IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO.

1, JANUARY 2020 763

Distributed Queueing-Based Random Access

Protocol for LoRa Networks
Wei Wu , Yan Li, Yanhe Zhang, Bin Wang, and Wennai Wang

Abstract—Among many others, LoRa/LoRaWAN is one of the layer over the LoRa PHY layer is based on pure Aloha fash-
most widely employed technologies to support Internet of Things ion, which is specified by LoRaWAN. However, the Aloha’s
applications for low-power and long-range wireless devices. scalability issue is notorious, i.e., it suffers from congestion
LoRaWAN is a medium access control (MAC) layer protocol
based on the traditional Aloha algorithm which reveals a great under high traffic load.
degraded throughput when a large number of devices attempt to An alternative solution to Aloha is distributed queue-
communicate over the shared channel at the same time. Aiming ing (DQ), which ensures high throughput regardless of the
at this scalability issue, we propose to introduce the distributed traffic load and traffic pattern [3]. DQ was first intro-
queueing (DQ) algorithm into the MAC protocol over the LoRa duced by Xu and Campbell as a novel MAC protocol for
physical layer, named DQ-LoRa. The details of DQ-LoRa are pro-
vided about the frame structure and access procedure, followed cable TV distribution [4]. The stability of its performance
by the performance evaluations on system throughput, average and near-optimum behavior in terms of channel utilization,
delay, and energy consumption. Our analytical results show that access delay, and energy consumption have been demonstrated
the maximum throughput of DQ-LoRa is independent of the in [5] and [6].
number of devices, and the delay is insensitive to the number With the motivation to eliminate the limitations of LoRa
of random access minislots for device contending. The numerical
computation shows the gain of throughput up to 2.6-fold, and the networks resulting from LoRaWAN Aloha MAC, in this arti-
savings of energy and latency up to 48% and 54%, respectively, cle, we introduce the DQ algorithm into the MAC layer
in comparing DQ-LoRa with a pure Aloha system. over LoRa, named DQ-LoRa. We believe that DQ-LoRa
Index Terms—Distributed queueing (DQ), LoRa, medium substituted for the Aloha-based MAC layer of commer-
access control (MAC), performance analysis. cial LoRaWAN devices is not so difficult, since most
LoRaWAN MAC functionalities are implemented in soft-
ware [7]. In addition to that our proposal is designed
I. I NTRODUCTION to be compatible to Class A by utilizing beacon signals
defined for Class B in traditional LoRaWAN [8]. Therefore,
OW-POWER wide-area networking (LPWAN) enables
L power efficient wireless communications over very long
distances, which offers a significant potential to support a large
our proposal requires LoRa end-devices (EDs) that can
switch between Classes A and B modes according to the
beacon signal.
number of Internet of Things (IoT) applications. LoRa, a phys- The following are our main contributions.
ical (PHY) layer modulation technique based on the chirp 1) DQ-LoRa is proposed to address the scalability issue
spread spectrum (CCS), is one of the most successful and of LoRaWAN, as a candidate for IoT applications with
adopted LPWAN technologies [1], [2]. Since battery powered unpredictable number of connected devices and bursty
nodes in LoRa networks share the unlicensed ISM band and traffic. Its frame structure as well as the access procedure
require a long lifetime, an effective medium access control are provided in detail.
(MAC) mechanism is critical for LoRa. A widely used MAC 2) The performance of DQ-LoRa is analyzed theoretically,
in terms of the system throughput, average delay, and
Manuscript received May 28, 2019; revised September 4, 2019; accepted
September 24, 2019. Date of publication October 3, 2019; date of current ver- average energy consumption. It is also compared with
sion January 10, 2020. This work was supported in part by the Postgraduate pure Aloha and CSMA.
Research and Practice Innovation Programm of Jiangsu Province under 3) The results demonstrate that with a large number of
Grant KYCX18_0897 and Grant SJCX18_0278. (Corresponding author:
Wennai Wang.) active devices, DQ-LoRa outperforms pure Aloha in
W. Wu, B. Wang, and W. Wang are with the Key Laboratory all the aspects of system performance. The through-
of Broadband Wireless Communication and Sensor Network put gain can be up to 2.6 times, as well as the
Technology, Ministry of Education, Nanjing University of Posts
and Telecommunications, Nanjing 210003, China, and also with the delay and energy savings up to 54% and 48%,
College of Telecommunication and Information Engineering, Nanjing respectively.
University of Posts and Telecommunications, Nanjing 210003, China The rest of this article is organized as follows. Related
(e-mail: [email protected]; [email protected]; wangwn@
njupt.edu.cn). works are discussed in Section II. In Section III, an overview
Y. Li and Y. Zhang are with the College of Telecommunication of LoRa and LoRaWAN is provided. Section IV is devoted
and Information Engineering, Nanjing University of Posts and to describing our design of DQ-LoRa in detail. Section V
Telecommunications, Nanjing 210003, China (e-mail: [email protected];
[email protected]). presents the performance analysis of DQ-LoRa. The conclu-
Digital Object Identifier 10.1109/JIOT.2019.2945327 sions are made in Section VI.
2327-4662 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 764 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 1, JANUARY 2020

II. R ELATED W ORKS TR is a heavy overhead, for example, its size is up to 26 sym-
A. Medium Access Control Protocols Over LoRa bols while the size of data packet is only 50 symbols (in LoRa
packet radios with payload size of 20 bytes). Additionally, in
Most of the existing researches focus on LoRaWAN Aloha
this protocol, a downlink transmission must be requested by
MAC and its throughput performance is analyzed in [1],
an uplink receive request, which increases the probability of
and [9]–[11]. In general, at a very low transmission rate the
collisions in the access procedure. The synchronization mech-
throughput of LoRa networks is impacted by packets colli-
anism of DQ-N is also absent, but strict timing mechanism is
sions, whereas at higher transmission rates it is impacted by
important since several events occur within a frame.
duty cycling [1]. Additionally, it has been shown that although
LoRaWAN uses Aloha, due to LoRa’s robust modulation tech-
nique only 32% more packet losses in the case of 1000 nodes III. L O R A /L O R AWAN OVERVIEW
per gateway (GW), whereas for the same case the losses are
A. LoRa Physical Layer
up to 90% in other pure Aloha-based networks [10]. A com-
prehensive model of LoRaWAN LPWANs in the ns-3 network LoRa is a PHY layer modulation technique, which makes
simulator is presented in [11], which forms the basis for a scal- the communication robust against noise and interference with
ability analysis of LoRaWAN. The simulation results show that the usage of CCS and forward error correction (FEC). It pro-
downstream traffic has a negative impact on the delivery ratio vides several customizable parameters, including spreading
of confirmed upstream traffic and increasing GW density can factor (SF), BW, code rate (CR), and transmission power (TP).
ameliorate but not to eliminate this effect. For each LoRa transmission, the available bit rate, resilience
In addition to that a few studies analyze the through- against interference, and the ease of decoding can be changed
put of LoRa networks utilizing other MAC protocols. by tuning these parameters. Three typical BWs (125-, 250-,
Farooq and Pesch [2] modified the Aloha-based mechanism by and 500 KHz) in the unlicensed frequency bands (433-, 868-,
using a simple systematic approach. They found that this mod- or 915 MHz) are available, as well as six different SFs (from
ification positively impacts the performance of the different 7 to 12) and four different CRs (4/5, 4/6, 4/7, and 4/8). The
communication settings. Besides, a range of MAC proto- higher SF implies lower bit rate but longer communication dis-
cols have been analyzed in order to search for a suitable tance, while the higher CR indicates higher protection against
MAC protocol for LoRa networks [12], namely, pure Aloha, interference and vice versa.
delay before transmit, random frequency hopping, and car- LoRa PHY packet structure includes a preamble, an optional
rier sensing multiple access (CSMA). The results demonstrate header, a data payload, and an optional CRC field. The
that CSMA has significant better performance and scalabil- preamble, whose length can be configured from 10 to 65 536
ity features in terms of the number of nodes and data traffic symbols, is used to synchronize the receiver with the trans-
generation models. mitter. The optional header contains the payload length, the
CR used for payload, and the header CRC. The presence of
the header depends on the header mode (explicit or implicit
B. DQ Extensions in Various Networks
mode), and it is disabled in the implicit mode when the pay-
DQ has been adapted to many types of the communica- load length, CR and CRC presence are fixed or known in
tion systems, such as wired centralized networks [13], [14], advance. More details of LoRa can be found in [8].
satellite communications [15], CDMA [16], WLAN [17]–[19],
body area networks [20], wireless ad-hoc networks [21], coop-
erative networks [22], massive machine type communications B. LoRaWAN
in LTE [23], and low-power wireless networks [7], [24]. The LoRaWAN is an MAC layer protocol and network archi-
results of all these works demonstrate that the superiority of tecture to be used with the LoRa PHY layer. As described in
DQ over Aloha-based protocols and its stable and scalable LoRaWAN specification, devices access the channel randomly
performance for densely loaded networks. in a pure Aloha fashion and must obey a strict duty cycle.
Low-power DQ (LPDQ) [24] is based on the low-power Utilizing the Aloha MAC protocol can be justified if to account
listening for network synchronization, DQ for channel access, for the hidden node problem and omitting listen-before-talk
and channel hopping against multipath propagation and exter- phase for energy savings [26].
nal interference. It is not intended for low-rate IoT networks, An LoRaWAN network has a star-of-star topology, in which
but for high-frequency, high-bandwidth (BW) wireless links. communications occur between three kinds of components:
In addition, the synchronization phase is a heavy overhead, 1) ED; 2) GW; and 3) network server (NS). GW transparently
which adversely affects channel utility and latency. relays the messages between EDs and NS. Communication is
DQ-N was proposed for crowdsourcing LPWAN [7] and its bidirectional, yet uplink communication is strongly preferred.
extension for LoRa (a demo) is described in [25]. Its main dif- Note that as GW receives every message from EDs within its
ference from LPDQ is that devices can request multiple data communication range, collisions occur when it receives more
slots in each frame by the transmission request (TR). The 4-bit than one message in the same channel at the same time.
ID field of TR was not explained in [7], and we infer from [25] Three types of EDs are defined by LoRaWAN [8].
that it is used to indicate device ID. However, the 4-bit length 1) Class A: Devices in this class open two short receive
is far from enough since IoT is expected to accommodate bil- windows following each uplink transmission. Downlink
lions of connected devices in the future. On the other hand, transmissions are only allowed after uplink messages.

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 WU et al.: DQ-BASED RANDOM ACCESS PROTOCOL FOR LoRa NETWORKS 765

Class A has the lowest power consumption and can be 2) pRQ Counter (Device Position in the CRQ): If the device
implemented by every device. is waiting in the CRQ, it must decrease the counter
2) Class B: The purpose of this class is to have a device by one at each frame. In addition, if the device has
available for reception at scheduled times (ping slots), in attempted an access in the previous frame and collided,
addition to the reception windows in Class A. It requires it sets the counter to the end of CRQ.
GW to broadcast a beacon (BCN) periodically to syn- 3) TQ Counter (DTQ Length): The counter is increased
chronize all the devices in the network. Class B has the by one for each success state accounted in the previous
medium power consumption. frame. After a DT, the counter is decreased by one.
3) Class C: Devices in this class have nearly continuously 4) pTQ (Device Position in the DTQ): If a device enters the
open receive windows, which can only be closed when DTQ, it sets the counter to the end of DTQ. If the device
transmitting. Hence, Class C has the maximum power is waiting in the DTQ, it must decrease the counter by
consumption. one every frame.
Note that all the EDs start and join the network in Class A, We introduce the DQ algorithm into the MAC protocol over
after that any ED can switch to Class B by searching for a the LoRa PHY layer, named DQ-LoRa. Similar to the opera-
BCN signal. If a network BCN was found, it switches to Class tion of DQ, this protocol requires EDs to contend the shared
B successfully, otherwise, it fails. Once in Class B mode, the channel by sending access requests before transmitting the data
ED should periodically receive BCN and inform its location packets. GW will act as a coordinator to broadcast the feed-
to NS. If no BCN has been received for a given period, the back information, then EDs can distributedly execute the rules
ED cannot synchronize with the network and it switches back of DQ-LoRa and autonomously decide when to transmit.
to Class A. At first, we make some assumptions.
1) Every ED has no more than a single packet to transmit
in one BCN period and they arrive at the beginning of
IV. DQ-BASED MAC P ROTOCOL FOR L O R A
this period.
A. DQ for Contention Resolution 2) Each BCN period is fixed and all the collisions occurring
DQ is an MAC protocol based on a tree-splitting algo- in one BCN period must be resolved before the next
rithm, whose performance is independent of the number of period.
devices sharing a common channel [4]. Different from Aloha, 3) The length of each data slot (packet) is fixed.
DQ mechanism requires active devices (i.e., the devices which 4) There are no transmission errors in the data slot.
have data to be transmitted) to contend in the contention slots 5) The number of contention slots is no less than 2.
before collision-free data transmission (DT). It uses a set of
rules are to be organize each device into one of the two log-
ical queues: 1) collision resolution queue (CRQ) and 2) DT B. Frame Structure
queue (DTQ). If collisions are detected, the devices are split The general frame structure of DQ-LoRa is composed of
into groups and organized into the CRQ, otherwise, they are three parts in the time domain: 1) a contention window (CW)
organized into the DTQ. The devices in the CRQ wait for consisting of m minislots (contention slots) for contention res-
contention resolution in the subsequent contention slots while olution; 2) a data slot for collision-free DT; and 3) a feedback
those in the DTQ wait for collision-free DTs. The splitting slot for downlink feedback packet (FBP) broadcasting. Apart
of the colliding devices reduces the probability of collisions from that a guard time—inter slot space (ISS), is reserved
in the later access procedures since the simultaneous attempts for the compensation of switching time between reception
are decreased. and transmission. It is worth to mention that the data slot is
The queues are distributed in the sense that each device optional and it can be disabled in any frame when the DTQ is
uses internal counters to represent the queue length and its empty. Hereafter, we call the frame which contains a data slot
position within the queue [23]. In order to update the internal “Type-I frame” (Fig. 1, upper side), and the frame without a
counters, the coordinator must provide feedback information, data slot “Type-II frame” (Fig. 1, lower side).
which includes the status of each contention slot as well as In order to accommodate the frame structure to the LoRa
the lengths of CRQ and DTQ. There are three states of each standard, the length of each part can be set as follows.
contention slot: 1) empty (no device contends); 2) success 1) CW Length: In order to significantly lower the overhead
(only one device contends); and 3) collision (two or more and inspired by the small preamble used in LTE [27],
devices contend). Upon receiving the feedback information, we design the access request as a short preamble of 2
the device checks the state of the slot it chooses. If it is symbols, referred to as random access preamble (RAP).
a success, the device enter the DTQ, otherwise, it enters If SF = 12, there exits 224 RAPs that can be allocated
the CRQ. At the same time it updates the internal counters to different devices. Therefore, the length of each min-
as follows [3]. islot must be at least two symbols and the length of
1) RQ Counter (CRQ Length): The counter is increased by CW is 2m symbols. Note that the number of minislots
one for each contention slot with collision state in the m affects system performance. The optimal setting of m
previous frame. At each frame, the counter is decreased in CW depends on the optimization objective (maximum
by one for the resolution attempt of the devices at the throughput or minimum energy consumption) rather than
head of the CRQ. the number of EDs, which will be demonstrated in

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 766 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 1, JANUARY 2020

TABLE I
C OUNTER VALUES

Fig. 1. Frame structures defined in DQ-LoRa. The upper side is Type-I frame
and the lower side is Type-II frame.

Section V. Therefore, a tradeoff between the through-

put and energy consumption must be considered when
setting m.
2) Data Slot Length: The length of the data packet is the
sum of preamble and PHY payload, and the number of
payload symbols can be calculated as [28]
  
8PL − 4SF + 28 + 16CRC − 20IH
max (CR + 4), 0
4(SF − 2DE) the DTQ using a first-in-first-out (FIFO) principle to arrange
where PL and SF are the size of payload (in bytes) their DT.
and the SF, respectively, IH indicates the type of header It is worth to mention that the BCN period is variable since a
mode (1 for implicit mode and 0 for explicit mode), DE new BCN period can start immediately after all active devices
indicates the usage of low data rate optimization (1 when in one BCN period completing their DTs. However, the BCN
used and 0 when disabled), CRC indicates the presence period is not unlimited, which is restricted by a minimum and
of the payload CRC (1 when on and 0 when off), CR a maximum thresholds. If there is no contending device in
is the programmed CR from 1 to 4. Considering a data one BCN period, another period will start after the elapse of
packet with explicit header mode, CR of 4/8, and pay- the minimum threshold. If there is a large number of con-
load size of 20 bytes, the length of the data packet is 50 tending devices and collisions have not been resolved until
symbols (10-symbol preamble and 40-symbol payload). the elapse of the maximum threshold, another BCN period
3) Feedback Slot Length: Each FBP must have 2m bits to will start immediately and the collided devices will recontend
indicate the results of m minislots since each slot has in that period. In order to simplify the performance analy-
three kinds of states. Apart from that the additional 4 sis, we make the assumption that each BCN period is fixed
bytes are required to indicate the length of CRQ and (Section IV-A). The performance analysis under variable BCN
DTQ [6]. We consider FBP with implicit header mode period is left as a future work.
to reduce the overhead and transmission time, then the As we mentioned in the previous section, the length of CW
length of FBP is (which relies on the number of minislots) is independent of
  the number of EDs, but a tradeoff between the throughput
0.25m + 4 − 5 and energy consumption must be considered when setting it.
LFBP = 18 + 8 . (1)
6 Besides, the configuration of minislots can either be adaptable
or determined. If it is adaptable, it can be broadcasted by the
It is known from (1) that LFBP is a step function of m
GW at the beginning of each BCN period, otherwise, it can
and the step height is eight symbols. More intuitively,
be optimized offline and preconfigured in the EDs.
the length of FBP is 18 symbols when m < 5, then it
The access procedure (Fig. 2) is as follows, and the counter
rises up to 26 symbols when 5 ≤ m < 29 and further
values of each device in each step are shown in Table I.
to 34 symbols when m = 29. The great step height will
1) At the beginning of the BCN period, the GW broadcasts
account for the steep drops of system throughput or the
a BCN for synchronization, the devices which receive
steep rises of energy consumption at some m, which will
the BCN switch to or maintain in Class B mode.
be exhibited in Section V.
2) At frame 1, three active EDs (in Class B mode) contend
in the CW. ED1 succeed in the second minislot and enter
C. Access Procedure in the first position of DTQ; ED2 and ED3 collide in
Fig. 2 shows an example to illustrate the access procedure the third minislot and enter in the first position of CRQ.
of DQ-LoRa, where three minislots are configured in the CW Note that there is no data slot in this frame because the
and three active devices arrive in the current BCN period. DTQ is always empty before the first FBP broadcast.
Colliding EDs enter the CRQ using a tree-splitting algo- 3) At frame 2, ED2 and ED3 contend in the CW as they
rithm to resolve the contention, whereas, succeeding EDs enter are in the first position of CRQ. They succeed in the

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 WU et al.: DQ-BASED RANDOM ACCESS PROTOCOL FOR LoRa NETWORKS 767

Fig. 2. Example to illustrate the access procedure of DQ-LoRa. Three minislots are configured in the CW and three active EDs arrive in this BCN period.

TABLE II
third and first minislots, respectively, thus they enter at E VALUATION OF n F ROM N UMERICAL E XPERIMENTS
the end of the DTQ according to the order of minislots.
Then, ED1 transmits its data in the data slot since it is
at the head of the DTQ. After that the GW broadcasts
an FBP.
4) At frame 3, no EDs contend in the CW since the CRQ
is empty. Subsequently, ED3 at the head of the DTQ
transmits its data, and an FBP from the GW follows.
5) At frame 4, the CW is still empty, followed by the DT
of ED2, then the FBP from the GW.
It is worth to mention that the influence of propagation delay
on the collision detection in minislots is negligible The reason
is that LoRa enables the long-range transmissions between
5 and 15 km [7], hence the maximum delay of the edge device
where TBCN , TRAP , and TFBP are the transmission time of
is about (15 × 103 )/(3 × 108 ) = 50 μs, which is much less
BCN, RAP, and FBP, respectively, and n is the number of
than the minislot duration 96 ms (when the bit rate is 250 b/s).
Type-II frames in one BCN period.
1) Simplified Expression for S: Since the access procedure
V. P ERFORMANCE E VALUATION is a stochastic process, the type of current frame depends on
We evaluate the performance of DQ-LoRa by the following the number of successful minislots in all the previous frames.
metrics. Therefore, it is cumbersome to obtain an accurate closed-form
1) System Throughput: The time occupation of the data expression for the number of Type-II frames n .
slots in a BCN period. As the authors demonstrated in [29], the speed of contention
2) Average Delay: The average time required for an ED to resolution is faster than the speed of DT when the number
complete its DT from its first contention. of minislots is no less than 3, which indicates that n < n
3) Average Energy Consumption: The average energy con- when m ≥ 3. According to the access procedure, we surmise
sumed by an ED in a BCN period. that n is much less than n when n is large. In addition to
that the traditional LoRaWAN can operate well under the low
traffic loads and only suffers from congestion as the traffic load
A. System Throughput increases [3], [10], and [11]. Van den Abeele et al. [11] showed
As we defined before, the system throughput can be repre- that when the number of EDs is greater than 100 the packet
sented as delivery ratio declines significantly. For that we evaluate n for
nTDATA n ≥ 100 in the numerical experiments, and show it in Table II.
S= (2) It can be observed from Table II that: when m = 2, n is
Toverhead + nTDATA
about 50% of n and it is much larger than n in other cases;
where TDATA is the transmission time of a data packet, n is the 2) when m = 3, n is about 10% of n under the condition of
number of active devices (i.e., the number of Type-I frames), n = 10 000; 3) when m ≥ 4, n /n is less than 10%; 4) the
and portion of n is reduced when n increases. Therefore, we can
assume that: 1) n + n ≈ n under the condition of m ≥ 4, or
Toverhead = TBCN + (n + n )(mTRAP + TFBP ) (3) the condition of m = 3 and n ≥ 10 000 and 2) n + n ≈ 1.5n

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 768 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 1, JANUARY 2020

(a) Fig. 4. Throughput comparison between DQ-LoRa and pure Aloha, with the
payload size of 20 bytes.

It is well know that the throughput of pure Aloha can be

expressed as SPA = G exp (−2G), where G = λTframe is the
traffic load. Here, λ = n/T is the input rate, T and Tframe are,
respectively, the BCN period and frame duration.
Fig. 4 compares the system throughput of DQ-LoRa with
pure Aloha. From Fig. 4, the throughput of DQ-LoRa is
independent of the number of EDs due to the simplified
expression (4). The result is consistent with the conclusion
made in [5] and [6]—DQ ensures high throughput regardless
of the traffic load. Besides, the throughput performance of DQ-
LoRa is always greater than that of pure Aloha. When m = 4,
(b) the minimum throughput gain is up to 2.6 times.
Fig. 3. Throughput performance of DQ-LoRa for any n. Throughput versus
(a) payload size and (b) number of minislots. B. Average Delay
The delay model of DQ-LoRa is similar to that of DQ [5],
which is composed of two queue subsystems: 1) collision res-
when m = 2. Hence, for n ≥ 100, (2) can be simplified to
olution (CR) subsystem and 2) DT subsystem. Therefore, the

TDATA
, m=2 average delay (T d ) for an active ED to complete its DT can
S = 3TRAP +1.5T FBP +TDATA
TDATA (4) be broken down into two terms: 1) the delay of CR subsystem
mTRAP +TFBP +TDATA , m ≥ 3.
(T CR ) and 2) the delay of DT subsystem (T DT ), that is,
The error introduced by the approximation of n + n is that the
T d = T CR + T DT . (5)
throughput calculated by (4) is greater than the actual value.
Nevertheless, the error is reduced when n increases. Besides,
1) Delay of CR Subsystem: We assume that the arrival
for m = 2 and m = 3, the throughput calculated by (4) can be
process of active EDs is memoryless and the input rate λ
used as a reference, but it is less accurate. Later in this section,
is Poisson distributed. Since λ = n/T, the higher input rate
we focus on analyzing the numerical results for m ≥ 4.
implies larger number of active devices in a BCN period.
2) Numerical Results: The impacts of the payload size and The probability of an arrival device finding a free minislot to
the number of minislots on the system throughput are shown access in a frame, P(λ), has been derived in [16]. However, the
in Fig. 3. It can be observed from Fig. 3(a) that the through- expression of P(λ) is inaccurate due to the wrong calculation
put performance is improved when the payload size increases. of the probability that nf devices contend in a given frame.
The trend of improvement becomes flat when the payload size We correct it as follows:
is larger than 150 bytes. In Fig. 3(b), it can be seen that the
maximum throughput occurs at m = 3 and the throughput at ∞
  
m = 4 is slightly lower. Additionally, the throughput expe- P(λ) = P success|k = nf P k = nf
riences degradation when m ≥ 4. One explanation for the nf =0
∞
   
degradation is that although increasing the number of minislots
−λ λnf −λ 1 1 nf −1
speeds up contention resolution, it also burdens the overhead, =e + e m 1−
nf ! m m
thus lowers the system throughput. Note that the steep drops nf =1
at m = 5 and m = 29 are caused by LFBP , which has been me−λ/m − e−λ
explained in the previous section. = (6)
m−1

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 WU et al.: DQ-BASED RANDOM ACCESS PROTOCOL FOR LoRa NETWORKS 769

Fig. 5. Impact of the number of minislots on the system throughput, under Fig. 6. Delay comparison of DQ-LoRa and pure Aloha.
different input rates.

traffic load G in a pure Aloha system consists of two parts:

where P(success|k = nf ) is the probability of an active device 1) the new arrival packets and 2) the retransmitted packets.
randomly choosing a free minislot when there are nf active For the Aloha access in LoRaWAN, however, it is limited
devices in a given frame, and P(k = nf ) is the probability that to the duty cycle regulations (e.g., 1% in EU 868), which
nf active devices contend in that frame. means if an ED transmits a packet and collides on the sub-
According to [16], the service time of CR subsystem is a band 868 MHz, it must wait for an off-period (99Tframe ) to
Poisson distributed random variable with mean retransmit. Therefore, the collided packets in one duty period
  −1 do not arrive immediately in this period, and the total traffic
1
t = ln . (7) can be considered to comprise only the new arrival packets,
1 − P(λ)
that is, G = nTframe /T. Besides, it is worth to mention that the
We can therefore model the system as an M/M/1. Thus, the expression for delay calculation is only applicable when the
total delay in CR subsystem is [30] system is stable (i.e., G ≤ 0.5), because the delay performance
  −1 degrade dramatically when G > 0.5 [31].
1 1
T CR = = ln −λ (8) The delay comparison between DQ-LoRa and pure Aloha
μCR − λ 1 − P(λ)
is shown in Fig. 6. Note that for the reasons we mentioned
where μCR = ln [1/(1 − p(λ))] is the service rate of CR above, the delay curve of pure Aloha is only plotted for
subsystem. n ≤ 500 (i.e., G ≤ 0.5). Fig. 6 exhibits that the delay of pure
2) Delay of DT Subsystem: Since both arrival and service Aloha increases exponentially with n, while that of DQ-LoRa
time processes are Poisson distributed, CR subsystem output increases slowly and keeps at a low level. When n = 500, the
traffic pattern will also be Poisson distributed with the same average delay of DQ-LoRa is about 54% lower than that of
rate λ as the input traffic [5]. This output traffic is directly pure Aloha. The results indicate that DQ-LoRa provides sig-
the input traffic of DT subsystem, hence the input rate of DT nificant improvement of delay performance comparing to pure
subsystem is also λ. Aloha.
As the service time for every ED in the DTQ is one frame,
DT subsystem can be modeled as an M/D/1, with a service C. Energy Consumption
rate μDT = 1. We can then obtain T DT immediately
From the perspective of an active device, the duration of a
1 ρ 2DT 2−λ BCN period can be represented as [6]
T DT = ρDT + = (9)
λ 2(1 − ρDT ) 2(1 − λ)
T = TBCN + T active + T sleep (10)
where ρDT = λ/μDT . Substituting (8) and (9) to (5), the
where TBCN is the transmission time of a BCN, T active is the
average delay can be obtained and it is in the unit of Tframe .
average time elapsed since a BCN round starts (after the BCN
3) Numerical Results: The curves of average delay against
transmission) until an active ED successfully transmit its data,
the number of minislots under different input rates are exhib-
and T sleep is the average time from the end of the successful
ited in Fig. 5. As we expected, the higher input rate results in
transmission to the end of the BCN round. The middle part
longer delay. However, under the same input rate, the number
can be expressed as
of minislots has little influence on the delay and the delay 
is almost constant when the number of minislots is larger T active = CRRAP + CRsleep Tframe
than 10. One explanation is that the duration of minislots, com- 
+ DTDATA + DTsleep Tframe (11)
pared to the data slot, only covers a small portion of the frame
duration. where CRRAP is the average number of frames, where the ED
The average delay of pure Aloha, according to [31], can be contends in the CW, CRsleep is the average number of frames,
expressed as T PA = exp (2G)Tframe . As we known, the total where the ED is sleeping in the CRQ waiting for the next

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 770 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 1, JANUARY 2020

TABLE III
contending, DTsleep is the average number of frames, where the S YSTEM PARAMETERS
ED is sleeping in the DTQ waiting for its DT, DTDATA is the
average number of frames, where the ED is in the first position
of DTQ to transmit data, and the frame duration Tframe =
mTRAP + TDATA + TFBP .
1) Expression of Average Energy Consumption: The aver-
age energy consumed by an ED in a BCN period can be
represented as

E = EBCN + Eactive + ρsleep T sleep (12)

where EBCN is the energy consumption during TBCN , Eactive

is the average energy consumption during T active , and ρsleep is
the power consumption in the sleeping mode. The middle part
can be further expressed as

Eactive = CRRAP ERAP + CRsleep + DTsleep
× ρsleep Tframe + DTDATA EDATA (13)

where ERAP and EDATA are the energy consumed by the ED

in a frame where it contends and transmits data, respectively.
At the beginning of each BCN period, every active ED
listens to the channel to receive the BCN from GW, hence (a)

EBCN = ρrx TBCN (14)

where ρrx is the power consumption in the receiving mode.

When the ED contends in a frame, it executes the following
actions: 1) transmits an RAP in one of m minislots, and keeps
in the standby mode in the other m − 1 minislots; 2) keeps in
the standby mode in the data slot; and 3) receives FBP in the
feedback slot. Therefore, ERAP can be represented as

ERAP = ρtx + (m − 1)ρstby TRAP + ρstby TDATA + ρrx TFBP
(15)

where ρtx and ρstby are the power consumption in the trans- (b)
mitting and standby modes, respectively.
Fig. 7. Impacts of the numbers of active EDs and minislots on average
When the ED transmits data in a frame, it executes the energy consumption, with payload size of 100 bytes. Energy consumption
following operations: 1) remains in the standby mode in the versus (a) number of active EDs and (b) number of minislots.
m minislots; 2) transmits data in the data slot; and 3) receives
FBP in the feedback slot. Then, EDATA can be expressed as
2) Numerical Results: System parameters are summarized
EDATA = mρstby TRAP + ρtx TDATA + ρrx TFBP . (16) in Table III, which are selected according to the LoRa SX1272
data sheet [28]. The number of active EDs and the number of
Since we assume that there are no transmission errors, minislots impact the average energy consumption, which is
DTDATA = 1. Additionally, CRRAP has already been derived shown in Fig. 7. The energy consumption is growing with
in [32], that is, the increased number of active devices, in Fig. 7(a), but the
growth trend tends to be flat when the number of active
 γ  1
dn = logm (n − 1) + 0.5 + + (17) devices is more than 6000. On the other hand, the energy
ln m 2n ln m consumption does not vary monotonically with the number of
minislots, shown in Fig. 7(b). There exists an optimal setting,
where γ ≈ 0.5772 is Eulers constant. Then,
m = 28, to achieve the minimum energy consumption and it
substituting (13)–(17) into (12)
is independent of the number of EDs. Additionally, the energy
 consumption increases exponentially up to a large value when
E = EBCN + dn EARS + EDATA + CRsleep + DTsleep m decreases to less than 3. Note that there exist steep rises at
× ρsleep Tframe + ρsleep T sleep . (18) m = 5 and m = 29, and the reason has been explained before.

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 WU et al.: DQ-BASED RANDOM ACCESS PROTOCOL FOR LoRa NETWORKS 771

TABLE IV
S YSTEM P ERFORMANCE C OMPARISON 3) DQ-LoRa is more energy efficient than pure Aloha and
it provides energy savings up to 48% when the number
of transmitted packets is large.
4) The optimal setting of the number of minislots con-
tributes to an optimum frame length. For the maximum
throughput the optimal setting is 4 (the frame length
is 76 symbols), while for the minimum energy con-
sumption it is 28 (the frame length is 132 symbols).
These two optimal values are not coincident, thus a
tradeoff between the system throughput and energy effi-
ciency must be considered when setting the number of
minislots.
It is worth to mention that according to Section V-A, the opti- In order to reduce the overhead, as we mentioned before,
mum setting of minislots is m = 4, which is not consistent with the design of RAP is a short preamble rather than a com-
the optimum setting here. Therefore, a tradeoff between the plete chirp signal. It, therefore, can be generated directly in
throughput and energy consumption must be considered while the PHY layer and the design of optimum minislot duration
optimizing the number of minislots. is one of the future directions. Additionally, we only focus on
System performance comparisons among pure Aloha, the performance analysis under fixed BCN period and leave
CSMA, and DQ-LoRa are listed in Table IV. Note that the the performance analysis under variable BCN period as a
energy consumptions and transmit to target ratios (TTRs) future work. Furthermore, the joint optimization problem of
of pure Aloha and CSMA were measured in [12] and the throughput and energy consumption is also a future direction.
performance of DQ-LoRa is calculated under the condition of
m = 4. TTR refers to transmit to target ratio, which is defined R EFERENCES
in [12] as the ratio of the total number of packets transmitted
[1] F. Adelantado, X. Vilajosana, P. Tuset-Peiro, B. Martinez,
to achieve the target. Hence, the closer the value of TTR to J. Melia-Segui, and T. Watteyne, “Understanding the limits of
1, the higher the reliability and efficiency of one protocol and LoRaWAN,” IEEE Commun. Mag., vol. 55, no. 9, pp. 34–40,
vice versa. Sep. 2017.
[2] M. O. Farooq and D. Pesch, “Analyzing LoRa: A use case perspective,”
We can observe from Table IV that DQ-LoRa consumes in Proc. IEEE 4th World Forum Internet Things (WF-IoT), Singapore,
about 48% lower energy than pure Aloha, whereas it consumes 2018, pp. 355–360.
more energy than CSMA. However, the TTR of DQ-LoRa [3] A. Laya, C. Kalalas, F. Vazquez-Gallego, L. Alonso, and
J. Alonso-Zarate, “Goodbye, ALOHA!” IEEE Access, vol. 4,
equals to 1 while the TTRs of pure Aloha and CSMA are pp. 2029–2044, 2016.
higher than 1. Therefore, DQ-LoRa is more reliable than pure [4] W. Xu and G. Campbell, “A near perfect stable random access pro-
Aloha and CSMA. For CSMA, its TTR is 0.3 times higher tocol for a broadcast channel,” in Proc. Disc. New World Commun.
(SUPERCOMM/ICC), vol. 1. Chicago, IL, USA, 1992, pp. 370–374.
than DQ-LoRa, which means 30% of packets are retransmit- [5] X. Zhang and G. Campbell, “Performance analysis of distributed queue-
ted. This indicates that the throughput and delay performance ing random access protocol—DQRAP,” DQRAP Res. Group, Comput.
of CSMA would be inferior to DQ-LoRa. Sci. Dept., Illinois Inst. Technol., Chicago, IL, USA, Rep. 93-1, 1994.
[6] F. Vázquez-Gallego, J. Alonso-Zarate, P. Tuset-Peiró, and L. Alonso,
“Energy analysis of a contention tree-based access protocol for machine-
to-machine networks with idle-to-saturation traffic transitions,” in Proc.
IEEE Int. Conf. Commun. (ICC), Sydney, NSW, Australia, 2014,
VI. C ONCLUSION pp. 1094–1099.
[7] K. Zhang and A. Marchiori, “Crowdsourcing low-power wide-area
LoRa is one of the most successful and adopted LPWAN IoT networks,” in Proc. IEEE Int. Conf. Pervasive Comput. Commun.
technologies, however, the Aloha-based MAC protocol used (PerCom), Kailua-Kona, HI, USA, 2017, pp. 41–49.
by LoRaWAN limits the scalability of LoRa networks. In this [8] LoRaWANTM Specification, V1.0.2, LoRa Alliance, Fremont, CA, USA,
2017.
article, we develop a DQ-based MAC protocol over LoRa, [9] M. C. Bor, U. Roedig, T. Voigt, and J. M. Alonso, “Do LoRa low-power
referred to as DQ-LoRa, in order to improve the scalability and wide-area networks scale?” in Proc. 19th ACM Int. Conf. Model. Anal.
reliability of LoRa networks. The details of DQ-LoRa are pro- Simulat. Wireless Mobile Syst. (MSWiM), 2016, p. 5967.
[10] J. Haxhibeqiri, F. Van den Abeele, I. Moerman, and J. Hoebeke, “LoRa
vided, including the frame structure and access procedure. To scalability: A simulation model based on interference measurements,”
evaluate its performance, we theoretically analyze the system Sensors, vol. 17, no. 6, p. 1193, 2017.
throughput, average delay, and average energy consumption. [11] F. Van den Abeele, J. Haxhibeqiri, I. Moerman, and J. Hoebeke,
Results show the following. “Scalability analysis of large-scale LoRaWAN networks in ns-3,” IEEE
Internet Things J., vol. 4, no. 6, pp. 2186–2198, Dec. 2017.
1) The throughput of DQ-LoRa outperforms pure Aloha [12] M. O. Farooq and D. Pesch, “A Search into a suitable channel access
and it is independent of the traffic load. With four min- control protocol for LoRa-based networks,” in Proc. IEEE 43rd Conf.
islots, the minimum throughput gain can be up to 2.6 Local Comput. Netw. (LCN), Chicago, IL, USA, 2018, pp. 283–286.
[13] C.-T. Wu and G. Campbell, “Extended DQRAP (XDQRAP). A
times. cable TV protocol functioning as a distributed switch,” in Proc.
2) The delay performance of DQ-LoRa is insensitive to the Int. Workshop Community Netw. Integr. Multimedia Services Home,
number of minislots. It is also superior to pure Aloha, Jul. 1994, pp. 191–198.
[14] H.-J. Lin and G. Campbell, “PDQRAP-prioritized distributed queueing
with the delay savings up to 54% in the case of 500 random access protocol,” in Proc. IEEE Conf. Local Comput. Netw.,
active devices. Minneapolis, MN, USA, Oct. 1994, pp. 82–91.

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.
 772 IEEE INTERNET OF THINGS JOURNAL, VOL. 7, NO. 1, JANUARY 2020

[15] C.-T. Wu and G. Campbell, “Interleaved DQRAP with global TQ,” Dept. [24] P. Tuset-Peiro, F. Vazquez-Gallego, J. Alonso-Zarate, L. Alonso, and
Comput. Sci., Illinois Inst. Technol., Chicago, IL, USA, Rep. 94-4, 1995. X. Vilajosana, “LPDQ: A self-scheduled TDMA MAC protocol for one-
[16] L. Alonso, R. Agusti, and O. Sallent, “A near-optimum MAC protocol hop dynamic low-power wireless networks,” Pervasive Mobile Comput.,
based on the distributed queueing random access protocol (DQRAP) for vol. 20, pp. 84–99, Jul. 2015.
a CDMA mobile communication system,” IEEE J. Sel. Areas Commun., [25] K. Zhang and A. Marchiori, “Demo Abstract: PlanIt and DQ-N for
vol. 18, no. 9, pp. 1701–1718, Sep. 2000. low-power wide-area networks,” in Proc. IEEE/ACM 2nd Int. Conf.
[17] L. Alonso, R. Ferrus, and R. Agusti, “WLAN throughput improve- Internet-of-Things Design Implement. (IoTDI), Pittsburgh, PA, USA,
ment via distributed queuing MAC,” IEEE Commun. Lett., vol. 9, no. 4, 2017, pp. 291–292.
pp. 310–312, Apr. 2005. [26] J. Petäjäjärvi, K. Mikhaylov, M. Pettissalo, J. Janhunen, and J. Iinatti,
[18] J. Alonso-Zarate, E. Kartsakli, A. Cateura, C. Verikoukis, and L. Alonso, “Performance of a low-power wide-area network based on LoRa tech-
“A near-optimum cross-layered distributed queuing protocol for wireless nology: Doppler robustness, scalability, and coverage,” Int. J. Distrib.
LAN,” IEEE Wireless Commun., vol. 15, no. 1, pp. 48–55, Feb. 2008. Sensor Netw., vol. 13, no. 3, pp. 1–16, Mar. 2017.
[19] E. Kartsakli, A. Cateura, L. Alonso, J. Alonso-Zarate, and C. Verikoukis, [27] L. Tello-Oquendo, I. Leyva-Mayorga, V. Pla, J. Martinez-Bauset, and
“Cross-layer enhancement for WLAN systems with heterogeneous traf- V. Casares-Giner, “Analysis of LTE-A random access procedure: A foun-
fic based on DQCA,” IEEE Commun. Mag., vol. 46, no. 6, pp. 60–66, dation to propose mechanisms for managing the M2M massive access
Jun. 2008. in wireless cellular networks,” in Proc. Workshop Innov. Inf. Commun.
[20] B. Otal, L. Alonso, and C. Verikoukis, “Highly reliable energy-saving Technol., 2015, pp. 95–101.
MAC for wireless body sensor networks in healthcare systems,” IEEE [28] Semtech SX1272. Accessed: Mar. 16, 2018. [Online]. Available:
J. Sel. Areas Commun., vol. 27, no. 4, pp. 553–565, May 2009. https://www.semtech.com/apps/product.php?pn=SX1272
[21] J. Alonso-Zárate, E. Kartsakli, L. Alonso, and C. Verikoukis, [29] W. Xu and and G. Campbell, “A distributed queueing random access
“Performance analysis of a cluster-based MAC protocol for wireless ad protocol for a broadcast channel,” ACM SIGCOMM Comput. Commun.
hoc networks,” EURASIP J. Wireless Commun. Netw., vol. 2010, no. 1, Rev., vol. 23, no. 4, pp. 270–278, 1993.
2010, Art. no. 625619. [30] L. Kleinrock, Queueing Systems. New York, NY, USA: Wiley, 1976.
[22] J. Alonso-Zarate, L. Alonso, C. Skianis, and C. Verikoukis, “Analysis [31] M. Reyes-Ayala, E. A. Andrade-Gonzlez, J. A. Tirado-Mendez,
of a distributed queuing medium access control protocol for coopera- and H. J. Aguilar, “Average packet delay in random multiple
tive ARQ,” in Proc. IEEE GLOBECOM, Miami, FL, USA, Dec. 2010, access for satellite systems,” in Proc. Int. Conf. Telecommun., 2008,
pp. 1–5. pp. 120–124.
[23] A. Laya, L. Alonso, and J. Alonso-Zarate, “Contention resolution queues [32] A. J. E. M. Janssen and M. J. de Jong, “Analysis of contention tree
for massive machine type communications in LTE,” in Proc. IEEE algorithms,” IEEE Trans. Inf. Theory, vol. 46, no. 6, pp. 2163–2172,
26th Annu. Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC), Sep. 2000.
Hong Kong, 2015, pp. 2314–2318.

Authorized licensed use limited to: Pt Dwarka Prasad Mishra Indian Inst of Info Tech. Downloaded on November 06,2020 at 22:20:57 UTC from IEEE Xplore. Restrictions apply.