How Agile is the Adaptive Data Rate Mechanism of LoRaWAN?
HHow Agile is the Adaptive Data Rate Mechanismof LoRaWAN?
Shengyang Li ∗† , Usman Raza ∗ , and Aftab Khan ∗∗ Toshiba Research Europe Ltd., Telecommunications Research Laboratory, Bristol, UK † University of Bristol, Bristol, UKEmail: [email protected], [email protected], [email protected]
Abstract —The LoRaWAN based Low Power Wide Area net-works aim to provide long-range connectivity to a large numberof devices by exploiting limited radio resources. The AdaptiveData Rate (ADR) mechanism controls the assignment of theseresources to individual end-devices by a runtime adaptationof their communication parameters when the quality of linksinevitably changes over time. This paper provides a detailedperformance analysis of the ADR technique presented in therecently released LoRaWAN Specifications (v1.1). We show thatthe ADR technique lacks the agility to adapt to the changinglink conditions, requiring a number of hours to days to convergeto a reliable and energy-efficient communication state. As a vitalstep towards improving this situation, we then change differentcontrol knobs or parameters in the ADR technique to observetheir effects on the convergence time.
I. I
NTRODUCTION
The Low Power Wide Area (LPWA) networking technolo-gies [1] are recent outcomes of breakthroughs in communi-cation technologies, as well as miniaturization and decreasingcosts of the Internet of Things (IoT) devices. Their fast-pacedadoption across all the seven continents aims to provide totens of billions of devices a range in the order of tens ofkilometers at a fraction of the cost and energy consumptionof legacy technologies. LoRaWAN has established itself asone of the leading LPWA technologies in the past few yearsalongside S IG F OX , Ingenu RPMA, and other cellular solutionssuch as NB-IoT and LTE-M. LoRaWAN is expected to connecta large number of static and mobile end-devices (EDs). Itssupport of the direct single-hop connection between EDs andthe gateways obviates the need for any complex and expensivemulti-hop mesh routing schemes.Nevertheless, LoRaWAN exploits a limited radio bandwidthavailable in the Sub-GHz part of the industrial, scientificand medical (ISM) band that is also shared among multipleco-existing technologies. In addition to this, the EDs andgateways must respect regional regulations related to the useof this spectrum, which restrict the time-on-air of transmis-sions. For these very reasons, efficient management of thescarce radio resources is essential for achieving very high network-wide performance measured in terms of scale andgoodput/reliability of a network. Apart from the problem oflimited radio resources, the EDs are often deployed in very far-flung areas in challenging radio environments, resulting in very high variability in the link quality over time due to multipledifferent reasons such as obstructions, device mobility and en-vironmental factors [2]. Thus, the responsibility of LoRaWANstack extends well beyond sharing of the limited resourcesamong a large number of EDs. It must also include intelligentmechanisms capable of adapting the communication settingsof individual EDs to cope up with the link changes and therebydeliver reliable connectivity at all times.In this paper, we study an Adaptive Data Rate (ADR) mechanism that is responsible for the radio resource man-agement and runtime link adaptation for individual EDsin LoRaWAN. This mechanism is recently updated in theLoRaWAN specifications v1.1 [3] and is gradually makingits way into commercial deployments. The performance ofthis mechanism is yet to be fully understood. Therefore, wemake a timely contribution in this paper by evaluating theperformance of the ADR mechanism as a first step towardsexposing different limitations and improving the underlyingtechnique. Unlike most other studies [4], [5], [6], [7], [8]that focus on network-wide performance metrics related toscalability, reliability, fairness, and throughput, we focus onperformance that the individual EDs receive while runningthe ADR mechanism. We provide an in-depth study of theagility of the ADR mechanism in adapting the communicationsettings of EDs in response to link changes.In this paper, we also provide very first insights into theruntime performance of the official ADR algorithm under dy-namic link conditions and various network sizes. Our detailedresults reveal that if link conditions change or network sizebecomes too large, the convergence time of ADR mechanismto a communication setting that provides good reliability andlow energy consumption is quite high. A large number ofpackets are therefore lost, motivating improvements in itsdesign. We then investigate if changing different control knobs,specifically ADR initialization parameters, can improve theconvergence rate, and if so to what extent. We also providea brief discussion of the useful insights gained through thisstudy that can help improve future ADR algorithms.This paper is organized as follows: Section II presents a verybrief primer on LoRa and LoRaWAN immediately prior to thedescription of the ADR mechanism in Section III. Section IVand Section V present simulation setup, the performance of
Accepted to appear at IEEE GLOBECOM 2018.Copyright c (cid:13) a r X i v : . [ c s . N I] A ug tart ADR_ACK_CNT= 0
ADRACKReq= ++ADR_ACK_
CNT
Downlink packet received?Yes
ADR_ACK_
CNT >=
ADR_ACK_LIMIT?No
No Send uplink packet with ADRACKReq= ++ADR_ACK_ CNTYes
Downlink packet received?
Yes No ADR_ACK_
CNT =
ADR_ACK_
LIMIT + ADR_ACK_
DELAY? No Yes
TP = TP max and SF = SF max ? Yes
End: Improvement not possible No TP < TP max ? TP = TP + 3SF = SF + 1YesNo
ADR_ACK_
CNT =
ADR_ACK_
LIMIT
If TP ≠ max: TP + 3 Else if SF ≠ max: ++SFCNT=LIMIT (a) ADR mechanism on ED side StartADR_ACK_CNT= 0
ADRACKReq= packet ++ADR_ACK_CNTDownlink packet received? YesADR_ACK_CNT >=
ADR_ACK_
LIMIT?No NoSend uplink packet with
ADRACKReq= ++ADR_ACK_CNTYes No ADR_ACK_CNT=ADR_ACK_LIMIT+ADR_ACK _DELAY? If TP ≠ max: TP + 3 Else if SF ≠ max: ++SFCNT=LIMIT No Yes
TP = TP max and SF = SF max ? YesEnd: Improvement not possible
StartUplink ADR bit = 1?Yes Network records
SNR of uplink packet Recorded N uplink packets? No Yes 1.Get SNR max (maximum SNR of these uplinks)2.Get SNR req (demodulation threshold based on current data rate) margin = SNR max - SNR req - margin_dBm
Step = round(SNR margin /3) Send MAC commands. ++Attempt
Yes EndSuccess? YesNoAttempt>3?
YesNetwork removes SNR of least recent uplink No N Step > 0 and
SF > SF min ? SF = SF - 1N
Step = N
Step - 1Yes No N
Step > 0 and
TP > TP min ? TP = TP - 3 N Step = N
Step - 1
YesNo N
Step < 0 and
TP < TP max ? TP = TP + 3N
Step = N
Step + 1No
TP < TP max ?YesTP = TP + 3 No SF = SF + 1No Downlink packet received? Yes
No Yes NoNoYes SF < SF max ? (b) ADR mechanism on network side Fig. 1: ADR algorithmthe ADR mechanism and the impact of various factors on it.Section VI quantifies the extent to which the agility of theADR mechanism can be improved by optimizing its differentparameters. We then conclude the paper.II. L O R A AND L O R A WAN
LoRa , a Chirp Spread Spectrum (CSS) technique, is theunderlying PHY layer used by LoRaWAN, the upper networkstack designed by LoRaWAN Alliance. The CSS techniquesupports multiple data rates. For a low data rate, it uses a largeSpreading Factor (SF) that puts a high level of redundancyand amount of energy in a signal. Therefore, the signal canreach long distances and still retain enough strength to besuccessfully received. The same explanation holds for a signalsent at a high Transmission Power (TP). From the radioresource management point-of-view, the use of high SF valueskeeps radio medium busy for a long duration due to low datarates. Therefore, it is desirable to use the lowest possible SFthat provides a good link between EDs and network. Differentorthogonal SFs enable multiple successful receptions whenused to send packets overlapping both in time and frequencychannel. In the context of this paper, we restrict our discussion to SF and TP, the two control knobs adjusted by the ADRalgorithm in pursuit of achieving long-distance, reliable, andenergy-efficient communication.
LoRaWAN defines the higher layer protocols and the networkarchitecture that enable EDs to directly connect with thegateways using an ALOHA based multiple access scheme overthe sub-GHz ISM bands. The gateways are then connectedto the network servers that perform device authentication,downlink transmission scheduling, and execution of a partof ADR algorithm among many other important network-level functions. LoRaWAN mechanisms respect the regionalregulation related to the use of the sub-GHz ISM spectrum,such as those governing maximum TP and duty cycles. Thispaper assumes operation of LoRaWAN in Europe where TPand duty cycle are limited to 14 dBm and 1% respectivelyfor the default frequency channels. The SF can be variedfrom 7 to 12 to adapt both the communication range and datarate. Three device classes are defined based on the applicationrequirements for overall energy-efficiency and downlink com-munication latency. The most energy-efficient (and typicallybattery-powered) EDs are the Class A devices that experiencehe longest latency to receive downlink messages that aresent by the network only shortly after an uplink transmission.The other device classes provide additional opportunities forreceiving downlink messages at the expense of higher energyconsumption.III. A
DAPTIVE D ATA R ATE M ECHANISM
The ADR mechanism is part of LoRaWAN specifications. Itaims to provide a fairly reliable and battery-friendly connectiv-ity by adapting SF and TP to changes in link conditions. BothEDs and the network play an important role in this process.If an ED observes that a large number of consecutiveuplink transmissions are not followed by a downlink responsefrom the network, it assumes lost connectivity and resolvesthis issue by gradually stepping up its TP to the maximumbefore doing the same for SF. These measures graduallyimprove the robustness of the link. Figure 1a explains the fulloperation of EDs for adapting their TP and SF according toLoRaWAN Specifications v1.1. The two parameters namelyADR ACK LIMIT and ADR ACK DELAY control the num-ber of uplink messages, after which if a downlink response isnot received, an ED must increase either TP or SF. The valueof these parameters along with the network size, deploymentenvironment, and the amount of link fluctuations, all affect thetime to converge to a state where ED is able to successfullyre-establish a reliable link to the network. Section V providesa detailed analysis of these aspects.The EDs adapt communication setting to establish a reliable,but not necessarily an energy-efficient communication with thenetwork. EDs can, however, request the network to step in andmonitor the quality of uplink receptions from the recent past.If the link quality calculated over the last N packets is toohigh compared to the minimum receiver sensitivity threshold,the network decides to reduce SF and/or TP. The new SFand TP values are set such that the expected signal-to-noiseratio of the future packets is above the minimum receiversensitivity threshold by a pre-configured margin. Reduction inSF and TP would enable faster (high data rate) transmissionsthat consume less energy. Semtech, the organization thatdesigned LoRa, provides recommendations for implementingthe network-side of the ADR algorithm, which is adopted bydifferent operators as well as The Things Network, a popularcrowd-sourcing LoRaWAN network. On the network-side, N ,the minimum number of received packets that the networkrequire to choose values of TP and SF, significantly affect theagility of the ADR algorithm as highlighted later in Section V.IV. S IMULATION S ETUP
We implement the ADR algorithm in LoRaWANSim [9], adiscrete event simulator that already includes LoRa as well asdetailed LoRaWAN MAC protocol features including supportfor downlink traffic, retransmissions and support for ClassA devices. Both EDs and the network are made capable ofexecuting their sides of the algorithm shown in Figure 1, oneof the contributions of this paper. TABLE I: Simulation Parameters
Parameter ValueCommunication range 670 mAverage message rate 10 minutes per messageCarrier frequency g1 sub-band (868.1, 868.3, 868.5 MHz)Bandwidth 125 kHzCode rate 4/5Spreading factor 7 to 12Transmission power {
2, 5, 8, 11, 14 } dBmPath loss values [4] d = 40 m, γ = 2.08, L pl (d ) = 127.41 dBChannel variation level Low to High: σ = {
0, 1.785, 3.57 } dB To evaluate the ADR mechanism, we simulated networksconsisting of a gateway with a varying number of randomlydistributed EDs that transmit one packet every ten minutes onaverage. An urban scenario with the same path loss modelas used by earlier works [9], [6] is assumed. The EDs aretuned to use the default three central carrier frequencies in theg1 sub-band of the European sub-GHz ISM band, which issubject to a 1% duty cycle limit. The simulation parametersused in following experiments are listed in Table I.Each experiment accounts for 12 days of simulated timeand the reported results are averaged over 30 repetitions. Theerror bars in all the performance figures represent the standarddeviations. V. P
ERFORMANCE E VALUATION
We are mainly concerned with how quickly an ED, afterexperiencing a change in its link quality, converge to a statewhere it is assigned the right SF and TP values by the network.To quantitatively measure this, we define convergence time asthe duration from the change in the link quality until whenthe network receives enough number of packets required tocompute the new SF and TP values. We are also interestedin the amount of energy consumed by the radio during thisperiod.Now, we evaluate the impact on the ADR mechanismdue to various factors such as network scale, deploymentenvironment, traffic type and link changes.
A. Impact of Network Size
Firstly, we start by looking at how network size affects theruntime performance of the ADR algorithm. For this particularexperiment, we initialize networks of varying sizes and let theADR algorithm run for all the EDs for some time so that theEDs acquire stable values of SF and TP. We then introduce tothe networks additional 100 EDs and measure and report theirconvergence times.Figure 2a shows that when the network size is increased, theconvergence time increases as well from around 200 minutesfor a 100-node network to more than 3000 minutes for a 4000-node network. The slow convergence in the large networks isdue to a very high contention between a large number of uplinktransmissions. This can be validated by Figure 2b that showsthat data loss due to collisions increases from approximately17% to 85% as network scales from 100 to 3000 EDs. Thus, A v e r ag ec o n v e r g e n ce t i m e ( m i n ) (a) Convergence time P ac k e tl o ss p e r ce n t a g e (b) Packet loss Fig. 2: Impact of network size on ADR algorithm. A v e r ag ec o n v e r g e n ce t i m e ( m i n ) Low variation in channelMedium variation in channelHigh variation in channel 0 1000 20000100200300400500600700800
Fig. 3: Impact of channel condition on convergence time ofADR algorithm.it takes more time for the network to receive N transmissionsfrom the EDs required to assign optimal TP and SF values.Furthermore, once the new values of TP and SF are calculated,the network is required to send downlink command messagesto inform EDs. This is often not possible for a larger networkdue to the 1% duty cycle limit on the transmissions from thegateway, resulting in additional delay. B. Impact of Deployment Environment
Different deployment environments cause different levelsof variation in the radio channel. We simulate three differentscenarios related to low, medium and high levels of channelvariation – consistent with [6]. This is achieved by changingthe value of the standard deviation that accounts for theshadowing effect in the log-normal path loss model used inour simulations.Figure 3 shows the convergence time of the ADR algorithmunder three different conditions. When network size is small,a high variation in channel slightly increases the convergencetime of ADR algorithm.Whilst one will naturally expect this to happen, highervariation in channel signifies that uplink packets are morelikely suffering from losses due to fading. Figure 4 highlightsdifferent reasons of packet loss. The total percentage packetloss under highly varying channel exceeds the total percentagepacket loss under a less varying channel. However, collisions P ac k e tl o ss p e r ce n t a g e UL Collision UL Fading (a) 100 network size P ac k e tl o ss p e r ce n t a g e UL Collision UL Fading (b) 3000 network size
Fig. 4: Packet loss during ADR algorithm
125 130 135 140 145 150 155Pathloss0100020003000400050006000 A v e r ag ec o n v e r g e n ce t i m e ( m i n ) (a) Increase in path loss A v e r ag ec o n v e r g e n ce t i m e ( m i n ) (b) Decrease in path loss Fig. 5: Impact of change in link quality on convergence timefor the ADR algorithm.happen more often under low variation. High packet losstranslates to longer convergence times.When the network size is large, we can observe that highervariation in a channel does not necessarily lead to an increasein the convergence time. Rather, it actually reduces the con-vergence time. As shown in Figure 4b, the uplink collisionshappen much more for low channel variation. Although uplinkfading under high varying channel occurs more often than theone in the low varying channel as expected. But under thelow varying channel, the increase in the uplink collisions isso high that the total packet loss outweighs the one under highvarying channel. One potential reason is that a high variation inchannel introduces randomness in the received signal strengthof the uplink packets. This leads to a large difference betweensignal strengths of overlapping packets, resulting in successfuldecoding of the strongest signal due to the capture effect.More uplink packets can therefore successfully reach thegateway. Another possible reason is high variation in channeleases the crowded network. High variations in channel causesome uplink packets to not reach the gateway, reducing thecontentions for others that do. Thus, the chances of collisionreduce at gateway. Overall, the success of uplink packetsreaching gateway is more influenced by the number of col-lisions rather than fading for large network size. This trend isreversed for network size.
25 130 135 140 145 150 155Pathloss0100020003000400050006000 A v e r ag ec o n v e r g e n ce t i m e ( m i n ) Confirmed UL=0%Confirmed UL=25%Confirmed UL=75% (a) Convergence time
125 130 135 140 145 150 155Pathloss051015202530 A v e r ag ee n e r g y c o n s u m p t i o n ( J ) Confirmed UL=0%Confirmed UL=25%Confirmed UL=75% (b) Energy consumption
Fig. 6: Impact of confirmed uplinks on ADR algorithm
125 130 135 140 145 150 155Pathloss0100020003000400050006000 A v e r ag ec o n v e r g e n ce t i m e ( m i n )
10 Past Frames20 Past Frames (a) Convergence time
125 130 135 140 145 150 155Pathloss0510152025 A v e r ag ee n e r g y c o n s u m p t i o n ( J )
10 Past Frames20 Past Frames (b) Energy consumption
Fig. 7: Impact of number of past frames collected on ADRalgorithm
C. Impact of Link Changes
LoRaWAN is a technology of choice for various smartcity applications that are often deployed in dense urbanenvironments. Wireless links may often degrade or improvein such environments suddenly. Smart parking application isone example where the parking sensors often are obstructedby vehicles causing links to degrade. Now, we simulate suchchanges in communication environment by altering the meanpath loss value of the communication between end devices andgateway. Figure 5a simulates a case when the link quality ofan ED degrades due to obstructions or mobility that increasethe mean path loss. Figure 5b, on the other end, shows a casewhere ED improves its link quality by reducing mean pathloss.Figure 5a clearly shows when the link quality degrades, thetime required by the ADR algorithm to converge to the rightcommunication parameter setting increases significantly. Thisoverhead mainly comes from the process running on ED toregain connectivity to the gateway. Unfortunately, this processrequires EDs to lose sufficient number of sent packets beforemoving to higher SF or TP values. In Figure 5b, when node hasa good link quality, initially ADR takes a slightly longer time,mainly because of channel variation and resulting additional
125 130 135 140 145 150 155Pathloss0100020003000400050006000 A v e r ag ec o n v e r g e n ce t i m e ( m i n ) ADR ACK LIMIT=32ADR ACK LIMIT=64 (a) Convergence time
125 130 135 140 145 150 155Pathloss0510152025 A v e r ag ee n e r g y c o n s u m p t i o n ( J ) ADR ACK LIMIT=32ADR ACK LIMIT=64 (b) Energy consumption
Fig. 8: Impact of ADR ACK LIMIT on ADR algorithm
125 130 135 140 145 150 155Pathloss0100020003000400050006000 A v e r ag ec o n v e r g e n ce t i m e ( m i n ) ADR ACK DELAY=16ADR ACK DELAY=32 (a) Convergence time
125 130 135 140 145 150 155Pathloss0510152025 A v e r ag ee n e r g y c o n s u m p t i o n ( J ) ADR ACK DELAY=16ADR ACK DELAY=32 (b) Energy consumption
Fig. 9: Impact of ADR ACK DELAY on ADR algorithmpackets due to loss of uplink packets. However as the linkquality continues to improve, gateway continues to receivethe packets for computing SF and TP. In our simulations, theconvergence time approaches 200 minutes, the minimum timerequired to receive N = 20 packets that are transmitted every10 minutes on average. D. Impact of Traffic Type
Some application traffic in LoRaWAN can be of morevalue than the rest and therefore should be acknowledged bythe network. For this purpose, LoRaWAN supports confirmedmessages. We now are interested in how the convergence timeis affected by the confirmed uplink traffic. Figure 6a shows theeffect of different percentage of uplink packets requiring ACKon the convergence time. If an ACK is lost, the frame willbe retransmitted after a short time. Retransmissions improvelink reliability, enabling the gateway to collect enough numberof packets quickly, thus reaching optimal SF and TP earliercompared to the cases when all the messages are unconfirmed.VI. O
PTIMIZATION FOR THE
ADR M
ECHANISM
After learning that the convergence of ADR is slow, wenow change different tunable parameters of the ADR algorithm(shown in Table II) to analyze the corresponding effect on itsperformance. We consider N , the number of packets requiredABLE II: Tunable parameters in the ADR mechanism Parameter Description N Number of packets required by the net-work to compute SF & TPADR ACK LIMIT Threshold on number of lost packets toforce DL ACKADR ACK DELAY Threshold on number of lost packets toincrease SF or TP by the network for ADR calculation, ADR ACK DELAY andADR ACK LIMIT for this purpose.From Figure 7 to Figure 9, we show convergence timeand energy consumption of ADR algorithm using differentparameter settings. Both Figures 7 and 8 show only marginalimprovement in convergence time. This can mainly be at-tributed to the fact that most time is consumed by the node toregain connectivity. This is achieved by increasing transmis-sion power and spreading factor step by step.Reducing the value of N requires the gateway to collectless number of packets in order to decide the values of SFand TP. This, in principle, should reduce the convergencetime. However, this does not bring significant reduction in theoverall time. This is because if the link is really bad, nodes arestill required to send and lose sufficient number of packets toincrease their SF and TP gradually to a reliable communicationsetting. In this case, the overall convergence time is mainlydominated by regaining connectivity rather than collection of N packets.Alternatively, if we reduce value of ADR ACK LIMIT, thiswill speed-up the process of requesting downlink response(ADRACKReq). Nevertheless, if the link is still bad, the EDsmay have to send multiple of ADR ACK DELAY packetsto gradually step up to right settings of SF and TP. There-fore, tuning N and ADR ACK LIMIT will not result in asignificant reduction in convergence time especially if the linkquality degrades quite a lot.Figure 9 shows a better improvement in both convergencetime and energy consumption of the algorithm if we reduceADR ACK DELAY. This is easy to understand if we lookback at the functionality of ADR on the node side. Nodeswill only increase either TP or SF every time when counterreaches (ADR ACK LIMIT + ADR ACK DELAY). Reduc-ing ADR ACK DELAY means we decrease the duration ofeach individual step that increases either TP or SF. The result-ing shorter duration will accumulate to speed-up the processof making link more robust against packet loses. Effectively,the shorter convergence time also brings the benefit of lessenergy consumption because of less number of transmissionattempts. VII. D ISCUSSION AND C ONCLUSION
The main objective of this paper was to understand theperformance of the official ADR mechanism proposed by LoRaWAN Alliance in LoRaWAN specifications v1.1. Weassessed the impact of different configurable parameters onthe performance of the ADR mechanism that runs on boththe EDs and the network. In this process, we attempted toanswer several questions i.e., how different factors impact theconvergence time. We provide useful insights into improvingthe algorithm: • The convergence of the ADR mechanism is slow, moresignificantly when the link quality degrades and EDs needto move from lower to higher value of SF or TP to regainconnectivity. This suggests that ADR should promptlyidentify the onset of lost connectivity and then increaseSF and TX power. • The convergence time is more sensitive toADR ACK DELAY compared to ADR ACK LIMIT.These observations should be taken into account whileimproving the ADR algorithm.The slow convergence rate of the ADR mechanism alsointroduces higher energy consumption and packet losses. Thelack of the necessary agility to adapt to changing link in ouropinion is a good research challange for current LoRaWANspecifications. Our observations about the effect of differentparamters on the convergence time provide a good directiontowards proposing the next generation of ADR algorithms that,in addition to being adaptive by definition, must also be agileand more reliable. R
EFERENCES[1] U. Raza, P. Kulkarni, and M. Sooriyabandara, “Low power wide area net-works: An overview,”
IEEE Communications Surveys Tutorials , vol. PP,no. 99, pp. 1–1, 2017.[2] M. Cattani, C. A. Boano, and K. Rmer, “An experimental evaluation ofthe reliability of lora long-range low-power wireless communication,”
Journal of Sensor and Actuator Networks
CoRR , vol. abs/1802.10338,2018. [Online]. Available: http://arxiv.org/abs/1802.10338[6] M. Slabicki, G. Premsankar, and M. D. Francesco, “Adaptive configura-tion of lora networks for dense iot deployments,” in
NOMS 2018 , 2018.[7] F. Cuomo, M. Campo, A. Caponi, G. Bianchi, G. Rossini, and P. Pisani,“Explora: Extending the performance of lora by suitable spreading factorallocations,” in , Oct2017, pp. 1–8.[8] V. Hauser and T. Hgr, “Proposal of adaptive data rate algorithm forlorawan-based infrastructure,” in , Aug 2017, pp. 85–90.[9] A.-I. Pop, U. Raza, P. Kulkarni, and M. Sooriyabandara, “Does bidirec-tional traffic do more harm than good in lorawan based lpwa networks?”in