Relational Consensus-Based Cooperative Task Allocation Management for IIoT-Health Networks
Carlos Pedroso, Yan Uehara de Moraes, Michele Nogueira, Aldri Santos
RRelational Consensus-Based Cooperative TaskAllocation Management for IIoT-Health Networks
Carlos Pedroso ∗ , Yan Uehara de Moraes ∗ , Michele Nogueira ∗† , Aldri Santos ∗††
Department of Computer Science - UFMG, Brazil ∗ Department of Computer Science - UFPR, BrazilEmails: { capjunior, yumoraes, michele, aldri } @inf.ufpr.br © IFIP, (2021). This is the author’s version of the work. It is posted here by permission of IFIP for your personal use. Not for redistribution. Thedefinitive version was published in Proceedings of IFIP/IEEE International Symposium on Integrated Network Management 2021 Abstract —IIoT services focused on industry-oriented servicesoften require objects run more than one task. IIoT objects posesthe challenge of distributing and managing task allocation amongthem. The fairness of task allocation brings flexible networkreconfiguration and maximizes the tasks to be performed. Al-though existing approaches optimize and manage the dynamics ofobjects, not all them consider both co-relationship between tasksand object capabilities and the distributed allocation over thecluster service. This paper introduces the ACADIA mechanismfor task allocation in IIoT networks in order to distribute taskamong objects. It relies on relational consensus strategies toallocate tasks and similarity capabilities to determine whichobjects can play in accomplishing those tasks. Evaluation onNS-3 showed that ACADIA achieved 98% of allocated tasks inan IIoT-Health considering all scenarios, average more than 95%of clusters apt to performed tasks in a low response time, andachieved 50% more effectiveness in task allocation compared tothe literature solution CONTASKI.
Index Terms —Task Allocation, Cooperative Management,IIoT-Health.
I. I
NTRODUCTION
The Internet of Things (IoT) is a heterogeneous networkwhose objects own characteristics like identity, physical at-tributes, computational and sensing capabilities [1], [2]. Forthe vast majority of IoT objects, it is important to reduce powerconsumption while communicating or making certain tasks.Meanwhile, objects must evenly share their resources andcooperate to support better network performance [3]. Thus, ob-jects that run a set of functions can collaborate to allocate andrealize different tasks [4]. In this context, Industrial Internet ofThings (IIoT) has drawn greater attention [5] since it focuseson connecting multiple objects with numerous capabilitieswithin an industrial environment, enabling thus every deviceto work in a synchronized and organized manner to performsensing and monitoring tasks. Health application use theadvanced technologies of IIoT to create interaction betweenpatients and medical staff, hospital, and medical devices bycreating a smart environment for the health domain [6], [7].Reaching fairness performance in the distribution managementof multiple tasks between IIoT objects is challenging due toconfigurations required by IIoT networks [8].Inefficient allocations of sensing tasks between IIoT objectscauses problems in distribution of computational resources,bad environmental setting, damaging the data collection andavailability [9], [10]. In this way, we must preserve theefficiency of the allocation of tasks so that the objects are always able in emergencies to perform the reconfiguration ofthe environment and also deal with the entire volume of datagenerated. The wrong configuration of the environment makesit difficult to classify objects and makes the volume of datagenerated by them available since it takes a long time to reachthe application and directly impacts its interpretation. In smarthospitals with high demand for multiple emergency services,the dynamic configuration between the environments andservices must be done in an optimized, fast, distributed, andflawless manner [11]. Thus, intelligent management optimizesthe resources of IIoT objects in the various tasks, by allowingobjects to interleave execution according to the needs of theapplication and by preserving their resources [10].Task allocation services have been extensively studied inWSN, usually treated as resource allocation for extending thenetwork life [12], in IoT networks, that issue becomes essentialfor the arise of new applications [4]. Among the works thatdeal with the task allocation, many of them apply objectvirtualization in task groups [3] or distributed consensus [12].Object virtualization applies assign tasks according to the sens-ing competencies of each object and its performance capacityin order to optimize the task execution to save resources. Thedistributed consensus applies to the equitable distribution ofresources based on the interactions of each group of objectspresent in the network, this approach is applied in networkswith a large number of participants [13].However, these solutions disregard the similarity relation-ship between objects and the executing tasks. Besides that,they do not take into account an allocation decision based onthe traits of the environment where objects are inserted. Hence,IIoT demands dynamic, adaptable, and fair solutions to handlea range of objects and to provide transparent configuration.Moreover, the solutions need to disseminate tasks among ob-jects aware of their relationship with the environment and alsothe object capabilities [6] for getting balanced management ofthe IIoT network resource. Though, it is crucial for IIoT toprovide mechanisms able to manage task allocation amongobjects by their relationships and capabilities for more robustand fairness management of available network resources.This paper introduces ACADIA (
Rel A tional C onsensus-B A se D Task Allocation for I IoT-He A lth ) a mechanism for sup-porting sensing task allocation among objects into IIoT-Healthnetworks. ACADIA arranges IIoT objects into similarity-basedclusters to address an effective distribution of sensing tasksbetween available objects. This work extends to the health1 a r X i v : . [ c s . N I] F e b omain our previous work of task allocation in [14], as well asaddressing the existence of multiple tasks. ACADIA employscollaborative relational consensus for better adaptation on thecontext, quick responses, and getting assertive decisions aboutthe capabilities of objects to make specific sensing tasks. Italso allows us to allocate simultaneously multiple and distincttasks among the clusters. In analysis against CONTASKI [14]on the NS-3 simulator, ACADIA achieved 98% of suitablyallocated tasks in a given IIoT-Health domain, average morethan 95% of clusters apt to realize sensing tasks in a lowresponse time and, achieved 50% more effectiveness in taskallocation compared to CONTASKI.This paper is organized as follows: Section II discusses therelated work. Section III defines the model and assumptionstaken by ACADIA. Section IV describes the ACADIA com-ponents and their operation. Section V shows the evaluationmethodology to analyze the performance, and the resultsobtained. Section VI presents conclusions.II. R ELATED W ORK
The demand for dynamic and distributed services based onthe resources and sensing capabilities of IoT objects has beenthe focus of several works [3], [12], [13], [15]. Though, mostof them still face many issues by managing the allocation ofthe IoT resources, like co-relationship between tasks and ob-ject capabilities, fairness in multi-tasks distribution, and flexi-ble network reconfiguration. In [3], an evolutionary algorithmbased on heterogeneity recognition heuristic in IoT networksaddresses to ensure greater stability and operational periods oftasks according to the current demands. The algorithm createscollaboration between the functions of objects by taking thetask’s demands and virtual objects groups selected to realizetasks and also reduce the energy consumption. In this solution,however, only a few objects are capable of performing thetasks attributed to network, as well as only two types of taskscan be done, which limits its use on networks including nodeswith diverse capacities. In [12], virtual objects (VO) in an IoTsmart health network realize the allocation of sensing tasksby a decentralized strategy, where VOs negotiate among themto reach a consensus on the resource allocation of the healthdevices. Despite it meets certain fairness in the task allocation,they employs only the same type of objects to perform thetasks, and hence ignoring the different sensing capacities, thevarieties of interactions, and the impact of the network size.In [14], we proposed the mechanism called CONTASKIfor allocating task in IIoT that arranges the network intosimilarity-based groups to handle the division of tasks to beallocated. Although CONTASKI makes use of a distributedconsensus strategy for decision making about the better taskdistribution for making a given service, we have ignored thesimultaneous allocation of multiple distinct tasks, even beinga condition expected in real networks. In [13], it is proposeda consensus-based heuristic approach to make decision onfault tolerance task allocation in IoT. The approach appliesthe concept of task groups and objects, so that in each taskgroup, objects are chosen as virtual and vice-virtual. The model partly gets a flexibility on the network configuration,but it needs periodically to exchange hello messages, beingcomputationally costly. In addition, they ignore the resourcescapabilities of nodes, which directly influence on the distri-bution of tasks. In [16], authors converted the task allocationissue in an IoT environment into an integration problem witha minimal degree variant to narrow the task allocation, andthus applying a genetic algorithm to reduce its execution time.Further, the objects only communicate each other by gatewaysservices responsible for managing the interaction. However,the gateways restricts the relationships between nodes and cancause communication bottleneck depending on the networksize. In [15], an algorithm decomposes sensing tasks in asensor network into distributed ones, by taking the energy con-sumption of each task in order to get better resource allocation.They also apply a centralizing entity for distribution of roles,making it costly during the network reconfiguration. In [17], analgorithm for adaptive task mapping in sensors works jointlywith the task scheduling based on a genetic algorithm toextend the network lifetime. Despite relating tasks and objectcapabilities, the centralization of distribution of tasks overloadsthe transmission channel and delays the message delivery,compromising the synchronization of tasks execution.III. II O T H
EALTH E NVIRONMENT
This section presents the structure of the IIoT health net-work, the manner how the objects communicate each other,and the model of sensing tasks in which ACADIA can runto realize the allocation management. We assume a hospitalsetting with multiple wings that can span over a number offloors and heterogeneous devices (IIoT objects) capable ofmaking different classes of sensing relative to the building’senvironment and patients’ physiological signals. Network model : An IIoT network composed by a setof objects denoted by N = { ob , ob , ..., ob n } in an area ( X x , Y y ) . All objects own an unique network identifier Id anddifferentiate each other by the sensing capabilities, representedby set C = { c , c , c , ., c n } , processing power and memory.The objects are fixed in the setting and evenly distributed inthe coverage area of the network. Also the objects do notsuffer from energy restriction due to the existence of energysource in the network. Objects take roles as common or leaderobjects and Access Points (AP). Communication model : Communication among the ob-jects takes place over the wireless medium via a sharedasynchronous channel, in which connections are reliable, andtherefore the objects do not present communication failures.Also all objects exchange messages on the network layer.Further, the data sensed by objects can be accessed by theapplication layer regardless of the location, using protocolssuch as CoAP. Sensing task model : Each task represents a demand forsensing the ambient and/or patients’ physiological signals andit requires a set of sensing capabilities, whose size is variable,that depends on the setting ACADIA runs. Those tasks takeplace in a programmed manner with predefined time duration.2 task T is a tuple { T id , C, τ, q } , so that T id is the task unique id ; C means the set of sensing capabilities required to makethe task, τ denotes the time needed to realize it and q the per- cluster quorum to perform the task. The sensing activitytakes into account two classes: the infrastructural one sensesthe environment values such as temperature and light; thephysiological one senses physiological signals like heartbeatand blood pressure as in [7]. The AP device keeps track oftasks’ status pending, when the tasks are queued; dispatched,when they are accepted by some cluster, and completed.IV. ACADIAThe ACADIA architecture comprises two modules, called Cluster Coordination ( CC ) and Task Allocation Control ( TAC ), as shown in Fig. 1. They act jointly to guarantee theconfiguration of clusters, as well as the dissemination andallocation of sensing tasks among IIoT-Health objects. The CC module arranges the network in virtual clusters and the TAC module controls the dissemination of multiple tasks tobe done by the network objects according to their sensingcapabilities. For achieving its goal, the modules exchangefive types of messages:
CapabilityDissemination that enable toconfigure the clustering;
LeaderRegister that make the leadersto register into AP;
TaskDispatch to dispatch tasks providedby the AP;
TaskAccept that support leaders to accept tasks;
LeaderToCluster that allow leaders to disseminate acceptedtask among the objects.
Cluster ManangementSimilarity
VerificationCapability Dissemination
CLUSTER COORDINATION
Capabilities
Dissemination Role
Assigner
Task
Verification
Task
DisseminationTask OperatorLeader To Cluster
TASK ALLOCATION CONTROLTaskDispatch
Fig. 1: The ACADIA ArchitectureThe CC module controls the creation and maintenance ofthe clusters by analyzing the neighboring objects using a sim-ilarity threshold of their capabilities in order to verify if theyare apt to participate in the same cluster. Therefore, every timethat CC receives a CapabilityDissemination message comefrom the neighbors it verifies such data about its identification,capabilities and number of neighbors. For that, CC accountwith three components: Capabilities Dissemination ( CD ) , re-sponsible for disseminating CapabilityDissemination messageswith the object’s identifier, its capabilities and number ofneighbors; Similarity Verification ( SV ) , which receives andverifies the fields of messages exchanged among the objects;and Cluster Management ( CM ) , which manages the clustercreation using the objects’ similarity, and the leader selection.The TAC module coordinates the sensing task allocation ac-cording to the object capabilities and dispatches multiple tasksto the objects of the IIoT network, deviating from [14], in orderto maximize and preserve their resources. It comprises the following components: Task Verification ( T V ) , Role Assigner ( RA ) , Task Dissemination ( T D ) and Task Operator ( T O ) . TVoversees which tasks should be done and what capabilities arerequired by them. Whereas RA monitors the type of tasksto be assigned to the objects, employing relational consensusbetween the leader and tasks to evaluate which ones should beallocated according to cluster’s capability. Next, DT dispatchesthe requested tasks considering the object capabilities. Lastly,TO takes care of the operation of the tasks came from theleader. Thus, the task allocation service becomes more fareand balanced, and does not overload the objects resources. A. Cluster configuration
As the IIoT-Health size involves objects with differentsensing capabilities, the CC module arranges the networkobjects in clusters based on leaders in order to create anetwork infrastructure capable to task allocate. Initially, ob-jects begin cluster configuration exchanging CapabilityDis-semination messages that carries the Id , capabilities andnumber of neighbors of the sender. Algorithm 1 describesthe cluster configuration process. Initially, each object sendsa CapabilityDissemination message in order to announceits Id , sensing capabilities ( M yCapabilities ) and numberof neighbors (
N eighborhoodSize ), through the procedure
WarmUp . When receiving a
CapabilityDissemination message,the receiver updates its neighbors (
N eighList ), alongsidewith their capabilities (
N eighCapabilities ), by the proce-dure
RecvCapabilities . The similarity verification takesinto account those information, being calculated using co-sine similarity. A neighbor can join into the cluster whenits similarity is within the threshold. This update proce-dure occurs dynamically in all objects, ensuring that eachone maintains its neighbor and cluster updated (procedure
SimilarityCalculation ).The cluster leader selection (procedure
SelectLeader )takes into account both the number of neighbors and individualcapabilities to choose the leader. After that, the leader selectedby the cluster informs to the AP that registers it as leader toguarantee the communication between AP, leaders and clustermembers and a better hierarchical network organization.Equation 1 computes the similarity value between twoobjects’ capabilities, being based on [18, Eq. 3]. The similaritytakes into account the object’s own capabilities ( C ob ) and theneighbor’s capabilities ( C ob ). In this division, the upper partcalculates the norm of the vector that means the intersectionbetween the capabilities. The bottom part takes the square rootof the multiplication of the norm of each capability vector. sim ( ob , ob ) = | C ob ∩ C ob | (cid:112) | C ob | ∗ | C ob | (1)The similarity value varies from to , being that closerto , more similar two objects are, being that, the similaritylevel is labeled S = Dissimilar, S = Neutral and S = Similaras a manner to show the levels of similarity objects and tasksget. This scale changes according to the previously established3 lgorithm 1: Cluster configuration procedure WarmUp while
60 seconds has not elapsed do B ROADCAST C APABILITIES () R ECEIVE C APABILITIES () end end procedure procedure BroadcastCapabilities Broadcast ( MyId, MyCapacities, NeighborhoodSize ) W aitInterval () end procedure procedure RecvCapabilities Id ← GetId () NeightList ← NeighList ∪ Id NeighCapabilities [ Id ] ← GetCapabilities () NeighSize [ Id ] ← GetNeighborhoodSize () end procedure procedure SimilarityCalculation foreach neighbor in NeighList do NeighborCapabilities ← NeighCapabilities [ neighbor ] sim = | MyCapacities ∩ NeighborCapabilities | √ | MyCapacities |∗|
NeighborCapabilities | if sim ≥ T reshold then cluster ← cluster ∪ neighbor end end end procedure procedure SelectLeader LeaderCandidates ← GetNeighborsGreatestNeighborhoodSize ( cluster ) Leader ← GetNeighborLargestCapabilitiesSet ( LeaderCandidates ) if Leader == MyId then SendLeaderRegisterT oAP ( MyId ) end end procedure capabilities before the IIoT is deployed and modifies accordingto the demand of application. B. Task Allocation
Tasks are made available to carry out through the AP,which keeps a list of pending tasks and dispatches themaccording to settings’ demands. The AP sends group leadersthe tasks via
T askDispatch messages.
TAC plays on anIIoT-Health infrastructure established by the cluster config-uration, and it runs guaranteeing resource maximization, i.e.,allowing the task dissemination according to the capabilitiesof each cluster. Algorithm 2 describes the task allocationprocess and how leaders and the AP negotiate the tasksbeing performed. In order to identify the leaders, the APmonitors the
LeaderRegister messages and keeps its leader listupdated (procedure
APRecvLeaderRegister ). Initially,the AP manages a collection of pending tasks (
T askList ).When dispatching tasks, the AP selects an amount ( sm ) ofmultiple pending tasks from the list and sends a TaskDispatch message to the cluster leaders announcing the task T tobe executed (procedure APSendTask ). After each dispatch,it waits for a time interval for receiving the confirmation(
TaskAccept messages) from the available and compatiblecluster leaders. When one confirmation is received at least,the AP removes it of the pending task list. Once the leaders receive the task T = ( T id , C, τ, q ) ,they verify the compatibility of their capabilities( M yCapabilities ) with the capabilities C needed toperform the task, and if the number of objects in the cluster isgreater or equal the quorum q needed. In case they meet thecriteria, the cluster leader confirms with the AP ( TaskAccept message) that it will perform the task and disseminatesthe task to the cluster. In case the cluster members cannotrealize the task, the leader doesn’t confirm this task with theAP (procedure
LeaderRecvTask ). Algorithm 2:
Task Allocation procedure APRecvLeaderRegister LeaderId ← GetId () LeaderList ← LeaderList ∪ LeaderId end procedure procedure APSendTask( sm ) foreach dispatch round do DispatchedT asks = DispatchMultipleT asks ( sm ) if W aitConfirmation () then T askList ← T askList − DispatchedT asks end end end procedure procedure LeaderRecvTask( T = ( T id , C, τ, q ) ) if C ⊆ MyCapabilities and | cluster | ≥ q then SendT askAccept ( AP ) SendLeaderT oCluster ( T ) end end procedure C. Operation
ACADIA’s task allocation acts dynamically and distributedin an IIoT-Health network on a hospital setting, whose objectsare embedded in both medical equipment’s and the structureof the building environment. The interactions between theIIoT objects occur over time and space dimensions, andobjects in the transmission radius of the others exchangecontrol messages in order to achieve a better configurationof the hospital activities. Fig. 2 illustrates how ACADIA actsfor supporting the formation of IIoT objects cluster, leaderselection, and allocation of tasks.The wireless signals meanobjects within the transmission radius of each other and thusapt to exchange control messages about sensing capabilities.Each object carries its identifier Id and a sensing capabilitiesset C that it can make. Furthermore, capabilities c , c , c and c correspond to the sensing of temperature, humidity, light-ing and body temperature, respectively. Besides, a similaritythreshold ranging between (weak) and (strong) was setupfor the formation of clusters, according to the capabilities ofeach object in the network.We show the ACADIA operation on three different mo-ments, said I t , I t , I t time instants. In I t , the set of objects( ob , ob , ob , ob ) exchange messages about its sensingcapability and neighborhood in order to realize the similaritycalculation according to Eq. 1. The four objects obtain thefollowing similarity values between them: sim ( ob , ob ) = sim ( ob , ob ) = sim ( ob , ob ) = 1 and sim ( ob , ob ) = sim ( ob , ob ) = sim ( ob , ob ) = 0 , . As the similarities4re within the range between S =Neutral and S =Similar,objects in that interval are clustered and it means objects withcapabilities c , c , c . 𝑰𝒏𝒔𝒕𝒂𝒏𝒕 𝒕𝟏 𝐨𝐛 𝐨𝐛 𝐨𝐛 𝑪 𝑪 𝑪 𝐨𝐛 𝑰𝒏𝒔𝒕𝒂𝒏𝒕 𝒕𝟐 𝐨𝐛 𝐨𝐛 𝐨𝐛 𝐨𝐛 Access point
𝑰𝒏𝒔𝒕𝒂𝒏𝒕 𝒕𝟑 𝐨𝐛 𝐨𝐛 𝐨𝐛 𝐨𝐛 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 𝑪 Fig. 2: Formation of clusters and distribution of tasksIn I t , ob is elected the cluster leader because it has highernumber of neighbors than the others. With the cluster coordi-nator operating in this way, each object keeps its neighborhoodinformation and capabilities updated through the exchange ofmessages. Thus, objects in the spatial neighborhood are seenas members of the same cluster, in addition to ensuring betterscalability to the network, since a hierarchy based on leadersaids in the quality of the information transferred. Moreover,it facilitates the distribution of tasks among the objects of thenetwork. In I t , AP dispatches a task to the leaders, that verifywhether their cluster is apt to do it. Over this view, ob , theleader, evaluates that the existing capabilities into the clusterare compatible and responds to the AP confirming that thecluster will realize such task. Other clusters might exist indifferent points in the network that are capable of carryingout other tasks sent along with task T .V. A NALYSIS
This section presents a performance evaluation of the ACA-DIA mechanism to assess its efficiency to the managementof simultaneous tasks. We implemented ACADIA in NS3-simulator, version 3.29, and make all simulations taking intoaccount an IIoT-Health scenario similar to a smart hospitalwith various levels. The object capabilities are classified instructural and health sensing functions. The former follows theones in [19] that consist of temperature , humidity , presence , light , position , and equipment condition . The latter consists of heartbeat monitoring , blood pressure , body temperature , oxy-genation , glucose level and electrocardiogram (ECG) [11]. Weevaluate six IIoT-Heath scenarios in the same aforementionedhospital setting, with 50, 100 and 150 objects, and the number of simultaneous dispatched tasks, either 2 or 4 tasks, addingup to six scenarios configurations. The objects are evenlydistributed in a area of
200 x 200 and communicate by IPv6over IEEE 802.15.4. To avoid packet loss and bottleneck, weadded a delay of to messages exchanged among objects.Scenarios with 2 dispatched tasks represent a configurationand sensing demand for an hospital wing with health servicessuch as Emergency Response (ER), named “
Demand A(
D-A ) ”. In contrast, scenarios with 4 dispatched tasks, thedemand relates to comprehensive hospital services such as ER,Intensive Care Unit and Infirmary, named “ Demand B (
D-B ) ”.We randomly generate each task and their duration employing std::minstd rand0 as random generator. Tasks’ total durationspan to account for multiple tasks being dispatched.Seeds for each round constitute the sum of number of objects,run number and number of simultaneous dispatched tasks. std::uniform int distribution derive the objects capabilitiesand tasks using the random generator. The objects’ capabili-ties, however, remained the same across the simulation rounds.Randomly producing those values brings variation to thesimulation to account for real-life differences. If tasks requiredthe same capabilities, given that objects do not have mobilityand their capabilities remained fixed across simulation rounds,then the error would be close to zero, and, therefore, notrepresentative. But, this randomization also reflects on higherstandard deviation, leading to high amplitude error bars.The system operates over and in the first there isan exchange of messages between all objects to disseminatetheir capabilities, followed by the similarity computation andleader register. The AP dispatches multiple tasks every ,from to up , and clusters perform them for or , according to tasks’ requirements. Moreover, each taskhas a random capabilities set, and all tasks require as structuralcapabilities at least temperature , humidity and presence and ashealth capabilities heartbeat monitoring , blood pressure , bodytemperature , oxygenation . The final capability set has upto three other structural capabilities and two other healthcapabilities, among the remaining ones of each type.Pending tasks are forwarded to leaders by AP, alwaysavailable, located in the center of the network, and equippedwith a strong internet signal to reach all objects. The similarityparameter varies from 0 to 1. Clusters are classified as apt and inapt on each task dispatch, so that those apt canmake the dispatched task, and ones considered inapt continuein an idle state saving resources for the next task. Also,considering the static scenario and transmission range of theobjects, they form clusters with a non-deterministic number ofparticipants. We also compare ACADIA with the CONTASKIsystem [14] to analyze both performance in the task allocation.We assess the two systems with the following metrics basedon [3]: number of clusters (NC) , Number of unallocatedtasks (NUT) , number of allocated tasks (NAT) , clustersapt to perform tasks (CPT) , clusters inapt to performtasks (CIT) , latency of task accept time (LAT) and, energyconsumption (EC) . All results correspond to the average of35 simulations with a confidence interval of 95%.5 . Results Fig. 3 (striped) shows the ACADIA performance for sup-porting the cluster formation, lighter striped bars represent theCPT and darker striped bars are the maximum. The CPT valuerelates to the similarity among objects that meets each capabil-ity set and the capabilities set of their neighbors, creating thusa consensual relationship between common objects and leadersto perform sensing tasks. ACADIA achieved an average CPTclose to the average NC in most of the scenarios, showing thatall clusters were apt to perform at least one of the dispatchedtasks. In the scenario with 50 and 100 objects with either D-A and D-B the NC remained close to 4, and the CPT closelyfollows that average. The 150 objects scenario showed an NCof 6 for both demands. However, the CPT value had an averageof 3.8 clusters, reaching up to 5 clusters. But, as it can be seenlater, most of the tasks were performed. As for CONTASKI,it achieved higher CPT values, 4, 2.6 and 5.4, close to theaverages of 5, 3 and 6. However, it did not translate in higherNAT values making explicit its inefficiency in task allocation.
ACADIA - Dem. A ACADIA - Dem. B CONTASKI
50 100 150 50 100 150 50 100 150024681012141618
Number of objects
NAT - Dem. ATotal - Dem. ACPT - Dem. ANC - Dem. A NAT - Dem. BTotal - Dem. BCPT - Dem. BNC - Dem. B NAT - CONT.Total - CONT.CPT - CONT.NC - CONT.
Fig. 3: Apt Clusters (CPT) & Task allocations (NAT)Fig. 3 (non-striped) exhibit the amount of sensing tasksdispatched by AP. ACADIA’s clusters were able to perform98% of dispatched tasks on average, with some rounds per-forming 100% of them. But, ACADIA could not sustain 100%of allocated tasks in all simulation rounds due to the randomcapabilities assigned to each object. Scenarios with
D-A hadover 98% of allocated tasks and scenarios with
D-B had over86% of allocated tasks. In contrast, CONTASKI allocated100% of tasks for the scenario with 50 objects and for othersit only allocated 60% of them, even though it had higher CPTvalues, revealing, again, inefficient task allocation. ACADIAwas able to allocate around 100% tasks, in all scenarios,obtaining 40% more of NAT than CONTASKI. Moreover, it isnoteworthy that CONTASKI dispatches only one task, whereasACADIA dispatches 2 and 4 tasks simultaneously.Fig. 4a shows the graphs of the latency of task accept timethat quantifies the difference between the task dispatch timeand the last accept time as seen by the AP. ACADIA’s
D-A and
D-B , with 50 objects, achieved similar LAT times around , given they had the same CPT. The scenarios with 100and 150 objects had the most variations between the demandsdue to factors mentioned previously. With 100 objects,
D-A achieved and
D-B , . In addition, with 150 objectsLAT achieved and , respectively. Such variationsobserved are associated to the distance between objects andthe AP, as well as the time taken by leaders to check it out iftheir capabilities are compatible to make the demand. Further,CONTASKI shows poor performance in LAT times, since itsstandard deviation is so high, the error bars reach values belowzero (not showed in the graph). ACADIA, in its turn, achievedconsistent LAT times in all scenarios. CONTASKI’s LATtimes are , , and . ACADIA’s higher LAT times– in the order of 50% higher than CONTASKI with 50 objectsto almost four times with 100 objects – however, translated ineffective task allocation, as showed in NAT analysis. Number of objects La t en cy ( m s ) Demand ADemand BCONTASKI (a) LAT
Nº of objects E ne r g y ( J ) Demand ADemand BCONTASKI (b) EC
Fig. 4b shows the graphs about ACADIA’s and CON-TASKI’s energy consumption for the task allocation manage-ment, only including energy spent on message exchange. Itis noteworthy that ACADIA remained stable in energy con-sumption with little variation between the demands. ACADIAspent , and J in scenarios with 50, 100 and 150objects in
D-A . The
D-B spent , and J, exhibitingless than 25% variation between the demands. While ACADIAexhibits a predictable trend in EC, CONTASKI displayed lessenergy consumption in all rounds. It spent , and J exhibiting 80% less consumption then ACADIA’s. HoweverACADIA’s higher consumption translated close to 100% oftasks allocated in most rounds.VI. C
ONCLUSION
This paper presented ACADIA for multiple tasks allocationon objects in an IIoT-Health network. It organizes the IIoTnetwork into clusters based on the similarity of the capabilitiesof the objects and the neighboring objects. The mechanismapplies the relational consensus to manage and distributetasks between the clusters, considering the capabilities theyinform. Results show the effectiveness of ACADIA in taskallocation among IIoT objects. ACADIA was also compared toanother task allocation mechanism, showing its effectivenessin multiple task allocation. As future work, we intend toevaluate scenarios with different types of mobility, priority inthe task execution by the clustering and security concerns.A
CKNOWLEDGMENT
We would like to acknowledge the support of the BrazilianAgency CNPq, grants
EFERENCES[1] J. Gubbi, R. Buyya, S. Marusic, and M. Palaniswami, “Internet of things(iot): A vision, architectural elements, and future directions,”
Futuregeneration computer systems , vol. 29, no. 7, pp. 1645–1660, 2013.[2] A. Botta, W. De Donato, V. Persico, and A. Pescap´e, “Integration ofcloud computing and internet of things: a survey,”
Future generationcomputer systems , vol. 56, pp. 684–700, 2016.[3] E. A. Khalil, S. Ozdemir, and A. A. Bara’a, “A new task allocationprotocol for extending stability and operational periods in internet ofthings,”
IEEE Internet of Things Journal , 2019.[4] Z. Ghanbari, N. J. Navimipour, M. Hosseinzadeh, and A. Darwesh,“Resource allocation mechanisms and approaches on the internet ofthings,”
Cluster Computing , pp. 1–30, 2019.[5] L. D. Xu, E. L. Xu, and L. Li, “Industry 4.0: state of the art and futuretrends,”
International Journal of Production Research , vol. 56, no. 8,pp. 2941–2962, 2018.[6] W. Qin, S. Chen, and M. Peng, “Recent advances in industrial internet:insights and challenges,”
Digital Communications and Networks , vol. 6,no. 1, pp. 1–13, 2020.[7] J. J. P. C. Rodrigues, D. B. D. R. Segundo, H. A. Junqueira,M. H. Sabino, R. M. Prince, J. Al-Muhtadi, and V. H. C. D.Albuquerque, “Enabling technologies for the internet of health things,”
IEEE Access , vol. 6, pp. 13 129–13 141, 2018. [Online]. Available:https://doi.org/10.1109/access.2017.2789329[8] M. Aazam, M. St-Hilaire, C.-H. Lung, I. Lambadaris, and E.-N. Huh,“Iot resource estimation challenges and modeling in fog,” in
FogComputing in the Internet of Things . Springer, 2018, pp. 17–31.[9] Y. Yuehong, Y. Zeng, X. Chen, and Y. Fan, “The internet of things inhealthcare: An overview,”
Journal of Industrial Information Integration ,vol. 1, pp. 3–13, 2016.[10] M. Z. Hasan and H. Al-Rizzo, “Task scheduling in internet of thingscloud environment using a robust particle swarm optimization,”
Concur.and Computation: Pract. and Experience , vol. 32, no. 2, p. e5442, 2020.[11] A. M. Rahmani, T. N. Gia, B. Negash, A. Anzanpour, I. Azimi, M. Jiang,and P. Liljeberg, “Exploiting smart e-health gateways at the edgeof healthcare internet-of-things: A fog computing approach,”
FutureGeneration Computer Systems , vol. 78, pp. 641–658, 2018.[12] V. Pilloni and L. Atzori, “Consensus-based resource allocation amongobjects in the internet of things,”
Annals of Telecommunications , vol. 72,no. 7-8, pp. 415–429, 2017.[13] G. Colistra, V. Pilloni, and L. Atzori, “Task allocation in group ofnodes in the iot: A consensus approach,” in . IEEE, 2014, pp. 3848–3853.[14] C. Pedroso, Y. U. de Moraes, M. Nogueira, and A. Santos,“Managing consensus-based cooperative task allocation for IIoTnetworks,” in . IEEE, Jul. 2020. [Online]. Available:https://doi.org/10.1109/iscc50000.2020.9219635[15] V. Pilloni and L. Atzori, “Deployment of distributed applications inwireless sensor networks,”
Sensors , vol. 11, no. 8, pp. 7395–7419, 2011.[16] M. Kim and I.-Y. Ko, “An efficient resource allocation approach basedon a genetic algorithm for composite services in iot environments,” in . IEEE, 2015,pp. 543–550.[17] Y. Jin, J. Jin, A. Gluhak, K. Moessner, and M. Palaniswami, “Anintelligent task allocation scheme for multihop wireless networks,”
IEEETransactions on Parallel and Distributed Systems , vol. 23, no. 3, pp.444–451, 2011.[18] O. B. Abderrahim, M. H. Elhedhili, and L. Saidane, “Ctms-siot: Acontext-based trust management system for the social internet of things,”in , 2017, pp. 1903–1908.[19] L. D. Xu and L. Duan, “Big data for cyber physical systems in industry4.0: a survey,”
Enterprise Information Systems , vol. 13, no. 2, pp. 148–169, 2019., vol. 13, no. 2, pp. 148–169, 2019.