localization-free and energy-efficient hole bypassing techniques for fault-tolerant sensor networks

Localization-free and energy-efficient hole bypassing techniquesfor fault-tolerant sensor networks

Onur Yilmaz a,b, Orhan Dagdeviren b,n, Kayhan Erciyes c

a New Jersey Institute of Technology, Electrical and Computer Engineering, University Heights, Newark, NJ 07102, USAb Ege University, International Computer Institute, Bornova, Izmir 35100, Turkeyc Izmir University, Computer Eng. Dept., Uckuyular, Izmir 35350, Turkey

a r t i c l e i n f o

Article history:Received 7 December 2012Received in revised form21 May 2013Accepted 9 September 2013

Keywords:Wireless sensor networksFault-tolerant routingFault toleranceBypassBypassing holesResilienceReliable routing

a b s t r a c t

Nowadays, since wireless sensor networks (WSNs) are increasingly being used in challenged environ-ments such as underground mines, tunnels, oceans and the outer space, fault-tolerance need has becomea major requirement for routing protocols. So far, the proposed fault-tolerance methods or algorithmsaim to recover the isolated failures which occur at different parts of the network in different times.However, there is another type of failure for WSNs which is more destructive for the applications.By collapsing sensor nodes as a group at the same time, a hole can appear at the network which may cutthe data delivery drastically. In the literature, previous studies for bypassing holes are based onlocalization which may have significant energy and economic costs. In this paper, two localization-freeand energy-efficient algorithms are proposed for bypassing the holes formed by group collapse. Werealized that when holes are modeled with clusters, hole bypassing can be solved by cluster bypassing.Our algorithms, intra-cluster bypass and inter-cluster bypass, aim to heal the corrupted communicationlinks in the presence of holes. We show the operation of the algorithms, analyze them and provideextensive simulation results in an ns-2 environment. We compare our proposed algorithms with theother approaches and show that our algorithms significantly improve the fault recovery percentageswhile consuming a reasonable amount of energy even in the presence of high collapse ratio.

Crown Copyright & 2013 Published by Elsevier Ltd. All rights reserved.

1. Introduction

Advances in hardware and wireless network technologies haveresulted in low-cost, low-power, multi-functional miniature sen-sor devices (Tubaishat and Madria, 2003; Mendes and Rodrigues,2011; Lin, 2013; Hadjidj et al., 2013; Pantoni and Brando, 2013;Dai et al., 2013). These sensor nodes are small, with limitedprocessing and computing resources, and they are inexpensivecompared to traditional sensors. These tiny sensor motes cansense, measure, and gather information from the environmentand, based on local decision processes they can transmit thesensed data to a base station (sink). By communicating with eachother and conveying information, which is gathered from theenvironment, they form a huge network which cannot be com-pared with the other type of networks. WSNs are generally largescale and distributed systems which are composed of ten tothousand of sensor motes that are communicating with eachother. They are used in diverse applications such as military target

tracking and surveillance, natural disaster relief and biomedicalhealth monitoring (Yick et al., 2008). Although WSNs are used forvarious purposes, they contain challenges such as limited energyand wireless communication problems.

WSNs are increasingly being used in challenged environmentssuch as underground mines, tunnels, oceans and the outer space.Wireless communication in challenged environments has trans-mission failures, mainly as a consequence of direct impact ofphysical world. In addition to energy constraints and wirelesscommunication problems, tiny sensor motes are prone to failures.In Paradis and Han (2007), sources of all fault types are proposedin detail. In many critical applications of WSNs, the communica-tion in challenged environments has to be reliable and thereforethese requirements bring a need for the faults to be detected andrecovered timely.

In particular, in a sensor network which is deployed in anextreme environment, each node may individually fail or a groupof geographically close located nodes may collapse which formsholes (Fang et al., 2006). A hole may prune the communication linksin a sensor network where data transfer from hole affected regionsof the network becomes impossible without necessary topologyrecreation. In this case, sensor network applications have to suspendtheir communications and data delivery during construction of

Contents lists available at ScienceDirect

journal homepage: www.elsevier.com/locate/jnca

Journal of Network and Computer Applications

1084-8045/$ - see front matter Crown Copyright & 2013 Published by Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.jnca.2013.09.002

n Corresponding author. Tel.: þ90 2322464949; fax: þ90 2322240909.E-mail addresses: [email protected] (O. Yilmaz), [email protected]

(O. Dagdeviren), [email protected] (K. Erciyes).

Please cite this article as: Yilmaz O, et al. Localization-free and energy-efficient hole bypassing techniques for fault-tolerantsensor networks. Journal of Network and Computer Applications (2013), http://dx.doi.org/10.1016/j.jnca.2013.09.002i

Journal of Network and Computer Applications ∎ (∎∎∎∎) ∎∎∎–∎∎∎

www.sciencedirect.com/science/journal/10848045

www.elsevier.com/locate/jnca

http://dx.doi.org/10.1016/j.jnca.2013.09.002



mailto:[email protected]








new paths. Thus, network recovery in the presence of holes is a veryimportant problem. Although, this problem is addressed and studiedby the researchers (Fang et al., 2006; Yu et al., 2007, 2008, 2009a;You et al., 2009b; Shiow-Fen et al., 2010), the main focus lies in nodelocalization based approaches where sensor nodes should beequipped with a position tracker or localization algorithms shouldbe executed a priori on these nodes. Equipping nodes with a positiontracker like a global positioning system (GPS) receiver may causesignificant costs. On the other hand, executing complicated localiza-tion algorithms may exhaust the batteries of sensor nodes whichmay cause new faults. In either cases, localization may introduce aconsiderable cost to the sensor network.

In this study, we aim to design localization-free and energyefficient hole bypassing techniques for fault-tolerant sensor net-works. Our idea is firstly to construct a cluster tree rooted at a sinknode where the network is partitioned into multi-hop clusters.By applying this strategy, we aim to model holes with clusters.Afterwards, to recover the communication links in the network,we propose an intra-cluster energy-efficient solution in the firststep and an inter-cluster robust solution in the second step. Byapplying these methods, we aim to avoid the cost of localizationand network-wide topology recreation.

The rest of this paper is organized as follow. In Section 2, wereview the related work on fault recovery techniques. We showthe network model and the hole problem formulation in Section 3.In Section 4, we introduce the proposed intra-cluster and inter-cluster based methods. We show the simulation results of theproposed methods and its performance is compared with therelated work in Section 5. The conclusions are drawn in Section 6.

2. Related works

Routing is an attractive research area in all networks. Like othernetworks, researchers have proposed various protocols and algo-rithms for conveying messages to sink and dealing with thechallenges of WSNs.

The early studies on routing in WSNs dealt with the energyproblem and they tried to optimize the energy consumption. One ofthe answers to this problem was the data-centric routing mechan-ism. Sensor protocols for information via negotiation (SPIN)(Heinzelman et al., 1999) is one of the earliest works to pursue thedata-centric routing mechanism in WSNs. Then, directed diffusion(Intanagonwiwat et al., 2003) was proposed which is an importantmilestone in data-centric routing mechanisms. After these studies,many approaches which pursue data-centric routing were proposedincluding the study in Braginsky and Estrin (2002). On the otherhand, clustering-based protocols were proposed around the sametime, in order to solve the energy problem. Low energy adaptiveclustering hierarchy (LEACH) (Heinzelman et al., 2000) formsclusters regarding received signal strength indicator (RSSI) of sensornodes which is one of the earliest studies in clustering-basedprotocols. However, LEACH is inefficient in terms of energy since itlacks multi-hop routing. Then, the threshold-sensitive energy-effi-cient sensor network protocol (TEEN) (Manjeshwar and Agrawal,2001) and the adaptive periodic threshold-sensitive energy-efficientsensor network protocol (APTEEN) (Manjeshwar and Agrawal, 2002)were proposed which are designed to respond reactively to suddenenergy changes efficiently.

Then, the focus of network community shifted to fault-tolerancebecause sensor nodes are prone to failures. Since the event packetsare conveyed over the sensor nodes hop by hop toward sink, anyfault of sensor nodes can cause event packet losses. In particular,this is a serious problem for critical systems. In Boukerche et al.(2006a), the periodic event-driven and query-based protocol (PEQ)and its variation clustering periodic event-driven query-based

protocol (CPEQ) which are fault-tolerant and low-latency routingprotocols are proposed. In Boukerche et al. (2008), the inter-clustercommunication based energy aware and fault-tolerant protocol(ICE) is proposed which is a cluster based, energy aware andfault-tolerant routing protocol for WSNs. In Boukerche et al.(2006b), the variable transmission range protocol (VTRP) is pro-posed for smart dust networks, a special type of WSN, which is anenergy-efficient and fault-tolerant protocol using a variation of thetransmission range. This is the first study for data propagation inthe literature which uses a varying transmission range technique.It is pointed out in this paper that additional knowledge, obtainedby increasing the transmission power in a distributed manner,improves the data propagation to sink in terms of fault tolerance.In Chatzigiannakis et al. (2007), a fault-tolerant and efficient datapropagation protocol for WSNs is proposed which uses the varyingtransmission range technique same as in Boukerche et al. (2006b).In Ganesan et al. (2001), the resilience of directed diffusion isincreased by constructing disjoint and braided multipaths.

When the fault-tolerant routing algorithms and protocols inthe literature are examined, they usually gather around the samefault-tolerant techniques including disjoint and braided multi-paths, instant recovery and reconstructing the routing paths.Although these techniques provide reasonable recoveries, theydo not intend to recover the holes which are formed by suddengroup collapse. Although there are some studies intended to solvehole problem (Fang et al., 2006; Yu et al., 2007, 2008, 2009a;You et al., 2009b; Shiow-Fen et al., 2010), they are mainly based onbypassing holes using localization techniques. Since our concern islocalization-free hole bypassing methods, we omit these localiza-tion based studies. The related methods are as follows:

� Disjoint multipath routing: In this routing scheme, the alternatepaths are node disjoint with the primary path and with eachother (Ganesan et al., 2001). Because of this property, a nodefailure in the primary path does not affect the alternate pathsand so on. Node disjoint paths can be constructed by executinglocalized algorithms like directed diffusion (Intanagonwiwatet al., 2003). Although node disjoint paths are fault-tolerant,it may not always be possible to construct disjoint pathsespecially in sparse networks. Besides, node disjoint pathscan be energy inefficient since alternate paths can be longerthan the primary path.

� Braided multipath routing: Node disjointedness requirement isrelaxed in braided multipath routing where alternate braidedpaths are partially disjoint from the primary path (Ganesanet al., 2001; Boukerche et al., 2008). Braided paths from asource node to the sink node can be constructed as follows: foreach node v on the primary path, find the best possible pathfrom source node to sink node that does not include node v.The alternate paths are expected to be geographically close tothe primary path, thus the technique is energy-efficient intui-tively. Alternate paths can be constructed with a localizedalgorithm similar to disjoint multipath routing. Although thistechnique provides a fault-tolerant infrastructure, it is boundedwith the parent–child relationships which are constructed apriori, and it may not reactively recover faults in the presenceof holes.

� Instant recovery: In this technique, when a node detects itsparent's fault, it reactively tries to recover the fault by searchingalternative parent (Boukerche et al., 2006a). In order to find asuitable alternative parent, the node m broadcasts a Search(m.level) message to its neighbors. When the neighbor node v

receives a Search(level) message, it replies with an acknowl-edgment if v:levelrSearch:level and it is not the child of thesender of Search message. In this manner, the loops areprevented. PEQ and CPEQ protocols use this technique to

O. Yilmaz et al. / Journal of Network and Computer Applications ∎ (∎∎∎∎) ∎∎∎–∎∎∎2





recover faults (Boukerche et al., 2006a). Although this techni-que provides a fast and low cost recovery, it is bounded withthe level of nodes and the event delivery percentage may notbe adequate when many of nodes collapse.

� Re-executing topology construction algorithms: Previously describedmethods can be inadequate to recover faults in challenged envir-onments where holes are present and fault rate is high. In thissituation, topology may be periodically reconstructed in order tocontinue data delivery. When topology reconstruction is applied,new links are chosen for data communication and parent–childrelationships are updated with the level information. Fault-tolerantprotocols like CPEQ and ICE use this method periodically toregenerate links (Boukerche et al., 2006a, 2008). Many othertopology construction algorithms offer this method for faultrecovery (Karl and Willig, 2005; Dagdeviren and Erciyes, 2010;Akkaya and Younis, 2005). Although this method may provideidealized recovery from the faults, it has important drawbacks.Firstly, in the duration re-execution is applied, nodes may not beable to transmit their data packets. This may cause too much delayin data delivery operation. Secondly, the energy cost of thisoperation can be very high. Assuming that each node at leastsends one message during re-execution, the message cost of thisoperation is ΩðNÞ where N is the total node count.

3. Problem formulation

3.1. Network representation

The following assumptions are made about the network:

� Each node has distinct node_id.� The nodes are stationary.� Links between nodes are symmetric. Thus if there is a link from

u to v, there exists a reverse link from v to u.� Nodes do not know their positions. They are not equipped with

a position tracker like a GPS receiver.� All nodes are equal in terms of processing capabilities, radio,

battery and memory.� Each node knows its neighbors.

Based on these assumptions, the network may be modeled as anundirected graph GðV ; EÞ where V is the set of vertices, E is the set ofthe edges. An example of undirected graph model is depicted inFig. 1 where ten ordinary sensor nodes and a sink node are locatedon the sensing area. Each node is labeled with its identifier, thetransmission ranges of the nodes are shown in dotted circles, andthe transmission edges between two nodes are shown in solid lines.

3.2. Hole problem

Holes may occur in WSNs located in various environmentshaving challenging conditions. A hole may stop the operation ofthe sensor network completely or may partition the network intodisjoint parts which may significantly reduce the event collection.To recover faults in the presence of holes, we propose to constructa multi-hop cluster tree architecture. This architecture is generic,it can be an arbitrary multi-hop cluster tree rooted at the sink andcan be constructed by topology generation algorithms such asgiven in Boukerche et al. (2006a), Boukerche et al. (2008), Akkayaand Younis (2005), Erciyes et al. (2008), Wagner and Wattenhofer(2007), Peleg (2000).

An example case is shown in Fig. 2 where the transmissionedges between two nodes are shown in dashed lines, the border ofclusters is shown in solid circles, the cluster edges are shown in

directed solid edges. Each cluster has an identification (id) varyingfrom A to M. Clusters are leveled from 0 to 4 by calculating theircluster count on the shortest path to sink node. The cluster levelscan be found by a graph traversal algorithm like breadth-firstsearch (BFS) where the level of sink node is 0, the levels ofneighbor clusters of sink node are 1, and the levels of other clustersare their cluster count on the path to sink node after BFS areexecuted. A node is identified by concatenating its cluster id andnode id, e.g. H5. Each cluster head (CH) has an id of 1 and is shownin black. The level of a node is its hop distance from its CH. Forexample, H5's level is 2, since it can reach H1 in two hops. The holeis shown in a bold closed spline. Hole includes all nodes in clustersA and D, node H2, node H3, node B2 and node E2 as shown in Fig. 2.

After the formation of the hole, H5 cannot transmit its packetsince its parent H2 has crashed. H5 cannot recover its link by usinginstant recovery and multipath routing techniques because H5'sremaining neighbor is H6 and its level is higher than H5's level.Besides, the packets coming from cluster L and cluster K cannot beforwarded to the sink node, because H5 is the parent of both L1 andK1. Although H6 can recover its link and set its new parent as H7,this healing operation will only benefit itself. Moreover, since D4 isthe parent of G1 and it has crashed, packets coming from cluster Gcannot be relayed to sink node. Instant recovery and multipathrouting techniques cannot recover the link of G1, since G2 and G3'slevels are higher than G1's level. We may state from this examplethat holes may significantly reduce the performance of sensornetworks using instant recovery and multipath routing fault-tolerant techniques.

To heal all the corrupted links which cannot be recovered byinstant recovery and multipath routing techniques, a networkwide re-execution of the clustering protocol can be applied. Onthe other hand, nodes can redundantly consume their energy byexecuting this technique. In Fig. 2, network area is divided into tworegions bounded by a bold dashed line. The Region 1 covers sensornodes which cannot send their data packets to the sink node. TheRegion 2 consists of the hole and the other working nodes whichcan send their data packets to sink. Although the links of nodes inRegion 1 are recovered by the re-execution, links of nodes inRegion 2 remain the same. Since Region 2 is much more greaterthan Region 1, we may state re-execution that may inefficientlyconsume the energy of sensor nodes. In addition, the nodes inRegion 2 cannot send event packet to the sink node duringreconstruction phase. Our example hole scenario depicted inFig. 2 which is explained so far may occur in various forms insensor networks and these problems can be generalized. In this

Fig. 1. Network model.

O. Yilmaz et al. / Journal of Network and Computer Applications ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3





manner, our problem is localization-free and is energy-efficient holebypassing techniques for multi-hop clustered and fault-tolerantsensor networks.

4. Proposed methods

In this section, we will describe two localization-free and energyefficient methods that we propose for fault tolerance in multi-hopclustered sensor networks.

4.1. Intra-cluster bypass

We propose the intra-cluster bypass technique which is designedto heal the corrupted intra-cluster links. Our aim is to recover thecommunication paths in a corrupted cluster by applying localoperations residing in the same cluster. Our method is both faulttolerant and energy-efficient in this manner.

In some cases, a hole may span just a partition of a cluster asdepicted with cluster H in Fig. 2. Instant recovery cannot recover thelink of a node whose neighbors have higher cluster level than thisnode as shown with node H5 in Fig. 2. Our goal is to reconstructintra-cluster paths in these cases.

When a node (node v) cannot find an alternate parent, itinitiates intra-cluster bypass operation by broadcasting an intra-bypass search message to its neighbors. The neighbors of node v

which are at the same cluster forward this message just once totheir neighbors by broadcasting where this message is relayed tonode v's CH and the nodes residing in different clusters do notrespond to this message. When the CH of node v firstly receivesthis message, it unicasts a reinforcement message back to theforwarder of search message where this message is relayed to thenode v by unicasting in the backward direction. In this way, thepath is reconstructed. The pseudocodes of the procedures for this

algorithm are given in Algorithms 1 and 2. The features of theproposed intra-cluster bypass technique are listed below:

� The method recovers links locally in the corrupted clusterwithout disturbing the operation on the other clusters.

� The method provides cycle-free routes and it is resource-efficient as proved in Theorems 1 and 2.

Algorithm 1. Intra bypass – search.

upon node r receives an intra bypass search message from node s;if search.clusterID ¼ clusterID thenif search message with search.searchID has firstly received then

store the sender and search.searchID as pair intableIntraSearchMessages;if the node is cluster head thenunicast a reinforcement message to node s;

elsebroadcast the same search message;

endend

endend upon

Algorithm 2. Intra bypass – reinforcement

upon node r receives an intra bypass reinforcement messagefrom node s;

set node s as new parent;if node r is the initiator of the intra bypass thencontinue to relay events because the path is repaired;

elsefind the sender of intra bypass search message fromtableIntraSearchMessages;

Fig. 2. Hole problem.






unicast reinforcement to the sender of intra bypass searchmessage;

endend upon

An example operation is shown in Fig. 3. The corrupted clusterH from Fig. 2 is depicted in Fig. 3. Nodes H2 and H3 are corruptedwhich cause that node H5 and node H6 cannot send their packetsto sink node as shown in Fig. 3(a). Because of this, node H5 initiatessending search message and this message is relayed to the nodeH1. Reinforcement message originated from node H1 is sent back-ward along the path of search message. These message transfersare shown in Fig. 3(b). Finally, the path is recovered asH5-H6-H7-H4-H1 as shown in Fig. 3(c).

Theorem 1. Intra-cluster bypass operation is free from cycle formation.

Proof. In an intra-cluster operation, each node sends searchmessage once. The reinforcement message can only be originatedby a CH once for each search message. Thus the corrupted pathcannot be recovered without the CH. Since CH's parent is not inthe same cluster and CH's level is smaller than its clustermembers, formation of a cycle is not possible. □

Theorem 2. Assume that Cm is maximum node count in a cluster,Dm is the maximum cluster diameter, the message and the timecomplexities of an intra-cluster bypass operation are O(Cm) and O(Dm) respectively. The message size is O(log2(N)) and space complex-ity is O(Cm) where N is the number of nodes.

Proof. All nodes in the corrupted cluster broadcast search mes-sage and nodes along the recovered path unicast reinforcementmessage. Assume that the corrupted node count in the cluster is fand the count of nodes along the recovered path is r, the totalnumber of messages in this case is OðCmþr� f ÞAOðCmÞ when0oroCm and 0o f oCm. At most, propagation of search messagein the corrupted cluster takes O(Dm) time.Since all fields in search and reinforcement messages may be in

(0,N) interval, the message size is O(log2(N)). Each node shouldstore a table including (sender,ID) pair where this table can have asize of at most O(Cm). □

4.2. Inter-cluster bypass

When a fault occurred in a cluster, the nodes in this cluster firstlytry to recover it by in-cluster fault-tolerant techniques. However, itmay not be possible to recover the fault in some cases especially whena hole may span a whole cluster or major part of it. For example, node

Fig. 3. Intra-cluster bypass example. (a) A cluster and a Hole inside the cluster, (b) transmission of reinforcement and search messages for intra-cluster bypass and(c) Reconstructed cluster connections.

Fig. 4. An inter-cluster bypass example.






G1 cannot send its packets directly to a cluster at the upper level sinceall nodes in cluster D have failed as shown in Fig. 4.

In the multi-hop clustered sensor network architectures, everycluster except sink node's cluster has an up cluster (parent cluster).These clusters are responsible to send their data as well as to relaythe packets of their down clusters to their up clusters. Whencluster's up cluster has failed, it can send its packet to sink nodeand it cannot use intra-cluster bypass to recover, as mentioned inthe previous paragraph. To overcome this situation, we proposethe inter-cluster bypass technique. Our aim is to provide the downclusters keep relaying message to sink. In our technique, the CH(s) of down cluster(s) search for an alternate CH for constructing apath toward to it and relaying its messages to sink with bypassingits up cluster. By relaying the messages over the alternate CH, thepaths of down clusters can pass its up cluster.

When a CH with cluster level (cl) cannot find an alternateparent, it initiates inter-cluster bypass operation by broadcastingan inter-cluster bypass search message to its neighbors. The CH'sneighbors residing in different clusters forward this message justonce to their neighbors. This message is forwarded once by theother nodes if their cluster level (nl) is clþmZnl4cl�m where mis a previously defined constant. When a CH with cluster level (bl)receives a search message, it originates a reinforcement messageonce to the forwarder of searchmessage, if clZblZcl�m. Also thisCH should not use a path constructed with inter-cluster bypass toavoid cycle formation. The reinforcement message sent by CH isrelayed to the search originator CH by unicasting it from thebackward direction where the path is recovered. There may bemore than one CH sending a reinforcement message, in this casethe first one constructs the path, operation of other CHs is omitted.The pseudocodes of the procedures of this algorithm are given inAlgorithms 3 and 4. The features of the proposed inter-clusterbypass algorithm are summarized below:

� The proposed method provides a recovery technique when ahole spans more than one cluster.

� The proposed method is cycle free as proved in Theorem 3. Thetime complexity depends on the diameter of the network, themessage complexity depends on the m constant, maximumcluster count at the same level (k), the maximum node count ina cluster (Cm) and the network diameter as proved in Theorem4. When k, Cm and DAOð

ffiffiffiffi

Np

Þ, the message complexity is linear.� In the worst case, the count of fault recovery alternatives

produced by the inter-cluster bypass technique is at least thecount of fault recovery alternatives produced by other methodsas proved in Theorem 5.

Algorithm 3. Inter bypass – search

upon node r receives an inter bypass searchmessage from node s;if search message with search.searchID has firstly received then

store the sender and search.searchID as pair intableInterSearchMessages;if search:clusterLevelZclusterLevel andðsearch:clusterLevel�mÞoclusterLevel thenif node r is cluster head and does not use inter bypass thenunicast a reinforcement message to node s;

elsebroadcast the same search message;

endelse if search:clusterLeveloclusterLevel andðsearch:clusterLevelþmÞZclusterLevel thenbroadcast the same search message;

endendend upon

Algorithm 4. Inter bypass – reinforcement

upon node r receives an reinforcement message from node s;if reinforcement message with the search id reinforcement.

searchID has firstly received thenstore the sender and reinforcement.searchID as pair intableByPassParents for bypass path;if node r is the initiator of the inter bypass then

continue to relay events through the new path because thefault is recovered;else

find the sender of inter bypass search message fromtableInterSearchMessages;unicast reinforcement to the sender of inter bypass search

message;end

endend upon

An example inter-cluster bypass operation with m ¼ 1 isdepicted in Fig. 4 where clusters from Fig. 2 are shown. The holein the figure spans cluster D and node E2. Since cluster G's upcluster is cluster D, node G1 cannot send its cluster's packet to sinknode. So node G1 starts inter-cluster bypass operation by broad-casting a search message to its neighbors and this message isrelayed to all nodes in clusters at level 2, level 3 and level 4. NodeH1 which used intra-cluster bypass to recover its residual clusterlinks, forwards a reinforcement message along the path of receivedsearch message. All CHs in clusters with level 2 and level 3 sendreinforcement messages same as node H1. Since node H1's reinfor-cement message is received before the other CHs' messages,node H1 constructs the inter-cluster bypass path which isG1-G2-G4-K1-K3-L2-L1-H5-H6-H7-H4-H1.

Theorem 3. Inter-cluster bypass operation is free from cycle formation.

Proof. In an inter-cluster bypass operation, each node sendssearch message once. The reinforcement message can be originatedby CHs with equal or smaller cluster levels, once for each searchmessage. Also these CH's should not have used inter-cluster bypasspreviously. By applying these constraints, formation of a cyclebecomes impossible. □

Theorem 4. The message complexity of an inter-cluster bypassoperation is O(mk(CmþD)) and the time complexity is O(D) wherek is the maximum number of clusters at the same level and D is thenetwork's diameter. The message size is Oðlog 2ðNÞÞ and the spacecomplexity is O(Cm).

Proof. When a CH with cluster level (lb) tries to bypass its uppercluster, it sends search message where this message is propagatedto the clusters with level (lo) providing lbþmZ lor lb�m. Thus,the message is propagated to 2m þ 1 levels of clusters excludingthe faulty cluster. The number of messages for this operation ismkCm. A mk�1 number of CHs may send reinforcement messages,so ðmk�1ÞD reinforcement messages may be transferred includingrouting operations. The number of total messages is mkCmþðmk�1ÞDAOðmkðCmþDÞÞ. The time complexity for the dissemination ofsearch messages is O(D) and similarly transmission of reinforcementmessages takes O(D) time, thus total time complexity is O(D). Allfields in messages may be in (0,N) interval, thus the message size isOðlog 2ðNÞÞ. Each node should store a table including (sender,ID) pairwhere this table can be at most of size O(Cm), same as intra-clusterbypass. □






Definition 1. A cluster border node has at least one neighborresiding in another cluster. Other nodes are cluster non-bordernodes.

Theorem 5. Assume that the count of node v's alternate parent thatcan be chosen by the instant recovery and the braided multipathtechnique is i, by the intra-cluster bypass is a and by applying theinter-cluster bypass technique is b. Also assume that lh is the count ofinter-cluster links which connect node v to higher level nodes. Forcluster border nodes with ðΔ� lhÞ4 lh and for all cluster non-bordernodes, the fault recovery order is: 0r irarbrΔ. For the othernodes the fault recovery order is 0rar irbrΔ.

Proof. The crashed parent can be the only neighbor of faultdetecting node, thus lower bound is 0. In all techniques the newparent should be chosen in the non-faulty neighbor set, thusupper bound is Δ. The non-faulty neighbor nodes with equal orsmaller level than the level of the fault detecting node can bechosen as the new parent in instant recovery and braided multi-path techniques. The non-faulty neighbor nodes in the samecluster can be chosen as the new parent in intra-cluster bypass.When ðΔ� lhÞ4 lh, as a result of lower level intra-cluster recoveryalternatives being higher, the order is ira. Obviously, this case istrue for all cluster non-border nodes. All non-faulty neighborswithout any restriction are the candidates for the new parent inthe inter-cluster bypass technique, thus arb. The final orderwhen ðΔ� lhÞ4 lh is: 0r irarbrΔ. When ðΔ� lhÞo lh is true,as the higher level inter-cluster alternatives are more than otheralternatives, the order becomes: 0rar irbrΔ. □

5. Performance evaluations

In order to evaluate the performance of proposed two fault-tolerant techniques in the presence of group collapse, we havecarried out an extensive simulation study using ns-2 simulator.We have conducted two simulation studies over two differentprotocols including a generic protocol and CPEQ (Boukerche et al.,2006a). In the simulation study with the generic protocol, wecompared the proposed techniques with well-known fault-toler-ant techniques including instant recovery (search), braided-multipath and global re-execution in terms of event deliverypercentage, energy consumption and delay. When the well-known fault-tolerant routing protocols (Boukerche et al., 2006a,2008, 2006b; Chatzigiannakis et al., 2007; Ganesan et al., 2001)are examined deeply, it can be clearly seen that these techniquesare mainly used for fault recovery purposes. Besides, thesetechniques are utilized in various protocols by employing eitheronly one of them or as a combination of multiple techniques. Onthe other hand, in the simulation study with the CPEQ protocol,we extended the CPEQ protocol with the proposed fault-toleranttechniques and evaluate its performance in terms of event deliverypercentages and energy consumptions. The CPEQ protocol(Boukerche et al., 2006a) is one of the widely accepted cluster-based fault-tolerant routing algorithms in the literature. The aimof this simulation study is also to show how the proposedtechniques can be implemented to a well-known cluster basedprotocol.

We designated the simulation parameters by taking into con-sideration the parameters in CPEQ and directed diffusion(Intanagonwiwat et al., 2003) for both simulation studies. Theinitial energy of sensor nodes was set to 100 joules. Randomgenerated topologies were used in different node counts rangingfrom 100 to 500 nodes for average node degrees varying between5 and 23. The position of sink was also set randomly. IEEE 802.15.4radio and medium access control (MAC) layer standards readilyavailable in the ns-2 simulator were chosen for lower layer

protocols. The transmission range of nodes was set to 20 m. Thetransmission, receiving and idle powers were set to 0.660 w,0.395 w and 0.035 w respectively. The simulation parameters aresummarized in Table 1.

5.1. Performance evaluations on the generic protocol

In this simulation study, we use a generic protocol in order toassess the performance of two proposed fault-tolerant techniquesin the presence of group collapses. We call it as a generic protocolsince many researchers used these type of topologies for datadelivery (Erciyes et al., 2008; Boukerche et al., 2008; Karl andWillig, 2005; Wagner and Wattenhofer, 2007; Peleg, 2000).Besides implementing the generic protocol, we evaluated theperformance of instant recovery, braided-multipath and globalre-execution, and compared with the two proposed techniques.The reason to choose these techniques is almost all localization-free fault-tolerant routing protocols just use these techniques andtheir very similar versions for fault recovery as mentioned inSection 2.

The paths are constructed toward the sink with the shortesthop algorithm in the generic protocol for relaying the messageenergy efficiently. The algorithm is based on flooding the networkwhich is initiated by the sink with a hop value. The neighbornodes of the sink store the received hop value, increment it andtransmit it to its neighbor nodes and so on until the whole sensornetwork is configured with different levels of hops. In order toconfine the message diffusion, a node does not always transmit anew hop message after it receives a hop message. When a nodereceives a hop message from its neighbor, it checks this valueagainst its local hop value. If the local hop value is greater than thereceived one, the node updates its hop, increment this value andretransmit it to its neighbors. Each node sets the sender ofminimum hop message as a parent node which is used forforwarding the messages toward sink. After the algorithm termi-nates, a shortest hop tree is formed whose root is the sink.

Since the proposed techniques are for cluster based protocols,clusters have to be formed in the generic protocol. Therefore, weused the method in Erciyes et al. (2008) for forming clusters over atree. This technique is very simple and it does not have anyoverhead. A constant depth value is defined in this method and if(hop level mod depth ¼ 0) then, the node becomes the clusterhead, otherwise it is an ordinary node in the cluster. Thistechnique can also be used as a tree based protocol in order toform simply clusters. Therefore, although the proposed techniquesare for cluster based protocols, they can be implemented to treebased protocols by simply forming clusters with this method.In order to assess the performance of the proposed techniques fordifferent cluster sizes, the simulation study is conducted by settingthe depth value to 3, 5 and 7 in order.

Table 1Experimental study parameters.

Simulation time (s) 1000Number of nodes 100–500Sink position Randomly placed in the areaPercentage of source nodes (%) 2Source data rate (eventMsgs/s) 2Radio range (m) 20MAC 802.15.4Transmit energy (w) 0.660Receive energy (w) 0.395Dissipation in idle (w) 0.035Initial energy (j) 100Node degrees 5–23Percentage of node failure (%) 10, 20 and 30






The delay, energy consumption and resilience of instant recov-ery, braided-multipath, global re-execution and proposed techni-ques are evaluated over a sample WSN application which is sameas in Boukerche et al. (2006a) and Intanagonwiwat et al. (2003).Two percent of the nodes of the network are selected randomlyper 0.5 second in order to generate a random traffic. The selectednodes generate event packets and these packets are relayed to sinkthrough the shortest hop tree. After a node transmits a packet toits parent in the tree, it waits an acknowledgement from its parentfor reliable communication. If a sender node cannot receive anacknowledgement in a constant time, it retransmits the packet.Otherwise, if it cannot receive any acknowledgement after tenretransmissions, the sender node detects a fault and performs arecovery algorithm.

During the simulation, group collapse is simulated by turningoff the fixed fraction of nodes. At a random time in the simulation,randomly selected nodes and their neighbors are turned off until afixed fraction of nodes are reached, and thereby, this type of faultsforms big and small holes in the network.

The implemented methods are utilized regarding their faultrecovery merits and energy consumptions in various routingprotocols. Generally, the cheapest method in terms of energyconsumption is firstly performed. However, it usually offers lowerfault recovery rate. Other costly methods are performed adaptivelyif they are required. Thus, in order to show how the proposedmethods can be used coherently with these methods and toprovide a comparison, we set up twelve different simulationseries. The series can be divided into two groups. In the firstgroup of series, each technique is executed solely in order tomeasure its performance. We implemented no-fault tolerancealgorithm, braided multipath, instant recovery, intra-clusterbypass and global re-execution techniques. Since inter-clusterbypass is designed for joint use with another technique, we didnot implement it alone. The series of the first group is as follows:

� No fault-tolerance algorithm (No Tolerance): In this simulationsetup, only the generic protocol is used which does not performany fault-tolerance algorithm in the presence of a fault. There-fore, No Tolerance is expected to show the worst event deliverypercentage.

� Braided multipath (BM): The braided multipath fault-toleranttechnique is only used in the generic protocol. Braided multi-paths are constructed while the shortest hop algorithm isperformed. In the presence of a fault, a node tries to recoverthe fault via using another path in the braided multipath.

� Instant recovery (search) (IR): The instant recovery fault-toleranttechnique is used in the generic protocol. The node discovers anew node from the lower layer instantly in order to recover thefault and keep relaying event messages.

� Intra-cluster bypass (IntraBp): The intra-cluster bypass fault-tolerant technique is used by a cluster member node in thepresence of a fault.

� Global re-execution (GR): The shortest hop algorithm is re-performed per 10 s for fault recovery. This technique is optimalin terms of fault recovery but it consumes high energy as wellas all nodes in the network have to wait for termination of thealgorithm in order to keep relaying events to the sink.

The second group includes integrated techniques for in-cluster andnetwork-wide fault tolerance. Braided multipath, instant recoveryand intra-cluster bypass are used for in-cluster fault tolerancewhereas inter-cluster bypass with m ¼ 1 and global re-executionare implemented to provide network-wide fault tolerance.

� Braided multipath þ inter-cluster bypass (InterBp): When acluster member node encounters a fault, it tries to recover

the fault by using a braided multipath. On the other hand,when a CH is informed or encounters a fault and it cannotrecover with a braided multipath, it uses the inter-clusterbypass technique.

� Braided multipath þ intra-cluster bypass þ inter-cluster bypass:This setup is same as the previous setup except that a clustermember node first tries to recover by braided multipath then ituses intra-cluster bypass.

� Braided multipath þ global re-execution: In the presence of afault, a node first tries to recover the fault via using braidedmultipath. Besides, the shortest hop algorithm is re-performedper 20 s for fault recovery. The difference between thistechnique and GR is that this technique provides fast recoverywith a braided multipath. Otherwise, a node can wait 20 s forrecovery and this can cause problems in emergency applications.

� Instant recovery þ inter-cluster bypass: When a node encoun-ters a fault, it tries to recover the fault by using braidedmultipath. When a CH cannot recover with braided multipath,it uses inter-cluster bypass.

� Instant recovery þ intra-cluster bypass þ inter-cluster bypass:This setup is same as the previous setup except that a clustermember node first tries to recover by instant recovery then ituses intra-cluster bypass.

� Instant recovery þ global re-execution: Instant recovery is usedfor in-cluster fault tolerance. For global re-execution, the shortesthop algorithm is re-performed for each 20 s.

� Intra-cluster bypass þ inter-cluster bypass: Only two proposedmethods in this paper are implemented together.

5.2. Event delivery percentages

Firstly, we measured the event delivery percentages of thealgorithms against various conditions in order to obtain faulttolerance quality of them. Figure 5 displays the affect of d clusterdepth to event delivery percentage at 10% node collapse rate for500 nodes. In fact, d parameter corresponds to the cluster size inthe generic protocol because clusters and CHs are determinedregarding the d parameter. According to Fig. 5, event deliverypercentage of IntraBp is increasing when the cluster size isincreased because the node which wants to bypass the hole inthe cluster has more chance to find alternative nodes. On the otherhand, the performance of InterBp with BM or IR is decreasingwhen the cluster size increases because the number of alternativeCHs for bypassing the holes decreases. When the IntraBp þInterBp with BM or IR are examined, it is shown that the resultsare quite different. Since cluster size has different effects onIntraBp and InterBp, the highest event delivery percentage wasmeasured in the middle of depth values. Thus, the cluster size

3 5 740

45

50

55

60

65

70

75

80

d Depth Value

Even

t Del

iver

y Pe

rcen

tage

(%)

BM+InterBpBM+IntraBp+InterBpIR+InterBpIR+ IntraBp +InterBpIntraBpIntraBp+InterBp

Fig. 5. Effect of d cluster depth on event delivery percentage.






should be selected as 5 for the case when IntraBp and InterBp areused together.

Figure 6(a) and (b) displays the event delivery percentage offault recovery techniques at 10% node collapse percentage and 5 ascluster depth for different node counts ranging from 100 to 500nodes. To compare these techniques with the ideal situation, weimplemented an ideal fault tolerance method called Full Tolerance.This method heals a corrupted path if it is possible, by applying acentral tree construction algorithm on the whole graph in anunrealistic manner. In Fig. 6(a) and (b), the event deliverypercentages of all fault-tolerant techniques are decreasing whilethe node count of network is increasing. The reason is that the sizeof formed holes increases with the node count in the networksince a fixed percentage of the nodes have collapsed. Thus, largerhole sizes cause more disrupted communication paths. The lowestevent delivery percentage which belongs to No Tolerance which isdecreasing from 83% to 38% whereas the highest, Full Tolerancewhich is decreasing from 99% to 86%. Figure 6(a) shows us that IRand BM provide nearly 3% more event delivery percentage than NoTolerance. It is clearly shown that IntraBp achieves better perfor-mance than BM and IR, and provides from 7% to 15% performanceincrease regarding No Tolerance. When the InterBp is performedwith BM, IR and IntraBp, event delivery percentage significantlyincreases as shown in Fig. 6(b). In particular, IR þ IntraBp þInterBp provides 97–78% which is very close to the performance ofFull Tolerance.

Figure 7(a) and (b) displays the event delivery percentages offault recovery techniques at 20% node collapse percentage and5 cluster depth for different node counts ranging from 100 to 500nodes. Figure 7(a) shows us that the event delivery percentages ofall techniques decrease compared with the results in Fig. 6(a) dueto an increase of node collapse percentage. Figure 7(a) shows usthat the event delivery percentage of No Tolerance significantlydecreases compared with the results in Fig. 6(a). The event

delivery percentage of No Tolerance is 77–28% whereas, thehighest, Full Tolerance, is decreasing from 98% to 79%. BM and IRprovide 4–5% more event delivery percentages than No Tolerance.IntraBp also achieves better performance than BM and IR at 20%node collapse percentage which provides up to 15% performanceincrease regarding No Tolerance. On the other hand, when Fig. 7(b) is examined, it is shown that the InterBp technique has morecontribution to the protocol in terms of event delivery percentageat this collapse percentage. Similarly in Fig. 6(b), IR þ IntraBp þInterBp provides the highest event delivery percentage with 93%to 64% among fault-tolerant techniques except Full Tolerance.

Figure 8(a) and (b) displays the event delivery percentage offault recovery techniques at 30% node collapse percentage and5 cluster depth for different node counts ranging from 100 to 500nodes. No Tolerance has 69–25% event delivery percentagewhereas, Full Tolerance has 95% to 61%. BM and IR provide 3–5%more event delivery percentages than No Tolerance. IntraBp alsoachieves better performance than BM and IR which is 10–14%more than No Tolerance. When the node collapse percentage is30%, the effectiveness of InterBp is significantly increasing.Figure 8(b) shows us that IR þ IntraBp þ InterBp provides thehighest event delivery percentage with 91–54% after FullTolerance.

Figure 9(a) and (b) displays the effect of node degree on eventdelivery percentage at 20% node collapse percentage and 5 clusterdepth for 500 nodes. Since a node can find more alternative nodesfor recovering the fault at high node degrees, event delivery of allfault-tolerant increases with the node degree. In particular, anincrease of IntraBp is more significant than the others. Therefore,the performance of IntraBp regarding BM and IR increases muchmore while node degree is increasing. The best performance isachieved when IntraBp is used with InterBp as shown in Fig. 9(b).

The most important obtained result is that the well-knownfault-tolerant techniques (IR and BM) are not capable of recovering

100 200 300 400 500

40

50

60

70

80

90

100

Node Count

Even

t Del

iver

y Pe

rcen

tage

(%)

BMIRNo ToleranceFull ToleranceIntraBp

100 200 300 400 500

40

50

60

70

80

90

100

Node Count

Even

t Del

iver

y Pe

rcen

tage

(%)

No ToleranceFull ToleranceBM+InterBpBM+IntraBp+InterBpIR+InterBpIR+IntraBp+InterBpIntraBp+InterBp

Fig. 6. (a) Event delivery percentage at 10% node collapse, (b) event delivery percentage at 10% node collapse (cont.).

100 200 300 400 500

30

40

50

60

70

80

90

100

Node Count

Even

t Del

iver

y Pe

rcen

tage

(%)


100 200 300 400 500

30

40

50

60

70

80

90

100

Node Count

Even

t Del

iver

y Pe

rcen

tage

(%)

No ToleranceFull ToleranceBM+InterBpBM+IntraBp+InterBpIR+InterBpIR+IntraBp+InterBpIntraBp+InterBp







the faults formed by group collapses when the node collapsepercentage increases. Since the IntraBp and InterBp techniques canreach more alternative nodes in the case of group collapses, theyoutperform the IR and BM in terms of event delivery percentages.Although IntraBp achieves better performance than IR and BM,it is inadequate at high node counts and collapse percentages. Atthis point, InterBp provides network wide fault-tolerance andreaches nearly Full Tolerance in terms of the event deliverypercentage. In addition, the event delivery percentage resultsconform to the Theorem 5.

5.3. Energy consumptions

Energy efficiency is an important metric for WSNs. We mea-sured the energy consumptions of proposed algorithms, whichoccur mostly by message transfers, and previous approaches. Inorder to accurately obtain the energy consumption of fault-tolerant techniques, the energy was measured for a time periodstarting right after the holes that are formed until each node relaysan event to sink. All energy figures display the energy consump-tion per node when the corresponding fault-tolerant technique isperformed. The measurements in the figures related to energyconsumption include both the energy consumption of fault-tolerant techniques and event relaying to the sink.

Figure 10 displays the effect of d cluster depth on energyconsumption per node when the corresponding fault recoverytechnique is performed at 20% node collapse percentage for 500nodes. Figure 10 shows us that energy consumptions of all fault-tolerant techniques except IntraBp increase from 3 to 5 and thenthey decrease from 5 to 7 cluster depth values. These energyconsumptions are related to both the event delivery percentageand the cluster size. From 3 to 5 cluster depth value, both thenumber of delivered events and the cluster size increase, thereby,the energy consumption increases. On the other hand, although

the cluster depth value increases from 5 to 7, since the event deliverypercentage decreases, the energy consumption also decreases. Theenergy consumption of IntraBp slightly increases when the clusterdepth value is increased because both the event delivery percentageand the cluster size increase.

Figure 11(a) and (b) displays the effect of node count on energyconsumption per node when the corresponding fault recoverytechnique is performed at 5 cluster depth for different node countsranging from 100 to 500 nodes. The lowest energy consumptionwhich belongs to No Tolerance increases from 1.57 J to 6.39 Jwhereas the highest, GR, increases from 2.21 J to 10.93 J.Figure 11(a) shows that BM and IR bring a few overhead in termsof energy consumption regarding No Tolerance which is between0.05 J and 0.3 J. IntraBp consumes slightly more energy than BMand IR as expected because the search messages of the method areflooded inside the cluster. Figure 11(b) shows that since the fault-tolerant techniques are performed adaptively, the energy con-sumption of the techniques which are combination of InterBpconsumes nearly the same energy ranging from 1.72 J to 8.44 J.Therefore, these techniques bring from 0.25 J to 2.0 J overhead pernode. On the other hand, BM and IR with GR consume from 1.9 J to9.68 J energy and bring from 0.42 J to 3.5 J energy overhead pernode. Therefore, although BM and IR with GR, and only GR providethe highest event delivery percentage, their energy consumption isfar from applicable. In addition, techniques with GR constantlydissipate energy while the WSN is operating but, on the otherhand, the other techniques consume energy only when they areperformed in order to recover the faults.

Figure 12(a) and (b) displays the effect of node collapsepercentage on energy consumption per node when the corre-sponding fault recovery technique is performed at 5 cluster depthand 500 nodes. Figure 12(a) and (b) shows that collapse percen-tage increase does not affect much energy consumption but, thereis still a decrease when it is increased. Since the event delivery

100 200 300 400 50020

30

40

50

60

70

80

90

100

Node Count

Even

t Del

iver

y Pe

rcen

tage

(%)


100 200 300 400 50020

30

40

50

60

70

80

90

100

Node Count

Even

t Del

iver

y Pe

rcen

tage

(%) No Tolerance

Full ToleranceBM+InterBpBM+IntraBp+InterBpIR+InterBpIR+IntraBp+InterBpIntraBp+InterBp


6 9 1235

40

45

50

55

Node Degree

Even

t Del

iver

y Pe

rcen

tage

(%)

IRIntraBpBM

6 9 1245

50

55

60

65

70

75

Node Degree

Even

t Del

iver

y Pe

rcen

tage

(%) BM+InterBp

BM+IntraBp+InterBpIR+InterBpIR+IntraBp+InterBpIntraBp+InterBp

Fig. 9. (a) Effect of node degree on event delivery percentage, (b) effect of node degree on event delivery percentage (cont.).






percentage is decreasing by the collapse percentage, the energyconsumption of delivered messages decreases. These figures showthat our proposed techniques consume far less energy than GR andconsume similar energy to IR and BM.

Figure 13(a) and (b) displays the effect of node degree onenergy consumption per node when the corresponding faultrecovery technique is performed at 20% node collapse percentage,5 cluster depth and 500 nodes. Energy consumption of all fault-tolerant techniques increases with node degree since a message isreceived by all neighbors of a sender node. We may conclude asbefore, for Fig. 13(a) and (b) that the energy consumption perfor-mance of our techniques outperforms GR and it has a similarperformance to IR and BM.

From obtained measurements, we can firstly state that theenergy consumptions of IntraBp and InterBp methods are scalableand stable against node count and degree. The reason for thisobservation is that their total message transfer is bounded by thecluster size and the count of clusters in the same level. SinceIntraBp and InterBp find more alternative paths than IR and BM asshown in the previous section, they consume slightly more energythan IR and BM. On the other hand, they consume far less energythan GR because they diffuse their messages relatively to thesmaller parts of the network whereas GR may cause depletion ofenergy in all nodes of the network.

5.4. Delays

One of the important criteria in sensor network applications isthe event delivery delay, so we measure the delay values of thealgorithms. The figures on delay include only the base fault-tolerant techniques except BM because it is usually formed withthe paths. Since BM does not need a time for recovery, it is notneeded to be displayed. Figure 14(a) displays the effect of d cluster

depth on delay of fault recovery techniques at 20% node collapsepercentage for 500 nodes. Figure 14(a) shows us that IR is notaffected by the cluster size, whereas IntraBp and InterBp areaffected by the cluster size. Since the cluster size increases whilethe depth value is increasing, IntraBp spends more time in order tofind the clusterhead. Similarly, InterBp spends more time, sincesearch messages disseminate over more number of nodes and theclusterhead count decreases. The delay increase in IntraBp andInterBp is linear so we may claim that the proposed approachesare scalable in terms of delay against varying depth values.

Figure 14(b) displays the effect of node degree on delay of faultrecovery techniques at 20% node collapse percentage for 500nodes. The delay of IntraBp and InterBp is not affected by nodedegree as shown in Fig. 14(b). On the other hand, time consump-tion of GR is very high especially for the sparsely connectednetworks. Figure 15(a) displays the effect of node collapse percen-tage on delay of fault recovery techniques at 20% node collapsepercentage for 500 nodes. In Fig. 15(a), the delay of IR and IntraBpis stable against varying node collapse percentages because thesetechniques are performed in a designated area. On the other hand,delay of InterBp is increasing while node collapse percentageincreases because probability of finding a close alternative clusteris decreasing with the node collapse percentage.

Figure 15(b) displays delay of fault recovery techniques fordifferent node counts ranging from 100 to 500 nodes. Node countmetric shows the scalability of fault-tolerant techniques fordifferent node counts. Figure 15(b) shows us that the delays ofIR, IntraBp and InterBp are nearly constant against node countwhereas the delay of GR is increasing with the node count sinceGR needs network-wide execution.

The results measured in this section show us that IntraBp andInterBp are scalable in terms of event delivery delay againstvarying node degrees, node counts, cluster counts and clustersizes which conforms to Theorems 2 and 4. The time consump-tions of the IntraBp are slightly more than those of IR since allcluster links are regenerated in IntraBp. InterBp performs far morebetter than GR because when the GR technique is used, the wholenetwork has to wait since all the paths are regenerated. But,InterBp is only performed only in a part of network and othernodes can continue to relay messages to sink.

5.5. Performance evaluations on the CPEQ protocol

In this simulation study, we extended the CPEQ protocol withthe proposed InterBp fault-tolerant technique. Then, we comparethe CPEQ protocol with the extended version of it which isdenoted by CPEQ þ InterBp. CPEQ does not need IntraBp becauseclusters are periodically re-formed. With this simulation study, weshow that how these proposed techniques can be easily beinserted into any cluster-based routing protocol in the literature.

3 5 76

6.5

7

7.5

8

8.5

9

d Depth Value

Ener

gy (j

oule

)

BM+InterBpBM+IntraBp+InterBpIR+InterBpIR+IntraBp+InterBpIntraBp+InterBpIntraBp

Fig. 10. Effect of d cluster depth on energy consumption.

100 200 300 400 500123456789

1011

Node Count

Ener

gy (j

oule

)

BMGRIRIntraBpNT

100 200 300 400 500123456789

1011

Node Count

Ener

gy (j

oule

)

NTBM+InterBpBM+IntraBp+InterBpIR+GRIR+InterBpIR+IntraBp+InterBpIntraBp+InterBpBM+GRGR

Fig. 11. (a) Effect of node count on energy consumption, (b) effect of node count on energy consumption (cont.).






10 20 306

6.57

7.58

8.59

9.510

10.511

11.5

Collapse Percentage

Ener

gy (j

oule

)

BMGRIRIntraBpNT

10 20 306

6.5

7

7.5

8

8.5

9

9.5

10

Collapse Percentage

Ener

gy (j

oule

) NTBM+InterBpBM+IntraBp+InterBpIR+GRIR+InterBpIR+IntraBp+InterBpIntraBp+InterBpBM+GR

Fig. 12. (a) Effect of node collapse percentage on energy consumption, (b) effect of node collapse percentage on energy consumption (cont.).

6 9 125

6

7

8

9

10

11

Node Degree

Ener

gy (j

oule

)

BMGRIRIntraBpNT

6 9 125

5.5

6

6.5

7

7.5

8

8.5

9

9.5

10

Node Degree

Ener

gy (j

oule

) NTBM+InterBpBM+IntraBp+InterBpIR+GRIR+InterBpIR+IntraBp+InterBpIntraBp+InterBpBM+GR

Fig. 13. (a) Effect of node degree on energy consumption, (b) effect of node degree on energy consumption (cont.).

3 5 70.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

d Depth Value

Del

ay (s

econ

d)

IRIntraBpInterBp

6 9 120

0.5

1

1.5

2

Node Degree

Del

ay (s

econ

d)

IRIntraBpInterBpGR

Fig. 14. (a) Effect of d cluster depth on delay, (b) effect of node degree on delay.

10% 20% 30%

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Collapse Percentage

Del

ay (s

econ

d)

IRIntraBpInterBp

0 100 200 300 400 5000

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

Node Count

Del

ay (s

econ

d)

IRIntraBpInterBpGR

Fig. 15. (a) Effect of node collapse percentage on delay, (b) delay of fault-tolerant techniques at different node counts.






CPEQ is a periodic, event-driven and query-based protocolwhich includes fault-tolerant and low-latency algorithms thatmeet sensor network requirements for critical conditions. It is acluster-based routing protocol that groups sensor nodes to effi-ciently relay the sensed data to the sink by uniformly distributingenergy dissipation among the nodes. CPEQ uses the publish/subscribe paradigm to disseminate requests across the network.

CPEQ starts with building the hop tree which is a floodingbased algorithm. This algorithm is similar to the shortest hopalgorithm used in the generic protocol. The only difference is thatthe algorithm in CPEQ does not construct any path toward sink. Itsaim is to configure nodes in the network by assigning hop levels.Then, the sinks in the network subscribe to the nodes bydisseminating subscription messages. While the subscriptionmessage is disseminated in the network, paths toward sink areconstructed. Each node maintains a subscription table and itforwards the events by looking at this table.

Clusters are re-formed per fixed time period in the CPEQprotocol. The clustering algorithm starts by designating aggregatornodes. In order to designate the aggregator nodes at each cluster-ing period, each node generates a random number between 0 and1. The nodes which generate a number lower than 0.05 willrequest the energy level from its immediate neighbors. The nodesthen select the neighbor with more energy as an aggregator andinform it. The newly selected aggregator node is responsible forforming the cluster. The cluster configuration is performedthrough the broadcasting of a notification packet. In order to limitthe size of a cluster, the notification packet carries a time to live(ttl) field. We set the value of ttl value as 3 in our simulations.

When a node detects an event in the environment, the senseddata is sent to the cluster aggregator node. Then, the aggregator noderelays the event toward the sink using the path constructed in thesubscription phase. Timer and acknowledgement mechanisms areenabled after a message is transmitted for reliable communication.

The percentage of aggregator nodes is selected as 5% in oursimulations.

Although the CPEQ protocol includes fault-tolerant and low-latency algorithms, its effectiveness decreases in the presence ofholes caused by group collapses because it uses instant recovery inthe protocol. Therefore, we extended the CPEQ protocol by insert-ing the proposed techniques in this paper. Figures 16 and 17(a),(b) display the event delivery percentage comparison of CPEQ andCPEQ þ InterBp during 10%, 20% and 30% node collapse percen-tages respectively. Three figures definitely show that the InterBpfault-tolerant technique increases the event delivery percentage ofCPEQ protocol. InterBp provides 7–18% event delivery percentageincrease in each node collapse percentage. Figures 18 and 19(a),(b) display the energy consumption per node comparison of CPEQand CPEQ þ InterBp during 10%, 20% and 30% node collapsepercentages respectively.

These figures show us that the InterBp technique increases theenergy consumption per node 4% in the worst case, 2% in theaverage case. In fact, since CPEQ þ InterBp recovers more fault andprovides a significant increase in event delivery, more events arerelayed toward sink. Therefore, in order to relay more events, moreenergy is consumed. However the consumed energy is notsignificant when compared to the increase in event delivery. Thereason of this observation is InterBp that tries to reconstruct a newpath only within a few levels of clusters and reconstructs a newpath immediately if it is found.

6. Conclusions

In this paper, we first model holes as clusters to convert theproblem of hole bypassing to cluster bypassing. Intra-cluster andInter-cluster bypass algorithms are proposed for bypassing the

100 200 300 400 50060

65

70

75

80

85

90

95

Node Number

Even

t Del

iver

y Pe

rcen

tage

(%)

CPEQCPEQ+InterBp

Fig. 16. Event delivery percentage comparison of CPEQ and CPEQ þ InterBp at 10%collapse.

100 200 300 400 50040

50

60

70

80

90

Node Number

Even

t Del

iver

y Pe

rcen

tage

(%)

CPEQCPEQ+InterBp

100 200 300 400 50030

40

50

60

70

80

Node Number

Even

t Del

iver

y Pe

rcen

tage

(%)

CPEQCPEQ+InterBp

Fig. 17. (a) Event delivery percentage comparison of CPEQ and CPEQ þ InterBp at 20% collapse, (b) event delivery percentage comparison of CPEQ and CPEQ þ InterBpat 30% collapse.

100 200 300 400 50016

17

18

19

20

21

22

23

24

25

Node Number

Ave

rage

Ene

rgy

Con

sum

ptio

n (jo

ule)

CPEQCPEQ+InterBp

Fig. 18. Energy consumption (per node) comparison of CPEQ and CPEQ þ InterBpduring 10% collapse.






100 200 300 400 50016

17

18

19

20

21

22

23

24

25

Node Number

Ave

rage

Ene

rgy

Con

sum

ptio

n (jo

ule)

CPEQCPEQ+InterBp

100 200 300 400 50016

17

18

19

20

21

22

23

24

25

Node Number

Ave

rage

Ene

rgy

Con

sum

ptio

n (jo

ule)

CPEQCPEQ+InterBp

Fig. 19. (a) Energy consumption (per node) comparison of CPEQ and CPEQ þ InterBp during 20% collapse, (b) energy consumption (per node) comparison of CPEQ and CPEQþ InterBp during 30% collapse.

holes inside the cluster and holes which cover more than onecluster. The algorithms are analyzed, evaluated with extensivesimulations and compared with well-known fault-tolerant algo-rithms. In the performance evaluations, two simulation studiesover a generic protocol and CPEQ (Boukerche et al., 2006a) havebeen conducted in order to show the effectiveness and applic-ability of algorithms, and how the algorithms increase the eventdelivery percentage of CPEQ. We also showed from the simula-tions that the event delivery percentages and resource usages ofthe proposed methods conform to the theoretical analysis.

Although fault-tolerant routing techniques have been studied inthe literature, localization-free and energy-efficient hole bypassinghas not been studied so far. Disjoint and braided multipaths, andinstant recovery algorithms aim to recovery the isolated failures.Unfortunately, these algorithms cannot recover the faults occurringby group collapses. Fang et al. (2006), Yu et al. (2007, 2008, 2009a),You et al. (2009b), Shiow-Fen et al. (2010) proposed algorithms andprotocols for bypassing the holes which are formed by groupcollapses, deployment etc. But, the algorithms in these studies usethe localization techniques. Therefore, they are not applicable tomany WSN applications.

The simulation results show that intra-cluster and inter-clusterbypass algorithms provide very close performance to Full Toler-ance in terms of fault-tolerance while they consume significantlylower energy than Full Tolerance. When they are compared withthe well-known fault-tolerant techniques, they provide up to 25%better fault-tolerance as well as reasonable delay and energyconsumptions.

In addition, intra-cluster and inter-cluster bypass algorithmscan be easily integrated into many protocols in the literature. Sincethey can be performed adaptively, they increase the fault-tolerance with very low energy consumption. In the simulationstudy on the CPEQ protocol, it is obviously seen that the inter-cluster bypass algorithm increases the event delivery percentageof CPEQ protocol with insignificant energy consumption.

References

Akkaya K, Younis MF. A survey on routing protocols for wireless sensor networks.Ad Hoc Networks 2005:325–49.

Boukerche A, Werner Nelem Pazzi R, Borges Araujo R. Fault-tolerant wirelesssensor network routing protocols for the supervision of context-aware physicalenvironments. Journal of Parallel and Distributed Computing 2006a;66:586–99.

Boukerche A, Chatzigiannakis I, Nikoletseas S. A new energy efficient and fault-tolerant protocol for data propagation in smart dust networks using varyingtransmission range. Computer and Communications 2006b;29:477–89.

Boukerche A, Martirosyan A, Pazzi R. An inter-cluster communication based energyaware and fault tolerant protocol for wireless sensor networks. MobileNetworks and Applications 2008;13:614–26.

Braginsky, D., Estrin, D. Rumor routing algorithm for sensor networks. In: Proceed-ings of the 1st ACM International Workshop on Wireless Sensor Networks andApplications, WSNA '02, ACM, New York, NY, USA; 2002. p. 22–31.

Chatzigiannakis I, Kinalis A, Nikoletseas S. Fault-tolerant and efficient datapropagation in wireless sensor networks using local, additional networkinformation. Journal of Parallel and Distributed Computing 2007;67:456–73.

Dagdeviren O, Erciyes K. Graph matching-based distributed clustering and back-bone formation algorithms for sensor networks. The Computer Journal2010;53:1553–75.

Dai H, min Zhu Z, Gu X-F. Multi-target indoor localization and tracking on videomonitoring system in a wireless sensor network. Journal of Network andComputer Applications 2013;36(1):228–34.

Erciyes, E., Ozsoyeller, D., Dagdeviren, O. Distributed algorithms to form clusterbased spanning trees in wireless sensor networks. In: Proceedings of the 8thInternational Conference on Computational Science, Part I, ICCS '08. Berlin,Heidelberg: Springer-Verlag; 2008. p. 519–28.

Fang Q, Gao J, Guibas LJ. Locating and bypassing holes in sensor networks. MobileNetworks and Applications 2006;11.

Ganesan, D., Govindan, R., Shenker, S., Estrin, D. Highly-resilient, energy-efficientmultipath routing in wireless sensor networks. In: Proceedings of the 2nd ACMInternational Symposium on Mobile Ad Hoc Networking & Computing, Mobi-Hoc '01. ACM, New York, NY, USA; 2001 p. 251–4.

Hadjidj A, Souil M, Bouabdallah A, Challal Y, Owen H. Wireless sensor networks forrehabilitation applications: challenges and opportunities. Journal of Networkand Computer Applications 2013;36(1):1–15.

Heinzelman, W.R., Kulik, J., Balakrishnan, H. Adaptive protocols for informationdissemination in wireless sensor networks. In: Proceedings of the 5th AnnualACM/IEEE International Conference on Mobile Computing and Networking.MobiCom '99, ACM, New York, NY, USA; 1999. p. 174–85.

Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H. Energy-efficient communica-tion protocol for wireless microsensor networks. In: Proceedings of the 33rdHawaii International Conference on System Sciences, vol. 8. Washington, DC,USA: HICSS '00, IEEE Computer Society; 2000. p. 8020.

Intanagonwiwat C, Govindan R, Estrin D, Heidemann J, Silva F. Directed diffusion forwireless sensor networking. IEEE/ACM Transactions on Networking 2003;11:2–16.

Karl H, Willig A. Protocols and architectures for wireless sensor networks. JohnWiley & Sons; 2005.

Lin, K. Research on adaptive target tracking in vehicle sensor networks. Journal ofNetwork and Computer Applications 2013;36(5):1316–23.

Manjeshwar A, Agrawal DP. Teen: a routing protocol for enhanced efficiency inwireless sensor networks. International Parallel and Distributed ProcessingSymposium 2001;3(30189a):195–202.

Manjeshwar, A., Agrawal, D. Apteen: a hybrid protocol for efficient routing andcomprehensive information retrieval in wireless sensor networks. In: Proceedings16th International Parallel and Distributed Processing Symposium, vol. 13. IEEEComput. Soc.; 2002. p. 195.

Mendes LD, Rodrigues JJ. A survey on cross-layer solutions for wireless sensornetworks. Journal of Network and Computer Applications 2011;34(2):523–34(Special issue on Efficient and Robust Security and Services of Wireless MeshNetworks).

Pantoni RP, Brando D. A gradient based routing scheme for street lighting wirelesssensor networks. Journal of Network and Computer Applications 2013;36(1):77–90.

Paradis L, Han Q. A survey of fault management in wireless sensor networks.Journal of Network and Systems Management 2007;15:171–90.

Peleg, D. Distributed computing: a locality-sensitive approach. Philadelphia, PA,USA: Society for Industrial and Applied Mathematics; 2000.

Shiow-Fen, H., Chia-Hsuan, Y., Yi-Yu, S., Chyi-Ren, D. Energy efficient hole bypassingrouting in wireless sensor networks. In: Computer Science and InformationTechnology (ICCSIT), 2010 3rd IEEE International Conference on ICCSIT'10.Piscataway, NJ, USA: IEEE Press; 2010. p. 576–80.



http://refhub.elsevier.com/S1084-8045(13)00190-2/sbref1












http://refhub.elsevier.com/S1084-8045(13)00190-2/othref0005



































































Tubaishat M, Madria S. Sensor networks: an overview. IEEE Potentials 2003;22(2):20–3.

Wagner D, Wattenhofer R. Algorithms for sensor and ad hoc networks: advancedlectures. Berlin, Heidelberg: Springer-Verlag; 2007.

Yick J, Mukherjee B, Ghosal D. Wireless sensor network survey. Computer Net-works 2008;52:2292–330.

You, J., Lieckfeldt, D., Reichenbach, F., Timmermann, D. Context-aware geographicrouting for sensor networks with routing holes. In: Proceedings of the 2009IEEE Conference on Wireless Communications & Networking Conference,WCNC'09. Piscataway, NJ, USA: IEEE Press; 2009. p. 2589–94.

Yu, F., Lee, E., Choi, Y., Park, S., Lee, D., Tian, Y., Kim, S.-H. A modeling for holeproblem in wireless sensor networks. In: Proceedings of the International

Conference on Wireless Communications and Mobile Computing. New York,NY, USA: IWCMC '07, ACM; 2007. p. 370–5.

Yu, F., Park, S., Tian, Y., Jin, M., Kim, S.-H. Efficient hole detour scheme forgeographic routing in wireless sensor networks. In: Proceedings of the 67thIEEE Vehicular Technology Conference, VTC Spring, 11–14 May 2008, Singapore.IEEE; 2008. p. 153–7.

Yu F, Park S, Lee E, Kim S-H. Hole modeling and detour scheme for geographicrouting in wireless sensor networks. Journal of Communications and Networks2009a:327–36.




























localization-free and energy-efficient hole bypassing techniques for fault-tolerant sensor networks

Documents