Bringing Fairness in LoRaWAN through SF Allocation

Optimization
Christelle Caillouet, Martin Heusse, Franck Rousseau

To cite this version:

Christelle Caillouet, Martin Heusse, Franck Rousseau. Bringing Fairness in LoRaWAN through SF
Allocation Optimization. ISCC 2020 - 25th IEEE Symposium on Computers and Communications,
Jul 2020, Rennes, France. �10.1109/ISCC50000.2020.9219611�. �hal-02780468v2�

HAL Id: hal-02780468

https://inria.hal.science/hal-02780468v2
Submitted on 19 Nov 2021

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Bringing Fairness in LoRaWAN through SF
Allocation Optimization
Christelle Caillouet† , Martin Heusse⇤ , Franck Rousseau⇤
⇤ Univ.Grenoble Alpes, CNRS, Grenoble INP, LIG, F-38000 Grenoble France
† Univ.
Côte d’Azur, CNRS, Inria, I3S, F-06900 Sophia Antipolis
Email: [email protected], [martin.heusse,franck.rousseau]@imag.fr

10
Abstract—We propose an optimization model for single-cell
LoRaWAN planning which computes the limit range of each
spreading factor (SF) in order to maximize the minimum packet
delivery ratio (PDR) of every node in the network. It allows to

5
balance the opposite effects of attenuation and collision of the
transmissions and guarantee fairness among the nodes. We show
that our optimization framework improves the worst PDR of

rep(0, 6)
the nodes by more than 13 percentage points compared to usual SF7

0
SF boundaries based on SNR threshold. A study of the tradeoff l5
between precision and resolution time of the model shows its
effectiveness even with a small number of possible distance limits, SF8 l4 l3
l2 l1

−5
and its scalability when the node density increases. SF9
l0
I. I NTRODUCTION
SF12
Low Power Wide Area Networks enable long range com- −10

munications for Internet of Things (IoT) applications such Figure−101: Annuli−5of SF allocation
0 around
5 the gateway
10

as nation-wide cellular networks or private data collection rep(0, 6)

systems. One of the main technology for LPWAN is LoRa

whose physical layer uses CSS (Chirp Spread Spectrum) for different perspective. We seek to balance the packet delivery
robust communication in the sub-GHz ISM band. There are ratio (PDR) in the network to ensure fairness among nodes
several spreading factors (SF) to choose from, which allows by properly allocating the spreading factors (SF), based on an
to trade data rate for range. LoRaWAN defines a channel accurate system model taking into account both attenuation
access method based on ALOHA with rare feedback from the and collisions to derive a representative PDR formula.
gateway. Transmissions using different spreading factors are The contributions of this paper are as follows: we first derive
quasi-orthogonal —in case of a collision, both frames succeed a simple accurate model for collisions and physical capture,
if one is not several orders of magnitude stronger than the which gives better insight into the dynamics of packet loss due
other. For collisions in the same SF, a frame succeeds if it is to ALOHA with capture effect. We consider a traffic model
significantly stronger than the other one. where all nodes have the same traffic intensity. Then, based
One of the main issues in LoRa and LoRaWAN concerns on this model, we derive an optimization framework to fairly
the radio channel capacity. Performances of LoRaWAN have determine the SF boundaries of a network cell in order to
recently been well investigated [1]–[7]. Different models have guarantee the largest PDR to all nodes, regardless of their
been proposed and try to capture the effects of collisions distance to the gateway.
between transmissions occurring in the same SF and the same We start by describing our model in the next section,
channel, that are very harmful to the network performance. before looking at the optimization of the SF boundaries in
Alleviating attenuation or collisions have antagonistic effects Section III. Numerical results are then presented in Section IV.
because switching to a larger SF results in more robust trans- We then review existing work of the literature in Section V
missions but with longer duration, which increases contention, and conclude.
hence collision probability. We will see that it is wise to
control the number of nodes using higher SFs because they II. S YSTEM MODEL AND PERFORMANCE M ETRIC
occupy much more channel capacity than nodes with lower We consider the packet delivery ratio (PDR) as performance
SFs. Similarly, a smaller cell can handle more nodes than metric. The PDR models the ratio of the number of messages
a larger one because SF allocation is less constrained, as received by the gateway over the total number of transmitted
basically all nodes are able to use a small SF. messages, taking into account the channel model, SNR, and
In order to improve LoRaWAN performance, a number collisions due to interferences.
of works have proposed resource optimization methods [8]– We consider a single cell LoRaWAN with one gateway
[10] for SF allocation. We study this question from a slightly located at the center with maximum transmission range l0 = R
and n nodes homogeneously scattered with spatial density Table I: Notations
⇢ = n/(⇡ ⇥ R2 ) on the cell disk of finite radius R. The
Spatial density of nodes ⇢
number of nodes allocated to SF(12 j) is a function of the Traffic generation intensity t
radii defining the annulus between lj+1 and lj around the Frame transmission duration at SF(12 j) ⌧j
gateway (see Figure 1). We neglect the presence of nodes Distance of farthest node at SF(12 j) lj
Traffic occupancy (in Erlang) at SF(12 j) vj
beyond l0 . Each node has the same traffic intensity that Average channel gain at distance d g(d)
depends on the data generation pattern of the application. We SNR threshold for SF(12 j) qj
consider a network with a single application, in which case Transmission power, in-band noise power P, N
Success probability, due to attenuation, fading, thermal noise H
all nodes produce the same traffic following a Poisson process Success probability, due to collisions Q
of intensity t . We base the choice of t on the duty cycle
limitations as explained below.
In LoRaWAN, nodes have to limit their occupation of each
For a given transmission attempt, provided that it has
frequency band to 1% of the time and there are 3 to 5
already met the signal strength reception condition, with
frequency channels in each band —for instance, in Europe,
probability H, we distinguish three cases:
3 in band h1.4 and 5 in h1.3. So we look at saturating h1.4,
with a 0.33% duty cycle per frequency channel. We set t to 1) if the transmission does not overlap with any other
the maximum intensity for any node in the system, which is frame, it succeeds;
determined by the “slowest nodes”, those operating at SF12, at 2) if a single frame with the same SF interferes with
their authorized duty cycle limit, and using 59 B frames —the the transmission, then the frame is still captured and
maximum size at this data rate. It corresponds to ⌧0 = 2.47 s decoded by the gateway if it is received with 6 dB more
of airtime, thus t = 2.47⇥3⇥100
1
s 1 for 0.33% duty cycle. power than the colliding frame [11];
All nodes transmit at power Ptx = 14 dBm in the 125 3) if more than two frames overlap the transmitted one,
kHz-wide band. The signal power at the gateway depends on we simply deem it lost. In fact —as the calculation
the distance and Rayleigh fading. The transmission power is for case 2 makes apparent—, collisions with successful
attenuated depending on the distance to the gateway: capture are already relatively rare, so meeting the capture
condition for two frames is effectively quite unlikely.
Prx = Ptx ⇥ g(di ) (1) For a tagged transmission, the probability of case 1 is
Q1 = exp ( 2vj ) , as there should be no other transmission
where g(di ) is the path-loss attenuation function based on event during 2⌧j to avoid overlap. Case 2 happens when a
the Okumura Hata model in a suburban environment with an single transmission occurs during this time with probability
antenna height of 15 m. We consider a Rayleigh channel, so 2 vj exp( 2vj ). If we neglect the variability of g(d) among
that the received signal power is affected by a random variable the nodes using the same SF, the received power in each
which follows an exponential distribution with unit mean. We annulus follows an exponential distribution. Successful frame
neglect shadowing, since its net effect would be small in our capture occurs for a difference of 6 dB in received power,
case: it would modify the channel gain from each node without which correspond to a factor of 4. Since the probability that
changing the general behavior of the system [6]. an exponential random variable is x times above another one
In case there is no collision, a transmission is successfully is x+1
1
, the success probability of frame capture is 15 and the
received if the SNR at the receiver is above the minimum SNR probability of success in case 2 is
for the corresponding spreading factor [11]. Therefore, the 2
success probability of an isolated frame reception at distance Q2 = vj exp( 2vj ). (3)
5
di is:
Thus, the presence of concurrent traffic impacts the packet
N qf
H = exp( ) (2) reception probability by a ratio Q:
Prx
2
where N is the thermal noise for a 125 kHz-wide band: N = Q = Q1 + Q2 = (1 + vj ) exp( 2vj ). (4)
5
174 + NF + 10 log(BW ), with BW the signal bandwidth
And, combining 2 and 4, we get the probability of successful
and NF the receiver noise figure that can be considered to
packet reception:
have a value of 6 dB in the hardware implementations, and qf
PDR = H ⇥ Q (5)
is the minimum SNR for the corresponding spreading factor
f . We assume an antenna gain of 6 dB.
For SF = 12 j, the number of nodes contending with each III. I MPROVING FAIRNESS WITH A BETTER SF
other is thus nj = ⇡⇢ (lj2 lj+1 2
) so that traffic occupancy ALLOCATION
is vj = nj ⌧j t (in Erlang), where ⌧j is the transmission We study SF allocation, which is the process of assigning a
duration. In this work, we adopt a simplifying assumption that spreading factor SF to each node. In practice, this is achieved
there are no inter-SF interferences, although it may have some using LoRa MAC Adaptive Data Rate (ADR) commands
impact [3], [6], [8]. issued by the network server, telling end nodes which SF to use
 Table II: SNR-based SF boundaries (km)

100
SF7 SF8 SF9 SF10 SF11 SF12 Cell

80
Htarget l5 l4 l3 l2 l1 l0 surface

60
99.4% 1.05 1.26 1.52 1.83 2.14 2.50 20

PDR
92% 2.10 2.53 3.05 3.67 4.28 5.00 79

40
74% 2.94 3.54 4.27 5.14 5.99 7.00 154
⌧j (ms) 102.7 184.8 328.7 616.5 1315 2466 "

20
Fair SF boundaries
qj (dB) -6 -9 -12 -15 -17.5 -20 (km2 ) SNR−based SF boundaries

0
0 1000 2000 3000 4000
Nodes ordered by distance to gateway

based on the observed SNR at the gateway. In the following,

we consider three scenarios with three different radio ranges Figure 2: Small cell radius (2.5km), 4000 nodes. PDR of nodes
of 2.5, 5 and 7 km (small, medium and large cell radius). in function of distance with the 2 SF boundaries policies.

A. SNR-based SF boundaries

100
Before taking into account medium access contention, it is

80
reasonable to start by setting the SF of each node so that
it meets a given target PDR. The idea is to have a similar

60
PDR
success probability Htarget across all annuli. For a given cell

40
radius, Htarget is fixed, for instance, in a medium range cell,
Htarget = 92%, which is the value of H at range 5 km. We

20
Fair SF boundaries
change SF each time the probability of success H drops below SNR−based SF boundaries
this target value. Solving Equation 2 for each j gives the range 0
0 500 1000 1500
for each SF shown in Table II. Nodes ordered by distance to gateway
Dashed red lines in Figures 2, 3 and 4 show the PDR for the
three cell sizes mentioned above. For the small, medium and Figure 3: Medium cell radius (5km), 1600 nodes. PDR of
large cell size, we consider between 1000 and 6500 nodes, nodes in function of distance with the 2 SF boundaries
500 and 4000 nodes, and 50 and 1500 nodes respectively. policies.
We present the evolution of the minimum achievable PDR in
Figure 10 and the PDR repartition in the network for 4000,
1600 and 400 nodes in Figures 2, 3, and 4. We order the nodes one distance d 2 D per lj . We introduce binary variables xjd
by distance to the gateway: e.g. the 4000th node is at a distance stating if lj is set to distance d or not. We derive a linear
of 2.5 km in the small cell case. program for the fair SF boundaries computation as follows.
All figures exhibit the same trend. Close to the gateway, for max minj=0,...,5 PDR(j) (6)
small SFs, there is little contention (vj ⌧ 1) and the PDR P j
is satisfactory. But farther away, two adverse effects combine. s.t. d2D xd = 1 8j = 1, ..., 5 (7)
First, higher SF transmissions take more airtime, resulting in a x0l0 =1 (8)
higher collision probability. Secondly, for a given SF, the area xjd 2 {0, 1} 8d 2 D, j = 0, ..., 5 (9)
of each annuli gets larger when going outward of the coverage
disk, thus including more nodes, resulting in more collisions. The constraints ensure that only one distance is allocated to
Almost half of the nodes end up using the two most robust each SF boundary (Constraints (7)), and we fix the limit range
modulations, SF11 or SF12, resulting in very low PDRs, due of SF12 to the maximum communication range l0 (Constraints
to a large number of collisions. (8)). Recall from Equation (5) the formulation of the PDR :

B. Fair SF boundaries N qj 2
PDR(j) = exp( ) ⇥ (1 + vj ) exp( 2vj ) (10)
Ptx ⇥ g(lj ) 5
To avoid the unwanted effects of the traffic concentration
in peripheral annuli, it is necessary to adjust the SNR at This formula depends on the boundaries lj and lj+1 that have
which the nodes change SF. To this end, we propose an to be determined by our model. We thus have to include the
optimization framework computing the SF boundaries lj such binary variables xjd and xj+1
d0 , with d < d, in the expression
0

that it maximizes the minimum PDR among all SFs. of the objective function (6) to compute the actual value of the
Given the node distribution ⇢, we seek to maximize the PDR if SF12 j and SF12 (j+1) boundaries are set to d and
minimum value of the PDR at the boundaries lj , j = 1, ..., 5. d0 . Objective function (6) can thus be rewritten as follows:
To avoid the complexity of finding values lj in a continuous
space, we discretize the cell radius segment [0, l0 ] into a set X (
N qj
)
X 2
max min e Ptx ⇥g(d) xjd ⇥ (1+ vj )e( 2vj ) j+1
x d0 (11)
D of possible distances (the number of samples |D| is a j=0,...,5
d2D d0 <d
5
parameter given as input of the model) and we try to allocate
 Table III: Fair SF boundaries (km)
100

SF7 SF8 SF9 SF10 SF11 SF12

80

Hmin l5 l4 l3 l2 l1 l0
60

95.7% 1.70 2.11 2.32 2.43 2.47 2.50

PDR

68.9% 3.03 3.77 4.30 4.68 4.88 5.00

40

57.2% 3.40 4.20 4.99 5.86 6.51 7.00

20

Fair SF boundaries
SNR−based SF boundaries
0

0 100 200 300 400

Nodes ordered by distance to gateway

Figure 4: Large cell radius (7km), 400 nodes. PDR of nodes

in function of distance with the 2 SF boundaries policies.

with vj = ⇡⇢ (d2 d02 )⌧j t . We finally linearize Equation (11)

j j j+1
by introducing a new set of binary variables ydd 0 = xd ⇥ xd0

to obtain a linear program solved using IBM Cplex solver.

Figure 5: Resolution time of the model in function of the
IV. N UMERICAL RESULTS number of samples of the cell distances |D|.

We present the results of our model for the three sizes of

LoRa cell presented in the previous section. However, we have A. Solution accuracy
developed the model so that it can be executed for various Our model relies on the discretization of the cell radius. The
scenarios. The only parameters needed to run the model are optimal solution is approximated by considering discrete val-
the number of nodes, the maximum range l0 of the cell, and ues for all the possible distances in [0, l0 ]. The smaller the step
the number of samples |D| to discretize the cell radius.
p Given between two distances d 2 D, the better the approximation.
|D|, we choose the distance values di = l0 ⇥ i/|D|, for The number of samples used (i.e. |D|) is a key parameter
i = 1, ..., |D|, so as to define equal-area based limits, so that for the model complexity. Increasing |D| drastically increases
the distance steps are closer from each other far from the the number of binary variables (|SF|(1 + |D|) more variables
gateway, for the large SFs. each time |D| increases by 1), and therefore the resolution
By balancing the impact of attenuation and collisions, our time of the model, as we can see on Figure 5.
optimization allows to achieve similar level of losses for all However, as shown on Figure 6, we quickly converge to
nodes, as can be seen on the solid black curves in Figures 2, 3 an accurate minimum PDR since the objective value does not
and 4. The SF boundaries optimizing fairness are also pre- increase much with the number of samples |D|. The difference
sented in Table III. In the small and medium cell scenario, of the minimum PDR achievable for all nodes in the network
in which losses due to collisions have more weight, there is between 50 and 300 samples is less than 1%. It allows us to
a dramatic improvement between the SNR-based and the fair compute very accurate SF boundaries in less than 100 seconds
SF boundaries. The minimum PDR goes from 0.21 to 63.6% of resolution time, to balance the PDR between the nodes and
in the small cell, and from 8.63 to 60.73% in the medium cell. ensure the best performances for the nodes independently of
This allows hundreds of nodes to reach satisfactory loss rates. their distance to the gateway.
In the large cell, the gain is less important but still represents Also, we note that increasing the number of samples does
13 percentage points for the most distant nodes (going from not always improve the objective value of the minimum
42 to 55.64%). achievable PDR in the network. Our optimization model finds
Interestingly, very few nodes use SF12 after optimization, the best combination of SF boundaries within a given sets of
and the range of SF7 always increases in comparison to the distances, and it can give slightly lower values of PDR even
SNR-based boundaries. The smaller the cell, the larger the SF7 though a larger set of distances is used. Two sets of distances
range, but consequently the larger the PDR loss of the nodes D and D0 with |D| < |D0 | can be completely disjoint, due to
closed to the gateway. This negative effect is counterbalanced the way samples are computed (presented in the beginning of
by the gain of distant nodes: in every scenario at least 50% Section IV).
of the nodes have a PDR greater or equal as in the SNR-
based case. The fact that as many nodes as possible should B. Scalability
use smaller SFs corroborates previous findings [12], although, We run our model with different node densities to analyze
as soon as channel attenuation is different between nodes, all how the SF boundaries and PDR evolve with the number of
SFs need to be used. nodes in the cell. We set the number of samples |D| to 100
 100
80
60
PDR
40
20
1600 nodes
1000 nodes
500 nodes

0
0 1000 2000 3000 4000 5000
Distance to gateway (m)

Figure 6: Minimum value of the PDR in function of the Figure 8: Medium cell range (5km). Evolution of the fair SF
number of samples of the cell distances |D|. boundaries and associated PDR in function of the distance to
the gateway.
100

100
80

80
60
PDR

60
PDR
40

40
20

4000 nodes

20
2000 nodes
1000 nodes 400 nodes
0

200 nodes
0 500 1000 1500 2000 2500
100 nodes
0

Distance to gateway (m)

0 1000 2000 3000 4000 5000 6000 7000
Distance to gateway (m)
Figure 7: Small cell range (2.5km). Evolution of the fair SF
boundaries and associated PDR in function of the distance to Figure 9: Large cell range (7km). Evolution of the fair SF
the gateway. boundaries and associated PDR in function of the distance to
the gateway.

since we show that the solutions are accurate and obtained

rapidly. and lifetime. The studies define analytical models for the
First, we analyze the evolution of the fair SF boundaries network capacity [1]–[6] evaluating the number of supported
with various node densities in the cell. For the small cell, the nodes in a LoRaWAN cell. In many recent papers focusing
boundaries do not vary much with the node density (Figure 7). on the LoRaWAN radio capacity, the assumption is that all
For the medium and large cases (Figures 8 and 9), the small nodes use up their allowed duty cycle regardless of their SF.
SF ranges increase: the greater the node density, the more area Practically, this means that they send much more packets
is covered with lower SFs. Incidentally, it reveals that SF12 when closer to the gateway, which is not realistic. In this
does not help much when the node density is high, since the case, when the capacity is expressed in terms of throughput,
number of collisions steeply increases with traffic at this data this assumption may be optimistic as it means being able to
rate. fill-up all available quasi-orthogonal subchannels [2]. On the
We then evaluate the evolution of the minimum achiev- contrary, if the capacity is expressed as a number of nodes [1],
able PDR in function of the number of nodes in the cell [6], [13], it gives dramatically small values because then,
in Figure 10. We show that our SF boundaries allocation moving a node to a smaller SF brings very little benefit, since
preserves an acceptable rate when the node density increases it immediately transmits much more traffic and thus creates
in comparison to the SNR-based SF allocation. Indeed, the similar levels of congestion.
PDR quickly drops to 0% with the SNR-based SF boundaries Fairness in LoRa networks has already been investigated
because the SF allocation does not take into account collisions so far by properly allocating the physical resources such as
which increase with the node density. Our policy balances the bandwidth (BW), data rates (DR), power control (PC), and
effect of collisions and ensures a minimum PDR of 60% with spreading factors (SF). Abdelfadeel et al. [12] propose a trans-
up to 4500 nodes in the small cell, 1600 nodes in the medium mission power control algorithm balancing the received signal
cell, and 260 nodes in the large cell. powers at the gateway and ensuring a fair collision probability
among the nodes, without taking into account Rayleigh fading,
V. R ELATED WORK so that they do not consider the necessary tradeoff between
Recent research on LoRa has mainly focused on perfor- fading and collisions losses. Reynders et al. also investigated
mance evaluation in terms of coverage, capacity, scalability fairness among packet error rate [14] and derived a min max
 It allows to balance the opposite effects of attenuation and
collisions.
However, we can see that the solution might still be
improved, in particular because the PDR for some SF is
significantly higher than its surrounding ones, for instance
SF11 in Figure 2. A careful adjustment of the boundaries
around these specific SF might lead to an even better node
repartition, while still guaranteeing the same minimum PDR.
Finally, we leave as future work the extension of the
model to networks with several gateways and macro-diversity
reception.
Figure 10: Minimum value of the PDR in function of the
R EFERENCES
number of nodes in the cell (|D| = 100).
[1] O. Georgiou and U. Raza, “Low power wide area network analysis: Can
LoRa scale?” IEEE Wireless Communications Letters, vol. 6, no. 2, pp.
optimization model based on log distance path loss model. 162–165, Apr. 2017.
[2] Z. Li, S. Zozor, J.-M. Brossier, N. Varsier, and Q. Lampin, “2D
Amichi et al. [9] consider SF imperfect orthogonality in their time-frequency interference modelling using stochastic geometry for
optimization problem to maximize the minimum achievable performance evaluation in Low-Power Wide-Area Networks,” in 2017
rate in the network. They propose an SF allocation algorithm IEEE International Conference on Communications (ICC), May 2017.
[3] A. Waret, M. Kaneko, A. Guitton, and N. El Rachkidy, “LoRa through-
based on matching theory to fairly allocate the SF. In their put analysis with imperfect spreading factor orthogonality,” IEEE Wire-
model, meeting the capture condition with probability Q in less Communications Letters, vol. 8, no. 2, pp. 408–411, 2019.
Equation (4) implies successful reception, whereas fading [4] J. Petäjäjärvi, K. Mikhaylov, M. Pettissalo, J. Janhunen, and J. Iinatti,
“Performance of a low-power wide-area network based on LoRa tech-
should still be taken into account. nology: Doppler robustness, scalability, and coverage,” International
Experimental measurements by Mikhaylov et al. [15] show Journal of Distributed Sensor Networks, vol. 13, no. 3, 2017.
that physical capture in presence of co-SF interference may [5] K. Mikhaylov, J. Petaejaejaervi, and T. Haenninen, “Analysis of Ca-
pacity and Scalability of the LoRa Low Power Wide Area Network
happen well below 6 dB of power margin. Compared to our Technology,” in 22th European Wireless Conference, May 2016.
model, this would tend to increase the actual packet delivery [6] A. Mahmood, E. Sisinni, L. Guntupalli, R. Rondon, S. A. Hassan, and
ratio. The same paper also questions the theoretical levels of M. Gidlund, “Scalability analysis of a LoRa network under imperfect
orthogonality,” IEEE Transactions on Industrial Informatics, 2018.
inter-SF interference: in particular, low SFs may be jammed [7] A. Duda and M. Heusse, “Spatial issues in modeling LoRaWAN
by other signals even when the power margin is of the order capacity,” in 22nd ACM International Conference on Modeling, Analysis
of -6 dB, whereas the expected values would be at least 10 and Simulation of Wireless and Mobile Systems (MSWIM), Nov. 2019.
[8] C. Caillouet, M. Heusse, and F. Rousseau, “Optimal SF Allocation in
dB lower [11]. This is not overly critical: first, interference LoRaWAN Considering Physical Capture and Imperfect Orthogonality,”
from higher SFs comes from nodes located farther away, so in IEEE Global Communications Conference (GLOBECOM), Dec. 2019.
they face greater attenuation and are thus relatively harmless. [9] L. Amichi, M. Kaneko, N. E. Rachkidy, and A. Guitton, “Spreading
factor allocation strategy for LoRa networks under imperfect orthog-
Secondly, for the fair SF allocation, channel utilization is onality,” in 2019 IEEE International Conference on Communications
much reduced for higher SFs, so that nodes within the lower (ICC), May 2019.
SF annuli face lesser levels of accumulated interference from [10] D. Zorbas, G. Z. Papadopoulos, P. Maillé, N. Montavont, and
C. Douligeris, “Improving LoRa network capacity using multiple spread-
farther nodes compared to distance-based or even SNR-based ing factor configurations,” in 25th International Conference on Telecom-
SF allocation for instance. munications, ICT, 2018, pp. 516–520.
Our Aloha model captures the fact that collisions happen [11] C. Goursaud and J. M. Gorce, “Dedicated networks for IoT: PHY /
MAC state of the art and challenges,” EAI Endorsed Transactions on
when other frames overlap within 2⌧j around a transmission Internet of Things, vol. 1, no. 1, Oct. 2015.
event, which is not the case in Georgiou’s model [1], [13] [12] K. Abdelfadeel, V. Cionca, and D. Pesch, “Fair adaptive data rate
(which would correspond to slotted Aloha or to continuous allocation and power control in LoRaWAN,” in World of Wireless,
Mobile and Multimedia networks (WoWMoM), Feb. 2018.
transmissions as in CDMA networks). [13] A. Hoeller, R. D. Souza, O. L. Alcaraz López, H. Alves,
M. de Noronha Neto, and G. Brante, “Analysis and Performance
VI. C ONCLUSION Optimization of LoRa Networks With Time and Antenna Diversity,”
In this paper, we present an optimization framework to IEEE Access, vol. 6, 2018.
[14] B. Reynders, W. Meert, and S. Pollin, “Power and spreading factor
fairly determine the SF boundaries of a network cell in control in low power wide area networks,” in 2017 IEEE International
order to guarantee the largest packet delivery ratio to all Conference on Communications (ICC), May 2017.
nodes, regardless of their distance to the gateway. We base [15] K. Mikhaylov, J. Petäjäjärvi, and J. Janhunen, “On LoRaWAN scalabil-
ity: Empirical evaluation of susceptibility to inter-network interference,”
our framework on an accurate system model for collisions in European Conference on Networks and Communications (EuCNC),
and physical capture, and with the same traffic intensity to June 2017.
all nodes. Results show the effectiveness of our approach.