Mathematical Model of LoRaWAN Channel Access

Dmitry Bankov, Evgeny Khorov and Andrey Lyakhov

Institute for Information Transmission Problems, Russian Academy of Sciences, Moscow, Russia
Email: {bankov, khorov, lyakhov}@iitp.ru

Abstract—While 3GPP has been developing NB-IoT, the mar- rate mathematical model which takes into account LoRaWAN
ket of Low Power Wide Area Networks has been mastered peculiarities related to retransmission policy.
by cheap and simple Sigfox and LoRa/LoRaWAN technologies.
Being positioned as having an open standard, LoRaWAN has II. L O R AWAN C HANNEL ACCESS D ESCRIPTION
attracted also much interest from the research community.
Specifically, many papers address the efficiency of its PHY A typical LoRaWAN [9] network consists of end devices,
layer. However MAC is still underinvestigated. Existing studies called motes, gateways (GWs), and a server. Motes are con-
of LoRaWAN do not take into account the acknowledgement nected to the GWs via wireless LoRa links. Gateways gather
and retransmission policy, which may lead to incorrect results. information from the motes, send it to the server via an IP
In this paper, we carefully take into account the peculiarities of
LoRaWAN transmission retries and show that it is the weakest network, and forward packets from the server to the motes.
issue of this technology, which significantly increases failure LoRaWAN devices operate in different ways. Depending
probability for retries. The main contribution of the paper is on operation, the standard describes three classes of devices.
a mathematical model which accurately estimates how packet The basic functionality for sporadic uplink data transmission
error rate depends on the offered load. In contrast to other is described as class A operation and is studied in this paper.
papers, which evaluate LoRaWAN capacity just as the maximal
throughput, our model can be used to find the maximal load, A LoRaWAN network simultaneously works in several
which allows reliable packet delivery. wireless channels. For example, in Europe they can use three
Index Terms—LoRa, LoRaWAN, LPWAN, Channel Access, main channels and one downlink channel. To transmit a data
Performance Evaluation, ALOHA frame, each mote randomly selects one of the main channels
(see Fig. 1). Having received the frame, the GW sends two
I. I NTRODUCTION ACKs. The first one is sent in the main channel, where the
frame was received, 𝑇1 after frame reception. The second ACK
LoRaWAN is a relatively new protocol designed to provide is sent in the downlink channel after timeout 𝑇2 = 𝑇1 + 1 s.
cheap and reliable wireless connectivity in various Internet If a mote receives no ACK, it makes a retransmission. The
of Things scenarios. Being a Low Power Wide Area Network standard recommends making a retransmission in a random
technology operating in the ISM band, it rapidly got popularity time drawn from [1, 1 + 𝑊 ] seconds, where 𝑊 = 2. Note
in both industry and academic communities. Literature review that the recommended 𝑊 is too small and, as we show in the
shows that in spite of numerous studies of its PHY layer [1]– paper leads to the “avalanche effect”.
[3], the MAC layer got little attention, even though it has At the PHY layer, LoRaWAN uses Chirp Spread Spectrum
multiple issues [4], [5] that limit its performance. However, as modulation. Its main feature is that signals with different
LoRaWAN is designed to support networks of thousands of spreading factors can be distinguished and received simultane-
devices, it is crucial not only to consider the performance of ously, even if they are transmitted in the same time on the same
this technology in point-to-point scenarios, but also to evaluate channel. Spreading factor, together with the channel width and
its applicability in case of highly-populated networks. the coding rate, determines the data rate. Lower data rates
To calculate throughput of LoRaWAN networks, in existing extend transmission range and improve transmission reliability.
studies of the MAC layer (e.g., see [6]), the authors typically For the first transmission attempt, the rate is determined by
use the classical approach for modeling ALOHA networks [7]. the GW. The standard also recommends decrementing data
The papers (e.g. [8]) also limit the study to unacknowledged rate every two consequent transmission failures, limiting the
mode, which has no control acknowledgements (ACKs). Thus, number of retransmissions by 𝑅𝐿 = 7. The first ACK is sent
with no control traffic the throughput increases. However the at a data rate that is lower than the data rate for the frame
reliability of transmission decreases. transmission by a configurable offset (it can be zero). The
In this paper, we provide a mathematical model for a second ACK should always be sent at a fixed data rate, by
LoRaWAN network operating in the acknowledged mode. We default the lowest one.
explain why the usage of classical ALOHA-like approach
underestimates the collision probability and develop an accu- III. P ROBLEM S TATEMENT
Consider a LoRaWAN network that consists of a GW and
The reported study was partially supported by RFBR, research project No. 𝑁 motes and operates in 𝐹 main channels and one downlink
15-37-70004 mol a mos. channel. The motes use data rates 0, 1, ..., 𝑅, set by the GW.
978-1-5386-2723-5/17/$31.00 ○2017
c IEEE Let 𝑝𝑖 be the probability that a mote uses data rate 𝑖.
 Main 𝑇1 to the beginning of the considered frame. For a Poisson
channel process of frame generation, such an event happens with
Mote: Data 𝐷𝑎𝑡𝑎
probability 𝑒−2𝑟𝑖 𝑇𝑖 . We consider that the GW cancels ACK
Downlink GW: ACK1 𝑡
transmission if it is receiving a data frame, so a collision can
channel GW: ACK2 happen only if the ACK is generated in the interval [−𝑇𝑖𝐴𝑐𝑘 , 0].
𝑇2 𝑡 The rate of ACK generation is 𝑃𝑖𝐷𝑎𝑡𝑎 𝑟𝑖 , so the probability to
𝐷𝑎𝑡𝑎 𝐴𝑐𝑘

Fig. 1: LoRaWAN channel access

avoid collision with an ACK is 𝑒−𝑟𝑖 𝑃𝑖 𝑇𝑖
. Finally, 𝑃𝑖𝐷𝑎𝑡𝑎
can be found from the following equation:
𝐷𝑎𝑡𝑎
+𝑃𝑖𝐷𝑎𝑡𝑎 𝑇𝑖𝐴𝑐𝑘 )𝑟𝑖
We consider that a frame collision occurs when two frames 𝑃𝑖𝐷𝑎𝑡𝑎 = 𝑒−(2𝑇𝑖 .
are transmitted in the same channel at the same data rate, and
they intersect in time. As for ACKs, the probability that at least one ACK arrives
The motes generate frames according to a Poisson process is calculated according to the inclusion-exclusion principle:
with total intensity 𝜆 (the network load). All motes transmit 𝑃𝑖𝐴𝑐𝑘 = 𝑃𝑖𝐴𝑐𝑘1 + 𝑃𝑖𝐴𝑐𝑘2 − 𝑃𝑖𝐴𝑐𝑘1 𝑃𝑖𝐴𝑐𝑘2 ,
frames with 51-byte Frame Payload which corresponds to the
biggest payload that can fit a f rame a t t he l owest d ata rate. where 𝑃𝑖𝐴𝑐𝑘1 and 𝑃𝑖𝐴𝑐𝑘2 are the probabilities that the first
The frames are transmitted in the acknowledged mode, and and the second ACK, respectively, is transmitted successfully,
ACKs carry no frame payload. We consider a situation, when provided that data was transmitted at rate 𝑖. The first ACK is
motes have no queue, i.e. if two messages are generated, a transmitted successfully if no data frame intersects it:
𝐷𝑎𝑡𝑎
mote transmits the most recent one. 𝑃𝑖𝐴𝑐𝑘1 = 𝑒−(min(𝑇1 ,𝑇𝑖 )+𝑇𝑖𝐴𝑐𝑘 )𝑟𝑖 .
For the described scenario, it is important not only to know
the nominal channel capacity, but also to find the maximal load Here we take the minimum of 𝑇𝑖𝐷𝑎𝑡𝑎 and 𝑇1 , because if a
at which the network can provide reliable communications. In frame exceeds 𝑇1 , it breaks the acknowledged frame, but such
other words, we need to find t he p acket e rror r ate ( PER) as an event is already taken into account by 𝑃𝑖𝐷𝑎𝑡𝑎 . The second
a function of network load 𝜆. ACK is transmitted successfully if no data frame is successful
in any other channel or at any other data rate, such that its
IV. M ATHEMATICAL M ODEL second ACK would intersect the considered one:
𝑝𝑖 ∑︀𝑅
To solve the problem, we develop a mathematical model of 𝐴𝑐𝑘 𝐷𝑎𝑡𝑎
𝑃 𝐴𝑐𝑘2 = 𝑒−𝑇0 𝜆(1− 𝐹 ) 𝑗=0 𝑃𝑗 𝑝𝑗 .
𝑖
the transmission process. As the first t ransmission attempts
are described by the Poisson process, to find t he P ER in B. Retransmissions
these assumptions, in Section IV-A, we consider the approach Consider a case, when two motes transmit frames with
used to evaluate ALOHA networks [7] and extended to take collision, as shown in Fig. 2. Let 0 be the time when the
into account ACKs. This approach is however inapplicable for frame of mote A begins, and 𝑥 be the offset for frame of
retransmissions, because they do not form a Poisson process, mote B. Motes choose a channel for retransmission randomly.
so in Section IV-B we propose another way to take them into If they choose different channels, the collision is resolved.
account and thus to improve the accuracy of the model. Otherwise, with probability 𝐹1 , they choose the same channel.
In this case, let 𝑦 and 𝑧 be the times when motes A and B start
A. The First Transmission Attempt their retransmission, respectively. The value of 𝑦 is distributed
The first transmission attempt is successful with probability uniformly in the interval [𝜏, 𝜏 + 𝑊 ], where 𝜏 is the frame
𝑅 duration 𝑇 plus the timeout for the ACK. The value of 𝑧 is
distributed uniformly in the interval [𝜏 + 𝑥, 𝜏 + 𝑥 + 𝑊 ]. The
∑︁
𝑃𝑆,1 = 𝑝𝑖 𝑃𝑖𝐷𝑎𝑡𝑎 𝑃𝑖𝐴𝑐𝑘 , (1)
𝑖=0 retransmission results in a new collision, if [𝑧, 𝑧 +𝑇 ] intersects
with [𝑦, 𝑦 + 𝑇 ], which happens with the probability
where 𝑃𝑖𝐷𝑎𝑡𝑎 is the probability that the data frame is transmit-
ted without collision at data rate 𝑖 and 𝑃𝑖𝐴𝑐𝑘 is the probability ∫︀𝑇 ∫︀𝑊 𝑊∫︀+𝑥 1(𝑦≤𝑧≤𝑦+𝑇 )+1(𝑧≤𝑦≤𝑧+𝑇 )
𝑟𝑖 𝑒−𝑟𝑖 𝑥 𝑊2 𝑑𝑧𝑑𝑦𝑑𝑥
that at least one ACK out of two is received by the mote, 0 0 𝑥
provided that the data frame is successful. 𝑃𝑥 = =
∫︀𝑇
Since the packets transmitted in different channels and at 𝑟𝑖 𝑒−𝑟𝑖 𝑥 𝑑𝑥
0
different rates do not collide, we need to consider separately (︃ )︃
each combination of channel and data rate. Specifically for 𝑇 3 2 1
= 2 2𝑊 − 𝑇 − + ,
rate 𝑖 and one of 𝐹 channels, the load equals 𝑟𝑖 = 𝜆𝑝 𝑖
𝐹 .
𝑊 2 2
𝑇 𝑟𝑖 𝑟𝑖 tanh( 𝑟𝑖2𝑇 )
A data frame transmission is successful if it intersects with
where 1(𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛) is the indicator function which equals 1
no transmission of another frame or an ACK sent by the
if 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛 is true and 0 otherwise.
GW as a response to previous frame. Let 𝑇𝑖𝐷𝑎𝑡𝑎 and 𝑇𝑖𝐴𝑐𝑘
Motes have the same probability of being the first and the
be the durations of a data frame and an ACK, respectively,
second one, so the probability that there is no collision equals
at rate 𝑖. Intersection with a frame does not occur if no
𝐷𝑎𝑡𝑎
frames are generated in the interval [−𝑇𝑖𝐷𝑎𝑡𝑎 , 𝑇𝑖𝐷𝑎𝑡𝑎 ], relative 𝑃𝑖,𝑅𝑒 = 1 − 2𝑃𝑥 /𝐹.
 frame of mote A frame of mote B 100

... 10-1
0 𝑥 𝑇 𝜏 𝑦 𝑧 𝜏 +𝑊 𝑡

10-2 PER, sim

PER
Fig. 2: Retransmission
PER, math
The average probability of a successful transmission 𝑃𝑆 is PER1 , sim
10-3
PER1 , math
𝑃𝑆 = 𝑃1 𝑃𝑆,1 + (1 − 𝑃1 )𝑃𝑆,𝑅𝑒 , λ∗
10-4 -3
where 𝑃𝑆,𝑅𝑒 is the probability of a successful retransmission, 10 10-2 10-1 100 101
𝐷𝑎𝑡𝑎
calculated as in eq. (1), using 𝑃𝑖,𝑅𝑒 instead of 𝑃𝑖𝐷𝑎𝑡𝑎 , and Network load, frames per second
𝑃1 is the probability that the transmission is the first one (not
Fig. 3: Dependency of PER on the network load
a retry). 𝑃1 is reverse to the average number of transmission
attempts per a frame:
(︃ 𝑅𝐿
)︃−1 because of a poor retransmission policy the PER rapidly tends
𝑃1 = 1 + (1 − 𝑃𝑆,1 )
∑︁ 𝑟 𝑟+1
(1 − 𝑃𝑆,𝑅𝑒 ) 𝑃𝑁 , to 1, when the load exceeds 10−1 packets per second.
𝑟=0 VI. C ONCLUSION
(𝑇𝑖𝐷𝑎𝑡𝑎 +𝑇2 +𝑇0𝐴𝑐𝑘 +⟨𝑇𝑤𝑎𝑖𝑡 ⟩)
∑︀𝑅 𝜆
−𝑁
where 𝑃𝑁 = 𝑖=0 𝑝𝑖 𝑒 is the In the paper, we develop the first a ccurate mathematical
probability that a new frame does not arrive during the model of acknowledged uplink transmissions in LoRaWAN
transmission and ⟨𝑇𝑤𝑎𝑖𝑡 ⟩ = 1 + 𝑊/2 is the average interval networks with class A devices. We have shown that leaving
that a mote waits before a retransmission. The packet error out of consideration retransmission process significantly over-
rate is calculated as 𝑃 𝐸𝑅 = 1 − 𝑃𝑆 . estimates efficiency o f a L oRaWAN n etwork. I n c ontrast, our
The model estimates PER correctly up to such network load, model takes into account peculiarities of the retransmission
that new frames arrive at the motes as quickly as the motes process and correctly estimates packet error rate when the
drop the frames due to inability to resolve collisions after 𝑅𝐿 load is lower than some threshold 𝜆* , which is found in
retransmission attempts. It means that the load equals the paper. However the area with the higher loads is not
(︃ 𝑅 )︃−1 interesting from a practical point of view. Indeed, after the
∑︁ (︀ load exceeds the described threshold, PER rapidly grows to
𝜆* = 𝐹 𝑝𝑖 𝑇𝑖𝐷𝑎𝑡𝑎 + 𝑇2 + 𝑇0𝐴𝑐𝑘 + ⟨𝑇𝑤𝑎𝑖𝑡 ⟩ 𝑅𝐿
)︀
.
1 because retransmissions form an “avalanche”. Thus in this
𝑖=0
area LoRaWAN cannot provide reliable communications.
V. N UMERICAL R ESULTS
Let us use the developed model to evaluate performance of a R EFERENCES
LoRaWAN network. As in [6], we consider a scenario, when [1] M. Centenaro, L. Vangelista, A. Zanella, and M. Zorzi, “Long-range
the motes are distributed uniformly in a circular area with Communications in Unlicensed Bands: The Rising Stars in the IoT and
Smart City Scenarios,” IEEE Wireless Communications, vol. 23, no. 5,
radius of 1 km around the GW, and the path-loss is described pp. 60–67, 2016.
by Okumura-Hata model for urban environment. We consider [2] L. Vangelista, A. Zanella, and M. Zorzi, “Long-Range IoT Technologies:
EU 863-880 MHz ISM band. In this case, the data rates are The Dawn of LoRaTM ,” in Future Access Enablers of Ubiquitous and
Intelligent Infrastructures, pp. 51–58, Springer, 2015.
distributed as follows: 𝑝0 = 0.28, 𝑝1 = 0.2, 𝑝2 = 0.14, 𝑝3 = [3] C. Goursaud and J.-M. Gorce, “Dedicated networks for IoT: PHY/MAC
0.1, 𝑝4 = 0.08, 𝑝5 = 0.2. We simulate a network with 1000 state of the art and challenges,” EAI endorsed transactions on Internet of
motes and compare the average PER and PER1 for the first Things, 2015.
[4] D. Bankov, E. Khorov, and A. Lyakhov, “On the Limits of LoRaWAN
transmission attempt with those obtained with the developed Channel Access,” in Engineering and Telecommunication (EnT), 2016
mathematical model. The results are shown in Fig. 3. Because International Conference on, pp. 10–14, IEEE, 2016.
of inefficient retransmission parameters the real PER is by [5] K. Mikhaylov, J. Petäjäjärvi, and T. Hänninen, “Analysis of the Capacity
and Scalability of the LoRa Wide Area Network Technology,” in Pro-
50% greater than PER1 . Thus, by taking into account re- ceedings of 2016 22nd European Wireless Conference, VDE, May 2016.
transmissions, we have significantly improved the accuracy [6] F. Adelantado, X. Vilajosana, P. Tuset, B. Martinez, J. MELIA-SEGUI,
of the model. From Fig. 3 we also see that we correctly and T. Watteyne, “Understanding the Limits of LoRaWAN,” IEEE Com-
munications Magazine, June 2017.
estimate 𝜆* which is the highest load when we can neglect [7] N. Abramson, “THE ALOHA SYSTEM: Another Alternative for Com-
high-order collisions and the “avalanche effect” inherent to puter Communications,” in Proceedings of the November 17-19, 1970,
the default retransmission parameters. Non-adaptive and small Fall Joint Computer Conference, AFIPS ’70 (Fall), (New York, NY,
USA), pp. 281–285, ACM, 1970.
retransmission window does not allow to resolve collisions [8] A. Augustin, J. Yi, T. Clausen, and W. M. Townsley, “A Study of LoRa:
with high number of packets, and involving new motes in Long Range & Low Power Networks for the Internet of Things,” Sensors,
collisions is faster than packet dropping or collision resolution. vol. 16, no. 9, p. 1466, 2016.
[9] LoRa Alliance, LoRaWAN Specification, 2015.
This significantly limits the capacity of a LoRaWAN network.
While the network can transmit several packets per second,