Ad Hoc Networks 143 (2023) 103121

Contents lists available at ScienceDirect

Ad Hoc Networks
journal homepage: www.elsevier.com/locate/adhoc

Probabilistic and distributed traffic control in LPWANs

Kawtar Lasri a,b ,∗, Yann Ben Maissa a , Loubna Echabbi a , Oana Iova b , Fabrice Valois b
a
National Institute of Posts and Telecommunications, INPT, STRS Laboratory, Rabat, Morocco
b
Univ Lyon, INSA Lyon, Inria, CITI, F-69621 Villeurbanne, France

ARTICLE INFO ABSTRACT

Keywords: Low-power, low data transmission rates, and long-range wireless networks, also known as LPWANs, are
LPWAN intended to work best with equipment that uses few resources and can be used for many years thanks to their
Traffic control long battery life operation. This type of networks can handle traffic from nearly 1,000 nodes while maintaining
Probabilistic transmission
a duty cycle of less than 1%. However, as the nodes become denser, the number of collisions increases and
Distributed approach
network traffic management becomes mandatory. To address this concern, we propose a Distributed and
LoRaWAN
Probabilistic Traffic Control algorithm (DiPTC) that allows nodes to change their traffic in response to the
application requirements (e.g., acquiring K measurements over a period of time) while being agnostic to the
number of nodes or the network topology. When this requirement is not achieved, the gateway sends a feedback
message to all the nodes so that they may adapt their traffic. We compare the proposed solution to LoRaWAN
and to a Centralized Optimal Traffic Control solution (COTraC), in simulation. Compared to LoRaWAN, our
algorithm proved successful in achieving the objective while minimizing collisions and extending the network
lifetime threefold.

1. Introduction a client may transmit more data than allowed and establish connec-
tions with additional end devices, resulting in network congestion.
One of the primary applications for deploying sensors in smart We believe that these situations will become increasingly common in
buildings and smart cities is data collection. Low Power Wide Area Net- LPWANs. While the goal of LoRaWAN Adaptive Data Rate (ADR) [7–9]
works (LPWANs), such as LoRaWAN [1] or Weightless [2], are quickly is to minimize energy consumption and adapt transmission data rates in
becoming a popular choice because they provide an infrastructure for response to radio link budget and environmental factors, this algorithm
data collection at a low cost while still covering large area [3–5]. is not capable of correctly addressing network congestion, nor impact
Although they are appealing, these networks have some major traffic across multiple nodes.
drawbacks, such as the used frequencies (e.g., 868 MHz in Europe, To solve this problem, we presented in [10] a new Distributed Prob-
915 MHz in USA) and medium access protocols (e.g., Aloha) [6]. abilistic Traffic Control algorithm (DiPTC) for LPWANs that enables the
For large-scale deployments that need to be able to support a lot of network manager and the applications to better control data collection.
traffic, controlling the traffic sent is necessary. This is especially true
In DiPTC, when the application requires a specific amount of data (e.g.,
for two different types of scenarios. The first one is in applications
K measurements over a period of time), a central server operates a con-
that aim to obtain a sampling of a certain situation in an area (e.g.,
trol loop to impose a traffic policy that complies with the requirements.
environmental monitoring): only a specific amount of measurements
The central server uses a downlink message to inform the nodes if the
per time unit is required. Obtaining more data has no effect and would
target was achieved or not. After receiving this control message, nodes
simply result in network congestion, collisions, and energy waste. On
use a local algorithm to modify their traffic intensity without having
the other hand, receiving less information results in an inefficient
application. The second scenario is when an LPWAN is operated by any neighborhood knowledge. We present here a major improvement
a telecommunication company providing coverage and connectivity to of DiPTC that considerably decreases the energy consumption of end
their customers. A Service-Level Agreement (SLA) specifies the max- devices. We also study the impact of the traffic control parameters on
imum capacity provided by the company and establishes the client DiPTC performance, and we make an extensive performance evaluation
traffic model. Unfortunately, LPWANs lack an algorithm for managing by comparing DiPTC against LoRaWAN and a Centralized Optimal
user traffic and the number of nodes linked to the network. As a result, Traffic Control solution (COTraC).

∗ Corresponding author at: National Institute of Posts and Telecommunications, INPT, STRS Laboratory, Rabat, Morocco.
E-mail address: [email protected] (K. Lasri).

https://doi.org/10.1016/j.adhoc.2023.103121
Received 7 October 2022; Received in revised form 25 January 2023; Accepted 12 February 2023
Available online 15 February 2023
1570-8705/© 2023 Elsevier B.V. All rights reserved.
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

The remainder of the paper is structured as follows: we discuss to automatically manage transmission times in order to reduce the
the state of the art for traffic control in wireless sensor networks and probability of packet collision while improving communication quality
LPWANs in Section Section 2. We introduce and explain our proposed in event-driven traffic. Q-learning is used to determine the transmis-
solution in Section Section 3. Section Section 4 describes the simulation sion time offset and the probability. Assuming an ideal downlink,
setup and scenarios used for the performance evaluation of DiPTC. In the transmission probability is automatically controlled based on the
sections Section 5, Section 6, and Section 7 we validate and compare downlink acknowledgment. They use transmission probability to limit
our results to those of the baseline LoRaWAN and the COTraC solutions. the number of nodes that transmit event packets, thereby reducing
We discuss our findings in Section Section 8 and conclude our work in the probability of packet collision. However, none of these schemes
Section Section 9. actually reduces the amount of traffic in the network and cannot be
used to control the generated traffic.
2. Related work (4) Traffic control. While not present in current IoT approaches
as deemed unnecessary, traffic control algorithms are heavily used
While traffic control can have several benefits in a network (e.g.,
in the Internet, where networks have to handle huge data flows.
reducing the number of collisions, decreasing energy consumption, and
Transport layer protocols such as TCP [25] and QUIC [26] use feedback
limiting congestion), there are several methods through which this can
control algorithms to adapt the sending rates to the network states
be achieved, as we can see next.
and limit congestion in the network. A popular one is the Additive-
(1) Data aggregation. Many researchers in wireless sensor networks Increase/Multiplicative-Decrease (AIMD) [27] algorithm that combines
propose spatial and temporal data aggregation as a solution for traffic linear growth of the transmission traffic when there is no congestion,
and thus collision reduction. Spatial data aggregation is based on with an exponential reduction when congestion is detected. Related
network node organization and is accomplished by selecting a leader or schemes exist, such as Multiplicative-Increase/Multiplicative-Decrease
a group of leader devices that are accountable for transmitting the gath- (MIMD) and Additive-Increase/Additive-Decrease (AIAD). However,
ered data to the gateway [11,12]. To collect the data, communication they do not reach stability. To the best of our knowledge, no such
between the messenger(s) and the nodes is essential. [13–15] describe traffic control algorithm exists today for LPWANs, and this is where
temporal data aggregation based on data prediction, with the authors our inspiration comes from.
focusing on temporal aggregation functions using ARIMA, ARMA, or
LMS-PCS prediction models. When a threshold error occurs, the nodes
3. Towards a distributed and probabilistic approach
modify their prediction model and send their new model coefficients to
the gateway. The gateway uses these coefficients to predict data.
To solve the stated problem using data aggregation methods, nodes In low power wide area networks, such as LoRaWAN, there is no
must have a large memory capacity, neighborhood knowledge, and the traffic control for the end-nodes, except the limitations due to the
ability to communicate with one another. Unfortunately, LPWANs can- duty cycle. No algorithms are provided to allow a network provider
not fulfill these criteria. As a result, temporal aggregation techniques to control the uplink traffic. Nevertheless, it is useful from a point of
cannot be applied to our scenarios: limiting traffic to collect only 𝐾 view of (1) an application that requires collecting a certain number
samples per time period 𝛥𝑇 . of samples periodically or (2) an IoT telecommunication operator re-
(2) Resource allocation. In LoRaWAN, researchers try to reduce the quiring a maximum sample reception rate through several service level
number of collisions and improve scalability by optimizing resource al- agreements. To that end, we propose a Distributed and Probabilistic
location, such as spreading factor and transmission power [8,9,16,17]. algorithm for the Traffic Control of end nodes (DiPTC).
Ta et al. present LoRa-MAB, a flexible decentralized learning resource
allocation approach based on the Multi-Armed-Bandit reinforcement
algorithm [9]. Their approach outperforms the traditional Adaptive 3.1. DiPTC: the big picture
Data Rate (ADR) algorithm used in LoRaWAN. Another important re-
source that can be optimized is the frequency channel, as transmissions The goal is to meet the application requirement, i.e., to collect K
in different channels do not interfere with each other. Shen et al. packets per time period 𝛥𝑇 over a given area, regardless of the number
propose to optimize the channel assignment and the backoff time of end-nodes, without adding excessive control which overloads the
between two consecutive transmissions [18]. Chinchilla et al. propose network, and without adding technical features at the end-node level.
a combination of channel allocation and spreading factors [19], and In LPWANs, the data transmission is only limited by the duty cycle.
Qin et al. combine channel allocation with the optimization of the In DiPTC, we propose to limit the data transmission in order to reach
transmission power [20]. However, these approaches only adapt the the application requirement, while respecting the duty cycle limitation.
physical layer parameters of the existing traffic to the environmental
To this end, rather than transmitting new data according to the duty
conditions. While they can indeed improve the capacity of the network,
cycle, in DiPTC, each node draws first a transmission probability, in
they do not control the amount of traffic sent by the nodes, and they
order to decide if it can send data.
do not ensure that the traffic respects the constraints of the application.
Assuming node 𝑛 is allowed to transmit, it uses its weighting co-
Still, they can easily be combined with traffic reduction algorithms for
better efficiency. efficient, 𝛼𝑛 , to determine the number of messages it should send
(3) Traffic scheduling. Collisions can be reduced by scheduling all (see Fig. Fig. 1, step (a)). At the end of the time period 𝛥𝑇 , the
the traffic in the network. This schedule can be done by a central network server compares the number of received packets k with the
entity that assigns transmission slots to each end-node based on several objective K: if lower (𝑘 < 𝐾), the network server sends a feedback
parameters [17,21,22], or in a distributed manner, where each end- message to the gateway, which then forwards it to all nodes in the
node chooses its transmission slot [23]. However, both approaches add area, to increase their weighting coefficient, therefore their number of
unwanted overhead in the network (due to the dissemination of the transmitted messages; if higher (𝑘 > 𝐾), the opposite occurs (weighting
transmission schedule) and demand high clock accuracy of end-nodes coefficient decreases — see Fig. Fig. 1, step (b)). When the feedback
(which is hard to meet in real life). A simpler approach would be message is received by the end nodes, according to the DiPTC algo-
to only offset the transmission timings, to avoid unnecessary packet rithm, an updated weighting coefficient, 𝛼𝑛 , is locally computed by
transmission, as proposed by Kaburaki et al. [24]. The authors propose each end node, and the number of messages to send is re-evaluated
an autonomous decentralized traffic control system that uses Q-learning (see Fig. Fig. 1 step (c)).

2
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

Fig. 1. DiPTC in a nutshell. (a) Initial transmission considering the initial weighting coefficient equal to 0.5 (for instance). (b) Feedback control send by the gateway to adapt the
uplink traffic intensity. (c) Local computation of the new weighting coefficient and the new number of messages to send, based on the control feedback and the previous weighting
coefficient.

3.2. Assumptions

In this paper, we examine the scenario of a single gateway covering

a given area. The multi-gateway scenario can be simply analyzed since
the network server has the DiPTC functionality, not gateways. We also
consider that the number of nodes associated to the network, denoted
by N, is unknown. This number may change over time as a result
of subsequent deployments, hardware failures, or depleted batteries.
Our proposal can be useful for monitoring applications that require
samples of a physical quantity over time, such as pollution [28]. These
applications only require K measurements over a certain area at regular
intervals (every 𝛥𝑇 ), regardless of how many nodes are in the network.
In LPWANs, the downlink is primarily used to provide control
information to change the physical layer configuration, as in LoRaWAN,
and to acknowledge correctly received packets by the gateway [1]. In
DiPTC, we propose to use also the downlink as a source of feedback
information, indicating if the application goal is reached. This feedback
may include the quantity of packets received over the previous period
𝛥𝑇 , as well as the quantity of K packets required by the application,
or it may include less detailed data, such as whether the quantity
of packets received falls within the application objective, or not. In
general, any other type of message could be considered in order to
notify goal achievement. Each node can adjust its traffic to match
the requirements of the application by relying on the information
received and always abiding by local laws, such as the duty cycle.
The implementation of the broadcast or multicast required to send this
feedback message in an LPWAN is not covered in this paper but will be
discussed in section Section 4 for the case of LoRaWAN. Fig. 2. Finite state machine of the DiPTC algorithm, for each node.

3.3. The DiPTC algorithm

using the coefficient 𝑥𝐷 . The finite state machine in Fig. Fig. 2 describes
We propose DiPTC, based on the additive increase multiplicative the DiPTC algorithm adopted by each node in the network. We define
decrease algorithm. We use a decrease factor, 𝑥𝐷 , to apply a reduction 𝑚𝑛 (𝑡 + 𝛥𝑇 ) as the number of messages to send by the node 𝑛 during the
strategy when the network server receives excessive amounts of data. next time period 𝛥𝑇 , 𝛼𝑛 (𝑡 + 𝛥𝑇 ) the weighting coefficient for node 𝑛
On the other hand, we use an increasing policy with an increase factor, which is used to compute the number of messages to send, 𝑀𝑎𝑥𝐷𝑇 is
given by 𝑥𝐼 , when the network server does not get enough data. This the maximum number of messages a node can transmit during 𝛥𝑇 w.r.t.
algorithm provides a prompt response by rapidly altering the number of its duty cycle. Initially, the number of messages to be sent is null for
transmitted messages by every node, depending on the information sent all nodes: it avoids a burst of uplink messages when new end nodes are
by the network server through the gateway in each time period 𝛥𝑇 . To deployed in the network. The node then waits for the feedback message
lighten the downlink message payload, which is highly constrained, the for the next time period 𝛥𝑇 before adjusting the number of messages
feedback message sent by the network server through the gateway is a to send. The network server only sends the feedback message via the
binary value. In other words, the gateway will either spread 0, meaning gateway if the number of received messages differs from the required
that too many packets were received, or 1 if not enough packets were number. This reduces downlink traffic.
sent. When the goal is met, the network server and gateway send no To avoid an event-based synchronization between nodes which
downlink, and the nodes continue to send the same number of messages would lead to a situation where all the nodes react identically with
for the next time period. The traffic intensity of each node rises linearly, the same number of messages to send, a node decides to take into
taking the coefficient 𝑥𝐼 into account, as long as the network server account the next feedback message following a Bernoulli distribution
does not receive enough K packets required by the application per (𝐵(𝑝), see the stochastic variable 𝑉𝑛 (𝑡)). If the outcome of the Bernoulli
time period 𝛥𝑇 . Otherwise, when the network server receives more distribution is positive, the node will wake up to receive the feedback
than K packets, the nodes reduce their traffic intensity exponentially message; otherwise, it will sleep and wake up to transmit its messages

3
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

without adjusting its traffic intensity. Since receiving a downlink mes- The values used for the parameters in these models are presented in
sage consumes a considerable amount of energy in LPWANs, in this Table 2. The LoRa parameters (SF, CR, BW, Freq.) are chosen randomly
new version of DiPTC, a node opens its reception window to receive the at the beginning, and they do not change during the simulation. In
feedback message from the gateway only if it should adapt its traffic as our setup, we do not use channel hopping to be robust in case of
a result of a favorable draw (𝑉𝑛 (𝑡) = 1), which saves energy compared interference and to avoid collisions, in order to study DiPTC behavior
to the previous version of the algorithm [10]. The finite state machine in the worst case (i.e., high collision probability).
of the new version (Fig. Fig. 2) shows how we modified these three
steps to improve the energy efficiency of end nodes and the network 4.3. Feedback message support in LPWANs
lifetime.
When the feedback message is received and taken into account, As discussed previously, the downlink message sent by the gateway
each node increases or decreases its number of transmitted messages is necessary to disseminate the feedback information to the end-devices
considering the weighting coefficient 𝛼𝑛 (𝑡) according to the binary value in order to increase/decrease the traffic intensity to meet the applica-
of the feedback message received from the gateway. On the one hand, tion requirement. In LoRaWAN, multicast is available for Class B and
each node increases the number of messages to transmit using 𝑥𝐼 , Class C, but not in Class A. How to implement a multicast downlink
𝑥𝐼 ∈ ]0, 1], if the value of the feedback message is equal to 1. On the in Class A is nowadays an open question. We could define a multicast
other hand, it uses the multiplication factor 𝑥𝐷 to reduce the number address for a group of end-devices belonging to the same application
of transmitted messages exponentially, 𝑥𝐷 ∈ ]0, 1]. The additive factor and we could use the reception windows RX1/RX2 to disseminate the
𝑥𝐼 and the multiplicative factor 𝑥𝐷 are constant in time and identical feedback message.
for every node. They represent respectively, the increase or decrease
in traffic compared to the last period. When the value of 𝑥𝐼 is greater,
5. Overall performance evaluation
the amount of traffic at each node increases more quickly; conversely,
when the value of 𝑥𝐷 is greater, the amount of traffic decreases more
This section covers the performance evaluation of DiPTC. We de-
slowly. Next, the node schedules its 𝑚𝑛 (𝑡 + 𝛥𝑇 ) messages, which is
scribe here the simulation scenarios, the DiPTC configurations, the
the integer portion of 𝛼𝑛 (𝑡 + 𝛥𝑇 ). A node is said to be active when
LoRa [31] settings, and comparative solutions. Then, considering sev-
𝑚𝑛 (𝑡 + 𝛥𝑇 ) > 0. Optimized scheduling is out of the scope of this work,
and any scheduling algorithm could be applied. eral metrics (success rate, collision rate, and network lifetime), we
compare the performance of DiPTC against the baseline LoRaWAN and
4. Simulation model an optimal solution.

While DiPTC can be used with any LPWAN technology, we focus 5.1. Scenarios and network parameters
here on applying it to LoRaWAN. To implement and test our proposal,
we used LoRaSim [29], a well-known discrete event-based network We consider the case of a single central gateway surrounded by N
simulator, as it allows us to simulate collision, capture effect, and inter- randomly distributed nodes with LoRa parameters chosen at random.
ference in the network. We improved this simulator by adding downlink We evaluate and compare the performance of DiPTC with baseline
communication and a battery depletion model, as we discuss below. LoRaWAN using the three simulation scenarios listed in Table 1. Each
scenario differs from the number of nodes N in the network, the number
4.1. Wireless environment of measurements K required by the application in the time period
𝛥𝑇 , the increasing and decreasing parameters (𝑥𝐼 , 𝑥𝐷 ), the adaptation
Since LoRaSim does not implement downlink communication in probability 𝑃𝑎𝑑𝑎𝑝𝑡 , and the simulation time Simtime. The adaptation
LoRaWAN, we first enhanced the simulator by simulating downlink parameters 𝑥𝐼 , 𝑥𝐷 , and 𝑃𝑎𝑑𝑎𝑝𝑡 depend on the number of nodes in
communication. Second, we used a random variable that follows a the network and the traffic intensity and were chosen accordingly.
Bernoulli distribution of parameter 𝑃𝐷𝐿 to describe the reliability of the We present a detailed study on how to choose these parameters in
downlink. Third, we took into account the work of Roedig et al. [29] Section Section 7.
to describe collisions and interference on the uplink: To evaluate the performance of DiPTC, we consider two different
scenarios: Intensive and Dense. In the Intensive scenario, we consider a
𝑃𝑟𝑥 = 𝑃𝑡𝑥 + 𝐺𝐿 − 𝐿𝑝𝑙 (𝑑) (1)
significantly higher traffic load to test the scalability of our proposal. In
the Dense scenario, we consider a significantly larger number of nodes
𝐿𝑝𝑙 (𝑑) = 𝐿̄ 𝑝𝑙 (𝑑0 ) + 10𝛾 log(𝑑∕𝑑0 ) + 𝑋𝜎 (2) to test the spatial scalability of our proposal. By testing DiPTC under
where 𝑃𝑟𝑥 is the power of the received signal, 𝑃𝑡𝑥 the transmission these different scenarios, we are confident that it will perform well in
power, 𝐺𝐿 the accumulated general gains losses along the communica- a variety of situations. To the best of our knowledge, we are the first
tion path, 𝐿𝑝𝑙 (𝑑) the path loss in dB at the communication distance 𝑑, ones to propose a traffic control algorithm for LPWANs, so there are
𝐿̄ 𝑝𝑙 (𝑑0 ) the mean path loss at the reference distance 𝑑0 , 𝛾 the path loss not other existing solutions against which we can compare. Still, we
exponent, and 𝑋𝜎 ∼ 𝑁(0, 𝜎 2 ) the normal distribution with zero mean, consider that it is important to have a comparison benchmark against
and 𝜎 2 the variance to account for shadowing. the existing LoRaWAN solution and against an optimal solution that
we call the Centralized Optimal Traffic Control algorithm (COTraC).
4.2. Energy consumption To propose a fair comparison and to respect the application constraints
(i.e., receiving K measurements every time period 𝛥𝑇 ), in baseline
No energy consumption model for the end nodes is implemented LoRaWAN, nodes send their packet following a Poisson distribution
by LoRaSim. We expanded the simulator by adding the one mentioned with the rate: AVG= 𝛥𝑇 𝐾
× 𝑁. In the COTraC solution, we assume that
in [30]. After sending 𝑚 messages to each node 𝑛, the energy 𝐸𝑐 used the network server is a central entity with access to all the information
depends on the time on air 𝑇 𝑂𝐴, the power used in the reception mode about the end devices in the network: residual energy, duty cycle, and
𝑃𝑤𝑅𝑥 , and the power used in the transmitter mode 𝑃𝑤𝑇 𝑥 . LoRa parameters. The network server establishes an optimal scheduler,
where in a round-robin way, each node sends no more than the K
𝐸𝑐 = 𝑇 𝑂𝐴 ∗ (𝑃𝑤𝑇 𝑥 ∗ 𝑚 + 𝑃𝑤𝑅𝑥 ) for DiPTC (3) messages per 𝛥𝑇 required by the application. If a node can no longer
send messages, because of energy depletion or because of the duty cycle
𝐸𝑐 = 𝑇 𝑂𝐴 ∗ 𝑚 ∗ (𝑃𝑤𝑇 𝑥 + 𝑃𝑤𝑅𝑥 ) for LoRaWAN (4) constraint, the next one takes over in sending the remaining ones.

4
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

Table 1 Table 3
Simulation scenarios. DiPTC, Baseline LoRaWAN and COTraC performance.
Scenario Intensive Dense Sc. Measures Baseline LoRaWAN DiPTC COTraC
DiPTC 𝜇𝑐 0.90% 58% 95.3%
Intensive 𝜏𝑐 0.70% 0.074% 0%
𝐍 150 500
𝑡𝑙 25 096 min 34 702 min 54 334 min
𝐊 10 1
𝛥𝐓 1 min 10 min 𝜇𝑐 28.99% 97.62% 98.6%
𝐱𝐈 0.5 0.5 Dense 𝜏𝑐 0.085% 0.026% 0%
𝐱𝐃 0.5 0.5 𝑡𝑙 8912.83 h 13 393.5 h 291 033.33 h
𝐏𝐚𝐝𝐚𝐩𝐭 0.06 0.5
𝑆𝑖𝑚𝑡𝑖𝑚𝑒 1 year 1 year
LoRaWAN
Interarrival time 15 min 5000 min
network lifetime, thanks to our traffic control mechanism which limits
the medium use.
However, the COTraC solution outperforms DiPTC and LoRaWAN.
Table 2 With the overall knowledge of the network parameters and the end
LoRa and network parameters.
nodes configuration in the COTraC solution, the network server sched-
Parameters Values
ules the nodes traffic with respect to their energy so that the next
Payload size (PL) 20 bytes
one takes over immediately after the active node’s death or before
Header length (H) 0
Preamble symbols 8
the excess of its duty cycle limit without going through a transient
regime. That explains the near-perfect success rate. Note that for all
Downlink reception probability 𝑃𝐷𝐿 0.99
scenarios the success rate in the COTraC is not 100% because the uplink
Transmission power 𝑃𝑇 𝑥 14 dBm
reliability is taken into consideration in the evaluation. The collision
Gain and Loss 𝐺𝐿 0
Path Loss exponent 𝛾 2.08 rate is zero in all scenarios because only one or two nodes are active
Reference distance 𝑑0 40 m simultaneously, and send the message(s) which lowers the collision
Max. distance to the gateway 300 m probability to zero. Furthermore, in this scenario, the application re-
Path Loss at the reference distance 𝐿𝑝𝑙 (𝑑0 ) −127.41dB
quires 10 messages every minute. Those ten messages are sent by one
Normal distribution 𝑋𝜎 N(0,3.57)
or two nodes in COTraC. In DiPTC, however, they are sent by one,
Current drawn during the receive mode 𝐼𝑅𝑥 11.2 mA
two, or even ten different nodes. As a result, reaching the application
Current drawn during the Transmission mode 𝐼𝑇 𝑥 90 mA
Current drawn during the sleep mode 𝐼𝑆𝑙𝑒𝑒𝑝 1 μA requirement in DiPTC takes more time and energy than in COTraC.
Supply voltage 3 V Furthermore, as previously stated, COTraC does not have a transient
Battery capacity 30 J regime, whereas DiPTC does. All these facts explain COTraC’s high
success rate when compared to DiPTC.
In the Dense scenario, DiPTC presents a success rate of 97%, which
is three times that of LoRaWAN. This demonstrates that, despite the
Table 2, summarizes the values of the different parameters used in
high number of nodes, DiPTC can meet the application requirements
our simulation. The values of the propagation model are determined
more often than baseline LoRaWAN. In addition, when compared to Lo-
empirically in [29], and those of the energy consumption model come
RaWAN, DiPTC reduces collision rates threefold and increases network
from the datasheet of the Semtech SX1272 LoRa transceiver [32].
lifetime. Compared to COTraC, DiPTC performs as well as the COTraC
Moreover, when a packet is lost, baseline LoRaWAN is configured
solution in terms of success rate and collision rate with a shorter
to re-transmit it a maximum of 8 times before dropping it, a value
network lifetime. In fact, in this scenario one message is required by
commonly used in different implementations given that the standard the application each period of time 10min. This message can be sent
specifies a maximum of 15 re-transmissions. In DiPTC and in the only by one node which lowers the nodes competition in DiPTC and
COTraC there are no re-transmissions, as the nodes do not receive an fits perfectly with the COTraC solution where the transmission is done
unicast acknowledgment from the gateway like in LoRaWAN. by one node all the time.

5.2. DiPTC vs. Baseline LoRaWAN vs. COTraC 6. Horizontal and vertical scalability

In this section, we compare the overall performance evaluation of In this section, we take a deep dive into the behavior of DiPTC in
DiPTC against baseline LoRaWAN and COTraC, using the following order to better understand the results that we obtained in the previous
metrics: section. To be more precise, we look at the number of sent and received
packets, number of collisions, dead nodes, and downlink/uplink losses
• Success rate 𝜇𝑐 : measures the number of times the base station throughout the whole network lifetime. We evaluate these metrics in
receives exactly the K required measurements per period 𝛥𝑇 . the scenarios that put the most strain on the network, allowing us to
• Collisions rate 𝜏𝑐 : measures the number of packet collisions in study the impact of network densification and traffic intensification on
the network divided by the number of transmitted packets. The DiPTC performance..
packet reception rate is 1 − 𝜏𝑐 .
• Network lifetime 𝑡𝑙 : measures the time the network is able to 6.1. Impact of the traffic intensity
support the application requirements (i.e., the nodes are able
to send the K required measurements per period 𝛥𝑇 before the In this section, we focus on the Intense scenario, to test the vertical
exhaustion of their battery). scalability of DiPTC. First, we want to ensure that using DiPTC does not
damage data collection by introducing a spatial bias; a problem that
Table 3 summarizes the simulation results for these metrics consid- could be extremely detrimental to the requirements of this application.
ering the two scenarios Intensive and Dense. To do so, we plotted out the node activity at different times throughout
Despite the high traffic in the Intensive scenario, DiPTC has a success the simulation in Fig. 3. A green-colored node indicates its activity. As
rate of 58%, which is 64 times higher than LoRaWAN. Furthermore, we can see from these figures, messages sent to the gateway come from
DiPTC has a collision rate 32 times lower than LoRaWAN and a longer all over the targeted area, which shows that DiPTC is able to provide a

5
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

Fig. 3. Spatial distribution of active nodes, 43200 min (1 month), Intensive scenario - DiPTC config. (0.5, 0.5, 0.06)..

Fig. 4. Traffic evolution in the Intensive scenario for DiPTC (0.5, 0.5, 0.06) and a downlink probability reception of 0.99.

Fig. 5. Traffic evolution in the Intensive scenario for DiPTC (0.5, 0.5, 0.06) and a downlink probability reception of 0.70.

form of spatial load-balancing. That is not the case in COTraC, where though the transient state could experience some delays, the stationary
one or two nodes take over data transmission during a time period 𝛥𝑇 . state is unaffected, showing that DiPTC is resilient to downlink loss.
Fig. 4 illustrates the evolution of the number of collisions, sent DiPTC performs significantly better than LoRaWAN in the Intensive
and received packets, dead nodes, downlink, and uplink losses, in the scenario, as shown in Figs. 4 and 6. Due to dead nodes and downlink
case of the Intensive scenario for DiPTC. In the latter, the results for losses, Baseline LoRaWAN sees a decrease due to an important amount
DiPTC show multiple alternations between long transient and short of uplink losses and collisions. The convergence of our proposal de-
stationary states. These oscillations are caused by instabilities brought pends mainly on the amount of lost uplink messages. Furthermore, in
by numerous variables, including frequent and consecutive uplink and this case, the maximum absolute error is 7, which means that at worst,
downlink losses, active node deaths, and a high collision probability. 17 messages are received at the gateway. With network density taken
The probability is higher in this scenario since we shorten the time into account, this error is rather small.
period (𝛥𝑇 = 1 min) while increasing the number of measurements Fig. 8 represents the application error over time in the case of
required by the gateway (K = 10). the Intensive scenario, respectively for baseline LoRaWAN and DiPTC.
If we look at the traffic evolution in the zoomed area (Fig. 7) we The application error in DiPTC has lower values and frequencies than
notice that the number of packets received fluctuates at around 10 LoRaWAN. An application error of zero means that the application
packets every minute with an error of at least three packets. These received exactly 𝑘 = 10 measures. Note that the success rate of DiPTC
oscillations are triggered by the loss of an uplink packet from node is 19 times higher than baseline LoRAWAN. This indicates that our
6 at 10 175 min and two other packets belonging to nodes 6 and algorithm can fulfill the demands of the application more often than
96 at 10 182 min. They persist for about twenty minutes, especially baseline LoRaWAN, even in the case of intensive traffic.
with the loss of downlink messages and collisions. At the end of the Finally, Fig. 9 shows the nodes traffic evolution in the Intensive
zoom, we can notice a 5 min delay in the transient regime due to the scenario for DiPTC and baseline LoRAWAN. In DiPTC (Fig. 9(a)),
downlink reliability. However, our proposal manages to converge with the traffic adaptation rate increases when the number of active node
an acceptable level of error given the density of the network. deaths increases, as expected. Unlike baseline LoRaWAN (Fig. 9(b)),
In order to better evaluate the effects of downlink loss on DiPTC, in DiPTC, the average network traffic increases with the nodes death
we lower the downlink reception probability to 𝑃𝐷𝐿 = 0.70. As we can accumulation. In baseline LoRaWAN, the average network traffic is
see in Fig. 5, we have the same general evolution pattern as in Fig. 4 still regular after the nodes death accumulation, because of a traffic
where the downlink reception probability is higher 𝑃𝐷𝐿 = 0.99. Even distribution generation that is fair on average.

6
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

Fig. 6. Traffic evolution in the Intensive scenario for Baseline LoRaWAN.

Table 4
DiPTC configurations.
𝑃𝑎𝑑𝑎𝑝𝑡 𝑥𝐼 𝑥𝐷
0.2 0.2
0.1 0.2 0.9
or 0.5 0.5
0.5 0.9 0.2
0.9 0.9

In comparison to DiPTC, we notice that the traffic in LoRaWAN

appears to be lower (Fig. 11). In fact, the maximum number of received
Fig. 7. Zoom on the traffic evolution in the Intensive scenario for DiPTC(0.5, 0.5, messages per period is only 28 (𝛥𝑇 = 10 min), while in DiPTC it
0.06).
is 138. The baseline LoRaWAN oscillations, on the other hand, are
more frequent. This is due to the fact that LoRaWAN packet genera-
tion follows a Poisson distribution of rate 0.0002, which implies that
one packet is sent every 5000 min on average, causing the nodes to
traffic less often but more regularly. As illustrated in Fig. 12, DiPTC
outperforms LoRaWAN when it comes to application error frequency.
The null application error (𝑒 = 0) occurs three times more frequently
with DiPTC than with LoRaWAN.
Finally, Fig. 13 shows node traffic evolution in the Dense scenario
for baseline LoRaWAN and DiPTC. It illustrates the necessity to have a
frame transmission control in LoRaWAN especially when the number of
nodes is important. We can notice that the maximum number of packets
a node can send in DiPTC remains significantly small (2 messages),
Fig. 8. Frequency error in the Intensive scenario for DiPTC(0.5, 0.5, 0.06) and Baseline although the number of nodes increases in the network, contrary to
LoRaWAN. LoRaWAN. In fact, compared to baseline LoRaWAN, the maximum
number of packets a node may send in DiPTC is 7 times lower, and
the network traffic average is generally smoother.
6.2. Impact of the nodes scalability
7. Impact of traffic control parameters
In this section, we focus on the Dense scenario, to test the horizontal
scalability of DiPTC. We increased the number of nodes from 150 In the previous sections, DiPTC parameters (𝑥𝐼 , 𝑥𝐷 , and 𝑃𝑎𝑑𝑎𝑝𝑡 )
to 500, and decreased the traffic intensity compared to the Intensive were chosen and fixed according to the characteristics of the simulated
scenario. network that took into account node density and traffic intensity. As
Note in Fig. 10(a) that there is an alternation of transient and a reminder, 𝑥𝐼 and 𝑥𝐷 represent how much the traffic will increase
stationary states, with the former only lasting for a short time due to or decrease from the previous period. The larger the value of 𝑥𝐼 , the
factors such as node death or packet loss. What this suggests is that faster the increase, and the greater the value of 𝑥𝐷 , the slower the
DiPTC can quickly converge to the desired number of measurements, decrease. The last parameter, 𝑃𝑎𝑑𝑎𝑝𝑡 , controls nodes adaptation traffic.
known as K, despite obstacles like fluctuating environments. As in the This section investigates the effects of these three parameters on DiPTC
Intensive scenario, the transient states are generally caused by issues performance. We investigate the different values described in Table 4.
including node death and uplink loss. Even if a node’s downlink is lost, Fig. 14 compares the error and transient regimes duration fre-
our algorithm will continue to function as the node itself keeps the same quencies, in the Intensive scenario for DiPTC, with a 0.5 adaptation
traffic intensity until it hears back from the gateway. The reason we probability. We notice that the five configurations have significant
see a lot of long stationary states is because of our algorithm which different results. In this scenario, the adapting probability is set to 0.5,
provides stability and reduces collisions. There are fewer collisions in meaning that at least 75 nodes are capable of adapting their traffic,
general because there is less traffic overall (since K = 1). Low traffic has hence competing to send the 10 required measurements. Fig. 14(a)
also a positive effect on the lifetime of a node, as it reduces the battery displays that the best configurations, with regards to receiving exactly
depletion process. Despite the significant increase in the number of 10 messages each period of time 𝛥𝑇 = 1 min from 150 nodes i.e.,
nodes, our approach can quickly converge to the target number of (𝑒𝑟𝑟𝑜𝑟 = 0), are in order (0.2, 0.9, 0.5), (0.2, 0.2, 0.5), (0.9, 0.9, 0.5)
measurements, K, and maintain stability for a considerable amount of and finally (0.9, 0.2, 0.5). Even with intensive traffic, the configurations
time. with a low increase give the best results in terms of error frequency.

7
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

Fig. 9. Nodes traffic evolution in the Intensive scenario.

Fig. 10. Traffic evolution in the Dense scenario for DiPTC(0.5, 0.5, 0.5).

Fig. 11. Traffic evolution in the Dense scenario for Baseline LoRaWAN.

frequency of the short transient regimes duration. In fact, in this

scenario, at least 9 nodes are competing to send the ±10 messages each
1 min. There are greater chances to meet the application requirements
with 9 (𝑃𝑎𝑑𝑎𝑝𝑡 = 0.1) nodes than with 75 nodes (𝑃𝑎𝑑𝑎𝑝𝑡 = 0.5). Note
that if the goal of the application is receiving exactly ±10 messages
each period of time 𝛥𝑇 = 10 min, then the best configurations are
in order (0.2, 0.2, 0.06), (0.2, 0.9, 0.06), (0.5, 0.5, 0.06), (0.9, 0.2,
0.06) and finally (0.9, 0.9, 0.06). With fewer competing nodes, the best
configuration is (0.2, 0.2, 0.06). However, if the application allows an
error of ±1, (0.2, 0.9, 0.06) is by far the best one.
Fig. 16 compares the performance measures of the different DiPTC
configurations in the Intensive scenario. As we already notice in the
Fig. 12. Error frequency in the Dense scenario for DiPTC(0.5, 0.5, 0.5) and Baseline
LoRaWAN. previous figures, the configuration (0.2, 0.9, 0.06) performs better in
comparison to the others (Fig. 16(a)). In this scenario (Fig. 16(b))
where at most 9 nodes compete to meet the application requirements
A weak decreasing coefficient boosts the performance. Since nodes (±10 message(s) at each period of time 𝛥𝑇 = 1 min), configurations
increase and decrease their traffic slowly, it is more likely to achieve the with a weak increasing coefficient 𝑥𝐼 traffic and consume less energy,
application requirements, i.e., receiving 10 measurements in a period therefore providing longer lifetime.
of time 𝛥 = 1 min. Those configurations also show more frequent short Fig. 17 compares the error and transient regimes duration frequen-
regimes than the other ones 14(b). cies, in the Dense scenario for DiPTC, with a 0.5 adaptation probability.
Fig. 15 compares the error and transient regimes duration fre- The best configurations, with regards to receiving exactly 1 message
quencies, in the Intensive scenario for DiPTC, with a 0.06 adaptation each time period 𝛥𝑇 = 10 min from 150 nodes, (i.e., 𝑒𝑟𝑟𝑜𝑟 = 0) are
probability. As we can see, decreasing the adaptation probability in- in order (0.2, 0.2, 0.5), (0.5, 0.5, 0.5), (0.9, 0.2, 0.5) and finally (0.2,
creases significantly not only the frequency of sending ±10 messages 0.9, 0.5). If the application allows an error of ±1, the configurations
per time period 𝛥𝑇 = 1 min (i.e., error 0, 1 and −1) but also the (0.2, 0.9, 0.5) and (0.2, 0.2, 0.5) and (0.5, 0.5, 0.5) take the lead. We

8
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

Fig. 13. Nodes traffic evolution in the Dense scenario.

Fig. 14. A comparison of error frequency and transient regimes duration frequencies, in the Intensive scenario for the DiPTC when 𝑃𝑎𝑑𝑎𝑝𝑡 = 0.5.

Fig. 15. A comparison of error frequency and transient regimes duration frequencies, in the Intensive scenario for the DiPTC when 𝑃𝑎𝑑𝑎𝑝𝑡 = 0.06.

Fig. 16. A comparison of performance measures of different DiPTC configurations (𝑥𝐼 , 𝑥𝐷 , 𝑃𝑎𝑑𝑎𝑝𝑡 ) in the Intensive scenario.

also observe an error increase and the apparition of longer transient nodes). Thus they make smaller errors, but take a longer time to
regimes. When the number of nodes increases, the traffic increases, adjust their sending. The best configurations are characterized by small
leading to more errors and longer transient regimes. frequencies for larger errors, important frequencies when the error is 0,
Fig. 18 compares the error and transient regimes duration frequen- and finally short transient regimes duration. In the Dense scenario and
cies, in the Dense scenario for DiPTC, with a 0.1 adaptation probability. for an adaptation probability of 0.1, the best configurations are (0.2,
As we can see, the decrease in the adaptation probability decreases 0.2, 0.1), (0.5, 0.5, 0.1), (0.9, 0.2, 0.1).
the errors and increases the duration of the transient regimes. If the Fig. 19 contrast the performance measures of the different DiPTC
adaptation probability has a small value 𝑃 = 0.1, the number of configurations in the Dense scenario. These figures display the presence
nodes allowed to adapt their traffic each period of time is smaller of a trade-off between the network lifetime and the success rate.
(50 nodes), compared to when the adaptation probability is 0.5 (250 Actually, in this scenario where the application objective is 1 or 2

9
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

Fig. 17. A comparison of error frequency and transient regimes duration frequencies, in the Dense scenario for the DiPTC when 𝑃𝑎𝑑𝑎𝑝𝑡 = 0.5.

Fig. 18. A comparison of error frequency and transient regimes duration frequencies, in the Dense scenario for the DiPTC when 𝑃𝑎𝑑𝑎𝑝𝑡 = 0.1.

Fig. 19. A comparison of performance measures of different DiPTC configurations (𝑥𝐼 , 𝑥𝐷 , 𝑃𝑎𝑑𝑎𝑝𝑡 ) in the Dense scenario.

message(s) at a period of time 𝛥𝑇 = 10 min, configurations with downlink reception probability. Even though the transitory state
a low increasing coefficient 𝑥𝐼 traffic less and consume less energy, could experience some delays, the stationary state is unaffected.
therefore they perform better, in terms of network lifetime. While the (2) An active node battery depletion leads to its death, which in turn
configurations with a strong increasing coefficient adapt their traffic creates a new, short-lived transient mode.
more often to the desired messages and perform better w.r.t. the success (3) The reliability of the uplink has a detrimental influence on the
rate measure. Notice that the packet reception rate and the average DiPTC convergence since it yields lengthy transient states and
channel inactivity time are equal for the five configurations in both brief stationary states.
figures. (4) DiPTC converges exactly towards K measurements per time
period in the Dense scenario. In comparison to the Intensive
8. Discussion & insights scenario, the stationary states have a relatively long duration.
(5) Despite the frequent and longer transient states in the Intensive
The algorithm we propose achieves excellent results in terms of scenario, DiPTC converges towards the required measurements
convergence, not just for high traffic networks, but also for those with (K) with a reasonable absolute error.
a large number of nodes. It shows that the use of a control traffic (6) The loss of transmission due to a collision or the propagation
management in LPWAN increase widely the overall performance and model has the same effect.
allow to respect strong application requirement. Our algorithm is able (7) In the Intensive scenario the (0.2, 0.9, 0.06) configuration is by
to converge towards the required measurements K with a reasonable far the best in terms of success rate and network lifetime.
absolute error margin owing to the adaptation algorithm stability fea- (8) Unlike the Intensive one, in the Dense scenario, the choice of
ture. In fact, DiPTC is capable of regulating LPWAN traffic such as DiPTC parameters requires a compromise between the network
LoRaWAN with low collision rates and minimal overhead. lifetime, the error tolerated by the application, and the success
Following our analysis, we can draw the next conclusions: rate.
(9) When compared to baseline LoRaWAN, DiPTC is able to achieve
(1) DiPTC is unaffected by downlink unreliability, hence it will the goal within reasonable deadlines, while maintaining a low
continue to operate as intended even in the presence of a low number of collisions with a longer network lifetime.

10
K. Lasri et al. Ad Hoc Networks 143 (2023) 103121

(10) The performance of DiPTC in the Intensive scenario is lower [6] B. Chaudhari, M. Zennaro, S. Borkar, LPWAN technologies: Emerging application
than the performance of the optimal centralized traffic control characteristics, requirements, and design considerations, MDPI Future Internet
(2020).
solution (COTraC), in terms of success rate and network lifetime.
[7] R.M. Sandoval, A.-J. Garcia-Sanchez, J. Garcia-Haro, T.M. Chen, Optimal policy
Note that COTRaC works without collision due to the optimal derivation for transmission duty-cycle constrained LPWAN, IEEE Internet Things
scheduling and without additional control traffic that uses the J. (2018).
downlink, which is not realistic. [8] A. Azari, C. Cavdar, Self-organized low-power IoT networks:a distributed learning
(11) Like the COTraC solution, in the Dense scenario, our DiPTC approach, in: IEEE GLOBECOM, Abu Dhabi, UAE, 2018.
[9] D.T. Ta, Khawam, et al., LoRa-MAB: a flexible simulator for decentralized
is capable of achieving the objective of K measurements per learning resource allocation in IoT networks, in: IEEE Wireless and Mobile
period. Networking Conference, Paris, France, 2019.
(12) Unlike the COTraC solution our DiPTC algorithm presents a [10] K. Lasri, Y. Ben Maissa, L. Echabbi, O. Iova, F. Valois, A new distributed and
spatial load-balancing characteristic. probabilistic approach for traffic control in LPWANs, in: Springer International
Conference on Advanced Information Networking and Applications, AINA, 2021.
[11] H. Ennajari, Y.B. Maissa, S. Mouline, Energy efficient in-network aggregation
9. Conclusion algorithms in wireless sensor networks: a survey, in: Springer UNet, Casablanca,
Morocco, 2016.
In this work, we propose a Distributed and Probabilistic algorithm [12] Y. Ma, Y. Guo, X. Tian, M. Ghanem, Distributed clustering-based aggregation
for Traffic control in LPWANs, called DiPTC. It is based on a multi- algorithm for spatial correlated sensor networks, IEEE Sens. J. (2010).
[13] L. Tan, M. Wu, Data reduction in wireless sensor networks: a hierarchical LMS
plicative increase and multiplicative decrease algorithm, as well as a prediction approach, IEEE Sens. J. (2015).
binary feedback message sent by a gateway. This simple yet effective [14] C. Liu, K. Wu, M. Tsao, Energy efficient information collection with the ARIMA
way of controlling traffic yields very good results. Moreover, DiPTC model in wireless sensor networks, in: IEEE GLOBECOM, St. Louis, MO, USA,
ensures convergence of the application requirements (K measurements 2005.
[15] J.-L. Lu, F. Valois, M. Dohler, Optimized data aggregation in WSNs using adaptive
per period) and reduces collision risk within acceptable delays. DiPTC
ARMA, in: IEEE Fourth International Conference on Sensor Technologies and
also shows a threefold increase in success rate over Baseline LoRaWAN Applications (SensorComm), Venice, Italy, 2010.
with a significant longevity advantage. The Centralized Optimal Traffic [16] J. Moraes, N. Matni, A. Riker, H. Oliveira, E. Cerqueira, C. Both, D. Rosário, An
Control algorithm COTraC outperforms our decentralized DiPTC be- efficient heuristic LoRaWAN adaptive resource allocation for IoT applications,
cause of the network knowledge and no feedback message is needed. in: IEEE Symposium on Computers and Communications, ISCC, 2020.
[17] Z. Xu, J. Luo, et al., S-MAC: achieving high scalability via adaptive scheduling
However, DiPTC ensures a network spatial load balancing which is not in LPWAN, in: IEEE INFOCOM, Virtual Conference, 2020.
the case for COTraC. [18] L.-H. Shen, C.-H. Wu, W.-C. Su, K.-T. Feng, Analysis and implementation for
In our future work, we plan to improve the formalization of our traffic-aware channel assignment and contention scheme in LoRa-based IoT
model by using advanced approaches such as game theory and rein- networks, IEEE Internet Things J. (2021).
[19] N. Chinchilla-Romero, J. Navarro-Ortiz, P. Muñoz, P. Ameigeiras, Collision
forcement learning, for success rate and network lifetime enhancement.
avoidance resource allocation for LoRaWAN, IEEE Sensors (2021).
[20] Z. Qin, J.A. McCann, Resource efficiency in low-power wide-area networks for
Declaration of competing interest IoT applications, in: IEEE Global Communications Conference, 2017.
[21] K.Q. Abdelfadeel, D. Zorbas, V. Cionca, D. Pesch, 𝐹 𝑟𝑒𝑒 —Fine-grained scheduling
The authors declare the following financial interests/personal rela- for reliable and energy-efficient data collection in LoRaWAN, IEEE Internet
Things J. (2020).
tionships which may be considered as potential competing interests: [22] J. Haxhibeqiri, I. Moerman, J. Hoebeke, Low overhead scheduling of LoRa
LASRI KAWTAR reports financial support was provided by Campus transmissions for improved scalability, IEEE Internet Things J. (2019).
France. LASRI KAWTAR reports financial support was provided by [23] D. Zorbas, K. Abdelfadeel, P. Kotzanikolaou, D. Pesch, TS-LoRa: time-slotted
inria. Oana Iova is an Area editor at Ad Hoc Networks. LoRaWAN for the industrial internet of things, Elsevier Comput. Commun.
(2020).
[24] A. Kaburaki, K. Adachi, O. Takyu, M. Ohta, T. Fujii, Autonomous decentralized
Data availability traffic control using Q-learning in LPWAN, IEEE Access (2021).
[25] U.o.S.C. Information Sciences Institute, RFC 793: Transmission Control Protocol,
Data will be made available on request. Darpa Internet Program Protocol Specification, Tech. Rep., IETF, 1981.
[26] J. Iyengar, M. Thomson, RFC 9000: Quic: A UDP-Based Multiplexed and Secure
Transport, Tech. Rep., IETF, 2021.
Acknowledgments
[27] D. Chiu, R. Jain, Analysis of the increase/decrease algorithms for congestion
avoidance in computer networks, June Comput. Netw. ISDN Syst. (1989).
This research was partially supported by CAMPUS FRANCE (PHC [28] A. Boubrima, W. Bechkit, H. Rivano, On the deployment of wireless sensor net-
TOUBKAL 2019, French-Morocco bilateral program), Grant Number: works for air quality mapping: optimization models and algorithms, IEEE/ACM
Trans. Netw. (2019).
41562UA.
[29] M.C. Bor, U. Roedig, T. Voigt, J.M. Alonso, Do LoRa low-power wide-area
networks scale? in: IEEE MSWiM, Malta, 2016.
References [30] M. Slabicki, G. Premsankar, M. Di Francesco, Adaptive configuration of LoRa
networks for dense IoT deployments, in: IEEE/IFIP NOMS, Taipei, Taiwan, 2018.
[1] N. Sornin, L.M. Eirich, T. Kramp, O. Hersent, LoRaWAN Specification, LoRa [31] A.J. Wixted, P. Kinnaird, H. Larijani, A. Tait, A. Ahmadinia, N. Strachan,
Alliance, 2015. Evaluation of LoRa and LoRaWAN for wireless sensor networks, in: IEEE
[2] Weightless aliance, https://www.weightless-alliance.org/. SENSORS, 2016.
[3] K. Mekki, E. Bajic, F. Chaxel, F. Meyer, A comparative study of LPWAN [32] SX1272/73, Semtech datasheet - 860 MHz to 1020 MHz low power long range
technologies for large-scale IoT deployment, Elsevier ICT Express (2019). transceiver, rev. 4, 2019.
[4] N. Nurelmadina, M.K. Hasan, I. Memon, R.A. Saeed, K.A. Zainol Ariffin, E.S.
Ali, R.A. Mokhtar, S. Islam, E. Hossain, M.A. Hassan, A systematic review on
cognitive radio in low power wide area network for industrial IoT applications, Kawtar Lasri is a doctoral student in joint supervision between National Institute of
MDPI Sustain. (2021). Posts and Telecommunications of Morocco and INSA Lyon of France. Her research is
[5] M. Swain, D. Zimon, R. Singh, M.F. Hashmi, M. Rashid, S. Hakak, LoRa-LBO: an focused on Traffic control in Low Power Wide Area Networks. She received her M.S
experimental analysis of LoRa link budget optimization in custom build IoT test degree in Automatic Computing and Embedded Systems at Mohammed V university in
bed for agriculture 4.0, MDPI Agron. (2021). Rabat, Morocco, in 2017.

11