Dealing with Nonuniformity in Data Centric Storage for Wireless Sensor Networks

Dealing with Nonuniformity in Data CentricStorage for Wireless Sensor Networks

Michele Albano, Member, IEEE, Stefano Chessa, Member, IEEE,

Francesco Nidito, and Susanna Pelagatti

Abstract—In-network storage of data in Wireless Sensor Networks (WSNs) is considered a promising alternative to external storage

since it contributes to reduce the communication overhead inside the network. Recent approaches to data storage rely on Geographic

Hash Tables (GHT) for efficient data storage and retrieval. These approaches, however, assume that sensors are uniformly distributed

in the sensor field, which is seldom true in real applications. Also they do not allow tuning the redundancy level in the storage according

to the importance of the data to be stored. To deal with these issues, we propose an approach based on two mechanisms. The first is

aimed at estimating the real network distribution. The second exploits data dispersal method based on the estimated network

distribution. Experiments through simulation show that our approach approximates quite closely the real distribution of sensors and

that our dispersal protocol sensibly reduces data losses due to unbalanced data load.

Index Terms—Wireless sensor networks, data centric storage, information dispersal, load balancing.

Ç

1 INTRODUCTION

WIRELESS Sensor Networks (WSNs) [6] are a recenttechnology suitable for unattended monitoring of a

wide range of environments, spanning infrastructures (suchas factories or public buildings), houses or even humans. In aWSN, a set of low-power, inexpensive, embedded devices(called sensors or nodes) spontaneously cooperate to constructa wireless network to support their monitoring activities.Each sensor is a microsystem combined with a radio interfacethat embeds a set of transducers aimed at measuring differentenvironmental parameters. A special sensor, called sink, actsas a gateway with the external networks, and it makes thesensed data available to external users.

In the early approaches, WSNs implemented an ex-ternal storage scheme, for example, using Directed Diffu-sion [12], where all sensed data are sent to the sink to bestored and analyzed outside the WSN. This schemeassumes that the sink has a permanent connection withthe network, and that it performs most of the dataanalyses, while the role of the WSN is limited to dataacquisition. However, the external storage approach is notfeasible in applications where the WSN has an intermittentconnection with the sink. For these reasons, the workdescribed in [21] introduced the Data Centric Storage

model with Geographic Hash Tables (DCS-GHT), in whichdata are stored within the WSN, for the sink to collectthem at a future time, maybe after aggregation orpreprocessing. Comparing this approach to the externalstorage approach, the authors observed that in-networkstorage contributes to save sensors’ energy and to improvenetwork lifetime. Since sensors have limited memorycapacity, the storage of all the sensed data in the WSNmay result impractical, however, with data centric storageit is possible to aggregate data thus reducing their size.

In this paper, we reconsider the DCS-GHT approach todata storage in WSN. In this approach, each datum isassociated with a metadatum; the metadatum is hashed to apair of coordinates ðx; yÞ on the WSN area and the datum isthen stored on the sensors forming a perimeter aroundðx; yÞ. DCS-GHT constructs the perimeter by means of ageographic routing protocol [13], [9], [11]. We first deeplyanalyze the behavior of DCS-GHT through simulation. Inparticular, we analyze the effects of using a uniformlydistributed hashing on nonuniformly distributed sensorsand the effects of using perimeters comprising an unpre-dictable number of sensors. Our results show that, even onuniformly distributed sensors, the amount of stored dataper sensor is extremely variable with DCS-GHT, which maylead to data losses in overburdened sensors. This phenom-enon is even worse if the sensors are not uniformlydistributed. For this reason, we introduce a novel approach,called Load Balanced Data Centric Storage (LB-DCS), to thestorage of sensed data in a WSN. In our approach, sensorsapply a distributed protocol to compute an approximation fof their actual distribution. Then f is used to bias the hashfunction in order to distribute coordinate pairs according tonetwork distribution (more data stored on densely popu-lated areas in the sensing field). Finally, a datum isreplicated according to a QoS level that depends on theimportance of the datum (as decided by the user). Weevaluate thoroughly our approach using simulations based

1398 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 22, NO. 8, AUGUST 2011

. M. Albano is with the Instituto de Telecomunicacoes—Polo de Aveiro,Campus Universitario de Santiago, 3810-193 Aveiro, Portugal, andDipartimento di Informatica—Universita’ di Pisa, Largo B. Pontecorvo3, 56127 Pisa, Italy. E-mail: [email protected].

. S. Chessa is with the Dipartimento di Informatica—Universita’ di Pisa,Largo B. Pontecorvo 3, 56127 Pisa, Italy, and ISTI-CNR, via Moruzzi 1,56124 Pisa, Italy. E-mail: [email protected].

. F. Nidito and S. Pelagatti are with the Dipartimento di Informatica—Universita’ di Pisa, Largo B. Pontecorvo 3, 56127 Pisa, Italy.E-mail: {susanna, nids}@di.unipi.it.

Manuscript received 26 Sept. 2009; revised 1 Apr. 2010; accepted 14 July2010; published online 3 Jan. 2011.Recommended for acceptance by X. Tang.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TPDS-2009-09-0469.Digital Object Identifier no. 10.1109/TPDS.2011.18.

1045-9219/11/$26.00 � 2011 IEEE Published by the IEEE Computer Society

on NS-2 [19]. As compared with DCS-GHT, our approachguarantees a much better load balancing of storage andgreatly reduces the loss of data due to overburdenedsensors. It should be stressed that, although our approachcan be applied to network topologies which changedynamically due to sensors’ movements or failures, ourproposal is thought for networks with limited mobility. Weare planning to extend this approach to WSNs with highermobility in our future work.

The rest of the paper is organized as follows: Section 2presents the related work and briefly describes DCS-GHT.Section 3 introduces the LB-DCS protocol and its mechanismsfor density sampling, metadata hashing, and data dispersal.In Section 4, we evaluate the performance of LB-DCS bymeans NS-2, and Section 5 draws the conclusions. Supple-mentary material is available on the Computer Society DigitalLibrary at http://doi.ieeecomputersociety.org/10.1109/TPDS.2011.18.

2 RELATED WORK

The Data Centric Storage (DCS) [21] defines a paradigmwhere data are stored within the network itself. Inparticular, each datum is associated to a metadatum andthe datum is stored in a set of sensors that is a function ofthe metadatum.

The first proposal of DCS in WSN is with GeographicHash Tables (DCS-GHT) [21]. In DCS-GHT, it is assumedthat the geographic coordinates of sensors are known, andthat each datum is described by a unique metadatum. The setof sensors selected to store a datum is computed by meansof a hash function applied to the corresponding metadatum.This function returns a pair of geographic coordinatesfitting in the area where the sensor network is deployed.

DCS-GHT exploits the primitive put for data storageand get for data retrieval. The put primitive takes in inputa datum d and its metadatum k. By hashing k, it produces apair of coordinates ðx; yÞ and it uses the GPSR routingprotocol [13] to find the sensor closest to the coordinatesðx; yÞ (called home node), and a set of sensors (called homeperimeter) forming a perimeter around ðx; yÞ (the details ofGPSR are presented in Section 2 of the supplementarymaterial. Then, to enforce data persistence against sensors’faults, the home node requires the sensors in the homeperimeter store a copy of ðk; dÞ. The get primitive hashesthe input parameter k (the metadatum) to obtain thecoordinate ðx; yÞ, then, by means of GPSR, it sends arequest for the data with metadatum k to the point ðx; y).When this request reaches the sensors in the homeperimeter around ðx; yÞ, they send back all the data theystore that correspond to k.

Although innovative, DCS-GHT presents a number oflimitations when deployed on real networks. It assumes auniform distribution of sensors and uniformly hashesmetadata on them. Moreover, if the WSN produces a largeamount of data associated to the same metadatum, all suchdata will be stored by DCS-GHT within the same homeperimeter, thus overloading sensors on that perimeter. Toavoid this problem DCS-GHT uses structured replication,which distributes data with the same metadatum more evenlyin the WSN [21]. However, as observed in our previous work

[4] and discussed in Section 3 of the supplementary material,this is not enough to ensure load balancing. In fact, the storageload can become unbalanced even if metadata are balancedand uniformly distributed.

Along this trend of research many alternative DCSmechanisms have been proposed. They are similar toDCS-GHT in the definition of the put and get primitives,but they differ in the internal mechanisms used toimplement routing, data dispersal and storage. In particu-lar, CHR [5] organizes the WSN into clusters of sensors inorder to address scalability issues related to routing andenergy efficiency. GEM [18] constructs a labeled graphspanning the network to assign addresses to the sensors;this enables the routing of data by using such addressesrather than on geographical coordinates. LHR [8] basesrouting on hierarchical location names of the sensors thatare manually assigned. RR [23] associates the data toregions of the WSN rather than to single points in order torelax the requirements for location accuracy. GLS [14]provides cluster-based location services for locating data ornodes in grids and, even if not related to WSNs, itimplements a geographic routing system relying on real-world geographic location information to route its queries.DIM [15] implements a geographic embedding of an indexstructure. It recursively divides the plane to assignaddresses to sensors, then it hashes metadata to thataddress space. Comb-needle [16] differs from the otherapproaches since it does not use metadata. In Comb-needle,each datum is replicated on a number of sensors belongingto a vertical stripe of the WSN deployment area, and theretrieval is ensured since the queries are directed to thesensors belonging to a horizontal stripe of the same area.The positions and sizes of the stripes are optimized toensure that data can be retrieved efficiently.

All these approaches neither consider nonuniformlydistributed WSN, nor consider QoS and load balancing inthe storage. The works in [7], [20] consider nonuniformWSN, however, their focus is on connectivity problems andbroadcast protocols. Effects of nonuniformity in WSN datastorage are taken into account in our preliminary works[2], [4]. In particular, [4] introduces a dispersal strategy thatexploits a nonuniform hash function and that introducesthe concept of QoS in the storage. However, it assumes thatthe network distribution is known a priori and it does notuse any strategy to infer the actual distribution of thesensors. This fact is particularly limiting considering thatthe network distribution may change due to sensor faults.The works in [2] builds over [4] by introducing a primitivemechanism for the estimation of the density of thenetwork. However, this mechanism is intended only as asetup of the network and it is assumed that the networkdensity does not vary with time. In this paper, we removethis assumption by introducing a mechanism for the on-demand estimation of the network density.

3 A NOVEL PROTOCOL FOR DATA CENTRIC

STORAGE

In our previous study [4], we observed that DCS-GHT issubject to load unbalance due to the fact that it relies on

ALBANO ET AL.: DEALING WITH NONUNIFORMITY IN DATA CENTRIC STORAGE FOR WIRELESS SENSOR NETWORKS 1399

the home perimeters for data replication. Since the size ofthe perimeters can be extremely variable, this greatlyaffects the storage load of the sensors, that is alsoextremely variable. In Section 3 of the supplementarymaterial, we show that this load unbalance results inconsistent data leakages, both in uniform and Gaussian-distributed WSN.

This ill behavior of DCS-GHT is mainly due to threeunderlying assumptions:

. the density of the sensors is supposed to be known,hence DCS-GHT does not provide any means toinspect the network density;

. network density is supposed to be constant all overthe WSN area, hence the hash function used by DCS-GHT to map metadata to WSN area locations is ausual uniform hash function; and

. there is no stress on load balancing, hence DCS-GHTselects all the sensors on the home perimeters fordata storage, regardless of their size. Consequently,the set of sensors selected for the storage of data canbe very large or very small, without any means tocontrol the size of this set.

The goal of coping with load balancing and nonuniformWSNs led us to the design of Q-NiGHT [4], [3], that uses ageneralized hash function to select the home node and thatdoes not resort to a home perimeter to select the nodes thatstore data. Moreover, our seminal work [2] used a primitivetechnique to sample and distribute the actual WSN density.This section proposes a more organic approach, called LoadBalanced Data Centric Storage (LB-DCS) to overcome thelimitations of previous DCS systems. LB-DCS exploits threenovel mechanisms to address the issues presented above.In particular:

. a density estimation protocol that provides thesensors with a network density estimation f to beused in the hashing function of metadata in the putand get protocols;

. a hashing module that uses a generalized hashfunction biased with f ; and

. a storage protocol that enforces QoS in the selectionof the sensors for data storage.

Summarizing, a put operation first acquires informationon the actual sensor density in the WSN (as described inSection 3.1), then it selects a home node in a load balancedway (as described in Section 3.2), and finally it selects a set ofnodes close to the home node, whose size is the QoS of themetadatum at hand (as described in Section 3.3). Similarly, aget operation first acquires information about the sensordensity (as described in Section 3.1), then it directs the querytoward the home node of the metadatum, which is computedusing the same generalized hash function used by the putoperation (as described in Section 3.2).

3.1 Sampling Density, and Providing It to theSensors

This section describes how we can estimate networkdistribution and provide it to the sensors needing it. Ingeneral, we can have static networks, in which sensors do notmove during the network lifetime and dynamic networks, in

which density varies over time. For static networks, oncedensity is computed it remains the same for all the networklifetime. In the case of dynamic networks, the lifetime of thenetwork is divided into epochs of fixed time length. Thedensity of the network is estimated at the beginning of eachepoch and when an epoch expires all the density data areconsidered obsolete. The length of an epoch is systemdependent and it may vary according to the network usage,to environmental conditions, to sensor mobility etc. In the restof this section, we detail the protocol assuming a staticnetwork, intending that all the steps needed to perform thenetwork density sampling should be repeated at the begin-ning of each epoch. On the other hand, each sensor uses atimer to understand when an epoch is finished and to discardoutdated data accordingly.

To the purpose of estimating sensor density, the WSN isdivided into n� n non overlapping square regions of side p(without loss of generality, we assume that the WSN area isa square). The point at the center of a region is called watchpoint, and the sensor closest to a watch point acts as asentinel for that region. An example of division of the WSNarea into nine regions is shown in Fig. 1. Here, we spot witha black circle the center of each region (the watch point) andwith a “+” each sensor.

Note that p should be large enough to ensure that thesentinels are not in the radio range of each other, otherwisethe same neighbor could be reported in two differentregions. For large WSNs, p will be, in general, much largerthan the sensors transmission range r.

The election of a sentinel in a region assures that thesentinel is closer than r=2 to the watch point and it works asfollows: First, each sensor computes its distance from thewatch point of the region where it belongs. This can beeasily done if we assume each sensor knows the size ofthe WSN area and the side p of each region. Then, eachcandidate sentinel (i.e., each sensor closer than r=2 to thewatch point of its region) broadcasts its coordinates and itsid to all its neighbors. As all candidate sentinels are withindistance r, they all receive the coordinates and id of theother candidates and compute the closest one. If two ormore candidates have the same distance from the watchpoint the smaller id wins.

After the election, each sentinel broadcasts a request toits neighbors to count them. The number of neighbors isthen used as an estimation of the local density in the region.Either proactive or reactive mechanisms can be used todeliver the estimates computed by sentinels. We considerone proactive protocol (Broadcast), and two reactive proto-cols (Stripes and FatStripes).


Fig. 1. WSN partitioning in squared regions.

In Broadcast, each sentinel sends its estimate to all thesensors in the network. This is done once for all at thestartup of each epoch.

In reactive protocols, when a sensor needs to perform aput/get operation it queries all the sentinels in the WSN toget their local density estimation. We can reach eachsentinel by sending a message toward the watch point ofits region using GPSR. If the closest node is a sentinel, itreplies with its local density estimation; otherwise (i.e., if itis farther than r=2 from the region watch point) it sendsback a negative answer and the corresponding region isassumed to have local density equal to zero.

In Stripes, each sensor along the unicast route back froma sentinel caches the density estimation received (Fig. 2).When a query for a sentinel arrives to a sensor, it firstchecks its cache. If an entry for that sentinel is present, thesensor sends back the cached density to the querying nodewithout forwarding the query any further. Otherwise, itsends the request toward the sentinel using GPSR.

FatStripes is a step forward in optimization with respectto Stripes. As depicted in Fig. 2, in Stripes only the relaynodes of GPSR cache the estimation received on the wayback from a sentinel. However, all the nodes in thetransmission range have actually spent energy to receivedensity information, even if they did not participate activelyin the protocol. FatStripes uses this observation to cachedensity information also on the passive receivers as in Fig. 3.We can observe that caching stripes are now larger than withStripes, thus increasing the probability of a hit during aquery. Note that the use of caches reduces the number ofmessages required in the protocol at the price of devotingsome memory to cache density estimates on each sensor. Ifan estimate is stored in b bytes and recalling that n2 is thenumber of sentinels, we need n2 � b bytes of memory tocache estimates on each sensor. As the accuracy of the globalestimate increases with the number of sentinels, for eachapplication there is a trade-off between the memory used forcaching and the accuracy required for density estimates.

Once a sensor has collected estimates for densities fromall sentinels, it needs to figure out the density of the entirenetwork. Here, we propose a simple algorithm to “rebuild”density from samples.

The rebuilding algorithm acts in two steps. The first stepcomputes an initial approximation for sensor density in eachregion. The second step refines the density approximationtaking into account approximations in neighbor regions.

Since we have n2 regions, we denote with wij the densityestimate provided by the sentinel in region ij (1 < i; j < n),and with d0ij and dij the density approximations for region ijcomputed by the first step and by the second step of thealgorithm, respectively.

The first step computes each d0ij as

d0ij ¼wijPij wij

:

This first approximation works quite nicely if sentinelsknow a close estimate of the number of nodes in theirregion. However, it can be misleading at least in twoopposite situations:

. false zeroes: if a region has got very sparse nodesnear the watch point and a concentration of manynodes along the border it can report a zero or verylow density which is not representative of thewhole region.

. over reporting: if a region has got very sparse nodenear the borders and the majority of the nodes nearthe watch point, it can report a much higher densitythan the real one.

To cope with these two problems, in the second step wecompute the final approximation dij as a weighted mean ofapproximations computed in the first step for region ij andfor its neighboring regions. The idea is to trust more d0ij thanthe neighbor approximations, thus if region ij has m

neighbors and we denote with Nij the set of indices ofneighbors regions, the second step computes the finalapproximation dij as

dij ¼m�d0ij þ

Pi0;j02Nij

d0i0j0

2�m :

The approximation of the network distribution is definedas the matrix D ¼ ðdijÞn�n of the estimated distributions.

Relative merits of Broadcast, Stripes, and FatStripesand effectiveness of rebuilding algorithm are assessed inSection 4.

3.2 Load Balancing the Choice of the Home Node

This section describes the hashing module that is employedby the DCS system to map each metadatum to the sensorthat acts as the home node for it. Traditional approaches areprone to load unbalance, since standard hash functionsdistribute the home nodes uniformly over all the WSN area,and hence sensors in crowded regions are assigned asmaller number of metadata.

On the other hand, once the network distribution isknown, it is possible to use a generalized hash function todistribute home nodes, thus scattering data approximatelywith the same distribution of the sensors. This way thehome node density is proportional to density of sensors ineach region, and the mean number of metadata assigned toeach sensor tends to be balanced over the WSN.


Fig. 2. Nodes that cache WSN density when Stripes is used. Fig. 3. Nodes that cache WSN density when FatStripes is used.

A generalized hash function hðk; fÞ is intuitively apseudorandom number generator, that receives in inputa probability distribution (the network distribution f) and aseed (the metadatum k) and produces an output coordinateðx; yÞ. The use of a pseudorandom number generatorensures that different keys are mapped w.h.p. on pairs thatare not correlated to each other. The pseudocode for thehash function used in this module is described in Table 1. Ituses a strategy similar to the one used in the rejection method[17] to produce random numbers following any probabilitydistribution in a limited domain. The probability function fis normalized such that the maximum value of f is one, thena triple ðx; y; zÞ is generated uniformly at random, ðx; yÞbeing a valid coordinate in the WSN area and z 2 ½0; 1�. Ifz < fðx; yÞ, the coordinate ðx; yÞ is accepted and returned.Otherwise, other triples ðx; y; zÞ are generated untilz < fðx; yÞ. The number of iterations of this method is inprinciple not limited, however in all of our experiments themethod has ended in a few iterations.

Clearly, the probability of some value ðx; yÞ to bereturned by the generalized hash function is proportionalto fðx; yÞ (that is, to the network distribution in ðx; yÞ),hence load balancing the number of assigned metadata overthe sensor nodes.

Note also that f can be obtained directly from matrix D.In particular, letting ij be the region that contains the pointðx; yÞ; f is obtained as fðx; yÞ ¼ dij. Hence to store f issufficient to store the matrix D of n2 values, one for eachregion. The memory overhead of the sensor grows linearlywith the number of watch points (corresponding to thenumber of regions), because in the worst case each sensorhas to cache the estimated density produced by each watchpoint. However, as discussed in Section 4.1, the error in thedensity decreases rapidly with the number of watch points,thus a limited number of regions (in our simulations around25) is sufficient to attain a good trade-off between theprecision of the protocol and the memory overhead.

3.3 Enforcing QoS in the Storage

Similarly to DCS-GHT, LB-DCS is built atop GPSR and itoffers the primitives put and get for data storage andretrieval. The put primitive takes three parameters: adatum d, its metadatum k, and a QoS parameter q thatexpresses the level of dependability required for the datumd. The parameter q may be expressed using different metricsand ranges according to the particular redundancy techni-que used. In our case, since we use data replication as

redundancy technique, q expresses the required number ofreplicas of the datum d. In particular, q ranges in ½1; qmax�,where qmax is the maximum number of replicas admitted.Note, however, that other redundancy techniques are alsopossible. For example, [1] investigates the use of erasurecodes in data centric storage.

Let s be the source node of a put(d; k; q) operation. sfirst computes ðx; yÞ ¼ hðk; fÞ as the destination of thepacket Pp ¼ <ðx; yÞ; <d; k; q>>. The packet in turn is sent tothe destination using GPSR. As in DCS-GHT, we call homenode the sensor H (of coordinates ðxH; yHÞ), which isgeographically nearest to the destination coordinates. Thus,H receives the packet as a consequence of applying GPSR.Upon the reception of packet Pp;H begins the dispersalprotocol, which selects a set of q sensors (called the replica set)to store the replicas of (k; d). The replica set includes H andit is unique for a given key k.

The dispersal protocol is iterative and uses the concept ofball. Given the home node H of coordinates ðxH; yHÞ, wedenote with BðxH;yH ÞðrÞ the ball centered in ðxH; yHÞ of radiusr, that is the set of sensors that are within a euclideandistance r from ðxH; yHÞ. Thus

BðxH;yH ÞðrÞ ¼ fsensors of coordinateðx; yÞ :

jðxH; yHÞ; ðx; yÞj < rg:

Sensor H then sends a request for storage to all thesensors within the ball. In turn, when a sensor in the ballreceives the request for storage of Pp, it acknowledges therequest to H. H accepts the q � 1 acknowledgmentsreceived from the nearest sensors, it confirms them, anddisregards the others. The confirmation requires an extrapacket sent by H. Sensors receiving the confirmation keepthe data while the others disregard them after a timeout. IfH receives q0 < q acknowledgments, then it executesanother iteration of the dispersal protocol with r ¼ 2r inwhich it considers only the sensors in Bðx0;y0Þð2rÞ �Bðx0;y0ÞðrÞ.The dispersal protocol stops as soon as q sensors have beenhired or the outermost perimeter has been reached. In thespecial case where at least q � 1 sensors are in thetransmission range of H;H computes the set of the sensorsin BðxH;yH ÞðrÞ, inserts their identifiers into the request forstorage, and only these sensors acknowledge the request forstorage and store the datum; this way, not only thedispersal protocol performs only one iteration, but thenumber of acknowledgments generated is limited to q � 1.It can be observed that our dispersal protocol is based on ageomulticasting protocol [22].

The complexity of the put protocol clearly depends onthe choice of r as this determines the number of iterationsmade to successfully place the q replicas. However, sincewe know the distribution of sensors f , for any given homenode H of coordinate ðxH; yHÞ and q it is possible to fix r insuch a way that, with high probability, at least q sensorsbelong to the ball BðxH;yH ÞðrÞ.

When a sensor s0 of coordinates ðx0; y0Þ executes get(k),it first computes ðx; yÞ ¼ hðk; fÞ, and sends a query packetPg ¼ <ðx; yÞ; <ðx0; y0Þ; k>> by means of GPSR. If the net-work topology is not changed since the execution of the putfor the same metadatum (i.e., the sensors have neithermoved nor failed) then GPSR guarantees that the querypacket will eventually reach the sensor closest to ðx; yÞ, i.e.,the sensor H that was selected by the put protocol. Note,


TABLE 1Rejection Hash

however, that in some cases the packet may reach anothersensor in the replica set before reaching H. In any case,either H or the sensor in the replica set of k receiving thequery packet will respond to s0 by sending back all thestored data that match with the metadata k.

However, if some sensors have failed, GPSR may fail toreach H and the other sensors in the replica set. Never-theless, LB-DCS can ensure the retrieval of data stored withQoS parameter q provided that up to q � 1 sensors in theWSN have failed and that the WSN is still connected (andhence the success of the get(k)). In particular, weanalyzed the cases which may prevent the retrieval of datastored with QoS parameter q when up to q � 1 sensors havefailed, and, by simulation, we proved that these cases areextremely unlikely. This result is discussed in detail inSection 6 of the supplementary material.

Note that the capability of LB-DCS to retrieve stored data isalso affected by sensors’ mobility that may significantlychange the network topology. To deal with mobility, weconsider the use of a Periodical Refresh Protocol (PRP) similarto the one introduced in [21]. Section 4 of the supplementarymaterial discusses the application of PRP to LB-DCS and itsperformance.

4 SIMULATION

This section presents the simulation results on the cost and theperformance of the density sampling protocols (Section 4.1),on the performance of the rejection hash technique in thedispersal of data in WSN with uniform and Gaussiandistributions (Section 4.2) and on the cost of the put andget protocols of LB-DCS and DCS-GHT (Section 4.3). To thispurpose, we have implemented both LB-DCS and DCS-GHTin the NS-2 simulator [19].

4.1 Simulations on Density Sampling

This section presents the simulation results on the error inthe approximation of the network density computed byLB-DCS and on the cost of Broadcast, Stripes, andFatStripes protocols.

For each simulation run, 100 WSNs are randomlygenerated, and the three protocols (Broadcast, Stripes, andFatStripes) are run on them to evaluate the error in thedensity estimation and the messages exchanged by theseprotocols. The simulation iterates the runs until the outputsof the simulator reach a 99 percent confidence interval that

is less than 1 percent. We report here the main resultsobtained with 200 sensors with a transmission range of 25 min a WSN area of 100� 100 m2, and network density(expressed as the average number of neighbors per sensor)equal to 39. Note that we obtained similar results fordifferent sizes of the WSN area and/or network density.

In Fig. 4, we show the errors measured by approximatingan uniform, Gaussian, and Hill distribution when varyingthe number of regions used. We measure the error using theMean Square Error (MSE), a scalar quantifying the distancebetween the real distribution and the distribution computedby the rebuilding algorithm. If we denote with DR ¼ðdRijÞn�n, the matrix of real region densities (and we recallthat D ¼ ðdijÞn�n is the matrix of the estimated regiondensities), the MSE is defined as the average of the square ofthe errors on each region,

MSEðDÞ ¼P

ij

�dij � dRij

�2

n2:

The error is under 0.0035 with nine regions and falls under0.0001 with 25 regions. A detailed evaluation of the errordue to the rebuilding of a Gaussian distribution with 10�10 regions is given in Section 5 of the supplementarymaterial.

Fig. 5 compares the behavior of Broadcast, Stripes, andFatStripes in terms of the average number of messagesgenerated, with respect to different numbers of sensors thatquery the sentinels. Since Broadcast is proactive, its perfor-mance is independent of the number of sensors’ requests: allthe sensors receive the sentinels’ data ahead of time, so no real“request” is generated. Stripes has a good behavior when asmall number of sensors ask for WSN density, but it suffersfrom its “unicast” communication when most of the sensorsrequest this information. On the other hand, FatStripes getsthe best of both worlds, since it is reactive and hence it is cheapwhen a few sensors query the sentinels, but it fully exploitsthe broadcasting characteristic of the physical medium todisseminate density information as much as possible. Hence,when many sensors request density information, most ofthem already have it because of the communication per-formed by past requests.

A detailed evaluation of the number of sent and receivedmessages of Broadcast, Stripes, and FatStripes can be foundin Section 5 of the supplementary material.


Fig. 4. Mean Square Error of estimated distribution (a scalar) measuredfor different values of p.

Fig. 5. Number of messages sent, against number of sensors thatrequested WSN density, 100� 100 m2 WSN area, 200 nodes.

4.2 Simulating Rejection Hash

The simulation results presented in this section are aimed atevaluating the data leakage of LB-DCS and to compare itwith the data leakage of DCS-GHT, which is analyzed indetail in Section 3 of the supplementary material. Dataleakage occurs when a sensor serves too many put

operations and it saturates its memory. In the simulations,we consider WSNs deployed in a square area of200� 200 m2, a transmission range of 10 m, and networkdensities (expressed as average number of neighbors persensor) in the range ½7; 30�. The simulations consider bothuniform and Gaussian distribution of the sensors. In thecase of LB-DCS, we also set the QoS parameter q ¼ 7.

We assume sensors with storage capacity of 512 KB (as itis the case of the Crossbow Mica family [10]). For eachnetwork generated in the experiments, the simulatorexecutes a number of put operations for each sensor, eachone accounting for 8 bytes (i.e., each put operation requiresthe storage of 8 bytes of data in each sensor in the homeperimeter or in the replica set). In these simulations, weassume that each sensor produces a total of 2,100 put

operations during its lifetime, i.e., each sensor produces anamount of data to be stored that corresponds approxima-tively to 1/30 of its memory capacity. We assume that, oncea sensor is requested to store a new datum but its memoryis full, it drops an older datum. This means that once asensor reaches its maximum storage capacity, it startsdropping data whenever it is requested to store new data.

With these settings, we performed the following experi-ment. In each simulation run, the simulator generates a newnetwork (according to the chosen network distribution), itsimulates all the put operations, and it computes thenumber of sensors that do not leak data and the totalquantity of data lost. The simulation iterates until theaverage number of sensors that leak data and the averagenumber of lost data reaches a 99 percent confidence intervalthat is less than 1 percent.

The result of these simulations is that with LB-DCS anegligible fraction (less than 0.00001) of the nodes leakssome data, and that a negligible fraction (less than 0.00001)of the data is lost. On the contrary, the fraction of data lostby DCS-GHT is around 0.7 (more details are available inSection 3 of the supplementary material and in [4]). Thismeans that, under this respect, LB-DCS significantly out-performs DCS-GHT.

Further simulations were thus performed to understandthe reasons of this behavior. The additional simulations

evaluate the average load of the sensors in WSNs withnetwork density equal to 14 and both uniform and Gaussiandistribution. Figs. 6 and 7 report, for different values of theload on the x-axis, the average number of sensors that havethat storage load in the cases where the sensors aredistributed according to a uniform distribution or to aGaussian distribution, respectively. The load on the x-axisrepresents the number of data stored by a sensor normalizedwith respect to the data that a sensor produces. In particular,a sensor with load equal to 1 stores the same amount of data itproduces (i.e., approximately 1/30 of its memory capacity).

The results of the two cases are similar. In particular,DCS-GHT shows a great load unbalance (a significantnumber of sensors have a heavy load).

On the other hand, LB-DCS evenly distributes the loadaccording to the estimated network distribution due to therejection hash. In particular, with uniform distribution mostof the sensors have a limited load (around seven data), andthe average number of sensors that have the higher load(which in this case is of 17 data) is very small (about 0.05).Similarly, in the case of Gaussian distribution, most of thesensors have a limited load (around 8 data), and the averagenumber of sensors with a high load (which in this case is of18 data) is very small (about 0.5).

4.3 Cost of the put and get Operations

This section evaluates the cost of the put and get

operations of DCS-GHT and of LB-DCS in terms of MAClayer send and receive operations.

In these simulations, we used the same parameters as inSection 4.2: WSN area of 200� 200 m2, communicationrange of 10 m, WSN density in the range ½7; 30�, and QoSparameter of LB-DCS set to q ¼ 7. Each simulation runiterates 1,000 put operations, each followed by a get

operation on the same metadatum. The simulator reportsthe average number of packet forwarded and received bythe sensors for each operation. The simulation runs arerepeated until the simulator outputs reach a 99 percentconfidence interval that is less than 1 percent. The simulatoralso reports on the correctness of each pair of put and get

operations, in particular, it computes the number of getoperations that were unable to retrieve the values storedwith the corresponding put. However, in our simulationexperiments the get was always successful.

The simulation results on the cost of the put and get arereported in Figs. 8 and 9. In particular, Fig. 8 reports the


Fig. 6. Stored data per node for LB-DCS, uniform distribution of sensors. Fig. 7. Stored data per node for LB-DCS, Gaussian distribution ofsensors.

number of MAC-level send for the storage and retrievaloperations of DCS-GHT and of LB-DCS, respectively, whileFig. 9 reports the number of MAC-level receives. From thesimulations, it results that the put operation is moreexpensive than the get operation. This happens becauseonce the put request has reached the home node, it shouldalso reach all the nodes in the home perimeter (in the case ofDCS-GHT) or in the replica set (in the case of LB-DCS), whichcan be quite large depending on the DCS scheme used and onthe desired redundancy of the data. Furthermore, it is seenthat the put operation is less expensive with LB-DCS thanDCS-GHT, due to the smaller cost incurred by the dispersalprotocol, while the cost of get is almost the same for the twoprotocols.

5 CONCLUSIONS

DCS systems are very effective in implementing an in-network data storage and retrieval system, since theyrequire only unicast communications. However, existingapproaches disregard issues related to load balancing of thesensors and QoS. In this paper, we have addressed suchaspects by proposing a new DCS system, LB-DCS, thatrelies on three mechanisms: a network density estimationprotocol, a rejection hashing technique that produces pairsof coordinates by taking into account the real networkdistribution, and a dispersal protocol that enforces QoS andload balancing. The simulation results show that ourapproach significantly balances the storage load on the

sensors and it adapts to different network distributions. Asfuture directions, we believe that having the opportunity ofestimating on the fly the distribution of the sensors may beexploited also in a better balancing of the routes and in theevaluation of the coverage of the sensing tasks. Future workincludes also the extension of our approach to release theassumption of limited mobility of the sensors.

ACKNOWLEDGMENTS

This work was supported in part by the ICT project ICT-248577 C2POWER, which is funded by the European Union.

REFERENCES

[1] M. Albano and S. Chessa, “Distributed Erasure Coding in DataCentric Storage for Wireless Sensor Networks,” Proc. 14th IEEESymp. Computers and Comm. (ISCC), pp. 67-75, 2009.

[2] M. Albano, S. Chessa, F. Nidito, and S. Pelagatti, “Data CentricStorage in Non-Uniform Sensor Networks,” Grid-Enabled RemoteInstrumentation, F. Davoli, N. Meyer, R. Pugliese, andS. Zappatore, eds., pp. 3-19, Springer, 2009.

[3] M. Albano, S. Chessa, F. Nidito, and S. Pelagatti, “Q-Night:Adding QoS to Data Centric Storage in Non-Uniform SensorNetworks,” Technical Report 06-16, Dipartimento di Informatica,Universita di Pisa, 2006.

[4] M. Albano, S. Chessa, F. Nidito, and S. Pelagatti, “Q-NiGHT:Adding QoS to Data Centric Storage in Non-Uniform SensorNetworks,” Proc. Eighth Int’l Conf. Mobile Data Management(MDM), pp. 166-173, 2007.

[5] F. Araujo et al., “CHR: A Distributed Hash Table for Wireless AdHoc Networks,” Proc. 25th IEEE Int’l Conf. Distributed ComputingSystems Workshop, 2005.

[6] P. Baronti et al., “Wireless Sensor Networks: A Survey on the Stateof the Art and the 802.15.4 and ZigBee Standards,” ComputerComm., vol. 30, no. 7, pp. 1655-1695, 2007.

[7] C. Bettstetter, “The Cluster Density of a Distributed ClusteringAlgorithm in Ad Hoc Networks,” Proc. IEEE Int’l Conf. Comm.,pp. 4336-4340, 2004.

[8] F. Bian, R. Govindan, S. Schenker, and X. Li, “Using HierarchicalLocation Names for Scalable Routing and Rendezvous in WirelessSensor Networks,” Proc. Second ACM Int’l Conf. EmbeddedNetworked Sensor Systems (SenSys ’04), pp. 305-306, 2004.

[9] P. Bose, P. Morin, I. Stojmenovic, and J. Urrutia, “Routing withGuaranteed Delivery in Ad Hoc Wireless Networks,” WirelessNetworks, vol. 7, no. 6, pp. 609-616, 2001.

[10] Crossbow Technology: http://www.xbow.com, 2011.[11] H. Frey and I. Stojmenovic, “On Delivery Guarantees of Face and

Combined Greedy-Face Routing in Ad Hoc and Sensor Net-works,” Proc. ACM MobiCom, pp. 390-401, 2006.

[12] C. Intanagonwiwat, R. Govindan, and D. Estrin, “DirectedDiffusion: A Scalable and Robust Communication Paradigm forSensor Networks,” Proc. ACM MobiCom, pp. 56-67, 2000.

[13] B. Karp and H.T. Kung, “GPSR: Greedy Perimeter StatelessRouting for Wireless Networks,” Proc. ACM MobiCom, pp. 243-254, 2000.

[14] J. Li, J. Jannotti, D.S.J. De Couto, D.R. Karger, and R. Morris, “AScalable Location Service for Geographic Ad Hoc Routing,” Proc.ACM MobiCom, pp. 120-130, 2000.

[15] X. Li, Y.J. Kim, R. Govindan, and W. Hong, “Multi-DimensionalRange Queries in Sensor Networks,” Proc. First ACM Int’l Conf.Embedded Networked Sensor Systems (SenSys ’03), pp. 63-75, 2003.

[16] X. Liu, Q. Huang, and Y. Zhang, “Combs, Needles, Haystacks:Balancing Push and Pull for Discovery in Large-Scale SensorNetworks,” Proc. Second ACM Int’l Conf. Embedded NetworkedSensor Systems (SenSys ’04), pp. 122-133, 2004.

[17] J.V. Neumann, “Various Techniques Used in Connection withRandom Digits,” Collected Works, A.H. Taub, ed., vol. 5, pp. 768-770, Pergamon Press, 1951.

[18] J. Newsome and D. Song, “GEM: Graph Embedding for Routingand Data-Centric Storage in Sensor Networks without GeographicInformation,” Proc. First ACM Int’l Conf. Embedded NetworkedSensor Systems (SenSys ’03), pp. 76-88, 2003.

[19] Network Simulator 2 (ns-2): http://nsnam.isi.edu/nsnam/, 2011.


Fig. 8. MAC-level sends for put and get in DCS-GHT and LB-DCS.

Fig. 9. MAC-level receives for put and get in DCS-GHT and LB-DCS.

[20] L. Orecchia, A. Panconesi, C. Petrioli, and A. Vitaletti,“Localized Techniques for Broadcasting in Wireless SensorNetworks,” Proc. Joint Workshop Foundations of Mobile Computing(DIALM-POMC ’04), 2004.

[21] S. Ratnasamy, B. Karp, S. Shenker, D. Estrin, R. Govindan, L. Yin,and F. Yu, “Data-Centric Storage in Sensornets with GHT, aGeographic Hash Table,” Mobile Networks and Applications, vol. 8,no. 4, pp. 427-442, 2003.

[22] K. Seada and A. Helmy, “Efficient and Robust GeocastingProtocols for Sensor Networks,” Computer Comm., vol. 29, no. 2,pp. 151-161, 2006.

[23] K. Seada and A. Helmy, “Rendezvous Regions: A ScalableArchitecture for Service Location and Data-Centric Storage inLarge-Scale Wireless Networks,” Proc. 18th Int’l Parallel andDistributed Processing Symp., 2004.

Michele Albano received the BSc degree inphysics, in 2003, and the BSc, MSc, and PhDdegrees in computer science, in 2004, 2006,and 2010, respectively, from the University ofPisa, Italy. He was a visiting researcher atUniversidad de Malaga, in 2007, at StonyBrook University, in 2009, and before being aresearcher he worked as a software engineerand wireless technologies specialist in privatecompanies in the period 2001-2006. In 2006

and 2007, he was involved in EU funded projects SMEPP andXtreemOs, and he is now a postdoc researcher at the Instituto deTelecomunicacoes—Polo de Aveiro, Portugal, working on EU fundedproject C2POWER. He has coauthored more than 25 paperspublished on international journals and conference proceedings, heis reviewer for several international journals and conferences, and heis an editor for European Transactions on Telecommunications andfor the Social Technologies Journal. His main research interests arein the areas of wireless networks and peer-to-peer networks. He is amember of the IEEE.

Stefano Chessa received the MSc and PhDdegrees in computer science from the Universityof Pisa, Italy, in 1994 and 1999, respectively.Since 2000, he is an assistant professor at theComputer Science Department of the Universityof Pisa and since 2003, he is also a researchassociate at the ISTI/CNR Institute (InformationScience and Technology Institute). He has beeninvolved in many national and European projects.In particular, he has coordinated the CNR project

Management of Data in Wireless Sensor networks (MaD-WiSe). He hasalso participated to the EU FP6 SatNex, SMEPP, InterMedia, PERSONAprojects and to the EU FP7 universAAL and Rubicon projects. He hascoauthored more than 80 papers published on international journals andconference proceedings; he is reviewer for several international journalsand conferences, and he has been member of several internationalprogram committees of conferences and workshops. His researchinterests are in the areas of wireless ad hoc networks, and wirelesssensor networks, security in wireless communications, and system-leveldiagnosis. He is a member of the IEEE.

Francesco Nidito received the PhD degree, in2008, from the University of Pisa, Departmentof Computer Science. He performed his re-search activity at the same University and as avisiting scholar at the Northeastern University,Boston (Massachusetts) collaborating with pro-fessor Stefano Basagni and Andras Farago.After the PhD, he has been working in thesearch engines field, first at Ask.com and thenat the Microsoft Search Technology Center of

London (United Kingdom).

Susanna Pelagatti received the MSc and PhDdegrees in computer science from the Universityof Pisa, Italy, in 1987 and 1993, respectively. In1995, she joined the University of Pisa, Diparti-mento di Informatica as an assistant professor.Since 2002, she is an associate professor in thesame department. Her main research interestsare in the areas of parallel computing, wirelessand ad hoc networks, and sensor networks. Shehas been involved in national and international

projects and she has coauthored more than 50 papers published ininternational journals and conference proceedings. She has beenmember of international program committees of conferences andworkshop.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Date post:	13-Dec-2016
Category:	Documents
Upload:	susanna
View:	213 times
Download:	0 times

Dealing with Nonuniformity in Data Centric Storage for Wireless Sensor Networks

Documents