspect to take into account when choosing the right and keep the paper self-contained, we start with a com-
technology for a given application use case. Other parameters prehensive description of MAC operation for each consid-
such as throughput, delay, reliability, coverage range, radiated ered technology in Section II. We then present the proposed
power, and cost can also be considered. energy consumption model and its validation in Section III.
For instance, it does not make sense to propose the use Section IV evaluates the impact of hardware parameters
of short-range BLE instead of long-range SIGFOX when an on the lifetime and provides the choice of the parameters
application requires large coverage. Similarly, there is always for the comparisons. Section V presents the results of our
a trade-off between the transmission power, the range, and the analyzer for different application scenarios and identifies
resulting bit error rate. However, to select the best technology the main factors that influence energy consumption. Finally,
given application requirements (performance, energy con- Section VI concludes the paper.
sumption) in a specific communication context (ranges, con-
ditions), it is necessary to understand the energy limitations II. BACKGROUND ON WIRELESS NETWORKS FOR IOT
of each technology and identify the factors that influence the While this section presents each of the IoT technologies
node lifetime for a given performance level. We provide some in detail, we need to discuss several specific assumptions
information on these aspects in Table 1. adopted in our analysis. First, we assume that the radio and
Our analysis also makes some simplifications that influ- the microcontroller are the most energy-consuming parts of
ence the precision of the energy consumption predictions: each device. Moreover, we consider the model of energy
a simple probabilistic model of packet loss, constant bit rate consumption in which the total energy consumption is com-
type of the application workload, and single link topology puted as the power required for each state over the time spent
setup. in the state. Our model assumes four different states for an

VOLUME 5, 2017 7099

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

TABLE 2. Hardware consumption states. according to the beacon-enabled mode can achieve a low
energy consumption because a node can safely turn its radio
off during the rest of the superframe and wake up at the next
beacon. Periodic beacons allow keeping time synchronization
between nodes.
Nodes in beacon-enabled mode use a slotted Carrier Sense
IoT platform: Rx, Tx, Idle, and Sleep illustrated in Table 2 Multiple Access with Collision Avoidance (CSMA/CA)
with the modes of each component (we give the details of the access scheme. In our analysis, we neglect the backoff mech-
power drawn in each state later on). anism to simplify the model. In this mode, transmissions
In addition, we consider a minimal packet structure for follow the superframe organization shown in Fig. 3 that
analysis, i.e. we do not consider the security overhead within presents: i) a timeline of a node transmitting data to its
data, beacon, and ack frames. coordinator (two acknowledged data frames), ii) the node
We summarize the analyzed IoT standards and technolo- state corresponding to different phases of the superframe. The
gies below and present their energy consumption patterns as superframe duration (SD) is defined as follows [1]: SD =
a function of the MAC operation (state, duration). 15360 ∗ 2SuperframeOrder (SO) µs, with SOmax = 14. Hence,
the maximum beacon interval is:
A. BEACON-ENABLED 802.15.4
Beacon Intervalmax = 251.658240 s.
802.15.4 [1] is the most popular standard for low-power
wireless networks, largely used in industrial applications and
B. 802.15.4e/TSCH MODE
subject to continuous development [18], [26], [27]. Its typical
TSCH is part of the 802.15.4-2012 amendment [8] that
targets are low-power embedded communication systems that
addresses the need for deterministic access in industrial appli-
require a low data rate and low latency.
cations: channel hopping helps the network to mitigate mul-
The 802.15.4 PHY layer uses the 2.4 GHz and sub-1 GHz
tipath fading and a periodic schedule using dedicated slots,
bands. In the 2.4 GHz band, 802.15.4 splits the spectrum in
defined by a channel and timeslot, avoids contention. TSCH
sixteen 5 MHz-wide channels with a 5 MHz inter-channel
mainly aims industrial networks that require an increased
space. 802.15.4 supports robust transmissions with DSSS
level of robustness, reliability, availability, and security.
(Direct-Sequence Spread Spectrum) and the O-QPSK (Offset
Quadrature Phase Shift Keying) modulation scheme giving
a bit rate of 250 kb/s. Multi-hop topologies can extend its
relatively short ranges (see Table 1).

FIGURE 2. Data frame structure of 802.15.4 (size in bits).

FIGURE 4. TSCH principles. (a) Example TSCH schedule. (b) Operation

states in a TSCH timeslot.

Fig. 4a shows an example of a TSCH schedule represented

FIGURE 3. Operation of beacon-enabled 802.15.4. by a slotframe that defines a couple (slot, channel) allocated
to a pair of nodes. The standard does not define how to build
Even if we consider a single link between a leaf node a schedule and some proposals for scheduling have already
and a sink, the beacon-enabled mode is interesting in multi- appeared [28]–[30]. The TSCH data frame is the same as the
hop topologies for saving energy at intermediate and leaf classical 802.15.4 one (see Fig. 2), except that the Sequence
nodes. In this mode, a coordinator node periodically sends Number field is omitted.
a beacon to delimit its superframes and invites neighboring Once a node has joined the network, there is no need
associated nodes to send their data frames (see Fig. 2) during of beacon frames to keep nodes synchronized (except for
the Contention Access Period (see Fig. 3). Nodes operating Extended Beacons to advertise the presence of the network).

7100 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

Whenever the transmitter wants to send a packet, it turns

its radio on at constant tsTxOffset after the beginning of
a timeslot (see Fig. 4b). On frame reception, the receiver
computes the difference between the frame arrival time and
the expected tsTxOffset to adjust its clock with respect to
the transmitter. In addition, an acknowledgment frame (ACK)
may also contain a timestamp to adjust the clock. However,
if there is no data traffic, nodes need to use specific keep-
alive messages for synchronization. The maximum interval
between keep-alive messages (empty data frame) is set so that
the clock drift never exceeds the guard time defined as: FIGURE 5. BLE data frame structure (size in bits, if not explicited).

guard time = tsTxOffset − tsRxOffset. (1)

Since we assume a symmetrical clock drift at the receiver and
the transmitter:
guard time
keep_aliveperiod ≤ (2)
2 ∗ clockAccuracy
Given a clockAccuracy of 40ppm and the default values of the
standard [8] (Table 52e), we used guard time = 1 ms, which
leads to keep_aliveperiod ≈ 12.5 s.

C. BLUETOOTH LOW ENERGY (BLE) FIGURE 6. BLE operation.

BLE is the low-energy version of the Bluetooth standard [4].

BLE is mostly used for monitoring applications, such as heart
rate monitors or remotely controlling the temperature of a mode, a joining node sends ADV frames on one or several
room. An important advantage of the BLE technology is of the three advertising channels scanned by a master node.
its popularity—BLE is already deployed in the majority of After its discovery by the master node, the slave node wakes
smartphones and it is largely used for mobile applications. up at the beginning of each connInterval and waits for a
It defines a master/slave relationship between two nodes in a unicast-poll frame from its master before sending its data
star topology. BLE does not yet support the mesh topology, packet (see Fig. 5) as shown in Fig. 6.
but a scatternet topology is already available in version 5.0 The slave and the master achieve time synchronization
(a collection of trees) since a node can be a slave and a through polling packets sent with a maximal period of 8.0 s by
master at the same time in different connections. The 6lo the master. If there is no data traffic during a long period, the
IETF working group defines the adaptation layer to transmit slave needs to wake up at least every 4 connIntervals, which
IPv6 packets over BLE [31]. corresponds to a period of 32 s, to listen to a poll packet from
BLE nodes split the 2.4 GHz ISM frequency band into its master and reply with an ACK.
37 dedicated channels, plus 3 advertising channels. The width
of all channels is 2 MHz with inter-channel spacing of 2 MHz.
BLE uses the GFSK (Gaussian Frequency-Shift Keying)
modulation to obtain a maximal 2 Mb/s bit rate and takes
advantage of frequency hopping spread spectrum for robust FIGURE 7. 802.11b PSM data frame structure (size in bits).

transmissions. The recent version 5.0 allows fitting different

ranges and robustness requirements with four coding schemes D. 802.11 PSM
leading to several PHY rates: 125 kb/s, 500 kb/s, 1 Mb/s, 802.11 PSM is an energy-optimized mode specified in the
and 2 Mb/s. In this analysis, we focus on the highest available 802.11b version. It targets energy efficiency by enabling
data rate that results in a minimal energy consumption for active/sleep duty cycles for associated nodes. In 802.11b,
similar overhead. BLE nodes can change frequency either nodes can transmit on one of the thirteen 22 MHz wide
periodically or at some logical time depending on the fre- channels separated by 5 MHz. The 802.11b revision speci-
quency hopping mode. fies Complementary Code Keying (CCK) of 8 bits/symbols,
BLE offers two different communication modes: asyn- which results in a bit rate of 11 Mb/s. 802.11 PSM supports a
chronous transmissions that involve advertisement pack- star topology with the Basic Service Set (BSS) and an Access
ets (ADV) and synchronous transmissions in connected Point (AP). Fig. 7 shows the data frame structure.
mode. Asynchronous transmissions only support sending a Fig. 8 presents the principles of the Distributed Coordi-
small number of bytes, so we focus on the connected mode nation Function (DCF) used in 802.11b PSM when a node
in which slave nodes discover and connect to a master. In this sends two data frames in a row to an always-on Access Point.

VOLUME 5, 2017 7101

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

announced range is around 40 km. SIGFOX devices can

transmit up to 140 messages per day to the base station
with a maximum user payload of 12 B, so we assume
that applications encapsulate their data directly in MAC
frames. Moreover, European Telecommunications Standards
Institute (ETSI) regulation imposes a capacity limitation:
SIGFOX operates in the sub-1 GHz band, which leads
to a maximum duty-cycle of 0.1% or 1% in 863 −
FIGURE 8. Operation in 802.11 PSM mode. 870 MHz (depending on the selected sub-band). Note that
140 messages per day corresponds to the limitation of 1%
DCF requires nodes to stay in Rx mode for a random period duty-cycle with a 100 b/s SIGFOX implementation.
of time before sending data, competing with other nodes to
access the channel. In our setup however, there is no con-
tention. The usual value of the beacon interval is 100 ms in
classical 802.11 networks. However, the listen interval field
of the beacon encodes the number of time units on 2 B, which FIGURE 10. SIGFOX Uplink packet structure (size in bits).

leads to a maximal value of ∼ 18.6 hours.

At the MAC layer, a device that wants to transmit data,
E. 802.11ah or Wi-Fi HaLoW encapsulates it within a packet (see Fig. 10) and trans-
The last version of 802.11ah draft standard [5] specifi- mits three times on different random frequencies with dif-
cally addresses main IoT requirements: an increased range, ferent channel encoding at each transmission. The network
increased reliability, and low energy consumption. It can be can send a downlink packet of 8 B, but we do not take
seen as an optimized 802.11ac PHY layer [32] on sub-1 GHz into account downlink transmissions in our analysis. While
frequency bands in the PSM mode. The choice of the sub- SIGFOX available data rates are low, the long range of the
1 GHz band comes from the objective of limiting interference technology and ease of deployment are important key features
in the crowded 2.4 GHz band. Moreover, the European limi- for many IoT applications.
tation of the duty cycle in this sub-band is not applicable for SIGFOX is an asynchronous technology (as is LoRa)—
802.11ah since the transmission scheme is based on Listen nodes do not need to wake up at a specific instance for
Before Talk and Adaptive Frequency Agility. As in 802.11ac, synchronization. Hence, the energy consumption model of
there are 10 available MCS operating in bands of different SIGFOX is simpler than for other technologies: nodes are
widths. We consider two extremes: MCS giving the minimum in Tx state while transmitting data and in Sleep state
and maximum data rates. The minimum data rate is given between transmissions. Note that SIGFOX defines an inter-
with MCS10 and 1 MHz band (denoted as 10_1), using a transmission time during which the node is in Idle mode.
BPSK modulation, with a 2*code repetition. The maximum As this time is negligible compared to the duration of the data
data rates are given by MCS8 in Europe, 2 MHz band (8_2) frame, we do not take it into account.
and MCS9 in the US, 16 MHz band (9_16), using a 256-QAM
G. LoRa
modulation.
For long-range IoT applications, the LoRa Alliance proposes
a cellular topology with base stations/gateways that receive
packets from devices and relay the data to a server on a
TCP connection. Actility (a LoRa partner) and other partners
enabled 6LoWPAN on top of LoRa, but for a fair comparison
FIGURE 9. 802.11ah data frame structure (size in bits).
with SIGFOX, we have assumed the transmission of applica-
tion data directly over the MAC layer.
802.11ah uses DCF as in 802.11 PSM (see Fig. 8). The LoRa operates in the same band as SIGFOX and 802.11ah,
Restricted Access Window (RAW) access scheme is also pos- and uses a proprietary modulation based on the Chirp Spread
sible [33], but as DCF and RAW obtain similar performance Spectrum (CSS) that trades data rate for sensitivity within a
in a small network [34], we only consider DCF. Fig. 9 shows fixed frequency band. The LoRa modulation offers data rates
the structure of a data frame. from 0.25 to 11 kb/s in the European frequency band. 250 b/s
corresponds to a spreading factor of 12 for a bandwidth of
F. SIGFOX 125 kHz, whereas 11 kb/s results from a spreading factor of
SIGFOX is an ultra-narrow-band technology that operates 7 and 250 kHz of bandwidth. Devices can also use FSK mod-
in the 868 MHz frequency band in Europe and 915 MHz in ulation to reach a higher data rate of 50 kb/s (not considered
the US. The available PHY rates are 100 b/s for a 100 Hz in the analysis).
bandwidth and 1000 b/s for a 1 kHz channel width in Europe In our analysis, we assume a Class A LoRa device that
(and 600 b/s in the US). With sensitivity of −140 dBm, its transmits frames as soon as the data is available (no Listen

7102 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

Before Talk): LoRa devices need to comply with the ETSI

regulations of the 868 ISM band. The energy consumption
model of LoRa includes the Idle state between transmissions
and ACK, and the Rx state during ACK.

FIGURE 11. LoRa packet structure (size in bits).

FIGURE 12. Variables used to compute energy consumption and device

Fig. 11 shows the LoRa data packet structure. The maxi- lifetimes. (a) Without synchronization scheme. (b) With synchronization
scheme.
mum user data size depends on the data rate. For the minimal
and the maximal data rate in Europe, the maximum data
sizes are respectively 59 and 250 B. As the LoRa preamble B. APPLICATION MODEL AND LIFETIME COMPUTATION
is composed of eight symbols, a node may spend long time Fig. 12 illustrates the definition of the main variables we
in Rx (resp. Tx) mode depending on the chosen modulation. use to compute the energy consumption while analyzing
It is however required to accurately synchronize the nodes to the operation of a given MAC protocol achieving the max-
achieve the LoRa coding gain necessary to demodulate the imum throughput. If a node has to fragment a packet, it will
LoRa signal under the noise floor. send fragments in consecutive frames. The synchronization
scheme corresponds to the communication for maintaining
III. PRINCIPLES OF ANALYZING ENERGY CONSUMPTION synchronization between two nodes based for instance on
We present below the fundamentals of our analyzer that beacon or poll frames.
computes energy consumption and the lifetime of the con- We define the following variables:
sidered wireless technologies. Compared to a simulator, an • ta , Application Period is the time interval between two
analytical approach presents the advantage of a shorter exe- instants of data generation by an application.
cution time and a lower development effort while providing • sa , Application Data Size is the size of the data gen-
sufficient precision. Note that chip manufacturers adopt a erated by the application each ta . sa is the useful data
similar approach, for instance Linear Technology proposes at the application layer. We then add the overhead of
an analyzer for its products [35]. Our analyzer is developed each protocol and take into account fragmentation if
in Python and available in the public domain.1 necessary.
• ra , Application Throughput corresponds to the generated
s
A. ENERGY CONSUMPTION MODEL application data rate in bits/s (b/s): ra = t a .
We use a model derived from the work by Vila- a
• tsyn , Maximal Synchronization Period corresponds to the
josana et al. [16] to express the energy consumption in interval after which a node needs to wake up for time
interval t as follows: synchronization. The period is required to keep an active
X association between devices. If there is no communica-
E(t) = PS × tS , S ∈ {Tx, Rx, Idle, Sleep} (3)
tion during an interval longer than tsyn , nodes lose their
S
association with the network and will need to rejoin,
where PS is the power consumption in state S and tS is the potentially at the cost of greater energy consumption.
time spent in state S during t. tS comes from the analysis • tCI , Check Interval is the time between two wake-ups
of the operation of each wireless technology presented in of a node defined by tCI = min(ta , tsyn ). A data trans-
section II. We assume a star topology with one link between mission may serve for synchronization, but if there is
an energy-constrained leaf node for which we want to find no application data to send, nodes need to communicate
the lifetime, and a main-powered sink. The analysis considers each tsyn (see right part of Fig. 12).
upward application traffic (such as periodic reporting) to the • ton , Active Period is the maximum time a node stays
sink. To take into account IP connectivity, we consider an awake, i.e. when it is in a state different from Sleep
adaptation layer (6LoWPAN or 6Lo) if needed, and IPv6 run- during tCI (see Figs. 3, 6, 8, and 4b).
ning on top of a given PHY/MAC layer for each technology • Lt , Lifetime of a node corresponding to the time a node
whenever possible (see Fig. 1). However, because of small can run on some initial energy E0 .
packet sizes, the long-range technologies support application To estimate device lifetimes for each protocol, we select the
data directly on top of the MAC layer. MAC parameters of each protocol to support ra , the applica-
Moreover, the analysis assumes the stationary state, i.e. the tion throughput with the minimal energy consumption, so we
leaf node is already associated with the sink. set ton to the minimal duration to transmit sa bytes of data.
For example, we choose the 802.15.4 SD parameter to reach
1 https://gitlab.imag.fr/morine/iot-analyzer the optimal ton period, given the value of sa per ta interval.

VOLUME 5, 2017 7103

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

If a technology cannot support given application through- probability PER, a node performs the following number of
put ra , it does not appear on the curves. The representation of transmissions on the average:
a technology stops when the application traffic corresponds 1
to the maximal capacity of the given protocol. Ntr = . (5)
1 − PER
In the comparisons, we will choose a given level of PER and
compute the energy consumption required to send Ntr data
frames followed by an ACK for one packet generated by the
application layer.

FIGURE 14. Algorithm of the ComputeEnergyToSendData function.

Fig. 14 gives the details of the ComputeEnergyToSendData

function in which we take into account the energy con-
sumption due to data retransmissions within each application
period ta .
FIGURE 13. Lifetime computation algorithm.
D. BATTERY MODEL
Based on the variables and the operation principles of each We assume an initial battery energy of E0 = 13.5 kJ corre-
MAC layer (explained in Section III-A), we compute the sponding to two AAA batteries (1250 mAh under 1.5 V each)
energy consumption over one ta with Eq. 3, in which PS , and a simple battery model with leakage Eleak of 5% per year,
the power consumption in state S, is multiplied by tS , the as well as a cutoff voltage when the residual energy reaches
time spent in state S during one ta . Our lifetime algorithm 10% of E0 . For fair comparisons, we assume the same battery
(see Fig. 13) computes the energy consumed per ta as a for all technologies even though a smaller battery can be used
function of application throughput ra and derives Lt , the for some protocols.
lifetime of a node. Lt is the number of ta a node can run until
the initial energy E0 is exhausted. IV. IMPACT OF HARDWARE PLATFORMS AND
CHOICE OF THEIR PARAMETERS
C. PACKET LOSS MODEL To compare the technologies on the same fair basis, we need
For technologies that use retransmissions, we take into to choose the hardware parameters for the power consump-
account lost data packets due to imperfect transmission con- tion values. The idea is to identify the parameters that result
f
ditions or collisions in the following way. Let per and prer in the smallest energy consumption for each technology.
be the packet error rate in the forward, respectively reverse,
direction: a data packet sent from a node will be lost with A. TECHNOLOGY-DEPENDENT PARAMETERS
f
probability per while its acknowledgment will suffer from the We start with the main parameters of the standards, presented
loss probability of prer . The probability of a lost transmission in Table 3. We also assume 802.15.4 CCADuration of 128 µs
at the MAC layer is thus: and use the default TSCH interval values (see TSCH-MAC
PIB attributes [8]):
PER = pfer + (1 − pfer ) × prer . (4)
• timeslotLength = 10 ms,
f
If prer
is small compared to (small ACK packets), the
per • tsTxOffset = 2.120 ms,
f
probability reduces to PER = per . Assuming packet loss • and tsRxAckDelay = 800 µs.

7104 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

FIGURE 15. Impact of hardware platforms on the lifetime, E0 = 13.5 kJ. (a) Varying sa , constant ta = 100 s. (b) Varying sa , constant ta = 1 s. The legend
applies to both (a) and (b).

TABLE 3. Technology parameters. We can notice that the lifetimes are not ranked by the
values of PTx and PRx , e.g. most of the 802.11 PSM plat-
forms perform better than 802.15.4 TelosB even though the
TelosB Tx and Rx power consumption is lower than that of
802.11 PSM. Actually, the power consumption in these states
is a decisive factor for high data rates (small ta ), but when the
data rate is lower (e.g. ta = 100 s), the Tx and Rx power
consumption becomes less important. In this case, contrary
to the common belief, the technologies are ranked by their
Psleep power, since nodes spend much more time in the sleep
TABLE 4. Hardware platform parameters. state. Hence, hardware designers should also concentrate on
this value in addition to minimizing the Tx and Rx power
consumption.

TABLE 5. Hardware parameters used for comparisons.

B. HARDWARE PLATFORM PARAMETERS

We consider several hardware platforms summarized in For a fair comparison, we have examined a range of hard-
Table 4 and evaluate the impact of their power consumption ware platform solutions and used the best-in-class param-
parameters of Table 2 on the lifetime. By a hardware platform eter value for each technology (denoted by the Selected
we mean a SoC with a microcontroller and a radio module. Min-energy in Tables 4, 5, and Fig. 15). In doing so, we have
Fig. 15 compares the lifetimes for the initial energy effectively modeled a composite platform that offers ideal-
E0 = 13.5 kJ, constant ta , and varying sa from 1 B to 10 kB. ized performance and results in the best trade-off between
This figure clearly shows an important impact of the hardware their Pidle , PTx , and PRx parameters. Here our assumption
parameters: for instance, the lifetime of 802.11 PSM varies is that while such an idealized device may not yet be on
from ∼ 1.5 to ∼ 4 years for different 802.11 PSM platforms, the market, it is feasible that a single platform offering such
when ta = 1 s (see Fig. 15b) and ra = 10 B/s. The effect best-in-class performance exists. Tables 5 and 6 summarize
is even more important for ta = 100 s (see Fig. 15a). the chosen hardware parameters that we use to generate the
Large differences for 802.15.4 and BLE platforms are also remaining results of the paper.
notable. Moreover, we have chosen to use the same parameters for
BLE, 802.15.4, and TSCH because their hardware design
2 assuming GreenNet STM32 CPU. is similar (same transmission power and modulation com-

VOLUME 5, 2017 7105

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

plexity). Hence, we adopt the lowest value of GreenNet period ta = 4 min. In an equivalent setup, our analyzer com-
for PIdle , the SmartMeshIP values of PRx and PIdle , along putes an average current consumption of 2.46 µA. The 18%
with the lowest Psleep value of BLE112. Choosing a com- difference is mainly due to a simplification inherent in our
mon platform presents the advantage of a fair comparison model that does not take into account the energy consumption
of the protocols instead of comparing hardware platforms. of state transitions, nor that of the temperature sensor used in
However, we cannot extend this approach to every the GreenNet measurements.
technology—we keep a large difference in Rx and Tx power In addition, we have compared the 100 B data frame perfor-
consumption between the 802.11 standard and the other mance for BLE and 802.15.4 with the performance reported
technologies. The difference in the output power (0 dBm by Siekkinen et al. [15]. We modify BLE parameters to fit
compared to 14 and 18 dBm) is necessary to reach 802.11 BLE4.0 specification [48], to compare equivalent scenarios.
ranges and therefore justifies a higher Tx power consump- Our results for BLE gives 460 KB/J, while 433 KB/J was
tion. Different modulation schemes explain the difference in reported by Siekkinen et al. Similarly, we find 180 KB/J
the Rx power consumption—demodulating a more complex for the 802.15.4 standard, while Siekkinen et al. reported
scheme costs more energy to achieve an acceptable Bit Error 168 KB/J. In both cases, the difference is less than 7%.
Rate (BER). To validate the analyzer for long range technologies,
we have compared its output with the values measured by
TABLE 6. SIGFOX and LoRa parameters. Martinez et. al [17] for transmission of a single 12-byte
data packet (i.e. the maximum size for the SIGFOX technol-
ogy). In the previous section, we have assumed data packets
without security for all technologies, but to set up a fair
comparison, we have added an additional overhead of 2 B to
the minimal SIFGOX packet presented in Fig. 10, because
SIGFOX imposes the use of HMACs for message authentica-
tion at the MAC layer. Martinez et al. measured a current load
of 50 mA for three consecutive transmissions of ∼2 s each.
Assuming a 3 V voltage, this value corresponds to 900 mJ.
Our analyzer computes a transmission time of 6.24 s with an
energy consumption of 917.28 mJ (less than a 5% difference).
Finally, we have compared our results with LoRa measure-
For long-range technologies, we adopt the parameters ments [17]: for transmission of a 15 B data packet and assum-
in Table 6. ing a maximum LoRa spread factor, the measured energy
consumption is ∼213 mJ. In a similar setup, we compute a
C. VALIDATION OF THE ANALYZER consumed energy of 224 mJ, which corresponds to a differ-
To validate our analyzer, we have compared its results with ence of ∼5%.
the SmartMeshIP power estimator [35] of TSCH provided by These comparisons show that our analyzer yields data
Linear Technology, using the lowest power mode. We have within 5%-7% of most published measured values and
configured a star topology with one node sending 10 bytes to broadly confirms the consistency of our approach.
its parent at each application period ta = 100 s (called report-
ing interval) with 100% path stability. The estimator assumes V. COMPARISONS OF DEVICE LIFETIMES
a keep-alive period of tsyn ∼ 4.083 s (∼ 25 keep-alive frames Having confirmed the validity of our analyzer, we now
during 100 s), but it does not exploit the possibility of time present the predicted device lifetimes for the following three
synchronization during data exchange. In this setup, the total cases:
average mote current estimated for a downlink slotframe of
1024 slots is 4.3 µA for a voltage of 3.6 V. In a comparable 1) assess the impact of application parameters without any
setup, our analyzer gives the average current consumption packet loss or clock drift.
of 4.317 µA, less than 1% error. 2) evaluate the impact of a limited packet loss probability
We have also compared the average energy consumption without clock drift.
of 802.15.4 with the measured value on the GreenNet plat- 3) assess the impact of imperfect clocks with clock drift
form [18]: the average current consumption was 3 µA for (but no packet loss).
BO = 4 (Beacon Interval of 240ms) with the possibility to
skip beacons by the leaf node with a maximum time of tsyn , A. IMPACT OF APPLICATION PARAMETERS,
SO = 1 (Superframe Duration of 15.36 ms), and application NO PACKET LOSS
3 Based on TD1202 [46] with GreenNet microcontroller and BLE112
Fig. 16 presents the comparison of the lifetime for fixed
power in sleep state.
values of application period ta = 1 day, 100 s, 1 s, and
4 Based on SX1272 [47] with GreenNet microcontroller and BLE112 10 ms, and for varying data size sa , which results in varying
power in sleep state. throughput ra . As there is no packet loss, PER = 0.

7106 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

FIGURE 16. Different ta and data size sa in bytes, leading to varying ra , impact lifetime for a starting energy E0 = 13.5 kJ, PER = 0. The bottom x-axis is
the data size in bytes per ta while the top x-axis presents the corresponding data rate ra in b/s. No packet loss or clock drift is assumed. (a) legend.
(b) varying sa , constant ta = 1 day. (c) varying sa , constant ta = 100 s. (d) varying sa , constant ta = 1 s. (e) varying sa , constant ta = 10 ms.

With the separation of application period ta and data size

sa , we can distinguish between the energy consumption of
an application that generates sa = 10 B every second and an
application sending sa = 600 B every minute, even though
for both cases ra = 80 b/s. In the first case, there are 60 trans-
missions with the corresponding protocol overhead, whereas
in the second case, the application only sends the minimum
number of frames, depending on the chosen technology maxi-
FIGURE 17. 6-State energy consumption model of TSCH.
mum data packet size. This effect explains, for instance, why
the lifetime for a given ra (e.g. 80 b/s) is higher in Fig. 16c
than in Fig. 16d.
The figures include the results for TSCH (marked as 868-868.6 MHz) when no Listen Before Talk and Frequency
6-State Model TSCH with dashed lines) based on a more Agility mechanisms are used. Hence, in our study, we dis-
sophisticated model of energy consumption with two addi- play the result for 0.1% as a solid line and the result
tional states available on recent 802.15.4 platforms that for 1% as a dotted line for the concerned technologies
can turn their microcontroller off when transmitting [16] (SIGFOX and LoRA). The duty cycle limit of 1% determines
(see Fig. 17). the maximal capacity of SIGFOX at 1 kb/s and 100 b/s. More-
For the sub-1GHz frequency band, the ETSI regula- over, the marketing limitation of SIGFOX to 140 packets per
tion recommends a certain duty cycle depending on the day further limits the capacity of SIGFOX at 1 kb/s and it
used sub-band (e.g. 0.1% for 863-870 MHz and 1% for corresponds to the 1% duty cycle at 100 b/s.

VOLUME 5, 2017 7107

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

1) VERY LOW TRAFFIC INTENSITY (ta = 1 DAY) this result with the lower energy consumption of our 802.15.4
Fig. 16b shows the lifetime comparison for very low traffic platform compared to TelosB used in the paper by Tozlu—as
intensity ta = 1 day corresponding to the operating condi- explained earlier, we have chosen the most efficient platform
tions of long-range technologies. We observe that BLE and to compute the lifetime of 802.15.4 (see Table 4). The lifetime
802.15.4 achieve the best lifetimes. LoRa performs remark- of 802.11 PSM becomes better than 802.15.4 only for very
ably well with an only slightly lower lifetime compared to the long packets (greater than 1.5 kB), not a usual size for IoT
short-range BLE and 802.15.4 standards. SIGFOX achieves applications.
similar lifetimes to LoRa only for extremely low ra , since the
available bit rate of SIGFOX is 11 times smaller than that of 3) MEDIUM AND HIGH TRAFFIC INTENSITY (ta = 1 s, 10 ms)
LoRa. For medium and high intensity traffic (see Fig. 16d and 16e,
SIGFOX and LoRa consume more energy as the data note also the logarithmic y-axis), BLE obtains the longest
size sa increases. Their lifetime drops because they can only lifetime although 802.15.4 is traditionally expected to be
handle small data packet sizes—fragmentation often occurs, better.
which increases the overhead, and consequently the energy We can identify mainly two reasons for this result:
consumption. Moreover, to reach the targeted long ranges, • The 802.15.4 PHY layer is less efficient than BLE due
they use high transmission power, which contributes to the to a spreading factor of 8 that reduces the bit rate of
increased energy consumption when they need to transmit for 802.15.4 from 2 Mb/s to 250 kb/s, whereas BLE uses
longer periods. the maximum rate of 2 Mb/s. Hence, the bit rate of BLE
As seen in Fig. 16b and 16c, SIGFOX and LoRa suffer from is 8 times higher than 802.15.4, which leads to a lower
capacity limitations due to the ETSI regulation in the sub- energy consumption.
1GHz band. In addition, ETSI regulation requires that nodes • Fragmentation in 802.15.4 starts from 120 B of useful
do not transmit more than a certain number of consecutive information, whereas BLE starts fragmenting at 245 B.
frames per hour. Thus, SIGFOX or LoRa nodes that must Since the overhead for a data frame does not depend on
transmit for more than 3.6 s (for 0.1% duty cycle) will have to the packet size, 802.15.4 is less efficient than BLE for
delay some packets to the next hour, which impacts latency. packets bigger than 120 B.
A SIGFOX node requires three transmissions to send one To validate this explanation, we have changed the bit
data frame, before it is acknowledged by the base station. rate of 802.15.4 to 2 Mb/s and set the data frame size to
In this case, the limit of messages per hour leads to an 245 B. In this setup, 802.15.4 comes close to BLE (as shown
incapacity to respect the duty cycle limitation in the 863 − in Fig. 19).
870 MHz with the rate of 100 b/s. Hence, we advise at least Moreover, Figure 19 displays the variation in terms of the
the use of the 1% tolerable sub-band to compete with other BLE energy efficiency performance when selecting different
IoT technologies. modes of Bluetooth Specification 5.0 [4]: the highest avail-
802.11ah performs significantly worse than LoRa for small able data rate leads to a lower energy consumption because
data sizes and outperforms it for larger data sizes. 802.11ah it results in the minimal time in radio consuming modes for
also outperforms 802.11 PSM and obtains similar perfor- similar overhead (see BLE data packets in Fig. 5).
mance to TSCH. Note that all 802.11ah curves are overlaid Our results confirm the findings by Siekkinen et al. [15].
in Fig. 16b: we have used the same Psleep value for the three They showed that, in the stationary phase, 802.15.4 transmits
802.11ah technologies (see Table 5). As Psleep value is the less data than BLE for the same amount of energy. Siekkinen
determining factor for a synchronized technology for very et al. also obtained their results on a hardware platform equiv-
low data rates, we understand the overlay of the curves. alent to BLE and 802.15.4. BLE is thus the best technology
in terms of the energy consumption for medium and high
data rates. Nevertheless, it suffers from capacity limitation
2) LOW TRAFFIC INTENSITY (ta = 100 s) for larger data sizes. Note that the case presented in Fig. 16e
For sporadic application traffic, BLE achieves the longest corresponds for some technologies to always-on nodes, when
lifetime with the 802.15.4 performance not too far behind operating at their maximum capacity (represented by the end
for small data sizes (see Fig. 16c). For larger data sizes in of a curve).
Fig. 16c, the US variant of 802.11ah with 16 MHz of band- For sa > 100 B, 802.11 PSM outperforms 802.15.4
width becomes an interesting solution after BLE. In these (see Fig. 16d) and becomes the second most energy-efficient
conditions, the important overhead of 802.11ah packets is technology up to sa > 500 B, for which US-only 802.11ah
mitigated by a larger maximum data size. For this reason, (MCS9, 16 MHz) with high modulation rates and large band-
the fragmentation happens less often than for BLE, which width becomes the most energy-efficient technology.
decreases the ratio overhead/data for 802.11ah. We note that the performance of 802.11ah with a smaller
Contrary to the findings by Tozlu regarding 802.15.4 and channel bandwidth, even with high modulation indices such
802.11 PSM [2], [49], Fig. 16c shows that 802.15.4 consumes as the maximum European rate (MCS 8, 2 MHz, 7.8 Mb/s),
less energy in most cases of low data traffic (up to throughput is worse than or equivalent to that of 802.11 PSM due to the
of 80 b/s) and obtains much longer lifetimes. We can explain difference in the data rate.

7108 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

FIGURE 18. Lifetime for varying ra , No clock drift, E0 = 13.5 kJ, Average Packet Error Rate (PER) = 0.2. (a) legend. (b) varying sa , constant ta = 1 day.
(c) varying sa , constant ta = 100 s. (d) varying sa , constant ta = 1 s. (e) varying sa , constant ta = 10 ms.

of 20% packet loss to represent a significant impact of chan-

nel conditions and interference on transmission quality. It is
also usually adopted as a threshold below which a wireless
link is considered of low quality and not used for transmis-
sions. We can only note a slight variation in the absolute
lifetime values: the relative positions of the curves do not
change. For instance, the BLE lifetime for ra = 800 b/s with
ta = 100 s drops from ∼ 15 years to ∼ 13.5 years, but still
stands as the less consuming technology.
FIGURE 19. Liftime for different BLE 5.0 modes and comparison with
theoretical 2Mb/s 802.15.4 (varying sa , constant ta = 1 s). C. IMPACT OF THE CLOCK DRIFT
To evaluate the impact of handling the clock drift between
We can also observe that the lifetime of 802.11ah MCS10,
nodes, we assume that nodes have imperfect clocks C(t)
1 MHz, 0.15 Mb/s is consistently worse than the other tech-
with a given bounded drift 1 such that | dC(t) dt − 1| ≤ 1.
nologies. To obtain a longer lifetime comparable to BLE
A typical value of 1 is 40 ppm for crystal clocks. As both
or 802.15.4, we will need to lower the default transmission
clocks at the transmitter and the receiver may diverge from the
power of this variant of 802.11ah to 0 dBm, which will
perfect clock, the maximum difference between their drifts
decrease the range, the key objective of 802.11ah.
is bounded by 21. To compensate the clock drift during
Check Interval tCI , we need to add a guard interval Tg such
B. ROBUSTNESS TO POOR CHANNEL CONDITIONS
that:
Fig. 18 presents the impact of the packet loss probability
PER = 0.2 on the lifetime. We have chosen the value Tg = 21tCI . (6)

VOLUME 5, 2017 7109

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

TABLE 7. Lifetimes of representative applications, 40 ppm Clock drift, PER = 0.2, and maximum European available rates.

For instance, if tCI = 100 s and 1 = 40 ppm, the guard specifies a default guard interval of 1 ms to compensate for an
interval is 8 ms. imperfect clock.
For ultra low traffic intensity, LoRa and SIGFOX achieve
the best lifetimes, the same ones as previously, thanks to
their asynchronous operation: they do not need to wake up
in advance to compensate for the clock drift.
These results call for an adaptive synchronization scheme.
Instead of consuming too much energy in guard time and
synchronization maintenance, it may be beneficial for a node
to lose association, and then reconnect when needed for a
data exchange. For example, to maximize efficiency over the
long term, a BLE node would disconnect from the network
during sleep mode, then reconnect during wake-up. There is
an additional overhead for reconnection, but it may be smaller
than the overhead due to clock synchronization. Nevertheless,
we note that some mechanisms of clock drift compensation
already exist to cope with the drift issue [50].
FIGURE 20. Timing relationships.
D. LIFETIMES OF REPRESENTATIVE APPLICATIONS
Our study aims at helping engineers and IoT application
Fig. 20 presents timing relationships between a transmitter
developers to select the best technology depending on given
and a receiver that need to be awake for a data transmission.
application requirements. Table 7 summarizes the lifetimes
The first case assumes perfect clocks so Tx and Rx happen
one can expect in a given representative application assuming
at the same time, which is the baseline case we used to
the optimal operating conditions of an IoT device with two
generate Fig. 16. Three other cases illustrate the operation
AAA batteries. The analyzer computes the lifetimes for the
with a guard interval Tg tolerating a maximal relative drift
corresponding application data rates (sa /ta ) presented in the
of 21 (see Eq. 6) and the situation with different values of
first column of Table 7.
relative drift (−21, 0, and +21). Note that the case of −21
corresponds, from the energy consumption point of view, to
E. ENERGY HARVESTING IoT DEVICES
the same situation as the case without the guard interval.
Hence, Fig. 16 also gives us the values for −21. In this section, we address the issue of IoT devices that
To show the impact of the guard interval and the clock drift harvest energy from the environment and store energy in a
on the energy consumption, we have chosen to generate the small capacity battery (e.g. 20 mAh). Such devices do not
results of Fig. 16 for the drift of +21 (with 1 = 40 ppm) have a fixed initial amount of energy that determines their
for which the receiving node needs to stay awake during lifetime, but instead, they harvest energy intermittently and
the longest interval: Fig. 21 shows the corresponding results. then consume it when operating.
We can see that for all synchronous technologies, the time
wasted during the guard interval Tg in Rx state leads to con- TABLE 8. Harvested power.
siderable shorter lifetimes, more than 20% shorter for ta =
1 day and BLE (approximately 23.5 years compared to 30).
Note also that for higher traffic intensity, the influence of
these factors is smaller, so Fig. 21c is almost the same as
Fig. 16d. In this case, small ta mitigates the wasted time of
Tg , since Tg is proportional to tCI that depends on ta .
The impact of the clock drift on the TSCH lifetime We take the example of a GreenNet node with a solar panel
is smaller than for other technologies: less than 10% for of '18 cm2 [18] with the indoor and outdoor values of the
ta = 1 day with the enhanced state model since TSCH already harvested power presented in Table 8.

7110 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

FIGURE 21. Lifetime for different ra , Clock drift of +21 = 80 ppm, E0 = 13.5 kJ, and PER = 0. (a) Legend. (b) ta = 1 day, 21 = 80 ppm.
(c) ta = 1 s, 21 = 80 ppm.

horizontal harvested power line are theoretically feasible, i.e.,

a device will have the sufficient power during ta to operate.
Our results show that only BLE and 802.15.4 can operate
with a typical indoor light intensity of 300 lx. The light inten-
sity of 8000 lx is sufficient for all technologies, but devices
need to be located 5 cm at most from the light source. Finally,
we can see that an outdoor, full-sun scenario can support all
technologies, but applications can be vulnerable to variations
in the levels of outdoor light.

VI. CONCLUSION AND FUTURE WORK

Our comparison reveals several results not previously brought
up in the literature:
1) Assuming ideal clocks, we found that:
• BLE offers the best lifetime for all traffic intensi-
ties in its capacity range.
• For ultra low traffic intensity, LoRa performs
remarkably well with an only slightly lower
lifetime compared to the short-range BLE and
802.15.4 standards. 802.11ah performs signifi-
cantly worse than LoRa for small data sizes and
outperforms it for larger data sizes.
• For low and medium traffic intensity, the rela-
tive ranking is: BLE, 802.15.4, TSCH, and 802.11
technologies with the latter becoming interesting
solutions for larger data sizes.
• For high data traffic intensity, 802.11ah with a
larger bandwidth and 802.11PSM are the less
FIGURE 22. Average consumed power per ta , PER = 0.2. (a) Legend. consuming technologies when BLE reaches its
(b) ta = 1 s, 21 = 80 ppm. (c) ta = 100 s, 21 = 80 ppm. capacity.
2) When we assume a guard time to compensate the clock
Fig. 22 presents the average consumed power per ta for all drift in synchronous technologies, the asynchronous
technologies. To the right, we mark the level of the intake LoRa and SIGFOX technologies obtain the best life-
power for a given light intensity. So, all curves below a given time for low intensity traffic.

VOLUME 5, 2017 7111

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

3) Contrary to the common belief that power consumption The main goal of the analysis was to evaluate energy
in Sleep state is negligible, Psleep becomes a deter- consumption for a given available throughput while tak-
mining factor of the device lifetime for low traffic ing into account the most important parameters and factors.
intensities while the consumption in Tx and Rx states Nevertheless, other aspects may impact the choice of a given
still remains important for higher traffic intensity. technology such as the range, total network capacity, spectral
4) Taking into account the energy spent in data frame efficiency, scalability, and latency. In the future, the energy
retransmissions due to corrupted frames and colli- consumption analysis can be extended along the following
sions does not change the relative ranking of the research directions:
technologies.
• Compare long range technologies with multi-hop net-
5) The current long-range technologies operating in the
works that may cover the same distances. The goal
sub-1 GHz frequency band with important duty cycle
is to take the approach adopted by Lampin [52] to
limitations, which reduces their capacity, are not ready
compare star networks with those supporting multi-hop
to support energy harvesting yet.
operation.
We have also shown that 802.15.4 consumes less energy
• Enhance the energy consumption model with additional
than 802.11 PSM for low intensity traffic, which restrains the
states and non-instantaneous transitions.
findings by Tozlu [2]: the lifetime of 802.11 PSM is only
• Integrate other long range technologies such as
better than 802.15.4 for higher traffic intensity and longer
802.15.4g [53], RPMA (Random Phase Multiple
packets.
Access) by Ingenu [54], or NarrowBand IoT [55].
With respect to the existing literature, our paper provides
• Optimize all aspects of energy consumption and perfor-
a unique contribution in several aspects. First, our paper
mance for a given application scenario under constraints,
goes beyond several existing comparisons of IoT technolo-
for instance a given limited frequency band.
gies and PHY/MAC layers for sensor networks [9]–[13]
that described the features of different solutions and their Beyond the analysis presented in this paper, the com-
limitations without proposing a comparison on an equal parison has also raised our interest in interoperability and
basis. Some papers limited their analysis to two technolo- mixing different technologies. We have shown that the cost
gies: 802.15.4 vs. 802.11ah [14] and BLE vs. 802.15.4 [15] of keeping an active connection, i.e. the synchronization cost,
by using different energy consumption metrics: energy to is sometimes too high. Hence, an interesting research issue
transmit a packet (mJ/packet) vs. energy utility (kB/J), which is to improve the device lifetime by proposing an adaptive
makes transitive comparisons difficult. scheme for time synchronization: BLE can transmit data in
Second, with respect to energy consumption models, much advertising packets, so it may be beneficial to disconnect
research focused on physical layer aspects without taking from a master and operate in non-connected mode for sparse
into account MAC operation [19]–[23]. Polastre et al. [25] traffic applications. This approach can also lower energy
initiated a realistic model of Mica2 under B-MAC further consumption in other technologies.
enhanced with the precise representations of energy con- Another issue is to explore how the key features of the
sumption for 802.15.4 [17], [18], TSCH [16], [17], and considered technologies can cooperate in the IoT context to
SIGFOX [17]. We have extended this approach to deal with fit the needs of different application traffic requirements such
all the considered technologies. as monitoring and delay-sensitive traffic, while obtaining suf-
Moreover, we can also note that the models of the physical ficiently long lifetimes. We plan to work on interoperability
layer operation [19]–[23] neglected the impact of the micro- between synchronous technologies such as 802.15.4e and
controller on energy consumption. This assumption is often asynchronous ones such as BLE in data-advertising mode or
justified by the low value of microcontroller power consump- non beacon-enabled 802.15.4.
tion compared to the radio. However, even if the value is
low, the accumulated energy consumption over long periods REFERENCES
makes it significant. Our results show that the value of Psleep ,
[1] Standard for Local and Metropolitan Area Networks—Part 15.4: Low-Rate
the microcontroller power consumption in sleep mode, can Wireless Personal Area Networks (LR-WPANs), IEEE Standard 802.15.4-
be a determining factor depending on the application traffic, 2003, 2003.
so Psleep must be taken into account in a precise analysis of [2] S. Tozlu, ‘‘Feasibility of Wi-Fi enabled sensors for Internet of Things,’’
in Proc. 7th Int. Wireless Commun. Mobile Comput. Conf. (IWCMC),
energy consumption. Jul. 2011, pp. 291–296.
Finally, our analyzer represents another progress beyond [3] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
the current state of the art. It implements the energy con- Specifications, IEEE Standard 802.11, 1999.
[4] Bluetooth SIG. [Online]. Available: https://www.bluetooth.com/
sumption model for all technologies and takes into account
specifications/adopted-specifications
several aspects missing in the previous studies: the over- [5] Draft Standard for Information Technologies Telecommunications and
all overhead of IP connectivity, clock drift compensation Information Exchange Between Systems Local and Metropolitan Area
mechanisms, and packet losses. It provides a means for Networks Specific Requirements—Part 11: Wireless LAN Medium
Access Control (MAC) and Physical Layer (PHY) Specifications—
quick estimation of the lifetime based on the most important Amendment 6: Sub 1 GHz License, IEEE Standard P802.11ah/D10.0,
parameters. Sep. 2016.

7112 VOLUME 5, 2017

É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

[6] LoRa Alliance, accessed on Jan. 10, 2017. [Online]. Available: http://lora- [29] A. Morell, X. Vilajosana, J. L. Vicario, and T. Watteyne, ‘‘Label switching
alliance.org/ over IEEE802.15.4e networks,’’ Trans. Emerg. Telecommun. Technol.,
[7] SIGFOX, accessed on Jan. 10, 2017. [Online]. Available: vol. 24, no. 5, pp. 458–475, Aug. 2013.
http://www.sigfox.com/en/ [30] N. Accettura, M. R. Palattella, G. Boggia, L. A. Grieco, and M. Dohler,
[8] IEEE Standard for Local and Metropolitan Area Networks—Part 15.4: ‘‘DeTAS: A decentralized traffic aware scheduling technique enabling IoT-
Low-Rate Wireless Personal Area Networks (LR-WPANs)—Amendment compliant multi-hop low-power and lossy networks,’’ in Proc. IEEE 2nd
1: MAC sublayer, IEEE Standard 802.15.4e-2012 (Amendment to IEEE WoWMoM Workshop Internet Things, Smart Objects Services (IoT-SoS),
Standard 802.15.4-2011), 2012. vol. 6. Jun. 2013, pp. 1–6.
[9] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash, [31] 6lo Status Pages, accessed on Jan. 10, 2017. [Online]. Available:
‘‘Internet of Things: A survey on enabling technologies, protocols, and https://tools.ietf.org/wg/6lo/
applications,’’ IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp. 2347–2376, [32] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
4th Quart., 2015. Specifications—Amendment 4: Enhancements for Very High Throughput
[10] T. Kim, I. H. Kim, Y. Sun, and Z. Jin, ‘‘Physical layer and medium access for Operation in Bands below 6 GHz, IEEE Standard 802.11, 2014.
control design in energy efficient sensor networks: An overview,’’ IEEE [33] E. Khorov, A. Lyakhov, A. Krotov, and A. Guschin, ‘‘A Survey on IEEE
Trans. Ind. Informat., vol. 11, no. 1, pp. 2–15, Feb. 2015. 802.11ah: An Enabling Networking Technology for Smart Cities,’’ Com-
[11] A. Bachir, M. Dohler, T. Watteyne, and K. K. Leung, ‘‘MAC essentials for put. Commun., vol. 58, pp. 53–69, Mar. 2015.
wireless sensor networks,’’ IEEE Commun. Surveys Tuts., vol. 12, no. 2, [34] O. Raeesi, J. Pirskanen, A. Hazmi, J. Talvitie, and M. Valkama, ‘‘Per-
pp. 222–248, 2nd Quart., 2010. formance enhancement and evaluation of IEEE 802.11ah multi-access
[12] C. Goursaud and J.-M. Gorce, ‘‘Dedicated networks for IoT : PHY/MAC point network using restricted access window mechanism,’’ in Proc.
state of the art and challenges,’’ EAI Endorsed Trans. Internet Things, IEEE Int. Conf. Distrib. Comput. Sensor Syst. (DCOSS), May 2014,
Oct. 2015. pp. 287–293.
[13] Q. Wang and J. Jiang, ‘‘Comparative examination on architecture and [35] Linear. Dust Networks SmartMesh Power and Performance
protocol of industrial wireless sensor network standards,’’ IEEE Commun. Estimator, accessed on Jan. 10, 2017. [Online]. Available:
Surveys Tuts., vol. 18, no. 3, pp. 2197–2219, 3rd Quart., 2016. http://www.linear.com/docs/42452
[14] B. B. Olyaei, J. Pirskanen, O. Raeesi, A. Hazmi, and M. Valkama, ‘‘Per- [36] RTX. RTX4100 Datasheet DS1, accessed on Jan. 10, 2017.
formance comparison between slotted IEEE 802.15.4 and IEEE 802.1 lah [Online]. Available: http://tinyurl.com/ooekop6
in IoT based applications,’’ in Proc. 9th IEEE Int. Conf. Wireless Mobile [37] Maxim Integrated. MAX2830 2.4GHz to 2.5GHz 802.11g/b RF Transceiver
Comput., Netw. Commun. (WiMob), Oct. 2013, pp. 332–337. with PA and Rx/Tx/Diversity Switch, accessed on Jan. 10, 2017.
[15] M. Siekkinen, M. Hiienkari, J. K. Nurminen, and J. Nieminen, ‘‘How [Online]. Available: http://tinyurl.com/q3xyynj
low energy is Bluetooth low energy? Comparative measurements with [38] STMicroelectronics. SPWF01SA Datasheet, accessed on Oct. 16, 2015.
ZigBee/802.15.4,’’ in Proc. WCNCW, Apr. 2012, pp. 232–237. [Online]. Available: http://tinyurl.com/or94b9l
[16] X. Vilajosana, Q. Wang, F. Chraim, T. Watteyne, T. Chang, and [39] Nordic Semiconductor. nRF51822 Product Specification v3.1,
K. S. Pister, ‘‘A realistic energy consumption model for TSCH networks,’’ accessed on Jan. 10, 2017. [Online]. Available: http://tinyurl.com/npuckqy
IEEE Sensors J., vol. 14, no. 2, pp. 482–489, Feb. 2014. [40] BlueGiga. BLE112 Datasheet, accessed on Jan. 10, 2017.
[17] B. Martinez, M. Monton, I. Vilajosana, and J. D. Prades, ‘‘The power of [Online]. Available: http://tinyurl.com/o8p8spj
models: Modeling power consumption for IoT devices,’’ IEEE Sensors J., [41] STMicroelectronics. BlueNRG Datasheet, accessed on Jan. 10, 2017.
vol. 15, no. 10, pp. 5777–5789, Oct. 2015. [Online]. Available: http://tinyurl.com/pyl5ms3
[18] L.-O. Varga et al., ‘‘GreenNet: An energy-harvesting IP-enabled wireless [42] P. Urard et al., ‘‘A self-powered IPv6 bidirectional wireless sensor & actu-
sensor network,’’ IEEE Internet Things J., vol. 2, no. 5, pp. 412–426, ator network for indoor conditions,’’ in Proc. IEEE Symp. VLSI Circuits,
Oct. 2015. Jun. 2015, pp. C100–C101.
[19] Q. Wang, M. Hempstead, and W. Yang, ‘‘A realistic power consumption [43] T. Watteyne, L. Doherty, J. Simon, and K. Pister, ‘‘Technical overview of
model for wireless sensor network devices,’’ in Proc. 3rd Annu. IEEE smartmesh IP,’’ in Proc. 7th IEEE Int. Conf. Innov. Mobile Internet Services
Commun. Soc. Conf. Sensor, Mesh Ad Hoc Commun. Netw. (SECON), Ubiquitous Comput. (IMIS), Jul. 2013, pp. 547–551.
Reston, VA, USA, Sep. 2006, pp. 286–295.
[44] Linear. SmartMeshIP Datasheet, accessed on Jan. 10, 2017. [Online].
[20] J. Li and P. Mohapatra, ‘‘Analytical modeling and mitigation techniques for Available: http://tinyurl.com/gqyeu5v
the energy hole problem in sensor networks,’’ Pervasive Mobile Comput. J.,
[45] J. Polastre, R. Szewczyk, and D. Culler, ‘‘Telos: Enabling ultra-low
vol. 3, no. 3, pp. 233–254, Jun. 2007.
power wireless research,’’ in Proc. 4th IEEE Int. Symp. Inf. Pro-
[21] Q. Wang and W. Yang, ‘‘Energy consumption model for power manage- cess Sensors Netw. (IPSN), 2005, pp. 364–369. [Online]. Available:
ment in wireless sensor networks,’’ in Proc. 4th Annu. IEEE Commun. Soc. http://tinyurl.com/hh6g6al
Conf. Sensor, Mesh Ad Hoc Commun. Netw., Jun. 2007, pp. 142–151.
[46] Telecom-Design. TD1202 Datasheet, accessed on Jan. 10, 2017. [Online].
[22] H.-Y. Zhou, D.-Y. Luo, Y. Gao, and D.-C. Zuo, ‘‘Modeling of node energy Available: http://tinyurl.com/oclnxyo
consumption for wireless sensor networks,’’ Wireless Sensor Netw., vol. 3,
[47] Semtech. SX1272 LoRa Datasheet, accessed on Jan. 10, 2017. [Online].
no. 1, pp. 18–23, Jan. 2011.
Available: http://tinyurl.com/qbbzn3d
[23] B. Kan, L. Cai, L. Zhao, and Y. Xu, ‘‘Energy efficient design of WSN
[48] Specification of Bluetooth Core v4.2, Dec. 2014.
based on an accurate power consumption model,’’ in Proc. IEEE Int. Conf.
Wireless Commun., Netw. Mobile Comput., Sep. 2007, pp. 2751–2754. [49] S. Tozlu, M. Senel, W. Mao, and A. Keshavarzian, ‘‘Wi-Fi enabled sensors
for Internet of Things: A practical approach,’’ IEEE Commun. Mag.,
[24] J. Polastre, J. L. Hill, and D. E. Culler, ‘‘Versatile low power media access
vol. 50, no. 6, pp. 134–143, Jun. 2012.
for wireless sensor networks,’’ in Proc. 2nd Int. Conf. Embedded Netw.
Sensor Syst., Baltimore, MD, USA, Nov. 2004, pp. 95–107. [50] M. Brzozowski, H. Salomon, and P. Langendoerfer, ‘‘On efficient clock
[25] J. Rahme, N. Fourty, K. A. Agha, and A. van den Bossche, ‘‘A recursive drift prediction means and their applicability to IEEE 802.15.4,’’ in
battery model for nodes lifetime estimation in wireless sensor networks,’’ Proc. IEEE/IFIP 8th Int. Conf. Embedded Ubiquitous Comput. (EUC),
in Proc. IEEE WCNC, Sydney, NSW, Australia, Apr. 2010, pp. 1–6. Dec. 2010, pp. 216–223.
[26] (2015). Introducing Thread: A New Wireless Networking Protocol [51] V. Raghunathan, A. Kansal, J. Hsu, J. Friedman, and M. Srivastava,
for the Home, accessed on Jan. 10, 2017. [Online]. Available: ‘‘Design considerations for solar energy harvesting wireless embedded
http://threadgroup.org/2014_07_Press_Release.aspx systems,’’ in Proc. IEEE 4th Int. Symp. Inf. Process. Sensors Netw.,
[27] OpenThread, an Open-Source Implementation of the Thread Net- Apr. 2005, p. 64.
working Protocol, accessed on Jan. 10, 2017. [Online]. Available: [52] Q. Lampin, ‘‘Urban wireless sensor networks: Applications, characteri-
https://github.com/openthread/openthread zation, and protocols,’’ (in French), Ph.D. dissertation, INSA De Lyon,
[28] M. R. Palattella, N. Accettura, M. Dohler, L. A. Grieco, and G. Boggia, Villeurbanne, France, 2014.
‘‘Traffic aware scheduling algorithm for reliable low-power multi-hop [53] K.-H. Chang and B. Mason, ‘‘The IEEE 802.15.4g standard for smart
IEEE 802.15.4e networks,’’ in Proc. IEEE 23rd Int. Symp. Pers. Indoor metering utility networks,’’ in Proc. IEEE 3rd Int. Conf. Smart Grid
Mobile Radio Commun. (PIMRC), Sep. 2012, pp. 327–332. Commun. (SmartGridComm), Nov. 2012, pp. 476–480.

VOLUME 5, 2017 7113

 É. Morin et al.: Comparison of the Device Lifetime in Wireless Networks for IoT

[54] Ingenu Website, accessed on Jan. 10, 2017. [Online]. Available: ROBERTO GUIZZETTI received the degree in
https://www.ingenu.com computer science from the University of Milano,
[55] ‘‘3GPP low power wide area technologies,’’ GSMA (GSM Italy, in 1988. He has been with STMicroelectron-
Association), White Paper, 2016. [Online]. Available: ics since 1988, in various design centers of the
http://www.gsma.com/connectedliving/wp-content/uploads/2016/10/ company in Italy, France, and Canada. He was
3GPP-Low-Power-Wide-Area-Technologies-GSMA-White-Paper.pdf involved in the field of digital designs and archi-
tectures for various application domains, audio,
video, telecommunication, with particular empha-
sis on low power.
ÉLODIE MORIN is currently pursuing the His main interests in the last few years have
Ph.D. degree with Grenoble Alps University, been in the Internet of Thing area, particularly in hardware and firmware
France. She is also a Research Engineer with for networks of autonomous wireless sensors and actuators.
STMicroelectronics and CEA-LETI, specializing
in interoperability for low layers of the Internet of
Things.
She received the Engineering Diploma degree in
telecommunications from INSA de Lyon in 2014.
Her research interests include low power Inter-
net of Things, and routing and wireless sensor
networks. ANDRZEJ DUDA received the Ph.D. degree from
the Université de Paris-Sud in 1984 and the Habil-
itation Diploma degree from Grenoble Univer-
sity in 1994. He was an Assistant Professor with
MICKAEL MAMAN received the Engineering the Université de Paris-Sud from 1983 to 1986,
degree in electronics, telecommunications and a Chargé de Recherche with CNRS from 1986
informatics from the Ecole Superieur de Chimie to 1995, and a Visiting Scientist with the MIT
Physique Electronique, Lyon, France, in 2006, and Laboratory for Computer Science from 1992 to
the M.Sc. degree in the field of telecommuni- 1994. From 2002 to 2003, he was an Invited Pro-
cations from the Institut National des Sciences fessor with the Swiss Federal Institute of Tech-
Appliquées, Lyon. He has been a Research Engi- nology, Lausanne. He is currently a Professor with the Grenoble Institute
neer with CEA-LETI, the Laboratory of Elec- of Technology–École Nationale Supérieure d’Informatique et de Mathéma-
tronic and Information Technologies of the French tiques Appliquées de Grenoble. He is also a Head of the Networks and
Atomic Energy and Alternative Energy Commis- Multimedia Group, Grenoble Informatics Laboratory. He has authored over
sion, since 2006. His research activity is concerned with MAC and network 150 papers in the areas of performance evaluation, distributed systems,
protocols for Internet of Things, wireless ad hoc, and sensor and body area multimedia, and networks.
networks.

7114 VOLUME 5, 2017