Computer Communications 33 (2010) 1334–1342

Contents lists available at ScienceDirect

Computer Communications

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

QoSNET: An integrated QoS network for routing protocols in large scalewireless sensor networks

Therence Houngbadji *, Samuel PierreDepartment 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 a b s t r a c t

Article history:Received 6 October 2009Received in revised form 12 March 2010Accepted 15 March 2010Available online 20 March 2010

Keywords:Sensor networksIterated function systemRoutingQuality of servicePercolation

0140-3664/$ - see front matter � 2010 Elsevier B.V. Adoi:10.1016/j.comcom.2010.03.017

* Corresponding author. Tel.: +1 514 340 4711x21273240.

E-mail addresses: therence.houngbadji@polymtl.cPierre@polymtl.ca (S. Pierre).

Numerous QoS routing strategies focus on end-to-end delays to provide time constrained routing proto-cols in wireless sensor networks (WSNs). With the arrival of wireless multimedia sensor networks, trafficcan be composed of time sensitive packets and reliability-demanding packets. In such situations, someworks also take into account link reliability to provide probabilistic QoS. The trade-off between the guar-antee of the QoS requirements and the network lifetime remains an open issue, especially in large scaleWSNs. This paper proposes a promising multipath QoS routing protocol based on a separation of thenodes into two sub-networks: the first part includes specific nodes that are occasionally involved in rout-ing decisions, while the remaining nodes in the second sub-network fully take part in them. The QoSrouting is formulated as an optimization problem that aims to extend the network lifetime subject toQoS constraints. Using the percolation theory, a routing algorithm is designed to solve the problem onthe respective sub-networks. Simulation results show the efficiency of this novel approach in terms ofaverage end-to-end delays, on-time packet delivery ratio, and network lifetime.

� 2010 Elsevier B.V. All rights reserved.

1. Introduction

Wireless sensor networks (WSNs) consist of an emergent tech-nology deployed for a large range of solutions, spanning military,civilian, environmental and commercial applications. They consistof a large number of low-cost sensing devices equipped with wire-less communication and computation capabilities. Given the re-cent advances in WSNs, it is expected that video sensors will besupported in such networks, for applications such as battlefieldintelligence, security monitoring, emergency response, and envi-ronmental tracking. For example, multimedia sensors may monitorthe flow of vehicular traffic on highways and retrieve aggregateinformation such as average speed and number of cars. Given thephysically small nature of sensors, and since multimedia applica-tions typically produce a huge volume of data that require hightransmission rates and extensive processing, power consumptionbecomes a fundamental concern in WSNs. More particularly, trans-mitting a video stream using the shortest path will drain the nodeof its energy along the path, thereby shortening the network life-time, for example. Most proposals for routing in WSNs are basedon a flat, homogenous architecture in which every sensor has iden-tical physical capabilities and can interact only with its neighbor-

ll rights reserved.

/3240x4685; fax: +1 514 340

a (T. Houngbadji), Samuel.

ing sensors. However, flat topologies are not always best suitedto handle the amount of traffic generated by multimedia applica-tions, including audio and video [1]. Therefore, developing newrouting strategies to maximize the network lifetime, while satisfy-ing the QoS requirements, represent a critical problem to beaddressed.

Some works attempt to address the issue through single pathrouting strategies which only guarantee delay constraints. Com-pared to multipath routing, single path routing algorithms in WSNsare simpler and they consume less energy. Moreover, aside frommeeting delay constraints, multipath routing is also reliable. Reli-ability can be characterized by way of a packet delivery ratio,which is defined as the ratio between the number of unique pack-ets successfully received by the sink over the number of packetsgenerated by source nodes. In WSNs, multipath routing is used toestablish multiple paths between all source–sink pairs in order toincrease the likelihood of reliable data delivery. Multiple copiesof data along different paths [2] are sent to reduce data deliverydelays by sharing transmission delays among the different pathsavailable from the source to the destination.

Thus, end-to-end delays, reliability requirements as well as theresidual energy of nodes in WSNs are considered to build a newmultipath routing strategy over a discretized network area. Thisperspective considers how underlying communication networkscan achieve hard QoS constraints while efficiently utilizing net-work resources. This work complements our previous work [3]on addressing systems in large scale WSANs. It extends routing

T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342 1335

capabilities by providing QoS for large scale WSNs that possibly in-clude powerful devices such as actors or mobile robots. In a firststep, the network area is discretized into cells using the fractal the-ory iterated function system (IFS) [4]. In each cell, a cell controller(CC) is elected to control the signaling operation (address alloca-tion, statistics collection, etc.). These CCs form the cell controllernetwork (CCN) which is also expected to integrate powerful de-vices such as actors and mobile robots. The remaining nodes com-pose the simple sensor network (SSN) which maintains theprimary data routing role of sensor nodes. The hard QoS con-straints are fulfilled by using a switching routing mechanism be-tween the two sub-networks in order to determine a forwardingset. The latter consists of the selected nodes where the incomingpacket will be forwarded. This switching mechanism is accom-plished by applying a percolation theory [5] to the discretized net-work area. The remainder of this paper is organized as follows:Section 2 discusses related works and Section 3 provides an in-depthlook at the proposed routing strategy. Performance evaluations aredetailed in Section 4 followed by a conclusion in Section 5.

2. Related work

Routing techniques are divided into three categories based ontheir underlying network structures: flat, hierarchical, and loca-tion-based [6]. Furthermore, these protocols can be classified intomultipath-based, query-based, negotiation-based, QoS-based, andcoherent-based. This section presents an overview of the existingQoS-based routing protocols used in WSNs. QoS-based protocolsallow sensor nodes to balance energy consumption and certainpredetermined QoS metrics before they deliver data to the sinknode. One of the first routing protocols for WSNs, sequentialassignment routing (SAR) [7], introduces the notion of QoS in rout-ing decisions according to three factors: energy resources, QoS oneach path, and the priority level of each packet. In order to avoidsingle route failures, a multipath approach is coupled with local-ized path restoration. SAR maintains multiple paths from nodesto a base station. Although, this ensures fault-tolerance and easyrecovery, the protocol suffers from high overhead to maintain ta-bles and states for all sensor nodes, especially in the presence ofa large number of nodes. SPEED [8], another QoS routing protocolfor WSNs, provides soft real-time end-to-end guarantees, yet it alsomaintains a desired delivery speed across sensor networks with atwo-tier adaptation method included to divert traffic at the net-working layer, and locally regulating packets sent to the MAC layer.MMSPEED [9], an extension of SPEED, supports service differentia-tion and probabilistic QoS guarantee. For delivery timeliness, mul-tiple network-wide packet delivery speed options are provided fordifferent types of traffic according to their end-to-end deadlines. Asis the case with SPEED, since all MMSPEED mechanisms work lo-cally without global network state information and end-to-endpath setup, they become scalable and adaptive to network dynam-ics. Both SPEED and MMSPEED share a common deficiency: theyfail to consider energy consumption metrics. Several delay-awarerouting techniques have been proposed [10–13] to improve delays.A routing algorithm proposed by [14] uses the global knowledge ofnode queue sizes by the gateway in order to calculate end-to-enddelays, lowest-cost paths and to generate routing tables. A heuris-tic solution to find energy-efficient paths for delay-constraineddata in WSNs has been designed in [15]. They employ topologycontrol and a model of contention delays caused by the MAC layer.Paths between source and sink nodes are identified and indexedaccording to the amount of energy they consume. End-to-end de-lays are estimated along each of the ordered paths and the onewith the lowest index that satisfies the delay constraint is selected.Multi-constrained multipath routing is proposed in [16,17], where

a node to sink packet delivery is achieved based on the so-calledSoft-QoS constraints expressed in terms of reliability and delays.This model addresses the issue of multi-constrained QoS in WSNsby considering the unpredictability of network topology andattempting to minimize energy consumption. However, with theseapproaches, packets may get routed to a node that is highly con-gested and/or whose energy is critical. Simulation results con-ducted for a small network size do not guarantee the sameperformance for large scale WSNs in which the number of selectedpaths can increase and become a source of wasted energy. A real-time power-aware routing (RPAR) protocol [18] is proposed toachieve application specified communication delays at low energycost, by dynamically adjusting transmission power and routingdecisions. In [19], a protocol called ReInForM is proposed to deliverpackets with the desired reliability level by sending multiple cop-ies of each packet along multiple paths between the source and thesink. The number of paths used is dynamically determined depend-ing on the channel error probabilities. Instead of using disjointpaths, GRAB [20] uses a path interleaving technique to achieve ahigh level of reliability.

To date, all multipath routing techniques address the QoS rout-ing in terms of the end-to-end delay and reliability in small WSNs,made up of homogeneous nodes whose capabilities are almostidentical. Furthermore, it seems that there is a need to addressthe same issues in large scale WSNs. The novelty of our proposalis then to provide QoS in large scale WSNs, and to deal with theheterogeneous properties of the deployed nodes by assuming thatpowerful nodes such as actors or mobile robots can cohabitatewith resource impoverished sensor nodes. Most of the multipathrouting protocols that aim to provide QoS in WSNs do not havethe ability to support the disparity among nodes in terms of capa-bility. Our proposal fills that gap by taking into account these dif-ferences through the integrated QoS network that allows the nodesto switch the routing decision between two separated networks.Moreover, some works especially those in [16,17] which are closerto ours use probabilistic QoS routing to meet the QoS requirementswith predefined probabilities. Instead of using the same approach,our work breaks the tie by meeting hard QoS requirements.

3. The system model

WSNs with randomly deployed nodes in a two-dimensionalnetwork area are being considered. Nodes are powered by batteryand they are assumed to be aware of their own location. Indeed,nodes can find their own approximate location coordinates usingtriangulation or multilateration methods. A grid network architec-ture is considered, as it discretizes the network into two-dimen-sional micro-scale areas that favor operation distributed over thenetwork. Other forms such as triangles or hexagons could be usedto discretize the network area; yet the grid representation is se-lected for its simplicity. Current network representations usinggrid structures [21] usually lead to an underutilization of the radiocoverage areas, as pointed out by [22]. This is due to the fact thatall nodes within a grid must be able to reach all nodes in adjacentgrids. Actually, nodes are forced to cover less than half the distanceallowed by the radio range. To deal with this issue, the fractal gridrepresentation, using a square form, discretizes the network areaby using the entire radio coverage of nodes. The proposed systemmodel follows the same IFS design strategy used in [3] for WSANaddressing systems. IFSs are simple methods to construct fractalsbuilt by uniting several copies of themselves ad infinitum [4]. Eachcopy is transformed by a set of functions which stand for a finitenumber of M transformations fcj; 0 6 j 6 M � 1g.

cjðx; yÞ ¼ ðajxþ bjyþ ej; cjxþ djyþ fjÞ ð1Þ

Cell ControllerNetwork

Sink

CC

1336 T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342

where aj; bj; ej; cj;dj; fj 2 R. For a given initial form S, the small affinecopies c1ðSÞ; c2ðSÞ; . . . ; cMðSÞ are produced and lead to a new formwðSÞ ¼

SNj¼1

cjðSÞ. The mapping process starts with a set S0 and com-putes recursively Sk ¼ wðSk�1Þ;8k > 1. The sequence fSkg convergesto a final set called the IFS attractor. Node coordinates are used bythe IFS to determine the cell where the node is located. In order toapply IFS to the network area, four linear affine transformationsthat represent the mapping functions are defined as follows:

c00ðx; yÞ ¼ x2 ;

y2

� �c01ðx; yÞ ¼ x

2 ;y2þ a

2

� �c10ðx; yÞ ¼ x

2þ a2 ;

y2þ a

2

� �c11ðx; yÞ ¼ x

2þ a2 ;

y2

� �

8>>><>>>:

ð2Þ

The initial set S0 ¼ fð0;0Þ; ða;0Þ; ða; aÞ; ð0; aÞg represents the net-work area and these affine transformations represent a contractionof the area by a factor s ¼ 1=2, followed by a translation. The in-dexes of these transformations define a code space C ¼ f00;01;10;11g. Indeed, / ¼ /1/2/3 � � � denotes an element of the attractorð/i 2 CÞ. Fig. 1(c) illustrates the resulting attractor. Sub-squares rep-resent the cell and the dots show the deployed sensor nodes. Thesensor node recursively computes the cell to which it belongs byapplying the defined transformations and it then compares its coor-dinates with the obtained sub-square bounds.

Fig. 1(a) shows a one-order attractor while Fig. 1(b) provides anoverview of the same attractor at a four-order level, in sub-square‘‘00”. Note that nodes are aware of the dimensions of the deploy-ment area, as such information can be gathered by means of net-work-wide flooding. Since sensor nodes cannot use an infinitealphabet sequence of the code space C to determine their cell ad-dress, the korder value that defines the number of iteration of the IFS,as well as the length of the cell address chosen, are finite andbounded. An address /cell of any cell is a finite sequence of ele-ments of C such as:

/cell ¼ /1/2/3 � � �/korder; /i 2 C ð3Þ

To allow communication between a sensor node and other nodes inits cell, a worst-case scenario is considered, in which the distancebetween two nodes in any cell equals the cell’s diagonal so thatthe cell’s side lcell is bounded as lcell 6

rffiffi2p . Since lcell ¼ a=2korder, the

number of sequences korder in the cell address can be chosen so that:

korder P12þ log2

ar

� �ð4Þ

where r denotes the sensor node radius. Thus, the cell address con-sists of the first part of the node address. Within each cell, nodeselect a dedicated entity that represents the Cell Controller (CC), tomaintain the global states within the cell. The only function ofCCs is to be signaling points within the cell. To discriminate sensorsin the same cell, the CC assigns an ID to each of them. Consequently,the complete address of a node can be written as:

/ : /id; 0 6 /id < ncell ð5Þ

where ncell denotes the number of nodes in the cell of address/ and /id, the node’s identifier in that cell. A periodic refresh, issued

00

01 10

11

(a) (b) (c)

Fig. 1. (a) One-order attractor. (b) Four-order cells of attractors. (c) Sensor nodesspanned in the attractor cell.

by the CC, triggers an update of the address pool by removing theIDs of unavailable nodes.

3.1. The QoS network model

With the assigned address, each sensor node can participate inthe network routing operations. As indicated above, each cell isequipped with a CC that manages the cell operations. Numerousstudies address QoS routing for WSNs that aims to allow each nodeto select a path based on QoS requirements. Although this seems tobe an interesting approach, we believe that, in general, the largescale of WSNs must allow using certain nodes to build a separatenetwork used to signal operations, and allow others to performregular network operations, such as routing. The resulting sub-net-work that includes the CCs can also integrate powerful entitiessuch as the actor overlay networks design presented in [3] or anyother resource-rich nodes. Doing so should almost free each nodeto periodically perform a dedicated role, as is the case in mostWSN clustering schemes. Such network is called cell controller net-works (CCNs). The remaining nodes in the network, i.e., thosewhich are neither elected CCs nor categorized as an actor or mobilerobot, make up the second type of sub-network, which mainly per-forms routing operations.

Fig. 2 illustrates both layers of the network plan. The lower levelplan represents the entire network, while the higher level plan isexclusively composed of CCs. Our conjecture is that by performingonly network signaling operations and occasional routing deci-sions, the CCs will support the same burden as a simple sensornode whose only role consists of performing data forwarding.Obviously, given the large number of nodes deployed, it is assumedthat it is not wasteful for sensor nodes to be exclusively dedicatedto performing signaling tasks. Before delving into the details ofhow QoS requirements are guaranteed, let us explicitly definethe parameters selected as QoS metrics.

3.2. End-to-end QoS parameters

Traditional networks specify QoS in a rigid manner, using met-rics such as the average bandwidth, bounded delays and jitterrequirements. In dynamic environments such as mobile wirelessnetworks, it is difficult to provision and support rigid QoS. Worksby [16,17], which are similar to our proposal, deal with the so-called Soft-QoS to fulfill QoS requirements by selecting end-to-end delays and path reliability as metrics.

In our proposal, we still strive to achieve hard-QoS constraints.For that purpose, and in order to integrate the energy consumption

Wireless SensorNetwork

Fig. 2. Two-layer overview of the WSN.

T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342 1337

issue, the End-to-End Path Battery Cost (EEPBC) is introduced asadditional metric. In fact, our hypothesis is that, extending the life-time of individual nodes will prevent early failures, thus helping tomaintain the hard QoS constraints. As a result, the proposed rout-ing scheme, in addition to the end-to-end delay and path reliabil-ity, considers the residual energy of the forwarding nodes. Let biðtÞdenote the battery capacity of a sensor si at time t. The EEPBCBs0sdð}Þ between the source node and the destination node sd is de-

fined as:

Bs0sdð}Þ ¼

Xp2}

Xsi2p

biðtÞ ð6Þ

The rationale for choosing the residual energy in this proposal isthat the information provided by the battery is easy to acquireand it offers an accurate level of energy consumed in the network.Since battery capacity is directly incorporated into the routing pro-tocol, this metric prevents nodes from being overused, therebyincreasing their lifetime.

Data can be reported at different rates depending on the type ofapplication used or the heterogeneity of the nodes in the network.The readings generated by detecting nodes can be subject to vari-ous time deadline constraints. Let dðsi; siþ1Þ denote the delay be-tween the sensor node si and neighbor siþ1. The end-to-end delaybetween any source node so and the destination node sd over theset of paths } is given by:

Dsosdð}Þ ¼min

p2}

Xsi2p

dðsi; siþ1Þ( )

ð7Þ

Indeed, Dsosdð}Þ is the minimal achievable delay when the generated

data are routed along the set of paths between so and sd. The delaydðsi; siþ1Þ between two nodes consists of the elapsed time for suc-cessfully transmitting a packet after having received it. Thus, it in-cludes queuing time, contention time, as well as transmission,retransmission and propagation times. Since our goal aims to allowCCs to participate in occasional routing decisions, one can expectthat the delay experienced to route packets in CCNs will be lowerthan in SSNs, as they are subject to less congestion.

Reliability, often computed to assess the probability of trans-mission failures, can be characterized by a packet delivery ratio,defined as the ratio of the number of unique packets successfullyreceived at the destination over the number of packets generatedby source nodes. The multipath end-to-end reliability between asource node so and a destination node sd on a set of paths } canbe written as follows:

Rsosdð}Þ ¼ 1�

Yp2}ð1� rðpÞÞ ð8Þ

where r(p) is the reliability of a path p, equal toQ

si2p

rðsi; siþ1Þ and rðsi; siþ1Þ denotes the reliability of the link ðsi; siþ1Þ.Since reliability is multiplicative, a variation in any links on p oran increasing number of hops on p will remarkably alter end-to-end reliability. That explains why single path routing cannot alwaysmeet reliability requirements. Once QoS parameters have beenidentified, one must define how such requirements are met. The fol-lowing section defines the strategy used to put together these QoSparameters in an optimization problem.

3.3. Switching QoS routing

In heterogeneous WSNs, a wireless node may also contain dif-ferent sensors such as audio, video and scalar sensors. As the prior-ity of such heterogeneous traffic differs, considering servicedifferentiation appears to be an additional issue. There are twotypes of traffic: real-time traffic has hard time constraints, suchas delays, although it is more tolerant of packet losses; non-real-

time traffic usually exists in networks and produces thousands ofpackets which are generated synchronously or asynchronously.In order to remain generic and not delve into the details of prioritylevels assigned to the different types of traffic, a general formula-tion of the QoS routing problem is considered. The joint goal ofreducing energy consumption and end-to-end delays, while maxi-mizing path reliability, appears to be very challenging. The routingobjective thus consists of finding a set of paths }ðso; sdÞ betweensource and destination that meet the specified QoS requirementsat data source, while maximizing the EEPBC. The QoSNet routingproblem can be formulated as follows:

QoSNet: Given the tuple ðDreq;RreqÞ of QoS requirements speci-fied by the source node, that represents the end-to-end delaysand the reliability level, find the set of paths }ðso; sdÞ from thesource node so to the destination node sd that maximizes theend-to-end battery power, subject to deadline and reliabilityconstraints.

max Bsosdð}Þ ð9Þ

Subject to:

Dsosdð}Þ 6 Dreq ð9aÞ

Rsosdð}ÞP Rreq ð9bÞ

Expression (9) maximizes the residual energy of the selected nodes,while constraints (9a) and (9b) express the delay and reliabilityrequirements of the data source node, respectively. The problemof determining a QoS route that satisfies multiple constraints isknown to be NP-hard [16]. The problem definition requires exactinformation regarding nodes on their paths to the destination,which is almost impossible to obtain in WSNs. Hence, to solve thischallenge, each source node periodically collects information fromits one-hop neighborhood. One-hop link metrics are much easierto acquire. Furthermore, such a strategy is scalable to the networksize. By uniformly partitioning the QoS requirements at all down-stream hops, the overall QoS requirements can be met. A nodecan satisfy the hop requirement by selecting next hop nodes basedon the link condition. As in [16], the link delay Ld

i , the reliability Lri

requirements and the local battery costs Lbi for each node, consider-

ing the hop count hi, are expressed as follows:

Ldi ¼ ðDreq � DiÞ=hi ð10Þ

Lri ¼

ffiffiffiffiRi

hip

ð10aÞLb

i ¼X

j2FwðsiÞbjðtÞ ð10bÞ

where Di is the actual delay experienced by a packet at node si fromthe source node, and Ri indicates the portion reliability requirementassigned to the path through node si, as decided by the upstreamnode of si. Moreover, FwðsiÞ denotes the forwarding set to be deter-mined by the node. Based on the above definitions, the QoSNet rout-ing problem can be reformulated for each node as follows:

QoSNet:

max Lbi ð11Þ

Xj2NðsiÞ

xjlogð1� RijÞ 6 logð1� Lri Þ ð11aÞ

xjDij 6 Ldi ð11bÞ

Expression (11) maximizes the residual energy of the forwardingset. Constraints (11a) express the fact that the selected links inthe forwarding set must all together satisfy the local reliabilityrequirement. Constraint (11b) expresses that the delay of the se-lected links must be lower than the local delay requirement. Theforwarding set Fw # NðsiÞwhich consists of the solution to this prob-

1338 T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342

lem, is determined by the set of neighbors NssnðsiÞ in the SSN, or theset NccnðsiÞ in the CCN, so that:

NssnðsiÞ [ NccnðsiÞ ¼ NðsiÞNssnðsiÞ \ NccnðsiÞ ¼ ;

�ð12Þ

The following section describes how nodes select the set that is sub-sequently considered to solve QoSNet. A typical forwarding scenariois described in Fig. 3. Feasible paths are found in the SSN when thedetected event is forwarded from the source node S, while node A,when solving the problem, finds feasible paths only in the CCN.When the node fails to find a feasible solution either inNssnðsiÞ or NccnðsiÞ, the packet is dropped. Each forwarding nodesolves QoSNet and the packet is forwarded until it reaches its des-tination. Intermediate nodes might receive duplicate packets forthe same session. For that purpose, each node maintains a table thatrecords the tuple h/ : /id; seq;Dii of a data reporting session. Asmentioned above, / : /id identifies the reporting node and seq theassociated sequence number. This table is emptied periodicallyand the node keeps a time stamp of its last clean out in order tomaintain coherent routing timing. The rationale behind this isguided by the fact that a forwarding node which receives a packettwice, forwards the latter only when the previous experienced delayto forward this packet is greater than the actual delay. Indeed, letDprev

i denote the delay experienced by the packet previously re-ceived by node si for the same session. The node solves QoSNet todetermine Fw only when the inequality Di < Dprev

i holds. So far, wehave shown how a packet is forwarded, but how the QoSNet prob-lem is actually solved has yet to be explained. In the following sec-tion, a promising strategy is developed in order to solve this issue,by applying site percolation to the discretized area.

3.4. Mapping QoSNet resolution to site percolation

The resolution of QoSNet is limited to the node’s neighborhoodNðSiÞ, and can be achieved in polynomial time OðjNðSiÞjÞ. However,if the nodes in all cells determine their forwarding set by consider-ing both their SSN and CCN neighbors, they will spend additionalenergy resources to those of the nodes in the CCN as they are al-ready involved in the network signaling operations. To occasionallyallow nodes in certain cells to consider both their SSN and CCNneighbors to solve QoSNet in such a situation, should help save en-ergy resources and extend the network lifetime. One way ofaddressing this issue is to take a probabilistic approach based onthe percolation theory. Introduced by Broadbent and Hammersley[5], the percolation theory has been widely applied in variousfields, including economics, biology, sociology and communication.In the field of wireless networks, this theory has been used in thepast to study node connectivity.

Fig. 3. Data forwarding strategy in QoSNet.

Site percolation is a kind of random graph in which the edgesare formed only between open neighboring sites. In site percola-tion, each site can be either open, with probability pc , or closed,with probability 1� pc . An edge exists only where two adjacentopen sites are connected. All sites are connected and become acluster if any pair of open sites is connected. A cluster that spansfrom one side of the plane to the other is said to be an infinite clus-ter. A rigorous mathematical study of site percolation has beenconducted in [4]. To model the QoSNet resolution to site percola-tion, consider the square grid in Fig. 4, in which each cell repre-sents a site. A site is considered open when the nodes in that sitesolve the QoSNet problem by considering their neighbor in theSSN, and in the CCN when they fail to find a solution with theSSN neighbors. A closed site corresponds to the situation whenthe node that has a packet to forward solves the QoSNet problemby using only its SSN neighbors. Obviously, when the value ofthe percolation probability is pc ¼ 1, all network sites are openedto solve the QoSNet problem in both SSN and CCN networks.Fig. 4 illustrates the way network nodes solve the problem. A darksquare represents the cells that determine their forwarding set byconsidering both their SSN or CCN neighbors exclusively, while theblank boxes represents the squares that solve it using the SSNneighbors only. First, the node solves the problem by consideringits SSN neighbors, and if a feasible solution is not found, the nodeattempts to solve it by considering the CCN neighbors if their cell isopen.

So far, we have shown how to model the QoSNet resolution byway of site percolation. However, the way the various cells are se-lected has yet to be defined. Indeed, let n(t) depict the number ofinstances the nodes in a given cell must solve QoSNet, to determinethe forwarding set Fw during a period of time t; nssnðtÞ the numberof successful attempts to find a feasible solution when the nodesconsider their SSN neighbors. These parameters are collected bythe CC on a periodic basis. The CC activates its cell by notifyingits members to also consider the CCN neighbors for QoSNet resolu-tion. The probability pc for a CC to be activated is computed asfollows:

pc ¼ 1� nssnðtÞnðtÞ ð13Þ

Illustration: assume nodes in cell / ¼ 100001 altogether have suc-cessfully found feasible solutions nssnðtÞ ¼ 100 during the elapsedperiod of time t, while receiving nðtÞ ¼ 120 to forward. The CC ofthat cell will activate its cell with the probability pc ¼ 0:16 to allowthe nodes in that cell to consider the neighbors in NccnðsiÞ for thesubsequent time period. If the cell is activated, its CC will notifyits member, in order to use this new setting to process the incomingpackets to be forwarded. Table 1 shows the QoSNet pseudo-codeexecuted by the node when it has packets to forward.

4. Performance evaluation

This section shows the effectiveness of the proposed routingscheme in terms of on-time packet delivery ratio, average end-to-end delays, and the network lifetime. For that purpose, exten-sive simulations were conducted in Qualnet [23] to assess the per-formance of QoSNet. Baseline comparisons were made with theMCMP [16] and the God routing. These schemes are selected forthe following reasons: (a) the God routing indicates the higherachievable performances of an ideal QoS routing protocol; (b) theMCMP relies on Soft-QoS constraints by setting the values of pre-defined parameters. Instead, our goal in this comparison is to showthat, hard QoS constraints can still be met when using our scheme.Moreover, the MCMP protocol is more tied to our work in term ofrouting strategy. Indeed, in the God routing, each node knows the

Fig. 4. Percolation stages for QoSNet resolution.

Table 1QoSNet routing algorithm.

(A) When the node receive a packet to forward1. TðiÞ NssnðiÞ; ForwardingsetFwðiÞ ¼ ;2. candidatesetCwðiÞ ¼ flij=hj < hi and Di þ dij < Dreq; J 2 TðiÞg3. If CwðiÞ–;4 compute Lr

i ¼ffiffiffiffiffiRi

hip

;5. do6. Select the node v 2 CwðiÞ with the maximum residual energy bv ;7. FwðiÞ FwðiÞ [ fvg;8. until constraint (9a) is met;9. else if ðCwðiÞ ¼ ; or FwðiÞ ¼ ;Þ then10. If the cell is activated then11. TðiÞ NccnðiÞ; goto 2;12. else13. update statistics;14. discards the packet and return;15. end if16. end if

(B) When the node is a Cell Controller1. At the expiration of the timer Tpercolation do2. collect the statistics from cell members;3. compute the percolation probability pc;4. activate the cell with probability pc and send notification to

members;End

Table 2Simulation parameters.

Parameter Value

Network area 200 m � 200 mSensors 256Transmission range 40 mBandwidth 250 kbpsPacket size 128 bytesSimulation time 1000 sReliability [0.8, 0.9]Reporting rate 1 packet/sPercolation period 50 s

Fig. 5. Effect of percolation on the delivery ratio.

T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342 1339

state of all network links and nodes. This scheme can be viewed asan upper bound of a QoS routing algorithm with end-to-end delaysand reliability requirements. The average delivery delay equals thelatency experienced by successfully received packets. On-timepacket delivery ratio means the number of successfully receivedpackets that satisfy the QoS requirements over the total numberof generated packets.

Our simulations assess the worst-case performance where linkdelays and reliability levels always change suddenly at any trans-mission instant, making them unpredictable. The success probabil-ity of each transmission is randomly picked from [0.8, 0.9], whichimplies that the link reliability ranges from 0.8 to 0.9. Link delaysare also randomly distributed over the range of [1, 50] ms. Link de-lays consist of the time elapsed to successfully transmit a packetafter it was received. It thus includes queuing time, contentiontime, as well as transmission, retransmission and propagationtimes. The sink is located on the upper right corner of the sensorfield. Over time, energy is dissipated at regular intervals until arouting hole appears after time T, which is known as the discon-nection time or the network lifetime. Let Ti denote the disconnec-tion time of an individual node. The network lifetime equals thetime when the residual energy at any one of the sensor nodereaches zero, and is given as:

T ¼min fTigni¼1

� �ð14Þ

The simulation implemented in Qualnet follows the parameter set-tings shown in Table 2. The delay constraints are taken in the rangeof [60,120] ms and the default reliability threshold is set at 0.7. Forthe MCMP routing, both parameters a and b are set at 95%. A data

packet contains 128 bytes. Different random seeds were appliedto generate different network configurations and all simulationslasted 1000 s. The results show an average of over 15 simulationruns.

To highlight how percolation affects the performance indexes, ascenario was simulated where each CC activates its cell with a gi-ven percolation probability. Figs. 5–7, respectively, illustrate theon-time packet delivery ratio, the average end-to-end delay, andthe network lifetime for various percolation probabilities. Indeed,allowing the participation of CCN nodes in routing decisions helpsimprove the QoS requirements. However, there is a price to pay forsuch performance, as evidenced by the network lifetime which de-creases when many CCN nodes are involved in routing decisions. Ifthe nodes in CCN are made of powerful devices like actors or mo-bile robots, as described by the AON in our previous work [3], thatwill results in better performances; in such scenario, wireless linksbetween such devices are more reliable and are subject to shorttransmission delay.

Figs. 8 and 9, respectively, present simulation results by com-paring other routing schemes for on-time delivery ratio and theaverage latency experienced by the delivered data packets. The

Fig. 6. Effect of percolation on the average packet delay.

Fig. 7. Effect of percolation on network lifetime.

Fig. 8. Average delivery ratio vs. delay requirement.

Fig. 9. Average packet delay vs. delay requirement.

Fig. 10. Network lifetime vs. delay requirement.

1340 T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342

QoSNet routing scheme performs better in terms of delivery ratiowhen compared to the MCMP scheme, since it takes the residualenergy of the routing nodes into account. This performance canbe explained by the fact that some nodes in the MCMP schemeare frequently chosen and soon run out of energy. Obviously theGod routing scheme outperforms all of the studied schemes dueto the fact that all network nodes are aware of the state of all net-work links.

The QoSNet scheme is more akin to the God routing as thenodes know the state of the local links, while the knowledge is glo-bal in the God routing scheme. Fig. 10 illustrates the performanceof the respective schemes in terms of network lifetime when thedelay requirement varies. As expected, since the QoSNet schemetakes into account the residual energy of the nodes for path selec-tion, it seems that the network lives longer than that of the MCMPscheme, which only minimizes the number of paths in order tosave energy. The QoSNet scheme outperforms the God routingscheme for low delay requirements as in God routing scheme,the node selects links with the least delays, potentially includingnodes with low residual energy. However, the QoSNet scheme onlyselects links with nodes endowed with high residual energy: thisexplains why it outperforms God routing in terms of low delayrequirements.

Figs. 11 and 12 illustrate the results of the respective schemeswhen the value of the reliability requirement varies. As for the

QoSNet scheme, the delivery ratio is tied to the God routingscheme and does not influence the performance indexes. This situ-ation can be explained by the fact that the paths selected satisfied

Fig. 11. Average delivery ratio vs. reliability requirement.

Fig. 13. Average packet delay vs. number of nodes.

Fig. 14. Average delivery ratio vs. number of nodes.

T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342 1341

together the reliability requirements. However, the gap betweenthe respective schemes in terms of the delivery ratio is due tothe fact that when the nodes do not find a feasible solution, theydrop the packet to be forwarded; MCMP thus drops more packetsthan other schemes.

Figs. 13–15, respectively, illustrate the average end-to-end de-lay, the on-time packet delivery ratio, and the network lifetimewhen the number of deployed nodes varies. It clearly appears thatusing the QoSNet routing protocol offers a good benefit comparedto the MCMP scheme in terms of delay and delivery ratio. As ex-pected, the node redundancy increases, so that the network parti-tion time is prolonged. The nodes in the paths selected by QoSNetlive longer as evidenced in Fig. 15. This resilience allows a diversityof choices among the nodes in the vicinity of the routing node, andthen favors the lower delay links to be considered as forwardinglinks. The QoSNet scheme only selects links with nodes endowedwith high residual energy and then outperforms the God routingin term of network lifetime. Indeed, with the God routing schemethe node selects links with the least delays, potentially includingnodes with low residual energy. Consequently, the same nodetransmits more packets overall with God Routing than QoSNet,and so the network depletes in advance. Meanwhile, the God rout-ing scheme performs better in terms of delay and delivery ratio, asit models the ideal case where each node is aware of the states of

Fig. 12. Average packet delay vs. reliability requirement.

Fig. 15. Network lifetime vs. number of nodes.

every node and every link in the network, and then it selects thepath accordingly.

When the number of nodes increases, the path diversity is en-hanced, and the multiplicity of the links that satisfy the QoSrequirements allows the schemes to select the best possible path.As it is impossible in reality for a node to apply the God routing

Fig. 16. Number of alive nodes overtime.

1342 T. Houngbadji, S. Pierre / Computer Communications 33 (2010) 1334–1342

scheme, the QoSNet scheme is a good alternative because it is clo-ser to the God routing scheme than MCMP. Fig. 16 illustrates thenumber of alive nodes over the time. The simulation runs 2000 sand the number of alive nodes is taken every 200 s. The nodes inMCMP are depleted more quickly because that scheme reducesthe number of paths in order to prolong the network lifetime,while QoSNet relies on the nodes residual energy and the switch-ing mechanism of percolation to achieve the same goal. The energyhole created by this depletion dramatically reduces the lifetime ofthe network in MCMP. This justifies the results depicted in Figs. 15and 16 on the network lifetime and the number of alive nodes inthe network.

5. Conclusion

This paper presents a new routing strategy to provide QoS inlarge scale WSNs. Existing routing algorithms that aim to provideQoS, focus only on end-to-end delays and link reliability in rela-tively small wireless networks. To deal with the scale factor, theproposed scheme splits WSNs into two sub-networks composedof nodes spread over a discretized area. The first includes cell con-trollers and possibly other powerful nodes such as actors, while thesecond is composed of the remaining sensor nodes. Consideringthe QoS metrics such as end-to-end delays and reliability, the pro-posed routing strategy is designed for large scale WSNs whose goalconsists of extending network lifetime. By using the percolationtheory, a switching mechanism allows the node that has a packetto forward to select forwarding nodes among its neighbors in thetwo sub-networks. Simulation results validate our scheme byshowing its efficiency in terms of mean end-to-end delays, on-timepacket delivery ratio, and most importantly, extended networklifetime.

References

[1] F.A. Ian, M. Tommaso, R.C. Kaushik, A survey on wireless multimedia sensornetworks, Computer Networks 51 (2007) 921–960.

[2] L. Ye Ming, W.S.W. Vincent, An energy-efficient multipath routing protocol forwireless sensor networks: research articles, International Journal ofCommunication System 20 (2007) 747–766.

[3] T. Houngbadji, S. Pierre, SubCast: a distributed addressing and routing systemfor large scale wireless sensor and actor networks, Computer Networks 53(2009) 2840–2854.

[4] H.-O. Peitgen, Jürgens, Hartmut, Saupe, Dietmar, Chaos and Fractals: NewFrontiers of Science, Springer, New York, 2004.

[5] S.R. Broadbent, J.M. Hammersley, Percolation processes I. Crystals and mazes,Proceedings of the Cambridge Philosophical Society 53 (1957) 629–641.

[6] J.N. Al-Karaki, A.E. Kamal, Routing techniques in wireless sensor networks: asurvey, IEEE Wireless Communications 11 (2004) 6–28.

[7] K. Sohrabi, J. Gao, V. Ailawadhi, G.J. Pottie, Protocols for self-organization of awireless sensor network, IEEE Personal Communications 7 (2000) 16–27.

[8] J.A. Stankovic, T.E. Abdelzaher, L. Chenyang, S. Lui, J.C. Hou, Real-timecommunication and coordination in embedded sensor networks, Proceedingsof the IEEE 91 (2003) 1002–1022.

[9] E. Felemban, L. Chang-Gun, E. Ekici, MMSPEED: multipath Multi-SPEEDprotocol for QoS guarantee of reliability and timeliness in wireless sensornetworks, IEEE Transactions on Mobile Computing 5 (2006) 738–754.

[10] B. Najet, S. Yeqiong, A new routing metric for satisfying both energy and delayconstraints in wireless sensor networks, Journal of Signal Processing andSystems 51 (2008) 137–143.

[11] P. Dario, M. Tommaso, F.A. Ian, Routing algorithms for delay-insensitive anddelay sensitive applications in underwater sensor networks, in: Proceedings ofthe 12th Annual International Conference on Mobile Computing andNetworking, ACM, Los Angeles, CA, USA, 2006.

[12] E. Sinem Coleri, V. Pravin, Energy efficient routing with delay guarantee forsensor networks, Wireless Networks 13 (2007) 679–690.

[13] R.M. Mahmood, G. Sathish, C.M.L. Victor, Multiobjective routing forsimultaneously optimizing system lifetime and source-to-sink delay inwireless sensor networks, in: Proceedings of the 29th IEEE InternationalConference on Distributed Computing Systems, IEEE Computer Society,2009.

[14] K. Akkaya, M. Younis, Energy and QoS aware routing in wireless sensornetworks, Cluster Computing 8 (2005) 179–188.

[15] P.K. Pothuri, V. Sarangan, J.P. Thomas, Delay-constrained, energy-efficientrouting in wireless sensor networks through topology control, in: Proceedingsof the IEEE International Conference on Networking, Sensing and Control,2006, pp. 35-41.

[16] H. Xiaoxia, F. Yuguang, Multiconstrained QoS multipath routing in wirelesssensor networks, Wireless Networks 14 (2008) 465–478.

[17] B.B. Antoine, G.M. Kuzamunu, Energy constrained multipath routing inwireless sensor networks, in: Proceedings of the Fifth InternationalConference on Ubiquitous Intelligence and Computing, Springer, Oslo,Norway, 2008.

[18] O. Chipara, Z. He, X. Guoling, C. Qin, W. Xiaorui, L. Chenyang, J. Stankovic, T.Abdelzaher, Real-time power-aware routing in sensor networks, in:Proceedings of the 14th IEEE International Workshop on Quality of Service,2006, pp. 83–92.

[19] B. Deb, S. Bhatnagar, B. Nath, ReInForM: reliable information forwardingusing multiple paths in sensor networks, in: Proceedings of the 28thAnnual IEEE International Conference on Local Computer Networks, 2003,pp. 406–415.

[20] Y. Fan, Z. Gary, L. Songwu, Z. Lixia, GRAdient broadcast: a robust data deliveryprotocol for large scale sensor networks, Wireless Networks 11 (2005) 285–298.

[21] Y. Xu, J. Heidemann, D. Estrin, Geography-informed energy conservation for adhoc routing, in: Proceedings of the Seventh Annual International Conferenceon Mobile Computing and Networking, Rome, 2001, pp. 70–84.

[22] G. Anastasi, M. Conti, M. Di Francesco, A. Passarella, Energy conservation inwireless sensor networks: a survey, Ad Hoc Networks 7 (2009) 537–568.

[23] Qualnet, Qualnet Simulator, 2007.