Computer Networks 53 (2009) 2840–2854

Contents lists available at ScienceDirect

Computer Networks

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

SubCast: A distributed addressing and routing system for large scalewireless sensor and actor networks

Therence Houngbadji *, Samuel Pierre 1

Department of Computer Engineering, École Polytechnique de Montréal, P.O. Box 6079, Centre-Ville Station, Montreal, Quebec, Canada H3C 3A7

a r t i c l e i n f o

Article history:Received 18 March 2008Received in revised form 6 February 2009Accepted 14 July 2009Available online 18 July 2009Responsible Editor: R. Sivakumar

Keywords:Wireless sensor and actor networksDistributed address allocationIterated function systemsPublish/subscribeRouting

1389-1286/$ - see front matter � 2009 Elsevier B.Vdoi:10.1016/j.comnet.2009.07.006

* Corresponding author. Tel.: +1 514 739 8302; faE-mail addresses: therence.houngbadji@polymt

Samuel.Pierre@polymtl.ca (S. Pierre).1 Tel.: +1 514 340 3240x4685; fax: +1 514 340 32

a b s t r a c t

Wireless sensor and actor networks (WSANs) are made up of a large number of sensingdevices which are resource-impoverished nodes and powerful actuation devices: bothare equipped with computation and communication capabilities. These devices cooperateto manage sensing and perform acting tasks. Numerous work conducted in the field ofWSANs assumes the existence of addresses and routing infrastructure to validate their pro-posals. However, assigning addresses and delivering detected events in these networksremains highly challenging, specifically due to the sheer number of nodes. To address theseissues, this paper proposes SubCast, a novel distributed address assignment and routingscheme based on a topic clustering system and fractal theory iterated function systems. Inorder to minimize data delivery costs among actors, the proposed architecture first buildsan actor overlay network before allocating addresses to network nodes. Location informa-tion in the allocated addresses allows establishing data delivery paths. Simulation resultsconfirm that the proposed system efficiently guarantees the allocation of unique addressesand performs efficient data delivery while reducing communication costs, delays as well asthe impact of imprecise locations.

� 2009 Elsevier B.V. All rights reserved.

1. Introduction

Recent innovations in wireless communication technol-ogies have raised interest in Wireless sensor and actor net-works (WSANs), which refers to a group of sensors andactors linked through a wireless medium to perform dis-tributed sensing and actuation tasks. Sensors gather infor-mation from the physical world (data regardingtemperature, pressure, movement, light, etc.), while actorsmake decisions and perform relevant actions upon theenvironment by means of actuators, which allow remote,automated interactions with the environment. Actors areresource-rich devices with networking-related functional-ities and extended battery life. They are more expensivethan sensors associated with low cost, inadequate power

. All rights reserved.

x: +1 514 340 3240.l.ca (T. Houngbadji),

40.

and short transmission range. Furthermore, the numberof sensor nodes deployed in a target area may be in the or-der of hundreds or thousands, while such a dense deploy-ment is usually unnecessary for actor nodes [1].

The collaborative sensing and acting tasks of WSANnodes enables sensors and actors to establish data paths.Also, a coordination mechanism between actors allowsthem to both estimate the characteristics of events and se-lect actors or subsets of actors which are best suited to per-form a given action [2,3]. As such, WSANs can be regardedas a self-organized, event-sensing, communicating anddecisive action loop. Aside from the WSANs coordinationframework, its network layer also deserves special atten-tion. Usually, node identifiers are assumed to exist. How-ever, the way addresses are allocated remains an openissue. Indeed, addresses can be used to establish a node po-sition in the physical world, which is useful to identify thephysical source of sensed data, especially with mobilenodes or unplanned node placements. The fact that con-trolled communication is considered more effective than

T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854 2841

flooding makes address allocations an outstanding routingsupport. That was the case for directed diffusion [4] andTAG [5], for which gradient and tree-based addressingand routing were used. Addressing is required to preventredundant ad-hoc communication between sensor nodes.Reducing intra-network transmission can save node en-ergy, thus extending network lifetime. During operationsand control, specific nodes need to be addressed in orderto update software modules, calibrate sensors and performlocalization tasks.

The special data reporting behavior and the need to se-lect a given actor to perform an actuation task makesWSAN address assignment problems significantly differentthan those of traditional networks. For applications inwhich events take place in different locations, it may alsobe necessary to send events to a set of actors which maybe far from or even totally outside the event area. The clos-est actor in the event area may not be the most appropriateone to handle properly the required action if commandsfrom suitable actors that can perform the set of tasks tobe accomplished in the event area are lacking. In such acontext, event dissemination to actors can be triviallyimplemented by flooding all events to all network actors.Although such a solution may seem trivial, its cost is veryexpensive as all actors are affected every time an event isdetected. This raises the issue of delivering events to themost appropriate network actors.

In order to tackle the issue of addressing and data deliv-ery in WSANs, this work proposes a distributed addressingand routing system. For that purpose, fractal theory iter-ated function systems (IFSs) [6] is used to discretize thenetwork area into micro-scale areas, referred as cells thatoffer a distributed low-cost address allocation to networknodes. In order to solve event dissemination issues amongactors, they become clustered according to their own inter-ests’ subscription. Actors are grouped into topic segmentsthat form what is referred as the actor overlay network(AON). Called SubCast, the overall architecture guaranteesunique identifications for every node, as well as routinginformation for all possible packet delivery. The remainderof this paper is organized as follows. Section 2 provides anoverview of existing work pertaining to node addressing.Section 3 offers an architectural overview of SubCast andSection 4 details its architecture functionalities. Section 5presents results ensuing from a performance evaluation,a simulation result analysis. In Section 6 conclusion shedslight on future work.

2. Related work

Although a few recent papers delve specifically intonodes addressing, the literature pertaining to addressingin wireless sensor networks (WSNs) is extremely limited,and that of addressing in wireless sensor and actor net-works remains quasi non-existent, in spite of being de-scribed as an open research issue at the network layer in[1]. A number of solutions have been proposed to allocateaddresses dynamically. The best-known approach is thedynamic host configuration protocol (DHCP) [7] used inIPv4 and IPv6 networks, which consists of a client–server

protocol composed of two major building blocks: a proto-col to deliver specific parameters to the client and an IP ad-dress selection mechanism. Although the dynamic profileof this protocol is adequate for WSANs, it is hindered bythe huge overhead needed to allocate addresses. Signalingoverhead represents a major source of misused energy. Inaddition, the centralized address management schemeused by this protocol cannot scale well in highly denseWSANs. Although a DHCP-Like addressing structure wasproposed for WSNs, it cannot be qualified as perfect. Theaddressing systems in [8–11] are not scalable nor energyefficient. Their allocation process includes the exchangeof numerous messages that cause serious energy burdens.More particularly, Yao et al. proposal [10], a round-basedallocation process, deals with duplicate detection of allo-cated addresses, causing the exchange of numerous mes-sages. The DHCP-like addressing approach can beappealing for low scale WSNs, yet it must be adapted forlarge networks, such as WSANs, in order to generate lowersignaling overhead. In [11], the authors propose a dynamicaddress allocation scheme for query-based sensor net-works. A temporary network-wide unique address is allo-cated on demand, solely to sensor nodes that report datain response to an explicit query from the sink. The ap-proach splits the geographic area into blocks and 2D logicalgrids, so that each node lies within a single grid. Uniqueaddresses are assigned locally to sensors positioned withina cell of a predetermined size. This approach fails to lightenthe heavy overhead used for allocating addresses. In addi-tion, the central approach adopted to allocate active ad-dresses to nodes may create bottlenecks as the networkexpands.

Others addressing systems [12–15] divide the networkinto layers, and the sensor nodes add a suffix to addressesof nodes in last layer to form their own addresses. In such alayered allocation process, addresses become increasinglylonger as the network expands. More particularly, theTreeCast addressing scheme [12], a global addressing andstateless routing architecture for sensor networks, adoptsthis strategy. In this scheme, nodes are organized in a treestructure and addresses are assigned according to nodepositions. Child nodes randomly generate their own ad-dresses, which are subsequently approved by parentnodes. TreeCast does not scale well in large networks asthe parents’ address strictly consists of the prefix of theirchildren’s address, thus increasing address lengths as nodelevels grow. Following the same design approach, the Zig-Bee network layer [13] provides a distributed addressassignment while routing management handles tree-rout-ing mechanisms as well as AODV-like reactive routingmechanism. Before forming a network, the ZigBee coordi-nator determines the maximum number of children for asingle router, the maximum number of child routers for asingle router, as well as the network depth. Devices’ ad-dresses are assigned in a top-down manner. For the coordi-nator, the whole address space is logically partitioned intoblocks: some blocks are assigned to the coordinator’s childrouters while others are reserved for the coordinators’ ownchild end devices. The hierarchical block addressingmechanism of ZigBee protocol is subject to the waste ofthe 16-bit address space (especially for small networks),

2842 T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854

the limitation of the network expandability (limitation onthe maximum number of children) and the lack of devicemobility.

Other works offer a new design avenue using cluster-based address allocation. Such a strategy is scalable andthus deserves a closer look. Works in [16–18] have adoptedthis design philosophy.

3. Design overview

By examining the existing work listed above, it appearsthat address assignment in WSANs remains largely anopen issue. This section provides the design rationale ofthe components used by the proposed addressing system.As pointed out by Akyildiz and Kasimoglu [1], in certainapplications, when events take place in different locations,it may also be necessary to pass events to the set of actorswhich may be far or outside the event area when the eventis detected, but rather to the set of actors which is closer tothe event when it is notified. Indeed, without commandsfrom suitable actors that can issue the necessary set oftasks, the closest actor in the event area may not be ableto handle properly the required action. In such a context,the event disseminated to a set of actors can be triviallyimplemented by flooding each event among actors, so thatthey can filter out events that do not match their interests’subscription. However, flooding among actors is costly, aseach actor is affected every time an event is detected. Toaddress such issues, an actor overlay network (AON) isbuilt, based on the publish/subscribe (PS) communicationparadigm. Indeed, the PS paradigm was proven to let infor-mation propagate from publishers to interested subscrib-ers in an anonymous and decoupled fashion [19]. Inorder to deliver events that match the actors’ specific sub-scriptions while keeping interest clustering in mind, it isnecessary to use a system with traffic confinement to buildan AON. Actors that share common interests are clusteredso that once an event reaches a cluster member, its dissem-ination can be limited to the cluster itself. This way, thecluster member that can act properly is immediatelyselected.

The second design goal of SubCast is to allocate ad-dresses to network nodes. As discussed above, previousworks on address allocation focuses only on address allo-cation in WSNs without integrating how the allocationprocess can affect the moving actor nodes in WSANs. Fur-thermore, numerous studies conducted in WSANs [2,3],Vedantham et al. [20,21] assume the existence of nodeidentifiers and routing protocol to validate their proposal.To solve the issue, this paper proposes a new addressingand routing protocols for WSANs, based on fractal theoryiterated function system (IFS) [6]. The usability in mobileor indoor environments of GPS-enabled devices appearsto be costly in real deployments even if the solution isattractive. Location approximation using localization algo-rithm can help achieve low-cost address allocation despitetheir lack of precision. The conjecture of this study is thatpartitioning the network into micro-scale areas with aminimal amount of physical location information, even ifinaccurate, can offer real-world decentralized addressing

system in large scale WSANs regarding the analysis ofexisting approaches. That is where the use of IFS becomesa valuable choice to discretize the network area with zerooverhead. Indeed, it offers the possibility of using approx-imate or relative location information to determine nodeareas. Accordingly, it is more robust to errors and impreci-sions in location measurements and estimates comparedto schemes that depend on exact location information. Asan outstanding feature, IFS generate an explicit area label-ing based on a predefined coded alphabet. The partitionsobtained provide a dampening factor that reduces the ef-fects of mobility. Local node movements, meaning move-ments within their own regions, do not affect theSubCast addressing structure. Finally, in addition to allow-ing the design of a distributed address allocation scheme,the use of IFS enables a new generic event routing strategyfor WSANs. Thus, the following design objectives are met:

(a) Low-cost address allocation: less signaling overheadis generated.

(b) A unique address is allocated to each node in thenetwork.

(c) Event routing is performed by using only the node’sneighborhood; path discovery to the destination androuting table maintenance are not needed.

Fig. 1 gives an overview of the building blocks that com-pose the SubCast network layer, as well as their respectivefunctionalities. In the following section, details are pro-vided on the way the illustrated functionalities are built.

4. System model

The assumptions pertaining to the SubCast addressingsystem are the following: the WSAN is deployed overtwo-dimensional areas and the architectures described in[1] are considered: a semi-automated architecture wheresensors route data back to the sink, that, in turn, may issueaction commands for actors. The automated architecture al-lows sensors to send their readings to the actors which, inturn, can trigger appropriate actions. It is assumed thatsensor nodes are static and aware of their own approxi-mate location through localization algorithms [22,23].Note that such an assumption is not specific to our pro-posal, as it appears in other WSAN works [20,21]. Forexample, actor nodes can cooperate with the sink in orderto help locate a node, if need be. Sensor nodes can reach ac-tors and sink through single or multi-hop communication.Actors can move freely and are assumed to have the sametransmission range as sensors, although they have greatercomputation resources. It is assumed that every node isassociated with a serial number as soon as it ismanufactured.

4.1. The actors’ overlay network (AON)

This section presents the design of the AON compo-nents. To this end, as mentioned above, the well-knowncommunication paradigm called publish/subscribe is con-sidered [19]. The PS paradigm consists of a collection of

Fig. 1. SubCast components and their functionalities.

T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854 2843

clients that interacts by publishing messages and subscrib-ing to classes of messages that perk their interest. Inter-ested parties specify the events in which they areinterested by using the subscribe function that lists theirtopics of interest. Objects of interest publish events notifi-cations via the publish function. Publishers do not keepsubscribers’ references and vice versa, as an entity, called‘’the broker”, coordinates interactions, making sure thatmatching events are delivered to the relevant parties. PSsystems are classified according to how their data descrip-tion styles are expressed: either topic-based or content-based systems [19]. Content-based PS systems are moreexpressive, although they often require a complex imple-mentation due to the nature of the filters. However, AONadopts on topic-based systems, a relatively simple class ofPS system. In topic-based PS systems, users subscribe totopics in order to be notified of matching events. The sub-scription syntax is limited to a simple test on a specificfield of the event.

4.1.1. Topic-based PS system for WSANsThe actor nodes which represent subscribers transmit

their subscription messages ‘‘Subscribe” to the sink node.Such messages are composed of identifiers that specifythe actor’s selected topics of interest. Subscriptions areregistered with the sink, which computes the topic seg-ments by gathering actors that share the same interests.Sensors, which represent the event sources, generate datain response to variable changes that they monitor in thereal-world, such as temperature changes, voice degrada-tion, etc, thus representing the publishers. ‘’Publish” mes-sages include the sensor address, the topic identifier, theevent data, as well as a sequence number. The event broker,an agent located on each sensor node, acts as an interfacebetween the event publisher and the subscriber agent forthe actor nodes. More specifically, the objective consistsof clustering actors which share common interests. Giventhis perspective, a mathematical formulation of the inter-est clustering problem is presented. The following sectiondefines certain notations. Let SA denotes the set of actorsðna ¼ jSAjÞ and T the set of topics. Each actor a 2 SA sub-scribes to ns topics, represented by a binary vector½xa;1; xa;2; . . . ; xa;ns � where:

xa;t ¼1; if the actor a subscribes to the topic t;

0; otherwise:

�

Let subscriptions for two different actors a and b, repre-sented by the vectors ½xa;1; xa;2; . . . ; xa;ns � and½yb;1; yb;2; . . . ; yb;ns

�, respectively. The similarity measure be-tween them can be defined as the total matches of corre-sponding topics for both vectors. The matching similaritymeasure between both actors’ subscriptions is defined asfollows:

dða; bÞ ¼Xns

t¼1

dðxa;t; yb;tÞ ð1Þ

where

dðxa;t ; yb;tÞ ¼1; xa;t ¼ yb;t;

0; xa;t–yb;t;

(1 6 t 6 ns ð2Þ

The lower the number of mismatches, the most similarboth objects.Let subset Z # SA of actors represent the clus-ters centers, and a subset D # SA of actors must be assignedto such centers. Furthermore, if actor nodes are clusteredaccording to their interests, one cluster can happen to bemuch bigger than the others, raising the costs of eventdelivery among cluster members. To solve this issue andfavor clusters of similar sizes, clustering entropy is takeninto account. Indeed, entropy quantifies the amount ofuncertainty present in the state of a variable or, in thisapplication, uncertain topic clustering. The entropy of Z isdefined as HðZÞ ¼ �

Pi2ZpilogðpiÞ where pi denotes the per-

centage of nodes in the cluster represented by its center zi.Let nc denote the number of clusters ðnc ¼ jZjÞ. Entropyreaches its maximal value logðncÞ when all clusters areequally probable, meaning of identical size. The relativeentropy function [30] is defined as follows:

HrðZÞ ¼ �P

i2ZpilogðpiÞlogðncÞ

; 0 < pi 6 1 ð3Þ

and has values between [0,1]. When the value of this func-tion is close to 1, cluster sizes are similar to one another. Infact, the objective of the problem on hand, denoted as topicclustering problem (TCP), consists of finding the subset Zcomposed of nc centroids and the connection vector thatmaximizes: (a) the cost of logical connections across actorsand (b) the relative entropy. The mathematical model isthus presented as follows:

Maximize FðW; ZÞ ¼ HrðZÞ þXi2Z

Xi2D

wijdði; jÞ ð4Þ

2844 T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854

which is subject to:

wij 2 f0;1g; 8 i 2 Z;8 j 2 D ð4aÞXi2Z

wij ¼ 1; 8 j 2 D ð4bÞ

0 <Xi2Z

Zi 6 na ð4cÞ

W ¼ ½wij� depicts a nc � byna binary matrix and Z ¼½z1; z2; . . . ; znc � denotes the cluster centers’ vector. Eq. (4)represents the objective function of the model. Expressions4a, 4b and 4c translate the model constraints: constraint(4a) defines the scope of variables wij; constraint (4b) en-sures that each actor belongs to a single cluster; and con-straint (4c) specifies the bounds of the number of clusters.

4.1.2. Algorithms and complexityFor large scale networks composed of numerous actors,

such a clustering issue cannot be solved in polynomialtime. Obviously, TCP belongs to the class of facility locationproblems known to be NP-hard [24]. In the literature, K-means algorithm (KM) is usually considered to solve thistype of problem, as is the case when mining data fromlarge sets of data clustering [25]. The pseudo-code of theKM algorithm appears hereafter.

k-means algorithm

1. Choose an initial Zð1Þ and maximize FðW;Zð1ÞÞ toobtain Wð1Þ. Set t ¼ 1.

2. Let cW ¼Wt and solve FðcW;ZÞ to obtain Zðtþ1Þ.

2.1. If FðcW;ZtÞ ¼ FðcW;Ztþ1Þ, output (cW;Zt)and stop; otherwise goto 3.

3. Let bZ ¼ Ztþ1 and solve FðW; bZÞ to obtain Zðtþ1Þ.

3.1. If FðWt ; bZÞ ¼ FðWtþ1; bZÞ, output and stop;otherwise let t ¼ t þ 1 and goto 2.

Note that with the KM, although the number of clustersis predefined, it has the shortcoming of converging to localoptima. To escape such local optima and find a more suit-able solution, Tabu Search (TS) algorithm becomes aninteresting alternative. The tabu search proposed by Glover[26] is a meta-heuristic that guides the solution searchprocedure to explore the solution space beyond local opti-mality. TS uses a local or neighborhood search procedureto iteratively move within S’s neighborhood, from SolutionS to Solution S1, until a certain stopping criterion is satis-fied. As our main objective consists of building clusterscomposed of actors that share common interests, the high-er the clustering costs, the better the algorithm. TS’s abilityto find a better solution compared to KM is the ultimatereason that motivated its selection when building theAON. The proposed TS algorithm, called Tabu–TCP for theproblem on hand, adopts as stopping criterion a certainnumber of iteration,D, which is controlled by the networkoperator. The larger the number of iterations, the longerthe search, and possibly the most cost efficient solution.Moreover, the designed Tabu–TCP algorithm uses a greedyalgorithm to build the initial solution from which thesearch starts. The entire pseudo-code is illustrated in the

Tabu–TCP algorithm below. Once clusters are computed,the sink broadcasts the results, composed of the membersof each cluster and their associated topic vectors. As such,each actor becomes an access point (AP) for the event thatoccurs in its action area, which matches a topic that be-longs to its cluster, while sensor nodes publish the match-ing data to the nearest AP.

Cluster computations make it possible to deliver eventsto actors according to the type of detected event. The issueto be addressed consists of setting the communicationinfrastructures used to deliver notifications among actors,while ensuring interactions remain possible between allnodes involved: sensors, actors, and sinks. The parts of thisinfrastructure, namely the distributed address assignmentand the routing strategy, are detailed in the next sections.

Tabu–TCP algorithm

1. (WZ), the current solution;2. (W�;Z�) the best-known solution and FðW�;Z�Þ

its value;3. N(Z) the neighborhood of Z;4. eNðZÞ, the admissible subset of N(Z) (i.e., non-

tabu);5. T ¼£ The tabu list6. Compute initial solution ðW0;Z0Þ with a greedy

algorithm. Set Z ¼ Z0;Z� ¼ Z0 andF�ðW�;Z�Þ ¼ FðW0;Z0Þ;

7. While the termination criterion is not met, do7.1. Select ðW;ZÞ ¼ argmax

Z02eNðZÞ½FðW;Z0Þ�;7.2. If FðW;ZÞ > F�ðW�;Z�Þ, then set

F�ðW�;Z�Þ ¼ FðW;ZÞ and,ðW�;Z�Þ ¼ ðW;ZÞ;7.3. Record tabu for the current move inT;7.4. Update T;

8. End

4.2. Distributed address assignment

4.2.1. Basic ideaIn geometric design, IFSs are defined as subdivision

methods that construct fractals by uniting several copiesof themselves ad infinitum [6]. Each copy is transformedby a set of functions which stand for a finite number ofM transformations, defined as follows:

cuðx; yÞ ¼ aðauxþ buyþ eu; cuxþ duyþ fuÞ0 6 u 6 M � 1

ð5Þ

where au; bu; eu; cu; du; fu 2 R. For a given initial form A (e.g.square, triangle, etc.), the small affine copiesc1ðAÞ; c2ðAÞ . . . cMðAÞ are produced and turned into a newform. The mapping process computes recursivelyAk ¼

SMu¼1cuðAk�1Þ;8k > 1.The sequence fAkg converges to

a final set called the IFS attractor and also define a codespace C composed of the affine transformations indexesu. A sequence of indexes defines the address of each partof the resulting attractor. An example of this kind of con-struction appears in Fig. 2. This illustration depicts whatis called a Sierpinski triangle. The set of IFS describingthe fractal process of the Sierpinski triangle maps a basetriangle into three children by splitting each edge of the

Fig. 2. A Sierpinki triangle [6].

T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854 2845

base triangle in half. Three affine transformations are ap-plied repeatedly with each new triangle obtained, untilthe size of the triangle falls below the desired resolution.Thus, at each step, smaller triangles are created by joiningthe midpoints of the triangles created in the previous iter-ation. Fig. 2 shows the results of applying this process to abase triangle five times. A detailed formalization of IFS andfractals is presented in [6].

4.2.2. Address allocationTo apply the IFS to a WSAN area, four linear affine trans-

formations are defined to represent the mapping functionsas follows:

c00ðx; yÞ ¼ x2 ;

y2

� �c01ðx; yÞ ¼ x

2 ;y2þ L

2

� �c10ðx; yÞ ¼ x

2þ L20

y2þ L

2

� �c11ðx; yÞ ¼ x

2þ L20

y2

� �

8>>>><>>>>: ð6Þ

The initial set A0 ¼ fð0;0Þ; ðL;0Þ; ðL; LÞ; ð0; LÞg represents theWSANs area and these affine transformations denote anarea contraction by a factor of 1/2, followed by a transla-tion. The indexes of these transformations define a codespace C ¼ f00;01;10;11g. A square cell belonging to theresulting form is represented by a sequence of code/ ¼ /1/2/3 � � � ð/u�CÞ. The address allocation essentiallyfollows a three-step process: (1) cell computing, (2) neigh-borhood discovery, and (3) identifier assignment.

4.2.2.1. Cell computing. By network-wide flooding, the sinkstarts the allocation process by broadcasting the set A0.Each node recursively computes the cell to which itbelongs, by applying the defined transformations andcomparing its coordinates with the bounds of obtainedsub-square. Since sensor nodes cannot determine their celladdresses using an infinite alphabet sequence of the codespace C, the value of korder; which defines the number ofIFS iterations and the cell’s address length, are chosen fi-nite and bounded by taking into account the sensors’ trans-mission range r. Indeed, to allow node communicationwithin a cell, the worst case scenario is considered: in acell, the distance between two nodes equals the cell’s diag-

Fig. 3. (A) One-order attractor, (B) four-order cells of attr

onal, so that the cell’s side lcell is bounded as lcell 6rffiffi2p . Since

lcell ¼ L=2korder , the number of sequences korder in the cell ad-dress can be chosen so that:

korder P12þ log2

Lr

� �: ð7Þ

The cell’s address thus consists of the first part of thenode’s address. An address /cell of any cell consists of a fi-nite sequence of elements of the code space C and is de-fined as follows:

/cell ¼ /1/2/3 � � �/korder;/u�C ð8Þ

Fig. 3 illustrates the construction process for r = 40 m andL = 200 m. Sub-squares represent the cells, and points de-ployed sensor nodes in the resulting attractor (Fig. 3C).With such physical parameters, the recursion orderamounts to korder ¼ 3 and the cell’s address length equals6 bits. Fig. 3A shows a one-order attractor while Fig. 3Bprovides an overview of the same attractor at a four-orderlevel in sub-square ‘‘00”. With this mechanism, nodes cell’saddresses become /cell ¼ 000001;100011; . . . ;111111.

4.2.2.2. Neighborhood discovery. When it finishes comput-ing its cell address, the node initiates a neighborhood dis-covery process by including its serial number and thecomputed cell’s address in a ‘‘Hello” messages. The costinherent to this neighborhood discovery approach is notspecific to our technique, since all solutions which allownode interactions must pay neighbor discovery costs.Nodes that share a same cell address are thus aware ofone another.

4.2.2.3. Identifier assignments. Within each cell, the nodewith the lower serial number becomes the AC. The AC isresponsible to maintain the global addressing states inthe cell: it assigns an identifier for each sensor node thatbelongs to its cell. This is performed with the simple broad-cast of a control message that specifies the cell’s addressand the list of both <serial number, assigned ID>. By doingso, each node that belongs to the cell obtains an identifier,rendering the node’s complete address as:

actor, and (C) nodes spanned in the attractor cells.

2846 T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854

/ : /id; 0 6 /id < ncell ð9Þwhere ncell denotes the number of nodes in the cell of ad-dress / and /id the node’s identifier in that cell. Note thatthe AC assigns itself /id ¼ 0. Periodic refresh messages is-sued by the AC trigger updates of address pools by remov-ing the IDs of unavailable nodes. The pseudo-code of theIFS algorithm appears below.

IFS Algorithm

1. Inputs:

2. A0 ¼ fðx; yÞ; ðxþ d; yÞ; ðxþ d; y þ dÞ; ðx; y þ dÞg:

Deployment corner points;

3. p0 ¼ ðx0; y0Þ: node coordinates. 4. korder: order of mapping;M: number of affine

transformations

5. Output:/cell

6. Begin

7. For i ¼ 0 to korder do

a. For u ¼ bf 0 to M doi. For each point p 2 Ai�1 do Ai ¼ Ai [ fcuðpÞg; End

For pii. /i ¼ /i þ u; Sq ¼ Sq [ fðAi;/iÞg

b. End For u

8. End For i

9. For each Square s 2 Sq do a. If p0 2 Sq then return /q; 10.End For s 11. End

The above described address allocation guarantees un-ique addresses since the discretization of the network areaallows the ACs to act in a disjoint manner. ACs control theidentifiers in their respective cell by maintaining the cell’sglobal state, ensuring that a node address is not assignedtwice in the network. The following section provides thedesign rationale that motivated using an iterated functionsystem to allocate node addresses.

4.3. Design rationale

Three keys features motivated the use of the IFS discret-ization fractal theory to allocate node addresses in WSANs.

(a) Location information: The feasibility of determininglocation by means of a localization algorithm isenhanced by the presence of actor nodes in the net-work. Indeed, actors are resource rich nodes and caneasily be involved in localization algorithm such asthose proposed for ad-hoc networks in [22,23].However, the delivery of detected events can beaffected significantly by the nodes’ imprecise loca-tion information. The usage of IFS offers the possibil-ity of using approximate or relative locationinformation to determine nodes’ cells. Accordingly,it is more robust to errors and imprecisions whenmeasuring and estimating locations than schemesthat rely on exact location information. Indeed, withimprecise location information, when location errors

are unimportant, the cell computed with an IF canbe the same as the one computed with the exactlocation information. This fact becomes salient uponviewing simulation results, shown in Section 5.

(b) Allocation costs: Each node can determine its cellwithout interacting with its neighbors. The cell iden-tifier assignment is performed with a single broad-cast advertisement, thus reducing allocation costsin terms of signaling overhead. The costs of 0ðmcellÞis lower than those of DHCP-Like or hierarchicalallocations, as shown with a comparative table pro-vided in Section 5.

(c) Routing: One of the targeted objectives consists ofensuring that the sensor nodes play the same rolewhen routing data to an actor or the sink, to avoidbottlenecks or a single point of failure. The shortageof hierarchical structure is that the task of the clus-ter head is heavy as it is responsible to manageand coordinate within the cluster nodes and maycreate network bottlenecks. For that purpose, theusage of IFS provides each node with the ability ofsending data to any destination throughout itsneighborhood, without referring to a dedicatednode. This is accomplished by using the locationinformation embedded in the cell’s address.

Now that features have been made explicit, the follow-ing section details how the allocated addresses serve therouting manager purposes through the fractal cell routingprotocol.

4.4. Routing

The proposed addressing system takes into account thetwo kinds of WSANs architectures described in [1]: theautomated architecture and the semi-automated architec-ture. Indeed, when a phenomenon is detected with theautomated architecture, sensors transmit their readingsto the actor nodes which then process all incoming dataand initiate appropriate actions. A semi-automated archi-tecture is used when data are routed back to the sink,which then issue action commands to actors. Such kindof communication architecture between sensors and actorsis well-described in [3,4], where sensor-actor communica-tion is performed through one or multiple hops. For thesetwo kinds of architecture, this section presents the fractalcell routing (FCR) strategy that uses the allocated addressesto carry data packets from a source to its destination.

4.4.1. Fractal cell routingWhen a node must send data to a specific destination, it

employs cell position information that is embedded in thecell’s address / to compute the inter-cell distance betweensource and destination. Let c0c1c2c3c4c5::c denote the ad-dress assigned to a given node. Its cell coordinates are gi-ven by ððc0c2c4Þ20 ; ðc1c3c5Þ2Þ. As such, from the addressesof any pair of cells u and v the inter-cell distance duv isequivalent to the Euclidian distance between their coordi-nates. From the neighborhood discovery process, eachnode acquires its neighbors’ cell addresses and can com-pute the inter-cell distance between themselves and the

T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854 2847

destination. Note that this neighborhood, called routeLet, iscomposed of nodes in adjacent cells as well as nodes with-in the same cell. Upon receiving a packet to be forwarded,the node selects, among the forwarding candidates set, thenodes in its routeLet with the minimum inter-cell distancewith the destination. Then, from that set, a node is selectedrandomly as forwarder. A node can have neighbors only inits own cell or have the same inter-cell distance with aneighbor in an adjacent cell. To break the tie, in a first in-stance, a node is selected randomly and, the second timearound, a node is selected in the most densely populatedcell in the routeLet. Fig. 4 illustrates the FCR strategy be-tween source and destination. Indeed, the source node Swith address /src

cell :: /srcid ¼ 0000 :: 1 must send a packet to

the destination node (an actor for example) whose addressis /dest

cell :: /destid ¼ 1111 : 2. Their respective cells’ coordinates

are ð00;00Þ and ð11;11Þ, equivalent in decimal base to(1,3) and (3,3). A possible path is composed of nodes se-lected in the cells, as follows: 0000 :: 1! ð0;0Þ !ð1; 1Þ ! ð1; 2Þ ! ð2; 2Þ ! ð3; 3Þ ! 1111 : 2.

4.4.2. Sensor-actor communicationActor nodes move freely from one point to another in

order to perform a given task, following a path on whichmany acting points are found. At each acting point, it col-lects data from sensors and ensures coordination withother actors according to such data. Actor coordinationand task assignment mechanisms are beyond the scopeof this paper. However, a coordination mechanism can beincorporated as an AON functionality to select the best sui-ted actor for the actuation task. Every time an actor movesto a new area, it requests a new address from the neighbor-hood ACs (if the movement is directed towards a new cell).It selects the address allocated by the closest AC and an-nounces itself by specifying its new address and interestsin a broadcast announcement beacon. The announcementis finally described by the tuple < /actor : /id;

½t1; t2; . . . ; tr � >. Sensor nodes that belong to a covered areacreate and configure a network interface according to thetopics to which advertising actors have subscribed to. Thisconfiguration includes the actor’s address and interests.When an event occurs, the node uses FCR to deliver thesensed data to the closest actor, even if the detected eventfails to match any of the actor’s interest subscription.Sometimes, when the source node belongs to an overlap-

Fig. 4. Fractal cell ro

ping area, the node reports data according to the actors’interests subscription.

4.4.3. Actor–actor communicationWhen an actor node receives data reported after an

event is detected, either of two situations occur: (1) theevent reported does not match its interest: in this case,the actor node selects a partner within its own cluster thatsubscribes to such interest; otherwise, the event is re-ported to the sink using the FCR strategy; (2) the reportedevent matches the actor’s subscription: in this case, theassociated data is disseminated by using a minimum span-ning tree (mST), rooted on the access point actor (APA). In-deed, the APA node is the first actor contacted by thesensor node that generates the event. The metric used tocompute this mST is the inter-cell distance between adja-cent actor nodes. Note that the actor holds a list of tuples< /actor : /id; ½t1; t2; . . . ; tr � > of its partner in the same clus-ter. From one actor to another in the mST, FCR is used tocarry data. Fig. 5 illustrates both communication cases.

4.4.4. Sensor–actor to sink communicationCommunication with the sink takes place in either of

these cases: (a) the event does not match any interest inthe cluster of the APA, which subsequently reports theevent to the sink using FCR; (b) for the semi-automatedarchitecture suggested in [3], sensors report event data tothe sink using FCR. The same communication pattern isused by the sink to issue commands for actors.

4.5. Handling topology changes

Topology changes in WSANs occur when nodes becomeunreachable, due to battery depletion, or when additionalnodes are deployed during network operations. Anotherimportant topology change in WSANs pertains to actors’mobility. Although sensor nodes are considered static mostof the time, actors are destined to navigate throughout thenetwork in order to fulfill their coverage mission. Both ofthese specific cases are thus considered in this study.

4.5.1. Joining nodesWhen a node, or a group of nodes, becomes unavailable

in a given cell, the AC updates the address pool accord-ingly. A newly arrived node in a given cell triggers athree-step process to get an address: (1) it waits for an

uting strategy.

Fig. 5. Sensor–actor communication.

2 The Zipf distribution is a well-known high-skewed distribution. Ifh P 0:9, it is very highly skewed; when h ¼ 0 the distribution is uniform.

2848 T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854

AC notification message, (2) upon receiving the AC notifi-cation message, the node issues an address request to theclosest AC in terms of radio signal strength indicator (RSSI),and (3) the AC responds by allocating its cell address andan identifier. A newly arrived node senses the AC notifica-tion before sending a request, since blind solicitation oftenleads to multiple responses, which are energy inefficient.In addition, a moving actor that obtains an address imme-diately notifies its cluster partners so that they can updatethe relevant profiles.

4.5.2. Departing nodesThe addresses assigned to nodes that become unreach-

able, due to a migration or eventual deaths, are recoveredby the AC when its notification fails to generate a responsefrom these nodes. AC notifications take place periodically.Two departure scenarios are considered: departures of reg-ular nodes or ACs. A regular node departure is detectedonce the AC notification does not generate a node acknowl-edgement. Indeed, failures to respond cause departuredetection, and the ID of the departed node is set in the poolfor future reuse. The AC departure is detected when its cellnodes fail to receive AC notifications. Once an AC departureis detected, a re-configuration phase is reinitiated amongthe nodes that share a common cell address. The pool ofthe newly elected AC is updated according to the numberof nodes that share the same cell address.

5. Performance evaluation

This section reports on the performance evaluation ofSubCast through extensive simulations using the Qualnetsimulator [27]. Prior to presenting simulation results andassessing the accuracy of SubCast, experimental parame-ters are listed in Table 1. Note that Melodia et al. [3] usedthe same parameter settings for their communicationarchitecture in WSANs. When conducting these experi-ments, key metrics depicting the behavior of the proposedscheme are measured. The results presented reflect anaverage of 15 simulations over different random topologieswith a confidence interval of 95%.

To evaluate the clustering efficiency of the AON, itsbuilding cost is measured using the KM algorithm and

compared to the proposed Tabu–TCP algorithm. KM andthe Tabu–TCP algorithms are both implemented on thesink. The termination criterion for the Tabu–TCP is set atD ¼ 50 iterations. The workload is constructed as follows:given the set of topics T, each topic t 2 T is associated withprobability qt ;

Pt2T qt ¼ 1, so that each node subscribes to

with a probability qt . The value of qt is distributed accord-ing to a Zipf distribution2 or a uniform distribution. The Zipffactor is set at h ¼ 0:5 according to studies provided in [28],based on the subject popularity. Each actor subscribes tons ¼ 8 of 20 available topics in T.

Fig. 6a depicts the AON clustering cost for various num-bers of clusters, using the Tabu–TCP and KM algorithms.The number of clusters varies from 2 to 48 and a very largeAON of na ¼ 50 actors is considered. Tabu–UNI and KM–UNI denote that actors subscribe to their interest in a ran-dom and uniform distribution, while Tabu–ZIPF and KM–ZIPF denote a Zipfian topic subscription. As illustrated,regardless of the subscription access, the TS algorithmachieves a better clustering cost than KM. This is due tothe fact that KM algorithms terminate at local minimumand, therefore, possibly not providing optimal solutionsto clustering problems. TS outperforms KM since it hasthe ability to provide means to escape from local optimalsolutions by exploring the solution space beyond localoptimality and attempts to find global optimal solutions.More particularly, at the maximal clustering cost, the opti-mal number of clusters that composes the AON is reached.For the case on hand (50 actors), the maximum cost isreached for nc ¼ 7 for Tabu–ZIPF,nc ¼ 9 for KM-ZIPF,nc ¼ 8for Tabu–UNI and nc ¼ 6 or 9 for KM–UNI. The Zipfian ac-cess costs are more appealing as actors share a greaternumber of common interests compared to the uniform ac-cess. When a tie occurs, for the same values of maximumclustering cost, the optimal number of clusters is selectedas the one with the maximum relative entropy. That isthe case for KM–UNI where an identical maximum cost isobtained for nc ¼ 6 and nc ¼ 9. The value of nc ¼ 6 shouldthus be selected as its relative entropy is greater (0.94),as illustrated in Fig. 6b. Indeed, this figure shows the

Table 1Simulation parameters.

Parameter Value

Network area 200 m � 200 mSensors 200Transmission range 40 mbandwidth 250 kbpsPacket size 128 bytesSimulation time 300 sActors variableNumber of topics 20Number of subscriptions 8Reporting rate 1 packet/s

T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854 2849

relative entropy variations as the number of clusters in-creases. This value increases with the number of clustersas clusters sizes are similar.

Using the above described methodology (clusteringcosts and relative entropy) to determine the optimal num-ber of clusters in the AON, the number of topics was variedin order to measure its influence on the communicationcosts in the AON. For that purpose, events are randomlygenerated from sensor nodes according to uniform distri-bution and Zipfian distribution over the number of topics.Note that for all experiments, the reporting rate equals1 packet/s, from 100 sensor nodes randomly selected inthe network. Fig. 7 depicts the comparative results ofAON architecture with the simple clustering architecture(SC) in terms of communication costs. In the SC architec-ture (similar to the automated architecture described in[1]), sensors report events to the nearest actor, and whenthe reported event fails to match the actor’s subscriptions,the actor forwards the events to the sink, which has theability to handle such situations. Fig. 7 shows that commu-nication costs and number of topics increase proportion-ally. This is explained by the increasing number of topicsthat the closest actor cannot handle in the SC case. In theAON case, growth is explained by the increasing numberof events that fail to catch interest in clusters, before being

Fig. 6. Topic cluster

forwarded to the sink. For a larger number of topics, com-munication costs increase as actors share too few commoninterests, leading to a large number of clusters, causing for-warded reported events, as with the SC architecture. Con-versely, for the zipfian interest accesses, communicationcosts are always better in the AON architecture due tothe large number of common interests shared by actors.

Fig. 8 illustrates the load distribution on sink and actors.It clearly shows that the sink load and the number of topicsincrease proportionally, regardless of the distribution ofthe actors’ subscriptions. The event load on actor nodes ismore significant when they share more common interests,as is the case in the zipfian distribution, especially with theproposed AON architecture. This highlights the ability ofactors to handle most events occurring in the network, aswell as the benefit of using the AON architecture.

To evaluate the fractal cell routing strategy (FCR) de-signed for SubCast, comparisons are made with the exist-ing ZigBee routing protocols [13] for tree and meshnetworks in terms of packet delivery ratio (PDR) andend-to-end delay. Fig. 9 depicts the packet delivery ratioand the end-to-end delay with an increasing number ofreporting nodes randomly selected in the network. As ex-pected, it is clear that the FCR strategy outperforms theZigBee routing protocols, since FCR is based on sensorscell’s coordinates and its forwarding strategy is quite sim-ilar to a classical geographic forwarding routing. More par-ticularly, prior to forwarding data packet with the FCRstrategy, the node does not refer to any cluster head as isthe case with the ZigBee routing protocols. With a growingnumber of reporting nodes, the PDR decreases abruptlywith the ZigBee routing protocols, which is not the casewith FCR as paths used to send packets are more disjointscompared to those in ZigBee routing protocols, which al-most cross one another. Given such characteristics, it isobvious that end-to-end delays follow the same trend.

Since the SubCast architecture is based on the assump-tion that nodes are aware of their own location through alocalization algorithm, the impact of localization errors

ing accuracy.

Fig. 8. Events load distribution (a) AON-UNI, (b) SC-UNI, (c) AON-ZIPF, and (d) SC-ZIPF.

Fig. 7. Communication costs.

2850 T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854

on routing performance is evaluated. For that purpose,comparisons are made with the geographic forwardingstrategy (GF) [29] which forward data based on the knowl-edge of nodes coordinates. To highlight the pertinence of

the results, a sink is deployed at the upper right corner ofthe area, no actors are deployed in this scenario and all de-tected events are reported to the sink. Fig. 10 shows theimpact of the location error on the PDR and the end-to-

Fig. 9. FCR versus ZigBee routing protocols.

Fig. 10. Impacts of localization errors on the routing strategy.

Fig. 11. Communication costs for different event delivery infrastructurewith moving actors.

T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854 2851

end delay. Indeed, the location error is generated from aRayleigh distribution, by generating a pair ðDxi;DyiÞ whereboth Dxi and Dyi are selected from a zero-mean Gaussiandistribution Nð0;r2

locÞ. With the average locationl ¼ rloc

ffiffiffiffiffiffiffiffiffip=2

perror, one can generate a location error dis-

tribution by fixing rloc. When l varies from 0% up to 30%,the FCR routing strategy is less sensitive to location errorscompared to the GF strategy, which uses node coordinatesto deliver data to the sink. When location errors increase,the PDR with GF decays significantly compared to theFCR strategy. This leads to higher end-to-end delay as de-picted in Fig. 10b. The delay increases as the location errorinduces the forwarding strategy to choose the wrong nodesand then, increases the number of hops to be crossed in or-der to reach the sink node. The performance of FCR over GFcan be explained by the fact that the same cells are alwaysselected as forwarding cells to deliver data to the sink, anddespite the location error, almost nodes belong to the samecell.

To evaluate the impact of actor node mobility, variousnumbers of actors are deployed, navigating in the network

Table 2Qualitative analysis of address allocation schemes.

Properties SubCast TreeCast [12] ZigBee [13] DHCP-Like[8–10]

Uniqueness Yes Yes Yes YesAllocation overhead OðmcellÞ OðmnÞ OðmnÞ OðmnÞAllocation mode Distributed Centralized/distributed Centralized/distributed CentralizedAddress length [bits] log2ðnÞ P log2ðnÞ 16 log2ðnÞRouting Flat Tree Tree and Mesh Flat

2852 T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854

following a random waypoint mobility model. Their speedis set in the range ½0;2 m=s� with a 30-s break. As previ-ously, AON communication costs are measured and com-pared to:

(a) The simple clustering (SC), described above, inwhich sensors are clustered with their closest actor.Events reported to the closest actor with mis-matched subscription are delivered to the sink nodeusing FCR;

(b) The simple infrastructure less clustering architec-ture (SIC), in which the event reported to the closestactors with a mismatched subscription is broad-casted among the actor nodes in order to find theone that does match the reported event. Note thatactors do not broadcast events when subscriptionsmatch.

Fig. 11 compares results pertaining to the communica-tion costs of such configurations. Obviously, the SICscheme is only efficient for a small number of deployed ac-tors while it turns out to be more costly with a greaternumber of actors. Communication costs remain constantwith the SC scheme while they decay with the AON infra-structure, which is due to the following two reasons: (1)the clustering increases the likelihood of find matching ac-tors, and (2) actors’ mobility favors the proximity of sensornodes and actors belonging to clusters with matchingevents. Moreover, with a growing number of actors, addi-tional clusters help keep dissemination costs low com-pared to SC and SIC. This would not be the case if thenumber of clusters was fixed, as is the case for the KMalgorithm.

Table 2 provides a qualitative comparison of the pro-posed addressing system with some related works reportedin the literature. In a cluster-based addressing scheme, theaddress length is expressed as Ccb ¼ log2ðnaÞ þ log2

nna

� ,

assuming that clusters headed by actors have an identicalnumber of sensors. The closed form of the address lengthis then Ccb ¼ log2ðnÞ, which is equivalent to the addresslength in the network-wide addressing scheme [11], wherethe address length also equals log2ðnÞ. The address length inSubCast system is expressed as CSubCast ¼ 2korder

þlog2n

22korder

� and its closed form is CSubCast ¼ log2ðnÞ. Thus,

with the simulation parameters, the address length in Sub-Cast equals 8 bits. Assume that there are types of messageto allocate addresses to the network nodes. To completethe allocation process, messages are needed in SubCastand m � n for the other cases.

6. Conclusion

This paper presents an addressing and routing systemfor large scale WSANs based on topic clustering amongactors, and fractal area discretization to allocate ad-dresses. To this end, actors are clustered according totheir interest. The overall clustering system forms the ac-tors overlay network that uses a multicast tree commu-nication pattern to deliver reported event to interestedactors. Such a process of assigning node addresses inthe networks depends on the nodes’ physical location,which may be prone to localization errors. The fractaltheory iterated function system appears to deal effi-ciently with such location errors and provides a low-costdistributed allocation that guarantees that unique ad-dresses are assigned to network nodes. The proposedarchitecture was tested with simulations and the resultswere compared with certain traditional schemes found inthe literature. Results demonstrate SubCast robustnessand, in fact, contribute to making it an attractiveaddressing system for WSANs. Future research avenuesinclude plans to grant the task allocation process to actornodes according to the reported events, a constituent ofthe actors’ overlay networks.

Appendix A. Notations used in this paper

Notations Description

n

number of nodes in the network na number of actor nodes D number of iterations for termination criterion r transmission range korder number of recursions for IFS cell computing ncell number of nodes in a cell nc number of clusters ns number of actors’ subscriptions Ucell cell’s address Uid node identifier in a cell L network area side lcell cell side C addressing alphabet SA actors set nt number of topics T set of topics HrðZÞ relative entropy of the clustering vector Z h Zipf factor mcell number of cells M number of affine transformations u transformation indexes

T. Houngbadji, S. Pierre / Computer Networks 53 (2009) 2840–2854 2853

Appendix B. Frame used for all exchanges

Sourceaddress

Nexthopaddress

Destinationaddress

Packettype

Sequencenumber

Payload

Source Address: network address of the source node;empty when the source identification does not matter.Next hop Address: network address of the next hopnode; empty for Advertisements;Destination Address: network address of the destina-tion node; empty for Actor/Sink Advertisements.Packet type: denotes the type of packet.‘‘0”: Subscription Packet;‘‘1”: Data Packet;”2”: Actor Advertisement Packet;‘‘3”: Sink Advertisement Packet.Sequence Number: a number that uniquely identifies anotification.Payload: the useful data.

References

[1] I.F. Akyildiz, I. Kasimoglu, Wireless sensor and actor networks:research challenges, Elsevier Journal of Ad Hoc Networks 2 (4)(2004) 351–367.

[2] T. Melodia, D. Pompili, V.C. Gungor, I.F. Akyildiz, Communication andcoordination in wireless sensor and actor networks, IEEETransactions on Mobile Computing 6 (10) (2007) 1116–1129.October.

[3] T. Melodia, D. Pompili, I.F. Akyildiz, A communication architecturefor mobile wireless sensor and actor networks, in: Proceedings of theThrid Annual IEEE Communications Society on Sensor and Ad HocCommunications and Networks (SECON’06), September, vol. 1, 2006,pp. 109–118.

[4] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, F. Silva,Directed diffusion for wireless sensor networking, IEEE/ACMTransactions on Networking 11 (1) (2003) 2–16.

[5] S. Madden, M.J. Franklin, J.M. Hellerstein, W. Hong, TAG: a tinyaggregation service for ad-hoc sensor networks, SIGOPS OperatingSystem Review 36 (2002) 131–146. December.

[6] H.-O. Peitgen, H. Jurgens, D. Saupe, Chaos and Fractals New Frontiersof Science, second ed., Springer, New York, 2004.

[7] R. Droms, RFC 1541: Dynamic Host Configuration Protocol.[8] R.C. Doss, D. Chandra, L. Pan, Lease based addressing for event-driven

wireless sensor networks, in: Proceedings of the 11th IEEESymposium on Computers and Communications, ISCC’06, 26–29June, 2006, pp. 251–256.

[9] R. Teng, H. Morikawa, T. Aoyama, Network-layer and MAC-layeraddress autoconfiguration in self-organized sensor networks, in:Proceedings of the Sixth International Conference on ITSTelecommunications, June, 2006, pp. 1005–1010.

[10] Y. Zheng, D. Falko, Dynamic address allocation for management andcontrol in wireless sensor networks, HICSS, in: Proceedings of the40th Annual Hawaii International Conference on System Sciences,2007, p. 292b.

[11] W. Fang, Y. Liu, J. Wu, D. Qian, An on-demand address allocationscheme for query based sensor networks, in: Proceedings of theEighth International Symposium on Autonomous DecentralizedSystems, ISADS’07, 2007, pp. 173–179.

[12] S. PalChaudhuri, S. Du, A.K. Saha, D.B. Johnson, TreeCast: a statelessaddressing and routing architecture for sensor networks, in:Proceedings of the 18th International on Parallel and DistributedProcessing Symposium, April, 2004, p. 221.

[13] Zigbee Alliance, ZigBee Specification: ZigBee Document 053474r06Version 1.0, 14 December, 2004.

[14] M.-S. Pan, H.-W. Fang, Y.-C. Liu, Y.-C. Tseng, Address assignment androuting schemes for ZigBee-based long-thin wireless sensornetworks, in: IEEE Vehicular Technology Conference, VTC Spring,11–14 May, 2008, pp. 173–177.

[15] L.-H. Yen; W.-T. Tsai, Flexible address configurations for tree-based ZigBee/IEEE 802.15.4 wireless networks, in: Proceedings ofthe 22nd International Conference on Advanced InformationNetworking and Applications, AINA, 25–28 March, 2008, pp.395–402.

[16] D. Zhigao, Q. Depei, S. Stanczak, R. Heras, L. Yi, Auto-configuration ofshared network-layer address in cluster-based wireless sensornetwork, in: IEEE International Conference on Networking, Sensingand Control, ICNSC, 6–8 April, 2008, pp. 148–153.

[17] R. Ashwini, S. Patil, V. Muthukumar, A. Abraham, HAUNT-24: 24-bithierarchical, application-confined unique naming technique, in:Proceedings of the Fifth International Conference on IntelligentSystems Design and Applications, ISDA’05, 8–10 September, 2005,pp. 375–380.

[18] L. Jia-Liang, F. Valois; D. Barthel, M. Dohler, Low-energy addressallocation scheme for wireless sensor networks, in: Proceedings of theIEEE 18th International Symposium on Personal, Indoor and MobileRadio Communications, PIMRC, 3–7 September, 2007, pp. 1–5.

[19] P. Th. Eugster, P.A. Felber, R. Guerraoui, A.-M. Kermarrec, The manyfaces of publish/subscribe, ACM Computing Surveys (CSUR) 35 (2)(2003) 114–131. June.

[20] R. Vedantham, Z. Zhuang, R. Sivakumar, Mutual exclusion inwireless sensor and actor networks, in: Proceedings of the ThirdAnnual IEEE Communications Society on Sensor and Ad HocCommunications and Networks, SECON’06, vol. 1, 28 September,2006, pp. 346–355.

[21] R. Vedantham, Z. Zhuang, R. Sivakumar, Hazard avoidance inwireless sensor and actor networks, in: Proceedings of the SecondInternational Conference on Broadband Networks, BroadNets, vol. 1,3–7 October, 2005, pp. 690–702.

[22] N. Bulusu, J. Heidemann, D. Estrin, GPS-less low cost outdoorlocalization for very small devices, in: IEEE PersonalCommunications, Special Issue on Smart Spaces and Environments,October, 2000, pp. 28–34.

[23] N. Bulusu, J. Heidemann, D. Estrin, Adaptive beacon placement, in:Proceedings of the Twenty First International Conference onDistributed Computing Systems, ICDCS-21, April, 2001.

[24] O. Kariv, L. Hakimi, An algorithmic approach to network locationproblems Part II: The p-medians, SIAM Journal of AppliedMathematics 37 (3) (1979) 539–560.

[25] T. Kanungo, D.M. Mount, N.S. Netanyahu, C.D. Piatko, R. Silverman, A.Y Wu, An efficient k-means clustering algorithm: analysis andimplementation, IEEE Transaction of Pattern Analysis and MachineIntelligence 24 (7) (2002) 881–892. July.

[26] F. Glover, M. Laguna, Tabu Search, Kluwer, Norwell, MA, 1997.[27] Qualnet Simulator. <http://www.qualnet.com>.[28] H. Liu, V. Ramasubramanian, E.G. Sirer, Client behavior and feed

characteristics of RSS, a publish-subscribe system for webmicronews, in: Proceedings of the Internet MeasurementConference, Berkeley, CA, USENIX Association, Berkeley, 2005, p. 3.

[29] B. Karp, H.T. Kung, GPSR: greedy perimeter stateless routing forwireless networks, in: Proceedings of the Sixth Annual internationalConference on Mobile Computing and Networking, MobiCom’00,06–11 August, 2000, 243–254.

[30] J. Vesanto, E. Alhoniemi, Clustering of the self-organizing map,IEEE Transactions on Neural Networks 11 (3) (2000) 586–600.May.

Therence Houngbadji received the EngineerDegree in Telecommunication Systems fromSwiss University of Applied Sciences in 2003(Switzerland) and M.Sc. Degree in computersciences from École Polytechnique de Mont-réal, Québec (Canada) in 2006. He is currentlya Ph.D. candidate at the Mobile Computingand Networking Research Laboratory. Hisresearch interest includes wireless networksand mobile computing systems.

ter Networks 53 (2009) 2840–2854

Samuel Pierre is currently a professor ofcomputer engineering at École Polytechnique

de Montréal, where he is the director of theMobile Computing and Networking ResearchLaboratory (LARIM) and the NSERC/EricssonIndustrial Research Chair in Next-generationMobile Networking Systems. He is the authoror coauthor of six books, 15 book chapters,and 16 edited books, as well as more than 300other technical publications, including journaland proceedings papers. He received the BestPaper Award of the Ninth International

Workshop in Expert Systems and their Applications (France, 1989), the

2854 T. Houngbadji, S. Pierre / Compu

Distinguished Paper Award from OPNETWORK 2003 (Washington, USA),

and special mention from Telecoms Magazine (France, 1994) for one ofhis coauthored books, Telecommunications et Transmission de Données(Eyrolles, 1992), among others. His research interests include wirelineand wireless networks, mobile computing, performance evaluation, andelectronic learning. He is an associate editor of the IEEE CommunicationsLetters, the IEEE Canadian Journal of Electrical and Computer Engineering,and the IEEE Canadian Review. He is also a regional editor of the Journal ofComputer Science and he serves on the editorial board of Telematics andInformatics published by Elsevier Science. He is a fellow of the Engi-neering Institute of Canada, a senior member of the IEEE, and a member ofthe ACM and the IEEE Communications Society.