CMIP: Clone Mobile-agent Itinerary Planning Approach for Enhancing Event-to-Sink Throughput in Wireless Sensor Networks
Huthiafa Q Qadori, Zuriati Ahmad Zukarnain, Zurina Mohd Hanapi, Shamala Subramaniam, Mohamed A. Alrshah
RReceived October 7, 2018, accepted November 7, 2018, date of publication November 19, 2018, date of current version December 18,2018.
Digital Object Identifier 10.1109/ACCESS.2018.2882018
CMIP: Clone Mobile-agent ItineraryPlanning Approach for EnhancingEvent-to-Sink Throughput in WirelessSensor Networks
HUTHIAFA Q QADORI , ZURIATI AHMAD ZUKARNAIN ,(Member, IEEE), ZURINA MOHDHANAPI , (Member, IEEE), SHAMALA SUBRAMANIAM (Member, IEEE), and MOHAMED A.ALRSHAH (Senior Member, IEEE) Department Communication Technology & Network, Faculty of Computer Science & Information Technology, Universiti Putra Malaysia, Serdang 43400,Malaysia Sports Academy, Universiti Putra Malaysia, Serdang 43400 UPM, Malaysia
Corresponding authors: Huthiafa Q Qadori ([email protected]) and Zuriati Ahmad Zukarnain ([email protected])This work is supported by Universiti Putra Malaysia, the Ministry of Higher Education-Iraq (University of Anbar)
ABSTRACT
In order to mitigate the problem of data congestion, increased latency, and high energyconsumption in Wireless Sensor Networks (WSNs), Mobile Agent (MA) has been proven to be a viablealternative to the traditional client-server data gathering model. MA has the ability to migrate amongnetwork nodes based on an assigned itinerary, which can be formed via Single Itinerary Planning (SIP)or Multiple Itinerary Planning (MIP). MIP-based data gathering approach solves problems associated withSIP in terms of task duration, energy consumption, and reliability. However, the majority of existingMIP approaches focus only on reducing energy consumption and task duration, while the Event-to-sinkthroughput has not been considered. In this paper, a Clone Mobile-agent Itinerary Planning approach(CMIP) is proposed to reduce task duration while improving the Event-to-sink throughput in real-timeapplications, especially when the MA is assigned to visit a large number of source nodes. Simulationresults show that the CMIP approach outperforms both Central Location-based MIP (CL-MIP) and GreatestInformation in Greatest Memory-based MIP (GIGM-MIP) in terms of reducing task duration by about 56%and 16%, respectively. Furthermore, CMIP improves the Event-to-sink throughput by about 93% and 22%as compared to both CL-MIP and GIGM-MIP approaches, respectively.
INDEX TERMS
Data gathering, Clone, Mobile Agent, Static Itinerary, Dynamic Itinerary, MIP
I. INTRODUCTION
The merging of low-cost wireless communication,computation, and sensing have spawned a new generationof small and cheap intelligent devices. The deployment oftens to thousands of these tiny devices in self-organizingnetworks has established a new network type knownas Wireless Sensor Networks (WSNs). A WSN can begenerally described as the deployment of a vast number oftiny sensor nodes that connect to each other via wirelesscommunication. The advantages of these sensor nodes;such as easy installation, low cost, small size and lowpower consumption; make this type of networks very usefulfor many applications in different fields of life such as agriculture, industrial automation, transportation, healthcare, and military [1].The main objective of WSN deployment is to gatherenvironmental data and deliver it to end users [2]. Typically,in a WSN, sensor nodes are deployed in a field of interestwhere every sensor node is able to perform several taskssuch as sensing, surveillance, environmental monitoring,processing and storing the sensed data. The sensed data atany sensor node needs to be transmitted to a collection pointin the network, called a processing unit or sink, for furtheranalysis. The transmission of data is done via single ormulti-hop wireless communication. This type of transmissionis based on the conventional client-server model, where every
VOLUME 4, 2016 a r X i v : . [ c s . N I] F e b uthiafa Q Qadori et al. : Preparation of Papers for IEEE Access involved sensor node transmits data packets to the sink.The WSNs have a very limited link bandwidth comparedto wired networks, a set of connected wireless links needto be established from those nodes to the sink in orderto forward data packets. Consequently, this could increasenetwork data traffic and, as a result, it could increase theconsumption of network resources. As a solution to thisoverwhelming data traffic, [3] has proposed an MA-modelfor data gathering in WSNs. In this model, a softwarecomponent, called MA, is forwarded throughout the networkand to visit source nodes one by one following an itineraryto perform data gathering process locally at each sourcenode. Upon completion of data gathering, the MA bringsback the aggregated data to the sink. Consequently, this datacombination decreases the number of packets transmittedto the sink, which results in a decrease in communicationcosts, bandwidth utilization, network congestion, and energyconsumption.In WSNs, MA-based data gathering employs two schemesof the itinerary; Single Itinerary Planning (SIP) and MultiItinerary Planning (MIP). SIP utilizes single MA to roamnetworks for data gathering [4]–[7], while MIP utilizesmultiple MAs that work in parallel to perform data gathering[8], [9]. In large-scale networks, SIP introduces manydrawbacks [10] such as long delays, increase in MA packetsize, low reliability, and high probability of MA packet loss,due to the migrating to a large number of source nodes.To overcome these drawbacks, MIP [8], [9] was proposed,in which several MAs are distributed into the network tovisit groups or partitions of source nodes concurrently, whereeach MA is assigned to only one group of source nodes.This process reduces the MA packet size, which further leadsto a lower energy consumption compared to SIP schemes.Moreover, the task duration in MIP is minimized, due to thedistribution of concurrent aggregation tasks among multipleMAs.As known, the MA itinerary is the source-visitingsequence that the MA needs to follow during its migration,where the MA migration itinerary planning is still achallenging issue in MA-based data gathering. In spite ofMIP advantages, it worth noting that finding an optimalitinerary for the MA is a Nondeterministic Polynomial(NP-complete or NP-hard) problem [7], [11]. A sub-optimalMA itinerary may lead to a highly inefficient overall networkperformance.In MA-based data gathering, the determination of MA’sitinerary can be classified as static, dynamic or hybriditinerary [11], [12]. In static itinerary, the sequence of sourcenodes that will be visited by MA is calculated at the sinkbefore the MA starts its migration, while in the dynamicitinerary, the sequence of source nodes that will be visitedby MA is determined on the fly at each source node. Inhybrid itinerary, the source nodes to be visited are selectedat the sink, but the visiting sequence is computed on the flyby the MA. However, dynamic or hybrid itineraries consumevaluable node energy resources and employ larger MA. This is because the MA needs to carry the next hop computationcode to be executed at each node during the migrationprocess. On the other hand, static itinerary consumes lessenergy compared to dynamic or hybrid itinerary since theMA carries only a pre-determined itinerary that has beencalculated at the sink.The applications of WSNs are designed to accomplishspecific objectives or desired tasks. These objectives or tasksare defined and optimized according to the application user.The main goal of WSNs is to deliver the sensed data tothe relevant processing unit (sink) for further analysis anddecision making, where the collected data must reflect thecurrent state of the targeted environment. In most cases, thecollected data is valid only for a limited period of time dueto the constant changes in the monitored environment. Thus,the importance of these data changes from accurate to lessaccurate until it becomes totally inaccurate to reflect thecurrent state of the environment as time progresses [13]. Inreal-time WSNs’ applications, it is very crucial to deliverthe collected data to the sink without delay to make timelyactions or decisions. Moreover, collecting as much data aspossible is also very important to the WSN application userto allow him or her to improve the quality of evaluation andanalysis to make precise decisions [14].Indeed, MIP approaches focus only on reducing energyconsumption and task duration, while neglecting the amountof delivered data to the sink within the time period. However,collecting as much data as possible requires the distributedMAs to visit a larger number of source nodes, whichdefinitely will increase the delay. To solve this issue, a MIPapproach that is able to minimize task duration and maximizecollected data is highly needed to support the real-timeapplications. However, this solution will be at the expenseof energy efficiency.This paper proposes a novel MIP-based data gatheringapproach in WSNs, namely CMIP, which aims to reducetask duration of data gathering process while maximizingthe volume of collected data. The remainder of the paperis organized as follows: Section II explains the relatedworks, and Section III presents the proposed CMIP approach,while Section IV presents the simulation setup. Further,the performance evaluation and experimental results areexplained in Section V. Finally, Section VI concludes thework and presents the future directions for the potentialimprovements. II. RELATED WORKS
In the last few years, several MIP-based data gatheringapproaches have been proposed. The main objective of thoseapproaches was to reduce the energy consumption and taskduration. This section reviews several MIP-based approachesproposed for data gathering, discusses their basic concepts,and highlights their advantages and shortcomings.In [15], a Near-Optimal Itinerary Design (NOID)algorithm was proposed to address the problem associatedwith calculating the number of near-optimal routes for VOLUME 4, 2016 uthiafa Q Qadori et al. : Preparation of Papers for IEEE Access
MAs. NOID algorithm adapts the Esau-Williams heuristicmethod [16] that were designed for the ConstrainedMinimum Spanning Tree (CMST) problem in networkdesigning. NOID algorithm iteratively groups sensor nodesin the network to separate sub-trees that are connectedprogressively to the Processing Element (PE) or sink. Finally,each sub-tree is assigned to an individual MA, where thenumber of needed MAs is equal to the number of trees inthe relevant network.Later, the Tree-Based Itinerary Design (TBID) algorithm[17] was proposed, which outperformed NOID algorithm interms of low-cost itineraries. TBID not only determines theoptimal number of MAs but also creates low-cost itinerariesfor each individual MA, by partitioning the area around thePE/sink into concentric zones. The number of nodes lyingwithin the radius of the first zone of PE/sink represents thestarting points of itineraries for the needed MAs. The radiusof the first zone can be calculated as ar max , where r max is the maximum transmission range of any sensor node andthe input parameter a ∈ (0 , . The subsequent zones widthis calculated as r max / , such that each sensor node in anyzone can only communicate with sensor nodes belonging tothe previous, current, and next outer zones. In each iteration,MAs itineraries start from inner zones and proceed to outerzones connecting the pair of adjacent sensor nodes u and v ,where u is linked to an itinerary and v is not yet linked. Thisminimizes the edge Potential Cost (PC), which is equal tothe cost of an itinerary derived in the case of incorporatingthe edge in the itinerary. This ensures low-cost itineraries tobe created for each individual MA. The end output of TBIDis k trees, where each tree is rooted at the nodes lying withinthe radius of the first zone of PE/sink. The itinerary of eachMA in these trees is derived by a post-order traversal.[18] introduced a novel algorithm for the energy-efficientitinerary planning of MAs. This algorithm adopts ameta-heuristic method called Iterated Local Search (ILS)to derive the hop sequence of multiple traveling MAsover the deployed source nodes. Like other tree-based MIPalgorithms, ILS determines the number of itineraries (MAs)by considering a circular zone around the sink. The nodesthat are lying in the sink zone will be the starting points ofeach MA itinerary. However, as opposed to other tree-basedMIP algorithms, ILS iteratively examines the energy cost ofpotential attachment of each candidate node u (among thosethat remain unconnected) with any pair of already attachedsubsequent nodes. Such that the minimum itinerary costamong the examined candidate nodes will be added to thecurrent MA’s itinerary. As a result, the ILS approach buildslow energy cost of MAs’ itineraries.Although NOID, TBID, and ILS perform better than SIPapproaches, the MA in these tree-based schemes consumestwice as much energy due to the reverse routes that the MAtake, especially when there are many branches. This resultsin an increase in the energy consumption and task duration ofMA’s migration.The central location-based MIP (CL-MIP) is another algorithm proposed by [19], where the determination of theoptimal number of MAs can be divided into four parts;(1) Visiting Central Location (VCL) selection algorithm,(2) Source grouping algorithm, (3) Source-visiting orderdetermination SIP-based algorithm, (4) Iterative algorithm;to ensure that all source nodes have been assigned to theirMAs. CL-MIP groups all source nodes according to the nodedensity (gravity algorithm). The basic idea of VCL algorithmis to distribute each source node’s impact factor to othersource nodes. Let n represent the source node number; theneach source node will receive ( n - 1) impact factors fromother source nodes, and one from itself. At each iteration,the location of the source node with the highest accumulatedimpact factor will be selected as a VCL. Then, all sourcenodes within the radius of VCL are grouped in a clusterand assigned to an MA. The above process repeats until allremaining source nodes in the network are assigned to theirMAs, then, the itinerary for each MA can be planned by a SIPalgorithm. In CL-MIP, the itinerary for each distributed MAis determined by local closest first (LCF) [4]. In LCF, the MAlooks for the next hop node with the shortest distance fromthe current location. However, CL-MIP algorithm assumesthat the relevant source nodes are geographically distributedin several clusters, which limits the use of this algorithm in abroad range of applications.In [20], the author proposed an Optimal Multi-agentsItinerary Planning (OMIP) algorithm similar to CL-MIP,where the source nodes are grouped into clusters. OMIPadopts Efficient Clustering Routing Protocol (ECRP) topartition the network into v clusters, select a medoid node M N i in each cluster, and then minimizes the averagedistance between
M N i and other nodes in the cluster. The
M N i in OMIP functions as the cluster-head as well as thestarting and arrival node of the MA. In OMIP, the sinkdispatches an MA to every cluster in the network. The MAstarts its data gathering process from
M N i , roams the sourcenodes within the cluster and returns back to the
M N i . Afterthe MA completes data gathering, it travels back to the sinkwith the aggregated data. OMIP approach suffers from thesame drawbacks of CL-MIP, where the clustering methodhighly depends on the distribution of source nodes. OMIPgenerates clusters with few source nodes to be visited bythe MA. As a result, MAs’ itineraries becomes imbalancedwhich directly affects the network performance in terms ofenergy consumption and task duration.Multi-mobile Agent itinerary planning-based Energy andFault aware (MAEF) was proposed by [21]. Similar to(OMIP), MAEF also groups the source nodes into clustersand each cluster has a Cluster Head (CH). It adopts the sameidea presented in CL-MIP in which a distribution of densityimpact factor is used to select the CHs. Also, all sourcenodes, which lie within the maximum transmission range ofeach selected CH, are grouped together. MAEF differs fromCL-MIP and OMIP, as the MAs’ itineraries are planned onlyamong the CHs and determined using the MST. Furthermore,the number of itineraries are equal to the number of nodes
VOLUME 4, 2016 et al. : Preparation of Papers for IEEE Access located within the sink’ transmission range. In MAEF, oncethe MA reaches the CH for the first time, it notifies the CH’ssource nodes to send their data to the CH. When the MAreaches the last CH in its itinerary, it starts gathering datafrom CHs on its way back to the sink.Although the MAEF algorithm deploys a new MIPstrategy, where the distributed MAs visit only CHs, it hastwo main drawbacks. First, the CHs’ transmission range(which used to group the source nodes into clusters) cangenerate overlap between the clusters which can increase theoverhead. Second, each MA needs to traverse its itinerarytwice; to notify the CH’s source nodes to send data, and togather data from the CHS. This would result in an increase ofenergy consumption and task duration.In most of the aforementioned MIP approaches, thegeographic information of the sensor nodes is the mainparameter used to determine the optimal number of MAsand their itineraries. In [22], the Greatest Information inthe Greater Memory-based MIP (GIGM-MIP) algorithm wasproposed. The GIGM-MIP algorithm considers geographicinformation to partition the network and takes into accountthe data size in each partition to formulate the optimalnumber of MAs.In GIGM-MIP, the data size in each partition plays animportant role in determining the number of MAs. Afterpartitioning the network, GIGM-MIP calculates the data sizeof the source nodes in each partition and starts with an initialnumber of MA. Then, GIGM-MIP updates this number byassigning source nodes to the MAs that have free payloadmemory size, such that each partition may have more thanone MA. Using this strategy, GIGM-MIP approach balancesthe data load among distributed MAs, thereby reducing taskduration. However, exchanging the assigned source nodeamong the MAs results in long MA itinerary, especiallywhen the location of the assigned source node is far fromthe current MA’s location. Moreover, distributing more thanone MAs to one partition means that each distributed MAshould carry its processing code (aggregation code). MultipleMAs carrying the same aggregation code within a singlepartition would result in an increase in the MA’s migrationhops, which subsequently increases the energy consumptionof the network.In the above-reviewed MIP approaches, many solutionshave been proposed to determine the number of MAsand their itineraries, which is the main challenging issueassociated with MIP paradigm. Majority of the proposedMIP approaches are focusing on minimizing the energyconsumption and task duration, while the amount of thedata delivered to the sink within the time period is omitted.In some WSNs applications (e.g. real-time applications) aMIP approach that is able to minimize the task duration andmaximize the data collected at the sink is really needed evenif it is not optimizing the energy consumption. The proposedCMIP approach in this paper adopts the cloning concept ofthe MA to reduce the task duration of data gathering processwhile maximizing data collection.
III. CMIP APPROACH
This section presents the proposed CMIP, which includes twoparts, the MA’s cloning mechanism and the mechanism ofdetermining the itinerary used by MA(s) to visit the sourcenodes.
A. CLONING MECHANISM OF MA IN CMIP
As mentioned in [8], [19], the task duration in MIP is thedelay experienced by the last MA, which returns to the sinknode. In other words, this MA has the longest delay amongthe other dispatched MAs. In [5], the delay of MA’s itineraryto complete its data gathering task ( T MA ) can be expressedas follows: T MA = T p + T roam + T back , (1)where T p is the delay of the MA migration from the sink tothe first source node, T roam is the MA migration delay fromthe first source node to the last source node, and T back is theMA migration delay from the last source node to the sink.The T p can be simplified as below: T p = (cid:16) pcD r + t ctrl (cid:17) × H, (2)where pc is the MA’s size (processing code plus MA packetheader), D r is the data rate at MAC layer, t ctrl is the delay ofcontrol messages and H is the number of hops between thesink and the first source node. The T roam can be expressedas follows: T roam = M (cid:88) i =1 (cid:16) t + S data D p + S ima + pcD r + t ctrl (cid:17) × H, (3)where M is the number of source nodes, t is the MA accessdelay (the time required to mount MA’s processing code inthe target source node), S data is the size of sensed data atthe source node, D p is the data processing rate, and S ima isthe size of the aggregated data carried by the MA at sourcenode i . The T back can be decomposed into: T back = (cid:16) S Mma + pcD r + t ctrl (cid:17) × H, (4)where S Mma is the size of MA at the last source node.Based on Equation 3, the delay of MA is highly associatedwith the number of source nodes and the number of MA’shops. The more source nodes to be visited by MA, thelarger MA’s hops which consequently implies a higher delay.Additionally, when the MA visits a large number of sourcenodes, its data payload size becomes large, which can furtherincrease the MA delay.The time required by an MA to collect data from a largenumber of source nodes can be minimized by dispatchingtwo MAs since each MA will be assigned to fewer sourcenodes. However, this strategy would lead to increased energyconsumption due to the large number of hops utilized bythe two separated itineraries. Moreover, based on Equation2, using two separated itineraries would mean each MAis required to carry its own processing code (aggregation VOLUME 4, 2016 uthiafa Q Qadori et al. : Preparation of Papers for IEEE Access
FIGURE 1: MA itinerary planing based on: (a) GIGM-MIP; (b) CMIPcode), which further leads to increase energy consumption.This excessive increase in energy consumption would lead toan increase in the probability by which MA could drop itsaccumulated data. Therefore, there is a need for a solutionthat distributes the data collection task while reducing thetask duration and maximizing the data collection.To mitigate the above-mentioned issues, this paper adoptthe MA cloning concept [23]–[25]. In this concept, the MAhas the ability to clone itself at a certain point resultingin an instance with the same capacities and capabilities. Inthe proposed CMIP, one MA is dispatched from the sink,where this MA then clones itself at a certain point (namedcloning point). Figure 1 illustrates the MA itinerary planningbased on the cloning mechanism in our proposed CMIP ascompared to GIGM-MIP approach.In Figure 1, there are seven source nodes (the red nodes , , , , , , ) to be visited by the MA. In GIGM-MIP(Figure 1(a)), the sink has determined one MA to visit theseven source nodes. Using LCF algorithm, the visiting orderis such that MA begins from source node , , , , , and finishes at source node . It is clear that the MA’s datapayload would have become large after the MA completesdata collection from source node . With this large MA’s datapayload, the MA needs to migrate 5 hops until it reaches thesink, which increases its delay. On the other hand, Figure 1(b)shows the MA itinerary planing based on cloning mechanismin CMIP, where the source nodes will be separated into twogroups. One group will be assigned to the Main MA (MMA),the MA that has the ability to clone itself, while the secondgroup will be assigned to the Cloned MA (CMA). At theinitial stage, the sink determines the Farthest Source Node(FSN), the source node located at the farthest distance fromthe sink. This FSN can be calculated using Equation (5), asbelow: F SN = (cid:113) ( CSN x − Sink x ) + ( CSN y − Sink y ) , (5)where ( CSN x , CSN y ) represent the position of a sourcenode points in MA’s itinerary, and ( Sink x , Sink y ) determinethe position of the sink. The FSN is selected to be the cloning point, where theMMA clones itself (for example node number in Figure1(b)). At the FSN, both MMA and CMA will start datacollection from the source nodes. The reason behind selectingthe FSN to be the cloning point because the operation of datagathering from the farthest source node first towards the sinkconsumes less energy and time as compared to the one startsfrom the nearest source node (more details are discussed inthe following section). B. MA’S ITINERARY DETERMINATION IN CMIP
The delay and energy consumption of an MA itinerarydepends on the efficiency of source nodes visiting order.In GIGM-MIP, this visiting order is determined by theLCF algorithm, which selects the nearest source node tothe current MA location as the next MA’s hop. However,selecting the closest source node as the next MA’s hop doesnot always ensure the optimal solution. For example, inFigure 1(a), the output of the LCF algorithm shows that theMA needs to migrate 5 hops after data collection until itreaches the sink. In this context, 5 hops will be at the costof high energy dissipation and will increase the delay sinceit carries a large amount of data over many hops. Indeed,it will be more efficient if the nearest source node to thesink becomes the last destination of MA (before reaching thesink). Therefore, in CMIP, the source nodes’ visiting orderis first determined by the LCF algorithm, then the outputitinerary is reversed, such that the last source node in theitinerary will be visited first.The energy consumption and delay between two sourcenodes is proportional to the hop count between them withrespect to the size of MA. Therefore, the hop count betweenthese two source nodes is the key metric to evaluate the MA’smigration cost in terms of energy consumption and delay. Inorder to evaluate the MA’s migration cost in the proposedCMIP, Figure 2 illustrates the comparison of cost betweenLCF itinerary and reversed itinerary. In this figure, there arefour source nodes to be visited by the MA. In Figure 2(a), thesource nodes’ visiting order determined by LCF is , , and VOLUME 4, 2016 et al. : Preparation of Papers for IEEE Access
FIGURE 2: A comparison of MA itinerary cost: (a) LCFitinerary; (b) Reversed itinerary , where source node will be visited first. Let assume thatthe size of the MA is 1 when dispatched from the sink and itwill increase by 1 after each source node visit. By multiplyingthe current size of MA by the number of hops, the cost of theMA’s itinerary can easily be calculated. The results show thatthe total cost of the LCF itinerary is 20 when source node is visited first. In contrast, from Figure 2(b), the total cost ofthe reversed itinerary is 16. Thus, the reversed itinerary canfurther decrease the energy consumption and delay.In the proposed CMIP, if the FSN is the last source nodein the itinerary, the whole itinerary is reversed and there is nocloning process. On the other hand, when the FSN is not thelast source node, the reversing procedure will be different.The first part of the itinerary from the sink to the FSN willbe reversed where the second part from the FSN to the sinkwill remain unchanged. Then, each part of the itinerary willbe assigned to an individual MA, where all MAs worksconcurrently. Figure 1(b) above demonstrate an example ofthis reversed process. In this figure, the FSN is the sourcenode , so the itinerary from the sink to the FSN is reversedwith a visiting order , , , , till it reaches the sink. Onthe other hand, the second part of the itinerary, which beginsfrom to , and , till it reaches the sink, will remainunchanged. This ensures that the data collection performedby the MA always starts from the farthest source nodeand proceeds towards the nearest nodes to the sink, whichminimizes the delay and energy consumption. Algorithm 1details the pseudo-code of the proposed CMIP approach. C. TASK DURATION IN CMIP
As known, the task duration of MIP approach equals thedelay of the MA from the dispatching time to it returns to thesink. However, since the CMIP consists of MMA and CMA,the task duration equals the delay from the dispatching timeof the MA to the arrival of both MMA and CMA to the sink,whichever is longer. Additionally, since the itinerary of theMMA begins and ends at the sink node and it is also based onEquation (1), the total delay of its itinerary can be calculatedas follows: T MMA = T MMAp + T MMAroam + T MMAback , (6) Algorithm 1:
Pseudo-code of CMIP approach Notation: F SN ← is the farthest source node from sink SN S ← denotes the set of source nodes to be visitedby MA N um iti ← denotes the number of MA itineraries f ( F SN, sink ) ← is a function that returns theshortest path between the sink and F SN f ( SN S , F SN ) ← is a function that returns the nextsource node determined by LCF Algorithm F SN S P ← denotes the MA hop sequencedetermined by f ( F SN, sink ) V SN S ← denotes the source nodes’ visiting orderdetermined by f ( SN S , F SN ) Initialization: N ← Number of sensor nodes M A
DP T ← The threshold value of the main MA datapayload Partitioning the network using k-means algorithmby calculating the distance among N: K ← Number of partitions (specified by user as inGIGM-MIP) p ← Set of source nodes in each partition for j = 1 to K do S ← Number of source nodes in set p N umM As ← Number of MAs in set p for m = 1 to N umM As do Determine the MA itinerary using LCFAlgorithm Find
F SN from the sink using Equation (5) if F SN is not the last SN in the MA’sitinerary then Split the itinerary into two itineraries from
F SN for each splited itinerary do F irSN ← F SN F SN S P ← f ( F SN, sink ) V SN S ← f ( SN S , F SN ) add F SN S P to V SN S end else Reverse the MA itinerary determined byLCF Algorithm end end N um iti ← Reversed and splited MA itineraries end Return
N um iti where T MMAp is the delay of the MMA migration from thesink to the FSN, and T MMAroam is the MMA migration delayfrom the FSN to the last source node, and T MMAback is theMMA migration delay from the last source node to the sink.As for the case of the CMA, since its itinerary begins at VOLUME 4, 2016 uthiafa Q Qadori et al. : Preparation of Papers for IEEE Access the FSN and ends at the sink, its total task duration can becalculated as follows: T CMA = T MMAp + T CMAroam + T CMAback , (7)where T MMAp is the delay incurred due to the MMA’smigration from the sink to the FSN plus the delay from FSNto the first source node in CMA’s itinerary. T CMAroam is theCMA migration delay from the first source node to the lastsource node, and T CMAback is the CMA migration delay fromthe last source node to the sink. Noting that the delay ofthe T MMAp is added to the delay of CMA in Equation (7)because the delay incurred to reach the FAN is the same forboth MMA and CMA.
IV. SIMULATION SETUP
The proposed CMIP approach has been implemented andtested on a simulation developed via MATLAB R2017b(student version). The same network model used in [8], [19],[22], [26] is adopted which is the most popular networkmodel in data gathering-based MIP. We used the same energyconsumption model as in [8], [27]. More specifically, alarge-scale network, consisting of 800 static nodes denseand uniformly deployed, is considered in the experiments inorder to validate the scaling of the proposed CMIP approach.Each node has a transmission range of 60 meters. The sinknode has a continuous energy supply and it is located at thecenter of the network. The generated MAs’ itineraries arestatically predetermined at the sink node before the MAsare dispatched to the network. In each data gathering task,a random number of source nodes, varying from 10 to 80 bythe step of 5, is selected. Each compared MIP algorithm wastested with the same number of selected source nodes. Thesimulation parameters are listed in Table 1.TABLE 1: Simulation parameters
Network Parameters Value
Size of network 1000 m ×
500 mNumber of deployed nodes 800Number of source nodes 10 - 80Transmission range 60 mRaw data size 2048 bits
MA Parameters Value
MA processing code 1024 bitsMA accessing delay 10 msRaw data reduction ratio 0.8Aggregation ratio 0.9Data processing rate 50 Mbps
V. PERFORMANCE EVALUATION
This section evaluates the performance of the proposed CMIPin terms of task duration, Event-to-sink throughput, andenergy consumption, where the GIGM-MIP and CL-MIPapproaches are chosen for benchmarking. The performanceevaluation is done based on two scenarios: (1) a scenariowith variable number of source nodes and (2) a scenario withvariable aggregation ratio.
A. IMPACT OF VARYING SOURCE NODES
In this scenario, the number of source nodes is increased from10 to 80 by a step of 5, while keeping all other parameters,as in Table 1, unchanged. In fact, the number of sourcenodes has a direct impact on the performance of the studiedalgorithms in terms of energy consumption, Event-to-sinkthroughput, and task duration because the more the sourcenodes to be visited, the larger the MA size. T a sk D u r a t i on ( s ) Number of Source NodesCL-MIPGIGM-MIPCMIP
FIGURE 3: Task DurationFigures 3, 4, and 5 show the impact of varying thenumber of source nodes on the task duration, Event-to-sinkthroughput and energy consumption, respectively. In Figure3, it is clear that the proposed CMIP achieves the minimumtask duration by about 56% and 16% reduction comparedto CL-MIP and GIGM-MIP approaches, respectively.Moreover, CMIP decreases the task duration linearlycompared to GIGM-MIP when the number of source nodesincreases, as a result of using the cloning mechanism, whichsplits the MA itineraries that have a large number of sourcenodes. This split operation helps each distributed MA to takeless number of source nodes, which in turn helps them tocomplete the data gathering journey much faster. E v en t - t o - S i n k T h r oughpu t ( KB / s ) Number of Source NodesCL-MIPGIGM-MIPCMIP
FIGURE 4: Event-to-Sink ThroughputMoreover, the reduction in task duration noticed in CMIPalso improves the Event-to-sink throughput. In WSNs, theEvent-to-sink throughput E is defined as the number ofpackets received at the sink over the time period T [28]. VOLUME 4, 2016 et al. : Preparation of Papers for IEEE Access
Since the MAs in MA-based data gathering algorithms areresponsible for collecting and delivering the aggregated datato the sink within the task duration, thus E can be calculatedas follows: E = T Data T Duration , (8)where T Data is the total aggregated data that has beensuccessfully delivered to the sink by all MAs and T Duration is the task duration from beginning till the end.Figure 4 shows that the CMIP is able to outperform thecompared schemes (GIGM-MIP and CL-MIP) by attaininghigher Event-to-sink throughput due to the ability of theinvolved MAs in CIMP to complete their round-trip ofdelivering the data to the sink with lower task duration. Incontrast, GIGM-MIP and Cl-MIP show lower Event-to-sinkthroughput due to the longer time needed for the MAs toaccomplish their data gathering process. In GIGM-MIP andCl-MIP, some MAs are assigned to a large number of sourcenodes, which in turn increases the number of hops and MAdata payload size that results in task duration increase. T a sk E ne r g y ( J ) Number of Source NodesCL-MIPGIGM-MIPCMIP
FIGURE 5: Energy consumptionFigure 5 presents the energy consumption needed toaccomplish the task of CMIP compared to GIGM-MIPand CL-MIP. In MIP, the task energy consumption is theaccumulated energy spent by all distributed MAs to performthe data gathering task from all source nodes [8], [18]. Itincludes the energy spent on transmitting, receiving, andexchanging control message. Thus, the total task energyconsumption can be calculated as follows: C total = | I | (cid:88) t =1 IC t , (9)where IC t is the energy cost of itinerary I t covered by theMA, and IC t can be simplified to: IC t = | I t | (cid:88) j =1 ( jdf + pc ) c i,j , (10)where | I t | represents the number of visited nodes in theitinerary I t by the relevant MA, j is the visited sensor node, jdf is the size of data collected by the MA at the sensor node j after it is aggregated by a ratio of f , pc is the MA’s initialsize (processing code plus MA packet header), and c i,j is theenergy consumption of the MA to migrate from sensor node i to sensor node j . Noting that j could act as a source node(has data to be collected by the MA) or as an intermediatenode (forwarding node).The proposed CMIP has a compatible performance withGIGM-MIP when there are 10-20 source nodes. However, ithas lower energy consumption compared to CL-MIP scheme.As the number of source nodes increase, CMIP starts toconsume higher energy compared to both GIGM-MIP andCL-MIP schemes because CMIP starts to clone its MA forthe itineraries that have a larger number of source nodeswhich in turn results in a slightly higher number of hops. It isworth noting that the increase in hops does not affect the MAround trip since the data collected by the MMA and CMA iswell distributed. Consequently, the MAs could successfullycomplete their tasks with less delay and high Event-to-sinkthroughput. B. IMPACT OF VARYING AGGREGATION RATIO:
This section evaluates the impact of varying data aggregationratio f on the performance of CMIP. In this scenario,the varied f ratio represents different redundancy andcompression function of the collected data. Thus, differentsizes of collected data can be delivered to the sink, whichhas a direct impact on the proposed CMIP in terms of taskduration, Event-to-sink throughput, and energy consumption.In this experiment, f ratio is varied from 0.1 to 0.9 for 80source nodes. T a sk D u r a t i on ( s ) Aggregation ratioCL-MIPGIGM-MIPCMIP
FIGURE 6: Task DurationFigures 6, 7 and 8 show the impact of varying theratio of f on task duration, Event-to-sink throughput andenergy consumption, respectively. As shown in Figure 6,the proposed CMIP outperforms GIGM-MIP and CL-MIPin terms of task duration, although the size of MA becomeslarger whenever the value of f becomes smaller. This isdue to the cloning strategy deployed by the CMIP, whichfacilitates the collaboration between the involved MAs(MMA and CMA) allowing them to collect larger datacollaboratively in a shorter duration of time. VOLUME 4, 2016 uthiafa Q Qadori et al. : Preparation of Papers for IEEE Access
On the other hand, Figure 7 shows that the Event-to-sinkthroughput of CMIP is higher compared to GIGM-MIP andCL-MIP schemes. Although the size of the collected datachanges on varying f , CMIP still performs better than theother schemes due to the large amount of data collected bythe MMA and CMA within a shorter time, which improvesthe Event-to-sink throughput. As the size of the collecteddata in GIGM-MIP and CL-MIP increases, the MA sizeincreases. Thus, receiving and transmitting an MA with largedata at any hop consumes a significant amount of time, whichincreases the total task duration and directly decreases theEvent-to-sink throughput.Additionally, receiving and transmitting an MA that carriesa large amount of data at any hop consumes more energy,which may shut down that hope due to having a low levelof energy. As a result, the MA with its carried data will belost, which negatively affects the Event-to-sink throughput.However, the increase in energy consumption caused by thelarge size of MA has a negligible impact on the Event-to-sinkthroughput in CMIP, as shown in Figure 8, due to the useof the cloning strategy that deploys two MAs to gatherthe data from a large number of source nodes. Thereby,the collected data will be divided over the distributed MAsallowing them to complete their task with minimum time andhigh Event-to-sink throughput. E v en t - t o - S i n k T h r oughpu t ( KB / s ) Aggregation ratioCL-MIPGIGM-MIPCMIP
FIGURE 7: Event-to-Sink Throughput T a sk E ne r g y ( J ) Aggregation ratioCL-MIPGIGM-MIPCMIP
FIGURE 8: Energy consumption
VI. CONCLUSION
In this paper, a Clone Mobile-agent Itinerary Planning(CMIP) approach has been proposed. The proposed CMIPadopt the cloning concept of MA to reduce the task durationwhen the MA’s itinerary has a large number of source nodesto be visited. The CMIP mitigates this issue by splitting theitinerary into sub-itineraries, each of which is assigned toan individual MA. Moreover, a reversed MA’s itinerary isproposed to further decrease the energy consumption andtask duration. Further, simulation experiments have beenconducted to evaluate the performance of CMIP. The resultsshow that CMIP significantly outperforms the comparedschemes in terms of task duration and Event-to-sinkthroughput.
REFERENCES [1] Daniele Puccinelli and Martin Haenggi. Wireless sensor networks:applications and challenges of ubiquitous sensing. IEEE Circuits andsystems magazine, 5(3):19–31, 2005.[2] Jamal N Al-Karaki and Ahmed E Kamal. Routing techniques in wirelesssensor networks: a survey. IEEE wireless communications, 11(6):6–28,2004.[3] Hairong Qi, Yingyue Xu, and Xiaoling Wang. Mobile-agent-basedcollaborative signal and information processing in sensor networks.Proceedings of the IEEE, 91(8):1172–1183, 2003.[4] Hairong Qi and Feiyi Wang. Optimal itinerary analysis for mobile agentsin ad hoc wireless sensor networks. Proceedings of the IEEE, pages147–153, 2001.[5] Min Chen, Taekyoung Kwon, Yong Yuan, Yanghee Choi, and VictorLeung. Mobile agent-based directed diffusion in wireless sensor networks.EURASIP Journal on Applied Signal Processing, 2007(1):219–219, 2007.[6] Qishi Wu, Nageswara SV Rao, Jacob Barhen, SS Iyenger, Vijay KVaishnavi, Hairong Qi, and Krishnendu Chakrabarty. On computingmobile agent routes for data fusion in distributed sensor networks.Knowledge and Data Engineering, IEEE Transactions on, 16(6):740–753,2004.[7] Min Chen, Victor Leung, Shiwen Mao, Taekyoung Kwon, and MingLi. Energy-efficient itinerary planning for mobile agents in wirelesssensor networks. In Communications, 2009. ICC’09. IEEE InternationalConference on, pages 1–5. IEEE, 2009.[8] Wei Cai, Min Chen, Takahiro Hara, Lei Shu, and Taekyoung Kwon. Agenetic algorithm approach to multi-agent itinerary planning in wirelesssensor networks. Mobile Networks and Applications, 16(6):782–793,2011.[9] Mostefa Bendjima and Mohamed Feham. Multi mobile agent itineraryfor wireless sensor networks. International Journal of Emerging Trends &Technology in Computer Science, 1(1):6–11, 2012.[10] Huthiafa Q Qadori, Zuriati A Zulkarnain, Zurina Mohd Hanapi,and Shamala Subramaniam. Multi-mobile agent itinerary planningalgorithms for data gathering in wireless sensor networks: A reviewpaper. International Journal of Distributed Sensor Networks,13(1):1550147716684841, 2017.[11] Min Chen, Sergio Gonzalez, and Victor Leung. Applications anddesign issues for mobile agents in wireless sensor networks. WirelessCommunications, IEEE, 14(6):20–26, 2007.[12] Aristides Mpitziopoulos, Damianos Gavalas, CharalamposKonstantopoulos, and Grammati Pantziou. Mobile agent middleware forautonomic data fusion in wireless sensor networks. Autonomic computingand networking, pages 57–81, 2009.[13] Ousmane Diallo, Joel JPC Rodrigues, and Mbaye Sene. Real-time datamanagement on wireless sensor networks: A survey. Journal of Networkand Computer Applications, 35(3):1013–1021, 2012.[14] Saeid Pourroostaei Ardakani. Wireless Sensor Network Routing Protocolsfor Data Aggregation. PhD thesis, University of Bath, 2014.[15] Aristides Mpitziopoulos, Damianos Gavalas, CharalamposKonstantopoulos, and Grammati Pantziou. Deriving efficient mobileagent routes in wireless sensor networks with noid algorithm. In Personal,Indoor and Mobile Radio Communications, 2007. PIMRC 2007. IEEE18th International Symposium on, pages 1–5. IEEE, 2007.
VOLUME 4, 2016 et al. : Preparation of Papers for IEEE Access [16] Larry R. Esau and Kenneth C. Williams. On teleprocessing system design:part ii a method for approximating the optimal network. IBM SystemsJournal, 5(3):142–147, 1966.[17] Charalampos Konstantopoulos, Aristides Mpitziopoulos, DamianosGavalas, and Grammati Pantziou. Effective determination of mobile agentitineraries for data aggregation on sensor networks. Knowledge and DataEngineering, IEEE Transactions on, 22(12):1679–1693, 2010.[18] Damianos Gavalas, Ioannis E Venetis, Charalampos Konstantopoulos,and Grammati Pantziou. Energy-efficient multiple itinerary planningfor mobile agents-based data aggregation in wsns. TelecommunicationSystems, pages 1–5, 2016.[19] Min Chen, Sergio Gonzalez, Yan Zhang, and Victor CM Leung.Multi-agent itinerary planning for wireless sensor networks. InInternational Conference on Heterogeneous Networking for Quality,Reliability, Security and Robustness, pages 584–597. Springer, 2009.[20] Amine Rais, KHALID BOURAGBA, and Mohamed Ouzzif.Determination of itinerary planning for multiple agents in wirelesssensor networks. International Journal of Communication Networks andInformation Security (IJCNIS), 10(1), 2018.[21] Mohamed El Fissaoui, Abderrahim Beni-Hssane, and Mostafa Saadi.Multi-mobile agent itinerary planning-based energy and fault aware dataaggregation in wireless sensor networks. EURASIP Journal on WirelessCommunications and Networking, 2018(1):92, 2018.[22] Imene Aloui, Okba Kazar, Laid Kahloul, and Sylvie Servigne. A newitinerary planning approach among multiple mobile agents in wirelesssensor networks (wsn) to reduce energy consumption. InternationalJournal of Communication Networks and Information Security, 7(2):116,2015.[23] Aristides Mpitziopoulos, Damianos Gavalas, CharalamposKonstantopoulos, and Grammati Pantziou. Cbid: a scalable methodfor distributed data aggregation in wsns. International Journal ofDistributed Sensor Networks, 6(1):206517, 2010.[24] Soheil Javadi, Mohammad H Hajiesmaili, Behzad Moshiri, and AhmadKhonsari. Clone-based mobile agent itinerary planning using separatetrees for data fusion in wsns. International Journal of Wireless & MobileNetworks, 4(4):227, 2012.[25] Govind P Gupta, Manoj Misra, and Kumkum Garg. Energy and trust awaremobile agent migration protocol for data aggregation in wireless sensornetworks. Journal of Network and Computer Applications, 41:300–311,2014.[26] Junfeng Wang, Yin Zhang, Zhuanli Cheng, and Xuan Zhu. Emip:energy-efficient itinerary planning for multiple mobile agents in wirelesssensor network. Telecommunication Systems, pages 1–8, 2015.[27] Min Chen, Laurence T Yang, Taekyoung Kwon, Liang Zhou, and MinhoJo. Itinerary planning for energy-efficient agent communications inwireless sensor networks. IEEE Transactions on Vehicular Technology,60(7):3290–3299, 2011.[28] Chen Wang and Li Xiao. Improving event-to-sink throughput in wirelesssensor networks. In International Conference on Distributed Computingin Sensor Systems, pages 50–63. Springer, 2007.
Huthiafa Q Qadori received his bachelor degreein computer science from Faculty of ComputerScience and Information Technology, Universitiof Al-Anbar, Iraq, in 2006, and his master ofInformation Technology from University TenagaNational (Uniten), Malaysia, in 2012. He iscurrently pursuing his Ph.D. degree at theDepartment of Communication Technology andNetworks, Universiti Putra Malaysia. His researchinterests are in computer networks, wireless sensornetworks, routing algorithms and the Internet of things.
ZURIATI AHMAD ZUKARNAIN (M’10)received her bachelor and master degrees inphysics and education from Universiti PutraMalaysia (UPM) in 1997 and 2000, respectively,and her Ph.D. degree in quantum computing andcommunication from the University of Bradford,U.K., in 2005. Now, she is a professor in theFaculty of Computer Science and InformationTechnology, UPM. She was appointed as Headof Department for Communication Technologyand Networks from 2006 to 2011. She also being appointed as the Headof the Section of High-Performance Computing, Institute of MathematicalResearch, UPM, from 2012 to 2015. She taught several courses forthe undergraduate students such as data communication and networks,distributed system, mobile and wireless networks, network security,computer architecture, and assembly language. For postgraduate students,she taught few courses such as advanced distributed computing and researchmethod. Also, she is a member of the IEEE. Her areas of interest arecomputer networks, distributed system, mobile and wireless networks,network security, quantum computing, and quantum cryptography.
Mohamed A. Alrshah (M’13–SM’17) receivedhis BSc degree in Computer Science from NaserUniversity - Libya, in 2000, and his MScand Ph.D. degrees in communication technologyand networks from Universiti Putra Malaysiain May 2009 and Feb 2017, respectively. Now,he is a Senior Lecturer in the Departmentof Communication Technology and Networks,Faculty of Computer Science and InformationTechnology, Universiti Putra Malaysia (UPM).Also, he is a senior member of the IEEE. He has published a numberof articles in high impact factor scientific journals. His research interestsare in the field of high-speed TCP protocols, high-speed wired andwireless network, parallel and distributed algorithms, WSN, IoT, and cloudcomputing.
Zurina Mohd Hanapi (M’10) received herbachelor in Computer and Electronic SystemEngineering from Strathclyde University in 1999,and her master in Computer and CommunicationSystems Engineering from UPM in 2004,and her Ph.D. in Electrical, Electronic, andSystem Engineering from Universiti KebangsaanMalaysia in 2011. Now, she is an AssociateProfessor in the Faculty of Computer Scienceand Information Technology, University PutraMalaysia (UPM). She has received an Excellence Teaching Awards in2005, 2006 and 2012 and she has received a silver medal in 2004 andbronze medal in 2012. She is a leader of some research projects and she haspublished many conference and journal papers and she is a member of theIEEE. Her research interests in Routing, Wireless Sensor Network, WirelessCommunication, Distributed Computing, Network Security, Cryptography,and Intelligent Systems. VOLUME 4, 2016 uthiafa Q Qadori et al. : Preparation of Papers for IEEE Access
Shamala Subramaniam (M’10) received herBachelor, Master, and Ph.D in Computer Sciencefrom University Putra Malaysia (UPM) in 1996,1999, and 2002, respectively. Now, she is aprofessor in the Department of CommunicationTechnology and Network, Faculty of ComputerScience and Information Technology, UniversitiPutra Malaysia. Also, she is a member ofthe IEEE. Her research interests are ComputerNetworks, Simulation and Modeling, Schedulingand Real-Time System.