fault-tolerant and efficient data propagation in wireless sensor networks using local, additional...

J. Parallel Distrib. Comput. 67 (2007) 456–473www.elsevier.com/locate/jpdc

Fault-tolerant and efficient data propagation in wireless sensor networksusing local, additional network information

Ioannis Chatzigiannakisa,b, Athanasios Kinalisa,b, Sotiris Nikoletseasa,b,∗aComputer Technology Institute, P.O. Box 1382, 26500 Patras, Greece

bDepartment of Computer Engineering and Informatics, University of Patras, Greece

Received 5 May 2006; received in revised form 15 December 2006; accepted 29 December 2006Available online 13 January 2007

Abstract

We propose a new data dissemination protocol for wireless sensor networks, that basically pulls some additional knowledge about thenetwork in order to subsequently improve data forwarding towards the sink. This extra information is still local, limited and obtained in adistributed manner. This extra knowledge is acquired by only a small fraction of sensors thus the extra energy cost only marginally affects theoverall protocol efficiency. The new protocol has low latency and manages to propagate data successfully even in the case of low densities.Furthermore, we study in detail the effect of failures and show that our protocol is very robust. In particular, we implement and evaluate theprotocol using large scale simulation, showing that it significantly outperforms well known relevant solutions in the state of the art.© 2007 Published by Elsevier Inc.

Keywords: Wireless sensor networks; Data propagation; Energy efficiency; Distributed protocols; Performance evaluation; Simultation

1. Introduction

Wireless Sensor Networks are very large collections of tinysmart sensor devices that form ad hoc distributed sensing anddata management networks that collect detailed informationabout the ambient environment. In a usual scenario, thesenetworks are largely deployed in areas of interest (such asinaccessible terrains or disaster places) for fine grained monitor-ing in various classes of applications [1,8]. The flexibility andself-organization, fault-tolerance, high sensing fidelity, low-costand rapid deployment characteristics of sensor networks createmany new exciting application areas for remote sensing. In

This work has been partially supported by the IST Programme of theEuropean Union under contract number IST-2005-15964 (AEOLUS), the Pro-gramme PYTHAGORAS under the European Social Fund (ESF) and Op-erational Program for Educational and Vocational Training II (EPEAEK II)and the Programme PENED under contract number 03ED568, co-funded75% by European Union—European Social Fund (ESF), 25% by GreekGovernment—Ministry of Development—General Secretariat of Researchand Technology (GSRT), and by Private Sector, under Measure 8.3 of O.P.Competitiveness—3rd Community Support Framework (CSF).

∗ Corresponding author. Computer Technology Institute, P.O. Box 1122,26110 Patras, Greece.

E-mail addresses: [email protected] (I. Chatzigiannakis), [email protected](A. Kinalis), [email protected] (S. Nikoletseas).

0743-7315/$ - see front matter © 2007 Published by Elsevier Inc.doi:10.1016/j.jpdc.2006.12.003

the near future, this wide range of application areas will makesensor networks an integral part of everyday life. In this pa-per we study the problem of efficient and fault-tolerant datapropagation in wireless sensor networks. We propose a newprotocol which is simple, local and uses limited extra knowl-edge of the network that is obtained in a distributed manner.The protocol uses local information regarding the surroundingactual network conditions, acquired by appropriately varyingthe range of wireless communication, and then plans a path ofpairwise adjacent sensor devices that are used in the forward-ing (i.e. propagation) of data towards the sink. Neighboringdevices decide individually on whether to participate in prop-agation of events. The demand-driven sequence of plan & for-ward phases aims at better performance, compared to typicalfixed transmission range data propagation, most needed in somefrequently occurring situations like the case of low local den-sities of faulty sensor devices where fixed range protocols maytrap in backtracking actions when no devices towards the sinkare found; our protocol, by increasing the transmission range,may find such devices and avoid extensive backtracking.

This role-based approach where a limited number of devicesdo the high cost planning, while the rest operate in a low coststate, leads to systems that have increased energy efficiencyand high fault-tolerance, since the planning phases allow to

http://www.elsevier.com/locate/jpdc

mailto:[email protected]



I. Chatzigiannakis et al. / J. Parallel Distrib. Comput. 67 (2007) 456–473 457

optimize propagation paths and bypass obstacles (where nosensors are available) or faulty sensors (e.g. due to physicaldamage, power failure). We show that the cost of forward plan-ning is amortized by the low energy short-range optimized hop-by-hop transmissions performed by our protocol; this selectivespending of energy increases the lifetime of the network andthe total number of events successfully reported to the controlcenter.

The basic idea of our approach is to trade-off the cost of ob-taining a certain amount of limited extra knowledge with theperformance gains achieved by using this additional knowl-edge in the subsequent propagation of data. By obtaining thisextra knowledge about the network conditions (e.g. energy ac-tually available at sensors, distance to the sink, faults etc.) at asomewhat global level, several performance measures (such asenergy dissipation, latency, fault-tolerance) can be improved.

We implement and evaluate our protocol using simulation,showing that it significantly outperforms existing, well estab-lished relevant solutions in the state of the art. In particular, wedemonstrate the above performance properties by comparingthe new protocol to the well known Directed Diffusion (DD)paradigm [26] for information dissemination in wireless sen-sor networks using several important efficiency measures witha focus on energy dissipation, success rate and delivery delay.The extensive simulations that we present here, highlight thebehavior of the internal mechanisms of the new framework andgive useful insight on the fine-tuning of the various networkparameters. The findings indeed demonstrate that our protocolachieves significant improvements in energy efficiency, highersuccess rates in faulty networks of low densities, and managesto disseminate data to their destination faster.

1.1. Related work and comparison

While there is an imperative need for reliable and low latencydissemination of data in wireless sensor networks, optimizeddata propagation paths are generally infeasible (or at least verydifficult) to obtain. Such optimized end to end solutions existfor other types of more structured networks, even though ob-taining paths satisfying certain constraints in terms of latencyand cost is an NP-complete problem. For example, [16] pro-poses algorithms that use low error discretization in order tocompute approximations of the optimal solution fast. However,due to the limitations in the processing capabilities of the sen-sors and more importantly due to the lack of global knowledge,such solutions may not be directly applicable in wireless sen-sor networks. Thus, many routing protocols use only local opti-mization criteria, avoiding the cost (or even the infeasibility) ofobtaining an approximation of a global solution. In [27] the costof locality is investigated, in the sense that for general cover-ing and packing problems the authors provide upper and lowerbounds that characterize the achievable trade-offs between theamount of local information and the quality of the global so-lution. However, even when using only local knowledge andlocal interactions several problems can be solved while opti-mizing some performance aspects. For example, in [7] a novelcoverage-preserving scheme is proposed, which can resolve the

off-duty conflict and guarantee network coverage while mini-mizing the number of active sensors and the amount of con-sumed energy, by exchanging only local information instead ofusing global information and synchronized time scheduling.

Local optimization protocols (like the Local Target Protocol[15]) evolve in a greedy fashion trying to make optimal choicesbased on network knowledge within the typical fixed trans-mission range of the sensor currently possessing data underpropagation. Such protocols tend to be more suitable in densenetworks, with rather “uniform’’ conditions, i.e. where localsamples of the network tend to be representative of it as a whole.Our protocol instead performs optimizations at a more global(yet limited) level, taking advantage of the extra knowledge ob-tained. Several protocols in the state of the art (most notablyDD [26]) try to maintain and update some global structure, suchas a set of paths towards the sink to pull down data. Such ap-proaches perform well in networks of low dynamics but theirefficiency may drop in networks with many frequent changesand failures. Our protocol tries to become aware of the current,actual network conditions and accordingly optimize; however,this is done at a relatively local level in order to avoid collectedknowledge becoming obsolete in the case of high dynamics.Furthermore, no structure or hierarchy are maintained by ourprotocol; once network information is obtained and optimizedpaths are chosen, data propagation happens in a hop-by-hopmanner.

Our multi-hop approach is also in contrast to clusteringprotocols such as LEACH [24]. In such protocols, sensors self-organize themselves into clusters; in each cluster, only a sin-gle cluster head transmits directly to the sink, while the rest ofthe sensors propagate data to their cluster head. Such protocolsperform well in small area networks of low event generationrate; however in larger networks of high event generation rates,transmissions happen at large distances and rotation of clusterheads may be too slow to avoid their energy depletion.

Probabilistic forwarding schemes (like PFR [11]) performredundant optimized multi-path transmissions to combineenergy efficiency and fault-tolerance. Such protocols, althoughwell suitable in sparse networks, tend to spend a lot of energyin the case of high densities. In [22] the authors present twopath formation algorithms that construct k disjoint or braidedadditional alternate paths such that no set of k node failures caneliminate all the paths. The redundant paths are constructedbased on local criteria using the path reinforcement techniqueof [26]. The authors of [23] attempt to minimize energy con-sumption while still achieving k fault-tolerance. Since optimallysolving this problem is NP-hard, the authors propose threeapproximation algorithms achieving O(k) approximation. Incontrast to these approaches, our protocol does not maintainend to end multiple paths so it incurs no additional com-munication or computational cost at the sensors. Instead, thefault-tolerance property of our protocol is based on the abilityto very quickly repair a path once a faulty intermediate nodeis detected. The price of this simplicity is that k fault-tolerancecannot be guaranteed.

The efficient detection of faults in distributed environmentsis examined in [20]. The authors are based on a comparison

458 I. Chatzigiannakis et al. / J. Parallel Distrib. Comput. 67 (2007) 456–473

diagnostic model where nodes in static wireless networks dis-seminate tests to their neighbors and compare the replies to theirown calculation. The proposed protocol then constructs anduses a spanning tree to disseminate the local diagnosis resultsin the whole network thus allowing the discovery of all failednodes. This work is significantly extended in [21] where faultdiagnosis is considered in MANETs. A mobility aware com-parison diagnostic protocol is proposed along with an adaptiveprotocol that maintains the spanning tree used to disseminatethe local diagnostic views. The proposed protocol reconfiguresthe tree as topology changes due to node mobility or due tonode failures. The above protocols identify both hard faults,where affected nodes seize to communicate with the rest of thenetwork, and soft faults, where affected nodes produce erro-neous output. In our work here we consider only hard faultsassuming no explicit mechanism to detect faulty sensors otherthan detecting a communication error. Thus, fault discovery inour case is done only on demand when a sensor attempts tocommunicate to a faulty sensor and in a purely local fashion.

For a survey of data propagation protocols, see [6]. A collec-tion of algorithms covering several aspects of wireless sensornetworks like deployment, coverage, routing and sensor fusioncan be found in [30]. Also, efficient protocols for fundamentalproblems in optical smart dust networks are proposed in [18],while routing communications methodologies are given in [2].

In [4], the computational complexity of the localization prob-lem is studied for the first time, proving that it is NP-hard insparse networks. We note that, also in view of this result, weavoid to solve a localization procedure since the network westudy may be sparse; instead we obtain some implicit locality-related measurements (such as distances) at a local level. Infact, we neither assume an a priori sense of orientation, sincewe only progressively build such a knowledge. Clearly, wecan assume an explicit sense of orientation mechanism. Forexample, [19] presents self-stabilizing procedures for broad-cast, flooding and sense of direction in wireless sensor net-works. Such a protocol providing a sense of direction can beused by our protocol in order to obtain local orientation refer-ences (i.e. for the sensors to know a general direction towardsthe sink).

1.2. Main findings

Our extensive performance evaluation indicates that theamount of local information (on the surrounding actual net-work conditions) that is available to the protocol plays a crucialrole in the overall performance of the network. This extraknowledge does not need to be accessible by all the devices ofthe network; allowing access to only a small group of devicessuffices to considerably improve the overall performance. In-terestingly, the additional energy spent by this small group (toobtain the extra knowledge) does not affect the overall energydissipation, which is dominated by the high number of short-range transmissions employed during the data dissemination.Our protocol uses a simple collision-resolution mechanism toimprove the network performance in cases of dense deploy-ment of sensor devices and/or when no suitable underlaying

MAC protocol is available. By carefully adjusting the protocolparameters we can trade-off latency with collision resolution.In fact, a limited increase to the delivery delays may lead todramatic reductions to the total number of dropped packets.

In order to acquire a more complete view on the perfor-mance of our protocol, we conduct a comparative study withDD, a representative global structure based approach. Theextensive experiments that we conducted, highlight the advan-tages of our approach that achieves significant improvementsin energy efficiency, higher success rates in networks of lowdensities, and manages to disseminate data to their destinationfaster. We move beyond the typical study of networks with nofailures, and investigate the performance of the two protocolsin the presence of permanent node failures. We show thateven under harsh conditions where more than 50% of the net-work becomes inoperable, our protocol still outperforms DDin all fundamental performance metrics (success rate, energydissipation and delivery delay).

An early version of some of the ideas of our work haveappeared in [14], in the second International Conference onDistributed Computing in Sensor Systems (DCOSS, 2006).

2. A simple model for sensor networks

We abstract the technological specifications of existing wire-less sensor systems [17,25]. Each node in our model is a fullyautonomous computing and communication device, is equippedwith a set of monitors (e.g. sensors for temperature, humid-ity etc.) and is characterized mainly by its available powersupply (battery) and the energy cost of computation and trans-mission of data. The communication equipment broadcastsmessages to nearby devices within range R that can vary (i.e.the transmission power can be set at appropriate levels). Fol-lowing [3,12,24,28], for the case of transmitting and receivinga message we assume the following simple model where theradio dissipates Eelec to run the transmitter and receiver cir-cuitry and amp for the transmit amplifier to achieve acceptableSNR (signal to noise ratio). We also assume that the signalattenuation over distance r from the source is proportional tor2. Thus to transmit a k-bit message at distance r in ourmodel, the radio expends

ET(k, r) = Eelec · k + amp · k · r2

and to receive this message, the radio expends

ER(k, r) = Eelec · k,

where Eelec and amp are constants that characterize the hard-ware radio module. Overall, there are three different kinds ofenergy dissipation: (a) ET, the energy dissipation for transmis-sion; (b) ER, the energy dissipation for receiving and (c) Eidle,the energy dissipation for idle state. For the idle state, we as-sume that the energy consumed for the circuitry is constant foreach time unit and equals Eelec (the time unit is 1 s).

We consider a simple sensor network for remote surveillanceof a region. In practice, such a network might consist of sev-eral hundreds or thousands of sensor devices deployed within


Sensor nodesSensor field

Control Center

Fig. 1. A wireless sensor network.

that region (see Fig. 1). Let n be the total number of sensordevices, that are present in an area of size A. In some cases,the devices may be deployed in a regular fashion (e.g. a lattice,or a linear array) within that region. More generally, however,communication and networking protocols cannot assume struc-tured sensor fields. Here, we assume that the sensor devices aredeployed in a random manner, do not move and that the settingis two-dimensional.

A user of this remote surveillance system, which we call thesink S, may contact the sensor devices in order to acquire infor-mation regarding the environmental conditions. In this sense,the user injects sensing tasks in the network, i.e. by broadcast-ing messages with a task description; the system can support avariety of task types [8]. Those sensor devices that match thetask description report to S using hop-by-hop wireless com-munication and routing mechanisms described in Section 3.

The networks that we consider in our model are prone to fail-ures due to the following reasons: (i) the components that makeup a sensor device are of low-cost and also of low-reliability;(ii) the area of deployment may be harsh and unfriendly (e.g.terrain with water puddles, animals that run over the sensors)thus many operational failures may occur. We here model halt-ing errors by introducing the failure rate F : the number of sen-sor devices that permanently fail to function per unit of time.For each unit of time, F failures occur at randomly chosennodes, instantly, and no further computation and/or communi-cation can be performed by these failed nodes.

3. Our data dissemination protocol

The basic idea of our approach is that dissemination of in-formation towards S is carried out within the wireless sensornetwork using a series of interchanging phases (see Fig. 2):(i) the listening phase (the device is sensing the environmentand passively listening for messages); (ii) the planning phase(the device is preparing to propagate data to the sink S) and(iii) the forwarding phase (the device is participating in datapropagation).

Given a particular environmental event that is sensed by adevice p, and a surveillance (sensing) task that is set by S,

Planning

Phase

Forwarding

Phase

Listening

Phase

Fig. 2. The protocol phases.

a new message M is generated by p. Our goal is to use alimited (by , a protocol parameter, that can be set by thenetwork implementor, which is described below) number oflong range transmissions to collect information regardingneighboring nodes and then plan a series of short range, lowpower transmissions between nearby particles, based on certainoptimization criteria, so that data is propagated to S. This plan& forward procedure provides: (i) high fault-tolerance as longrange transmissions allow to select paths that bypass obsta-cles (where no sensors are available) or faulty sensors (thathave been disabled e.g. due to power failure); (ii) increasedenergy-efficiency because of the path optimization performedover long range and also as short-range hop-by-hop transmis-sions can effectively overcome some of the signal propagationeffects in long-distance transmissions and (iii) enhanced se-curity as the low energy transmissions protect from undesireddiscovery of the data propagation operation.

In our protocol, each sensor uses two small-sized data struc-tures: (i) the neighbors cache that stores a small set of infor-mation about the active neighboring devices and (ii) the pathcache, a list of node IDs that keeps track of the last path usedto propagate data to S. The size of the neighbors cache is basedon the density of the network while the path cache is boundedby the protocol parameter . These structures are maintainedduring the listening phase and are extensively used during theplanning phase.

3.1. Initialization phase

We assume that there exists an initialization phase of thewireless sensor network during which all devices initialize theirlocal caches and execute an underlying localization protocoll . Since, in our model, the sensor devices cannot move, thisphase is executed only once and does not impose any furtheroverheads to the execution of the network. The protocol l isused by the sensors so that they can be able to estimate theirdistances within a certain accuracy factor, that depends on thecurrent technology advance and the actual protocol l . Let d(i, j)

be the Euclidean distance of sensor devices i, j and des(i, j) bethe estimation of this distance measured by sensor devices. Notethat des is not necessarily an exact value but rather an estimate ofthe real distance d; we however assume that measured distancesare analogous to real ones.


Protocol l may operate without any common sense of orien-tation or any geolocation abilities; obviously, assuming specialhardware equipment (e.g. smart antennae or GPS) makes thistask even easier. Such a protocol is presented in [29] (and iscompatible with our model of Section 2) that assigns fictitiousvirtual coordinates to all the devices of the network.

3.2. The listening phase

In this phase, sensor devices stay idle passively listeningfor nearby devices that transmit announcements (see planningphase below) or that respond to the announcements, until: (i)a new event is sensed that matches the interests given by S or(ii) a message M is received from another device.

When a device p′ listens an announcement from p, it firstchecks if des(p

′, p)Rclose, a constant set by the protocolimplementor (see Fig. 5). If this is true, it adds a new pathP = p in the path cache. Then, it individually decides (basedon local criteria) whether to respond to this announcement ornot. The incentive here is to allow each device to control theenergy consumption by ignoring some low-priority tasks, ordeciding not to join a forwarding path when the network isdense and many neighboring devices have already joined. Thisdecision can be based on a mechanism that considers variouscriteria regarding the conditions of the device (e.g. availableenergy, current load levels, etc.), the local conditions of thenetwork (e.g. average neighborhood energy, local density, etc.)and even global conditions imposed by the network controller(e.g. operation-rule: all devices must join to increase successrate).

In [14], we propose a mechanism that allows the devicesto react locally on environment and context changes by us-ing a set of rules that are based on response thresholds thatrelate individual-level plasticity with network-level resiliency,motivated by the nature-inspired method for dividing labor, ametaphor of social insect behavior for solving problems [5].We plan to include this mechanism in extended versions of ourprotocol.

The device p′ continues to passively listen to any otherdevice p′′ that responded to the announcement of p, and ifdes(p

′′, S) < des(p′, S) (i.e. p′′ is closer to S than p′), it adds

p′′ in the neighbors cache. This passive listening allows thedevices to update their cache with additional network topol-ogy information. In some sense, devices take advantage of anylong range announcements conducted by nearby devices thatundergo the planning phase to better understand the surround-ing network conditions, and essentially reduce their own (fu-ture) needs to discover the neighborhood.

We here note that it is not necessary for the devices to con-stantly listen the radio channel, a very energy-consuming pro-cess. Our protocol can be combined with a lower-layer powerconservation scheme like the one proposed in [13] or easilyextended by incorporating sleep–awake schemes into the lis-tening phase as done in [28]. Based on the performance evalua-tion presented in Section 5, our protocol operates well even forsparse networks, or otherwise, sensor networks implementingaggressive sleep–awake strategies.

3.3. The planning phase

A sensor enters the planning phase when data needs to bepropagated. During this phase, p first examines the path cache.If the cache contains a valid path Pcache of intermediate de-vices, it concludes and proceeds to the forwarding phase. If nosuch path exists, then p examines the neighbors cache in or-der to construct a new path P that will be used to forward Mtowards S.

If the neighbors cache is outdated or empty, or because thecache contains very limited data regarding neighboring devicespreventing p from constructing a sufficiently long path, p triesto discover all neighboring devices. Given a transmission rangeR, p performs a high power data transmission with range ·R (is a protocol parameter) to announce its interest to disseminatedata.

Remark that during an announcement, it is possible that pwill not manage to discover all the neighboring devices becauseof message collisions occurring due to the concurrent responsesof the nearby nodes. In order to tackle this problem, we imple-ment a simple random back-off scheme during which p, aftermaking the announcement, waits for a predetermined amountof time ts. The nearby devices delay their response by a randomperiod tr, where 0 tr < ts. Of course, this mechanism can beavoided if the MAC protocol can properly handle collisions.In [10] distributed, contention-free self-organizing MAC proto-cols which do not assume a global time reference are proposed.A distributed, local approach like in [10] can be assumed torun in combination with our protocol at a lower level to resolvemedium access conflicts. However, even in the case where anefficient MAC layer is used, a small ts waiting period is stillneeded since a device does not know in advance the number ofneighbors that try to respond to the announcement.

Still, it is possible that the device cannot detect any neigh-bor (e.g. because of low density, routing obstacles, high rateof failures, etc.). In this case the protocol has reached a Dead-end situation [29]. A possible way to overcome this situationis by repeating the transmission of the announcement (in casesome devices decide to participate this time or some previouslyinactive devices are now awake). This solution can be moreeffective if it is combined with a Range Variation operation,then the new announcement will be transmitted in range ( +i) ·R, where i is a counter of the number of re-tries. Of coursethe sensor device cannot extend it’s transmission range beyonda hardware and energy imposed limit, thus in extreme casesfurther propagation of M will be impossible. This concept issimilar to the one presented in [3] where the sensor device mod-ifies its transmission range R according to a change-function.Another possibility is to use a Backtracking mechanism, sim-ilar to the one presented in [15], and send the message to adevice that can construct another path that bypasses the nodethat reached the dead end.

Given that p has acquired enough information about the sur-rounding network conditions, it selects a path P such that Mis delivered to another sensor device p′′′ that is closer to Sthan p. This selection can be optimized in several ways, e.g. byselecting the particle with the higher available energy resources,


Cache Lookup

Neighbors

DiscoveryPath Selection

L

M,Pcache .

M,P .

M .

Deadend

Recovery

Listening

PhaseForwarding

Phase

Planning

Phase

Fig. 3. The flow of the planning phase.

the particle that has the lowest message load, or even the particlethat has the most up-to-date cache. Clearly, the length of thepath P is characterized by the locality of the information kept inthe cache of p. If the knowledge about the neighboring devicesis limited, path P will be short.

As soon as p′′′ is selected, p separates all the neighbor-ing nodes, for which it has information in the cache, in sublists (L1, L2, . . . ,L) in a way such that ∀pj ∈ Li :(i − 1) · R < des(p, pj ) i · R. Then p chooses one sensordevice from each sublist Li (i ∈ 1, − 1) so that the pathP = p1, p2, . . . , p−1, p

′′′ is defined. We here consider anoptimization criterion for selecting one device from eachsublist that is based on the relative distances of the nodes.

The sensor device executes a preparation procedure duringwhich a bipartite multi-stage graph of stages is generatedbased on L. Each stage i of the graph G(V, E) contains verticesthat correspond to the sensor devices of Li and the edges of Gare between vertices of consecutive stages. Weights are assignedto the edges of G to reflect the estimated physical distance of thesensor devices that correspond to the adjoining vertices. Then,based on G, the protocol calculates the shortest path joining pand p′′′. The intuition for using a bipartite multi-stage graphis to reduce the total number of edges m = |V | and thereforereduce the complexity of the shortest path operation given thelimited processing capabilities of the devices.

Note that it is possible that the operation of splitting L in sublists (L1, L2, . . . ,L) may result in some sublists Li

being empty, probably due to low-density of sensor devices onthe particular sector of the neighborhood area or due to thepresence of a routing obstacle. In this case the protocol willproduce a path P of length l < . However, empty sublistsare bypassed in the constructed path, the path will contain onenode from each non-empty sublist and the final node will al-ways be in the non-empty sublist closest to L or in L whenL = ∅. This may force the protocol to perform hops (transmis-sions) in distance greater than the nominal transmission rangeR. Certainly, there might exist other strategies for selecting thepath P that emphasize other aspects (such as available energy,distance from the S) and/or may also include randomizationtechniques.

Fig. 4. Transmission example.

Rclose

Sp2

p’

p1p3

Fig. 5. Nearby sensor devices react to Announcements.

As soon as the decision on such a path P is made, the pro-tocol enters the forwarding phase, a schematic of the variousstages of the planning phase can be seen in Fig. 3. In the for-warding phase, a message containing the actual data as wellas the constructed path(M, P), is transmitted to the first sen-sor device in P (i.e. in the example of Fig. 4, p1). Then,every device pj that receives (M, P) forwards

(M, P − pj )

to pj+1. When device p′′′ (i.e. in the example of Fig. 4, p3)receives (M, ·) the forwarding phase concludes and the pro-tocol enters a new planning phase. Now p′′′ is responsible tofurther disseminate M towards S. The constructed paths forthe example presented in Fig. 5 can be seen in Table 1.

3.4. The forwarding phase

In the forwarding phase, given a message of type (M, P),the sensor device does the following:

P is not empty: The message contains information about apath of sensor devices. If the path P ′ = P −pj is not empty,P ′ is stored in the path cache (for an example see Table 1)


Table 1The contents of the path cache for the examples of Figs. 4 and 5

Particle Path cache

p p1, p2, p3p1 p2, p3p2 p3p3 emptyp′ p

and p sends (M, P ′) to p1 and sends a success message tosender(M) (i.e. to the device it originally received the infor-mation from); in case p generated M, no success message issent. Otherwise, if P ′ is empty, the protocol enters the planningphase.

P is empty: The message contains no information about thepath of sensor devices to use in order to propagate M towardsS. The protocol enters the planning phase.

After transmitting the packet, p will wait for p1 to send asuccess message in order to ensure that M was received prop-erly and the dissemination continues as planned. If p1 does notrespond within a predefined period of time, p assumes that thetransmission fails and retransmits the packet to p1. This pro-cess is repeated until either p responds with a success messageor until the maximum number of retries has been reached (aprotocol parameter). In this case, p decides that p1 is no longeractive, updates its cache (by removing p1 and P) and enters theplanning phase. Also, note that since p has an estimation of it’sdistance from p1, i.e. the hop distance is more or less known,we allow p to reduce or increase accordingly it’s transmissionrange to a level that the message will be received by p1. Thismechanism allows our protocol to perform long distance hopsas required by the calculated paths, thus effectively bypassingcertain routing obstacles or routing through areas of low nodedensity.

3.5. Protocol design aspects

In this section we discuss several aspects that affect the per-formance of routing protocols and explain in greater detail thedesign choices in our protocol (which we hereafter call the CKNprotocol) for dealing with these aspects, as well as their rela-tion to other solutions presented in established relevant work.The primary objective of our protocol is to construct low cost(especially in terms of energy consumption) and low latencydata propagation paths while taking into consideration the lim-itations of the sensor devices. The way CKN achieves this goalmight at first may seem counter-intuitive since several longrange, and therefore high cost, data transmissions are requiredin order to obtain the network information around a device.However, the cost, of transmitting an announcement and theresponses to that announcement, is mitigated by three factors:

(i) The constructed path is optimized in such a way that itminimizes the paths length and at the same time max-imizes the traversed distance towards the destination.

Combined with the fact that devices can reduce their trans-mission range to the lowest level (within a safety margin)where the message is audible by the receiver, more energycan be saved on the short-range transmissions.

(ii) The constructed path is transmitted as a part of the mes-sage. Intermediate nodes store this path in their path cacheand use it not only for forwarding the current messagebut also for subsequent messages. In this way paths aremerged and used for propagating several messages thus,the initial energy spend for the path construction is amor-tized.

(iii) Nodes overhear responses to announcements and updatetheir neighbor cache accordingly. Also, nodes with anempty path cache can immediately select their path to bethe initiator of the announcement (see Section 3.2). In thisway neighbor information collection and path construc-tion can be done passively without additional announce-ment transmission cost. Initially sensors may gain onlypartial knowledge and constructed paths will probably beof poor quality, but subsequently, enough neighbors willbe discovered and the quality of the constructed paths willimprove.

Note also that there is a small overhead on the amount oftransmitted data as the path is piggybacked in each message.The required storage in the message in bits is in the orderO() which is small enough to be manageable by the sensors.The same applies for the size of the path cache, assuming onesink, while for many sinks the memory requirement increaseslinearly according to the number of sinks. However, for theneighbor cache the storage space required can be O((R · )2).So it is possible that in some configuration sensor devices maynot be able to accommodate the full size of the neighbor cache.In that case, the protocol can use a mechanism (such as leastrecently used) to keep only enough nodes in the cache so thatpaths can be constructed. Since this is more likely to happen invery dense networks, even only a fraction of the neighbors willbe representative of the network topology and the constructedpath will be of good quality.

Thus our protocol achieves energy efficiency and low latencyusing extended network knowledge obtained locally, in a feasi-ble manner for the constrained hardware of the sensor devices.This approach differentiates CKN from other local optimiza-tion protocols (like [15]) where limited, to one hop, networkknowledge is used to greedily route data. In [31] three greedyalgorithms for wireless networks that achieve loop-free routingare studied: GEDIR, MFR and DIR. GEDIR always moves thepacket to the neighbor whose distance to the destination is thesmallest. The MFR method chooses the next hop that makesthe greatest progress towards the sink, this approach yields sim-ilar results to GEDIR. In the DIR method, the selected nexthop has the minimum angular distance from the imaginaryline joining the current node and the destination. The authorspropose variations of these algorithms that perform the samenext hop selection criteria but in an extended two-hop area.Flooding messages is proposed when these protocols reach adead-end situation. Likewise, our protocol uses information in


an extended area but unlike [31] we explicitly provide mecha-nisms to obtain such information in an energy efficient manner.Furthermore, instead of simply selecting a more suitable des-tination node, we optimize the whole path towards that node.Also, since our protocol always selects the next node closestto the sink than itself, we also achieve loop-free routing. InSection 3.3 we already discussed the possibility of reaching adead-end situation where the current node is the closest to thesink than all it’s neighbors. Instead of using a flooding mech-anism that raises many efficiency issues (i.e. broadcast stormproblem, collisions, high energy dissipation), we proposed arange variation scheme. Range variation can also be used toovercome worst case scenarios where most of the sensor de-vices near the sink have depleted their energy reserves, in thiscase our protocol can locate the few remaining nodes in dis-tance ( + i) · R and perform hops in distance greater than Rin order to reach these sensors.

Local path optimization but in a non-greedy manner is alsoperformed in [26]. During the network initialization a set ofpaths towards the sink is constructed throughout the network,forming a global data propagation structure. Subsequently thesepaths are maintained and optimized based on local criteria(latency, reliability of path). On the contrary, CKN constructsand maintains paths on demand only when data is propagated;however, this is done at a relatively local level in order to avoidcollected knowledge becoming obsolete in the case of high dy-namics. Also, in contrast to clustering protocols such as [24];no structure or hierarchy are maintained by our protocol; oncenetwork information is obtained and optimized paths are cho-sen, data propagation happens in a hop-by-hop manner.

We will see in Section 5 that our protocol is also fault-tolerant; during data propagation the acknowledge mechanismdescribed in Section 3.4 can detect unresponsive sensors, dueto hardware failure, power depletion, etc. This mechanism isvery simple, uses only local interactions and induces very lowcost, as opposed to [20,21] where nodes disseminate tests totheir neighbors in order to discover failures and then compareand propagate the list of failed nodes throughout the network,thus constructing a global list of failed nodes. When a failureis detected CKN reinitiates the planning phase and a new pathis selected. In relevant literature, fault-tolerance is achievedby proactively constructing and maintaining several, preferablydisjoint, paths between source and destination in such a waythat no set of k node failures can eliminate all the paths. Suchan approach is followed in [23,22], on the other hand in [11]a randomized solution is presented that uses multipath datapropagation to achieve fault-tolerance. While these solutionsachieve good fault-tolerance, they also impose the significantcost of constructing, maintaining and/or using multiple paths.In CKN another approach is followed, instead of proactivelyconstructing and maintaining multiple paths, we reactively con-front failures by repairing the paths as needed. Thus, we im-pose very low overhead on the sensors but as a trade-off theprotocol cannot guarantee k fault-tolerance.

Additionally, to the low cost and good efficiency of our pro-tocol, we feel that CKN is also highly versatile and config-urable since it allows many design choices to be changed. As

we discussed earlier there are several strategies for selectingthe path P that can emphasize other aspects and perform differ-ent optimizations. Below we present three alternative path con-struction methods, which we consider evaluating in the future,each using a different amount of network knowledge aiming atachieving various desirable behaviors.

(a) Energy-aware: In this case, the intermediate sensordevice pi is chosen such that among all devices found ina sublist Li contains the highest energy levels. To be ableto use this criterion, the devices responding to the beaconof p along with their ID they include an indication on thelevels of their available energy (i.e. Ei) when responding toannouncements.

(b) Randomized: Towards a good average case performanceof the protocol, we use randomization to avoid bad behaviordue to the worst case input distributions for each selection(i.e. sensor devices with high energy sources being far awayfrom each other in the energy-aware case and sensor devicesoptimally placed but with very low available energy in thedistance-aware case. Thus, in this variation each device pi isselected uniformly randomly from Li .

(c) Latency-aware: In this case, the optimization processconsiders network characteristics that affect the throughput ofpaths and tries to select the most appropriate least congestedpath. More specifically, nodes responding to announcementsinclude an estimation of congestion, measured as the numberof pending packets in their queues, and the number of dis-covered nodes in their cache. Then from each sublist the sen-sor device with the lowest congestion and the higher numberof neighbors is selected. In this way more reliable and lowerlatency paths are constructed since congested nodes, which aremore likely to drop packets, are avoided while also the inter-mediate nodes will have a high number of alternative paths incase the constructed path fails.

4. Performance metrics

In this work we wish to evaluate the performance of our pro-tocol based on the following three fundamental metrics: the suc-cess rate, the energy dissipation and the delivery delay. Theseperformance metrics characterize the ability of the protocol tocoordinate the sensor devices so that all messages regardingthe realization of environmental phenomena are transferred toS, in an energy efficient way and with minimum delay. Theimportance of each of the above metrics depends on the natureof the application since there are inherent trade-offs betweensuccess rate, energy and latency.

In the previous sections we described the basic design prop-erties of our protocol considering the existence of a single mon-itoring task and the dissemination of messages related to thistask. In real world scenarios, the system will serve many sens-ing tasks. Thus, reporting data will require a large number ofmessages to be propagated throughout the network.

Let K be the total number of crucial events (E1, E2, . . . , EK )that need to be reported to a particular S in the area and let usconsider that a data dissemination protocol manages to reportk number of these events.


Definition 1 (Success rate). Ps, is the fraction of the numberof events successfully propagated to the sink over the totalnumber of events, i.e. Ps = k

K.

Definition 2 (Total energy dissipated). Etot = ∑ni=1

(Einit

i

−Ei), where Einiti is the initial energy of sensor device i and

Ei the available energy of device i at the end of the systemoperation.

Definition 3 (Delivery delay). Let D be the total period of timethat elapsed since the realization of a crucial event E until itwas finally delivered to the sink S.

Furthermore, we consider three protocol specific metricsthat measure: (i) the Total number of announcements; (ii) theAverage Path Length and (iii) the Number of Collisions interms of dropped packets. These performance metrics char-acterize the network management overhead imposed by theprotocol stack (a combination of data dissemination andcollision-resolution protocols) given a set of monitoring tasks.These metrics provide useful insights on the effect of thevarious network and protocol parameters to the overall perfor-mance of the system.

Based on the energy cost model used (see Section 2), theenergy consumption for the transmission of a message is relatedto the distance that it is required to travel and its size in bits.Since announcement messages sent during the planning phasesare very short (e.g. a small, constant number of ka bits), theextra energy spent due to increasing the transmission rangeis not increased a lot; in any case, this extra energy is worthspending since the additional knowledge obtained allows formuch better path selection. Also, the propagation of the actualdata packets during the forwarding phase (which may be longerthan the short announcements, i.e. ki bits) is still performed ina multi-hop way, thus the energy spent in each hop is in theorder of R2.

In fact, based on the energy cost model, the total energy spentin each sequence of plan & forward phases is proportional toka · ( · R)2 + · ki · R2. If <

ki

kathen the energy cost of

the announcements is smaller than the actual data propagationenergy cost, i.e. the total energy is · k′

i · R2, where k′i is

constant, similar to other, common multi-hop approaches.Assuming that the sensor devices are random uniformly dis-

tributed on the area, the density can be calculated according

to [9] as (R) = (nR2)A . Basically, (R) gives the number

of sensor devices within the transmission radius of each de-vice in region A. Therefore, when R is set to R, the effec-tive density ((R)) of the devices becomes 2 times larger,which leads to many more devices responding to the announce-ment. To scan the same area and number of devices, a muchgreater number of short-range transmissions (at least 2) wouldbe needed. Of course, the increased number of concurrent re-sponses to the long-range announcement will potentially resultin a high number of collisions, random back-off (or other col-lision resolution mechanisms) are more efficient in our case(as shown by the low latency achieved and the high successrate) since devices found during a single scan can be better

coordinated with respect to many nearby announcements ofshorter range.

5. Performance evaluation

In this section we present a comparative evaluation studyof our protocol with the well established DD paradigm forinformation dissemination in wireless sensor networks [26].We implement our protocol at the same level of the networkstack with DD and use a higher layer sensing application thatinjects sensing tasks to the sensor network. A number of eventsis generated, corresponding to the sensing tasks, for propaga-tion to the sink S. The experimental evaluation is conductedbased on the commonly used Network Simulator (ns-2 ver-sion 2.26), that provides a quite detailed implementation of thephysical and MAC layers and allows detailed measurements ofmany variables (such as the energy dissipation) in simulationsof wireless networks.

In Table 2 we present the setup for the first series of theconducted simulation experiments. Note that with these topol-ogy settings, the corresponding (R) ranges from 6.28 up to12.57. In [26] the experimental study conducted also consid-ered networks of different sizes and number of sensor devicesbut with an almost fixed density (R) ≈ 9.817. Also, theenergy available to the sensor devices was set to high lev-els (Einit

i = 20 J), to create good enough initial conditions(in terms of available energy) where all the events can be de-livered. Thus, this setting allows us to compare the energydissipation of the protocols in a fair way, without risking anarbitrary number of failures due to strained energy reserveswhich would affect the results and make more difficult toreach meaningful conclusions. Also note that the assumed highlevel monitoring application produces events at a steady rate ,and each event is being sensed by a randomly chosen sensordevice.

We start the evaluation of our protocol by investigating theimpact of the various parameters on the performance of thenetwork. In the first set of experiments we examine the impact

Table 2Simulation setup for the experiments presented in Fig. 6

Simulation parameter Values for Fig. 6

Network area A Rectangle 500 m × 500 mSink position Lower left corner (0, 0)

Number of sensors n 200, 300, 400Nominal transmission range R 50 mRange for selecting next hop Rclose 50 mEnergy available on sensors Einit

i 20 JDiscovery range multiplier 1, 2, 3, 4Search time ts 75 msNumber of generated events 1000Event generation rate 2 events/sSimulation time 515 sMetrics Success rate, energy dissipation,

delay, announcements,path length, dropped packets


10097.595908580

70

60

50

40

30

20

104321

Success R

ate

%

Beta

200 particles300 particle s400 particles

0.5

0.45

0.44321

Energ

y C

onsum

ption (

Joule

s)

Beta

200 particles300 particles400 particles

0.2

0.15

0.10.09

4321

Dela

y (

sec)

Beta


0

200

400

600

800

1000

4321

Num

ber

of A

nnouncem

ents

Beta


4

3.5

3

2

1

4321

Path

Length

Beta


1000

10000

100000

4321

Dro

pped P

ackets

Beta


Fig. 6. Success rate (Ps), energy dissipation (Etot), delay (D), average number of announcements, average path length and average number of dropped packetsfor different values of ∈ 1, 2, 3, 4, various number of devices (n ∈ 200, 300, 400) and fixed search time (ts = 75 ms),

of parameter , i.e. the parameter that controls the transmis-sion range for the announcements and in extend the lengthof the path generated during the planning phase. Fig. 6 de-picts the six efficiency metrics discussed in Section 4. In this

first set, we fix the search time to ts = 75 ms and inject aset of sensing tasks that generate 2 events/s. The simulationduration is calculated according to the event rate and is longenough to allow all messages to be generated. Another 15 s





Number of sensors n 300Nominal transmission range R 50 mRange for selecting next hop Rclose 50 mEnergy available on sensors Einit

i 20 JDiscovery range multiplier 1, 2, 3, 4Search time ts 25, 50, 75, 100 msNumber of generated events 1000Event generation rate 2 events/sSimulation time 515 sMetrics Success rate, energy dissipation,

delay, announcements,path length, dropped packets

of simulation time are added to allow the arrival of delayedmessages.

The results depicted in Fig. 6 demonstrate the ability to im-prove the success rate of our protocol in the cases of sparsedeployment of sensor devices by adjusting . For all networkdensities considered, setting = 2 suffices to achieve a 100%success rate. However, when = 1 the protocol operates in apurely greedy fashion. Many announcements have to be trans-mitted and data propagation fails when a node does not haveany closer neighbor to the sink than itself. In terms of en-ergy consumption, as it was discussed in Section 4, the costof the long range transmission of announcement messages isindeed amortized by the high number of short-range transmis-sion of messages and the effective use of cache regardless ofthe density of the network and . However, we are pleasedto report that the latency of the network is improved when increases. This is related to the fact that the devices trans-mit fewer announcements as increases, thus reducing theoverall delay as the devices spend less time waiting for thenearby devices to respond to their announcements. Therefore,although the long-range transmission of announcements leadto a higher number of collisions, this does not critically affectthe performance of the network while the protocol managesto devise “good’’ paths based on the additional informationacquired.

In the second set of experiments (see Table 3 for the sim-ulation setup and Fig. 7 for the results) we examine the im-pact of parameter ts, that is the time period that a device waitsafter making an announcement so that nearby devices can re-spond; we refer to ts as the search time. The central idea foradjusting ts is to allow the responses to spread over a longerperiod of time and in this way increase the effectiveness ofthe collision resolution protocol. Of course, by increasing ts,the delivery delay of the network is also affected. However, asshown in Fig. 7, the overall degradation of the latency is lim-ited while the number of dropped packets (i.e. the number ofcollisions) is dramatically reduced (notice that the figure is inlogarithmic scale). This allows the devices to collect more in-formation regarding the neighboring devices and thus devise

longer paths. Interestingly, even if, by increasing ts, the needto make an announcement is reduced, the overall energy dissi-pation seem to remain fixed, implying that the overall energyconsumption is dominated by the short-range transmissions ofdata messages rather than by the long-range transmissions ofannouncements.

We now proceed with the comparative study of our protocol(CKN) with DD. In order to highlight the differences betweenthe two different approaches, we first evaluate the two protocolsin a “controlled’’ environment. In the following sets of experi-ments, since our protocol essentially variates the transmissionrange (based on ) in order to make long-range announcements,to make the comparison fair, when the network executes DD,we set the transmission range of the devices to · R. Notehowever that unlike our protocol, DD does not vary the trans-mission range through out the execution of the experiment. Forthese experiments we consider only the three efficiency metricsdiscussed in Section 4.

In the third set of experiments (see Fig. 8), we measure theperformance of the protocols when only one message needs tobe disseminated to S. This message is generated by the devicepositioned at (500, 500), i.e. the device that has the greatestdistance from S. In the second set of experiments (see Fig.9), we measure the performance of the two protocols wheneach device generates one message that needs to be dissem-inated to S, i.e. the two protocols must disseminate a totalof n messages. These two different cases allow us to investi-gate the performance of the two protocols in the extreme casewhen a message is generated far away from S and in the av-erage case when messages are sent from all possible positions(Table 4).

The two different sets of experiments show that in the worstcase, DD manages to deliver the message with higher successrate when = 1, while for higher values of , both protocols al-ways succeed. Furthermore, still in the worst case scenario, forall different values of considered, it seems that DD managesto deliver messages in shorter time, but at a higher energy cost.This is somewhat expected since DD has already computedavailable paths during network initialization, whereas CKNneeds to construct the path on demand. For the average case (seeFig. 9), again DD manages to deliver more messages than ourprotocol when = 1, however, in terms of energy consumptionand delivery delays, it is clear that our protocol significantlyoutperforms DD. In this scenario all nodes will be forced to de-vise a path at some point, seemingly DD has an advantage sinceit pre-constructs all paths. However, CKN overtakes DD thanksto the caching mechanism that allows the reuse of paths as wellas path formation without issuing announcements, thus the pathformation latency is effectively amortized. These experimentsalso indicate the impact of on the success rate and latency ofour protocol, while the overall energy dissipation is affected at avery limited way that almost suggests that it is independent of (Table 5).

In the set of experiments shown in Fig. 10a we evaluatethe two protocols for the general case, i.e. when we assumea set of sensing tasks that generate 2 events/s and each eventis being sensed by a randomly chosen sensor device. In this


10097.5

95908580

70

60

50

40

30

20

100.10.0750.050.025

Success R

ate

%

Search Time (sec)

CKN β=1CKN β=2CKN β=3CKN β=4

0.5

0.45

0.40.10.0750.050.025

Energ

y C

onsum

ption (

Joule

s)

Search Time (sec)


0.2

0.15

0.10.09

0.10.0750.050.025

Dela

y (

sec)

Search Time (sec)


0

200

400

600

800

1000

0.10.0750.050.025

Num

ber

of A

nnouncem

ents

Search Time (sec)


0

1

2

3

4

5

0.10.0750.050.025

Path

Length

Search Time (sec)


1000

10000

100000

0.10.0750.050.025

Dro

pped P

ackets

Search Time (sec)


Fig. 7. Success rate (Ps), energy dissipation (Etot), delay (D), average number of announcements, average path length and average number of dropped packetsfor different values of ∈ 1, 2, 3, 4, different search times ts ∈ 25, 50, 75, 100 ms and fixed number of devices (n = 300).

setting we generate a total of 1000 events and the simula-tion duration is calculated according to the event rate andis long enough to allow all messages to be generated. An-

other 15 s of simulation time are added to allow the arrivalof delayed messages. In contrast to the previous two sets ofexperiments (see Figs. 8 and 9), in this set of experiments


10097.5

95

90

85

80

70

60

50

404321

Success R

ate

%

Beta

CKNDirectedDiffusion

0.006

0.007

0.008

0.009

0.01

0.011

0.012

4321

Energ

y D

issip

ation (

Joule

s)

Beta

CKNDirected Diffusion

0

0.5

1

1.5

2

4321

Dela

y (

sec)

Beta


Fig. 8. Success rate (Ps), energy dissipation (Etot) and delay (D) of CKN and DD in the case of 1 event (at point (500, 500)), for ∈ 1, 2, 3, 4, fixedn = 300 and fixed ts = 75 ms.

we used the original implementation of DD, as described in[26], i.e. the transmission range is always set to R regardlessof .

This experiment clearly shows the superiority of ourapproach in all three efficiency metrics considered here. Bysetting the parameter = 2, our protocol achieves a 100%success rate, and delivers all messages to S with significantlylower delays and by spending fewer energy than DD.

In our last set of experiments we investigate the (more real-istic) scenario where stopping failures occur at the sensing de-vices. In contrast to the previous settings where the operationof nodes was guaranteed, in this set we examine the behaviorof the protocols under the presence of node failures. We use thefailure rate F (defined in Section 2) to control the harshnessof the environment and evaluate the fault-tolerance achievedby our protocol. We deploy n = 300 devices and set F =·nTsim

, where ∈ 0.1, 0.2, 0.3, 0.4, 0.5, 0.6 is a parameter thatcontrols the fraction of nodes that fail within the first Tsim =500 s of simulation time. Essentially, we allow the 10 . . . 60%

of the nodes to fail during the simulation period. Based on theresults shown in Fig. 10b, we observe that the failure rate mainlyaffects the success rate of the protocols while the energy con-sumption and delivery delay seems to be unaffected. Note thatwhen 0.3 the number of alive nodes at the end of the sim-ulation is similar to the case of deployment of n = 200 nodeswithout failures, examined in Fig. 6. However, in contrast to thecase of low densities, does not control the performance of theprotocol; although for = 2 the performance improves, furtherincreases lead to reduced efficiency in terms of the achievedsuccess rate. This is explained by the fact that higher values of lead our protocol form to longer paths. Since all nodes areequally likely to fail, the probability for a node failure to dam-age a path increases for long paths. When a path is discontin-ued the protocol is forced to construct new paths thus leadingto slightly increased delivery delays and energy consumption.Although DD seems to follow a similar pattern of behavior, ourprotocol achieves higher fault-tolerance for all cases of failurerates considered.


10097.5

95

90

85

80

70

60

50

404321

Success R

ate

%

Beta


0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

0.28

0.3

4321

Energ

y D

issip

ation (

Joule

s)

Beta


0

0.2

0.4

0.6

0.8

1

4321

Dela

y (

sec)

Beta


Fig. 9. Success rate (Ps), energy dissipation (Etot) and delay (D) of CKN and DD in the case of 300 events (1 per each sensor), for ∈ 1, 2, 3, 4, fixedn = 300 and fixed ts = 75 ms.




Number of sensors n 300Nominal transmission range R 50 m, 50, 100, 150, 200 for DDRange for selecting next hop Rclose 50 mEnergy available on sensors Einit

i 20 JDiscovery range multiplier 1, 2, 3, 4Search time ts 75 msNumber of generated events 1 at upper right corner (500, 500)

Simulation time 15 sMetrics Success rate, energy dissipation, delay

To further analyze the fault recovery mechanism of CKNin Fig. 11 we present the number of announcement messagestransmitted by our protocol in the same setting as Fig. 10b (seeTable 6). We deliberately omitted the case of = 1 to focus




Number of sensors n 300Nominal transmission range R 50 m, 50, 100, 150, 200 for DDRange for selecting next hop Rclose 50 mEnergy available on sensors Einit

i 20 JDiscovery range multiplier 1, 2, 3, 4Search time ts 75 msNumber of generated events 300 one at each sensorEvent generation rate 2 events/sSimulation time 165 sMetrics Success rate, energy dissipation, delay

on the cases presenting the most interesting results. We can seethat when = 2 the protocol performs the highest number ofannouncements in all examined settings of F , a behavior con-sistent with previous results (see Figs. 6 and 7). However, as


10097.5

95908580

70

60

50

40

30

20

10400300200

Success R

ate

%

Number of Particles


Directed Diffusion

0.3

0.4

0.5

0.6

0.7

0.8

400300200

Energ

y C

onsum

ption (

Joule

s)

Number of Particles


Directed Diffusion

0.5

0.3

0.2

0.15

0.10.09

400300200

Dela

y (

sec)

Number of Particles


Directed Diffusion

100

95908580

70

60

50

40

30

20

106050403020100

Success R

ate

%

Failed Particles %


Directed Diffusion

0.8

0.7

0.6

0.5

0.4

0.36050403020100

Energ

y C

onsum

ption (

Joule

s)

Failed Particles %


Directed Diffusion

0.5

0.3

0.2

0.15

0.1

0.05

6050403020100

Dela

y (

sec)

Failed Particles %


Directed Diffusion

97.5

a b

Fig. 10. (a) On the left, success rate (Ps), energy dissipation (Etot) and delay (D) of CKN and DD in the case of 1000 events, for n ∈ 200, 300, 400, ∈ 1, 2, 3, 4 and fixed ts = 75 ms; (b) on the right, success rate (Ps), energy dissipation (Etot) and delay (D) of CKN and DD in the case of 1000 events,when 0 . . . 60% of the n = 300 nodes fail (F ∈ 0.0, 0.6, 0.12, 0.18, 0.24, 0.30, 0.36 failures/s), ∈ 1, 2, 3, 4 and ts = 75 ms.


100

90

80

70

60

50

40

30

20

6050403020100

Num

ber

of A

nnouncem

ents

Failed Particles %

CKN β=2CKN β=3CKN β=4

Fig. 11. Number of announcements of CKN in the case of 1000events, when 0 . . . 60% of the n = 300 nodes fail (F ∈ 0.0, 0.6, 0.12,

0.18, 0.24, 0.30, 0.36 failures/s) within the Tsim = 500 s of simulation; ∈ 2, 3, 4 and ts = 75 ms.


Simulation parameter Values for Fig. 10a Values for Fig. 10b

Network area A Rectangle 500 m × 500 m Rectangle 500 m × 500 mSink position Lower left corner (0, 0) Lower left corner (0, 0)

Number ofsensors n

200, 300, 400 300

Nominal trans-mission range R

50 m 50 m

Range forselecting nexthop Rclose

50 m 50 m

Energy availableon sensors Einit

i

20 J 20 J

Discovery rangemultiplier

1, 2, 3, 4 1, 2, 3, 4

Search time ts 75 ms 75 msNumber ofgenerated events

1000 1000

Event genera-tion rate

2 events/s 2 events/s

Simulation time 515 s 515 sFailure rate F 0.0 0.0, 0.6, 0.12, 0.18,

0.24, 0.30, 0.36Metrics Success rate, energy

dissipation, delaySuccess rate, energydissipation, delay, an-nouncements

the failure rate increases, the case of = 3 tends to performalmost as many announcements as the case of = 2, while for = 4 the minimum number of searches is conducted. To ex-plain this fact we need to consider the extensive use of cacheby our protocol. Recall that an announcement is not necessaryif a sensor node has enough neighbor information stored in itscache to construct a new path. Even though as we explained

earlier in the cases of = 3 and = 4 paths are more likely tofail, nodes will also have collected enough information to con-struct a new path without the need to issue an announcement.However, as failure rate increases the information stored in thecache becomes obsolete, thus constructed paths may containfailed nodes, hence more announcements are required to buildan up to date path. This behavior may also contribute to thelower percentage of success rate for = 3 and = 4. Con-sequently, even though in cases of high failure rate the infor-mation provided by the cache mechanism may be outdated forhigh values of , overall CKN still achieves very good fault-tolerance.

6. Closing remarks

In this work, we presented and evaluate a new protocol forenergy efficient, fault-tolerant and scalable dissemination ofdata in Wireless Sensor Networks, that uses local informationregarding the surrounding actual network conditions, acquiredby appropriately varying the range of wireless communication,and then plans an optimized path of pairwise adjacent sensordevices that are used in the forwarding (i.e. propagation) of datatowards the sink. We have implemented the new protocol andconducted an extensive simulation study on networks of largesize to validate its performance and investigate its scalabilityon the size of propagated data. Our results basically show thesuperiority of the protocol over well-established relevant solu-tions in the state of the art, as it achieves high success rates,dissipates fewer energy and delivers messages in shorter peri-ods of time.

We wish to extend our protocol by introducing adaptivemechanisms that will allow sensor devices to self-adjust thevarious parameters (e.g. , ts) in terms of the actual networkconditions. We plan to compare the performance of our proto-col with other, existing protocols and also using different net-work shapes, various distributions used to drop the sensors inthe area of interest.

References

[1] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless sensornetworks: a survey, J. Comput. Networks 38 (2002) 393–422.

[2] C. Alvarez, J. Diaz, J. Petit, J. Rolim, M. Serna, Efficient andreliable high level communication in randomly deployed wireless sensornetworks, in: Third International Mobility and Wireless Access Workshop(MOBIWAC), 2004, pp. 106–110.

[3] T. Antoniou, A. Boukerche, I. Chatzigiannakis, G. Mylonas, S.Nikoletseas, A new energy efficient and fault-tolerant protocol for datapropagation in smart dust networks using varying transmission range,in: 37th Annual Simulation Symposium (ANSS 2004), IEEE Press, NewYork, 2004, pp. 43–52.

[4] J. Aspnes, D. Goldberg, Y. Yang, On the computational complexityof sensor network localization, in: First International Workshop onAlgorithmic Aspects of Wireless Sensor Networks (ALGOSENSORS),Springer, Lecture Notes on Computer Science, vol. 3121, 2004, pp.32–44.

[5] E. Bonabeau, M. Dorigo, G. Theraulaz, Swarm intelligence: from naturalto artificial systems, in: A Volume in the Santa Fe Institute Studies inthe Sciences of Complexity, Oxford University Press, 1999.


[6] A. Boukerche, S. Nikoletseas, Protocols for data propagation in wirelesssensor networks: a survey, in: Wireless Communications Systemsand Networks, Kluwer Academic Publishers, Dordrecht, MA, 2004,pp. 23–51.

[7] A. Boukerche, F. Xin, R. Araujo, P. Patnaik, Local information exchangebased coverage-preserving protocol for wireless sensor networks, in:IEEE International Conference on Communications (ICC), 2006.

[8] A. Boukerche, R. Pazzi, R. Araujo, A supporting protocol to periodic,event-driven and query-based application scenarios for critical conditionssurveillance, in: First International Workshop on Algorithmic Aspects ofWireless Sensor Networks (ALGOSENSORS), Springer, Lecture Noteson Computer Science, vol. 3121, 2004, pp. 137–146.

[9] N. Bulusu, D. Estrin, L. Girod, J. Heidemann, Scalable coordinationfor wireless sensor networks: self-configuring localization systems, in:International Symposium on Communication Theory and Applications(ISCTA), 2001.

[10] C. Busch, M. Magdon-Ismail, F. Sinrikaya, B. Yener, Contention-freeMAC protocols for wireless sensor networks, in: 18th InternationalWorkshop on Distributed Algorithms (DISC), Springer, Lecture Noteson Computer Science, vol. 3274, 2004, pp. 245–259.

[11] I. Chatzigiannakis, T. Dimitriou, S. Nikoletseas, P. Spirakis, Aprobabilistic forwarding protocol for efficient data propagation in sensornetworks, in: Fifth European Wireless Conference on Mobile andWireless Systems Beyond 3G (EW), 2004, pp. 344–350 (Also, in theJournal of Ad-Hoc Networks) 4 (5) (2006) 621–635.

[12] I. Chatzigiannakis, A. Kinalis, S. Nikoletseas, Wireless sensor networksprotocols for efficient collision avoidance in multi-path data propagation,in: ACM Workshop on Performance Evaluation of Wireless Ad Hoc,Sensor, and Ubiquitous Networks (PE-WASUN 2004), 2004, pp. 8–16,Performance Evaluation J. 63 (2006) 1074–1093 (Also, in the SpecialIssue on Performance Evaluation of Wireless Ad Hoc, Sensor, andUbiquitous Networks).

[13] I. Chatzigiannakis, A. Kinalis, S. Nikoletseas, An adaptive powerconservation scheme for heterogeneous wireless sensor networks withnode redeployment, in: 17th Annual Symposium on Parallel Algorithmsand Architectures (SPAA), 2005, pp. 96–105, (Also, in the Theory ofComputing Systems Journal (TOCS): Special Issue on Selected Papersof SPAA05), 2007, to appear.

[14] I. Chatzigiannakis, A. Kinalis, S. Nikoletseas, Efficient and robustdata dissemination using limited extra network knowledge, in: SecondInternational Conference on Distributed Computing in Sensor Systems(DCOSS), 2006, pp. 218–233.

[15] I. Chatzigiannakis, S. Nikoletseas, P. Spirakis, Efficient and robustprotocols for local detection and propagation in smart dust networks,J. Mobile Networks Appl. 10 (1) (2005) 133–149 (Special Issueon Algorithmic Solutions for Wireless, Mobile, Ad Hoc and SensorNetworks).

[16] S. Chen, M. Song, S. Sahni, Two techniques for fast computationof constrained shortest paths, in: IEEE Global TelecommunicationsConference (GLOBECOM), 2004, pp. 1348–1352.

[17] Crossbow technology inc., MICA motes 〈http://www.xbow.com/Products/productsdetails.aspx?sid = 71〉.

[18] J. Diaz, J. Petit, M. Serna, Evaluation of basic protocols for optical smartdust networks, IEEE Trans. Mobile Networks 2 (3) (2003) 186–196.

[19] S. Dolev, T. Herman, L. Lahiani, Polygonal broadcast, secret maturityand the firing sensors, in: Third International Conference on Fun withAlgorithms, 2004, pp. 41–52.

[20] M. Elhadef, A. Boukerche, H. Elkadiki, Performance analysis ofa distributed comparison-based self-diagnosis protocol for wirelessad-hoc networks, in: Ninth ACM International Symposium on Modeling,Analysis and Simulation of Wireless and Mobile Systems MSWiM,Spain, 2006, pp. 165–172.

[21] M. Elhadef, A. Boukerche, H. Elkadiki, Diagnosing mobile ad-hocnetworks: two distributed comparison-based self-diagnosis protocols, in:ACM International Workshop on Mobility Management and WirelessAccess (MobiWac), Spain, 2006, pp. 18–27.

[22] D. Ganesan, R. Govindan, S. Shenker, D. Estrin, Highly-resilient, energy-efficient multipath routing in wireless sensor networks, Mobile Comput.Commun. Rev. 4 (5) (2001).

[23] M.T. Hajiaghayi, N. Immorlica, V.S. Mirrokni, Power optimizationin fault-tolerant topology control algorithms for wireless multi-hopnetworks, in: Ninth ACM/IEEE Annual International Conference onMobile Computing (MOBICOM), 2003, pp. 300–312.

[24] W.R. Heinzelman, A. Chandrakasan, H. Balakrishnan, Energy-efficientcommunication protocol for wireless microsensor networks, in: 33rdIEEE Hawaii International Conference on System Sciences (HICSS),2000.

[25] S. Hollar, COTS Dust, in: MSc thesis, Engineering-MechanicalEngineering, University of California, Berkeley, USA, 2000.

[26] C. Intanagonwiwat, R. Govindan, D. Estrin, Directed diffusion: a scalableand robust communication paradigm for sensor networks, in: SixthACM/IEEE Annual International Conference on Mobile Computing(MOBICOM), 2000, pp. 56–67.

[27] F. Kuhn, T. Moscibroda, R. Wattenhofer, The price of being near-sighted,in: Proceedings of the 17th Annual ACM-SIAM Symposium on DiscreteAlgorithms (SODA), 2006, pp. 980–989.

[28] S. Nikoletseas, I. Chatzigiannakis, H. Euthimiou, A. Kinalis, T. Antoniou,G. Mylonas, Energy efficient protocols for sensing multiple events insmart dust networks, in: 37th Annual Simulation Symposium (ANSS),IEEE Press, New York, 2004, pp. 15–24.

[29] A. Rao, S. Ratnasamy, C. Papadimitriou, S. Shenker, I. Stoica,Geographic routing without location information, in: Ninth ACM/IEEEAnnual International Conference on Mobile Computing (MOBICOM),2003, pp. 96–108.

[30] S. Sahni, X. Xu, Algorithms for wireless sensor networks, Int. J.Distributed Sensor Networks, Invited Paper, Preview Issue (2004) 35–56(also in 1(1) (2005) 35–56).

[31] I. Stojmenovic, X. Lin, Loop-free hybrid single-path/flooding routingalgorithms with guaranteed delivery for wireless networks, IEEE Trans.Parallel Distributed Systems 12 (10) (2001) 1023–1032.

Ioannis Chatzigiannakis is currently a Re-searcher of Research Unit 1 (“Foundations ofComputer Science, Relevant Technologies andApplications”) at the Computer TechnologyInstitute (CTI), Patras, Greece and also an Ad-junct Professor at the Computer Engineeringand Informatics Department of Patras Univer-sity, Greece. His main research interests includeDistributed and Mobile Computing, WirelessSensor Networks, Cooperative Mobile RoboticSystems and Algorithmic Engineering. He re-ceived his B.Eng. degree from the ComputerScience and Engineering Department of the

University of Kent at Canterbury, UK in 1997 and his Ph.D. degree fromthe Computer Engineering and Informatics Department of Patras University,Greece in 2003. He has published scientific articles in international confer-ences and journals. He has participated in EU funded R&D projects andprojects funded by the Private Section.

Athanasios Kinalis graduated from the Depart-ment of Computer Science, University of Ioan-nina in 2002. He was awarded his M.Sc. inComputer Science and Engineering from theDepartment of Computer Engineering and In-formatics, University of Patras in 2005. Cur-rently, he is a Ph.D. student and continues hisresearch at the Department of Computer En-gineering and Informatics, University of Pa-tras. Since 2003, he is also a Junior Researcherat the Research Unit 1 “Foundations of Com-puter Science, Relevant Technologies and Ap-plications” of the Research Academic Computer

Technology Institute (CTI). His main research interests are Distributed Com-puting, Algorithmic Engineering, Large Scale Simulation and Wireless SensorNetworks.

http://www.xbow.com/Products/productsdetails.aspx?sidmathsurround =2ptunhbox voidb@x hbox $mathbin =$71

http://www.xbow.com/Products/productsdetails.aspx?sidmathsurround =2ptunhbox voidb@x hbox $mathbin =$71


Sotiris Nikoletseas is currently an AssistantProfessor at the Computer Engineering andInformatics Department of Patras University,Greece and a Senior Researcher at the Com-puter Technology Institute (CTI). His researchinterests include Algorithmic Techniques inDistributed Computing (focus on wireless sen-sor networks and ad hoc mobile networks),Probabilistic Techniques and Random Graphs(with applications to reliable network comput-ing), Algorithmic Engineering and Large Scale

Simulation. He has co-authored over 80 publications in Journals and refereedConferences, 8 Chapters in Books by major publishers and a Book onthe Probabilistic Method. He has served as a Program Committee Chairof many Conferences and as Editor of Special Issues and Member of theEditorial Board of major Journals. He has co-initiated several internationalevents (including ALGOSENSORS, DCOSS). His research has got more than170 citations currently. He has delivered several invited talks and tutorials.(www.cti.gr/RD1/nikole)

http://www.cti.gr/RD1/nikole

fault-tolerant and efficient data propagation in wireless sensor networks using local, additional...

Documents