Indonesian J our nal of Electrical Engineering and Computer Science V ol. 40, No. 2, No v ember 2025, pp. 687 ∼ 699 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v40.i2.pp687-699 ❒ 687 MQTT li v e perf ormance on the IN A-CBT communication system: a measur ement-based e v aluation A. A. N. Ananda K usuma 1,2 , T ahar Agastani 1 , Rifqi F . Giyana 1 , Sakinah P . Anggraeni 1 , Arfan R. Hartawan 1 , T oto B. P alok oto 1 , W idrianto S. Pinastik o 1 1 Research Center for T elecommunication, National Research and Inno v ation Agenc y , Gd. T eknologi 3 KST BJ Habibie, T angerang, Indonesia 2 Department of Informatics, Uni v ersitas Multimedia Nusantara, T angerang, Indonesia Article Inf o Article history: Recei v ed Oct 30, 2024 Re vised Jul 9, 2025 Accepted Oct 15, 2025 K eyw ords: Latenc y Li v e measurement MQTT QoS le v el Tsunameter ABSTRA CT Cable-based tsunameters ha v e been deplo yed in Indonesia under the name of the IN A-CBT project. Currently , the system operated at the Lab uan Bajo land- ing station w orks well and sends aggre g ated data from the seaoor sensors to a central or read do wn station in Jakarta for further proces sing. The current scheme mak es use of a publish and subscribe indirect communication among the landing station (LS) as the publisher and v arious clients as subscribers for the sensor data. Message queue telemetry transport (MQTT) w as selected as the application-layer protocol for implementing this communication scheme. This paper presents a measurement-based e v aluation of the MQTT li v e performance by observing the MQTT messages’ latencies rece i v ed at the subscriber of the IN A-CBT’ s MQTT brok er . The results gi v e insight on the general achie v able performance of the IN A-CBT communication system in pro viding reliable data for the tsunami detection system. Furthermore, the results obtained can be used as communication parameters for making a more realistic virtual testbed for de- signing a more appropriate and scalable CBT system. This is an open access article under the CC BY -SA license . Corresponding A uthor: A. A. N. Ananda K usuma Department of Informatics, Uni v ersitas Multimedia Nusantara Jl. Scientia Boule v ard, Gading Serpong, T angerang 15810, Indonesia Email: ananda.kusuma@lecturer .umn.ac.id 1. INTR ODUCTION A disaster early w arning system is necessary for e v ery country to mitig ate the disaster’ s se v ere and unpredictable impact. Indonesia is an archipelagic country with man y v olcanoes and s eismic sites; thus, it has a high potential for se v eral disasters, such as earthquak es, tsunamis, and v olcanic eruptions. In response, Indonesia has implemented se v eral systems for disaster early w arning system, and one of them is a tsunami detection system based on ber optic cables on the seaoor wi th se v eral earth-monitoring and tsunami-related detection sensors. This is referred to as Indonesia’ s cable-based tsunameters (IN A-CBT), and the y were de- plo yed at Lab uan Bajo and Rokatenda sites in 2021. Figure 1 sho ws the locations of IN A-CBT LSs at Lab uan Bajo and Rokatenda in the pro vince of East Nusa T engg ara. These sites are at distance of approximately 1,400 km and 1,700 km respecti v ely from the read do wn station (RDS) in Jakarta, the capital city of Indonesia. Se v eral countries ha v e similar systems for natural disas ter monitoring and seaoor observ ation. Japan, through the National Research Institute for Earth Science and Disaster Resilience (NIED), operates the seaoor observ ation netw ork for earthquak es and tsunamis along the Japan trench (S-NET) [1], the ne w S-NET [2], the J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
688 ❒ ISSN: 2502-4752 dense ocean oor netw ork system for earthquak es and tsunamis (DONET) [3], and the Nankai trough seaoor observ ation netw ork for earthquak es and tsunami (N-Net) [4]. Those systems basically co v er the P acic Coast in Eastern Japan and the coast in western side of the country . T aiw an also operates a system called the marine cable hosted observ atory (MA CHO), which is used t o monitor acti v e v olcanoes and detect earthquak es and tsunamis occurrences of f the coast in the northeast of the country [5]. Other systems that are still in operation include the Canadian North-East P acic underw ater netw ork ed e xperiments (NEPTUNE) [6], considered as the w orld’ s rst multi-node cabled ocean observ atory , and the European multidisciplinary seaoor and w a- ter column observ atory (EMSO) [7]. A recent de v elopment sho ws the interest for inte grating en vironmental sensors in the repeaters of submarine telecommunication cables, which results in the system called scientic monitoring and reliable telecommunications (SMAR T) [8]. Such system pro vides data stream for v arious earth observ ation for seismic, tsunami and other early w arning scenarios, alongside re gular telecommunication traf c. Thus, future cabled-based tsunameter projects need collaboration with telecommunication industries. Figure 1. The IN A-CBT with LS at Lab uan Bajo and Rokatenda The IN A-CBT is still at its early stage b ut already sho wed some promises on its use of local engineer - ing kno wledge during its de v elopment. The IN A-CBT’ s w orking scale, consisting of one LS and tw o ocean bottom units (OB Us), is considered much smaller than the systems described abo v e, b ut an y studies performed on it are w orth it for learning e xperience, and will be useful to v arious related kno wledge. Se v eral studies ha v e reported v arious aspects of IN A-CBT , e.g. po wer supply considerations [9], f ault-tolerance analysis in its switching netw orks [10], [11], data acquisition and its tsunami detection algorithm (TD A) [12], testbed de v elopment for its sub-communication system [13], and a related seabed morphology characterization [14]. The majority of literature, especially for lar ge systems as described abo v e, concern more about tar gets on connecting lar ge numbers of OB Us through oceanoor optical netw orks, whereas interconnecting systems to public wide area netw orks for transporting sensor data to a central or read do wn station (RDS) did not get much attention. A particular w ork on modeling and testbed de v elopment of message queue telemetry transport (MQTT) t ransmission on the IN A-CBT’ s LS to RDS sub-communication in [13] in v estig ated the impact of bottleneck bandwidth on the achie v able message latencies of OB U’ s sensor data. Message latencies are critical for performance measure, especially for the ef fecti v eness of the processing algorithm for detecting possible tsunamis. Ho we v er , this w ork is considered lack of reality as it needs to be assessed and impro v ed by the kno wledge obtained from a li v e system. The MQTT protocol is widely used for iternet of things (IoT) services and sho wn to be rob ust in se v eral conte xts [15]–[19], and its performance for transporting CBT’ s data in a li v e system needs to be justied. Research w ork conducted on a li v e system is not only useful for CBT interest b ut also in general data communication systems that use an application layer’ s publi sh-subscribe mechanism lik e MQTT protocol. This paper then aims at addressing MQTT transmission issues on the IN A-CBT’ s LS to RDS sub-communication, and pro vides some contrib utions as follo ws: presenting a specic sub-communication component of IN A-CBT , discussing the netw ork and application performance metrics that can be measured from a li v e netw ork, describing measurement for MQTT mess age latencies and interpreting their results related to the requirement of TD A. The remainder of this paper is or g anized as follo ws. Section 2 pro vides a general o v ervie w of the IN A-CBT communication system. In section 3 e xperiment design and measurement procedures are e xplained. Section 4 presents some results, and their implications are discussed. Finally , this paper concludes with some remarks and future w ork. Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 687–699 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 689 2. IN A-CBT COMMUNICA TION SYSTEM The IN A-CBT communication system consists of the OB U to LS and LS to RDS sub-communi cation systems. Figure 2 sho ws a simplied vi e w of the IN A-CBT communication system at a particular LS. The system has tw o OB Us where each OB U hos ts three sensors: 3-ax es accelerometer (Acc), bottom pressure recorder (BPR), and en vironment (En v). Acc measures three-dimensional displacement as an indication of ground shaking, BPR measures pressure that is correlated to w ater column height, and En v measures OB U’ s internal conditions. Currently , the tsunami detection algorithm used is based on BPR data [12], using the well- kno wn and popularly used D AR T algorithm [20]. Data from OB Us’ sensors are aggre g ated at LS and then transported to RDS in Jakarta using an MQTT protocol for further processing. Figure 2. The IN A-CBT communication system at a particular LS. Some icons used in this gure were retrie v ed from Flaticon.com [21] In the case of Lab uan Bajo, the OB Us are positioned at about 37 km and 57 km respecti v ely from the shore; the one nearer to the shore w as laid do wn around 2,110 meters of w ater depth and the other w as laid do wn around 4,120 meters of w ater depth. The LS and the tw o OB Us are ph ysically connected in a ring topology , so that redundant links are a v ailable if incident occurs, so data can still be sent using the alternate path. Pro viding redundant links requires a loop-a v oidance mechanism, i.e. creating a spanning tree, to pre v ent looping in the reconguration process. The spanning tree mechanism in the conte xt of IN A-CBT has been studied and reported in [10], [11]; both of them in v estig ated f ailo v er and f ailback times of OB Us’ switches. Early in v estig ation w as conducted using real equipments and a proprietary turbo-ring protocol as the ones deplo yed by the IN A-CBT project, whereas these later w orks used simulator and open spanning tree protocols for e xibility in future de v elopment [10], including arbitrary number of OB Us [11]. The deplo yed OB U’ s BPR and Acc sensors transmit data through their serial ports; their data in seri al frame format are then con v erted to ethernet frame format by a serial-to-ethernet (S/E) con v erter . Based on the operational setting, BPR and Acc send data at frequencies 1 Hz and 125 Hz respecti v ely . By ha ving a switch connecting sensor de vices, each of them can be identied by its IP address and associated port number . Sensor data acquisition can then be controlled by the related program run at the landing station computer (LS PC) based on IP address and port number; data from OB U’ s sensors are aggre g ated at the LS. The program running at LS updates the timestamp of sensor data and creates MQTT messages based on their OB U’ s number as the topic for the utilized publish-and-subscribe mechanism. The publisher program sends MQTT messages to the MQTT brok er operated at RDS Jakarta via Internet. An y program that acts as a subscriber to the MQTT brok er can recei v e MQTT m essages sent from LS, and then retrie v e sensor data based on the subscribed topics for further processing. The IN A-CBT utilizes mosquitto v ersion 1.6.9 as the brok er supporting MQTT v ersion 5.0/3.1.1; client softw are, publisher and subscriber , are custom-b uilt programs based on paho-mqtt library v ersion 1.6.1. There are se v eral alternati v es for connect ing LS to the Internet, depending on its location and a v a il- able infrastructure. At Lab uan Bajo the primary link is through the Internet service pro vider which pro vides ber -optic connecti vi ty , whereas backup links are pro vided by cellular and satellite operators. The IN A-CBT MQTT live performance on the IN A-CBT communication system: ... (A. A. N. Ananda K usuma) Evaluation Warning : The document was created with Spire.PDF for Python.
690 ❒ ISSN: 2502-4752 project also has a mission to contr ib ute to local communities around its LS, so i ts polic y is to shar e its Internet connecti vity . At Lab uan Bajo, the school community around LS is allo wed to access Internet via the LS’ s W iFi access point. This situation gi v es a mix ed traf c scenario where MQTT -based disaster -related data blend with general Internet traf c [22]. The MQTT transmission between a publisher and subscribers needs to be analyzed as it impacts the quality of service (QoS) of IN A-CBT’ s sensor data. The critical parameter for data processing is the timeliness of the recei v ed data; thus, MQTT message latenc y can be considered as the performance objecti v e. MQTT supports three QoS le v els: QoS 0 (at most once), QoS 1 (at least once), QoS 2 (e xactly once); this QoS le v el gi v es an application’ s reliability option to users connecting in unreliable netw orks. MQTT performance in deli v ering data based on their QoS le v el has been in v estig ated from v arious points of vie w , such as the analyses that correlate QoS le v els to pack et errors [23], the in v estig ation of control pack ets’ beha vior to communication delays and their impacts to w ards MQTT’ s data deli v ery [24], the use of deep-learning with MQTT to correlate its QoS with potential intrusion [25], adding additional MQTT’ s payloads for security or reliability reasons that increases latencies [26], [27]. Ho we v er , there is still lack of information about MQTT performance for transporting CBT’ s sensor data. One study aims at sho wing MQTT latencies for all QoS le v els as a function of bottleneck bandwidth, and it w as conducted in an idealized and simplied testbed of CBT’ s sub-communication system [13]. A related w ork using a virtual testbed for general MQTT transmission in [28] sho ws that latencies using QoS 2 increase signicantly with respect to netw ork delays. Results abo v e sho w that in v estig ation on data deli v ery’ s latencies due to QoS le v el selection and netw ork parameters, such as bottleneck bandwidth and delay , needs further attention. Measurement results in li v e CBT system and their analyses are needed for better understanding on MQTT performance and its feasibility for transporting tsunami detection-related sensor data. 3. EXPERIMENT DESIGN The e xperiment design is sho wn in Figure 3 that sho ws the LS-RDS part of the IN A-CBT system at Lab uan Bajo and the remote monitoring and measurement station set-up at ANP Lab, KST BJ Habibie, T angerang Selatan (a satellite city southern of Jakarta). From the perspecti v e of measuring MQTT perfor - mance, it can be seen that in general there are four major components to be considered: the MQTT publisher at LS, Internet connecti vity , the MQTT brok er at RDS, and the MQTT subscriber at ANP Lab . T o control e xperiments by v arying MQTT publishing parameters, a VPN connection w as set-up between Control PC at ANP Lab and LS PC at Lab uan Bajo’ s LS. The VPN remote access at LS is part of the remote monitoring and management system of IN A-CBT infrastructure. F or a limited time and research purposes, remote access to LS and data retrie v al were permitted. The MQTT subscriber at ANP Lab (VM Subscriber) w as set-up to subscribe a particular OB U’ s topic from the MQTT brok er at RDS. Figure 3. Li v e measurement setting. Some icons used in this gure were retrie v ed from Flaticon.com [21] In this e xperiment, both Internet and MQTT brok er are considered as blackbox es that our concerns are only on their interf aces; in other w ord, only end-to-end aspects of LS to subscribers are under consideration. Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 687–699 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 691 Both netw ork and application related parameters are needed for understanding the system performance, and in this e xperiment the follo wing end-to-end parameters measured are as follo ws: a v ailable bandwidth, com- munication delay , and MQTT message latenc y . Note that the a v ailable bandwidth and communication delay are parameters estimated at LS interf ace and based on pack et transmission delay between LS and ANP Lab . The y are not related directly to the path tak en by MQTT messages from LS to subscribers via an MQTT brok er at RDS. Ne v ertheless, the measured a v ailable bandwidth gi v es indication on the link quality at LS, and the communication delay gi v es the latenc y’ s lo wer bound. T o accommodate e xtra traf c due to acti vities from communities around LS, measurements were conducted in three separate sessions: morning session (8.00 to 11.00), afternoon session (14.00 to 17.00), and e v ening session (20.00 to 23.00); these times are in W estern Indonesian time zone. At deplo yment in 2021, IN A-CBT’ s landing station at Lab uan Bajo w as designed to use an optical netw ork as its primary means of communication with RDS, as sho wn in Figure 2, and the bandwidth contract with the pro vider w as 2.5 Mbps for both upstream and do wnstream. Because bandwidth is dedicated, at an y gi v en time the v alue must be v ery close to the agreement; ho we v er , measurement is still needed to v erify its real- time v alue more accurately . The communication delay is estimated by taking end-to-end measurement between Lab uan Bajo site and ANP Lab . Separate delay measurement w as conducted to accommodate for the possibility of asym metrical delay between uplink and do wnlink due to dif ferent load on upstream and do wnstream traf c. Each delay af fects dif ferent part s of MQTT message transmission used for transporting sensor data; that is, the uplink delay af fects MQTT data and signaling pack ets while do wnlink delay only af fects MQTT signaling pack ets. The est imated a v ailable bandwidth at Lab uan Bajo site and the estimated end-to-end communicat ion delay between Lab uan Bajo site and ANP Lab pro vide indication on the path ef fecti v eness for transport ing sen- sor data. Ho we v er , what m atters to the applications that use sensor data is the latenc y of sensor data deli v ery to the application layer . Also, note that there e xist e xtra transit time to RDS site, processing time at the MQTT brok er , and e xtra time for nal tr ansmission and processing at subscriber site. Therefore, estimating latenc y at MQTT le v el is needed, and this is achie v ed by taking the dif ference of the subscribed MQTT messages’ times- tamps at the recei ving end and their associated published timestamps. The published timestamps correspond to the data processing step at LS that adds timestamps to sensor data from each OB U [12]. Since all OB Us’ data are aggre g ated at LS before being timestamped and published using MQTT , one may focus on one OB U’ s sensor data only , as their LS to RDS transmission characteristics are similar . Message latenc y measurement w as tak en in the morning, afternoon, and e v ening session; at each session, operating steps for g athering MQTT messages are sho wn in Figure 4. Based on measurement data, for each QoS le v el message latenc y samples from each sensor were obtained by taking the dif ference between subscribed timestamps and their associated published tim estamps. Note that in contrast to research on testbed where all components can be fully-controlled [13], in li v e measurement study there are se v eral things must not be interrupted, e.g. turning- of f or restarting MQTT brok er; therefore, an y data anomaly recei v ed, typically at the be ginning of transition, e.g. change in QoS le v el, will be considered as outliers. Figure 4. Operating steps for g athering MQTT messages at each session of measurement MQTT live performance on the IN A-CBT communication system: ... (A. A. N. Ananda K usuma) Evaluation Warning : The document was created with Spire.PDF for Python.
692 ❒ ISSN: 2502-4752 4. RESUL TS AND DISCUSSION Measurement studies took a v ailable bandwidth, communication delay , and MQTT message latenc y; ho we v er , MQTT mess age latenc y is the main parameter that requires more discussion. The quality of sensor data streams processed by the associated applications is hea vily af fected by the achie v able latenc y . By using iperf3, the follo wing results at LS were obtained: 2.88 Mbps do wnstream bandwidth and 2.53 Mbps upstream bandwidth. This conrmed that dedicated bandwidth is pro vided in the do wnlink and uplink link from LS at Lab uan Bajo site. Note that measurement w as tak en as end-to-end connection from LS to ANP Lab, and one needs t o ensure at ANP Lab site, the associated bandwidth is much lar ger . By making use of speedtest tool and its public serv er , the follo wing results were obtained: 12.07 Mbps do wnstream bandwidth and 14.42 Mbps upstream bandwidth. Thus, LS to RDS links need further attention as their bandwidth is the potential bottleneck along the prospecti v e end-to-end path. Measurement results also sho w that the a v ailable upstream bandwidth for CBT purposes is steady; the upstream is fully a v ailable for sensor data transmission. Dif ferent results were observ ed on the do wnstream; the a v ailable bandwidth drops in morning session. It can be inferred that intensi v e Internet acti vities in school communities around LS during s chool hours contrib ute to this drop. Not much do wnstream traf c w as observ ed in afternoon and e v ening session. Bandwidth measurement sho ws an e xpected traf c and bandwidth usage of the CBT’ s LS to RDS links. Measurement results sho w that communication delays are generally symmetrical at each session of measurement. Morning session data sho ws the statistical v alues for do wnstream are 28.080 ms on a v erage, 117 ms on maximum, and 26 ms on minimum, whereas for upstream the y are 28.106 ms on a v erage, 163 ms on maximum, and 26 ms on minimum. Afternoon session data sho ws the statistical v alues for do wnstream are 27.650 ms on a v erage, 40 ms on maximum, and 25 ms on minimum, where as for upstream the y are 28.468 ms on a v erage, 223 ms on maximum, and 26 ms on minimum. Lastly , e v ening session data sho ws the statistical v alues for do wnstream are 26.753 ms on a v erage, 42 ms on maximum, and 24 ms on minimum, whereas for upstream the y are 28.272 ms on a v erage, 126 ms on maximum, and 25 ms on minimum. It can be inferred that communication delays on LS to RDS sub-communication are considered stable, and not af fected by surround- ing traf c. The a v erage communication delays are in the range of 26 to 28 ms, with some spik es due to netw ork glitches along the path. Ho we v er , e v en with these spik es, delays are considerably acceptable to tsunami detec- tion as e xplained in later section. These delays are the lo wer bound for higher layer applications, e.g. latencies observ ed by MQTT me ssage reception, and can be utilized as the addit ional communication parameters for IN A-CBT testbed presented in [13]. Not e that communication delays af fect the beha vior of TCP , which is the underlying protocol for MQTT , and ha v e an impact on the characteri stics of MQTT mes sage latenc y [24]. The probability model re g arding MQTT message latenc y also depends hea vily on the underlying communication delay parameters [29]. The results of MQTT message latenc y are only based on messages from OB U 1 that were published and recei v ed on April 11, 2023 during three sessions of measurement based on the general procedure sho wn in Figure 4. It is understood that only certain time windo ws were allo wed for doing li v e measurements in order to minimize potential disruption. These sensor data can be used in v arious IN A-CBT related purposes, b ut in this paper the y were used only for estimating MQTT message latenc y . T aking payload of MQTT messages sho wed that BPR and En viro data size are 95 bytes and 92 bytes r especti v ely . In general, it can be e xpected that BPR and En viro data load the system almost similarly due to their comparably data size. On the other hand, Acc data size is much lar ger and in general v aries across MQTT messages with the range from 7793 to 8049 bytes. The payload of Acc is about 82 to 85 times that of BPR. These sensor data are considered small, b ut its reliability and timeliness need to satisfy the application requirement, i.e. in this case a tsunami detection system. Figures 5 to 7 sho w measured latencies for BPR, Acc, and En viro data deli v ery . The latencies plotted are their a v erage v alues with their condence interv al, measured in each session. Numerical v al ues for these a v erage v alues are presented in T able 1. As e xpect ed, the selected QoS le v el af fects the achie v able latenc y; a higher QoS le v el results in higher latenc y . This trend has also been observ ed in se v eral other studies, although in dif ferent conte xts, because the latenc y performance of MQTT depends on the application and use of the MQTT protocol itself. S e v er al other related studies generally use testbed implementations [13], [23], [27], [28], [30], netw ork simulations [31], [32], or a mathematical model [29], generally using local area netw ork (LAN) scenarios, although in certain cases a simulated wide area netw ork (W AN) is used. Meanwhile, in this paper , measurements were carried out by connecting the client to a real netw ork, which is managed by the IN A- CBT project, so that a real measurement en vironment is obtained on a W AN scale. This research contrib utes to MQTT performance measurements on tsunami-related W AN sensor netw orks. Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 687–699 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 693 In QoS 0, no guarantee in message deli v ery is pro vided; a publisher simply sends an MQTT m essage only once and does not check whether the message arri v ed at its destination. Ev en though QoS 0 pro vides the f astest message deli v ery , it is not advisable to use it for pro viding reliable sensor data transmiss ion; some data might be lost when an y kind of error occurs in the w ay , and an e xtra precaution in the application might be needed. In QoS 1, a publisher sends a message, and stores it until it gets a PUB A CK message from the MQTT brok er that ackno wledges receipt of the message. P ack et identier in each message is used to match a published message to the corresponding PUB A CK. Ho we v er , when the PUB A CK message is lost, it is possible that the same m essage being deli v ered twice. In terms of reliability , QoS 1 is more superior than QoS 0; ho we v er , e xtra protocol steps added result in higher latenc y . Note that for applications that mak e use of MQTT messages recei v ed using QoS 1, an e xtra program for reordering messages is needed. The most reliable one is QoS 2 where MQTT guarantees that each message is recei v ed only once by the intended recipients. It is accomplished by a four -w ay handshak e between the sender and recei v er pair in the path, e.g. the publisher and the brok er . When a recei v er gets a message from a sender , it processes the message accordingly and replies to the sender with a PUBREC message that ackno wledges receipt of the message. If the sender does not get a PUBREC message, it sends the published message ag ain with a duplicate (DUP) ag until it recei v es an ackno wledg- ment. Once the sender recei v es a PUBREC message from the recei v er , the sender can safely discar d the initial published message. The sender stores the PUBREC message from the recei v er and responds with a PUBREL message. After the recei v er gets the PUBREL message, it can discard all stored states of the recei v ed message and answer with a PUBCOMP message. In this w ay , the recei v er a v oids proces sing the mes sage a second time; thus, it ensures message deli v ery e xactly once. After the sender recei v es the PUBCOMP message, the message deli v ery is complete. It can be seen that e xtra protocol steps are required for QoS 2 to achie v e reliable and or - dered message deli v ery; consequently , longer latenc y is e xpected. Recei ving complete and in-order messages is useful for the application as it can focus more on its tar geted computation. Using QoS 2 is desirable, as long as its achie v able latenc y is within the application specication, as it will be discussed later . Figure 5. A v erage latencies for BPR data Figure 6. A v erage latencies for Accelerometer data MQTT live performance on the IN A-CBT communication system: ... (A. A. N. Ananda K usuma) Evaluation Warning : The document was created with Spire.PDF for Python.
694 ❒ ISSN: 2502-4752 Figure 7. A v erage latencies for En viro data T able 1. A v erage latenc y results for three measurement sessions Message latenc y (ms) Session BPR Acc En viro QoS 0 QoS 1 QoS 2 QoS 0 QoS 1 QoS 2 QoS 0 QoS 1 QoS 2 Morning 111.4 159.8 445.8 685.1 863.6 1083.5 120.2 214.0 559.4 Afternoon 183.0 278.1 828.6 944.4 1055.7 1489.4 268.1 349.8 963.5 Ev ening 380.7 363.7 935.4 1212.8 1200.3 1613.5 419.1 444.8 1033.5 Comparing latencies for BPR and En viro data from Figures 5 and 7, it can be seen that their results are relati v ely similar . Note that the data payload for BPR and En viro is almost the same size. When comparing the latencies for Acc data in Figure 6, it can be seen that much higher latencies are observ ed for Acc data. This is due to much lar ger data payload for Acc, i.e. about 82 to 85 times lar ger than BPR’ s data payload. All measured latencies sho w lo wer v alues in morning session, and their v alues increase further in afternoon and e v ening session. Since the LS to RDS sub-communicati o n is considered steady with much a v ailable bandwidth for sensor data transmission, the plausible e xplanation for increasing latencies as the day progresses is the increase in Internet traf c o v er the RDS to ANP Lab path. Analyzing traf c at RDS t o ANP Lab path w as not conducted due to its comple x and heterogeneous en vironment. Each QoS le v el responds dif ferently for dif ferent kind of sensor data. Latencies for BPR and En viro respond almost similarly for each QoS le v el, and the y need to be compared with Acc’ s latencies . The follo wing discussion considers BPR and Acc only . F or BPR data, increasing QoS le v el from the unreliable one (QoS 0) to the most reliable one (QoS 2) results in the follo wing increase of latencies: 295.7% (morning session), 352.8% (afternoon session), 145.7% (e v ening session). In contrast, for Acc data much less latenc y performance de gradation is observ ed. F or Acc data, increasing QoS le v el from the unreliable one (QoS 0) to the m ost reliable one (QoS 2) results in the follo wing increase of latencies: 58.2% (morning session), 57.7% (afternoon session), 41.1% (e v ening session). The reason that the impact of changing QoS le v el is w orse for BPR data is due to the polic y of aggre g ating all sensor data in publishing MQTT messages; each topic for MQTT transmission is set based on OB U’ s number . In this polic y , BPR and Acc data are dif ferentiated based on the el d data type in MQTT messages; all sensor data from the same OB U will share MQTT re sources, e.g. queues at the participating nodes (publisher , brok er , an subscriber). Sharing resources lik e this is not f air for lo w-rate sensor data lik e BPR and En viro. This transmission mechanism of IN A-CBT needs to be re vised so that each sensor data at each OB U gets its o wn MQTT resources. MQTT beha vior for each QoS le v el can also be inferred from the number of queued messages at the LS’ s publisher , as sho wn in Figure 8. Using QoS 0, a publisher sends messages without storing them, thus queued messages are zero. F or higher QoS le v el, queued messages b uild-up since messages need to be stored for completing the required handshak es. It can be seen from Figure 8 that e v en for QoS 2 le v el, the a v erage queued mess ages are still lo w; this indicates stability in MQTT transmission on the LS to RDS sub- communication system. Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 687–699 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 695 Figure 8. A v erage number of queued messages at the LS’ s publisher Sensor data, particularly BPR data, are used by the TD A to detect the possible occurrence of a tsunam i. The currently used algorithm is based on the algorithm de v eloped by Mofjeld for the U.S. NO AA ’ s deep-ocean assessment and reporting of tsunamis ( D AR T) program [20]. This algorithm tak es BPR data as input, and predicts incoming tides; if the predicted tide i s abo v e a certain threshold, then a tsunami alert is issued. The prediction uses a cubic polynomial based on BPR data stored o v er the past three hours, and updated with fresh data e v ery 15 seconds. More adv anced algorithm, e.g. the one that uses an articial neural netw ork (ANN) [33], also requires data updating e v ery 15 seconds. Thus, one must ensure that BPR data latencies are much less than 15 seconds for satisfying the BPR-based TD A algorithms. In terms of a v erage latencies, Figure 5 sho ws that all a v erage latencies are well belo w 15 seconds; the w orst case for QoS 2 sho ws a v erage latenc y equal to 0.94 seconds. T o ensure no data sample is abo v e 15 seconds, cumulati v e distrib ution functions (CDFs) for latencies were created, and the y are sho wn in Figure 9; subgures are presented o v er three separate measurement seasons. Figure 9(a) sho ws the morning session, Figure 9(b) sho ws the afternoon session, and Figure 9(c) depicts the e v ening session. Only CDFs f o r QoS 2 are sho wn as QoS 2 le v el is the most reliable message deli v ery with the highest latenc y . It can be seen from Figure 9 that with certainty (probability equal to one) the highest latenc y data sample is less than 3 seconds. Thus, the current system has satised the data latenc y requirement for some popular BPR-based TD A algorithms. BPR data latencies can be further impro v ed by making separate MQTT resources for BPR and Acc, as discussed before. F or Acc data transmission using QoS 2, CDFs for latencies are presented in Figure 10. The subgures sho ws the latenc y during three separate measurement sessions. Figures 10(a) to 10(c) sho w the latenc y during three separate measurement se ssions, namely the morning, afternoon, and e v ening sessions. It can be s een that latenc y data samples are belo w approximately 4 seconds. This will be suitable for algorithms that use both BPR and Acc data for tsunami prediction [34]. In general, measurement results sho w the ef fecti v eness of using MQTT transmis sion for wi de-area sensor n e tw orks for tsunami detection deplo yed by the IN A-CBT project. Some aspects of the system may need further optimization, and consequently one needs to kno w the performance limit of the IN A-C BT com- munication system. This research w ork is ideally conducted in a testbed, in line with the w ork reported in [13]; making v arious modications in a testbed will not disrupt the li v e system. (a) (b) (c) Figure 9. Cummulati v e distrib ution functions (CDFs) for recei v ed BPR data latencies o v er three separate sessions: (a) morning, (b) afternoon, and (c) e v ening MQTT live performance on the IN A-CBT communication system: ... (A. A. N. Ananda K usuma) Evaluation Warning : The document was created with Spire.PDF for Python.
696 ❒ ISSN: 2502-4752 (a) (b) (c) Figure 10. Cummulati v e distrib ution functions (CDFs) for recei v ed Acc data latencies o v er three separate sessions: (a) morning, (b) afternoon, and (c) e v ening 5. CONCLUSION This paper has presented li v e measurement results of transporting tsunami-related sensor data using MQTT protocol o v er the IN A-CBT’ s wide-area operational netw ork. T o the best of our kno wledge, this is the rst paper that presents MQTT performance measurement of latenc y-critical sensor data for predicting tsunami o v er public Internet. It has been sho wn that the IN A-CBT communication sub-system that transports sensor data from Lab uan Bajo w orks well, and measured latencies ha v e satised the tar get requirement of som e TD A algorithms. The currently deplo yed MQTT application is not optimal yet, and can be impro v ed by ensuring f airness for each type of sensor data. Future research w ork will be conducted in a laboratory-scale testbed by making use of some measured parameters during li v e measurement. Scalability of the system can be assessed by increasing sensor data payload, e.g. increasing the number of OB Us and sensors, and in v estig ating its impact to w ards performance. FUNDING INFORMA TION Authors state no funding in v olv ed. CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY Data a v ailability is not applicable to this paper as no ne w data were created or analyzed in this study . REFERENCES [1] T . Kanaza w a, “Japan trench earthquak e and tsunami monitoring netw ork of cable-link ed 150 ocean bottom observ atories and its impact to earth disaster science, ” in 2013 IEEE International Underwater T ec hnolo gy Symposium (UT ), Mar . 2013, pp. 1–5, doi: 10.1109/UT .2013.6519911. [2] M. Shinohara, T . Y amada, S. Sakai, H. Shiobara, and T . Kanaza w a, “Ne w ocean bottom cabled seismic and tsunami observ ation system enhanced by ICT , ” in 2014 Oceans - St. J ohn’ s , Sep. 2014, pp. 1–6, doi: 10.1109/OCEANS.2014.7003045. [3] N. T akahashi et al. , “Real-time tsunami prediction system using DONET , ” J ournal of Disaster Resear c h , v ol. 12, no. 4, pp. 766–774, Aug. 2017, doi: 10.20965/jdr .2017.p0766. [4] S. Aoi et al. , “De v elopment and construction of Nankai trough seaoor observ ation netw ork for earthquak es and tsunamis: N-net, ” in 2023 IEEE Underwater T ec hnolo gy (UT) , Mar . 2023, pp. 1–5, doi: 10.1109/UT49729.2023.10103206. [5] N.-C. Hsiao, T .-W . Lin, S.-K. Hsu, K.-W . K uo, T .-C. Shin, and P .-L. Leu, “Impro v ement of earthquak e locations with the marine cable hosted observ atory (MA CHO) of fshore NE T aiw an, ” Marine Geophysical Resear c h , v ol. 35, no. 3, pp. 327–336, Sep. 2014, doi: 10.1007/s11001-013-9207-3. [6] C. R. Barnes and V . T unnic lif fe, “Building the w orld’ s rst multi-node cabled ocean observ atories (NEPTUNE Canada and VENUS, Canada): science, realities, challenges and opport unities, ” in OCEANS 2008 - MTS/IEEE K obe T ec hno-Ocean , Apr . 2008, pp. 1–8, doi: 10.1109/OCEANSK OBE.2008.4531076. [7] P . F a v ali and L. Beranzoli, “EMSO: European multidisciplinary seaoor observ atory , ” Nuclear Instruments and Methods in Physics Resear c h Section A: Acceler ator s, Spectr ometer s, Detector s and Associated Equipment , v ol. 602, no. 1, pp. 21–27, Apr . 2009, doi: 10.1016/j.nima.2008.12.214. Indonesian J Elec Eng & Comp Sci, V ol. 40, No. 2, No v ember 2025: 687–699 Evaluation Warning : The document was created with Spire.PDF for Python.